INTERNATIONAL TRADE, FOREIGN DIRECT INVESTMENT

AND PRODUCTIVITY: AN EMPIRICAL INVESTIGATION

by

JOEL RODRIGUE

A thesis submitted to the

Department of Economics

in conformity with the requirements for

the degree of Doctor of Philosophy

Queen’s University

Kingston, Ontario, Canada

June, 2008

Copyright © Joel Rodrigue, 2008

Abstract

This dissertation investigates the effects of foreign direct investment (FDI) and interna-

tional trade on firm level decisions and examines the effect of policy changes on aggregate

productivity and welfare in open economies. The first two chapters build an open economy

model of FDI and trade where heterogeneous firms make simultaneous foreign investment

and export decisions. The theoretical model is estimated using Indonesian manufacturing

data. The estimated model is then used to perform a variety of counterfactual experiments

to assess the positive and normative effects of international barriers to trade and FDI. I

find that the impact of FDI on aggregate productivity is at least three times the impact of

international trade on aggregate productivity.

The third chapter, co-authored with Hiroyuki Kasahara, investigates whether importing

intermediate goods improves plant performance. While addressing the issue of simultaneity

between productivity shocks and the decision to import intermediates, we estimate the

impact of the use of foreign intermediates on plants’ productivity using plant-level Chilean

manufacturing panel data. We find that by importing foreign intermediates, manufacturing

plants can improve productivity.

i

Co-Authorship

Chapter 4 of this thesis was co-authored with Hiroyuki Kasahara.

ii

Dedication

To my wife Patricia: For your constant support, encouragement and love. This would not

have been possible without you.

iii

Acknowledgments

I am greatly indebted to Beverly Lapham and Hiroyuki Kasahara for their guidance, sup-

port, insights and encouragement. I am also very grateful for the generous support offered

by Gregor Smith throughout both my MA and PhD degrees.

I thank the the seminar participants of the Numerically Intensive Economic Policy Anal-

ysis Meetings, Midwest Trade Meetings, Midwest Macroeconomics Meetings, Canadian

Economic Association Meetings, North American Summer Meetings of the Econometric

Society, Empirical Investigations in International Trade Conference, Laurier Conference on

Empirical International Trade, Queen’s University, Bank of Canada, University of Western

Ontario, Pennsylvania State University, McMaster University, Carleton University, Univer-

sity of Ottawa, University of Victoria, HEC Montreal, Arizona State University, Purdue

University, Vanderbilt University, Iowa State University and Duke University.

Finally, I very grateful for to my wife, sons, and parents for their love and, most of all,

their patience.

iv

Table of Contents

Abstract i

Co-Authorship ii

Dedication iii

Acknowledgments iv

Table of Contents v

List of Tables viii

List of Figures xi

Chapter 1:

General Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Chapter 2:

Foreign Direct Investment, Exports and Aggregate Productivity . . 6

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2 Empirical Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.3 A Model of FDI and Exports . . . . . . . . . . . . . . . . . . . . . . . . . 16

v

2.4 Empirical Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2.5 Estimation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

2.6 Conclusions and Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . 56

Chapter 3:

Foreign Direct Investment, Exports and Aggregate Productivity:

Sunk vs. Fixed Costs . . . . . . . . . . . . . . . . . . . . . . . . . . 58

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

3.2 An Augmented Cost Structure . . . . . . . . . . . . . . . . . . . . . . . . 60

3.3 Structural Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

3.4 Estimation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

3.5 Conclusions and Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . 77

Chapter 4:

Does the Use of Imported Intermediates Increase Productivity?

Plant-Level Evidence . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4.2 The Theoretical Framework . . . . . . . . . . . . . . . . . . . . . . . . . . 82

4.3 Econometric Specification . . . . . . . . . . . . . . . . . . . . . . . . . . 86

4.4 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

4.5 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

4.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

Chapter 5:

Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . 106

vi

Appendix A:

Appendix for Chapter 2 . . . . . . . . . . . . . . . . . . . . . . . 117

A.1 Transition Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

A.2 Counterfactual Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 118

A.3 Fixed Cost Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

A.4 Additional Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

A.5 Measurement of Aggregate Productivity and Welfare . . . . . . . . . . . . 126

A.6 Export and Ownership Premia: Fixed Effects and the Chemicals Industry . . 128

A.7 Mark-Ups & Productivity . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

A.8 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

A.9 Variation in Plant-Level Productivity . . . . . . . . . . . . . . . . . . . . . 134

Appendix B:

Appendix for Chapter 3 . . . . . . . . . . . . . . . . . . . . . . . 137

B.1 Transition Probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

B.2 Counterfactual Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 139

B.3 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

Appendix C:

Appendix for Chapter 4 . . . . . . . . . . . . . . . . . . . . . . . 143

C.1 Estimation Procedures: Selection and Adjustment Costs . . . . . . . . . . . 143

C.2 Additional Estimation Results . . . . . . . . . . . . . . . . . . . . . . . . 149

vii

List of Tables

2.1 Foreign Plants Over Time . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.2 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.3 Export & Ownership Premia . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.4 Production and Sales Combinations . . . . . . . . . . . . . . . . . . . . . 24

2.5 Indonesian Export Destinations . . . . . . . . . . . . . . . . . . . . . . . . 42

2.6 1993-96 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . 43

2.7 Structural Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

2.8 Distribution of Plants by Ownership/Export Status . . . . . . . . . . . . . . 45

2.9 Export/Domestic Market Share by Ownership/Export Ownership . . . . . . 47

2.10 Average Productivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

2.11 Transition Probabilities - Textiles . . . . . . . . . . . . . . . . . . . . . . . 52

2.12 Counterfactual Experiments - Textiles . . . . . . . . . . . . . . . . . . . . 53

3.1 Structural Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

3.2 Distribution of Plants by Ownership/Export Status . . . . . . . . . . . . . . 69

3.3 Export/Domestic Market Share by Ownership/Export Ownership . . . . . . 70

3.4 Average Productivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

3.5 Transition Probabilities - Textiles . . . . . . . . . . . . . . . . . . . . . . . 75

3.6 Counterfactual Experiments - Textiles . . . . . . . . . . . . . . . . . . . . 75

viii

4.1 Descriptive Statistics in 1980 . . . . . . . . . . . . . . . . . . . . . . . . . 92

4.2 Transition Probability of Import Status and Exit . . . . . . . . . . . . . . . 93

4.3 Estimates of Production Function: Discrete Import Variable . . . . . . . . . 95

4.4 Estimates of Production Function: Continuous Import Variable . . . . . . . 100

4.5 OLS Regression of TFP on Import and Export: Discrete Variables . . . . . 103

4.6 OLS Regression of TFP on Import and Export: Continuous Variables . . . . 103

A.1 Transition Probabilities - Food . . . . . . . . . . . . . . . . . . . . . . . . 117

A.2 Transition Probabilities - Metals . . . . . . . . . . . . . . . . . . . . . . . 118

A.3 Counterfactual Experiments - Food . . . . . . . . . . . . . . . . . . . . . . 119

A.4 Counterfactual Experiments - Metalsa . . . . . . . . . . . . . . . . . . . . 120

A.5 Export & Ownership Premia: Output per Worker 1993-1996 . . . . . . . . 129

A.6 Export & Ownership Premia: Chemicals . . . . . . . . . . . . . . . . . . . 130

A.7 Top Corporate Tax Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

A.8 Mark-Ups Across Firms and Industries . . . . . . . . . . . . . . . . . . . . 132

A.9 Structural Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

A.10 Counterfactual Experiments - Food . . . . . . . . . . . . . . . . . . . . . . 135

A.11 Counterfactual Experiments - Metalsa . . . . . . . . . . . . . . . . . . . . 135

A.12 Counterfactual Experiments - Textiles . . . . . . . . . . . . . . . . . . . . 135

A.13 Revenue Bracket Transition Matrix - Domestic Plants . . . . . . . . . . . . 135

A.14 Revenue Bracket Transition Matrix - Foreign Plants . . . . . . . . . . . . . 136

B.1 Transition Probabilities - Food . . . . . . . . . . . . . . . . . . . . . . . . 137

B.2 Transition Probabilities - Metals . . . . . . . . . . . . . . . . . . . . . . . 138

B.3 Counterfactual Experiments - Food . . . . . . . . . . . . . . . . . . . . . . 139

B.4 Counterfactual Experiments - Metalsa . . . . . . . . . . . . . . . . . . . . 140

ix

B.5 Structural Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

B.6 Counterfactual Experiments - Food . . . . . . . . . . . . . . . . . . . . . . 141

B.7 Counterfactual Experiments - Metalsa . . . . . . . . . . . . . . . . . . . . 142

B.8 Counterfactual Experiments - Textiles . . . . . . . . . . . . . . . . . . . . 142

C.1 Additional Estimates of Production Function: Discrete Import Variable . . . 151

C.2 Estimates of Production Function for Food and Metal Industries . . . . . . 152

C.3 Panel OP/LP Estimates: Energy vs. Materials . . . . . . . . . . . . . . . . 155

C.4 Descriptive Statistics in 1980 (Extended Sample) . . . . . . . . . . . . . . 158

C.5 Transition Probability of Import Status and Exit (Extended Sample) . . . . 160

C.6 Export and Import Status Change . . . . . . . . . . . . . . . . . . . . . . . 160

x

List of Figures

2.1 Foreign Country Thresholds . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.2 Home Country Thresholds . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.3 Productivity Distributions of Domestic Entrants and Incumbents (Actual

vs. Predicted) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

2.4 Productivity Distributions of Foreign Entrants and Incumbents (Actual vs.

Predicted) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

2.5 Productivity Distributions of Domestic and Foreign Firms by Export Status

(Actual vs. Predicted) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

3.1 Productivity Distributions of Domestic Entrants and Incumbents (Actual

vs. Predicted) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

3.2 Productivity Distributions of Foreign Entrants and Incumbents (Actual vs.

Predicted) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

3.3 Productivity Distributions of Domestic and Foreign Firms by Export Status

(Actual vs. Predicted) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

4.1 Productivity Dynamics and Import Status . . . . . . . . . . . . . . . . . . 98

4.2 Productivity Dynamics Heterogeneity across Different Import Shares . . . . 101

A.1 Indonesian Firms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

xi

A.2 Foreign Firms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

A.3 Food . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

A.4 Metals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

A.5 Textiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

C.1 Monotonicity Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157

C.2 Productivity Dynamics for the 50 Percentile Importing Plants . . . . . . . . 159

xii

Chapter 1

General Introduction

Merchandise trade worldwide has grown by more than three hundred percent over the past

twenty years. The growth in worldwide trade has been intimately related to the diffusion

of foreign direct investment and the trade in intermediate inputs, which have both grown

at even faster rates over the same time period (UNCTAD, 2002). Consequently, the in-

ternational trade literature has sought to incorporate these developments into trade theory

and empirically quantify their size and importance. My dissertation adds to this literature

in three separate but related studies investigating the effects of foreign direct investment

(FDI) and international trade on firm level decisions and examining the effect of policy

changes on aggregate productivity and welfare in open economies.

It is well-documented that there exists substantial firm level productivity differences

across firms in the same narrowly defined industries (see Clerides, Lach and Tybout (1998)

and Aw, Chung and Roberts (2000), for examples). The most productive firms in an in-

dustry are typically those that interact with markets abroad, either through foreign direct

investment or international trade (Arnold and Javocik (2005)). As such, policies that en-

courage the most productive firms in an economy to grow into international markets have

1

the potential to raise industry productivity by reallocating resources towards the most pro-

ductive firms in an industry. As empirically shown by Baily, Hulten, and David (1992) and

Pavcnik (2002), it is vital to examine plant-level changes in order to understand changes

in aggregate productivity levels. Furthermore, recent developments in trade theory suggest

that understanding the plant-level response to trade policy is a crucial factor in understand-

ing its impact on aggregate productivity and welfare (e.g., Melitz, 2003; Bernard, Eaton,

Jensen and Kortum, 2003).

Similarly, if a plant benefits from foreign inputs or knowledge, trade liberalization may

also improve firm-level and aggregate productivity. For instance, countries that are more

open to trade may benefit more from foreign R&D because they are better able to access

improvements in technology. Importing intermediate goods is a potential example of such

a mechanism. Among recent papers discussing the impact of foreign intermediate inputs

on productivity at the micro-level are van Biesebroeck (2003), Muendler (2004), Amiti

and Koenings (2005), and Halpern, Koren, and Szeidl (2006). The empirical findings in

the literature are mixed. While van Biesbroeck (2003) and Muendler (2004) find little

evidence for strong productivity improvements at the plant level through imported inter-

mediates, both Amiti and Konings (2005) and Halpern, Koren and Sziedl (2006) find that

there are substantial productivity benefits to the use of the imported intermediates at the

plant-level. This dissertation adds to this literature by investigating the effects of foreign

direct investment (FDI) and international trade on firm level decisions in Indonesia and

Chile and examines the effect of policy changes on aggregate productivity and welfare in

open economies.

2

Chapter 2 builds an open economy model of international trade and foreign direct in-

vestment where heterogeneous firms make simultaneous foreign investment and export de-

cisions. The model extends the Melitz, Helpman and Yeaple (2003) international trade and

FDI environment to capture export-platform FDI. Specifically, in my model, firms invest in

foreign countries not only as a means to access foreign markets but also to potentially take

advantage of lower production costs.

The theoretical model is estimated using Indonesian plant-level manufacturing data and

broadly replicates many of the features of the Indonesian manufacturing sector. The esti-

mated model is then used to perform a variety of counterfactual experiments to assess the

positive and normative effects of international barriers to trade and investment. I find that

the impact of FDI and international trade on aggregate productivity ranges from seventeen

to thirty-eight percent across the food, metals and textiles industries. While there exists

substantial differences across industries, the impact of FDI on aggregate productivity is

at least three times the impact of international trade on Indonesian productivity. In gen-

eral, the results imply that FDI policy has a much larger impact on aggregate productivity

relative to trade policy in Indonesia.

In contrast, the estimates imply that FDI and trade restrictions reduce aggregate welfare

by one to one and a half percent across industries. The explanation for the small impact on

welfare is due to the fact that FDI and trade flows partially insure each other. For example,

when firms face higher trade (FDI) costs, they are more likely to access the Indonesian

market through FDI (trade). As such, the Indonesian consumer is partially insured from

adverse changes in trade (FDI) policy by the existence of the FDI mechanism. Moreover,

my results imply that models without both FDI and trade will tend to overstate the welfare

costs associated with increases in trade or FDI restrictions.

3

Chapter three extends the empirical model presented in the preceding chapter to cap-

ture sunk export and FDI costs. The model’s predictions demonstrate that the additional

dynamic features of the model improve the fit of the model relative to the observed data and

are necessary to match plant level dynamics in Indonesia. The estimates maintain the result

that the impact of FDI on aggregate productivity is several times greater than the impact

of international trade. I find that the model with sunk export and FDI costs suggests that

accounting for the dynamic nature of entry and exit in international markets can amplify

the impacts of FDI and trade policy on aggregate productivity. Due to sunk costs, foreign

firms are even less likely to exit Indonesia due to change in trade policy. This mitigates the

fall in aggregate productivity due to a rise in trade barriers. The impact of FDI and inter-

national trade on aggregate productivity ranges from thirty-one to forty-one percent across

the food, metals and textiles industries. Again, the impact of FDI on aggregate productivity

is at least thirteen times the impact of international trade on Indonesian productivity while

the welfare costs of restricting trade and FDI are always less than one and a half percent

across industries.

The fourth chapter, co-authored with Hiroyuki Kasahara, examines whether importing

intermediate goods improves plant performance. While addressing the issue of simulta-

neous productivity shocks and decisions to import intermediates, we estimate the impact

foreign intermediates have on plants’ productivity using plant-level Chilean manufacturing

panel data. The results across different estimators indicate a statistically significant, often

substantial, positive impact from the use of imported intermediates on plant productivity.

The point estimates suggest a positive and significant impact from the use of imported in-

termediates on productivity of at least 2.6 percent across different estimators. In addition,

4

we find some evidence for a positive dynamic effect from the use of imported intermedi-

ates, or “learning by importing." Our results suggest that past import status has a positive

impact on current productivity. We also examine the sensitivity of our results to export sta-

tus and time-varying industry-specific effects. Overall, the results indicate the robustness

of the effect of intermediate imports on output across various estimation methods.

The remainder of the thesis proceeds as follows: Chapter two presents a simple model

of exports and foreign direct investment, structurally estimates the theoretical model and

presents the counterfactual policy experiments. Chapter three extends the model presented

in Chapter two to allow for sunk export and FDI costs and highlights the differences in

model performance and counterfactual results. The fourth chapter provides a model of

import behaviour at the plant-level and empirically estimates this relationship between im-

ports and plant-level productivity in Chile. The fifth chapter concludes.

5

Chapter 2

Foreign Direct Investment, Exports and

Aggregate Productivity

2.1 Introduction

This chapter builds an open economy model of international trade and foreign direct in-

vestment with heterogeneous firms. The model extends Melitz (2003) by allowing firms to

offshore production to foreign low-wage countries. While firms can set up plants in for-

eign countries in order to access foreign markets as in Helpman, Melitz and Yeaple (2004),

the model also allows firms to locate plants abroad in order to export back to the country

of origin. The model highlights how differences in production, location and export costs

affect the structure of international trade. The theoretical model is estimated using Indone-

sian plant-level manufacturing data. The estimated model is then used to perform a variety

of counterfactual experiments to assess the positive and normative effects of international

barriers to trade and foreign direct investment.

Recent evidence suggests that the current volume and direction of trade are intimately

6

related to multinational production decisions. It is estimated that total sales from multina-

tional firms accounts for 60% of world GDP and over 35% of world trade in 2001.1 At

the same time, empirical studies have repeatedly confirmed that trade can substantially im-

pact resource allocation across firms and aggregate productivity within an industry. Both

Pavcnik (2002) and Trefler (2004) find that increasing openness induces productivity gains

among exporters, while Pavcnik (2002) also finds increases in aggregate productivity.2 Yet,

few studies have examined the role of foreign direct investment decisions (FDI) on trade

flows, resource allocation and their effect on aggregate productivity.

A separate but related set of empirical evidence suggests that among firms producing in

a given country, multinational firms are the most productive. Domestic exporters are report-

edly less productive than multinationals, but significantly more productive than domestic

non-exporters.3 In this paper, I confirm these productivity differences across foreign and

domestic firms in Indonesia and provide new evidence emphasizing important differences

across foreign owned firms. The data suggests that firms who invest in Indonesia to serve

the Indonesian market are substantially more productive than those who invest in Indonesia

as a platform for exports.

Economists have traditionally separated FDI into horizontal and vertical components.

Horizontal FDI, the largest of the two components, usually refers to firms which set up af-

filiate plants in multiple countries to serve the domestic market in each of those countries.

By investing in a foreign country, a firm saves the export transportation costs, but incurs1Ramondo (2006).2Pavcnik (2002) finds substantial productivity gains due to resource reallocation from less to more pro-

ductive after trade liberalization in Chile. Similarly, Trefler (2004) finds productivity improvements amongboth exporters and importers after the Canada-U.S. free trade agreement.

3See Helpman, Melitz and Yeaple (2004) for a comparison of multinationals, domestic exporters anddomestic non-exporters. Arnold and Javocik (2005) provide evidence that foreign owned firms are moreproductive than domestic firms. Similarly, Aw, Chung, and Roberts (2000), Bernard and Jensen (1999),Bernard et. al. (2003), Clerides, Lack and Tybout (1998), and Eaton, Kortum, and Kramarz (2004) find thatmore productive firms are more likely to export.

7

additional fixed production costs by operating multiple plants (c.f. Brainard (1997), for ex-

ample). Helpman, Melitz and Yeaple (2004) extend the traditional proximity-concentration

framework to examine the choice between exports and horizontal FDI in an economy where

firms are characterized by heterogeneous productivity. They demonstrate that only the most

productive firms will produce in multiple countries. Most importantly, they show that al-

lowing productivity to vary across firms generates richer trade patterns and provides an

explanation for the co-existence of exports and multinational production across developed

countries.

In contrast, vertical FDI often represents the production of goods in foreign countries

for re-export back to the domestic parent. Though smaller in aggregate than horizontal FDI,

vertical FDI is often reported to be growing quickly, particularly in developing countries.4

Typically, vertical FDI is driven by a desire to take advantage of low foreign production

costs in an environment where contracting with local producers is either impossible or sub-

optimal. For example, Antras and Helpman (2004) combine the insight of Melitz’s (2003)

heterogeneous firms framework with the vertical integration model of Antras (2003) to

explore the effects of incomplete contracts and multiple stage production.

Several studies include both types of foreign investment. However, models encompass-

ing the multistage nature of vertical FDI, along with the proximity-concentration trade-

off of horizontal FDI are inherently complex. Ekholm et al. (2003), Yeaple (2003) and

Grossman, Helpman and Szeidl (2006) illustrate that as one varies the determinants of in-

ternational trade or investment,5 it is possible to generate a myriad of different equilibria.6

4Yeats (2001) argues that trade in intermediate inputs has grown much faster than trade in final goods.Moreover, he estimates that trade in intermediates accounts for 30% of world trade in manufactures. Sim-ilarly, Hummels, Ishii and Yi (2001) assert that within firm transfers of goods across national borders ac-counted for one-third of world export growth between 1970 and 1990.

5Such as transport costs, set up costs, country size, wages, etc.6The Grossman, Helpman and Szeidl (2006) model is particularly appealing because it includes firm level

8

Unfortunately, none of these models are well suited to empirically examine the plant-level

decisions to export and/or invest abroad.7

One objective of this paper is to provide a framework that is rich enough to describe

the principal channels for international capital and trade flows, while remaining suitable

for estimation using plant-level data. The advantage of this approach is that all of the

model’s predictions can be readily tested. Moreover, the model is used to empirically

assess the influence of policy on plant-level decisions, and thus provides a framework for

evaluating economic policy and predicting changes in aggregate productivity, exports and

foreign investment.8

I begin by extending the international trade and investment framework of Helpman,

Melitz and Yeaple (2004) by allowing firms to offshore production in a foreign country.

In my environment, firms may set up plants in foreign countries for two reasons. First,

as in Melitz, Helpman and Yeaple (2004), firms can set up plants solely to access the lo-

cal market in foreign countries. Relative to a model without FDI, such as Melitz (2003),

FDI provides an additional avenue to for firms to make sales to foreign consumers. Sec-

ond, firms can set up plants in a foreign country in order to export back to the country of

origin. While there are a number of papers that address offshoring,9my model simultane-

ously allows firms to engage in horizontal FDI. I examine how the structure of fixed and

transport costs affect the firm level location and export decisions across high and low wage

heterogeneity and captures a wide variety of firm structures and trade flows across countries. Moreover, it cangenerate predictions for firm behaviour in relation to the determinants of trade and investment. Unfortunately,while it is possible to test the model’s aggregate implications across countries, it is would be practicallyimpossible to evaluate the plant level predictions using even the most detailed modern data sets.

7As noted by Grossman, Helpman and Szeidl (2006), the distinction between horizontal and vertical FDIis increasingly blurred. Feinberg and Keane (2003) report that 69% of all multinational firms in the US haveelements of both horizontal and vertical FDI. A similar pattern is found here for Indonesia.

8It is important to note, however, that a disadvantage of this approach is that it requires abstracting fromthe details of contractual arrangements that cannot be observed. To the extent that incomplete information isimportant in determining the types the structure of FDI, the estimates may be biased.

9See Antras and Helpman (2004) and Grossman, Helpman and Sziedl (2006) for examples.

9

countries. I show that my model can generate productivity differences across plants with

different ownership and export status which are consistent with the observed differences in

the Indonesian manufacturing data.

Using the theoretical model and a panel of Indonesian manufacturing plants, I develop

and estimate a structural empirical model of exports and FDI. Indonesia is a country of

particular interest because it is one of the largest economies in South East Asia and one

of the largest hosts of multinational corporations (Ramstetter and Sjoholm (2006)).10 The

estimated model captures the pattern of productivity, exports and market share across firms

with different ownership and export status. I find that the mean of domestic productiv-

ity at the steady state is substantially higher than the estimated mean at entry, indicating

that endogenous exit decisions play an important role in determining aggregate productiv-

ity. Moreover, to examine the effects of trade and foreign investment policies, I perform a

number of counterfactual policy experiments which demonstrate that restrictions on invest-

ment and trade can potentially have widely different effects on aggregate productivity. I

find that the impact of FDI restrictions account for a fall in average total factor productivity

between 8 and 27 percent across industries. Trade restrictions, in contrast, are estimated to

have a smaller impact on average productivity. Across the food, manufactured metals and

textiles industries average total factor productivity is estimated to fall by 1 to 4 percent.

Since foreign and domestic plants respond differently to policy change, failing to account

for these differences will lead to biased estimates of the impact of trade restrictions. The

results suggest that policies which induce inwards flows of FDI will have a much larger im-

pact on aggregate productivity relative to those that encourage exports. The welfare impact

on Indonesia of simultaneous trade and FDI restrictions are estimated to be approximately10In fact, the World Bank (1993) identified rapid industrialization, export growth and inflows of foreign

direct investment as key elements to the accelerated growth experienced by Indonesia and surrounding coun-tries in the preceding years.

10

1% across industries. However, when trade or FDI are individually restricted, the welfare

impacts are very small. This last result is because trade (FDI) flows provide some insurance

against FDI (trade) restrictions.

The next section outlines the differences across and within foreign and domestic pro-

ducers. The third section presents a theoretical model of exports and FDI with heteroge-

neous firms and countries, while the fourth describes the model’s empirical analog. The

fifth section presents empirical results and the sixth concludes.

2.2 Empirical Motivation

In this section I briefly describe the Indonesian manufacturing data and provide summary

statistics to characterize patterns across plants with varying degrees of international inte-

gration.

2.2.1 Data

I use the Indonesian manufacturing census for 1993-1996. The data set reports the total

value of domestic and export sales, the percentage of foreign ownership and the number

of workers. I determine the export status of each firm by checking whether the value of

export sales is positive or zero. Likewise, I determine the ownership status of each firm by

checking whether foreign investors hold a positive amount of equity in the plant.11 Nominal

values are converted to real values using the industry output price deflators. I identify the

entry/exit decisions by observing the number of workers across years. After cleaning the

data, I use an unbalanced panel of 24,519 plants. Each plant is observed for at least one

11Using other threshold values (e.g. 10% or 50% of equity) yielded similar results.

11

year between 1993 and 1996.12

2.2.2 Exports and Foreign Ownership

It is well known that multinational and/or foreign-owned firms are typically the largest firms

in a country. Table 2.1 documents this fact for Indonesian manufacturing plants between

1993-1996. Although only six percent of all firms have any foreign ownership, foreign

firms account for more than one quarter of total output and over one third of all exports

in manufacturing. Moreover, foreign firms are not solely export oriented but also capture

one quarter of the Indonesian domestic market for manufactured goods. This suggests

Indonesia may also be an important market for foreign plants.

Table 2.1: Foreign Plants Over Time1993 1994 1995 1996 1993-1996 avg.

Exports 0.30 0.38 0.35 0.38 0.36Output 0.23 0.27 0.29 0.34 0.28Domestic Market Share 0.20 0.22 0.27 0.31 0.25% of Firms 0.06 0.06 0.06 0.06 0.06

Notes: The top three rows document the percentage of the total manufacturing sales attributed to foreign firms. The bottom rowdocuments the percentage of foreign manufacturing plants.

A common explanation for these findings is that multinational or foreign firms are sub-

stantially larger and more productive than their domestic counterparts.13 Comparing the

the second and third rows with the fourth and fifth rows of Table 2.2 it is evident that for-

eign plants are not only much larger than domestic firms but they also appear to be more

capital-intensive, use a higher fraction of skilled employees and produce more output per

worker than their domestic counterparts.

12I omit 729 plants which are wholly owned by the government.13See Helpman, Melitz and Yeaple (2004) or Arnold and Javorcik (2005) for an example.

12

Similarly, it is often stated that domestic exporters are relatively large, capital and skill-

intensive, and more productive relative to domestic non-exporters.14 The bottom two rows

of Table 2.2 are consistent with this result for the Indonesian manufacturing sector. Al-

though the Indonesian data confirms that foreign plants are always larger and more produc-

tive than their domestic counterparts with the same export status, the disparity in differences

across export groups is striking. For instance, while foreign exporters are approximately

twice as large and productive as domestic exporters, foreign non-exporters are ten to fifteen

times larger than domestic non-exporters and produce eight times the output per worker.

Table 2.2: Descriptive StatisticsTotal Export Labor Skill Capital K/L Output/ No. of

Salesa Intensitya,b Ratioc Ratioa Worker Obs.All 35.47 0.70 185.57 0.14 6.00 0.02 0.12 72,732Plants (261.88) (0.33) (613.13) (0.15) (22.00) (0.03) (0.59) —Foreign 166.38 0.71 700.14 0.18 30.13 0.06 0.35 2,563Exporters (381.88) (0.34) (1,047.91) (0.17) (47.30) (0.05) (0.58) —Foreign 167.76 — 411.23 0.28 29.09 0.09 0.48 1,803Non-Exporters (361.14) — (649.24) (0.20) (49.08) (0.06) (1.08) —Domestic 81.14 0.70 445.81 0.15 14.53 0.03 0.16 10,784Exporters (495.38) (0.33) (1,192.23) (0.14) (37.84) (0.03) (0.37) —Domestic 11.77 — 74.25 0.10 1.81 0.01 0.06 57,582Non-Exporters (139.75) — (286.21) (0.14) (8.43) (0.02) (0.50) —

Notes: Reported number are sample means with standard deviations in parentheses. (a) In millions of Indonesian Rupiahs. (b) Computedusing the sample of exporting plants. (c) The ratio of skilled workers to total workers.

The differences between foreign exporters and non-exporters are less obvious. Several

studies have previously examined the differences across a variety of attributes for domes-

tic exporters and non-exporters (e.g. Bernard and Jensen, 1999 or Kasahara and Lapham,

2007), but few have examined differences across ownership status,15 and even fewer have

14See Kasahara and Lapham (2006) for an example.15A notable study examining the differences across ownership status using the Indonesian manufacturing

census is Arnold and Javorcik, 2005.

13

examined differences across foreign firms. Although foreign exporting plants employ al-

most twice as many workers, they earn little more revenue, employ a smaller percentage

of skilled workers, are less capital-intensive and produce less output per worker than for-

eign non-exporters on average. The difficulty in comparing foreign exporters and non-

exporters arises from the fact that exporters serve multiple markets from one plant while

non-exporters only serve the Indonesian market. On average foreign exporting plants sold

63.5 million rupiahs worth of goods to the Indonesian market per year. Thus, foreign non-

exporters sold approximately 2.5 times more on average than foreign exporters sold to the

Indonesian market.

Table 2.3: Export & Ownership PremiaPooled OLS: 1993-1996

Export/Ownership Status Domestic Exporters Foreign Exporters Foreign Non-ExportersOutput per Worker 0.217 0.710 0.871

(0.012) (0.021) (0.024)Average Wage 0.114 0.443 0.604

(0.007) (0.013) (0.015)Non-Production/Total Workers 0.086 0.006 0.246

(0.010) (0.018) (0.021)Capital per Worker 0.351 0.712 0.932

(0.011) (0.019) (0.022)Domestic Sales -0.789 -0.043 0.848

(0.016) (0.030) (0.029)Total Sales 0.250 0.737 0.883

(0.014) (0.024) (0.027)Total Employment 1.062 1.528 1.045

(0.010) (0.019) (0.022)No. of Observations 72732

Notes: Standared errors are in parentheses.

While the mean differences between firms with different export and ownership status

are striking, it is not clear that they are statistically significant. In fact, the standard devia-

tions are often large, particularly for foreign firms where there are fewer observations. Fol-

lowing Bernard and Jensen (1999) and Kasahara and Lapham (2007) I estimate the export

and ownership premia using a pooled ordinary least squares regression over the 1993-1996

14

period:

ln Xit = α0 + α1dxit(1− dfdi

it ) + α2dxitd

fdiit + α3(1− dit

x)dfdiit + Zitβ + εit, (2.1)

where Xit is a vector of plant attributes such as total employment, domestic sales, out-

put per worker, average wages, the percentage of non-production workers and capital per

worker. A firm’s export status is captured by the dummy variable dxit, while dfdi

it is a dummy

variable capturing whether any of the plant’s equity is held by foreign investors. Last, Zit

is matrix of control variables including industry dummies at the 5-digit ISIC level, year

dummies, province dummies, municipality dummies and total employment to control for

size. The export premium α1 is the average log point difference between exporters and

non-exporters among domestic plants. The foreign exporter premium α2 is the average log

point difference between domestic non-exporters and foreign exporters, while the foreign

non-exporter premium α3 is the average log point difference between domestic and foreign

plants who do not export.

The results in Table 2.3 show that there are not only substantial differences between do-

mestic exporters and non-exporters but also between foreign exporters and non-exporters.

Column 3 indicates that foreign non-exporters tend to demonstate the highest premia across

all measures other than those related directly to size. In particular, foreign non-exporting

15

plants tend to demonstrate higher productivity foreign exporters or domestic plants.16 Sim-

ilar results are found at the industry level.17 It is natural to ask why the larger, more produc-

tive foreign non-exporters do not export while the smaller, less productive foreign exporters

do. A likely reason is that foreign non-exporters are part of a larger multinational corpo-

rations that serve multiple markets though multiple plants located in different countries. I

will further develop this hypothesis to explain the effects of foreign barriers to trade and

investment across different plants in Indonesia.18

2.3 A Model of FDI and Exports

The model extends the environment of Helpman, Melitz and Yeaple (2004) to include FDI

and exports back to the country of origin to capture vertical FDI.19

16Another possibility is that differences in prices across countries causes spurious correlation in our mea-surement of productivity across firms. In particular, if mark-ups are higher in domestic markets than exportsmarkets the productivity measurement of foreign exporters would be biased downwards. In contrast, it islikely that the mark-up in export markets is higher than in domestic markets due to income differences acrosscountries. This would imply that Tables 2.2 and 2.3 underestimate the differences between foreign exportersand non-exporters. Table A.8 in Appendix A reports imputed “approximate mark-ups” across producers andindustries. A last possibility is that foreign firms use transfer pricing to earn profits abroad due to high corpo-rate tax rates. However, the World Tax Database reports that over 1993 to 1996 the top Indonesian corporatetax rates were slightly lower than those in Japan, the United Kingdom and the United States on average. TableA.7 in Appendix A reports top corporate tax rates across countries.

17With the exception of the chemicals industry. The chemicals industry in Indonesia tends to display adistinctly different pattern where foreign exporters are the most productive firms. This is not surprising sincethe chemicals industry is dominated by fuel and natural gas related products for which there are large exportmarkets. Results for the chemicals industry are reported in Appendix A.

18Although the estimates between foreign exporters and non-exporters in Table 2.3 are statistically signif-icant, these differences could be driven by unobserved plant-specific differences. Discussion of fixed effectsestimation in this context and results controlling for plant-specific heterogeneity are reported in Appendix A.

19I exclude strict vertical integration here. I do not consider the case where foreign firms set up a subsidiaryor subcontract with a domestic firm solely to supply the parent firm further up the supply chain. The reasonfor this choice is two-fold. First, strict vertical integration represents at most 2% of domestic firms and 7-9%of foreign firms. Second, strict vertical integration likely involves more complex contractual arrangementswhere incomplete information between plants may play a larger role in determining the direction of FDI andtrade. See Antras and Helpman (2004) for an example.

16

2.3.1 Environment

Consider two countries, Home and Foreign, which are endowed with non-depreciating

stocks of labour, L.20 In each country there are two sectors: a homogeneous good sector

(agriculture) and a differentiated good sector (manufacturing).

2.3.2 Consumers

Consumers supply capital and labour inelastically. Their preferences are defined by a Cobb-

Douglas utility function over the agriculture product z and a continuum of manufactured

goods indexed by v:

U = z1−δ

[∫v∈V

q(v)αdv

]δ/α

.

The elasticity of substitution between different varieties of manufactured goods is given by

ε = 1/(1− α) > 1.21

2.3.3 Producers

Agriculture

There is a continuum of potential firms that can freely enter the agricultural sector and

produce a homogeneous agricultural product, z, with linear technology, z = φll. Producers

hire labour on perfectly competitive markets which pins down wages in each country. I

assume that the foreign agricultural technology is more productive than the technology

employed in the home country, φ∗l > φl, so that wages in the foreign country are greater

than those in the home country.

20All foreign country variables are starred.21The set of available goods is denoted by V and all goods are substitutes, 0 < α < 1.

17

Manufacturing

The model allows for a wide variety of potential outcomes across countries. To keep the

model as transparent as possible, I focus on the case in which both foreign and domestic

exporters and non-exporters are present. Specifically, I allow each country to have four

types of firms: non-exporters, exporters, multinational non-exporters and multinational

exporters who are owned by either domestic residents or those of the other country.22

I denote variables for non-exporting plants by d and exporting plants by x. Following

the nomenclature in the literature I will refer to multinational exporters as “vertical multi-

nationals” and will denote their variables by v while I refer to multinational non-exporters

as “horizontal multinationals” and will denote their variables by h.23

To enter the market each firm must pay a fixed cost, fe. Once the entry cost is paid,

each firm receives a productivity draw a from the distribution, Ga(a), and an extreme cost

shock with constant probability ξ. A firm’s productivity is constant over the life of the firm.

If the firm suffers the extreme cost shock it is forced to exit the industry regardlessly of its

productivity.

Conditional on survival, each firm can decide to exit immediately or to produce accord-

ing to the linear production function:

q =l

a(2.2)

where l is labour hired on competitive markets. It is well known that the Dixit-Stiglitz

(1977) framework generates a demand function γRP 1−ε/pε for each variety where R is

the total revenue earned in each country, P is an index of manufacturing prices and p is the

22I exclude the possibility of joint ownership.23I exclude the possibility that a firm owned by foreign investors will produce in the home country solely

for the home market without producing in the home market. While it is possible to extend the model to allowfor this to occur, I omit it here for tractability. See Appendix A for conditions under which no firm willoptimally choose to only operate abroad for the foreign market.

18

price chosen by each individual producer.24 Since demand is exogenous to each individual

producer the optimal pricing rule for each firm depends only the firm-specific productivity

level:

p(a) =aw

α

where w is the wage in the home country.

To produce domestically, each firm must pay a fixed overhead cost f each period. If

the firm also decides to export abroad it bears additional fixed cost fx and iceberg transport

costs τ > 1 per unit shipped to the foreign country. As in Helpman, Melitz and Yeaple

(2004) the firm may choose to set up production in the foreign country rather than export to

that market. While the firm saves on the fixed export costs, fx, and on the transport costs,

τ , by choosing to produce abroad, it incurs the additional fixed overhead cost, ffdi.25 In my

model, firms may choose not to produce domestically at all. In this case, each firm sets up a

plant abroad and exports back to its country of origin.26 The firm then incurs the fixed costs

f , fx and ffdi and the transport cost τ . If ffdi > 0, any firm that produces abroad to export

back home incurs higher fixed costs and the same transport costs of a home exporter. Thus,

to give firms an incentive to produce abroad and export there must be some difference in

factor prices across countries.

I interpret fx as the cost of forming and maintaining distribution and service networks

in export markets. It is important to note that firms must pay this cost even if they are

exporting back to their home market.27 Similarly, ffdi includes the distribution, servicing,

24R = wL where L is the total amount of labour and w is the wage in the home country. The price indexis P = [

∫V

p(v)1−εdv]1/(1−ε)

25In the multi-country setting each firm would pay fx or ffdi for each export/production destination.26This can be thought of as “offshoring”. However, as described below, no firm will ever choose to produce

domestically and solely export back to its country of origin. Although it is possible to extend the model toinclude this possibility, for simplicity I omit it here.

27The empirical model allows for differences in fixed export costs across firms.

19

overhead and set up costs of operating a subsidiary abroad. The relative size of the fixed

costs plays an important role in determining each firm’s optimal production and export

decisions.

2.3.4 FDI and Exports

I assume that each firm that chooses not to exit will always serve the domestic market. It

accomplishes this through domestic production or exports from the foreign country. Sim-

ilarly, firms will never choose to produce solely for export. As argued below, casual em-

piricism suggests that a small number of firms produce solely for export markets.28 In

equilibrium, I rule out the possibility that multinational firms produce solely for the foreign

market or that any firm will serve the same market by both exports and FDI.29

The operating profits for any firm with productivity level a producing in the home

country solely for domestic consumption are

πd = (awB)1−ε − f, (2.3)

where B = [(1 − α)γwL]1/(1−ε)/(αP ). Since market size, B, is positively related to

country size L, larger countries will have a larger market, for a given price level. In addition

to profits earned at home, exporters earn profits from selling abroad. Thus, the operating

profits for an exporter are

πx = πd(a) + (τawB∗)1−ε − fx. (2.4)

Similarly, firms that produce in both countries for each domestic market earn the additional

28It would be possible to generate production solely for export by allowing fixed costs to vary acrosscountries.

29Rob and Vettas (2001) use uncertainty in a dynamic setting to explain this phenomenon.

20

operating profits from their plants abroad. The operating profits for a horizontal multina-

tional are

πh = πd(a) + (aw∗B∗)1−ε − (f + ffdi). (2.5)

Lastly, vertical multinationals do not produce at home for the home market, but instead

produce abroad for both home and foreign markets. The profits these firms earn in either

country will depend upon the factor prices in the foreign country. The operating profits for

a vertical multinational are

πv = (aw∗B∗)1−ε + (τaw∗B)1−ε − (f + fx + ffdi) = π∗x − ffdi. (2.6)

Equations (2.3)-(2.6) indicate that export and investment decisions will depend primar-

ily on both firm-specific characteristics (productivity) and the differences across countries

(wages, fixed costs, size). Suppose that the foreign country is a large, high-wage country,

while the home country is relatively small and characterized by low-wages. Consider the

export and investment decisions facing foreign country firms. The first decision is whether

to produce domestically and export abroad versus producing abroad and exporting back to

the country of origin. This decision will clearly depend upon the cost of exporting in either

direction. However, even if transport costs are equal there will be substantial differences in

these two production and export decisions. When transport costs are equal, the advantage

to exporting from the foreign country is that the foreign firm incurs lower fixed overhead

costs. The disadvantage is that labour costs are high.30 Thus, foreign firms that produce in

the home country must be productive enough to afford the higher fixed costs.

Similarly, consider a firm deciding whether to produce all units abroad and export back

to the foreign country versus opening a plant in each country. By producing all units abroad

30Country size also plays an important role here. Firms will be more likely to enter larger markets by FDI,ceteris paribus. However, to focus on the effect of factor prices I abstract from country size differences here.

21

the firm incurs the lowest marginal costs on each unit of output and saves the extra fixed

costs from operating multiple plants. However, by producing all units abroad the firm

incurs the transport cost on each unit exported. This is particularly costly when the foreign

country is large.

The model has several intuitive implications. First, since there is a fixed cost associ-

ated with entering export markets, firms that export will be more productive than domestic

non-exporters. Second, higher fixed costs associated with investing abroad suggests that

multinational firms will be more productive than their domestic counterparts. Finally, if

country size is equal across countries, only firms from the foreign (high wage) country will

invest in the home (low wage) country and export back to the foreign country. In other

words, the model predicts that vertical multinationals will originate in the high wage coun-

try and produce in the low wage country. This does not mean that FDI only flows from the

foreign country to the home country. Highly productive home firms may still want to invest

in the foreign country to save the transport costs from exporting. It will never, however, be

profit maximizing for home firms to use the foreign country as an export platform for home

country sales.

It is well known that multinational firms are more productive than domestic exporters

and that domestic exporters are more productive than domestic non-exporters.31 The evi-

dence presented in Section 2 demonstrates that we can further divide multinational firms

into more productive horizontal multinationals and less productive vertical multination-

als.32 Similar to Helpman, Melitz and Yeaple (2004), I must place restrictions on the size

of fixed and transport costs to ensure that firms can be partitioned into four groups. These

31See Helpman, Melitz and Yeaple (2004), for example.32Since the evidence in Section 2 is not conclusive, I consider the possibility of vertical multinationals as

the most productive firms in the appendix.

22

conditions are as follows:

f

(τB

B∗

)1−ε

< fx (C1)

fxw1−ε(B1−ε + (τB∗)1−ε)− w∗1−ε(B∗1−ε + (τB)1−ε)

(w∗τB)1−ε)< ffdi (C2)

ffdi <

[(wB)1−ε − (w∗τB∗)1−ε

(w∗B∗)1−ε − (wτB∗)1−ε

](f − fx) (C3)

w∗ < τw (C4)

The first condition is identical to that of Melitz (2003) except that it allows for dif-

ferences in country size. It ensures that the continuum of domestic firms is partitioned

between domestic non-exporters and exporters. In essence, the fixed cost to exporting must

be large enough that not all domestic firms are willing to export. The second condition

implies that the fixed costs of production abroad must be large enough to ensure that not all

exporting firms are more profitable by offshoring production. The third condition implies

that the fixed cost to domestic production cannot be so low that no firm would ever want

to offshore production.33 Moreover, for condition (C3) to hold the fixed cost of maintain-

ing a plant in the home country must be greater than the fixed cost of exporting from the

foreign country.34 Finally, the last condition states that the marginal production cost in the

home country must be lower than joint production and transport costs in the foreign coun-

try. If (C4) did not hold, production in the foreign country would be so inexpensive that no

multinational firm would ever choose to produce in the home country.

Condition (C3) also allows us to refine the model’s prediction regarding the direction

of vertical FDI. It is consistent with vertical FDI flows from large, high wage countries

33Note: If B∗ ≥ B then the term in square is positive. However, if B > B∗, then (C3) can only hold if thedifference between w∗ and w is sufficiently big.

34Although, I maintain the symmetry of fixed export costs across countries, one might expect MNCs tohave much lower fixed export costs than domestic exporters. I allow for this possibility in the estimation ofthe model.

23

to small, low wage countries. However, it also suggests that vertical FDI may flow from

small, high wage countries to large, low wage countries if the wage gap is large enough.35

There are a variety of other production and export possibilities that I have purposely

omitted here. For instance, it is possible that foreign investors may open plants in the home

country exclusively for sales to the home country consumers without operating a plant in

any other country. I exclude this possibility for the following two reasons: First, I cannot

empirically separate plants which are multinationals and those which stand-alone but have

foreign ownership. Second, Indonesia is a small country relative to its major sources of

FDI.36 If foreign investors are able to profitably operate a plant which sells exclusively to

Indonesian consumers, it is likely they are also able to the same in their home countries.

Table 2.4: Production and Sales CombinationsProduction Sales LocationLocation Home Foreign BothHome Yes∗ No YesForeign No Yes∗ YesBoth No No Yes

*Only if the ownership is the same as the country where production is located.

Table 2.4 documents the other production and sales combinations that I have purposely

restricted from the set possibilities by placing bounds on the fixed and transport costs.

The first row of Table 2.4 shows that firms can locate in the home country and produce

only for the home country or both markets, but not just the foreign market. As in Melitz

(2003) I exclude this possibility since less than two percent of all Indonesian plants receive

all revenues from export sales. Similarly, in the second row I exclude all foreign plants

35Similarly, the inequality w∗B∗ > wB also suggests vertical FDI to flow from large, low wage countriesto small, high wage countries if the difference in country size is large enough. However, in this case, condition(C4) would be violated.

36Japan, the United States and Europe.

24

that receive all revenues from export sales. Although seven percent of foreign firms in

the data strictly export all output, it is likely that these plants do so to protect technological

advantages, patents or product features from local rivals.37 I exclude this possibility since it

is not observable in the Indonesian manufacturing data.38 The last row simply indicates that

if a firm produces in both countries, then it must sell in both countries. Since firms must pay

all of the fixed costs to produce in a given country and the Dixit-Stiglitz framework implies

some residual demand for each variety, selling to local consumers can always increase

profits.39

2.3.5 Exit & Entry Decisions

Upon drawing their firm-specific productivity level each firm must decide whether or not

to enter each market. For convenience, first consider the decisions of foreign country firms.

The least productive firms expect negative profits and choose not to produce. Marginal

firms earn just enough profits to cover the fixed cost f :

(aw∗B∗)1−ε = f (2.7)

since ε > 1 and all of the profit functions are monotonically increasing in 1/a. For con-

venience, I denote productivity by the index a1−ε and let the productivity of the marginal

foreign firm be a∗1−ε. Figure 2.1 graphs the profit functions for all firms originating in the

foreign (high wage) country under the assumption that the demand level is the same in both

countries. Any firm with a productivity level below a∗1−ε will earn negative profits and

choose to exit the market altogether. Similarly, as shown in Figure 2.2, the same is true for

37Since the large majority of plants (93%) have positive amounts of domestic sales it is not likely that thesefirms are subject to the same restrictions.

38See Antras and Helpman (2004) for a model with incomplete contracts and vertical integration.39Over 98% of Indonesian producers and 93% of foreign producers sell to the local Indonesian market.

25

all home country firms with productivity below a1−ε.

Two important features of the model are highlighted in Figure 2.1. First, the profit

functions for foreign country firms are increasingly steep across domestic non-exporters,

domestic exporters, foreign exporters and foreign non-exporters. Second, across the same

four groups of firms, the total fixed costs required for production increase.40 The slope of

the profit function is determined by factor prices, transport costs and country size. Increases

in factor prices reduce the profitability of every unit sold and thereby reduce the slope of

profit function. Similarly, increases in transport costs reduce the profitability of every unit

sold in the export market which will lower the slope of exporters’ profit functions. For

example, foreign exporters have the lowest marginal cost of production for every unit sold,

while foreign non-exporting MNCs only benefit from low marginal costs of production in

the foreign plant. However, foreign non-exporting MNCs do not pay the transport costs

in the home country. Conditional on productivity, both firms are equally profitable in the

foreign country. Which firm is more profitable in the home country will depend on the

relative size of marginal costs across countries and the transport cost. If the total marginal

cost of producing a unit for sale abroad by a foreign exporter τw is greater than the marginal

cost of the foreign non-exporter w∗, the latter will earn higher profits in the home country.

On the other hand, if w∗ > τw and fixed costs are higher for the foreign non-exporting

MNC41 then no firm will ever choose to be a horizontal multinational.

The last effect on the slope of the production function is country size. Although Figures

2.1 and 2.2 abstract from differences in country size, it will always increase the profitability

of any firm producing for that market. Thus, the greater the country size, the greater the

amount of entry from any type of firm. This is particularly true for home country firms,40As demonstrated in conditions (C1)-(C4) the fixed costs do not have to have this strict of a ranking. I

assume such a ranking here to improve the exposition of the cutoff values.41Technically, condition (C3) holds.

26

π∗H

π∗X

a∗1−ε0

π∗

−f

−(fD + fX)

−(fD + fI + fX)

−(fI + 2fD)

a∗1−εD a∗1−ε

X a∗1−εV a∗1−ε

H

π∗V

π∗D

Figure 2.1: Foreign Country Thresholds

πh

πx

a1−ε0

π

−f

−(f + fx)

−(f + ffdi + fx)

−(ffdi + 2f )

a1−ε a1−εx a1−ε

h

πv

πd

Figure 2.2: Home Country Thresholds

27

which would otherwise suffer substantial transport costs by exporting to the large country.

Figure 2.1 demonstrates that the differences in the slopes and intercepts create four

distinct cutoff levels between foreign country firms. The first cutoff a∗1−ε partitions the

active from the non-active firms. Between cutoffs a∗1−ε and a∗1−εx all firms have positive

profits in the domestic market, but are not productive enough to engage in exports or FDI.

Likewise, all firms with productivity levels between a∗1−εx and a∗1−ε

v are productive enough

to export abroad, but not productive enough to establish home plants and capture lower

foreign wages. Any firm with productivity above a∗1−εv will establish plants abroad, but

only those with productivity above a∗1−εh establish them in both countries. Thus, given a

firm’s productivity index a∗1−ε its optimal production strategy can be read off the highest

of all four profit functions at that point.

Figure 2.2 demonstrates a similar pattern for home country firms except for the fact that

the profit function for a vertical multinational always lies below at least one of the other

profit functions. This implies that home firms who produce abroad and export back home

are not maximizing profits.42 This is intuitive since foreign producers are already located in

the country with lower marginal costs. However, as long as wτ > w∗, the most productive

home country firms will still want to invest horizontally in the foreign country in order to

avoid paying the transport costs.

2.3.6 Equilibrium

I focus on a stationary equilibrium in which aggregate variables are constant over time. The

value of a potential entrant is given by the maximum of its exiting value, which is assumed

42Thus, for home country firms I use an economy very similar to that in Helpman, Melitz and Yeaple(2003).

28

to be zero, and the discounted sum of expected profits

V (a) = max

{0,

∞∑t=0

(1− ξ)tE

(max

s∈{d,x,v,h}πs(a)

)}= max

{0, max

s∈{d,x,v,h}

πs(a)

ξ

}(2.8)

where unstarred functions are replaced with starred functions for foreign firms. Following

Melitz (2003) I can write the revenue of any firm as a function of the productivity of the

marginal firm with productivity a

rs(a) = λs

(a

a

)1−ε

f (2.9)

where I can solve for

λx = 1+(τB/B∗)1−ε, λv = (w∗/w)(τ 1−ε+(B∗/B)1−ε) and λh = 1+((w∗B∗)/(wB))1−ε

by multiplying and dividing each revenue function by rd(a) = (awB)1−ε = f .43

Let average profits within each group of firms be denoted by πs(a) and the fraction of

firms within each group as νs(a). Average overall profit, π(a), is then

π(a) =∑

s

νs(a)πs(a). (2.10)

Equation (2.10) relates average profits to the marginal productivity level in equilibrium.

For foreign firms average profits can be expressed as

π∗(a∗) =∑

s

ν∗s (a∗)π∗

s(a∗) (2.11)

where a = a∗(

w∗B∗

wB

).

A second set of equilibrium relations is provided by the free entry conditions. The

free entry conditions imply that the ex-ante expected value of a potential entrant in either

country must be zero

(1−G(a))

(π(a)

ξ

)= fE and (1−G∗(a∗))

(π(a∗)

ξ

)= f ∗

e . (2.12)

43For the foreign country the same three multipliers exist, λ∗x, λ∗v , and λ∗h, the only difference being thatthe starred and unstarred variables are switched.

29

I solve equations (2.10)-(2.12) to determine the equilibrium values of a and π(a) (and hence

a∗ and π∗(a∗)). Unlike Melitz (2003), the lack of symmetry across countries does not allow

me to ensure the existence or uniqueness of equilibrium, however, it can be demonstrated

numerically that such an equilibrium exists for reasonable parameter values.44

2.4 Empirical Model

2.4.1 Environment

In this section I construct an empirical analog to the theoretical model presented in the

previous section. I extend the theoretical model to incorporate stochastic fixed cost shocks

to the exit, export and foreign investment decisions captured in the theoretical model.45

I extend Rust’s (1987) framework to examine the nature of exit, export and FDI de-

cisions in the presence of stochastic fixed costs. I index each producer by the subscript

i = 1, ..., N . Entrants and incumbents draw idiosyncratic shocks (εχit(0), εχ

it(1))χ at the

beginning of each period. I interpret these shocks as unobserved state variables. The shock

εχit(0) is the return if the firm chooses to exit, while the shock εχ

it(1) along with continuation

value, specified in the Bellman equation below, is the return from producing. Conditional

on the productivity level a and the idiosyncratic shocks εχit, the firm decides whether or not

to continue to operate. I assume that the shocks εχit are independent of alternatives and is

drawn from the extreme-value distribution with scale parameter %χ. Without these shocks

44A numerical example in Matlab is available from the author upon request. The example shows that suchan equilibrium exists, but does not necessarily imply that it is unique. See Helpman, Melitz and Yeaple(2004) or Kasahara and Lapham (2006) for other examples of similar models that exploit symmetry to provethe existence and uniqueness of equilibrium.

45The cost shocks are necessary to capture certain patterns in the data. For instance, the theoretical modelpredicts that all firms below a certain cutoff will exit the market. However, the data confirms the existence ofmany small firms.

30

the model predicts that all firms below a certain threshold will exit which is inconsistent

with the large number of small firms observed in the data.

If a firm chooses to produce in period t, it then draws cost shocks associated with FDI

and international trade. Firms first receive a shock associated with the decision to maintain

a plant abroad (εfdiit (0), εfdi

it (1)) ≡ εfdiit . The return from not maintaining a plant abroad is

the sum of the shock εfdiit (0) along with the continuation value of not engaging in FDI, while

the sum of the shock εfdiit (1) along with the continuation value of engaging in FDI is the

return from FDI. Conditional on its FDI decision, a firm then receives a shock associated

with its trade decision (εxit(0), εx

it(1)) ≡ εxit where (εx

it(0) and the continuation value from

not trading is the return from choosing not to trade and (εxit(1) and the continuation value

from trading is the return from choosing to trade. Note that if the firm engages in FDI,

the foreign trade cost shock is drawn from a different extreme-value distribution. The FDI,

domestic trade and foreign trade shocks are all independently drawn from extreme-value

distributions with the respective scale parameters %fdi, %x and %x∗.

Thus, firms differ with respect to their productivity level a, their current FDI and export

status and their past FDI and export status. Denote a firm’s decision by dit ≡ (dfdiit , dx

it) ∈

{(0, 0), (0, 1), (1, 0), (1, 1)}. where dfdiit is the firm’s decision to maintain a plant abroad

and dxit is the firm’s decision to export. If a firm decides to engage in FDI or export that

decision takes a value of 1, and is 0 otherwise.

Given the structure of the cost shocks, I can characterize the optimization problem for

an incumbent Indonesian firm with productivity level ai by the following set of Bellman

31

equations:

V (ai) =

∫max{εχ(0), W (ai) + εχ(1)}dHχ(εχ), (2.13)

W (ai) =

∫max{J(ai, 0) + εfdi(0), J(ai, 1) + εfdi(1)}dHfdi(εfdi) (2.14)

J(ai, 0) =

∫ (maxdx′

π(ai, dx′) + βV (ai) + εx(dx′)

)dHx(εx) (2.15)

J(ai, 1) =

∫ (maxdx′

π(ai, dx′) + βV (ai) + εx∗(dx′)

)dHx∗(εx∗) (2.16)

Using the properties of extreme-value distributed random variables46 along with the so-

lution to the functional equations (2.13)-(2.16), the conditional choice probabilities follow

the familiar nested logit formula (c.f. McFadden, 1978).47 The probability of producing

this period (χ = 1) is calculated as:

P χ(χ = 1|a) = (1− ξ)

(exp(W (a)/%χ)

exp(0) + exp(W (a)/%χ)

)(2.17)

where ξ is the exogenous probability of exit and W (·) is defined as in equation (2.14).

Similarly, the conditional choice probabilities for all possible FDI and export decisions

follow the nested logit formula.48 Conditional on operating, the probability of maintaining

a plant abroad is given by

P fdi(dfdi′ = 1|a, χ = 1) =exp(J(a, (1, dx)/%fdi)∑

dfdi exp(J(a, (dfdi, dx))/%fdi)(2.18)

46See Ben-Akiva and Lerman (1985) for an example.47As described in Rust (1994) there are important differences between the static nested logit models and

dynamic nested logit models, like the one described here. First, in static models the independence from irrel-evant alternatives (IIA) property holds within each nest. However, in a dynamic setting the IIA property typ-ically cannot hold even within a nest because the continuation value depends on alternatives outside the nest.Second, a static model usually has a closed-form specification in parameters, such as linear-in-parametersspecification. Dynamic models, such as the one here, do not have a closed-form expression in parameters andinstead require the solution to the functional equations (2.13)-(2.16). Evaluating the conditional choice prob-abilities in a dynamic setting is a computationally intensive task. Fortunately, the extreme-value specificationadopted here substantially simplifies the computation by avoiding the need for multi-dimensional numericalintegration in (2.13)-(2.16).

48See McFadden (1978).

32

while the conditional probability of producing everything in the country of origin is given

by P fdi(dfdi′ = 0|a, χ = 1) = 1−P fdi(dfdi′ = 1|a, χ = 1) and J(·, ·) is given in equations

(2.15)-(2.16). Simlarly, conditional on operating and choosing to maintain a plant abroad,

the firm’s export choice probability can be calculated as

P x(dx′|a, χ = 1, dfdi′ = 1) =exp([π(a, (dx′ , 1)) + βV (a)]/%x∗)∑

dx′ exp([π(a, ((dx′ , 1)) + βV (a)]/%x∗). (2.19)

The choice probilibility for exporting conditional on not engaging in FDI is analagous to

(2.19) where the scale parameters are replaced with their unstarred counterparts.

I focus on a stationary equilibrium where the distribution of a is constant over time. I

assume that the logarithm of plant-specific productivity, ln a, is drawn from the N(0, σa)

distribution where the variance of the distribution, σa varies across Indonesian and foreign

plants.49 As in the theoretical model, I further assume that these draws are independent of

each other and are constant over the life of the firm. Let ga(a) denote the density function

of a for Indonesian plants.

The above assumptions imply that we can write down the expected value of the firm,

which, under free entry, must be equal to the fixed entry cost fe:∫V (a)ga(a)da = fe. (2.20)

Denote the stationary distribution of a among incumbents as µ(a). A stationary equilibrium

requires that the number of exiting firms with productivity a must equal the number of

successful entrants with the same productivity level. Specifically,

MP (χ = 0|a)µ(a) = MeP (χ = 1|a)ga(a) for all a (2.21)

49I assume that the initial mean of the distribution of initial productivity draws is the same across foreignand domestic plants. It is difficult to identify the initial mean for foreign plants since the data only captures asubset of all foreign plants.

33

where M is the mass of incumbents and Me is the total mass of entrants that attempt to

enter the market. Rearranging terms, the stationary distribution of productivity µ(a) can be

computed as

µ(a) =Me

M

P (χ = 1|a)

P (χ = 0|a)ga(a) (2.22)

whereMe

M=

1∫ P (χ=1|a)P (χ=0|a)

ga(a)da

since∫

µ(a)da = 1. A similar procedure is followed for foreign plants.

2.4.2 The Likelihood Function

I define the following function of iceberg shipping costs50

ϕτ ≡ (1− ε) ln τ. (2.23)

I assume that total revenue is measured with error and that exogenous technological change

occurs at rate ρ. By modifying the profit functions to include measurement error and a time

trend, I use equations (2.3)-(2.6) to write the logarithm of observed revenue for any plant i

as

ln rit = ρtt + ln ϕB(1− dXit ) + ln[ϕB + ϕW exp(ϕτ )]d

Xit − ln ai + νit (2.24)

where rit is observed revenue, ϕW measures the ratio of wages across countries (w∗/w)1−ε,

ϕB is a function of the country sizes and prices (B∗/B)1−ε and νit is the associated mea-

surement error. Equation (2.24) highlights an important limitation of the data: I only ob-

serve the revenue, exports and ownership from plants located in Indonesia. I do not observe

any variables for plant parents or subsidiaries abroad. Estimating the model requires im-

posing some consistency across plants located in different countries. In particular, I assume50Allowing transport costs to vary across foreign and domestic plants makes little difference to the final

results.

34

that every foreign non-exporting plant also produces in a separate plant located in the for-

eign country with the same firm-specific productivity level as the plant located in Indonesia.

Using the empirical specification, a firm’s detrended net profit may be expressed in

terms of reduced-form parameters as:51

π(a, di,t−1, dit) = r(a, dit)− F (di,t−1, dit) (2.25)

Given an Indonesian firm’s current FDI and export decisions they may incur the fol-

lowing set of fixed costs in a given period:

F (dit, di,t−1) =

f for (dfdiit , dx

it) = (0, 0)

f + fx for (dfdiit , dx

it) = (0, 1)

f + f ∗ + f ∗fdi for (dfdi

it , dxit) = (1, 0)

f ∗ + ζfxf∗x + f ∗

fdi for (dfdiit , dx

it) = (1, 1)

where starred variables denote costs incurred in the foreign country.

Due to limitations of the data it is not possible to identify the parameters ϕW , f ∗, and f ∗x .

The first parameter represents differences in wages across countries. I calibrate this param-

eter using data on manufacturing wages in different countries.52 The last two parameters

are fixed cost parameters in the foreign country. To calibrate the fixed cost of operating in a

foreign country, f ∗ I use an index of labour rigidity to estimate that f ∗ = 0.51f . Similarly,

I calibrate that the fixed cost of exporting from the foreign country is fx∗ = 0.38fx using

the number of days needed to process export applications across countries. The informa-

tion for this calibration is taken from the World Bank’s Doing Business Report. To test for

possible misspecification around the fixed cost parameters, I check the robustness of the51The detrended firm’s problem uses a trend-adjusted discount factor β exp(ρ) when solving the Bellman’s

equation.52The wage data is taken from the International Labour Organization Bureau of Statistics. The foreign

wage is a weighted average of foreign wages using the share of FDI in Indonesia as weights.

35

results with regards to this calibration by estimating the model under various alternative

fixed cost assumptions.53

The vector of remaining parameters θ is estimated by the method of maximum likeli-

hood where54

θ = (ρ, ϕB, ϕτ , f, fx, ffdi, ζfx , ξ, %χ, %x, %x∗, %fdi, σa, σ

∗a).

Define the variable χit be a variable that takes a value of 0 if a plant exits the data and 1

otherwise. Thus, the probability of χ = 1 for domestic plants is simply the probability that

χ = 1, P χθ (χit = 0|ai) = P χ

θ (χit = 0|ai). However, a foreign plant may exit because of a

shock that causes the firm to die entirely or because it chooses to produce in its country of

origin instead:

P χ∗

θ (χ∗it = 0|ai) = P χ

θ (χit = 0|ai) + P χθ (χit = 1|ai)P

fdiθ (dfdi

it = 0|ai).

Similarly, define the probability of FDI and export as

P dθ (dit|ai) = P fdi

θ (dfdiit |ai)P

xθ (dx

it|ai, dfdii,t )

and P fdiθ (dfdi

it = 1|ai) = 0 for Indonesian plants by assumption.

Denote Ti,0 as the first year the firm appears in the data. Then, conditional on ai the

likelihood contribution of plant i in year t > Ti,0 is

Lit(θ|ai) =

P χ

θ (χit = 0|ai) for χit = 0

P χθ (χit = 1|ai)︸ ︷︷ ︸

Stay/Exit

P d(dit|ai, χit = 1)︸ ︷︷ ︸FDI/Export

gν(νit(ai))︸ ︷︷ ︸Revenue

for χit = 1

53In particular, I re-estimate the model assuming that the fixed operation costs are equal across countriesand that fixed costs are much higher in Indonesia.

54The discount factor is not estimated and is set to 0.95. It is difficult to identify the discount factor β indynamic discrete choice models.

36

where starred values replace unstarred values for foreign plants. Note that the endogeneity

of the export, FDI and exiting decisions are controlled by simultaneously considering the

likelihood contribution from each decision.

In the first year of the sample, Ti0, I only observe plants that stay in the market. Thus, I

calculate the likelihood contribution of these plants in the initial year conditional on χit =

1:

Lit(θ|ai) = P d(dit|ai, χit = 1)gν(νit(ai)) (2.26)

Let Ti,1 denote the last year plant i appears in the data. Then, the likelihood contribution

from each plant i is

Li(θ|ai, di,Ti,0) =

Ti,1∏t=Ti,0+1

Lit(θ|ai).

It is well known that the distribution of unobserved productivity a depends crucially on

whether entering plants are observed in the initial year of the sample (Heckman (1981)).

In the initial year of the sample I assume that ai is drawn from the stationary distribution

µ(a). If plant i enters the sample after the initial year, I assume that ai is drawn from the

distribution of initial draws upon successful entry into Indonesia

gea(a) =

P χ(χ = 1|a)∫P χ(χ = 1|a′)ga(a′)da′

ga(a). (2.27)

I use equation (2.27) along with the choice probabilities (2.17)-(2.19) to build the likelihood

function.

The likelihood contribution from each plant i is calculated by numerically intergrating

out unobserved plant-specific productivity ai as

Li(θ) =

∫

Li(θ|a′)µ(a′)da′ for Ti,0 = 1993,∫Li(θ|a′)P d

θ (di,Ti,0|a′)ge

a(a′)da′ for Ti,0 > 1993,

37

where the starred distributions are used in place of the unstarred distributions for foreign

firms and the stationary distribution of foreign plants is conditional on entering Indone-

sia, µ∗(a′|dfdii,Ti,0

= 1). The parameter vector θ can then be estimated by maximizing the

logarithm of the likelihood function

L(θ) =N∑

i=1

ln Li(θ). (2.28)

The evaluation of the log-likelihood function involves solving the dynamic program-

ming problem that approximates the Bellman equations (2.13)-(2.16) by discretization of

the state space. I first fix the fixed and sunk export costs in the foreign country as a fraction

of the costs in Indonesia. Then for each candidate choice of parameter vector, I solve the

discretized dynamic programming problem (2.13)-(2.16), calculate the conditional choice

probabilities (2.17)-(2.19) and the stationary distributions. Using the conditional choice

probabilities and the stationary distributions, I evaluate the log-likelihood function (2.28).

Searching over the parameter space of θ, I maximize (2.28) to find the estimates.

2.4.3 Reduced-Form and Structural Parameters

It is not possible to identify all of the parameters of the model. Equation (2.24) is a reduced-

form specification where the reduced-form parameters represent the structural parameters

as follows.

ϕB =

(wB

w∗B∗

)1−ε

, (2.29)

ϕW =( w

w∗

)1−ε

. (2.30)

It is important to note that policy changes may affect the value of reduced-form pa-

rameters if the underlying structural parameters change. For instance, any change to the

38

aggregate price level P will lead to a change in B = [(1−α)γE]1/(1−ε)/(αP ) and ϕB. The

counterfactual experiments in this paper explicitly account for equilibrium price changes on

the reduced-form coefficients using the relationship between the reduced-form coefficients

and the aggregate prices.55

2.4.4 Identification

The identification of the revenue function (2.24) parameters follows from the within-plant

variation in export status along with the moment restrictions E[νit − νi(t−1)|dit] = 0 where

i ∈ {h, f}. The latter condition is obtained by first-differencing the plant-level revenue

function (2.24). I further assume that the panel is long enough that given the parameters

identified in the revenue function, I can identify the value of plant-specific productivity a

for each plant.56

Since the exiting probabilities are strictly increasing in the fixed cost, f , I can identify

f by relating the probability of exit to the variation in a. At the same time, the elasticities

of the exiting probabilities tend to decrease as the variance of the exiting shocks increases.

The variation in the differences between different a’s across firms and the difference in

exiting probabilities identify the scale parameter %χ separately from f .57 We may similarly

identify the fixed cost and scale parameters by relating the variation in a to the variation in

export and investment probabilities.

Lastly, the scale of the profit function cannot be identified because multiplying the profit

function by a constant leads to the same optimal choice. Thus, for identification I normalize

55See Appendix A for details.56Even though I only use four years of data, the distributional assumptions on a allow me to identify each

plant’s likelihood of having a particular value of a.57The variance of the exit decision cost shocks is calculated as V ar(εχ(χ)) = (%χπ)2

6 .

39

the profit function (2.25) by κ = ε/(w∗B∗)1−ε.58

2.4.5 Data

I employ data from the Indonesian manufacturing census for 1993-1996.59 The census

enumerates all plants with at least 20 employees.60 I focus on Indonesian food, textile

and manufactured metals industries since they are among Indonesia’s largest industries and

receive substantial foreign direct investment.61 I omit all plants that are owned entirely

by the Indonesian government.62 The food, textile and manufactured metals industry data

consists of unbalanced panels of 6,042, 4,491 and 2,497 plants, respectively, where each

plant is observed for at least one year between 1993 and 1996.

The advantage of this data set relative to many other plant-level data sets is that I am

able to observe the percentage of foreign ownership for each individual plant. I identify

a foreign plant as any plant that has positive foreign ownership as a foreign plant. It is

possible that if foreign investors own a small minority of plant equity the plant may not

be foreign controlled. However, in over 66% of the foreign firms in the sample, foreign

investors own at least 50% of the equity, while foreign investors own at least 25% of foreign

firms in 95% of the sample.63 Similarly, I identify exporters as plants that receive any

positive revenues from export sales. As discussed in the Section 2.2 this definition is the

58Specifically, by multiplying the profit function by ε, I estimate the parameters κf , κfx, κffdi, κ%χ, κ%x,κ%x∗ and κ%fdiinstead of f , fx, ffdi, %χ, %x, %x∗ and %fdi.

59Two-digit ISIC classification.60It is believed that over this period the coverage was close to complete since regional offices had financial

incentives to enumerate plants (Blalock and Gertler, 2005). A description of the Indonesian manufacturingindustry over this period can be found in Blalock and Gertler (2005).

61Approximately 58% of all foreign plants in the Indonesian manufacturing census are in the food, textileand manufactured metals industries.

62Overall, I eliminate 669 plants over the 1993-1996 period.63Blalock and Gertler (2005) argue that firms with significant foreign ownership are likely to be foreign

controlled even if they are not majority owned. Moreover, this variable is highly correlated with a governmentstatus variable that indicates whether a firm if a plant has official foreign status by the Indonesian government.

40

most conservative of all possible thresholds.

The unit of observation is that of the individual plant, not the firm. This is particularly

important to this empirical exercise because I do not observe if the plant is a parent or a

subsidiary, but only if it is foreign or domestically owned. To estimate the model I will

have to make an assumption that all foreign plants are part of a multinational firm and that

all domestic plants are strictly national firms.64 Ramstetter and Sjoholm (2006) analyze the

same Indonesian data and indicate that foreign owned plants in Indonesia are typically part

of multinational corporations. Moreover, to the extent that the model captures the decision

of foreign plants to enter Indonesia, the model’s implications for Indonesia will remain

valid.65

Since I only observe plants located in Indonesia, I calibrate the foreign country param-

eters using data the International Labour Organization Bureau of Statistics and the World

Bank’s Doing Business Report. Specifically, I use the wage data from the International

Labour Organization Bureau of Statistics for manufacturing wages and the World Bank’s

Doing Business Report to estimate the relative size of fixed operation and export costs

across countries.

Another limitation of the data is that I do not observe the export destinations of each

firm. Thus, I cannot identify plants that export to developed markets versus those who use

Indonesia as an export platform for nearby regional markets. However, Table 2.5 suggests

that Indonesian industries that earn a higher percentage of revenues from exports are more

64Several other empirical papers use data that directly connects parents and subsidiary plants (see Helpman,Melitz and Yeaple (2004), for example). Unfortunately, these data sets are inadequate to estimate the modelpresented here, since it requires observing all plants in the industry. Baily, Hulten and Campbell (1992) showthat accounting for productivity differences across plants is particularly important when evaluating policy.Thus, in order to perform the policy experiments in this paper I must observe all plants in the industry.

65The OECD Direct Investment Statistics (2006) reports that Indonesia received over 14,352 million USdollars of foreign investment from 1993-1996, while it only supplied just over 40 millions US dollars worthof FDI to the world economy.

41

Table 2.5: Indonesian Export DestinationsIndustry Export % of Industry Exports No. of

Intensitya to Developed Nations Obs.Wood 0.67 58% 1,860Textiles 0.55 75% 1,816Food 0.46 64% 956Manufactured Metals 0.41 49% 892Minerals 0.41 47% 275Chemicalsb 0.35 47% 1,065Basic Metals 0.33 32% 144Paper 0.27 19% 147

Notes: Data compiled from the United Nations Commodity Trade Statistics 1994. (a) Export Intensity is the mean export intensity ofall exporting firms in the industry. (b) I have omitted firms in the petroleum industry as they were large outliers. This resulted in theremoval of twelve plant-year observations.

likely to export to developed countries. Along with the discussion in Section 2, this would

suggest that vertical multinationals are more likely to be export-intensive foreign firms. To

be conservative I assume that all foreign exporters export to developed countries.66

I focus on the following four variables for each plant i in year t: χit, rit, dxit and dfdi

it . I

convert the nominal value of total sales by the manufacturing output price deflator to cal-

culate the real value of total sales, rit. The binary variables dxit and dfdi

it are constructed by

checking the value of export sales and foreign ownership in each year. The entry/exit deci-

sion, χit is identified by checking whether plants employed a positive number of workers

in each year.

Table 2.6 provides descriptive statistics for all five variables over the sample period. The

large standard deviations indicate substantial variation across plants in terms of total sales,

export sales and labour. The percentage of foreign plants is highest in the manufactured

metals industry and lowest in the food industry. Table 7 also demonstrates high exit and

66The results would not change if I included firms that only receive a small percentage of revenues fromexport sales in the “non-exporting" group.

42

Table 2.6: 1993-96 Descriptive StatisticsTotal Export Labour Mark-Up % Foreign Entry Exit No. of

Salesa Salesa,b Ratec Plantsd Ratese Ratesf PlantsFood 25.51 44.67 106.64 0.26 0.02 0.15 0.09 6,042

(410.97) (312.42) (703.44) (0.18) — — — —Metalsg 45.77 45.70 168.87 0.33 0.10 0.16 0.06 2,497

(210.26) (102.67) (325.05) (0.19) — — — —Textiles 38.35 56.76 245.09 0.26 0.04 0.15 0.08 4,491

(198.91) (144.16) (730.25) (0.16) — — — —

Notes: Reported numbers are sample means with standard deviations in parentheses. (a) In millions of Indonesian Rupiahs. Thepercentage change is calculated as the mean percentage change across plants. (b) Computed using the sample of exporting plants. (c)The mark-up rate is computed as (revenue-variable cost)/revenue where variable cost is measured as the sum of materials, energy, fueland the wages paid to production workers. (d) The average is computed as the percentage of plants with foreign ownership. (e) Thenumber of new entrants divided by the total number of plants in 1993. (f ) The number of exiting plants divided by the total number ofplants. (g) Metals refers to manufactured metals rather than basic metals.

entry rates which are important for identifying the parameters that affect the choice prob-

abilities as well as the initial distribution of productivity shocks. On average 906, 674 and

400 new plants enter the Indonesian food, textiles and manufactured metals industries each

year, respectively, while 544, 359, 150 incumbents exit. A substantial amount of turnover

is important for identifying the parameters that determine the exiting choice probabilities

and the distribution of initial draws.

2.5 Estimation Results

Table 2.7 presents the maximum likelihood parameter estimates of the empirical model

along with the associated asymptotic standard errors. The standard errors are computed

using the outer product of gradients estimator. The parameters are evaluated in millions of

Indonesian rupiahs in 1983.

43

Table 2.7: Structural EstimatesIndustry Food Metals Textiles

κ%χ 52.416 (4.481) 31.780 (2.815) 35.490 (2.053)κ%x 10.287 (0.587) 4.897 (0.373) 7.902 (0.464)κ%x∗ 3.428 (6.545) 1.055 (16.411) 0.455 (0.619)κ%fdi 0.0001 (1.423) 0.001 (4.421) 0.0001 (1.667)

κf 2.379 (1.167) 3.567 (0.830) 5.870 (0.680)κfx 42.860 (2.307) 21.923 (1.548) 23.513 (1.324)

κffdi 28.041 (3.361) 29.293 (10.881) 46.737 (4.9860ζfx 0.078 (0.110) 0.088 (0.082) 0.034 (0.093)ρ -0.009 (0.010) 0.011 (0.013) 0.008 (0.009)

ϕB 0.853 (0.020) 1.531 (0.048) 1.486 (0.027)ϕτ -5.346 (0.060) -3.799 (0.078) -5.369 (0.060)σa 1.041 (0.007) 0.996 (0.008) 1.011 (0.003)σa∗ 1.139 (0.010) 0.971 (0.009) 0.941 (0.007)ξ 0.006 (0.001) 0.021 (0.002) 0.014 (0.001)

ε = 1/mark-up 3.8 3.0 3.8log-likelihood -34,806.143 -16,819.565 -31,866.159No. of Obs. 17,786 7,549 13,287

Notes: Standard errors are in parentheses. The parameters are evaluated in units of millions of Indonesian Rupiahs in 1983. Metalsrefers to manufactured metals.

2.5.1 Fixed Costs

The average fixed cost of operation in Indonesia ranges from 12 and 19 thousand 1983 US

dollars in the food and manufactured metals industries, respectively, to 25 thousand 1983

US dollars for the textiles industry. Although the average fixed cost is typically a large

percentage (70 to 105% across industries) of the average domestic non-exporters revenue it

is important to recall that the fixed costs incurred are substantially lower since the estimated

model predicts many plants only produce when they receive beneficial cost shocks.

The average fixed export cost ranges from 101 and 114 thousand 1983 US dollars in the

manufactured metals and textiles industries, respectively, to 222 thousand 1983 US dollars

in the food industry. The differences in fixed costs are likely picking up larger differences

in productivity between exporters and non-exporters in the food industry. The prevalence of

less productive exporters in the textiles and metals industries may potentially be attributed

to government export subsidies in those industries. Similarly, the fixed FDI costs range

44

Table 2.8: Distribution of Plants by Ownership/Export StatusActual Domestic Non-Exporters Domestic Exporters Foreign Non-Exporters Foreign ExportersFood 0.931 0.046 0.016 0.008Metals 0.824 0.071 0.058 0.047Textiles 0.839 0.114 0.025 0.023PredictedFood 0.949 0.028 0.018 0.006Metals 0.864 0.032 0.058 0.047Textiles 0.865 0.088 0.021 0.026

Note: Metals refers to manufactured metals.

from 145 thousand US dollars in the food industry, to 152 thousand 1983 US dollars in the

manufactured metals industry and 202 thousand 1983 US dollars in the textiles industry.

Again, the large differences in fixed FDI costs may not only reflect differences in fixed FDI

costs, but also government subsidies across industries.

The estimates imply that plants that engage in FDI and exports incur substantial per-

period fixed costs to continue these activities. However, the large estimated size of the

fixed export and FDI costs are likely capturing the fact that the model does not allow for

sunk costs associated with export or investment behaviour and may be biased upwards.67 I

will address the potential presence of sunk export or FDI costs in chapter three. Also, the

parameter capturing the fixed export cost savings of foreign exporters in Indonesia indicates

that foreign plants save an average 91 and 92 percent of the fixed export costs relative to

domestic exporters in the manufactured metals and food industries, while foreign textile

exporters save 97 percent of the fixed export costs. This may be indicative of the prevalence

of trade occurring across plants within the same multinational firm.

45

2.5.2 Exports and FDI

The parameters ϕτ and ϕB indicate that the impact of exports is substantial. In fact, the esti-

mates imply that the average plant can increase their revenues by 64, 65 and 113 percent in

the textiles, manufactured metals and food industries. These numbers are particularly large

in light of the fact that the transport cost is estimated to be approximately 6.7 to 6.8 across

industries. The large increase in revenues reflect the substantial productivity differences

between exporting and non-exporting firms and the small size of Indonesian economy rel-

ative to that of the rest of the world.68 The estimates also imply that on average 39 percent

of a domestic exporters revenues are from export sales in the manufactured metals and tex-

tiles industries, while exporters in the food industry receive 53 percent of revenues from

export sales. Relative to the data, the model underpredicts the average percentage of rev-

enues from exports in the manufactured metals and textiles industries by 1 and 14 percent,

respectively, while overpredicting the average percentage of revenues from exports in the

food industry by 4 percent.

We can examine the fit of the model by comparing its predictions of the actual distribu-

tion to plants in the data. Table 2.8 documents the predicted distribution of plants export and

ownership status for all three industries.69 The model’s predictions match the distribution

of plants across export and ownership status very closely in all industries. Table 2.9 re-

ports the models predicted domestic market and export shares across export and ownership

status. While one would expect the model to predict the dominant role of foreign plants

in the domestic and export markets, Table 2.9 demonstrates that it often overstates their

67See Kasahara and Lapham (2007) and Das, Roberts and Tybout (2007).68The parameters φB and φW jointly imply that Indonesia’s economy is roughly 0.4 to 3.5 percent of the

world economy.69The estimated model does not provide a prediction in terms of the total number of foreign firms relative

to the total number of domestic firms. As such, I take the percentage of foreign in the data as given.

46

Table 2.9: Export/Domestic Market Share by Ownership/Export OwnershipExport Share Domestic Market Share

Actual Domestic Foreign Domestic Non-Exp. Domestic Exp. Foreign Non-Exp. Foreign Exp.Food 0.837 0.163 0.551 0.289 0.126 0.032Metals 0.458 0.542 0.462 0.099 0.300 0.138Textiles 0.717 0.283 0.602 0.238 0.097 0.062PredictedFood 0.605 0.395 0.703 0.084 0.159 0.055Metals 0.247 0.753 0.487 0.066 0.248 0.200Textiles 0.452 0.548 0.550 0.136 0.149 0.165

Note: Metals refers to manufactured metals.

importance in both domestic and export markets. Because of the substantial productivity

difference in foreign and domestic plants, small differences in the distribution of plants or

the predicted productivity differences can have large impacts at the aggregate level.

2.5.3 Productivity

The model predicts that only the most productive firms are able to produce profitably in

Indonesia. Figure 2.3 shows the importance of survival selection among domestic plants

in Indonesia. The actual productivity distribution for incumbents in the top three panels

is skewed to the right relative to the distribution of new entrants for all three industries.

The bottom three panels of Figure 2.3 show that the model’s predicted productivity distri-

butions for entrants and incumbents capture this selection mechanism. Similarly, Table 11

confirms that domestic incumbents are on average 20 to 37 percent more productive than

new domestic entrants across industries.

Figure 2.4 shows the productivity distributions for foreign entrants and incumbents.

The top and bottom panels demonstrate both an important similarity and difference. First,

the top panel shows that the distributions of foreign entrants and incumbents into Indonesia

are very close. As Table 2.10 indicates, the estimated difference between foreign entrants

47

−6 −4 −2 0 2 4 60

0.05

0.1

0.15

0.2

0.25

log(1/a)+νit

Act

ual F

requ

ency

Food

EntrantsIncumbents

−6 −4 −2 0 2 4 60

0.05

0.1

0.15

0.2

0.25

0.3

log(1/a)

Pre

dict

ed F

requ

ency

Food

EntrantsIncumbents

−6 −4 −2 0 2 4 60

0.05

0.1

0.15

0.2

0.25

log(1/a)+νit

Act

ual F

requ

ency

Metals

EntrantsIncumbents

−6 −4 −2 0 2 4 60

0.05

0.1

0.15

0.2

0.25

0.3

log(1/a)

Pre

dict

ed F

requ

ency

Metals

EntrantsIncumbents

−6 −4 −2 0 2 4 60

0.05

0.1

0.15

0.2

0.25

log(1/a)+νit

Act

ual F

requ

ency

Textiles

EntrantsIncumbents

−6 −4 −2 0 2 4 60

0.05

0.1

0.15

0.2

0.25

0.3

log(1/a)

Pre

dict

ed F

requ

ency

Textiles

EntrantsIncumbents

Figure 2.3: Productivity Distributions of Domestic Entrants and Incumbents (Actual vs.Predicted)

and incumbents is relatively small. The predicted productivity distributions match this

feature well in all 3 industries since the distribution of entrants and incumbents in Indonesia

are almost identical. Second, there is a much smaller predicted productivity variance among

foreign firms than we observe in the actual distribution. The reason for this is that both

the exit and investment shocks are estimated to be very low relative to the profitability

of foreign plants. As such, if a plant is productive enough to invest abroad there is little

chance that they will receive a stochastic shock that will induce it to leave Indonesia. The

difference between the predicted and actual distributions is suggestive of the presence of

48

one-time sunk costs. That is, if a firm realizes a beneficial stochastic cost shock to a one-

time sunk cost that allows it to invest in Indonesia, lower productivity foreign firms may

be more likely to choose to invest in Indonesia. However, in a model with only fixed costs,

one-time beneficial cost shocks will encourage many low productivity firms to temporarily

enter Indonesia. However, this will quickly cause the exit rate of foreign firms to rise to

levels which are inconsistent with the observed exit rate in the data.

−5 0 5 100

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

log(1/a)+νit

Act

ual F

requ

ency

Food

EntrantsIncumbents

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

log(1/a)

Pre

dict

ed F

requ

ency

Food

EntrantsIncumbents

−5 0 5 100

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

log(1/a)+νit

Act

ual F

requ

ency

Metals

EntrantsIncumbents

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

log(1/a)

Pre

dict

ed F

requ

ency

Metals

EntrantsIncumbents

−5 0 5 100

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

log(1/a)+νit

Act

ual F

requ

ency

Textiles

EntrantsIncumbents

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

log(1/a)

Pre

dict

ed F

requ

ency

Textiles

EntrantsIncumbents

Figure 2.4: Productivity Distributions of Foreign Entrants and Incumbents (Actual vs. Pre-dicted)

As shown in Section 2.2 domestic exporters tend to have higher productivity than do-

mestic non-exporters, while the opposite is true for foreign plants. Also, foreign plants

49

Table 2.10: Average ProductivityFood Metals Textiles

Indonesian Foreign Indonesian Foreign Indonesian ForeignMean at entry trial 1.000 1.000 1.000 1.000 1.000 1.000Mean at successful entry in Indonesia 1.080 3.619 1.161 4.650 1.088 3.619Mean at steady state in Indonesia 1.300 3.739 1.585 4.837 1.421 3.676Non-Exporters 1.278 3.752 1.533 4.846 1.373 3.745Exporters 2.060 3.703 2.983 4.824 1.901 3.621

Note: Metals refers to manufactured metals.

tend to be more productive than domestic plants. Figure 2.5 shows the actual and pre-

dicted productivity distributions for domestic non-exporters, domestic exporters, foreign

non-exporters and foreign exporters, while Table 2.10 reports their average productivities.

The top panel shows that in all three industries the plants follow the same productivity

ranking that we observed in the Section 2.2. In particular, it is notable that the productivity

distribution of foreign non-exporters is skewed to the right of the productivity distribution

of foreign exporters, though the differences across foreign plants are much less clear in the

manufactured metals and textiles industries. However, the large difference between foreign

and domestic firms is quite evident. The predicted distribution matches these rankings and

Table 2.10 confirms that on average foreign non-exporters are the most productive plants,

followed by foreign exporters, domestic exporters and domestic non-exporters.

2.5.4 Dynamics

Table 2.11 documents the actual and predicted transition probabilities of FDI, export and

exit in the textiles industry.70 It is noteworthy that despite the model’s restriction that no

domestic plant can become a foreign plant, the model captures many of the transition proba-

bilities between investment and export status relatively well. The data indicates that foreign

70Similar tables can be found in the appendix for the food and manufactured metals industries.

50

−5 0 5 100

0.05

0.1

0.15

0.2

0.25

0.3

log(1/a)+νit

Act

ual F

requ

ency

Food

Domestic Non−ExportersDomestic ExportersForeign ExportersForeign Non−Exporters

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

log(1/a)

Pre

dict

ed F

requ

ency

Food

Domestic Non−ExportersDomestic ExportersForeign ExportersForeign Non−Exporters

−5 0 5 100

0.05

0.1

0.15

0.2

0.25

0.3

log(1/a)+νit

Act

ual F

requ

ency

Metals

Domestic Non−ExportersDomestic ExportersForeign ExportersForeign Non−Exporters

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

log(1/a)

Pre

dict

ed F

requ

ency

Metals

Domestic Non−ExportersDomestic ExportersForeign ExportersForeign Non−Exporters

−5 0 5 100

0.05

0.1

0.15

0.2

0.25

0.3

log(1/a)+νit

Act

ual F

requ

ency

Textiles

Domestic Non−ExportersDomestic ExportersForeign ExportersForeign Non−Exporters

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

log(1/a)

Pre

dict

ed F

requ

ency

Textiles

Domestic Non−ExportersDomestic ExportersForeign ExportersForeign Non−Exporters

Figure 2.5: Productivity Distributions of Domestic and Foreign Firms by Export Status(Actual vs. Predicted)

ownership and export status are quite persistent. Although the model captures much of the

persistence in export and ownership status, it still underpredicts the degree persistence for

both foreign and domestic exporters. This could be indicative of the presence of the sunk

export or investment costs.71

71Kasahara and Lapham (2007) also find that fixed export costs alone cannot explain the persistence inexport status and suggest the potential presence of sunk export costs.

51

Table 2.11: Transition Probabilities - TextilesActual Dom. Non-Exporters Dom. Exporters For. Non-Exporters For. Exporters ExitDom. Non-Exporters at t 0.945 0.051 0.003 0.001 0.088Dom. Exporters at t 0.309 0.679 0.001 0.011 0.068For. Non-Exporters at t 0.138 0.031 0.581 0.256 0.030For. Exporters at t 0.027 0.066 0.178 0.732 0.016PredictedDom. Non-Exporters at t 0.912 0.088 — — 0.144Dom. Exporters at t 0.820 0.180 — — 0.094For. Non-Exporters at t — — 0.487 0.513 0.031For. Exporters at t — — 0.414 0.586 0.029

2.5.5 Counterfactual Experiments

I present the results from a series of counterfactual experiments intended to examine the

effect of trade and FDI barriers for the textiles industry. The results for the food and man-

ufactured metals industries are presented in Appendix A. In particular, to determine the

quantitative implications of barriers to trade and FDI, I conduct the following three coun-

terfactual experiments by manipulating three parameters:

1. Autarky: fx, fx∗ , ffdi →∞,

2. No Trade: fx, fx∗ →∞,

3. No Investment in Indonesia: ffdi →∞,

To determine the full impact of trade or investment barriers on the Indonesian economy

it is important to consider the effect policy changes have on the aggregate price level. To do

this I employ the free entry conditions (2.20) and solve for the new price levels in Indonesia

and the rest of the world which satisfies (2.20) under the policy change.72

Table 2.12 presents the results from the counterfactual experiments in the textiles in-

dustry. The effect of trade and investment on average productivity can be best understood72See Appendix A for details.

52

Table 2.12: Counterfactual Experiments - TextilesBase Autarky No Trade No FDI

Avg. Productivitya 2.772 2.200 2.728 2.268−(ε − 1)δ∆ ln P — −δ0.053 0 0

Exit/Entry Rate of Foreign — -1 0.017 -1Firms in Indonesia% of For. Non-Exporters 0.021 0 0.052 0% of Dom. Exporters 0.088 0 0 0.092%∆ in Dom. Exports — -1 -1 0

Dom. Mkt. Shr. of Dom. Non-Exp. 0.551 1 0.665 0.802Dom. Mkt. Shr. of Dom. Exp. 0.136 0 0 0.198Dom. Mkt. Shr. of For. Non-Exp. 0.149 0 0.335 0Dom. Mkt. Shr. of For. Exp. 0.165 0 0 0

Notes: a) Average productivity of all plants located in Indonesia in the steady state and is calculated using plant-level revenue shares asweights.

by comparing the steady state level of average productivity between the estimated (base)

model and that of the counterfactual experiments in the first row. The second row compares

welfare across the experiments and will be left to the end.

Eliminating trade and investment reduces average productivity in the textiles industry

by 20.6 percent. As I discuss below, much of the fall in productivity can be attributed

to the removal of foreign firms from the economy. I find a 18.1 percent fall in average

productivity when investment barriers alone are raised, while trade barriers cause a smaller

1.6 percent fall in average productivity. The explanation behind the large difference in the

magnitude of the fall in average productivity with trade and FDI restrictions lies in the

substitutability of trade and FDI. On one hand, trade restrictions reduce the incentive to

invest in Indonesia and use it as an export platform, hence, foreign firms are less likely to

invest in Indonesia. On the other hand, trade restrictions cause resources to be released

from all exporting plants, raising the average price level and increasing the profitability of

the Indonesian domestic market. In fact, column 3 suggests that the number of foreign

plants located in Indonesia rises by by 1.7 percent when there is no trade in Indonesia

53

with rest of the world. In this sense, one can think of FDI as a substitute for exports to

Indonesia. The entry and growth of foreign firms mitigates the resource reallocation from

domestic exporters to domestic non-exporters.

It is important to note that the increase in foreign firms rests heavily upon the assump-

tion of normality for the productivity distribution of foreign firms. However, even if there

was no new foreign entry into Indonesia due to the trade restrictions, our measure of aver-

age productivity in an economy with no trade would fall only by another 0.2 percent. Thus,

the model suggests that it is the growth of existing incumbents that largely mitigates the fall

in average productivity due to the trade restrictions. The economic implication is that the

Indonesian market, though small relative to the rest of the world, is large and profitable

enough that existing multinational firms would continue serving Indonesian consumers

even if they were unable to continue using Indonesia as an export platform. Moreover,

failing to account for the presence of foreign firms after the policy change would cause one

to overestimate the impact of trade restrictions on aggregate productivity.

The fourth column presents the results from a counterfactual experiment where trade

is allowed but FDI is not permitted. Average productivity drops by 18.1 percent without

FDI. Although the fall in productivity is almost as large as that under autarky it is largely

attributed to two features of the model. First, the fall is partly due the removal of the top

5% of the most productive firms, the foreign firms, in the industry.73Second, I construct

average productivity by weighting plant level productivity by revenue shares. Thus, while

foreign plants might only account for less than 5% of all plants, their productivity levels

receive almost one quarter of the total weight in the average productivity calculation. Due

to the direct impact that FDI restrictions have on the entry decision of foreign firms it is

73Note that when firms leave they take all of their knowledge and technology with them so their are nospillovers to the domestic industry.

54

clear that FDI policy can have a much larger impact on aggregate productivity in Indonesia

than trade. The estimates imply that the impact of FDI on aggregate productivity is 11

times that of international trade.

The welfare results are reported in the second row of Table 2.12. Welfare is measured as

the change in the inverse price level since increases in prices reduce the purchasing power

of consumers. The parameter δ captures the size of the textiles sector in the Indonesian

economy. Blalock and Gertler (2005) estimate that manufacturing composes approximately

one quarter of the Indonesian and I find that the textiles industry accounts for almost one

quarter of manufacturing which implies δ ≈ 0.055. Although this will greatly reduce the

size of the welfare impact it is clear from Table 2.12 that if similar changes occur across all

Indonesian manufacturing sectors there would be significant drop in overall welfare.

Table 2.12 documents a relatively large fall in welfare under autarky, but very small

changes to welfare when there are restrictions to trade or FDI. The intuition behind this

result is that trade (FDI) flows partially “insure” FDI (trade) flows in the presence of FDI

(trade) restrictions. For instance, a restriction to FDI is insured by the trade flows into

Indonesia so that Indonesian consumers can continue to access foreign goods through trade.

Similarly, trade restrictions are insured by foreign firms that enter and continue to produce

in Indonesia. Since the foreign firms tend to be much larger and more productive they tend

to greatly reduce the fall in welfare.

In sum, the estimates imply that the model’s predictions broadly match the features

of the Indonesian manufacturing data. Moreover, the estimates confirm the ranking of

productivity across plants with different ownership and export status as shown in Section

2. The counterfactual experiments indicate that FDI restrictions have a much larger im-

pact on aggregate productivity than international trade. The robustness results, presented

55

in Appendix A, document the same qualitative results across industries and very similar

quantitative results across estimation assumptions.

2.6 Conclusions and Extensions

This chapter presents and estimates a model of foreign direct investment and exports with

heterogeneous firms. I show that the model can generate productivity differences across

plants with different ownership and export status which are consistent with the observed

differences in the Indonesian manufacturing data. Using the theoretical model and a panel

of Indonesian manufacturing plants, I develop and estimate a structural empirical model of

exports and foreign direct investment.

The model’s empirical predictions broadly match the features of the Indonesian man-

ufacturing data. In particular, the model captures export decisions at the plant level and

documents the differential export behaviour across foreign and domestic firms. Moreover,

the estimates confirm the ranking of productivity across plants with different ownership and

export status as shown in Section 2. The model emphasizes that accounting for FDI flows is

essential to recovering accurate estimates of the impact of trade on aggregate productivity.

In particular, the counterfactual experiments imply that the impact of trade on productivity

is greatly mitigated by FDI flows and that trade restrictions may even encourage FDI flows

to Indonesia.

The counterfactual experiments imply that FDI restrictions can reduce aggregate pro-

ductivity by 3 to 11 times more than trade restrictions across FDI-intensive industries in

Indonesia. I find that the impact of FDI restrictions account for a fall in average total factor

productivity between 8 and 27 percent across industries. Trade restrictions, in contrast, are

estimated to have a smaller impact on average productivity. Across the food, manufactured

56

metals and textiles industries average total factor productivity is estimated to fall by 1 to 4

percent.

The results suggest that policies which induce inwards flows of FDI will have a much

larger impact on aggregate productivity relative to those that encourage exports. Moreover,

the results imply that the differential impact international policies have on foreign and do-

mestic firms can lead to drastically different results. These results are particularly important

for policymakers in developing countries where the interaction between trade and foreign

direct investment policy has been largely unexamined.

I also find that the welfare implications differ substantially across international inte-

gration policies. Autarky causes substantial reductions in welfare, while trade and FDI

restrictions cause much smaller welfare impacts on an economy such as Indonesia. The

model suggests that trade and FDI act as substitutes for each other and reduce the welfare

impact of trade or FDI restrictions.

The results suggest a number of extensions and interesting questions for future work.

There is growing evidence that entering foreign markets may require firms to pay both

sunk costs and per-period fixed costs (see Das, S., M.J. Roberts, and J.R. Tybout (2007)

and Kasahara and Lapham (2007) for examples). Additional sunk costs may also improve

the model’s ability to match the observed persistence export and ownership status. This

extension will be further examined in the next chapter.

57

Chapter 3

Foreign Direct Investment, Exports and

Aggregate Productivity:

Sunk vs. Fixed Costs

3.1 Introduction

This chapter extends the model presented in Chapter 2 to allow for sunk costs associated

with export and FDI decisions. The magnitude of these costs are estimated from the In-

donesian manufacturing data for both foreign and domestic plants. The estimated sunk

costs play an important role in the entry and exit decisions of producers in international

markets, as emphasized in the counterfactual experiments.

A number of recent papers suggest that firms who access foreign markets are typically

subject to large, one-time sunk costs. Das, Roberts and Tybout (2007), Ruhl and Willis

(2007) report large initial sunk costs associated with export behaviour, while Kashara and

58

Lapham (2007) find large initial export and import sunk costs. Sunk costs have proven

particularly important role in the ability of modern empirical models to match the plant-

level dynamics of entry and exit in international markets.

The results from Chapter 2 suggest that sunk costs may play a particularly important in

explaining the entry and exit dynamics of foreign firms. As shown in the preceding chapter

the model had difficulty simultaneously replicating the distribution of foreign firms along

with the low exit rate of foreign firms. For a model with fixed costs alone to match the dis-

tribution of foreign firms it must encourage some low productivity firms to enter Indonesia.

In the model presented in Chapter 2 this would occur if low productivity foreign firms

received large beneficial cost shocks. However, since there is no persistence in the cost

shocks, and fixed costs are high, low productivity foreign firms that enter due to a benefi-

cial cost shock are very likely to exit in subsequent years. This drives up the observed exit

rate among foreign firms which is inconsistent with the low observed exit rate of foreign

firms.

Sunk FDI costs allow for the persistent presence of low productivity foreign owned

firms in Indonesia. Low productivity foreign firms that enter when they receive beneficial

cost shocks are substantially less likely to exit in future periods because the sunk invest-

ment, and higher future profitability, in Indonesia will be lost. Moreover, by allowing for

sunk costs, the estimated fixed costs are much smaller, reinforcing the persistence of for-

eign firms in Indonesia.

The rest of this chapter is organized as follows. The next section describes the modi-

fications to the model presented in Chapter 2. The third section describes the augmented

structural estimation of the model with sunk costs, while the fourth section presents the

results and compares them to the model without sunk costs. The fifth section concludes

59

and discusses future work.

3.2 An Augmented Cost Structure

The economic environment presented in the preceding chapter will remain largely un-

changed with the exception that firms will now be subject to sunk costs associated with

their export and FDI decisions. Given a firm’s current and past FDI and export decisions

they may incur the following set of fixed and sunk costs in a given period:

F (dit, di,t−1) =

f for (dfdiit , dx

it) = (0, 0)

f + fx + cx(1− dxi,t−1) for (dfdi

it , dxit) = (0, 1)

f + f∗ + f∗fdi + c∗fdi(1− dfdii,t−1) for (dfdi

it , dxit) = (1, 0)

f∗ + ζfxf∗x + f∗fdi + [ζcxc∗x(1− dxi,t−1) + c∗fdi](1− dfdi

i,t−1) for (dfdiit , dx

it) = (1, 1)

where starred variables denote costs incurred in the foreign country. The sunk cost of

exporting is denoted by cx, while cfdi is the sunk cost of maintaining a plant aborad. I

interpret the sunk and fixed export and FDI costs as forming and maintaining distribution

and service networks in export markets. It is important to note that firms must pay this cost

even if they are exporting back to their home market, but the sunk and fixed cost may be

reduced through the parameters ζcx and ζfx . These parameters capture the complementarity

between foreign ownership and exporting to foreign markets. The fixed and sunk FDI costs

include the distribution, servicing, overhead and set up costs of operating a subsidiary

abroad. The relative size of the fixed and sunk costs plays an important role in determining

each firm’s optimal production and export decisions.

60

3.2.1 Exit & Entry Decisions

I now examine the dynamic component of a manufacturer’s decision problem. The sunk

costs are an extra state variable for firms when they make their export and FDI decisions.

At the beginning of each period, a producer’s state is given by their productivity level ai

and their previous FDI and export decisions, di,t−1. The Bellman equations for a producer

originating in the home country are written as follows:

V (ai, di,t−1) =

∫max{εχ(0), W (ai, di,t−1) + εχ(1)}dHχ(εχ), (3.1)

W (ai, di,t−1) =

∫max{J(ai, di,t−1, 0) + εfdi(0),

J(ai, di,t−1, 1) + εfdi(1)}dHfdi(εfdi) (3.2)

J(ai, di,t−1, 0) =

∫ (maxdx′

π(ai, di,t−1, dx′)

+βV (ai, d′) + εx(dx′)

)dHx(εx) (3.3)

J(ai, di,t−1, 1) =

∫ (maxdx′

π(ai, di,t−1, dx′)

+βV (ai, d′) + εx∗(dx′)

)dHx∗(εx∗) (3.4)

where Hχ, Hfdi, Hx and Hx∗ represent the cumulative distributions of εχ, εfdi, εx and εx∗ ,

respectively. As before, the value function V (·, ·) characterizes the firm’s exit decision,

while W (·, ·) describes the firm’s decision whether to engage in FDI. Similarly, J(·, ·, 1)

is the continuation value of a firm which has decided to maintain a plant abroad, while

J(·, ·, 0) is the continuation value of a firm that only produces in its country of origin. Note

that export decisions are made after the FDI decision.

While the firm’s decision problem appears very similar to the problem presented in the

preceding chapter, there is an important difference. A firm’s decision to export or maintain

a plant abroad will now depend on its own history of decisions. Thus, in the presence of

large sunk costs, plants who have previously exported are more likely to continue exporting

61

in the presence of sunk export costs, even if they receive a negative cost shock, since the

firm has already invested in the export market.

3.2.2 Equilibrium

As before, I focus on a stationary equilibrium where aggregate variables and the distribution

of productivity are constant over time. I denote the stationary distribution of productivity

across incumbent firms by µ(a, d), where I am now considering the joint distribution of

productivity and plant-level decisions. Although firms exit home and foreign markets as

they receive the various firm-specific shocks outlined above, these changes are exactly off-

set by new entrants of each type in equilibrium. For convenience, I drop the time subscript

and denote the firm’s state as (ai, di).

The free entry condition from the preceding chapter is unchanged:∫V (a, (0, 0))ga(a)da = fe (3.5)

where ga(a) is the distribution of initial productivity draws.

The stationary equilibrium requires that the number of exiting firms must equal the

number of successful new entrants. Specifically,

M

∫ (∑d

P χ(χ = 0|a, d)µ(a, d)

)da = Me

∫P χ(χ = 1|a, (0, 0))ga(a)da, (3.6)

where the choice probabilities now depend on both the plant productivity, a, and its FDI

and export decisions d. The probability of producing this period (χ = 1) is calculated as:

P χ(χ = 1|a, d) = (1− ξ)

(exp(W (a, d)/%χ)

exp(0) + exp(W (a, d)/%χ)

)(3.7)

where ξ is again the exogenous probability of exit and W (·, ·) is defined as in equation

(3.2).

62

The final condition required for a stationary equilibrium is that the measure of firms

with state (a, d) is constant over time. Conditional on operating, the probability of main-

taining a plant abroad is given by

P fdi(dfdi′ = 1|a, d, χ = 1) =exp(J(a, d, (1, dx)/%fdi)∑

dfdi exp(J(a, d, (dfdi, dx))/%fdi)(3.8)

while the conditional probability of producing everything in the country of origin is given

by P fdi(dfdi′ = 0|a, d, χ = 1) = 1 − P fdi(dfdi′ = 1|a, d, χ = 1) and J(·, ·, ·) is given in

equations (3.3)-(3.4). Simlarly, conditional on operating and choosing to maintain a plant

abroad, the firm’s export choice probability can be calculated as

P x(dx′|a, d, χ = 1, dfdi′ = 1) =

exp([π(a, d, (dx′ , 1)) + βV (a, d′)]/%x∗)∑dx′ exp([π(a, d, ((dx′ , 1)) + βV (a, d′)]/%x∗)

. (3.9)

The choice probability for exporting conditional on not engaging in FDI is analagous to

(3.9) where the scale parameters are replaced with their unstarred counterparts. As in the

previous chapter, define the probability of FDI and export as

P dθ (dit|ai, di,t−1) = P fdi

θ (dfdiit |ai, di,t−1)P

xθ (dx

it|ai, di,t−1, dfdii,t−1)

where P fdiθ (dfdi

it = 1|ai, di,t−1) = 0 for Indonesian plants by assumption. The equilibrium

condition can therefore be written as

Mµ(a, d) = MP χ(χ = 1|a, d)∑d′

P d(d|a, d′, χ = 1)µ(a, d)

+ MePχ(χ = 1|a, (0, 0))P d(d|a, (0, 0), χ = 1)ga(a) (3.10)

The first term on the right-hand side of (3.10) is the measure of incumbent producers who

do not exit and have state (a, d) in the current period. The second term is the measure of

new entrants with the same state.

63

3.3 Structural Estimation

The structural estimation of the model presented above will closely follow that described in

the previous chapter. The following sections will highlight the differences made to identify

and accommodate the sunk costs. In particular, the presence of the sunk costs substan-

tially increase the computational burden of the estimation by increasing the state space of

dynamic programming problem.

3.3.1 The Likelihood Function

A firm’s detrended net profit now depends on both its current and past decisions:

π(a, di,t−1, dit) = r(a, dit)− F (di,t−1, dit) (3.11)

As before, it is not possible to identify all of the cost components. In particular, it is

not possible to identify the additional sunk export costs parameter, c∗x. As with the other

unidentified parameters, ϕw, f ∗d and f ∗

x , I will calibrate the sunk export cost parameter as

c∗x = 0.38cx using information from the World Bank Doing Business Report. With these, I

test the sensitivity of this assumption and present robustness results in Appendix B.

There are three additional parameters to be estimated: the sunk export cost, cx, the sunk

FDI cost, cfdi, and the parameter capturing the complementarity between foreign ownership

and sunk export costs, ζcx . As before, the vector of parameters θ is estimated by the method

of maximum likelihood where θ is now1

θ = (ρ, ϕb, ϕτ , f, fx, ffdi, cx, cfdi, ζfx , ζcx , ξ, ε, %χ, %x, %x∗ , %fdi, σa, σa∗).

1The discount factor is not estimated and is set to 0.95. It is difficult to identify the discount factor β indynamic discrete choice models. I also assume that the mean of the distribution of initial productivity drawsis the same for Indonesian and foreign firms.

64

As in Chapter 2, let the variable χit be a variable that takes a value of 0 if a plant exits

the data and 1 otherwise. The probability of χ = 1 for domestic plants is then P χθ (χit =

0|ai, di,t−1) = P χθ (χit = 0|ai, di,t−1), while the probability that a foreign plant exits is:

P χ∗

θ (χ∗it = 0|ai, di,t−1) = P χ

θ (χit = 0|ai, di,t−1)+P χθ (χit = 1|ai, di,t−1)P

fdiθ (dfdi

it = 0|ai, di,t−1).

Recall that Ti,0 is the first year the firm appears in the data. Then, conditional on ai and

di,t−1 the likelihood contribution of plant i in year t > Ti,0 is

Lit(θ|ai, di,t−1) =

P χ

θ (χit = 0|ai, di,t−1) for χit = 0

P χθ (χit = 1|ai, di,t−1)︸ ︷︷ ︸

Stay/Exit

P d(dit|ai, di,t−1, χit = 1)︸ ︷︷ ︸FDI/Export

gν(νit(ai))︸ ︷︷ ︸Revenue

for χit = 1

where starred values replace unstarred values for foreign plants. Note that the likelihood

contribution is now conditional on the plant’s previous export and ownership status.

Following the procedure described in Chapter 2, the likelihood contribution from plants

in the initial year is

Lit(θ|ai, di,t−1) = P d(dit|ai, di,t−1, χit = 1)gν(νit(ai)). (3.12)

The overall likelihood contribution from plants observable in the initial year is

Li(θ|ai, di,Ti,0) =

Ti,1∏t=Ti,0+1

Lit(θ|ai, di,t−1)

where Ti,1 is the last year plant i appears in the data. I numerically integrate out unobserved

plant-specific productivity ai to calculate the likelihood contribution from each plant i as

Li(θ) =

∫

Li(θ|a′, di,Ti,0)µ(a′, di,Ti,0

)da′ for Ti,0 = 1993,∫Li(θ|a′, di,Ti,0

)P dθ (di,Ti,0

|a′, di,Ti,0−1)gea(a

′)da′ for Ti,0 > 1993,

where the starred distributions are used in place of the unstarred distributions for foreign

firms and the stationary distribution of foreign plants is conditional on entering Indonesia,

65

µ∗(a′, di,Ti,0|dfdi

i,Ti,0= 1). The parameter vector is again estimated by maximizing the likeli-

hood function. The rest of the estimation procedure is identical to that described in Chapter

2.

3.3.2 Sunk Cost Identification

The sunk cost parameters are identified by two different sources of variation. The sunk ex-

port cost is identified from the differences in export frequencies across plants with similar

productivity levels but different export histories. The sunk FDI costs are identified by the

entry and exit dynamics of foreign firms, given the parametric assumption on the distribu-

tion of productivity. Specifically, a model without sunk costs cannot match the distribution

of productivity among foreign owned firms in Indonesia, because both high fixed costs and

low costs shocks are required to match the higher productivity and low exit rates across

foreign owned plants. Thus, there are few low productivity, foreign owned plants. A model

with sunk costs can do a better job of simultaneously matching these features since the

previous investment in Indonesia will encourage low productivity foreign plants to stay in

Indonesia even in the presence of cost shocks.

3.4 Estimation Results

Table 3.1 presents the maximum likelihood parameter estimates of the empirical model

along with the associated asymptotic standard errors. The standard errors are computed

using the outer product of gradients estimator. The parameters are evaluated in millions of

Indonesian rupiahs in 1983.

66

Table 3.1: Structural EstimatesIndustry Food Metals Textiles

κ%χ 52.361 (4.567) 51.940 (6.215) 80.406 (2.2e-06)κ%x 8.812 (0.541) 8.234 (0.901) 6.634 (0.542)κ%x∗ 0.669 (1.794) 0.225 (2.067) 0.220 (0.116)κ%fdi 0.959 (0.214) 3.360 (0.168) 3.337 (0.186)κcx 35.619 (2.194) 29.258 (3.352) 20.921 (1.802)

κcfdi 38.907 (2.254) 50.062 (2.185) 47.054 (3.044)κf 1.221 (0.881) 0.692 (1.313) 5.749 (0.880)κfx 2.217 (0.659) 1.342 (0.795) 1.166 (0.376)

κffdi 1.842 (2.000) 0.792 (2.590) 0.048 (0.376)ζfx 0.098 (0.507) 0.028 (2.154) 0.051 (1.394)ζcx 0.004 (0.010) 0.009 (0.081) 0.006 (0.006)ρ 0.004 (0.008) 0.005 (0.012) 4.7e-06 (0.003)

ϕB 0.975 (0.015) 1.956 (0.051) 1.692 (0.024)ϕτ -5.331 (0.052) -3.803 (0.084) -5.437 (0.070)σa 0.778 (0.005) 0.742 (0.006) 0.779 (0.004)σa∗ 0.779 (0.007) 0.712 (0.010) 0.829 (0.015)ξ 0.001 (0.0001) 0.007 (0.001) 0.001 (0.0001)

ε = 1/mark-up 3.8 3.0 3.8log-likelihood -34795.114 -16347.454 -31431.706No. of Obs. 17,786 7,549 13,287

Notes: Standard errors are in parentheses. The parameters are evaluated in units of millions of Indonesian Rupiahs in 1983. Metalsrefers to manufactured metals.

3.4.1 Sunk and Fixed Costs

The estimates of sunk export and FDI costs across industries are substantial. The estimated

sunk cost of export for a domestic exporter ranges from 188 thousand 1983 US dollars in

the metals industry to 276 and 278 thousand US dollars in the food and textiles industries,

respectively. Across industries, the sunk costs are greater than one year’s worth of export

revenue. In contrast, foreign plants are estimated to only pay approximately 1% of the

sunk export costs faced by their domestic counterparts. Sunk FDI costs are estimated to

be even larger. Across industries the sunk FDI costs range from 301 thousand 1983 US

dollars (food) to 329 thousand 1983 US dollars (textiles). However, these costs account for

a smaller percentage of total revenue; across industries the sunk FDI costs account for 65%

(food) to 112% (metals) of total revenue earned by foreign plants in Indonesia. Although

the average sunk costs are typically a large percentage revenue it is important to recall that

67

the sunk costs incurred are substantially lower since the estimated model predicts many

plants only export or invest abroad when they receive beneficial cost shocks.

The fixed costs of operation are typically much smaller than the estimated sunk costs.

However, they also tend to display much more variation across industries and are much

more sensitive to the estimation assumptions than the model without sunk costs. The av-

erage fixed cost of operation in Indonesia ranges from 3 to 23 thousand 1983 US dollars

in the manufactured metals industry, 10 to 17 thousand dollars in the food industry, and

10 to 40 thousand 1983 US dollars for the textiles industry. The fixed export costs show

less variation across different estimates. The average fixed export costs in the base estima-

tion ranges from 8 thousand 1983 US dollars in the textiles and manufactured metals to 17

thousand 1983 US dollars in the food industry. As with the sunk costs, foreign firms are

estimated to only pay a small percentage of the fixed export costs (1 to 10% across indus-

tries). The fixed FDI costs are less than 6 thousand 1983 US dollars in all three industries

and are often insignificant

There are large differences between the fixed cost estimates presented here and those

presented in the preceding chapter. In particular, I note that the estimated average fixed

costs fall dramatically once sunk costs are introduced. This reflects both the persistent

nature of export status and the low exit rates among foreign plants. The model without

sunk costs inflated the fixed cost parameters to match the large productivity differences but

could not capture the dynamic elements of the data.

3.4.2 Exports and FDI

The performance of the sunk cost model along the cross-sectional attributes of the data is

similar to the model without sunk costs. The estimates imply that exporting will increase

68

Table 3.2: Distribution of Plants by Ownership/Export StatusActual Domestic Non-Exporters Domestic Exporters Foreign Non-Exporters Foreign ExportersFood 0.931 0.046 0.016 0.008Metals 0.824 0.071 0.058 0.047Textiles 0.839 0.114 0.025 0.023PredictedFood 0.932 0.044 0.017 0.007Metals 0.838 0.057 0.049 0.055Textiles 0.832 0.120 0.018 0.030

Note: Metals refers to manufactured metals.

the average plant’s revenues by 51, 52 and 101 percent in the manufactured metals, textiles

and food industries. The estimates also imply that on average 34 percent of a domestic ex-

porters’ revenues are from export sales in the manufactured metals and textiles industries,

while exporters in the food industry receive 50 percent of revenues from export sales. Rel-

ative to the data, the model underpredicts the average percentage of revenues from exports

in the manufactured metals and textiles industries by 7 and 19 percent, respectively, while

overpredicting the average percentage of revenues from exports in the food industry by 1

percent.

Table 3.2 documents the predicted distribution of plants export and ownership status

for all industries. Again, the model’s predictions match the distribution of plants across

export and ownership status very closely in all industries. Table 3.3 reports the models

predicted domestic market and export shares across export and ownership status. While

one would expect the model to predict the dominant role of foreign plants in the domestic

and export markets, Table 3.3 demonstrates that it often overstates their importance in

both domestic and export markets. Because of the substantial productivity difference in

foreign and domestic plants, small differences in the distribution of plants or the predicted

productivity differences can have large impacts at the aggregate level. The model performs

similarly to the model without sunk costs presented in the preceding chapter.

69

Table 3.3: Export/Domestic Market Share by Ownership/Export OwnershipExport Share Domestic Market Share

Actual Domestic Foreign Domestic Non-Exp. Domestic Exp. Foreign Non-Exp. Foreign Exp.Food 0.837 0.163 0.551 0.289 0.126 0.032Metals 0.458 0.542 0.462 0.099 0.300 0.138Textiles 0.717 0.283 0.602 0.238 0.097 0.062PredictedFood 0.542 0.458 0.550 0.110 0.248 0.093Metals 0.165 0.835 0.444 0.051 0.248 0.257Textiles 0.643 0.357 0.450 0.274 0.125 0.152

Note: Metals refers to manufactured metals.

3.4.3 Productivity

As in the preceding chapter, the model predicts that only the most productive domestic

firms are able to produce profitably in Indonesia. Figure 3.1 shows the importance of sur-

vival selection among domestic plants in Indonesia. The actual productivity distribution

for incumbents in the top three panels is skewed to the right relative to the distribution of

new entrants for all three industries. The bottom three panels of Figure 3.1 show that the

model’s predicted productivity distributions for entrants and incumbents capture this selec-

tion mechanism. Similarly, Table 3.4 confirms that domestic incumbents are on average 12

to 21 percent more productive than new domestic entrants across industries.

Figure 3.2 shows the productivity distributions for foreign entrants and incumbents.

One striking difference between foreign and domestic firms is that the model predicts that

foreign entrants will be more productive than foreign incumbents. This is also in contrast to

the model presented in Chapter 2 and is due to the additional sunk cost feature of the model.

Low productivity foreign plants are less willing to leave Indonesia when they receive a

negative cost shock since they expect much higher future returns from the plant. Moreover,

because they are low productivity plants they are unlikely to receive another cost shock in

the future that will make re-investment in Indonesia profitable if they leave. As such, there

70

−6 −4 −2 0 2 4 60

0.05

0.1

0.15

0.2

0.25

0.3

log(1/a)

Pre

dict

ed F

requ

ency

Food

−6 −4 −2 0 2 4 60

0.05

0.1

0.15

0.2

0.25

log(1/a)+νit

Act

ual F

requ

ency

Food

EntrantsIncumbents

EntrantsIncumbents

−6 −4 −2 0 2 4 60

0.05

0.1

0.15

0.2

0.25

log(1/a)+νit

Act

ual F

requ

ency

Metals

EntrantsIncumbents

−6 −4 −2 0 2 4 60

0.05

0.1

0.15

0.2

0.25

0.3

log(1/a)

Pre

dict

ed F

requ

ency

Metals

EntrantsIncumbents

−6 −4 −2 0 2 4 60

0.05

0.1

0.15

0.2

0.25

log(1/a)+νit

Act

ual F

requ

ency

Textiles

EntrantsIncumbents

−6 −4 −2 0 2 4 60

0.05

0.1

0.15

0.2

0.25

0.3

log(1/a)

Pre

dict

ed F

requ

ency

Textiles

EntrantsIncumbents

Figure 3.1: Productivity Distributions of Domestic Entrants and Incumbents (Actual vs.Predicted)

is an accumulation of lower productivity foreign plants.

Relative to Figure 2.4, it is clear that allowing for sunk costs substantially flattens out

the distribution of productivity among foreign plants, but still does not capture all of the

heterogeneity in productivity. A likely explanation for these results is that there are other

sources of heterogeneity that are not captured by the model. For instance, exporting firms

vary widely in the percentage exported and the use of imported materials. This is an im-

portant source of heterogeneity in previous work (see Kasahara and Lapham (2007) for

example).

71

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

log(1/a)+νit

Act

ual F

requ

ency

Food

EntrantsIncumbents

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

log(1/a)

Pre

dict

ed F

requ

ency

Food

EntrantsIncumbents

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

log(1/a)+νit

Act

ual F

requ

ency

Metals

EntrantsIncumbents

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

log(1/a)

Pre

dict

ed F

requ

ency

Metals

EntrantsIncumbents

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

log(1/a)+νit

Act

ual F

requ

ency

Textiles

EntrantsIncumbents

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

log(1/a)

Pre

dict

ed F

requ

ency

Textiles

EntrantsIncumbents

Figure 3.2: Productivity Distributions of Foreign Entrants and Incumbents (Actual vs. Pre-dicted)

As shown in Section 2.2, domestic exporters tend to have higher productivity than do-

mestic non-exporters, while the opposite is true for foreign plants. Also, foreign plants

tend to be more productive than domestic plants. Figure 3.3 shows the actual and pre-

dicted productivity distributions for domestic non-exporters, domestic exporters, foreign

non-exporters and foreign exporters, while Table 3.4 reports their average productivities.

The top panel shows that in all three industries the plants follow the same productivity

ranking that we observed in the Section 2. In particular, it is notable that the productivity

distribution of foreign non-exporters is skewed to the right of the productivity distribution

of foreign exporters, though the differences across foreign plants are less clear.

72

Table 3.4: Average ProductivityFood Metals Textiles

Indonesian Foreign Indonesian Foreign Indonesian ForeignMean at entry trial 1.000 1.000 1.000 1.000 1.000 1.000Mean at successful entry in Indonesia 1.044 4.242 1.080 4.463 1.045 3.511Mean at steady state in Indonesia 1.168 3.951 1.302 4.204 1.233 2.971Non-Exporters 1.138 3.996 1.280 4.295 1.145 3.186Exporters 1.786 3.843 1.617 4.123 1.846 2.841

Note: Metals refers to manufactured metals.

Again, the large difference between foreign and domestic firms is quite evident. The

predicted distribution matches these rankings and Table 3.4 confirms that on average for-

eign non-exporters are the most productive plants, followed by foreign exporters, domestic

exporters and domestic non-exporters. Moreover, the estimated model with sunk costs pre-

dicts larger productivity differences between foreign exporters and foreign non-exporters

than the model without sunk costs. This result occurs because sunk FDI costs allow for a

larger percentage of low productivity foreign plants in Indonesia.

3.4.4 Dynamics

Table 3.5 documents the actual and predicted transition probabilities of FDI, export and

exit in the textiles industry.2 The model captures the persistence in export and ownership

status and does a substantially improves the performance of the model along this dimension

relative to the model without sunk costs.

3.4.5 Counterfactual Experiments

I repeat the set of counterfactual experiments presented in the preceding chapter. Specifi-

cally, to determine the quantitative implications of barriers to trade and FDI, I conduct the2Similar tables can be found in Appendix B for the food and manufactured metals industries.

73

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

log(1/a)+νit

Act

ual F

requ

ency

Food

Domestic Non−ExportersDomestic ExportersForeign ExportersForeign Non−Exporters

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

log(1/a)

Pre

dict

ed F

requ

ency

Food

Domestic Non−ExportersDomestic ExportersForeign ExportersForeign Non−Exporters

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

log(1/a)+νit

Act

ual F

requ

ency

Metals

Domestic Non−ExportersDomestic ExportersForeign ExportersForeign Non−Exporters

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

log(1/a)

Pre

dict

ed F

requ

ency

Metals

Domestic Non−ExportersDomestic ExportersForeign ExportersForeign Non−Exporters

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

log(1/a)+νit

Act

ual F

requ

ency

Textiles

Domestic Non−ExportersDomestic ExportersForeign ExportersForeign Non−Exporters

−5 0 5 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

log(1/a)

Pre

dict

ed F

requ

ency

Textiles

Domestic Non−ExportersDomestic ExportersForeign ExportersForeign Non−Exporters

Figure 3.3: Productivity Distributions of Domestic and Foreign Firms by Export Status(Actual vs. Predicted)

following three experiments by manipulating three counterfactual parameters:

1. Autarky: fx, fx∗ , ffdi →∞,

2. No Trade: fx, fx∗ →∞,

3. No Investment in Indonesia: ffdi →∞,

Table 3.6 presents the results from the counterfactual experiments in the textiles indus-

try.3 Again, restricting FDI causes a large fall in average productivity, while trade barriers

3Results for the manufactured metals and food industries are presented in Appendix B.

74

Table 3.5: Transition Probabilities - TextilesActual Dom. Non-Exporters Dom. Exporters For. Non-Exporters For. Exporters ExitDom. Non-Exporters at t 0.945 0.051 0.003 0.001 0.088Dom. Exporters at t 0.309 0.679 0.001 0.011 0.068For. Non-Exporters at t 0.138 0.031 0.581 0.256 0.030For. Exporters at t 0.027 0.066 0.178 0.732 0.016PredictedDom. Non-Exporters at t 0.927 0.073 — — 0.110Dom. Exporters at t 0.479 0.521 — — 0.054For. Non-Exporters at t — — 0.556 0.444 0.001For. Exporters at t — — 0.272 0.728 0.011

Table 3.6: Counterfactual Experiments - TextilesBase Autarky No Trade No FDI

Avg. Productivitya 2.404 1.865 2.251 1.949−(ε − 1)δ∆ ln P — −δ0.048 −δ0.014 -δ0.008

Exit/Entry Rate of Foreign — -1 -0.021 -1Firms in Indonesia% of For. Non-Exporters 0.018 0 0.051 0% of Dom. Exporters 0.120 0 0 0.107%∆ in Exports — -1 -1 -0.615

Dom. Mkt. Shr. of Dom. Non-Exp. 0.450 1 0.655 0.715Dom. Mkt. Shr. of Dom. Exp. 0.274 0 0 0.285Dom. Mkt. Shr. of For. Non-Exp. 0.125 0 0.345 0Dom. Mkt. Shr. of For. Exp. 0.152 0 0 0

Notes: a) Average productivity of all plants located in Indonesia in the steady state and is calculated using plant-level revenue shares asweights.

75

cause a much smaller fall in average productivity. Eliminating trade and investment to-

gether reduces average productivity in the textiles industry by 22.4 percent. I find a 18.9

fall in average productivity when FDI barriers alone are raised, while trade barriers cause

a smaller 6.4 percent fall in average productivity. Column 3 suggests that the number of

foreign plants located in Indonesia falls by 2.1 percent when there is no trade between In-

donesia and the rest of the world. The counterfactual experiments from the model with

sunk costs confirm the previous result that the entry and growth of foreign firms mitigates

the resource reallocation from domestic exporters to domestic non-exporters.4

The fourth column presents the results from a counterfactual experiment where trade is

allowed but FDI is not permitted. Again, the fall in productivity is as almost large as that

under autarky. This is because the top 5% of the most productive firms, the foreign firms,

are removed from the industry and these firms are weighted very heavily. The estimates

imply that the impact of FDI on aggregate productivity is 11 times that of international

trade.

The welfare results are reported in the second row of Table 3.6. As in Chapter 2, there

is a relatively large fall in welfare under autarky, but smaller changes to welfare when there

are restrictions to trade or FDI. These results confirm the finding that models that of trade

or FDI may greatly overestimate the cost of trade or FDI barriers if they do not take into

account the substitutability of these flows.

The results presented here confirm those documented in Chapter 2. It is not necessarily

surprising that the qualitative implications of the counterfactual experiments are similar

across the two models since the counterfactual experiments focus on steady-states rather

than transitional dynamics. However, the model with sunk costs improves the fit of the

4Note that even if there was no new foreign entry into Indonesia due to the trade restrictions, our measureof average productivity in an economy with no trade would fall only by another 0.5 percent.

76

model to the observed data.

3.5 Conclusions and Extensions

This chapter estimates a model of foreign direct investment and exports with heterogeneous

firms and sunk costs. The model’s empirical predictions broadly match the features of the

Indonesian manufacturing data and improves the model’s dynamic performance relative

to the model without sunk costs. The estimates again confirm the ranking of productivity

across plants with different ownership and export status and emphasizes that accounting for

FDI flows is essential to recovering accurate estimates of the impact of trade on aggregate

productivity.

The counterfactual experiments imply that FDI restrictions reduce aggregate produc-

tivity by at least three times more than trade restrictions across FDI-intensive industries in

Indonesia. I find that the impact of FDI restrictions account for a fall in average total factor

productivity between 18 and 41 percent across industries. Trade restrictions, in contrast,

are estimated to have a smaller impact on average productivity. Across the food, manufac-

tured metals and textiles industries average total factor productivity is estimated to fall by

less than 6 percent.

I confirm that the welfare implications differ substantially across international inte-

gration policies. Autarky causes substantial reductions in welfare, while trade and FDI

restrictions cause much smaller welfare impacts on an economy such as Indonesia. The

model suggests that trade and FDI act as substitutes for each other and reduce the welfare

impact of trade or FDI restrictions.

This chapter’s results suggest a number of interesting directions for future work. The

77

data demonstrates that manufacturing plants in Indonesia demonstrate substantial hetero-

geneity across export intensity and import behaviour. Extending the model to allow richer

export and import patterns may uncover other dimensions of interaction across foreign di-

rect investment, export decisions and import decisions. I intend to address these issues in

my future research.

78

Chapter 4

Does the Use of Imported Intermediates

Increase Productivity?

Plant-Level Evidence

This chapter was co-authored with Hiroyuki Kasahara.

4.1 Introduction

International trade is one of the primary avenues for the diffusion and adoption of new

technologies worldwide. This is particularly true for developing nations where it is be-

lieved that importing new technologies is a significant source of productivity and economic

growth. Through adoption and imitation of imported technologies, countries can take ad-

vantage of research and development (R&D) abroad to improve the efficiency of domestic

production.

79

Previous empirical work using aggregate cross-country data shows that importing in-

termediate goods that embody R&D from an industrial country can significantly boost

a country’s productivity (cf., Coe and Helpman, 1995; Coe, Helpman and Hoffmaister,

1997). Countries that are more open to trade benefit more from foreign R&D because they

are better able to access improvements in technology by importing intermediate goods.1

Aggregated data, however, do not capture heterogeneity across different plants in the econ-

omy. As empirically shown by Baily, Hulten, and David (1992) and Pavcnik (2002), it is

vital to examine plant-level changes in order to understand changes in aggregate produc-

tivity levels. Furthermore, recent developments in trade theory suggest that understanding

the plant-level response to trade policy is a crucial factor in understanding its impact on ag-

gregate productivity and welfare (e.g., Melitz, 2003; Bernard, Eaton, Jensen and Kortum,

2003).

The goal of this chapter is to test whether the use of foreign intermediate goods in-

creases plant productivity, using a detailed panel data set on Chilean manufacturing plants

from 1979-1996. The data set captures heterogeneity in terms of import status across plants

and across time: some plants import most of their intermediate goods, some change their

import status, others do not import at all. While importers are larger and more produc-

tive than non-importers in the data, the direction of causality between importing foreign

intermediates and plant’s performance is not immediately obvious.

Does the use of foreign intermediate goods directly increase productivity or do inher-

ently high productivity plants tend to use foreign intermediate goods? To answer these

questions, we estimate both the immediate and long-run effects from the use of imported

intermediates on plant’s productivity while addressing the important econometric issues

1Keller (2001) provides the industry-level empirical evidence for the role of R&D spillovers throughimports.

80

of simultaneity and endogenous selection using the Within-Group estimator, the System

GMM estimator (cf., Blundell and Bond, 1998), and the estimator developed by Olley and

Pakes (1996) and Levinsohn and Petrin (2003) (OP/LP Proxy estimator, hereafter).

The results across different estimators indicate a statistically significant, often substan-

tial, positive impact from the use of imported intermediates on plant productivity. The

System GMM and the OP/LP Proxy point estimates suggest a positive impact from the use

of imported intermediates on productivity ranging from 12.9 to 22.0 percent. There exists

substantial uncertainty over the different estimators in regards to the precise magnitude of

the positive effect from importing. However, even the Within-Groups estimates—which

are likely to be biased downward due to measurement errors in imported intermediates (cf.,

Griliches and Hausman, 1986)—show that the use of imported intermediates leads to an

immediate increase in productivity of 2.6 percent. In addition, we find some evidence for

a positive dynamic effect from the use of imported intermediates; the evidence suggests

that past import status has a positive impact on current productivity (i.e., “learning by im-

porting"). We also examine the sensitivity of our results to export status and time-varying

industry-specific effects. Overall, the results indicate the robustness of the effect of inter-

mediate imports on productivity across various estimation methods.

Among recent papers discussing the impact of foreign intermediate inputs on produc-

tivity at the micro-level are Biesebroeck (2003), Muendler (2004), Amiti and Koenings

(2005), and Halpern, Koren, and Szeidl (2006). The empirical findings in the literature are

mixed. Biesebroeck (2003) finds that productivity improvements do not happen through

more advanced inputs in Columbia and, similarly, Muendler (2004) shows only a small

contribution of foreign materials and investment goods on output for Brazil. In contrast,

Amiti and Koenings (2005) find that the productivity gains arise from reducing input tariffs

81

especially for importing firms during a trade reform for Indonesia, which is consistent with

the findings of this paper. Halpern, Koren, and Szeidl (2006) use a panel of Hungarian

firms to examine two different mechanisms, a quality and a variety channel, through which

imports can affect firm productivity and find that importing inputs increase firm productiv-

ity by 14 percent, of which about two thirds is attributed to an increase in the variety of

intermediates used in production.

This chapter is organized as follows. The next section proceeds to describe the analyti-

cal framework used to study the relationship between productivity and imported intermedi-

ates. Section three outlines the empirical specification, while sections four and five explain

the estimation procedure and data set, respectively. The sixth section presents the results.

The last section concludes.

4.2 The Theoretical Framework

4.2.1 Production Function

For each period t, the ith plant’s production, Yit, is given by:

Yit = eωitKβkit Lsβs

it Luβu

it Eβe

it

[∫ N(dit)

0

x(j)θ−1

θ dj

] βxθθ−1

, (4.1)

where ωit represents a serially correlated productivity shock, Kit is capital input, Lsit is

skilled labour input, Luit is unskilled labour input, Eit is energy input, and intermediate

materials are horizontally differentiated. The elasticity of substitution between any two

material inputs is given by θ > 1. The variable N(dit) denotes the range of intermediate

inputs which are employed in the ith plant; it is a function of a plant’s discrete choice,

denoted by dit ∈ {0, 1}, to import from abroad or not: N(dit)(1−dit)Nh,t +ditNf,t, where

82

Nh,t is the range of intermediate inputs produced in this country and Nf,t is the range of

intermediate inputs available in the world. There are a range of intermediate inputs that are

not produced domestically in this country but are produced in foreign countries and thus

available through imports.

Horizontally differentiated materials in the production function is a common specifica-

tion used to analyze a change in the total factor productivity in the international trade and

the growth literatures (e.g., Ethier (1982), Romer (1990), and chapter 3 of Grossman and

Helpman (1991)). An alternative approach would include vertically differentiated inputs

with foreign inputs of higher quality (e.g., chapter 4 of Grossman and Helpman (1991)).

Given that our data set does not contain any information on firm-specific product price nor

the range of the variety of intermediate inputs a firm uses, it is difficult to empirically dif-

ferentiate between the quality or variety effect of foreign intermediates on productivity. For

tractability we employ the former model here but our estimates are likely to capture both

the variety and the quality effects.

Consider an equilibrium in which all intermediate goods are symmetrically produced at

level x. Substituting x(j)x into equation (4.1) leads to

Yit = eωitN(dit)βx

θ−1 Kβkit Lsβs

it Luβu

it Eβe

it Xβx

it , (4.2)

where XitN(dit)x.

Total factor productivity (TFP) is defined as Ait = Yit

Kβkit Lsβs

it Luβuit Eβe

it Xβxit

. Then, from

(4.2),

ln A(dit, ω) =βx

θ − 1ln(N(dit)) + ωit. (4.3)

This equation indicates that productivity is positively related to the range of employed

intermediate inputs. Plants importing intermediate inputs from abroad employ a larger

variety of intermediate inputs and hence exhibit higher productivity than those employing

83

domestic intermediate inputs only; for example, had there been no difference in the value

of ω across plants, then ln A(1, ω)− ln A(0, ω) = βx

θ−1ln(N(1)/N(0)) > 0.

4.2.2 Exit, Import, and Learning by Importing

The behavioural framework of Olley and Pakes (1996) is extended by incorporating im-

port decisions into their dynamic model. Consider a risk-neutral plant that maximizes the

expected present value of the sum of net cash flows. At the beginning of every period,

after observing the current productivity shock, ωt, the plant makes the following decisions.

First, it makes a discrete decision to exit, χt, by comparing a sell-off value of Φ with its

continuation value. If it continues in operation, it chooses its import status (dt), and then

variable factors (labour, materials, fuels) and investment level (ıt). Capital is accumulated

according to the law of motion Kt+1 = (1 − δ)Kt + ıt. It is assumed that this year’s

investment becomes productive the next year. Let kt ≡ ln(Kt).

Past import status may have an impact on the evolution of productivity. In particular,

importing materials may bring plants into close contact with foreign suppliers in developed

countries, which may lead to the positive dynamic externalities, or “learning by importing."

To examine the possibility of “learning by importing," we allow the distribution of ωt+1

conditional on information available at t to be dependent not only on past productivity, ωt,

but also on past import status, dt.2

Consider a fixed cost for importing materials, which may depend not only on the current

import choice but also on past import status because of a sunk start-up cost of importing

materials. We denote the fixed import cost—which we may think as the sum of the per-

period fixed cost and the sunk start-up cost—by Γ(dt−1, dt). Since the profit and the value

2Ericson and Pakes (1995) consider the model in which the distribution of ωt+1 depends on the amountof R&D investment.

84

functions depend on the time specific factors, such as factor prices, we index the profit and

the value functions by time. The Bellman equation for a plant can be written as

Vt(ωt, kt, dt−1) = max{Φt, max

dt,ıt

{πt(ωt, kt, dt)− c(ıt, kt)− Γ (dt−, dt)

+βE[Vt+1(ωt+1, kt+1, dt)|Jt]}}

,

where Φt is the sell-off value of the plant, πt(·) is the profit after maximizing out the variable

factors, c(ıt, kt) is the cost of investment, Γ(dt−1, dt) is the fixed cost of importing materials,

and Jt represents information available at time t.

The policy functions associated with the fixed point of the Bellman equation specify

an exit rule, a discrete import decision rule, and an investment decision rule. In particular,

when the profit function πt(·) is strictly increasing in ωt, the plant exit rule is characterized

by the threshold value ωt(kt, dt−1) as:

χt =

1, for ωt ≥ ωt(kt, dt−1),

0, otherwise(4.4)

The import decision rule and the investment demand equation are written, respectively, as:

dt = d∗t (ωt, kt, dt−1) (4.5)

ıt = ı∗t (ωt, kt, dt−1) (4.6)

Note that the decisions to exit, import, and invest crucially depend not only on the capital

stock but also past import status because past import status is one of the state variables.

Accordingly, we will modify the “standard" OP/LP estimation procedure by incorporating

past import status as an additional state variable.

85

4.3 Econometric Specification

Equation (4.2) suggests the following specification of the Cobb-Douglas production func-

tion augmented by the term representing the use of imported intermediates:

yit = βkkit + βslsit + βul

uit + βeeit + βxxit + βddit + ωit + ηit, (4.7)

where lowercase variables denote log values. A plant’s discrete choice to import from

abroad is denoted by dit while ηit is an i.i.d. shock that is not known to plants at the

time of input decisions. To examine the possibility of dynamic effects of import status on

productivity through “learning by importing," we consider the following stochastic process

of ωit:

ωit = ξt + γdi,t−1 + ρωi,t−1 + uit, (4.8)

where ξt is a year-specific productivity shock, uit is independent of di,t−1 and ωi,t−1 with

the cumulative distribution Fu(·).

We examine whether the use of imported intermediates leads to higher productivity by

testing whether βd > 0. A positive estimate of βd provides plant-level evidence for R&D

spillovers through trade in intermediate goods. On the other hand, a positive value of γ in

equation (4.8) is evidence for “learning by importing" and the long-run effect of “learning

by importing" is measured by γ1−ρ

.

The benefit from importing intermediates may differ across plants because the available

range of intermediate goods is different across plants.3 Assuming that all intermediate

goods are symmetrically produced at level x, the ratio of total intermediates to domestic

intermediates measures the ratio of the range of intermediate inputs available in the world3The benefit from importing intermediates may also be different when plants face heterogeneous trans-

portation costs of foreign intermediate goods. Kasahara and Lapham (2007) show that, in such a case,the higher the ratio of total intermediates to domestic intermediates, the larger the productivity effect fromimporting.

86

to the range of domestically produced intermediate inputs since Xit

Xhit

= Nit(1)xNit(0)x

= Nit(1)Nit(0)

,

where Xit is total intermediates and Xhit is domestic intermediates. From this viewpoint

the productivity effect of importing depends on how much imported intermediates are used

relative to domestic intermediates.

We examine whether the higher ratio of total intermediates to domestic intermediates

leads to higher productivity by considering the following alternative specification:

yit = βkkit + βslsit + βul

uit + βeeit + βxxit + βnnit + ωit + ηit, (4.9)

ωit = ξt + γni,t−1 + ρωi,t−1 + uit,

where nit = ln Xit

Xhit

. From the estimate of βn and βx, we may compute the elasticity of

substitution across different varieties of intermediate goods, denoted by θ, using βn = βx

θ−1.

4.4 Estimation

4.4.1 The OP/LP Proxy Estimator

One of the main econometric issues in estimating equations (4.7)-(4.9) is the simultaneity

of a productivity shock ωit and input decisions. If inputs are chosen on the basis of the pro-

ductivity shocks, a plant with a higher productivity shock may use more inputs. Since the

regressors are positively correlated with the error term, the coefficients estimated by ordi-

nary least squares (OLS) tend to be upwardly biased for variables that are more responsive

to a contemporary productivity shock than other variables.

Endogenous exit decisions may also induce bias in the estimated coefficients due to

sample selection. When a profit function is increasing in kt and dt−1, the threshold value

of productivity that induces exiting, ωt(kt, dt−1), is decreasing in kt and dt−1. Specifically,

87

plants with larger capital stocks and previous experience importing intermediates expect

larger future profits and hence stay in the market for lower realized values of ωt; the OLS

estimates may lead to biases in the coefficients of capital and imported materials.

To control for simultaneity and self-selection, we apply the framework developed by

Olley and Pakes (1996) and extended by Levinsohn and Petrin (2003).4 Our specification

for production technology differs from theirs in that we introduce import status, dit, as an

additional state variable and import status has a dynamic effect on productivity as specified

by (4.8).

Suppose that capital kit and the import decision dit are the state variables but lit, xit,

and eit are freely variable inputs. Then, the material’s demand function is given as xit =

x∗t (ωit, kit, dit), where the function x∗t (·) is time-dependent, reflecting its dependence on

time-specific common shocks in productivity and relative prices.

Assuming that x∗t (·) is strictly increasing in ωit, we can invert this function to obtain

the productivity shock ωit as a function of (xit, kit, dit):

ωit = ω∗t (xit, kit, dit). (4.10)

Replacing ω∗t (xit, kit, dit) for ωit in the equation (4.7) leads to a partial linear function:

yit = βslsit + βul

uit + βeeit + φt(xit, kit, dit) + ηit, (4.11)

where φt(xit, kit, dit) = βkkit +βxxit +βddit +ω∗t (xit, kit, dit). In the first stage, following

Levinsohn and Petrin (2003), we consistently estimate βs, βu, and βe from (4.11).

4Levinsohn and Petrin’s estimator is developed based on the investment proxy estimator of Olley andPakes (1996). In the Chilean data, there are a substantial number of zero investment observations (perhapsdue to the presence of fixed investment cost). For these observations, the investment proxy estimator of Olleyand Pakes cannot be used because they do not satisfy the monotonicity condition (and thus the investmentfunction is not invertible with respect to shocks). Given this feature of the Chilean data, we choose to usethe Levinsohn and Petrin intermediate proxy estimator rather than the Olley and Pakes investment proxyestimator while we address the selection bias using the method suggested by Olley and Pakes.

88

In the second stage, the rest of the parameters βk, βx, and βd are estimated as follows.

Define the innovations in productivity conditional on last year’s productivity, last year’s

import status, and survival:

νit = ωit − E[ωit|ωi,t−1, di,t−1, χit = 1], (4.12)

where χit = 1 if the ith plant continues in operation at t and χit = 0 if it exits. The

productivity innovations νit in (4.12) are orthogonal to all information available at time t−1

and, together with ηit in (4.11), can be used to construct the orthogonality conditions. In

fact, for each candidate parameter vector β∗ = (β∗x, β

∗k , β

∗d), we may construct an estimate

for the residual as:

( ˆνit + ηit)(β∗) = yit − βsl

sit − βul

uit − βeeit − β∗

kkit − β∗xxit − β∗

ddit

−E[ωit|ωi,t−1, di,t−1, χit = 1], (4.13)

where the estimate of E[ωit|ωi,t−1, di,t−1, χit = 1] is obtained using the procedure sug-

gested by Olley and Pakes (1996); in particular, we control for selection bias by considering

the expectation conditional on the survival χit = 1.5

The parameters β∗ = (βx, βk, βd) are estimated by minimizing the GMM criterion

function that is based on nine orthogonality conditions using the predetermined variables

(kit, ki,t−1, dit−1, di,t−2, xit−1, xi,t−2, lsi,t−1, l

ui,t−1, ei,t−1) as instruments for the residual νit +

ηit. The standard errors are obtained by the bootstrap.

5Appendix C discusses how exactly the selection issue is addressed in the context of the LP approach.While our approach controls for selection bias, Levinsohn and Petrin (2003) find that selection is unimportantwhen using an unbalanced panel.

89

4.5 Data

The data set is based on an annual census of Chilean manufacturing plants, which covers

all plants with more than 10 workers, by Chile’s Instituto Nacional de Estadistica (INE) for

1979-1996.6 Previous empirical studies using (a subset of) this data set includes Lui (1993),

Pavcnik (2002), and Levinsohn and Petrin (2003). The data set includes gross revenue, the

number of blue- and white-collar workers, various types of investment, imported materials,

total materials, electricity and fuels. Each variable is deflated by using the corresponding

annual price deflator to real 1980 Chilean pesos.

We exclude plants for which any of the data for capital stocks, unskilled labour, skilled

labour, energy, and domestic intermediates are either not available or reported as zero val-

ues. In particular, plants that do not report book values of their capital stocks in any year

are initially excluded since constructing capital stocks for these plants is impossible.7 After

cleaning the data, the unbalanced panel data set contains 3598 plants. A substantial num-

ber of plants are eliminated from the sample due to a missing initial capital stock. Because

this may lead to a sample selection problem, we also report the results under the extended

sample of 4508 plants in which the capital stock in 1980, if it is missing, is imputed by a

projected capital stock based on other reported plant observables.8 Hereafter, the sample

that excludes the plants missing book values of their capital stocks is called the “Basic6The unit of observation in our empirical analysis is “plant" rather than “firm." This is due to limitations

of our data set. Firm-level analysis might be particularly important to address the issue of “learning-by-importing"; the dynamic learning through importing might be more important at firm-level than at plant-level.

7The book values of capital are only reported, if any, in 1980 and 1981. Some plants did not report thebook values of capital in either 1980 or 1981. Since it is not possible to construct capital stock without thesereports, the plants missing their book values of capital were excluded from the sample. Notably, plants enterinto the market after 1982 are not included in the sample. We focus on the sample of plants that operated inboth 1979 and 1980 so that we may use the variables that are two period lagged in our regression analysis.

8Three types of capital (i.e., machinery, transportation, and buildings) are available in the data. We firstimpute a missing capital stock separately across types and then combine three types of imputed capital stockinto one. We use 4-digit industry dummies, location dummies, and business-type dummies (e.g., corporationvs. public) to do projection.

90

Sample" while the sample with imputed capital stocks is called the “Extended Sample."

Since the main features of the descriptive statistics of these two samples are similar, we

focus on the statistics of the Basic Sample in this section.

Output is total revenue adjusted for inventory change. Real output (Y ) is constructed

by deflating nominal output using an industry output price deflator. Real domestic mate-

rial (Xh) is constructed by subtracting the nominal value of imported materials from the

total materials and then deflated using an industry price deflator.9 Real imported materials

is constructed by deflating the nominal imported materials by the import price index (in

Chilean peso) reported in International Financial Statistics. The real value of total mate-

rials (X) is the sum of the real domestic materials and the real imported materials. The

number of blue- and white-collar workers are used for skilled and unskilled labour input

(Ls and Lu). The energy input (E) is the sum of the real purchased value of electricity and

that of fuels. The value of imported materials is reported separately from the total value of

materials.

The capital stock is constructed separately for buildings, machinery and equipment,

and vehicles from the 1980 book value of capital (the 1981 book value if the 1980 book

value is not available) using perpetual inventory method.10 The nominal net investment

variable is constructed, separately for buildings, machinery and equipment, and vehicles,

and then deflated using the construction deflator for buildings and the machinery deflator

for machinery and equipment, and vehicles to obtain the real net investment (ı).11 Buildings

9For both the output price deflator and the intermediate price deflator, we have used a 3-digit industrydeflator for 1979-1986, which is contained in the original data set as described in Lui (1993), and a 2-digitindustry deflator for 1987-1996 obtained from Yearbook of National Accounts by the Central Bank of Chile.As far as we know, the material price deflator at 3-digit levels are not available after 1987.

10Since the reported book values are evaluated at the end of year t, the book values of capital are deflatedby the (geometric) average deflator of machinery and equipment for years t and t+1. Depreciation rates areset to 5 % for building, 10 % for machinery and equipment, and 20% for vehicles.

11The data contains information on five types of investments: (i) purchases of new capital, (ii) purchases

91

Table 4.1: Descriptive Statistics in 1980Interme- Import Output/ No. of

Output Capital Labor Energy diates Ratios Workers PlantsAll 95.58 45.20 54.45 3.35 49.36 0.08 1.16 3598

Plants (437.46) (253.38) (105.09) (26.13) (221.81) (0.18) (1.63) —

Importing 445.64 194.32 177.17 11.20 201.87 0.37 2.56 273Plants (1021.01) (431.44) (256.26) (34.72) (407.35) (0.25) (3.73) —

Non-Importing 20.84 9.08 26.18 0.66 12.60 — 0.74 2017Plants (38.29) (51.95) (29.51) (5.52) (25.28) — (0.72) —

Switchers 137.77 69.78 72.44 5.86 74.23 0.13 1.50 1308(521.03) (355.70) (103.35) (39.38) (303.84) (0.22) (1.68) —

Survivors 170.97 76.44 77.74 6.28 84.93 0.11 1.50 1348(79.34) (376.54) (139.50) (40.20) (338.10) (0.21) (1.96) —

Quitters 50.41 26.48 40.50 1.60 28.05 0.05 0.95 2250(155.63) (129.72) (74.09) (10.77) (94.93) (0.16) (1.36) —

Notes: Standard errors are in parentheses. The statistics are based on the “Basic Sample" that excludes plants for which the initial capital

stock are not reported. “Importing Plants" are plants that continuously imported foreign intermediates in the sample. “Non-Importing

Plants" are plants that never imported foreign intermediates in the sample. “Switchers" are plants that switched their import status in the

sample. “Survivors" are plants that did not exit during the sample period (1980-1996) while “Quitters" exit during the sample period.

“Output," “Capital," “Energy," and “Intermediates" are measured in millions of 1980 pesos. “Labor" is the number of workers. “Import

Ratios" are the ratios of imported intermediate materials to total intermediate materials.

are likely to be rented rather than owned by plants, since zero values are found frequently

for buildings, especially for small plants. We add the capitalized rental value measured at

plant level to current year capital stock.12 The total capital stock (K) is the sum of the

real capital stock for building, machinery and equipment, and vehicles, and the capitalized

rental value. Note that the capital stock in year t does not include the investment in year t.

Table 4.1 reports descriptive statistics for variables in the year 1980. A comparison

between “Importing Plants," ‘Non-Importing Plants," and “Switchers" in Table 4.1 reveals

substantial differences between the three types of plants. Importing plants are substantially

of used capital, (iii) production of capital for own use, (iv) improvements in own capital by third parties, and(v) sales of capital. The net investment is the sum of (i)-(iv) minus (v).

12The data on rental rate is not available. To obtain a crude measure of rental rate, assuming the ag-gregate Cobb-Douglas production, we compute (rental rate)=(the share of capital)× GDP/(Capital Stock)-(depreciation rate)≈ 0.15 on average for 1980-1996 using the data on Chilean GDP and Capital Stockconstructed from the Chilean national accounting data with (the share of capital)=0.3 and (depreciationrate)=0.05. The capitalized rental value is computed as (rental value)/0.15 using rental value reported atplant level.

92

Table 4.2: Transition Probability of Import Status and ExitYear t status No Imports Imports

Year t + 1 status No Imports Imports Exit No Imports Imports Exit1981-1985 ave. 0.844 0.054 0.102 0.170 0.788 0.0421986-1990 ave. 0.885 0.055 0.061 0.173 0.805 0.0231991-1995 ave. 0.874 0.067 0.058 0.119 0.860 0.0211981-1995 ave. 0.868 0.059 0.074 0.154 0.818 0.028

Notes: The statistics are based on the “Basic Sample" that excludes plants for which the initial capital stock are not reported.

larger and have higher labour productivity while ‘Non-Importing Plants" are smaller and

least productive among those three types of plants. On the other hand, as shown in the

last two rows of Table 4.1, “Survivors" which do not exit before 1996 are larger, more

productive, and tend to import more in 1980 than “Quitters" that exit within the sample

period of 1980-1996.

Out of 3598 plants, 273 plants (7.6%) continuously import foreign intermediates through-

out the sample period (i.e., “Importing Plants" in Table 4.1), while 2017 plants (56.1%) are

“Non-Importing Plants" that never import intermediates from abroad. This suggests that

plant import status is persistent over time. There are, nevertheless, 1308 plants (36.4%)

that switch between importing and not importing over the period and, among them, 757

plants switch import status more than once. This within-plant variation of import status is,

thus, an important source of identification of the import variable coefficient.

Table 4.2 presents transition rates across import status together with exit rates. The last

row indicates the average transition rates for 1981-1995. The persistence in import status

is also clear here; among the plants that did not import in year t, more than 85 percent of

them did not import in year t + 1, while, among the plants that did import in year t, about

82 percent of them did import in year t + 1. Comparing plants across import status in year

t, we notice that importers are more likely to survive than non-importers.

93

4.6 Results

Tables 4.3 and 4.4 present the results from various estimators using the discrete choice

import variable; columns (1)-(5) of Table 4.3 and Table 4.4 report the results of the Basic

Sample while columns (6)-(8) of Table 4.3 and Table 4.4 report those of the Extended

Sample. To address the simultaneity issue, we also consider the Within-groups estimator

and the system GMM estimator (Blundell and Bond, 2000). The results from the OLS and

the Within-groups estimators for the Extended Sample are omitted because they are very

similar to those for the Basic Sample.

The most important finding is the significance and often large size of the current discrete

import variable coefficient across different estimators. The OLS point estimate implies that

a plant only using domestic intermediates can increase its productivity by 11.1 percent if

it starts importing intermediates. The OLS estimate is, however, likely to be biased due to

correlation between an unobserved plant productivity shock and inputs.

The within-estimator is robust against the simultaneity between a permanent plant-

specific shock and input decisions. Column (2) of Table 4.3 demonstrates that although

estimate of βd is smaller using the within-estimator relative to OLS, at 2.6 percent, it is

still positive and significant. The smaller estimate of βd may be due to the downward bias

induced by classification error in the discrete import variable dt.

While the within-estimator controls for correlation between inputs and a permanent

shock, it does not address the simultaneity between inputs and a persistent shock that varies

within-plant over time. To correct for such simultaneity in panel data, we further provide

the results from two alternative estimators: the system GMM estimator and the OP/LP

Proxy estimator. The system GMM estimates in columns (3) of Table 4.3 also indicate that

imports have a strong, positive effect on plant productivity. The model finds 18.0 percent

94

Table 4.3: Estimates of Production Function: Discrete Import Variable

The Data Set Basic Sample Extended SampleEstimators OLS Within GMM OP/LP Proxy GMM OP/LP Proxy

(1) (2) (3) (4) (5) (6) (7) (8)Skilled labour 0.139 0.062 0.034 0.137 0.038 0.127

(0.003) (0.004) (0.031) (0.006) (0.027) (0.008)Unskilled labour 0.143 0.175 0.251 0.145 0.243 0.142

(0.004) (0.006) (0.032) (0.008) (0.028) (0.008)Energy 0.052 0.062 0.092 0.043 0.002 0.057

(0.002) (0.003) (0.025) (0.005) (0.002) (0.006)Capital 0.091 0.054 0.108 0.058 0.064 0.139 0.065 0.076

(0.002) (0.003) (0.019) (0.009) (0.010) (0.016) (0.011) (0.012)Materials 0.647 0.581 0.612 0.549 0.509 0.655 0.575 0.525

(0.003) (0.005) (0.023) (0.025) (0.029) (0.020) (0.024) (0.037)Disc. Import (βd) 0.111 0.026 0.180 0.214 0.220 0.161 0.139 0.129

(0.005) (0.005) (0.049) (0.035) (0.067) (0.045) (0.032) (0.039)γ — — 0.009 0.041 — 0.008 0.024 —

(0.013) (0.011) (0.011) (0.009)ρ — — 0.245 0.892 — 0.271 0.900 —

(0.022) (0.027) (0.016) (0.116)Implied γ

1−ρ— — 0.012 0.379 — 0.011 0.235 —

P-value for over- 0.593 0.759 0.874 0.427identification test

No. of Obs. 33200 45518

Notes: Standard errors are in parentheses. Columns (1)-(5) use the “Basic Sample" that excludes plants for which the initial capital stockare not reported. Columns (6)-(8) use the “Extended Sample" in which a missing initial capital stock is imputed by a projected initialcapital stock based on other reported plant observables. The System GMM estimator in columns (3) and (6) use a lag length of 2 and3 for instruments in the first-differenced equations and a lag length of 1 in the level equations. The OP/LP estimators in columns (4)and (7) specify the stochastic process of ωt of the equation (4.8), while the OP/LP estimators in columns (5) and (8) uses the third orderpolynomials in (ωt−1, dt−1).

95

increases in productivity from a switch to imports for the Basic Sample.

Columns (4)-(5) of Table 4.3 provide the results of the OP/LP Proxy estimator which

controls for both selection and correlation between inputs and an unobserved productivity

shock by using intermediate inputs as proxies for unobserved productivity shocks.13 The

over-identification restrictions are not rejected for all four cases.14Column (4) reports the

OP/LP estimates under the AR(1) specification for the ωit process using the Basic Sample.

It indicates a large productivity effect (21.4 percent) arising from the usage of imported in-

termediate goods. To examine the robustness of the results with respect to the specification

of the productivity process, we report the OP/LP estimates under the flexible specification

for ωt using a third-order polynomial in (ωt−1, dt−1) with selection. The results are reported

in column (5) of Table 4.3. Once again, it indicates a large productivity effect (22.0 percent

for the Basic Sample) from the usage of imported intermediates. We also estimate the pro-

duction function using the technique proposed by Ackerberg, Caves and Frazer (2005) and

find that the results on the effect from using imports are robust with respect to the potential

identification issues they raise in respect to the OP/LP estimation technique.15

Comparing columns (3)-(5) with columns (6)-(8) in Table 4.3, we notice that control-

ling for sample selection due to missing capital stocks may be potentially important; the

estimates of import coefficient from the Extended Sample—which deals with the sample

selection due to missing capital stocks—range from 12.9 to 16.1 percent, which is smaller13Since both the investment and the import policy functions (4.5)-(4.6) may differ across years, due to the

macroeconomic cycles and the changes in trade policies, we allow for φ(·) to differ across the following sixperiods: 1979-1981, 1982-1983, 1984-1986, 1987-1989, 1990-1992, and 1993-1996.

14In addition to the over-identification test, we conducted two other specification tests suggested by Levin-sohn and Petrin (2003). First, to be consistent with the model, productivity shock should be monotonicallyincreasing in the materials, holding state variables (i.e., capital and import status) constant. By plotting theproductivity proxy ωit as a function of capital and intermediate inputs separately for importers and non-importers, we found that this is indeed the case. Second, we use the energy variable in place of the materialsas an input proxy and found that the estimated impact of imported materials on productivity is even largerunder the energy proxy.

15The result is available in Appendix C.

96

and perhaps more reasonable than those from the Basic Sample.

One might wonder why the estimates for the import coefficient from the GMM and

OP/LP estimators are substantially larger than the OLS estimates. In a multivariate context,

however, even if import variables are positively correlated with contemporary shocks, the

OLS estimates for import variables could be downwardly biased when import decisions are

less responsive to a shock than other inputs (cf., Levinsohn and Petrin, 2003).16 This is

potentially the case as import status is persistent over time in the data. It might be also

surprising that the capital coefficient from the OP/LP estimation are lower than the OLS

estimates given that a reason for the development of the OP/LP estimators was to remove a

suspected downward bias in OLS capital coefficients. However, relative to the “standard”

specification, our specification includes an additional persistent variable—import status—

which is positively correlated with capital and, consequently, it is difficult to assess ex ante

the direction of OLS bias on the capital coefficient in our case.17

Although we cannot be as certain about the magnitude of the effect given the wide range

of estimates across different estimators, even the within-estimator—which is likely to be

downwardly biased—suggests that the productivity gain from importing is 2.6 percent.

Overall, the results reported in Table 4.3 suggest a positive static effect from importing on

productivity since the estimates are quantitatively important even at the lower bound.18

16Other possible explanation for the downward OLS bias is the presence of measurement (or classification)errors in import variables. Using past import variables as instruments may correct for the bias due to mea-surement errors. This could be an explanation for the difference between the OLS and the GMM estimates.On the other hand, the implication of measurement errors for the OP/LP estimator is not well understood inthe literature.

17When we estimate the production function without the import variable using the Basic Sample, theOLS estimate of capital coefficient becomes 0.050 while the OP/LP estimate for capital becomes 0.086.Comparing them with the corresponding estimates in Table 4.3, we notice that the exclusion of import variablesubstantially decreases the OLS estimate of capital coefficient while it increases the OP/LP estimates.

18There are caveats regarding the validity of the GMM estimator as well as the OP/LP estimator, however.The GMM estimator may potentially suffer from weak instruments problem while the maintained assump-tions for the OP/LP estimator might be violated in reality. To the extent that the validity of the assumptions

97

−5 0 5 10 15 20 25

0

0.1

0.2

0.3

0.4

0.5

Year

ln(TFP)t

ln(TFP)t = β

d d

t + ω

t

ωt = γ d

t−1 + ρ ω

t−1

− − − 90 % bootstrap confidence interval

dt = 0 d

t = 1

Figure 4.1: Productivity Dynamics and Import Status

Another interesting finding is that, throughout columns (3)-(4) and (6)-(7) of Table

4.3, the estimated values of γ are positive and often significant. This suggests a positive

dynamic effect of the use of imported materials (i.e., “learning by importing") although the

evidence is not as strong as the case for the static effect given the small and insignificant

estimates of γ for the GMM estimator as reported in columns (3) and (6).

Figure 4.1 shows what would happen to total factor productivity, defined as βddit +ωit,

for a plant that is not importing at the steady state (i.e., di0 = 0) before Year 0 and, for some

exogenous reason, starts importing intermediates at Year 0.19 The figure is produced using

the estimates from the extended sample reported in column (7) of Table 4.3. The solid

line indicates the dynamic path of productivity implied by the point estimates reported in

underlying these estimators is uncertain, we should take the results from the GMM and the OP/LP estimatorswith caution.

19Idiosyncratic shocks, uit, are set to zero for all periods. Or, alternatively, we may interpret the solid lineas the path of “average" productivity among plants that start importing at Year 0 and keep importing afterYear 0 for some exogenous reason.

98

column (4) while the dashed lines represent a 90 percentile bootstrap confidence interval.20

At Year 0, a plant starts importing foreign intermediates, leading to an immediate increase

in productivity by 13.9 percent (static effect). After Year 0, a plant gradually achieves addi-

tional 23.5 percent productivity increase (dynamic effect). Note, however, that a 90 percent

bootstrap confidence interval suggests that there exists substantial uncertainty regarding the

precise magnitude of the dynamic effect of importing materials.

Table 4.4 presents estimates using the continuous measure of import usage, measured

by the ratio of total intermediates to domestic intermediates. In the table, the estimated

coefficients for skilled labour, unskilled labour, energy, capital, and materials are all signif-

icant and similar to those reported in corresponding columns of Table 4.3; that is, the use

of the continuous import variable in place of the discrete import variable leads to similar

estimates for these variables.

Across various estimators, the coefficients for the continuous import variable are often

significant and large across different estimators, indicating the importance of foreign inter-

mediates in explaining productivity differences across plants and over time. The system

GMM estimates in columns (3) and (6) imply that a 100 percent decrease in the share of

domestic intermediates in total intermediates could increase productivity between 5.8 and

7.2 percent although the estimate from the Basic Sample in column (3) is not significant.

The OP/LP estimates reported in columns (4)-(5) and (7)-(8) again support a substantial

impact of an increase in the share of imported intermediates on productivity, finding that a

100 percent decrease in the share of domestic intermediates increases productivity by 17.7

to 27.0 percent.

The positive estimates of γ throughout all columns in Table 4.4 are suggestive of the20To construct a 90 percentile bootstrap confidence interval, we repeatedly compute the dynamic path of

productivity under different bootstrap estimates for (βd, γ, ρ) and take a 5 and a 95 percentile of βddit + ωit

for each year.

99

Table 4.4: Estimates of Production Function: Continuous Import Variable

The Data Set Basic Sample Extended SampleEstimators OLS Within GMM OP/LP Proxy GMM OP/LP Proxy

(1) (2) (3) (4) (5) (6) (7) (8)Skilled labour 0.147 0.062 0.056 0.138 0.046 0.128

(0.003) (0.004) (0.031) (0.006) (0.026) (0.008)Unskilled labour 0.144 0.174 0.278 0.148 0.253 0.145

(0.004) (0.006) (0.033) (0.008) (0.028) (0.009)Energy 0.051 0.062 0.074 0.044 0.002 0.056

(0.002) (0.003) (0.025) (0.005) (0.002) (0.006)Capital 0.094 0.054 0.121 0.066 0.074 0.145 0.074 0.089

(0.002) (0.003) (0.019) (0.009) (0.010) (0.016) (0.009) (0.016)Materials 0.651 0.582 0.603 0.616 0.577 0.653 0.608 0.548

(0.003) (0.005) (0.024) (0.021) (0.027) (0.020) (0.023) (0.026)Cont. Import (βn) 0.096 0.034 0.058 0.246 0.270 0.072 0.177 0.182

(0.006) (0.007) (0.042) (0.052) (0.061) (0.032) (0.043) (0.062)γ — — 0.004 0.030 — 0.001 0.026 —

(0.016) (0.010) (0.014) (0.008)ρ — — 0.241 0.822 — 0.271 0.871 —

(0.023) (0.031) (0.016) (0.027)Implied γ

1−ρ— — 0.005 0.169 — 0.001 0.199 —

Implied θ 7.78 18.12 11.40 3.50 3.14 10.07 4.44 4.01P-value for over- 0.824 0.759 0.834 0.995identification test

No. of Obs. 33200 45518

Notes: Standard errors are in parentheses. Columns (1)-(5) use the “Basic Sample" that excludes plants for which the initial capital stockare not reported. Columns (6)-(8) use the “Extended Sample" in which a missing initial capital stock is imputed by a projected initialcapital stock based on other reported plant observables. The System GMM estimator in columns (3) and (6) use a lag length of 2 and 3for instruments in the first-differenced equations and a lag length of 1 in the level equations. The OP/LP estimators in column (5) and(8) specify the stochastic process of ωt using the third order polynomials in (ωt−1, nt−1).

100

−5 0 5 10 15 20 25

0

0.1

0.2

0.3

0.4

0.5

Year

10 percentile25 percentile50 percentile75 percentile90 percentile

ln(TFP)t

dt=0 d

t=1

Figure 4.2: Productivity Dynamics Heterogeneity across Different Import Shares

positive dynamic effect of an increase in the usage of imported intermediates although, as

in the case of the discrete import variable, the estimates from GMM estimator are insignifi-

cant. The OP/LP estimates of γ1−ρ

in columns (4) and (7) indicate that the long-run dynamic

effect of a 100 percent decrease in the share of domestic intermediates are 16.9 and 19.9

percent, respectively. From the estimated coefficients of materials and continuous import

variable, we can also compute an estimate of the elasticity of substitution as θ = 1 + βx

βn.

Using the OP/LP estimates in Table 4.4, we obtain point estimates of θ of 3.14 to 4.44

which are in line with those found by Feenstra, Markusen, and Zeile (1992).

Even among importing plants, there exists substantial heterogeneity in the use of im-

ported intermediates; the share of imported intermediates in total intermediates for the

bottom 10 percentile of importing plants is only 2.8 percent while the share of imported

intermediates for the 90 percentile importing plants is as high as 67.9 percent. In the view

of the specification for continuous import variable (4.9), when plants are heterogeneous

101

in the use of imported intermediates, they will experience different productivity dynamics

when they start importing intermediates.

Figure 4.2 plots the dynamic paths of total factor productivity before and after plants

start importing intermediates at Year 0 for five hypothetical plants with different import

shares.21 Here, we assume that the import shares when plants import are plant-specific and

permanently fixed. The thin line represents the productivity dynamics for the bottom 10

percentile importing plants with 2.8 percent import shares while the thick line corresponds

to the dynamics for the 90 percentile importing plants with 67.9 percent import shares;

other three lines represent the productivity dynamics for 25, 50, and 75 percentile import-

ing plants.22 Figure 4.2 highlights that the productivity effect from importing intermediates

may substantially differ across plants because of the difference in import shares. The bot-

tom 10 percentile importing plants hardly benefit from importing intermediates while the

90 percentile importing plants experience an immediate increase in productivity by 20.1

percent at Year 0 (static effect) and gradually achieve additional 22.6 percent productivity

increase after Year 0 (dynamic effect).

While the continuous import variable estimates provide important additional evidence

for the impact of imported intermediates on plant-level productivity, we also check the

sensitivity of the results with respect to the following potentially important controls: ex-

port behaviour and industry-year dummies.23 An omission of the export variable from the

regressors might lead to the coefficient of imported intermediates to be biased upwards

since “good” firms often both export and import (cf., Kasahara and Lapham, 2007). On

the other hand, the trade environment may change differently across industries over time;21The estimates in column (7) of Table 4.4 are used to produce the figure.22The shares of imported intermediates in total intermediates for 25, 50, and 75 percentile importing plants

are 8.1, 22.2, and 44.8 percent, respectively.23Although this procedure is common in the international trade literature, unlike the results in Tables 4.3

and 4.4, it is not robust to the presence of sunk imports costs. See Biesebroeck (2003) for an example.

102

Table 4.5: OLS Regression of TFP on Import and Export: Discrete Variables

Basic Sample Extended Sample(1) (2) (3) (4) (5) (6)

Discrete Import 0.153 0.128 0.007 0.098 0.076 0.004(0.004) (0.004) (0.003) (0.003) (0.003) (0.004)

Discrete Export 0.135 0.111 0.014 0.091 0.072 0.021(0.004) (0.005) (0.003) (0.003) (0.003) (0.006)

Industry-Year Dummies No Yes Yes No Yes YesPlant Fixed Effects No No Yes No No Yes

No. of Obs. 11027 11027 1694 12014 12014 3241

Notes: Standard errors are in parentheses. The estimates are based on the observations for the 1990-1996 period since those are the onlyyears for which export behaviour is observed. “Industry-Year Dummies” includes a full set of interactions between 4-digit ISIC industrydummies and year dummies. Total factor productivity is calculated as E[ωit|ωi,t−1, χit = 1].

Table 4.6: OLS Regression of TFP on Import and Export: Continuous Variables

Basic Sample Extended Sample(1) (2) (3) (4) (5) (6)

Continuous Import 0.139 0.112 0.012 0.081 0.062 0.005(0.008) (0.007) (0.003) (0.005) (0.004) (0.003)

Continuous Export 0.008 0.007 0.001 0.006 0.005 0.002(0.001) (0.001) (0.0003) (0.0003) (0.0003) (0.0004)

Industry-Year Dummies No Yes Yes No Yes YesPlant Fixed Effects No No Yes No No Yes

No. of Obs. 11027 11027 1890 12014 12014 3591

Notes: Standard errors are in parentheses. The estimates are based on the observations for the 1990-1996 period since those are the onlyyears for which export behaviour is observed. “Industry-Year Dummies” includes a full set of interactions between 4-digit ISIC industrydummies and year dummies. Total factor productivity is calculated as E[ωit|ωi,t−1, χit = 1].

for instance, an increase in tariffs in the mid-1980’s may have differential impacts across

industries. Industry-year dummies may capture the time-varying industry-specific trade

environment.24

To examine the robustness of these findings, we first estimate the production function

(4.7) by the OP/LP procedure but without including import variable as a regressor and

obtain the estimates of total factor productivity. Then we regress the estimated total factor

24Other potentially important controls include foreign investment goods and foreign ownership. Due todata limitations, we cannot examine the robustness against the inclusion of these controls and our estimatesof the productivity effect from importing intermediates potentially capture the effect of foreign investmentgoods or foreign ownership.

103

productivity on an import variable, an export variable, as well as a full set of interactions

between 4-digit ISIC industry dummies and year dummies. We also control for plant fixed

effects. For export status, we consider both a discrete export variable that takes the value of

one for an exporter (and zero for a non-exporter) and a continuous export variable measured

by the ratio of export sales to total revenues. Since export behaviour is observed only for

the 1990-1996 period in the data, the estimates are obtained based on the 1990-1996 period

sample.

Table 5 presents the results for the discrete import/export variables. As reported in

columns (1)-(2) and (4)-(5), the import effect is still positive and significant, ranging from

7.6 to 15.3 percent, even after controlling for export status and industry-year dummies.

The estimates controlling for plant fixed effects in columns (3) and (6) are lower, indicating

0.4-0.7 percent positive productivity effect from importing; the estimates for the plant fixed

effects may be downwardly biased because of classification errors and it is not significant

in column (6) possibly because there is not enough within-plant variation in the import

variables given the short nature of the panel data as well as the substantial persistence in

import status.25 On the other hand, the export effect is positive and significant in all cases,

indicating a possible productivity effect from becoming an exporter.

Table 4.6 reports the results for the continuous import/export variables. The results

imply that a 100 percent decreases in the share of domestic intermediates increases pro-

ductivity by 0.5 to 13.9 percent. Thus, not only whether plants import but also how much

they import is important in determining plant productivity. In contrast, the results for the

continuous export variable suggest a small productivity effect from increasing the share of

25More than 80 percent of plants are dropped from the fixed effects regression because these plants did notchange their export/import status throughout the sample period of 1990-1996.

104

exports in total output across all estimates. One possible interpretation is that the produc-

tivity effect of exporting may work mainly through the extensive margin of whether plants

export or not rather than the intensive margin of how much they export.

We also examined how the results change across industries by estimating the production

functions for two of the largest 3-digit industries, Food and Metals, and found that focusing

on one particular industry does not alter our basic finding.26

4.7 Conclusion

The results in this chapter demonstrate significant plant-level evidence that imported in-

termediates improve a plant’s productivity. We find that by switching from being a non-

importer to an importer of foreign intermediates a plant can immediately improve produc-

tivity; although the point estimates substantially differ across different estimators, even the

estimate from the Within-Group estimator, which is probably downwardly biased, indicates

a 2.6 percent positive productivity effect from importing. We also find some evidence of a

positive dynamic effect from the use of imported materials.

These results have important implications for both government policy and plant pro-

duction strategy. Often governments and policymakers focus on the potential benefits of

export-led growth when discussing trade policy. Our results suggest that manufacturing

plants may be able to substantially benefit from foreign imports and should be an important

part of trade policy or negotiations.

26The industry-level results are reported in Appendix C.

105

Chapter 5

Summary and Conclusions

The large increase in the volume of international trade and FDI over the past 20 years

has increased the attention paid to the impact of international integration on productivity

and welfare across plants and industries. This thesis provides further evidence that foreign

direct investment and international trade can substantially impact both plant-level and ag-

gregate productivity. The first two chapters present and estimate models of foreign direct

investment and exports with heterogeneous firms, while the third examines the impact of

plant-level imports on plant-level productivity

In the first chapter, I show that the model can generate productivity differences across

plants with different ownership and export status which are consistent with the observed

differences in the Indonesian manufacturing data. Using the theoretical model and a panel

of Indonesian manufacturing plants, I develop and estimate a structural empirical model of

exports and foreign direct investment. The counterfactual results emphasize that account-

ing for FDI flows is essential to recovering accurate estimates of the impact of trade on

aggregate productivity. In particular, the counterfactual experiments imply that the impact

of trade on productivity is greatly mitigated by FDI flows.

106

The model without sunk costs implies FDI restrictions can reduce aggregate productiv-

ity by 3 to 11 times more than trade restrictions. I find that the impact of FDI restrictions

account for a fall in average total factor productivity between 8 and 27 percent across in-

dustries. Trade restrictions, in contrast, are estimated to have a smaller impact on average

productivity. Across the food, manufactured metals and textiles industries average total

factor productivity is estimated to fall by 1 to 4 percent.

In the second chapter, I extend the model presented in the preceding chapter to allow for

sunk FDI and export costs. While the impact of FDI restrictions on aggregate productivity

are even greater in the model with sunk costs, the qualitative implications are unchanged.

The model with sunk costs finds that FDI restrictions cause aggregate productivity to fall by

18 to 41 percent across industries, while trade restrictions cause aggregate productivity to

fall by less than 6 percent across industries. The results suggest that policies which induce

inwards flows of FDI will have a much larger impact on aggregate productivity relative to

those that encourage exports. Moreover, the results suggest that changes in international

policies can have very different impacts across foreign and domestic producers. In partic-

ular, policymakers concerned with aggregate productivity should be particularly sensitive

the role policy has on the exit decision of highly productive, foreign-owned plants.

The third chapter examines the relationship between imported intermediates and plant-

level productivity. In that chapter we estimate the impact of using imported intermediates

on plant-productivity while controlling for potential biases from simultaneity, plant-level

selection and attrition. Across all of the estimates, we find that plants that begin importing

foreign intermediates can immediately improve productivity, although the point estimates

substantially differ across different estimators. However, even at the lower bound we find

that plants receive an increase in productivity by 2.6 percent when they begin to import

107

intermediates from abroad. We also find some evidence of a positive dynamic effect from

the use of imported materials.

There are a number of interesting questions that extend from this thesis. The manufac-

turing plants in Indonesia demonstrate substantial heterogeneity across export intensity and

import behaviour. Extending the model to allow richer export and import patterns across

heterogeneous plants may uncover other dimensions of interaction across foreign direct

investment, export decisions and import decisions.

Moreover, heterogeneous plants can have widely different responses to changes in tax

and/or tariff environment. This environment could be extended to examine the responses of

firms with different ownership and trade history to changes in tax or tariff policy. I intend

to address these issues in my future research.

108

Bibliography

Ackerberg, D. A., Benkard, C. L., Berry, S., and Pakes, A. (2006). Econometric Tools for

Analyzing Market Outcomes. mimeo.

Ackerberg, D. A., Caves, K., and Frazer, G. (2005). Structural Identification of Production

Functions. mimeo.

Amiti, M. and Konings, J. (2005). Trade Liberalization, Intermediate Inputs, and Produc-

tivity: Evidence from Indonesia. IMF Working Paper.

Antras, P. (2003). Firms, Contracts, and Trade Structure. Quarterly Journal of Economics,

118(4):1375–1418.

Antras, P. and Helpman, E. (2004). Global Sourcing. Journal of Political Economy,

112(3):552–580.

Arellano, M. and Bond, S. R. (1991). Some Test of Specification for Panel Data: Monte

Carlo Evidence and an Application to Employment Equations. Review of Economic

Studies, 58:277–297.

Arnold, J. and Javorcik, B. S. (2005). Gifted Kids or Pushy Parents? Foreign Acquisi-

tions and Plant Performance in Indonesia. CEPR Discussion Papers, Working Paper

5065:1–36.

109

Aw, B., Chung, S., and Roberts, M. (2000). Productivity and Turnover in the Export

Market: Micro-level Evidence from the Republic of Korea and Taiwan (China). World

Bank Economic Review, 14:65–90.

Baily, M., Hulten, C., and Campbell, D. (1992). Productivity Dynamics in Manufacturing

Plants. Brookings Papers on Economic Activity: Microeconomics, pages 67–119.

Bank, W. (1993). The East Asian Miracle: Economic Growth and Public Policy. Oxford

University Press, New York.

Bank, W. (2006). Doing Business: Comparing Business Regulations.

http://www.doingbusiness.org/.

Ben-Akiva, M. and Lerman, M. (1985). Discrete Choice Analysis: Theory and Application

to Travel Demand. MIT Press, Cambridge, MA.

Bernard, A. B., Eaton, J., Jensen, J. B., and Kortum, S. (2003). Plants and Productivity in

International Trade. American Economic Review, 93(4):1268–1290.

Bernard, A. B. and Jensen, J. B. (1999). Exceptional Exporter Performance: Cause, Effect,

and or Both? Journal of International Economics, 47:1–25.

Bernard, A. B. and Jensen, J. B. (2004). Why Some Firms Export. The Review of Eco-

nomics and Statistics, 86(2):561–569.

Biesebroeck, J. V. (2003). Revisiting some Productivity Debates. NBER Working Papers,

Working Paper 10065.

Blalock, G. and Gertler, P. (2005). Learning from Exporting Revisited in a Less Developed

Setting. Journal of Development Economics, 75(2):397–416.

110

Blonigen, B. A. and Wilson, W. W. (2006). New Measures of Port Efficiency Using Inter-

national Trade Data. NBER Working Papers, Working Paper 12052.

Blundell, R. W. and Bond, S. R. (1998). Initial Conditions and Moment Restrictions in

Dynamic Panel Data Models. Journal of Econometrics, 68:29–51.

Blundell, R. W. and Bond, S. R. (2000). GMM Estimation with Persistent Panel Data: an

Application to Production Functions. Econometric Reviews, 19(3):321–340.

Bond, S. R. and Soderbom, M. (2005). Adjustment Costs and the Identification of Cobb

Douglas Production Functions. mimeo.

Brainard, S. L. (1997). An Empirical Assessment of the Proximity-Concentration Trade-off

Between Multinational Sales and Trade. American Economic Review, 87(4):520–544.

Chamberlain, G. (1982). Multivariate Regression Models for Panel Data. Journal of Econo-

metrics, 18:5–46.

Clerides, S., Lach, S., and Tybout, J. R. (1998). Is Learning by Exporting Important?

Micro-Dynamic Evidence from Columbia, Mexico and Morocco. Quarterly Journal

of Economics, 113:903–947.

Coe, D. T. and Helpman, E. (1995). International R&D Spillovers. European Economic

Review, 39:859–887.

Coe, D. T., Helpman, E., and Hoffmaister, A. W. (1997). North-South R&D Spillovers.

Economic Journal, 107:134–149.

Das, S., Roberts, M. J., and Tybout, J. R. (2007). Market Entry Costs, Producer Hetero-

geneity, and Export Dynamics. Econometrica, 75(3):837–873.

111

Eaton, J., Kortum, S., and Kramarz, F. (2004a). An Anatomy of International Trade: Evi-

dence from French Firms. NBER Working Papers, Working Paper 10344.

Eaton, J., Kortum, S., and Kramarz, F. (2004b). Dissecting Trade: Firms, Industries,

and Export Destinations. American Economic Review, Papers and Proceedings,

94(2):150–154.

Editorial (2002). Shunned by Investors. The Jakarta Post, May 22,.

Ekholm, K., Forslid, R., and Markusen, J. (2003). Export-Platform Foreign Direct Invest-

ment. NBER Working Papers, Working Paper 9517.

Ethier, W. J. (1982). National and International Returns to Scale in the Modern Theory of

International Trade. American Economic Review, 72(3):389–405.

Feenstra, R. C. (1994). New Product Varieties and the Measurement of International Prices.

American Economic Review, 84(1):157–177.

Feenstra, R. C., Markusen, J. R., and Zeile, W. (1992). Accounting for Growth with New

Inputs: Theory and Evidence. American Economic Review, 82(2):415–421.

Feinberg, S. and Keane, M. (2006). Accounting for the Growth of MNC-Based Trade using

a Structural Model of U.S. MNCs. American Economic Review, 96(5):1515–1558.

Griliches, Z. and Hausman, J. (1986). Errors in Variables in Panel Data. Journal of Econo-

metrics, 31:93–118.

Griliches, Z. and Mairesse, J. (1998). Production Functions: The Search for Identification.

In Stom, S., editor, Econometrics and Economic Theory in the 20th Century: The

112

Ragnar Frish Centennial Symposium, pages 169–203. Cambridge University Press,

England.

Grossman, G. and Helpman, E. (1991). Innovation and Growth in the Global Economy.

MIT Press, Cambridge, MA.

Grossman, G., Helpman, E., and Szeidl, A. (2006). Optimal Integration Strategies for the

Multinational Firm. Journal of International Economics, 70:216–238.

Halpern, L., Koren, M., and Szeidl, A. (2006). Imports and Productivity. mimeo, Federal

Reserve Bank of New York.

Helpman, E., Melitz, M. J., and Yeaple, S. (2004). Export versus FDI with Heterogeneous

Firms. American Economic Review, 94:300–316.

Horowitz, J. L. (2001). The Bootstrap. In Heckman, J. J. and Leamer, E., editors, Handbook

of Econometrics, volume 5, pages 3159–3228. Elsevier Science, Oxford.

Hummels, D., Ishii, J., and Yun, K. (2001). The Nature and Growth of Vertical Specializa-

tion in World Trade. Journal of International Economics, 54:75–96.

International Labour Organization Bureau of Statistics (2007). International Labour Office

database on labour statistics. http://ilo.laborsta.org/.

Kasahara, H. and Lapham, B. (2007). Productivity and the Decision to Import and Export:

Theory and Evidence. mimeo, Queen’s University.

Keller, W. (2002). Trade and the Transmission of Technology. Journal of Economic

Growth, 1:5–24.

113

Levinsohn, J. and Petin, A. (2003). Estimating Production Functions Using Inputs to Con-

trol for Unobservables. Review of Economic Studies, 70:317–341.

Lui, L. (1993). Entry, Exit, and Learning in the Chilean Manufacturing Sector. Journal of

Development Economics, 42:217–242.

Marschak, J. and Andrews, W. H. (1944). Random Simultaneous Equations and the Theory

of Production. Econometrica, 12(3,4):143–205.

McFadden, D. Modeling the Choice of Residential Location. In et al., A. K., editor, Spatial

Interaction Theory and Residential Location. North-Holland, Amsterdam.

Melitz, M. J. (2003). The Impact of Trade on Intra-Industry Reallocations and Aggregate

Industry Productivity. Econometrica, 71:1695–1725.

Muendler, M.-A. (2004). Trade, Technology, and Productivity: A Study of Brazilian Man-

ufacturers, 1986-1998. mimeo, UCSD.

Mulkay, B., Hall, B. H., and Mairesse, J. (2000). Firm Level Investment and R&D in

France and the United States: A Comparison. NBER Working Papers, Working Paper

8038.

Nations, U. (1994). Commodity Trade Statistics 1994. United Nations Publishing Division,

New York.

OECD (2006). International Direct Investment Statistics Online Database.

http://www.sourceoecd.org/.

Olley, S. and Pakes, A. (1996). The Dynamics of Productivity in the Telecommunications

Equipment Industry. Econometrica, 65(1):292–332.

114

Pavcnik, N. (2002). Trade Liberalization, Exit, and Productivity Improvements: Evidence

From Chilean Plants. Review of Economic Studies, 69(1):245–276.

Ramondo, N. (2004). The Effects of Foreign Plants on Industry Productivity: Evidence

from Chile. mimeo, The University of Texas at Austin.

Ramondo, N. (2006). Size, Geography, and Multinational Production. mimeo, The Uni-

versity of Texas at Austin.

Ramstetter, E. and Sjoholm, F. The Issues: Multinationals and the questions they raise. In

Ramstetter, E. and Sjoholm, F., editors, Multinational Corporations in Indonesia and

Thailand. Palgrave MacMillan, New York.

Rob, R. and Vettas, N. (2003). Foreign Direct Investment and Exports with Growing De-

mand. Review of Economic Studies, 70(3):629–648.

Roberts, M. J. and Tybout, J. R. (1997). Decision to Export in Columbia: An Empirical

Model of Entry with Sunk Costs. American Economic Review, 87:545–564.

Romer, P. (1990). Endogenous Technological Change. Journal of Political Economy,

98(5):S71–S102.

Ruhl, K. and Willis, J. (2007). Convexities, Nonconvexities, and Firm Export Behaviour.

mimeo.

Rust, J. Estimation of Dynamic Structural Models, Problems and Prospects: Discrete

Decision Processes. In Sims, C. A., editor, Advances in Econometrics Sixth World

Congress, volume II. Cambridge University Press, New York.

115

Rust, J. (1987). Optimal replacement of GMC bus engines: an empirical model of Harold

Zurcher. Econometrica, 55(5):999–1033.

Trefler, D. (2004). The Long and Short of the Canada-U.S. Free Trade Agreement. Ameri-

can Economic Review, 94(4):870–895.

Wie, T. K. Intra-Regional Investment and Technology Transfer in Indonesia. In Yanagi,

K., editor, Symposium on Intra-Regional Investment and Technology Transfer. Asian

Productivity Organization, Kuala Lumpur.

Witular, R. A. (2006). Government blamed for messy port management. The Jakarta Post,

February 5,.

Yeaple, S. R. (2003). The Complex Integration Strategies of Multinationals and Cross

Country Dependencies in the Structure of FDI. Journal of International Economics,

60:293–314.

Yeats, A. J. Just How Big is Global Production Sharing? In Arndt, S. W. and Kierzkowski,

H., editors, Fragmentation: New Production Patterns in the World Economy. Oxford

University Press, Oxford.

Yi, K. (2003). Can Vertical Specialization Explain the Growth of World Trade? Journal

of Political Economy, 111(1):52–102.

116

Appendix A

Appendix for Chapter 2

A.1 Transition Probabilities

In tables A.1 and A.2 I report the transition probabilities for the food and manufactured

metals industries. Similar to the textiles industry, the model captures a substantial portion

of the persistence in non-exporter status for both foreign and domestic firms, but cannot

match the persistence in exporter status. As noted in the text, this may be indicative of the

presence of sunk export costs.

Table A.1: Transition Probabilities - FoodActual Dom. Non-Exporters Dom. Exporters For. Non-Exporters For. Exporters ExitDom. Non-Exporters at t 0.979 0.018 0.003 0.0005 0.094Dom. Exporters at t 0.292 0.694 0.003 0.010 0.071For. Non-Exporters at t 0.087 0 0.734 0.179 0.024For. Exporters at t 0.022 0.034 0.225 0.719 0.033PredictedDom. Non-Exporters at t 0.974 0.026 — — 0.105Dom. Exporters at t 0.873 0.127 — — 0.069For. Non-Exporters at t — — 0.735 0.265 0.028For. Exporters at t — — 0.732 0.268 0.029

117

Table A.2: Transition Probabilities - MetalsActual Dom. Non-Exporters Dom. Exporters For. Non-Exporters For. Exporters ExitDom. Non-Exporters at t 0.956 0.033 0.007 0.005 0.062Dom. Exporters at t 0.272 0.699 0.014 0.014 0.037For. Non-Exporters at t 0.071 0.009 0.750 0.170 0.040For. Exporters at t 0.022 0.017 0.191 0.774 0.041PredictedDom. Non-Exporters at t 0.968 0.032 — — 0.134Dom. Exporters at t 0.796 0.204 — — 0.069For. Non-Exporters at t — — 0.552 0.448 0.046For. Exporters at t — — 0.551 0.449 0.046

Note: Metals refers to manufactured metals.

A.2 Counterfactual Experiments

A.2.1 Counterfactual Results for the Food and Metals Industries

The results in the food and manufactured metals industries are similar to those in the textiles

industry. Table A.3 demonstrates that average productivity in the food industry falls by

16, 4 and 12 percent in the autarky, no trade and no FDI experiments. Also, the welfare

impact is several times larger under autarky than the other experiments due to the lack of

substitution of FDI for trade (or vice-versa).

Table A.4 demonstrates that average productivity in the manufactured metals industry

falls by 30, 3 and 28 percent in the autarky, no trade and no FDI experiments. Again, the

welfare impact is several times larger under autarky than the other experiments due to the

lack of substitution of investment for trade (or vice-versa).

118

Table A.3: Counterfactual Experiments - FoodBase Autarky No Trade No FDI

Avg. Productivitya 2.804 2.357 2.697 2.473−(ε − 1)δ∆ ln P — −δ0.041 −δ0.00004 0

Exit/Entry Rate of Foreign — -1 0.210 -1Firms in Indonesia% of For. Non-Exporters 0.017 0 0.024 0% of Dom. Exporters 0.028 0 0 0.028%∆ in Exports — -1 -1 -0.395

Mkt. Shr. of Dom. Non-Exp. 0.703 1 0.773 0.894Mkt. Shr. of Dom. Exp. 0.084 0 0 0.106Mkt. Shr. of For. Non-Exp. 0.159 0 0.227 0Mkt. Shr. of For. Exp. 0.055 0 0 0

Notes: a) Average productivity of all plants located in Indonesia in the steady state and is calculated using plant-level revenue shares asweights.

A.2.2 Price Indices

I denote the Indonesian aggregate price under the parameter vector θ by P (θ) and the for-

eign aggregate price by P ∗(θ). Suppose that I am interested in a counterfactual Indonesian

policy experiment with the parameter vector θ which is different from the estimated pa-

rameter vector θ. First, recall that the reduced form parameter ϕB is a function of the

Indonesian price level and the foreign price level:

ϕB =E

E∗

(P (θ)

P ∗(θ)

)ε−1

where the aggregate price levels are written as a function of the estimated parameter vector

θ. Writing the revenue functions in this fashion is equivalent to normalizing E∗P ∗(θ) = 1

in the estimation routine. Denoting φ∗B as

ϕ∗B =

E∗

E∗

(P ∗(θ)

P ∗(θ)

)ε−1

= 1

119

Table A.4: Counterfactual Experiments - Metalsa

Base Autarky No Trade No FDIAvg. Productivityb 3.920 2.728 3.785 2.854−(ε − 1)δ∆ ln P — −δ0.022 −δ0.00002 0

Exit/Entry Rate of Foreign — -1 0.039 -1Firms in Indonesia% of For. Non-Exporters 0.058 0 0.109 0% of Dom. Exporters 0.032 0 0 0.036%∆ in Exports — -1 -1 -0.753

Dom. Mkt. Shr. of Dom. Non-Exp. 0.487 1 0.541 0.881Dom. Mkt. Shr. of Dom. Exp. 0.066 0 0 0.119Dom. Mkt. Shr. of For. Non-Exp. 0.248 0 0.459 0Dom. Mkt. Shr. of For. Exp. 0.200 0 0 0

Notes: a) Metals refers to manufactured metals. b) Average productivity of all plants located in Indonesia in the steady state and iscalculated using plant-level revenue shares as weights.

I can rewrite the firm level revenue function as follows:

ri(ai, dit) = exp(

ln(ϕB)(1− dxit)(1− dfdi

it ) + ln(ϕ∗B + ϕB/ϕW exp(ϕτ ))(1− dfdi

it )dxft

+ ln(ϕB + ϕW ϕ∗B exp(ϕτ ))d

fdiit dx

it + ln(ϕ∗B + ϕB)dfdi

it (1− dxit)− ln ai

)where starred (unstarred) parameters replace unstarred (starred) parameters for foreign

firms.

At the counterfactual price P (θ) and P ∗(θ), the reduced form coefficients ϕB and ϕ∗B

take the values

ϕB =E

E∗

(P (θ)

P ∗(θ)

)ε−1

= ϕBk(θ, θ)

ϕ∗B =

E∗

E∗

(P ∗(θ)

P ∗(θ)

)ε−1

= ϕ∗Bk∗(θ, θ) (A.1)

where k(θ, θ) = (P (θ)/P (θ))ε−1 and k∗(θ, θ) = (P ∗(θ)/P ∗(θ))ε−1 represent the equilib-

rium price changes. We can then evaluate the revenue functions using the counterfactual

coefficients ϕB and ϕ∗B which have been adjusted for the change in the counterfactual prices

120

levels P (θ) and P ∗(θ):

ri(ai, dit) = exp(

ln(ϕB)(1− dxit)(1− dfdi

it ) + ln(ϕ∗B + ϕB/ϕW exp(ϕτ ))(1− dfdi

it )dxft

+ ln(ϕB + ϕW ϕ∗B exp(ϕτ ))d

fdiit dx

it + ln(ϕ∗B + ϕB)dfdi

it (1− dxit)− ln ai

)The equilibrium price change is determined so that the free entry conditions hold∫

V (a; θ, k(θ, θ))ga(a)da = fe∫V ∗(a∗; θ, k(θ, θ))ga∗(a

∗)da∗ = f ∗e

where V (a; θ, k(θ, θ)) and V ∗(a∗; θ, k(θ, θ)) are the solutions to the Bellman equations

(2.13)-(2.16) using the counterfactual adjusted revenue function and ga(a) and ga∗(a∗) are

the normal probability density functions from which the initial productivity level are drawn.

A.2.3 Mass of Firms

The estimated model does not provide a prediction for the relative mass of foreign and

domestic firms in Indonesia after the counterfactual change. This is important since the

overall change in average productivity or welfare in Indonesia will depend not only on the

probability of foreign firms investing in Indonesia, but also on the number of foreign firms

relative to the size of the domestic economy. Recovering the mass of firms in Indonesia

is difficult because the policy change in Indonesia may induce a change in the equilibrium

mass of firms worldwide (domestic or foreign) and the probability of entering Indonesia.

Fortunately, we can use the equilibrium conditions (3.10) along with the assumption that

any policy change in Indonesia does not affect the mass of “potential” foreign or domestic

entrants to the economy so that we can calculate the mass of firms in the economy as:

121

M =

∫ P χ(χ=1|a)

P χ(χ=0|a)ga(a)da∫ P χ(χ=1|a)

P χ(χ=0|a)ga(a)da

M

where M indicates the mass of firms calculated under the estimated model and the M indi-

cates counterfactual values. Starred variables replace unstarred variables when calculating

the mass of foreignfirms.

A.3 Fixed Cost Bounds

If exporting multinationals are more productive than non-exporting multinationals than the

following bounds on fixed costs must hold

f

(τB

B∗

)1−ε

< fx (C1a)

( w

w∗τ

)1−ε

<f + ffdi

fx

(C2a)

f + ffdi − fx

fx − f<

(w1−ε − (w∗τ)1−ε

(wτ)1−ε − w∗1−ε

)B

B∗ (C3a)

τw < w∗ (C4a)

A.4 Additional Figures

A.4.1 Decision Trees

The following two figures outline the order of decisions in the empirical model for foreign

and Indonesian firms:

Home firms first draw a cost shock associated with their decision to enter the market or

to exit. Then, conditional on entering the market, each firm draws a cost shock associated

122

Exitχ = 0

Enter/Stayχ = 1

NoExportdx = 0

Exportdx = 1

Figure A.1: Indonesian Firms

with its export decision. Given the realized export cost shock each firm must then decide

whether to export this period or not. Simiarly, foreign firms first draw a cost shock associ-

ated with their decision to enter the market or to exit. Conditional on entering the market

they draw a cost shock associated with its FDI decision and decide whether to invest in

Indonesia or not. Conditional on the investment decision they draw an export cost shock

and make their export decision.

A.4.2 5-Digit Industry Differences Across Export Status

Here I report the distributions foreign exporters and foreign non-exporters across 5-digit

industries (for selected 2-digit industries). Each dark column represents the percentage

of all foreign non-exporters in a particular 5-digit industry. Similarly, each light column

represents the percentage of all foreign exporters in a particular 5-digit industry.

Figure A.3 shows the percentage of foreign exporters and non-exporters across 5-digit

industries in the food industry. Across the large majority of industries it is evident that

industries that have a higher percentage of all foreign exporters also tend to have a higher

123

Exitχ = 0

Enter/Stayχ = 1

No FDIdi = 0

FDIdi = 1

NoExportdx = 0

NoExportdx = 0

Exportdx = 1

Exportdx = 1

Figure A.2: Foreign Firms

percentage of all foreign non-exporters. There are two notable exceptions at the right-

side of the graph representing the soft drink and mineral water industry and the industry

classified as “other food products n.e.c” where foreign non-exporters are notably more

prevalent.

Figures A.4 and A.5 graph the same distributions for the manufactured metals and tex-

tiles industries, respectively. A similar pattern emerges from these diagrams: 5-digit indus-

tries that receive a higher percentage of the total number of foreign exporters and also tend

to receive a higher percentage of the total number of foreign non-exporters. However, there

are differences in both industries. We tend to see a slightly higher percentage of foreign

non-exporters among industries that concentrate on spinning and weaving and a slightly

higher percentage of foreign exporters among industries that fabricate clothing. Similarly,

there is a higher number of foreign non-exporters in the transport industry. This is likely

124

Percentage of Foreign Plants in Each 5−Digit Industry

5−Digit Industries0

%

5

10

15

Foreign Exporters Foreign Non−Exporters

Figure A.3: Food

Note: Each pair of columns along the x-axis represents a different 5-digit industry (ISIC codes).

due to the Indonesian government’s sponsorship of foreign enterprises in the transport in-

dustry during this period. The differences across industries may account for some of the

observed productivity differences across foreign exporters and non-exporters. However, as

shown is Section 2.2, there are still statistically and economically significant differences

across foreign exporters and non-exporters even once we control for the industrial classifi-

cation.

125

Percentage of Foreign Firms in Each 5−Digit Industry

5−Digit Industries0

4

8

12

%Foreign Exporters Foreign Non−Exporters

Figure A.4: Metals

Note: Each pair of columns along the x-axis represents a different 5-digit industry (ISIC codes).

A.5 Measurement of Aggregate Productivity and Welfare

A.5.1 Productivity

To calculate aggregate productivity I weight each plant’s productivity level by its revenue

share in the Indonesian market. I denote total revenue earned by plants located in Indonesia

from Indonesian sales as

RIND =∑

i

r(ai)

where r(ai) = (aiwB)1−ε. The average productivity A is

A =∑

i

r(ai)

RIND

1

ai

126

Percentage of Foreign Plants in Each 5−Digit Industry

0

10

20

30

%

5−Digit Industries

Foreign Exporters Foreign Non−Exporters

Figure A.5: Textiles

Note: Each pair of columns along the x-axis represents a different 5-digit industry (ISIC codes).

A.5.2 Welfare

Following Melitz (2003) I define welfare as W = wL/PF where PF is the price index for

the full Indonesian economy

PF = w1−δP δ.

I define P as the price index for the manufacturing sector alone

P =

[∫v∈V

p(v)ε−1dv

] 1ε−1

.

127

A.6 Export and Ownership Premia: Fixed Effects and the

Chemicals Industry

In this section I report the estimated export and ownership premia estimated from fixed

effects regressions for the entire sample excluding the chemicals industry and pooled OLS

results for the chemicals industry alone. Although the estimates between foreign exporters

and non-exporters in Table 2.3 are statistically significant, these differences could be driven

by unobserved plant-specific differences. Fixed effects panel estimation is the most com-

mon regression used to control for plant-specific effects. However, there are two difficulties

with implementing a fixed effects regression here. First, ownership and export status are

very persistent over the short panel. Since the source of identification in a fixed effects

regression comes from variation in ownership and export status within each firm, there is

little variation from which to identify the coefficients. Second, because the identification is

coming from plants which switch export status and/or ownership status, the definition used

to identify these variables is particularly important. In general, the results are robust to

changes in the percentage of equity held by foriegn investors, but sensitive to the definition

of an exporter.

I observe in the data that plants that receive a relatively small percentage of total

revenues from export sales can be broadly characterized much more like foreign non-

exporters. If we define an exporter as one that has positive export sales, then the plants that

switch export status most often are going to plants that are in fact foreign non-exporters

in some years and receive only a small amount of revenue from export sales. If most of

the identification comes from these plants we would expect to observe very little difference

between foreign non-exporters and exporters in a fixed effects regression. Suppose one

128

changes the definition of an exporter so that exporters must receive at least a 25 percent of

total revenues from export sales. In that case, most of the identification will come from ex-

porting firms that cross the 25 percent threshold. A natural reason for arbitrarily increasing

the percentage of export revenues required for export status is that export intensive firms

are more likely to export to developed countries, while firms that export only a small per-

centage of revenues are more likely to emerging market countries near Indonesia. In fact,

in Table 2.5 I show that industries where firms export intensively have a higher percentage

of exports destined for Japan, the US and Western Europe. As such, I estimate equation

(2.1) by both pooled OLS and fixed effects and report the results for the measure of output

per worker as I vary percentage of export sales required for a foreign firm to have export

status.

Table A.5: Export & Ownership Premia: Output per Worker 1993-1996Pooled OLS Fixed Effects

Export Domestic Foreign Foreign Domestic Foreign ForeignThreshold Exporters Exporters Non-Exporters Exporters Exporters Non-Exporters0 0.141 0.534 0.800 0.007 0.294 0.306

(0.012) (0.023) (0.027) (0.011) (0.031) (0.030)25 0.176 0.490 0.869 0.028 0.253 0.288

(0.014) (0.028) (0.028) (0.012) (0.037) (0.035)50 0.174 0.435 0.869 0.027 0.209 0.309

(0.014) (0.029) (0.027) (0.012) (0.038) (0.034)No. of Obs. 57518 10801

Notes: Standared errors are in parentheses. Plants in the chemicals industry are excluded from the estimation and results for thechemicals industry are reported in the Appendix.

Table A.5 reports the results of the pooled OLS and fixed effects regressions. It shows

that as we vary the defintion of a foreign non-exporter the differences in the output per

worker premia grow.1 Moreover, the differences are increasingly significant even when

using a fixed effects estimator.2

1Similar results are found for other measures of premia.2A potential concern is that foreign exporters and foreign non-exporters produce very different products.

129

I also report the pooled OLS results for the chemicals industry alone. While the results

for export and ownership premia measured in terms of average wages, the ratio of non-

production to total workers, capital per worker, domestic sales and total sales and employ-

ment are consistent with the results presented in Section 2, the output per worker measure

is not consistent with the premia shown in other industries. In fact, in the chemicals indus-

try, output per worker is highest for foreign exporters, rather than non-exporters. Part of the

difference may be attributed to the fact that the mean difference total employment between

foreign exporters and non-exporters is much smaller in the chemicals industry than in other

industries. This may be indicative of increasing returns to scale in an industry that is largely

influenced by the production of natural gas for export.

Table A.6: Export & Ownership Premia: ChemicalsPooled OLS: 1993-1996

Export/Ownership Status Domestic Exporters Foreign Exporters Foreign Non-ExportersOutput per Worker 0.702 1.493 1.122

(0.038) (0.056) (0.053)Average Wage 0.122 0.776 0.900

(0.022) (0.033) (0.031)Non-Production/Total Workers 0.012 0.099 0.392

(0.029) (0.042) (0.041)Capital per Worker 0.381 0.814 1.055

(0.034) (0.049) (0.047)Domestic Sales -0.462 0.521 1.060

(0.045) (0.068) (0.060)Total Sales 0.694 1.476 1.100

(0.040) (0.059) (0.056)Total Employment 1.034 1.000 0.717

(0.031) (0.047) (0.046)No. of Observations 8,478

Notes: Standared errors are in parentheses.

While all pooled OLS regressions include 5-digit ISIC industry level dummies, it is possible that these dum-mies are not capturing fixed industry level differences for the foreign firms alone. Suppose that the foreigntechnology in one 5-digit industry is more productive that the foreign technology in another. If foreign non-exporters are more likely to be in the most highly productive industries then the output per worker premiawill be biased upwards. Fortunately, using plant-level fixed effects will control for this possibility. Since thesignificance is much lower on the fixed effects regressions than it is on the pooled OLS regressions I alsoreport the empirical distributions of foreign exporters and foreign non-exporters in the Appendix. As shownin the empirical distributions the mass of each type of plant in each 5-digit industry is highly correlated.

130

A.7 Mark-Ups & Productivity

A potential concern in the measurement of productivity in Section 2.2 is that firms that

foreign exporters may charge lower mark-ups on average relative to foreign non-exporters.

This may happen if firms that export back to Europe, Japan or the United States may charge

a parent company a lower price than they would an arms-length buyer in order to claim

higher profits in the parent’s country (transfer pricing). Moreover, if the mark-up in export

markets is lower than that in Indonesia due to higher competition we might observe lower

mark-ups in export markets. If this is the case, our measurement of the productivity/output

per worker of foreign exporters in Section 2.2 would be downwards biased. It would not

affect the other premia.

In contrast, if export markets are more profitable than Indonesian domestic markets,

(e.g. due to pricing to market, for example), then the bias would operate in the opposite

direction. In this case, the reported productivity differences between foreign exporters are

non-exporters would be smaller than the actual productivity differences.

Table A.7 reports the top corporate tax rates for Indonesia and the countries which are

its main sources of foreign investment of the 1993-1996 period. It shows that Indonesia

had, on average, the lowest corporate tax rates across this group of countries. As such,

Table A.7 suggests that foreign firms likely had more incentive to claim profits in Indonesia

rather than abroad over this period.

Similarly, if exporters were charging lower prices due to transfer pricing or highly com-

petitive export markets, one might expect to be able to observe those differences in esti-

mates of the firms mark-up behaviour. In particular, one can estimate the mark-up across

firms as

mark-up =revenues− variable costs

revenues.

131

Table A.7: Top Corporate Tax Rate1993 1994 1995 1996 1993-96 Avg.

Indonesia 0.350 0.350 0.300 0.300 0.325France 0.333 0.333 0.333 0.333 0.333

Germany 0.500 0.450 0.450 0.300 0.425Japan 0.375 0.375 0.375 0.375 0.375

United Kingdom 0.330 0.330 0.330 0.330 0.330United States 0.350 0.350 0.350 0.350 0.350

Notes: Taken from the World Tax Database at the University of Michigan. Data refer to the top marginal tax rate on domestic corpora-tions. Wie (1994) reports that all restrictions to FDI had been removed before 1993 and in fact Indonesia offered tax incentives/holidaysto many foreign firms over the 1993-1996 period. As such, the numbers presented in this table may be interpreted as an upper bound onthe top corporate tax rate in Indonesia.

Using the sum of production labour, intermediate inputs, electricity and fuel expenses as a

measure of variable costs, I estimate the average mark-up for each group of plants. Table

A.8 reports that foreign exporters tend to charge slightly higher mark-ups than foreign

non-exporters in the food and manufactured metals industries and slightly lower mark-ups

in the textiles industry, although the standard deviations are very high across all groups.

Moreover, while Table A.8 may suggest that the productivity difference between foreign

exporters and non-exporters may be slightly smaller than suggested by the estimated premia

in Section 2, there is little evidence to suggest that it is in fact due to mark-up heterogeneity

across exporters and non-exporters.

Table A.8: Mark-Ups Across Firms and IndustriesDomestic Non-Exporters Domestic Exporters Foreign Non-Exporters Foreign Exporters All Plants

Food 0.254 0.311 0.384 0.413 0.260(0.170) (0.210) (0.226) (0.189) (0.175)

Metals 0.315 0.410 0.401 0.425 0.332(0.182) (0.213) (0.208) (0.213) (0.191)

Textiles 0.253 0.275 0.317 0.300 0.257(0.159) (0.169) (0.185) (0.194) (0.162)

Standard deviations are in parentheses.

132

A.8 Robustness

In this section I present the structural estimates and the counterfactual results under dif-

ferent fixed cost assumptions. Specifically, I provide estimates for the model under the

following sets of assumptions:

A1. f ∗ = 0.510f and f ∗x = 0.380fx;

A2. f ∗ = f and f ∗x = fx;

A3. f ∗ = 0.255f and f ∗x = 0.190fx.

Assumption (A1) is the assumption used in the main text and the coefficients on the f and

fx terms are estimated from the World Bank Doing Business Report. Assumption (A2)

considers the case where fixed costs are symmetric across countries while assumption (A3)

estimates the model under the assumption that the fixed costs in the foreign country are one

half of that assumed in the baseline estimation.

133

Table A.9: Structural EstimatesIndustry Food Metals Textiles

Assumption A1 A2 A3 A1 A2 A3 A1 A2 A3κ%χ 52.416 39.774 55.199 31.780 40.1528 41.483 35.490 92.837 24.046

(4.481) (1.347) (4.983) (2.815) (3.233) (4.155) (2.053) (8.468) (1.324)κ%x 10.287 13.725 15.063 4.897 7.7448 6.570 7.9016 8.842 11.480

(0.587) (0.353) (0.828) (0.373) (0.2759) (0.440) (0.464) (0.493) (0.566)κ%x∗ 3.428 0.00001 0.482 1.055 0.00005 0.502 0.455 1.729 0.507

(6.545) (0.0002) (1.714) (16.411) (36.469) (17.987) (0.619) (12.001) (6.299)κ%fdi 0.0001 0.001 0.0002 0.001 0.0009 0.0002 0.0001 0.001 0.001

(1.423) (0.020) (1.458) (4.421) (28.567) (4.103) (1.667) (3.295) (3.642)κf 2.379 5.884 4.630 3.567 4.5498 8.087 5.870 2.010 6.390

(1.167) (0.0003) (1.255) (0.830) (0.8336) (1.384) (0.680) (2.020) (0.514)κfx 42.860 48.8470 57.304 21.923 26.808 27.933 23.513 26.507 39.071

(2.307) (1.069) (2.966) (1.548) (0.7574) (1.677) (1.324) (1.416) (1.811)κffdi 28.041 17.638 28.305 29.293 27.525 23.282 46.737 44.320 25.114

(3.361) (0.026) 5.783 (10.881) (20.425) (3.658) (4.986) (5.657) (11.122)ζfx 0.078 0.120 0.011 0.088 0.117 0.075 0.0339 0.056 0.037

(0.110) (0.003) (0.036) (0.082) (0.064) (0.056) (0.093) (0.145) (0.052)ρ -0.009 0.010 0.014 0.011 0.004 0.024 0.008 0.013 0.012

(0.010) (0.001) (0.009) (0.013) (0.010) (0.012) (0.009) (0.010) (0.008)ϕB 0.853 0.777 0.768 1.531 1.546 1.165 1.486 1.347 0.965

(0.020) (0.014) (0.015) (0.048) (0.406) (0.033) (0.027) (0.028) (0.016)ϕτ -5.346 -5.312 -5.327 -3.799 -3.864 -3.796 -5.369 -5.320 -5.316

(0.060) (1e-07) (0.056) (0.078) (0.004) (0.068) (0.060) (0.057) (0.0494)σa 1.041 1.097 0.961 0.996 0.947 0.996 1.011 1.300 0.982

(0.007) (0.008) (0.007) (0.008) (0.006) (0.008) (0.003) (0.006) (0.004)σa∗ 1.139 1.364 1.239 0.971 0.795 1.062 0.941 1.399 0.946

(0.010) (0.014) (0.010) (0.009) (0.006) (0.009) (0.007) (0.014) (0.007)ξ 0.006 0.010 0.001 0.021 0.011 0.005 0.014 0.009 0.006

(0.001) (0.001) (0.0001) (0.002) (0.001) (0.001) (0.001) (0.001) (0.0004)ε = 1/mark-up 3.8 3.0 3.8log-likelihood -34,806 -34,803 -34,651 -16,820 -17,126 -16,612 31,866 -30,379 -31,918No. of Obs. 17,786 7,549 13,287

Notes: Standard errors are in parentheses. The parameters are evaluated in units of millions of Indonesian Rupiahs in 1983. Metalsrefers to manufactured metals.

A.9 Variation in Plant-Level Productivity

A potential concern with the model and estimation is that plant-level productivity is not

persistent, as it is modelled, but variable over time. If plant-level productivity displayed

a great deal of variation over time, it would likely be reflected in variation in plant-level

revenue. Tables A.13 and A.14 show that in fact both domestic and foriegn plants in In-

donesia display substantial persistence in the revenues earned over time, even in a growing

economy.

134

Table A.10: Counterfactual Experiments - FoodExperiment Autarky No Trade No FDIAssumption A1 A2 A3 A1 A2 A3 A1 A2 A3% ∆ Avg. Productivitya -0.159 -0.138 -0.171 -0.038 -0.026 -0.027 -0.118 -0.087 -0.153−(ε − 1)∆ ln P b -0.041 -0.081 -0.085 -0.00004 -0.0005 -0.00006 0 0 -0.0002

Notes: a) The percentage change in average productivity for all plants in the steady state. Calculated using plant-level revenue shares asweights and evaluated relative to the estimated model. b) The welfare impact must be multiplied by industry share δ.

Table A.11: Counterfactual Experiments - Metalsa

Experiment Autarky No Trade No FDIAssumption A1 A2 A3 A1 A2 A3 A1 A2 A3% ∆ Avg. Productivityb -0.304 -0.247 -0.247 -0.034 -0.016 -0.028 -0.272 -0.227 -0.201−(ε − 1)∆ ln P c -0.022 -0.051 -0.035 -0.00002 0 0 0 0 0

Notes: a) Metals refers to manufactured metals. b) The percentage change in average productivity for all plants in the steady state.Calculated using plant-level revenue shares as weights and evaluated relative to the estimated model. c) The welfare impact must bemultiplied by industry share δ.

Table A.12: Counterfactual Experiments - TextilesExperiment Autarky No Trade No FDIAssumption A1 A2 A3 A1 A2 A3 A1 A2 A3% ∆ Avg. Productivitya -0.206 -0.172 -0.232 -0.016 -0.037 -0.016 -0.182 -0.111 -0.209−(ε − 1)∆ ln P b -0.053 -0.068 -0.067 0 0 -0.000002 0 0 0

Notes: a) The percentage change in average productivity for all plants in the steady state. Calculated using plant-level revenue shares asweights and evaluated relative to the estimated model. b) The welfare impact must be multiplied by industry share δ.

Table A.13: Revenue Bracket Transition Matrix - Domestic PlantsYear t + 1

Year t 0 to 10 10 to 20 20 to 30 More than 30 Exit0 to 10 0.967 0.024 0.004 0.005 0.09110 to 20 0.246 0.503 0.139 0.066 0.04920 to 30 0.092 0.215 0.391 0.302 0.037More than 30 0.040 0.037 0.067 0.856 0.031

Notes: The revenue brackets are in millions of Indonesian Rupiahs. One million Indonesian Rupiahs was worth approximately 452.55US dollars in 1983.

135

Table A.14: Revenue Bracket Transition Matrix - Foreign PlantsYear t + 1

Year t 0 to 10 10 to 20 20 to 30 More than 30 Exit0 to 10 0.780 0.126 0.045 0.049 0.16810 to 20 0.202 0.470 0.202 0.125 0.02220 to 30 0.122 0.218 0.340 0.319 0.015More than 30 0.023 0.034 0.061 0.882 0.013

Notes: The revenue brackets are in millions of Indonesian Rupiahs. One million Indonesian Rupiahs was worth approximately 452.55US dollars in 1983.

136

Appendix B

Appendix for Chapter 3

B.1 Transition Probabilities

In tables B.1 and B.2 I report the transition probabilities for the food and manufactured

metals industries. Similar to the textiles industry, the model captures a substantial portion

of the persistence in non-exporter status for both foreign and domestic firms. Although

the model captures much more of the persistence in export and ownership status than the

model withou sunk costs, it still misses some of the persistence in exporter status in the

metals industry.

Table B.1: Transition Probabilities - FoodActual Dom. Non-Exporters Dom. Exporters For. Non-Exporters For. Exporters ExitDom. Non-Exporters at t 0.979 0.018 0.003 0.0005 0.094Dom. Exporters at t 0.292 0.694 0.003 0.010 0.071For. Non-Exporters at t 0.087 0 0.734 0.179 0.024For. Exporters at t 0.022 0.034 0.225 0.719 0.033PredictedDom. Non-Exporters at t 0.972 0.028 — — 0.094Dom. Exporters at t 0.556 0.444 — — 0.043For. Non-Exporters at t — — 0.716 0.284 0.001For. Exporters at t — — 0.684 0.316 0.001

137

Table B.2: Transition Probabilities - MetalsActual Dom. Non-Exporters Dom. Exporters For. Non-Exporters For. Exporters ExitDom. Non-Exporters at t 0.956 0.033 0.007 0.005 0.062Dom. Exporters at t 0.272 0.699 0.014 0.014 0.037For. Non-Exporters at t 0.071 0.009 0.750 0.170 0.040For. Exporters at t 0.022 0.017 0.191 0.774 0.041PredictedDom. Non-Exporters at t 0.962 0.036 — — 0.061Dom. Exporters at t 0.729 0.271 — — 0.032For. Non-Exporters at t — — 0.556 0.444 0.007For. Exporters at t — — 0.523 0.477 0.007

Note: Metals refers to manufactured metals.

138

B.2 Counterfactual Experiments

B.2.1 Counterfactual Results for the Food and Metals Industries

The results in the food and manufactured metals industries are similar to those in the textiles

industry. Table B.3 demonstrates that average productivity in the food industry falls by

36, 1 and 28 percent in the autarky, no trade and no FDI experiments. Also, the welfare

impact is several times larger under autarky than the other experiments due to the lack of

substitution of FDI for trade (or vice-versa).

Table B.3: Counterfactual Experiments - FoodBase Autarky No Trade No FDI

Avg. Productivitya 2.686 1.707 2.664 1.945−(ε − 1)δ∆ ln P — −δ0.022 −δ0.0001 -δ0.001

Exit/Entry Rate of Foreign — -1 2.6640 -1Firms in Indonesia% of For. Non-Exporters 0.017 0 0.088 0% of Dom. Exporters 0.044 0 0 0.041%∆ in Exports — -1 -1 -0.682

Mkt. Shr. of Dom. Non-Exp. 0.550 1 0.357 0.806Mkt. Shr. of Dom. Exp. 0.110 0 0 0.194Mkt. Shr. of For. Non-Exp. 0.248 0 0.643 0Mkt. Shr. of For. Exp. 0.093 0 0 0

Notes: a) Average productivity of all plants located in Indonesia in the steady state and is calculated using plant-level revenue shares asweights.

Table B.4 demonstrates that average productivity in the manufactured metals industry

falls by 41, 3 and 41 percent in the autarky, no trade and no FDI experiments. Again, the

welfare impact is several times larger under autarky than the other experiments due to the

lack of substitution of investment for trade (or vice-versa).

139

Table B.4: Counterfactual Experiments - Metalsa

Base Autarky No Trade No FDIAvg. Productivityb 3.277 1.921 3.187 1.929−(ε − 1)δ∆ ln P — −δ0.038 −δ0.002 -δ0.002

Exit/Entry Rate of Foreign — -1 0.224 -1Firms in Indonesia% of For. Non-Exporters 0.049 0 0.133 0% of Dom. Exporters 0.057 0 0 0.061%∆ in Exports — -1 -1 -0.861

Dom. Mkt. Shr. of Dom. Non-Exp. 0.444 1 0.446 0.906Dom. Mkt. Shr. of Dom. Exp. 0.051 0 0 0.094Dom. Mkt. Shr. of For. Non-Exp. 0.248 0 0.554 0Dom. Mkt. Shr. of For. Exp. 0.257 0 0 0

Notes: a) Metals refers to manufactured metals. b) Average productivity of all plants located in Indonesia in the steady state and iscalculated using plant-level revenue shares as weights.

B.3 Robustness

In this section I present the structural estimates and the counterfactual results under differ-

ent fixed and sunk cost assumptions. Specifically, I provide estimates for the model under

the following sets of assumptions:

A1. f ∗ = 0.510f , f ∗x = 0.380fx and cx∗ = 0.380cx;

A2. f ∗ = f , f ∗x = fx and cx∗ = cx;

A3. f ∗ = 0.255f , f ∗x = 0.190fx and cx∗ = 0.190cx.

Assumption (A1) is the assumption used in the main text and the coefficients on the f ,

fx and cx terms are estimated from the World Bank Doing Business Report. Assumption

(A2) considers the case where fixed and sunk costs are symmetric across countries while

assumption (A3) estimates the model under the assumption that the fixed and sunk costs in

the foreign country are one half of that assumed in the baseline estimation.

140

Table B.5: Structural EstimatesIndustry Food Metals Textiles

Assumption A1 A2 A3 A1 A2 A3 A1 A2 A3κ%χ 52.361 35.212 30.449 51.940 35.421 57.646 61.860 61.085 43.280

(4.567) (2.792) (1e-05) (6.215) (2.422) (1.824) (5.343) (2.700) (0.00001)κ%x 8.812 6.610 6.036 8.234 6.219 8.251 11.413 10.535 5.456

(0.541) (0.387) (0.341) (0.901) (0.282) (0.492) (0.957) (0.469) (0.288)κ%x∗ 0.669 0.721 0.311 0.225 0.190 0.425 0.278 0.274 0.283

(1.794) (0.673) (0.810) (2.067) (0.239) (1.163) (0.529) (0.213) (0.657)κ%fdi 0.959 2.244 1.746 3.360 1.970 3.542 4.115 0.317 3.168

(0.214) (0.108) (0.076) (0.168) (0.074) (0.131) (0.267) (0.494) (0.232)κcx 35.619 27.896 29.780 29.258 26.466 31.262 36.496 33.783 18.646

(2.194) (1.678) (1.816) (3.352) (1.437) (2.159) (3.185) (1.669) (1.062)κcfdi 38.907 50.857 35.451 50.062 53.628 57.326 45.647 42.547 46.406

(2.254) (2.270) (1.480) (2.185) (1.828) (1.486) (3.632) (2.613) (3.779)κf 1.221 1.591 2.244 0.692 0.490 4.493 2.455 1.385 3.023

(0.881) (0.630) (0.154) (1.313) (0.514) (0.142) (0.883) (0.464) (0.336)κfx 2.217 2.565 0.963 1.342 0.735 1.156 0.834 0.774 1.011

(0.659) (0.503) (0.492) (0.795) (0.664) (0.856) (0.628) (0.584) (0.315)κffdi 1.842 3.416 1.023 0.792 1.265 0.535 0.799 0.212 0.082

(2.000) (1.414) (1.036) (2.590) (0.615) (1.091) (2.815) (0.847) (1.350)ζfx 0.098 0.140 0.025 0.028 0.090 0.317 0.014 0.221 0.112

(0.507) (0.491) (1.463) (2.154) (0.110) (1.863) (0.002) (1.242) (1.514)ζcx 0.004 0.003 0.013 0.009 0.018 0.037 0.008 0.009 0.011

(0.010) (0.006) (0.035) (0.081) (0.023) (0.104) (0.017) (0.009) (0.027)ρ 0.004 -0.0001 2e-08 -0.005 0.010 0.015 0.0001 4.4e-10 0.023

(0.008) (0.009) (1e-07) (0.012) (0.005) (0.002) (0.006) (0.004) (0.002)ϕB 0.975 0.968 1.044 1.956 1.519 1.793 1.530 1.818 1.285

(0.015) (0.016) (0.010) (0.051) (0.031) (0.031) (0.020) (0.016) (0.016)ϕτ -5.331 -5.365 -5.330 -3.803 -3.801 -3.803 -5.345 -5.333 -5.328

(0.052) (0.050) (0.046) (0.084) (0.009) (0.023) (0.063) (0.017) (0.037)σa 0.778 0.743 0.726 0.742 0.692 0.719 0.803 0.803 0.682

(0.005) (0.005) (0.005) (0.006) (0.005) (0.006) (0.004) (0.003) (0.003)σa∗ 0.779 0.659 0.602 0.712 0.518 0.645 0.945 1.252 0.760

(0.007) (0.006) (0.007) (0.010) (0.005) (0.008) (0.015) (0.014) (0.011)ξ 0.001 0.001 0.003 0.007 0.002 0.001 0.014 0.023 0.001

(0.0001) (0.0001) (0.001) (0.001) (0.0002) (0.0001) (0.002) (0.001) (0.0001)ε = 1/mark-up 3.8 3.0 3.8log-likelihood -34795 -35045 -35587 -16347 -16769 -16169 -31649 -31708 -32509No. of Obs. 17,786 7,549 13,287

Notes: Standard errors are in parentheses. The parameters are evaluated in units of millions of Indonesian Rupiahs in 1983. Metalsrefers to manufactured metals.

Table B.6: Counterfactual Experiments - FoodExperiment Autarky No Trade No FDIAssumption A1 A2 A3 A1 A2 A3 A1 A2 A3% ∆ Avg. Productivitya -0.364 -0.297 -0.254 -0.025 -0.131 -0.177 -0.324 -0.262 -0.228−(ε − 1)∆ ln P b -0.022 -0.029 -0.011 -0.002 -0.029 -0.011 -0.014 -0.009 -0.001

Notes: a) The percentage change in average productivity for all plants in the steady state. Calculated using plant-level revenue shares asweights and evaluated relative to the estimated model. b) The welfare impact must be multiplied by industry share δ.

141

Table B.7: Counterfactual Experiments - Metalsa

Experiment Autarky No Trade No FDIAssumption A1 A2 A3 A1 A2 A3 A1 A2 A3% ∆ Avg. Productivityb -0.414 -0.486 -0.344 -0.003 -0.050 -0.032 -0.414 -0481 -0.351−(ε − 1)∆ ln P c -0.038 -0.021 -0.045 -0.036 -0.021 -0.045 -0.001 -0.001 -0.014

Notes: a) Metals refers to manufactured metals. b) The percentage change in average productivity for all plants in the steady state.Calculated using plant-level revenue shares as weights and evaluated relative to the estimated model. c) The welfare impact must bemultiplied by industry share δ.

Table B.8: Counterfactual Experiments - TextilesExperiment Autarky No Trade No FDIAssumption A1 A2 A3 A1 A2 A3 A1 A2 A3% ∆ Avg. Productivitya -0.224 -0.497 -0.251 -0.064 -0.002 -0.003 -0.189 -0.495 -0.229−(ε − 1)∆ ln P b -0.067 -0.053 -0.044 -0.001 -0.0003 -0.0003 -0.001 -0.001 -0.009

Notes: a) The percentage change in average productivity for all plants in the steady state. Calculated using plant-level revenue shares asweights and evaluated relative to the estimated model. b) The welfare impact must be multiplied by industry share δ.

142

Appendix C

Appendix for Chapter 4

C.1 Estimation Procedures: Selection and Adjustment Costs

C.1.1 The issue of selection in the LP approach

In this section we outline how we control for endogeneous selection while using LP inter-

mediate proxy approach. The idea is essentially the same as the one used in Olley and Pakes

(1996); namely, we first identify the state variables that are relevant for endogneous exiting

decisions and approximate the survival probabilities using the polynomials in the observ-

able variables. Then, we can control for the endogenous exiting decision by including the

polynomials in the survival probabilities when the moment conditions are constructed.

First, the state variables that are relevant for the plant exit decision are the predeter-

mined level of capital kit and the past import decision di,t−1. The model in section 2 implies

that a plant chooses to continue to produce if the current realization of productivity term

ωit is higher than the threshold value ωt(kit, di,t−1). One might think that the intermediate

143

proxy approach is not applicable to control for the selection bias because we cannot “re-

cover” ωit from observables given that we do not observe the current period intermediates

if the plant chooses to exit. Note, however, that ωit follows the first order Markov process

ωit = ξt + γdi,t−1 + ωi,t−1 + uit (equation (4.8) in Chapter 4) and, thus, it is possible to

approximate ωit using the observable variables (di,t−1, ωi,t−1), where ωi,t−1, in turn, can be

proxied by the past value of intermediates, the past capital, and the past import decision so

that ωi,t−1 = ω∗t−1(xi,t−1, ki,t−1, di,t−1) (equation (10) in Chapter 4).1

Specifically, the plant chooses to stay if ωit ≥ ωt(kit, di,t−1), or using ωit = ξt +

γdi,t−1 + ωi,t−1 + uit and ωi,t−1 = ω∗t−1(xi,t−1, ki,t−1, di,t−1), the plant stays if

uit ≥ ωt(kit, di,t−1)− ξt − γdi,t−1 − ω∗t−1(xi,t−1, ki,t−1, di,t−1)

≡ ut(kit, ki,t−1, di,t−1, xi,t−1),

where ut(kit, ki,t−1, di,t−1, xi,t−1) is the threshold value of uit that induces a plant to exit

at t. Since a plant continues in operation if uit ≥ ut(kit, ki,t−1, di,t−1, xi,t−1), the survival

probabilities are given by

Pr{χit = 1|ut(kit, ki,t−1, di,t−1, xi,t−1)} = 1− Fu(ut(kit, ki,t−1, di,t−1, xi,t−1))

= Pt(kit, ki,t−1, di,t−1, xi,t−1)

≡ Pit, (C.1)

where Fu(·) is the cumulative distribution of uit. Equation (C.1) corresponds to equation

(10) in Olley and Pakes (1996). We can approximately estimate (C.1) by probit using the

polynomials in (kit, ki,t−1, di,t−1, xi,t−1) as explanatory variables for the survival decision.

1The linearity assumption can be relaxed. For instance, suppose that ωit follows the stochastic processωit = ht(di,t−1, ωi,t−1, uit). Even in this case, we may apply the similar logic to control for the selectionbias as long as ht(di,t−1, ωi,t−1, uit) is strictly increasing in uit.

144

Once the survival probabilities are estimated in terms of the observables, the rest of the

procedure for controlling for the selection bias is essentially the same as that of the OP

approach. By inverting (C.1), we may obtain uit as a function of Pit and write this inverse

function as uit = u∗(Pit). Then, the conditional expectation of ωit given ωi,t−1, di,t−1, and

χit = 1 can be expressed as

E[ωit|ωi,t−1, di,t−1, χit = 1] = ξt + γdi,t−1 + ρωi,t−1 + E[uit|uit ≥ u∗(Pit)]. (C.2)

Here, the term E[uit|uit ≥ u∗(Pit)] controls for the selection bias. For instance, if we know

uit is normally distributed, this term becomes the inverse Mill’s ratio.

We obtain the estimate of E[ωit|ωi,t−1, di,t−1, χit = 1] by the pooled OLS regression

of ( ˆωit + ηit)(β∗) ≡ yit − βsl

sit − βul

uit − βeeit − β∗

kkit − β∗xxit − β∗

ddit on the past im-

port status di,t−1, the estimate of the previous period’s productivity shock ωi,t−1(β∗) ≡

φt−1(xi,t−1, ki,t−1, di,t−1)−β∗kki,t−1−β∗

xxi,t−1−β∗ddi,t−1, and a third-order polynomial series

of the survival probability (C.1) which approximates the term E[uit|uit ≥ u∗(Pit)]. Here,

φt(·) is the estimate of φt(·) obtained by the OLS regressions of yit− βslsit− βul

uit− βeeit on

a third-order polynomial series of (xit, kit, dit) while the survival probability is estimated

by the probit with a third-order polynomial series in (kit, ki,t−1, di,t−1, xi,t−1) as regressors.

In estimating (C.2), we also allow for year-specific constant terms, ξt, to control for the

year-specific aggregate productivity shocks.

Establishing the procedure to estimate E[ωit|ωi,t−1, di,t−1, χit = 1], we consistently

estimate βk, βx, and βd while controlling for the selection bias as discussed in Chapter 4.

145

C.1.2 Alternative Estimators: the Within-Groups and the GMM esti-

mators

To address the simultaneity issue, we also consider the following two alternatives: the

within-groups estimator and the system GMM estimator. The within-groups estimator only

uses the within-plant variation so that it is robust against the simultaneity arising from the

correlation between an unobserved plant-specific productivity shock and inputs. It is not

robust, however, against the simultaneity due to the correlation between a transitory shock

and inputs. Furthermore, the between-plant variation often plays an important role in iden-

tifying the parameters; this is especially true for coefficients of capital and imported inter-

mediates where the within-plant variation is much less than the between-plant variation due

to their slow adjustment over time. The within-estimator may lead to imprecise estimates

especially for capital and imported intermediates. This issue becomes more pronounced

when there is idiosyncratic measurement error in inputs; within-transformation lowers sig-

nal to noise ratio and magnifies the bias induced by measurement errors (cf., Griliches and

Hausman, 1986).

In order to control for simultaneity in panel data, Blundell and Bond (1998, 2000)

propose the system GMM estimator by extending the first-differenced GMM estimator

(cf., Arellano and Bond, 1991). Consider the equation (4.7) in Chapter 4 with the following

stochastic process of ωit:

ωit = ξt + γdi,t−1 + ρωi,t−1 + αi + vit, (C.3)

where ξt is a year-specific effect, αi is a plant-specific effect, vit is an i.i.d. productivity

shock. Using a dynamic common factor representation, equation (4.7) in Chapter 4 with

146

(C.3) can be rewritten as:

yit = βkkit − ρβkki,t−1 + βslsit − ρβsl

si,t−1 + βul

uit − ρβul

ui,t−1 + βeeit − ρβeei,t−1

+βxxit − ρβxxi,t−1 + βddit + (γ − ρβd)di,t−1 + ρyi,t−1 + ξt + αi + µit (C.4)

where µit = ηit − ρηi,t−1 + vit.

Following Blundell and Bond (2000), we first estimate the unrestricted parameter vector

of (C.4) by the one-step GMM and then obtain the restricted parameter vector (βk, βs, βu, βe, βm, βd, γ, ρ)

using minimum distance (cf., Chamberlain, 1982). The following moment conditions are

used:

E[zi,t−s∆µit] = 0 for s = 2, 3, (C.5)

E[∆zi,t−s(α∗i + µit)] = 0 for s = 1, (C.6)

where zit = (yit, kit, lsit, l

uit, xit, dit) and ∆zit = zit − zi,t−1. The first set of the moment

conditions (C.5) comes from the first differenced equations with lagged levels of the vari-

ables as instruments. Blundell and Bond (1998) find that exploiting the additional moment

conditions (C.6), based on the level equations with lagged differences of the variable as

instruments, may lead to dramatic reductions in finite sample bias. Recently, however,

some researchers have found that even the system GMM estimator could lead to imprecise

and possibly biased estimates due to weak instruments (e.g., Griliches and Mairesse, 1998;

Mulkay, Hall, and Mairesse, 2000; Levinsohn and Petrin, 2003).

C.1.3 Adjustment Costs: Ackerberg, Caves and Fraser (2005)

Suppose that skilled labor, unskilled labor, and energy are subject to adjustment costs so

that the past variables for skilled labor, unskilled labor, and energy are also this period’s

147

state variables. Denote sit = (lsit, luit, eit). Then, the demand functions are written as:

lsit = ls∗t (kit, dit, ωit, si,t−1), luit = lu∗t (kit, dit, ωit, si,t−1), and eit = e∗t (kit, dit, ωit, si,t−1). Or

we can write

sit = s∗t (kit, dit, ωit, si,t−1),

where s∗t (·) is a vector-valued function. We add the subscript t since the demand functions

depend on prices which are time-dependent.

We maintain the assumption that materials are not subject to adjustment costs. Since xit

is a freely variable input and, thus, the adjustment costs for skilled labor, unskilled labor,

and energy affect the demand for materials only through their effects on the choices of

sit, we may consider the demand function for materials conditioned on sit = (lsit, luit, eit):2

xit = x∗t (kit, dit, ωit, sit). In the Cobb-Douglas case, it is straightforward to verify that

this demand function is strictly increasing in ωit and we get the function ω∗t (kit, dit, xit, sit)

which corresponds to the equation (4.10) in Chapter 4.3

The rest of the estimation procedure is similar to the one discussed in Chapter 4. In

particular, we have an estimate of the residual for each candidate parameter vector as:

( ˆνit + ηit)(β∗) = yit−β∗

s lsit−β∗

uluit−β∗

eeit−β∗kkit−β∗

xxit−β∗ddit−E[ωit|ωi,t−1, di,t−1, χit = 1].

Based on the residual, we can construct the GMM estimator using the instrument Zit =

(kit, ki,t−1, di,t−1, di,t−2, xi,t−1, xi,t−2, lsi,t−1, l

si,t−2, l

ui,t−1, l

ui,t−2, ei,t−1, ei,t−2). That is, the pa-

rameters β∗ = (β∗k , β

∗s , β

∗u, β

∗e , β

∗x, β

∗d) are estimated by minimizing the GMM criterion

2Alternatively, as Ackerberg, Caves and Fraser (2005) suggest, we may also consider the demand functionfor materials as xit = x∗t (kit, dit, ωit, si,t−1) but, in this case, it is not easy to verify the strict monotonicitycondition.

3For example, consider a simplified version of production function: Yit = eωitKβk

it Lβl

it Xβx

it . In thiscase, sit = lit and we omit dit. Then, from the first order condition for Xit, we have x(kit, ωit, lit) =constt + (βk/(βx − 1))kit + (βl/(βx − 1))lit + ωit, where “constt" depends on prices and parameters. It’sclear that this function is strictly increasing in ωit. Inverting it with respect to ωit, we have ω∗t (kit, xit, lit) =constt − (βk/(βx − 1))kit − (βl/(βx − 1))lit + xit.

148

function Q(β∗) =∑12

h=1[∑N

i=1

∑Ti

t=1981( ˆνit + ηit)(β∗)Zit,h]

2, where Ti is the last year the

ith firm is observed. Note that, by using past, rather than current, labor and energy variables

as instruments to identify βs, βu and βe we allow for the possibility that a plant makes these

decisions after observing this period’s innovation in productivity.

C.2 Additional Estimation Results

C.2.1 Ackerberg, Caves, and Frazer (2005)

There are several reasons to suspect the coefficients estimated in the first stage of the OP/LP

technique. One possibility is that the level of skilled labor cannot be altered without incur-

ring extra adjustment costs; in this case, they are more properly defined as state variables

in the firm’s problem, rather than freely variable inputs. Columns (1) and (3) of Table

C.1 present the results from the OP/LP estimator where skilled labor is treated as a state

variable for the Basic and Extended Sample, respectively. Because skilled labor is a state

variable its coefficient is estimated in the second stage. The results show that the coeffi-

cients across all variables are reasonably similar to those found in the original experiment

and again indicate the substantial static and dynamic effects from using imports.

Recently, Ackerberg, Caves and Frazer (2005) show that the OP/LP estimation tech-

nique may not properly identify the coefficients estimated in the first stage. Their critique

argues that the demand functions for skilled labor, unskilled labor, and energy can be writ-

ten as functions of the state variables (kit, dit, ωit) and, since we have ωit = ω∗t (xit, kit, dit)

in equation (10), they are fully written as functions of (xit, kit, dit). Then, looking at the

first stage regression (11), we notice that lsit − E[lsit|xit, kit, dit] = 0 etc.. That is, there

would be no variability left in the regressors to identify the first stage coefficients.

149

To deal with this identification issue, we estimate all of the coefficients in the second

stage of the OP/LP technique, extending a method proposed by Ackerberg et al. (see also

Bond and Söderbom, 2005). Our method assumes that skilled labor, unskilled labor, and

energy are also subject to adjustment costs. Then, their demand functions depend on their

past values, which in turn provide variations in their current values that are independent of

(kit, dit, ωit) to identify their coefficients. In this case, βs, βu, and βe are identified from the

moment conditions E[(νit + ηit)lsit−1] = 0, E[(νit + ηit)l

uit−1] = 0, and E[(νit + ηit)eit−1] =

0, respectively, while (βx, βk, βd) are identified from the moment conditions E[(νit +

ηit)xit−1] = 0, E[(νit + ηit)kit] = 0, and E[(νit + ηit)dit−1] = 0. We also include six over-

identifying conditions using the predetermined variables (ki,t−1, di,t−2, xi,t−2, lsi,t−2, l

ui,t−2, ei,t−2).

The results are reported in columns (2) and (4) of Table C.1. In column (2) we see that

the coefficient for unskilled labor is close to zero and the standard errors are considerably

wider on skilled labor and energy. Similarly, in column (4) the coefficient on skilled labor

is close to zero and the standard errors are again wider on skilled labor, unskilled labor and

energy. While this may be indicative of model misspecification, it could likely to point to

a lack of good instruments for those variables. However, the size and significance of the

coefficients measuring the static and dynamic effect from using imports does not change

qualitatively across the estimation procedure. Both βd and γ are positive and significant

even when all coefficients are estimated in the second stage of the OP/LP procedure.

C.2.2 Industry-level Results

We also check how the results change across industries. We estimate the production func-

tions based on the Basic Sample for two of the largest 3-digit level industries (ISIC codes):

Food (311) and Metals (381). Table C.2 presents the results from the OP/LP estimators,

150

Table C.1: Additional Estimates of Production Function: Discrete Import Variable

The Data Set Basic Sample Extended SampleEstimators OP/LP Proxy ACF OP/LP Proxy ACF

(1) (2) (3) (4)Skilled labor 0.111 0.049 0.076 0.0001

(0.032) (0.181) (0.010) (0.084)Unskilled labor 0.147 0.0001 0.138 0.107

(0.008) (0.017) (0.009) (0.021)Energy 0.041 0.199 0.048 0.097

(0.004) (0.143) (0.004) (0.096)Capital 0.060 0.036 0.075 0.058

(0.009) (0.012) (0.010) (0.013)Materials 0.603 0.528 0.632 0.655

(0.023) (0.049) (0.020) (0.034)Disc. Import (βd) 0.260 0.228 0.178 0.162

(0.037) (0.047) (0.035) (0.031)γ 0.035 0.051 0.024 0.028

(0.009) (0.007) (0.008) (0.007)ρ 0.815 0.868 0.816 0.821

(0.024) (0.035) (0.024) (0.028)Implied γ

1−ρ0.190 0.388 0.130 0.157

P-value for over- 0.744 0.719 0.975 1.000identification test

No. of Obs. 33200 45518

Notes: Standard errors are in parentheses. Columns (1)-(2) use the “Basic Sample" that excludes plants for which the initial capital stockare not reported. Columns (3)-(4) use the “Extended Sample" in which a missing initial capital stock is imputed by a projected initialcapital stock based on other reported plant observables. The OP/LP estimators in columns (1) and (3) specify the stochastic process ofωt of the equation (4.8) in Chapter 4. The OP/LP estimators in column (1) and (3) estimate the coefficient of skilled labor in the secondstage by treating it as an additional state variable. The ACF estimator in column (2) and (4) treats skilled labor, unskilled labor, andenergy as additional state variables and, hence, estimates all of these coefficients in the second stage.

151

Table C.2: Estimates of Production Function for Food and Metal IndustriesDiscrete Import Variable Continuous Import Variable

Industry Food Metals Food Metalsωit process AR(1) Series AR(1) Series AR(1) Series AR(1) SeriesSkilled labor 0.073 0.139 0.072 0.138

(0.009) (0.017) (0.008) (0.017)Unskilled labor 0.088 0.199 0.091 0.203

(0.015) (0.023) (0.011) (0.023)Energy 0.070 0.051 0.072 0.050

(0.010) (0.011) (0.007) (0.010)Capital 0.051 0.050 0.074 0.075 0.051 0.050 0.099 0.092

(0.009) (0.009) (0.032) (0.033) (0.009) (0.010) (0.036) (0.036)Materials 0.664 0.658 0.400 0.378 0.753 0.722 0.434 0.408

(0.055) (0.060) (0.097) (0.099) (0.025) (0.040) (0.096) (0.104)Discrete Import 0.257 0.191 0.243 0.227

(0.103) (0.111) (0.110) (0.124)Continuous Import 0.370 0.422 0.258 0.234

(0.167) (0.209) (0.157) (0.181)γ 0.064 — 0.060 — 0.068 — 0.037 —

(0.025) (0.030) (0.107) (0.056)ρ 0.837 — 0.881 — 0.748 — 0.884 —

(0.067) (0.073) (0.107) (0.093)Implied γ

1−ρ0.393 — 0.504 — 0.270 — 0.319 —

Implied θ 3.035 2.711 2.682 2.748P-value for over-identification test 0.784 0.769 0.809 0.834 0.724 0.871 0.844 0.839

No. of Obs. 12273 3733 12273 3733

Notes: Standard errors are in parentheses. The estimates are based on the “Basic Sample" that excludes plants for which the initial capitalstock are not reported. The OP/LP estimators that specify ωit processes by “Series" use the third order polynomials in (ωt−1, dt−1)for discrete import variable and (ωt−1, nt−1). for continuous import variable.

152

where ωit processes are specified using either AR(1) or a third-order series approximation.

Probably reflecting a difference in the sample sizes, the standard errors for Metal Industry

are generally larger than those for Food Industry.

Using the discrete import variable, the estimated coefficients on the import variables

under the AR(1) specification are reported in columns (1) and (3) and they indicate a large

positive static effect on productivity (24.3-25.7 percent). When we specify ωit processes by

series in columns (2) and (4), the estimates are slightly lower but still large (19.1-22.7 per-

cent) although they are marginally significant at a 10 percent level. The estimated values of

γ for the discrete import variable are positive and significant for both industries, suggesting

a positive dynamic effect of the usage of imported materials.

As for the results from using the continuous import variable, all the estimated coeffi-

cients for the continuous import variable are positive and of large size, ranging from 23.4 to

42.2 percent, but the estimates from the Metal industry are not as significant; for the Metal

industry, the estimate from AR(1) specification in column (7) is barely significant at a 10

percent level while the estimate from series in column (8) is not significant even at a 10

percent level. Although the relatively large standard errors for the Metal industry might be

due to its small sample size, the insignificance of the static productivity effect adds a caveat

on the positive impact of changing the amount of imported materials on productivity.

The estimated values of γ for the continuous import variable are positive but not signif-

icant for both industries. Thus, the evidence for the positive dynamic effect of an increase

in the usage of imported intermediates is, at best, weak. Compared to the result for the

discrete import variables, the relative insignificance of the dynamic effect of the usage of

imported materials in the regression using continuous import variables might indicate that,

it is not the intensive margin of how much a plant imports but the extensive margin of

153

whether a plant imports or not that determines the dynamic effect of importing materi-

als. This could be the case, for instance, if importing intermediates from foreign countries

per se—regardless of how much a plant imports—provides an opportunity to learn foreign

technologies and thus leads to a positive dynamic effect on productivity.

C.2.3 Energy As A Proxy

Table C.3 presents the results from the OP/LP Proxy estimator using energy as a proxy

(instead of materials) on the Basic Sample. Columns (1)-(4) use the discrete import vari-

able, while columns (5)-(8) present the results for the continuous import variable. We

have included the results for both energy and materials to ease comparison. Columns

(1),(2),(5) and (6) use materials as the proxy for productivity in the OP/LP estimation,

whereas columns (3),(4),(7) and (8) use energy as the proxy. Columns (1), (3), (5) and (7)

present the results under the assumption that the plant-specific productivity shock, ωit, fol-

lows an AR(1) process. Columns (2), (4), (6) and (8) use the OP/LP estimation technique

without making any assumptions on the structure of ωit and estimate ωit as a third order

polynomial in (ωi,t−1, Pit, di,t−1) where di,t−1 is the lagged decision to import.

The most important finding is the significance and large size of the current discrete

import variable across proxies. Comparing columns (1) and (3), (2) and (4), (5) and (7),

and, (6) and (8) it is clear that using the energy proxy increases the estimated impact of

switching to imported materials. Similarly, a positive and significant γ coefficient across

proxies strengthens the evidence of a dynamic import effect.

As a consequence of the strong dynamic effect, the long run effect, γ/(1 − ρ), is also

substantially higher using the energy proxy compared to the materials proxy. None of the

specifications are rejected by the bootstrapped overidentification test.

154

Table C.3: Panel OP/LP Estimates: Energy vs. MaterialsThe Data Set Basic Sample

Import Variable Discrete ContinuousThe Proxy Materials Energy Materials Energy

(1) (2) (3) (4) (5) (6) (7) (8)ωit process AR(1) Series AR(1) Series AR(1) Series AR(1) SeriesSkilled labor 0.137 0.137 0.138 0.138

(0.006) (0.022) (0.006) (0.006)Unskilled labor 0.145 0.143 0.148 0.147

(0.008) (0.024) (0.008) (0.007)Energy 0.043 0.057 0.061 0.044 0.098 0.084

(0.005) (0.016) (0.022) (0.005) (0.024) (0.022)Capital 0.058 0.064 0.026 0.029 0.066 0.074 0.047 0.041

(0.009) (0.010) (0.009) (0.010) (0.009) (0.010) (0.014) (0.011)Materials 0.549 0.509 0.643 0.616 0.577 0.643

(0.025) (0.029) (0.101) (0.021) (0.027) (0.006)Import 0.214 0.220 0.431 0.448 0.246 0.270 0.520 0.495

(0.035) (0.038) (0.039) (0.045) (0.052) (0.061) (0.158) (0.086)θ — — — — 3.505 3.115 2.237 1.170

(1.428) (2.361) (0.035) (0.064)γ 0.041 — 0.072 — 0.030 — 0.086 —

(0.011) (0.011) (0.010) (0.035)ρ 0.892 — 0.808 — 0.822 — 0.837 —

(0.027) (0.027) (0.031) (0.052)Implied γ

1−ρ0.380 — 0.373 — 0.169 — 0.528 —

(0.166) (0.107) (0.086) (0.246)P-value for over-identification test 0.593 0.759 0.930 0.980 0.824 0.759 1.000 0.915

No. of Obs. 33200

Notes: Standard errors are in parentheses. Columns (1)-(8) use the “Basic Sample" that excludes plants for which the initial capital

stock are not reported. Columns (1), (2), (5) and (6) use the materials variable as the OP/LP proxy, while columns (3), (4), (5) and (6)

use energy as a proxy. The OP/LP estimators in columns (2), (4), (6) and (8) specify the stochastic process of ωt using the third order

polynomials in (ωt−1, dt−1).

155

C.2.4 Monotonicity Condition

In this section we graphically examine the monotonicity condition required for the OP/LP

estimation procedure to be valid. The monotonicity condition is essentially the same as

that in Levinsohn and Petrin (2002): conditional on capital and the decision to import

intermediates, profit maximizing behavior must lead more productive firms to use more

intermediate materials.

Figure C.1 shows the relationship between productivity, capital and materials for each

group of years. On the left hand side we show this relationship for non-importers and on the

right hand side we show this relationship for importers. In all cases we see that the mate-

rials variable is increasing in both capital and productivity indicating that the monotonicity

condition is satisfied.

C.2.5 Other Omitted Results

In Chapter 4, we omitted some results from the Extended Sample because these results are

qualitatively very similar to those from the Basic Sample. In this section we report the

omitted results.

Tables C.4 and C.5, corresponding to Table 4.1 and 4.2 in Chapter 4, report descriptive

statistics for variables in the year of 1980 and transition rates across import status together

with exit rates from the Extended Sample. The descriptive statistics as well as the transition

probabilities across import status from the Extended Sample are similar to those from the

Basic Sample.

Table C.6 shows the frequency of export/import status change over the sample period

of 1990-1996 among continuosly operating plants. More than 80 percent of plants did not

change export/import status throughout the sample period.

156

6 8 10 125

1015

05

101979−81

kt

xt

ωt

5 10 155

1015

51015

1979−81

kt

xt

ωt

6 8 10 125

1015

05

101982−83

kt

xt

ωt

8 10 12 145

1015

01020

1982−83

kt

xt

ωt

6 8 10 125

1015

05

101984−86

kt

xt

ωt

8 10 12 145

1015

51015

1984−86

kt

xt

ωt

6 8 10 128

1012

51015

1987−89

kt

xt

ωt

8 10 12 145

1015

01020

1987−89

kt

xt

ωt

6 8 10 128

1012

01020

1990−92

kt

xt

ωt

8 10 12 145

1015

01020

1990−92

kt

xt

ωt

6 8 10 125

1015

01020

1993−96

kt

xt

ωt

8 10 12 1410

150

1020

1993−96

kt

xt

ωt

Non−Importers Importers

Figure C.1: Monotonicity Condition157

Table C.4: Descriptive Statistics in 1980 (Extended Sample)Interme- Import Output/ No. of

Output Capital Labor Energy diates Ratios Workers PlantsAll 98.33 40.76 54.33 0.64 50.65 0.07 1.18 4502

Plants (468.41) (233.12) (127.34) (7.65) (235.82) (0.18) (1.93)

Importing 442.28 180.44 168.44 2.91 203.11 0.37 2.58 308Plants 1003.26) (430.15) (231.34) (14.46) (404.55) (0.26) (3.69)

Non-Importing 22.03 10.91 25.93 0.08 13.02 0.00 0.75 2626Plants (57.58) (74.60) (29.39) (0.57) (32.36) (0.00) (0.84)

Switchers 158.55 63.32 79.48 1.13 83.72 0.13 1.62 1568(625.17) (323.55) (173.49) (11.18) (343.34) (0.22) (2.43)

Survivors 201.20 77.06 84.34 1.46 99.60 0.11 1.65 1460(784.12) (377.72) (192.22) (12.87) (388.51) (0.22) (2.64)

Quitters 48.95 23.34 39.93 0.25 27.16 0.05 0.95 3042(149.14) (105.10) (75.05) (2.59) (90.49) (0.15) (1.41)

Notes: Standard errors are in parentheses. The statistics are based on the Extended Sample, where a missing initial capital stock is

imputed by a projected initial capital stock based on other reported plant observables. “Importing Plants" are plants that continuously

imported foreign intermediates in the sample. “Non-Importing Plants" are plants that never imported foreign intermediates in the sample.

“Switchers" are plants that switched their import status in the sample. “Survivors" are plants that did not exit during the sample period

(1980-1996) while “Quitters" exit during the sample period. “Output," “Capital," “Energy," and “Intermediates" are measured in millions

of 1980 pesos. “Labor" is the number of workers. “Import Ratios" are the ratios of imported intermediate materials to total intermediate

materials.

For brevity, 90 % bootstrap confidence intervals are omitted in Figure 4.2 in Chapter

4. As a example, Figure C.1 plots the productivity dynamics before and after plants start

importing together with 90 % bootstrap confidence interval for the 50 percentile importing

plants with 22.2 percent import shares (nt = 0.251). Note that the solid line in Figure C.1

of this appendix corresponds to the dotted line in Figure 4.2 of Chapter 4.

158

−5 0 5 10 15 20 25

−0.02

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

Year

90 % bootstrap confidence interval

ln(TFP)t =β

n n

t + ω

tω

t = γ n

t−1 + ρ ω

t−1

dt = n

t =0 d

t = 1, n

t>0

Figure C.2: Productivity Dynamics for the 50 Percentile Importing Plants

159

Table C.5: Transition Probability of Import Status and Exit (Extended Sample)Year t status No Imports Imports

Year t + 1 status No Imports Imports Exit No Imports Imports Exit1981-1985 ave. 0.832 0.052 0.116 0.169 0.785 0.0461986-1990 ave. 0.877 0.052 0.071 0.176 0.801 0.0231991-1995 ave. 0.866 0.064 0.070 0.126 0.852 0.0231981-1995 ave. 0.858 0.056 0.086 0.157 0.813 0.031

Notes: The statistics are based on the Extended Sample, where a missing initial capital stock is imputed by a projected initial capital

stock based on other reported plant observables.

Table C.6: Export and Import Status Change

Exports ImportsStatus Changes No. % No. %0 1837 0.835 1765 0.8021 207 0.094 190 0.0862 109 0.050 163 0.0743 33 0.015 55 0.0254+ 15 0.007 28 0.013No. of Plants 2,201

Notes: Based on the sample of plants that continuously operated over the 1990-1996 period.

160