BEHIND THE LEARNING CURVE: LINKINGLEARNING ACTIVITIES TO WASTE REDUCTION

byM. A. LAPRE*

A. S. MUKHERJEE**

andL.N. VAN WASSENHOVE t

98/66/TM (CIMSO 5)(Revised version of 96/24/TM)

* Professor at Boston University, School of Management, 595 Commonwealth Avenue, Boston, MA02215, USA.

** Arthur D. Little, Inc., Cambridge, MA.

t Henry Ford Chaired Professor of Manufacturing and Professor of Operations Management at INSEAD,Boulevard de Constance, 77305 Fontainebleau Cedex, France.

A working paper in the INSEAD Working Paper Series is intended as a means whereby a faculty researcher'sthoughts and fmdings may be communicated to interested readers. The paper should be considered preliminaryin nature and may require revision.

Printed at INSEAD, Fontainebleau, France.

Behind the Learning Curve: Linking LearningActivities to Waste Reduction

Michael A. Lapre • Amit Shankar Mukherjee • Luk N. Van Wassenhove

Boston University, School of Management, 595 Commonwealth Avenue, Boston, MA 02215

Arthur D. Little, Inc., Cambridge, MA

INSEAD, Boulevard de Constance, 77305 Fontainebleau Cedex, France

August 11, 1998

Abstract

This exploratory research on a decade of Total Quality Management in one factory opens

up the black box of the learning curve. Based on the organizational learning literature we

derive a quality learning curve that links different types of learning in quality improvement

projects to the evolution of the factory's waste rate. Quality improvement projects that

acquired both know-why and know-how accelerated waste reduction, while other projects

either impeded or did not affect waste reduction. In complex, dynamic production environ-

ments locally acquired knowledge is difficult to disseminate. The combination of know-why

and know-how facilitates its dissemination.

(Learning Curve, Organizational Learning, Quality, TQM, Technological Knowledge, Exper-

imentation, Knowledge Transfer)

1 Introduction

Knowledge has become a critical resource for competition. Firms have to manage organiza-

tional learning efforts directed at building knowledge they can use in the future (Adler 1989).

One approach which is particularly learning-oriented is to embark on a Total Quality Man-

agement (TQM) program. The link between learning and quality, however, is ill-understood.

Little is known about the learning processes in TQM programs and their impact on bottom-

line organizational performance (Hackman & Wageman 1995).

1

In a previous paper (Mukherjee et al. in press) we started to explore this link between

learning and quality. We analyzed 62 quality improvement projects undertaken in one factory

over a decade. We identified two dimensions of the learning process that took place in these

projects: conceptual and operational learning. Conceptual learning consists of assessing

cause-and-effect relationships and designing concepts in response, i.e. the acquisition of know-

why. Operational learning consists of implementing a concept and observing the results, i.e.

the acquisition of know-how. We found that conceptual and operational learning played a

crucial role in changing factory personnel's attention as a result of the project. For example,

projects which acquired know-why or know-how were more likely to yield new Standard

Operating Procedures or changes in Statistical Process Control.

This paper extends the link between learning and quality from a cross-sectional, project

level analysis to a longitudinal, factory level analysis. In it, we open up the black box of

the learning curve, explicitly introducing the knowledge acquired by quality improvement

projects. Given the role of conceptual and operational learning in changing factory person-

nel's attention after project completion, we use these two learning dimensions to identify

four different types of knowledge acquisition: firefighting, artisan skills, unvalidated theories

and operationally validated theories. We construct the cumulative number of projects of

each type. We then use these cumulative number of project variables to explain changes

in the factory's waste rate (measured by the ratio of wasted material to total material

released to the process). We show that projects which acquired both know-why and know-

how—operationally validated theories—accelerated waste reduction. Other projects either

impeded or did not affect waste reduction.

We explain our findings in the context of complex, dynamic production environments.

Learning in such environments is often impeded by dynamic complexity, ambiguity and

erroneous myths (Sterman 1994, March & Olsen 1975). If the production process is not

completely understood, dissemination of locally acquired knowledge is problematic. We

argue that the combination of know-why and know-how facilitates its dissemination.

2

The paper is organized as follows. In section 2, we review drawbacks of the traditional

learning curve and ideas from the organizational learning literature that will help us un-

derstand the learning process behind the learning curve. In section 3, we derive a learning

curve model for waste reduction that addresses the drawbacks of the traditional learning

curve mentioned above. Section 4 describes the research site, its quality improvement ef-

forts, our previous study of these efforts and the data collection. Section 5 describes the

econometric results. Sections 6 and 7 discuss implications for managers and scholars.

2 Behind the Learning Curve

The learning curve phenomenon has been observed frequently. Firms realize large cost

reductions as they gain experience in production (see reviews by Yelle 1979, Dutton &

Thomas 1984). The functional form which has traditionally been suggested for the learning

curve is the power form

—6,C = Coq , (1)

where c is the cost to produce the q-th unit; co the cost to produce the first unit; and b the

learning rate.

Despite its frequent use, scholars have established fundamental shortcomings of the power

form (1). It is an entirely empirical phenomenon (Levy 1965), and as Lieberman (1984) notes

the appropriate functional form of the learning curve has never been rigorously tested. This

is peculiar in light of the following observations that do not support (1).

• The traditional power form does not accomodate two often observed patterns: initial

downward concavity and the plateau effect: after some amount of production no further

improvements are made (Muth 1986).

• Estimated learning rates vary widely, within an industry, within a plant, and over

time (Levy 1965, Dutton & Thomas 1984, Hax & Majluf 1982). Therefore, Dutton

& Thomas (1984) have advocated that the learning rate be treated as a dependent

3

variable as opposed to a given constant. In dynamic production environments with in-

complete technological knowledge, factory personnel should deliberately try to enhance

improvement rates (Jaikumar & Bohn 1992).

• Investments have resulted in shifts to steeper learning curves, Hax & Majluf (1982)

observed. Mishina's (1992) study of Boeing's production of B-17 heavy bombers cor-

roborates the importance of investments. His findings indicate that the scale-up of

production triggers learning. His rationale is that learning occurs only if there is a

challenge. Scale-up of production provides such challenge. Mishina used proven effec-

tive capacity to measure learning by new experiences. Findings by Epple et al. (1996)

and von Hippel & Tyre (1995) are consistent with Mishina's results.

Probably the most important reason why the traditional power form does not accomodate

these observations is that it lacks an underlying theory. It does not provide any insight

into the learning process behind the results. It assumes that cumulative volume is the

only source of learning and it ignores deliberately undertaken learning efforts (Bohn 1994).

Dutton & Thomas (1984) made this distinction between autonomous and induced learning.

Induced learning requires efforts or resources that are not present in the current operating

situation. These efforts are deliberately undertaken to create better knowledge. Autonomous

learning, on the other hand, does not involve deliberately undertaken learning efforts; it is

a much less cognitive process that occurs automatically "on-the-job". Levy (1965) and

Adler & Clark (1991) are the only two papers we are aware of that include learning process

variables, to measure induced learning, in a learning curve analysis. Yet, as the latter

authors acknowledge, variables like training hours and engineering activity are still proxies

of the actual induced learning process.

Levy (1965) assumed that for a new process, a firm has a maximum rate of output P it

would like to achieve. The rate of output after q units have been produced is Q(q) < P. His

crucial assumption is that the rate of improvement in Q(q) is proportional to the amount

4

the process can improve. From this assumption follows the adaptation curve:

Q(q) = P[1 — e-(a+Pq)],

where a represents the initial efficiency of the process, and it the process's rate of adaptation,

which can be affected by various factors y i , ... , yri like prior training and experience:

n

li = 00 + /3i i (2)i.i

Equation (2) is used to explain differences in learning rates across workers.

The adaptation curve overcomes many of the problems associated with the power form

(1). However, it does not address the dynamic nature of the learning rate and the impact

of changes in the production system. Furthermore, it raises new questions: (i) how should

the target level P be determined objectively, and (ii) what theory underlies the assumption

that the rate of improvement is proportional to the amount the process can improve?

One cannot obtain a value for P from estimation nor from hard data. Levy inferred

values for P from interviews. However, even the most ambitious values inferred for P can

be overtaken, thereby violating the assumption Q(q) < P. At Chaparral Steel, for exam-

ple, management sets targets for production rates "considerably beyond current production

capabilities." The company not only achieves these very ambitious goals, but continues to

improve performance unabated (Leonard-Barton 1992). Likewise, Boeing more than quadru-

pled the target it set initially at the planning stage (Mishina 1992).

Levy's assumption that the rate of improvement is proportional to the amount the pro-

cess can improve can be based on what the organizational learning literature refers to as

"performance gaps". A performance gap is the discrepancy between actual performance

and aspired performance (managerial targets). This performance gap induces organizational

members to search for alternative actions that might reduce the performance gap. March &

Simon (1958) postulated that a larger performance gap spurs organizations to intensify their

search efforts. Cangelosi & Dill (1965) confirmed that performance gaps trigger learning.

Yet, "learning is sporadic and stepwise rather than continuous and gradual." (p.175) Duncan

5

& Weiss (1979, p.52) noted that if the performance gap cannot be attributed to factors out-

side the locus of control or to improper implementation, "... it must be considered a failure

of organizational knowledge". Consequently, faced with a performance gap the organization

should obtain better knowledge about action-outcome relationships.

Adler & Clark (1991) found that two types of induced learning, engineering activity and

training activity, can be enhancing as well as disruptive. Their field research suggested some

explanations for their results, but they acknowledge that further research on identifying and

understanding induced learning variables is necessary. In prior work, we used research on

organizational learning to identify dimensions of the induced learning process. In this paper,

we use these dimensions to classify different types of induced learning.

Although there exist many different definitions of organizational learning, Fiol & Lyles

(1985) note that scholars typically distinguish two dimensions of the learning process: cog-

nitive development and behavioral development. They define organizational learing as the

process of improving actions through better knowledge and understanding. This process is

essentially a feedback process: "Decision makers compare ... information about the state

of the real world to goals, perceive [performance gaps], and take actions (they believe will)

cause the real world to move toward the desired state" (Sterman 1994, p.293) They revise

subsequent actions based on feedback received. Sterman identifies several barriers to learn-

ing that disrupt this feedback process in complex, dynamic environments. These include

dynamic complexity (separation of cause and effect in time and space), ambiguity (simulta-

neous existence of equally plausible but mutually contradictory explanations of a situation),

misperception of feedback, and poor inquiry and scientific reasoning skills. These barriers to

learning make it hard for people to create better knowledge about action-outcome relation-

ships. March & Olsen (1975), for example, discuss how ambiguity leads individuals to create

their own potentially inaccurate beliefs called "myths" (potentially inaccurate beliefs based

on subjective interpretations of data), and stick to these myths. They are skeptical of lo-

cally acquired knowledge. Dissemination of knowledge therefore becomes problematic. This

6

is further aggravated if all relevant knowledge is not fully understood and codified (Teece

Pisano 1994). In the next section, we enhance Levy's model using the organizational

learning ideas outlined above.

3 A Quality Learning Curve

In this section, we derive a learning curve model for quality improvement in a dynamic

production environment. We focus on waste, a key driver of both quality and productivity.

We make four modifications to Levy's (1965) model: a theoretical foundation; Mishina's

experience variable; a natural, objective aspiration level; and a dynamic learning rate.

Let xt be the production output in month t, zt = max,-< t XT the proven effective capacity

(Mishina's experience variable), W(z) the waste rate after z has been proven to be feasible

production output, and P the desired waste rate. W(z) — P is the performance gap. In our

context Levy's assumption becomes

dW(z) = /1 [W( z ) p],

dz

where y denotes the learning rate. However, instead of merely assuming (3), we can interpret

this equation in light of the organizational learning literature on performance gaps. The

performance gap W (z) — P induces the organization to search for alternatives to reduce

this gap. A larger discrepancy spurs the organization to exert more effort in searching for

better knowledge (e.g. March & Simon 1958). The effectiveness of acquiring new knowledge

is determined by the learning rate µ. Consequently, we can model the rate of improvement

as the product of the learning rate and the performance gap.

Mishina's findings suggest that scale-up of production triggers learning. In a dynamic

production environment scale-up of production can be achieved by adding new machines or

increasing machine speeds. In both cases factory personnel need to acquire new technological

knowledge on how to control their process in the changed production environment. We

therefore use Mishina's experience variable.

(3)

The TQM literature provides a natural choice for P: zero defects (Deming 1982). As our

research site aimed for zero defects, we will—without loss of generality—employ P = 0 in

the remainder of the paper (cf. Schneiderman 1986, 1988). Contrary to Levy's model, the

TQM context of waste reduction provides a value for P which is not subjective and which

can never be overtaken by actual performance.

Our final modification to Levy's model concerns the learning rate Following Jaikumar

& Bohn (1992), we assert that the learning rate is fundamentally dynamic in nature. Factory

personnel in a dynamic production environment have to create new knowledge unabated to

adapt to new situations, and deliberately try to enhance rates of improvement.

Formally, let t be the time index. Let y it , , ynt be managerial factors that affect

the learning rate, like the cumulative number of quality improvement projects. The most

parsimonious formulation for the learning rate is

n

At = 00 +

(4)

i=1

Solving (3) with condition W(z) > P gives (recall P = 0)

W(z) = ea+Az, (5)

Substituting (4) in (5) and taking logarithms, we obtain

In W(zt) = a + (60 + E (3iYit)Zt• (6)i=1

Equation (6) can be estimated if we have data for the waste rate, production, and the

managerial factors yit . Estimating (6) is essentially an investigation of Dutton & Thomas's

(1984) autonomous/induced learning dimension (cf. Adler & Clark 1991): 0 0 measures the

autonomous part of the learning rate, whereas E7 oiyit measures the induced part.

Note that we do not restrict the range of p t , because the induced learning variables can be

enhancing as well as disruptive (Adler & Clark 1991). Moreover, Dutton & Thomas (1984)

observed both positive and negative learning rates. Equation (4) assumes that projects

conducted long ago do not lose their impact. In this paper we are not looking at forgetting.

8

We are, for now, merely interested in identifying factors that affect the rate of learning.

Future research could look at forgetting of both autonomous and induced learning.

Equation (5) is equivalent to Schneiderman's (1988) half-life curve specification if z is

replaced by t and it is kept constant. The assumption underlying the half-life curve is that

waste "decreases at a constant rate, so that when plotted on semi-log paper against time, it

falls on a straight line." (p.53) Not only do our field observations contradict this assumption,

the disappointing managerial implication is that the rate of waste reduction, i.e. the learning

rate, is given and cannot be accelerated. This is precisely what (4) allows for. Moreover, in

addition to the reasons for using z instead of t advanced above, z incorporates the Cangelosi

& Dill (1965) notion that learning is a stepwise process as opposed to a continuous process.

Finally, Scheiderman actually observes that depending on the scope of quality improvement

projects learning rates fall in different ranges. Our specification (6) allows for the learning

rate to change if quality improvement projects change over time. We now turn to the context

in which we estimate equation (6).

4 Research Design and Data

This methodology section discusses the research site, its quality improvement projects, mea-

surement of the learning constructs in the projects, and the longitudinal data. The first two

subsections are largely based on Mukherjee et al. (in press).

4.1 The Research Site

Several considerations prevailed in choosing a research site. First, the site had to provide

access to detailed data about the systems used to improve quality. Second, it had to have

incomplete knowledge so as to faciliate a study on knowledge acquisition. Third, for our

findings to be generalizable, the site had to have implemented standard TQM methods.

Fourth, it had to have an established record of successful quality improvement.

These considerations convinced us to choose N.V. Bekaert, S.A., a Belgian multinational

9

corporation. Bekaert is the world's largest independent producer of steel wire. In particular,

its Steel Cord Division, which hosted our research, produces about one-third of the world's

output of the steel wire (called "tire cord") used in the production of steel belted radial tires.

Our project received the enthusiastic backing of then CEO Karel Vinck and current CEO

Rafael Decaluwe, providing us with unlimited access to people and documents.

Bekaert's basic process flow is deceptively simple: Thick wire are pulled ("drawn")

through dies which progressively reduce their diameter. Very thin wire ("filaments") are

wrapped around each other to form tire cord. The simplest cord has two filaments; the most

complex, hundreds. See figure 1.

Figure 1 about here

Of course, reality is much more intricate. First, each plant has a handful of huge draw-

ing machines upstream and hundreds of small drawing and filament wrapping machines

downstream. Second, swimming pool sized soap pits supply lubricants (for the dies) to the

drawing machines downstream. Thus, while these machines are run independently, the soap

system links them to each other. Third, the wire is heat treated at two intermediate points

to make it ductile. At one of these points, a chemical process also coats the wires with brass.

Despite the use of sophisticated controls, wires which are heat treated and coated together

do not necessarily have identical properties. Fourth, Bekaert's suppliers, which include some

of the best known steel companies in Europe and Japan, cannot guarantee homogeneity of

properties across the thousands of tons of wire they deliver. Microscopic flaws in the wire

can cause fractures (the industry's biggest quality problem) at any process stage.

As a result, plant personnel have to contend with very high levels of dynamic complex-

ity and ambiguity (Sterman 1994). The former arises from the ease with which effects of

problems experienced at any machine or production stage could be transmitted to other

machines and stages. The latter arises as production personnel often disagree on causes

for a particular production problem like fractures. One person could blame raw materials,

another intermediate process settings, and there would be no way to tell who is right.

10

During the 1980s, Bekaert's customers—tire manufacturers—experienced traumatic

change, and responded by making simultaneous demands of high quality, short lead times,

low cost and product line flexibility. Bekaert responded by initiating a pilot quality improve-

ment program at its flagship Aalter plant (Belgium) in 1981. At that time, little scientific

knowledge existed concerning the production process of tire cord. However, over the next ten

years, it introduced and institutionalized among others a structured approach to problem

solving, a functional TQM organization, Statistical Process Control, Standard Operating

Procedures, TQM project teams, information systems providing standardized daily, weekly

and monthly production and quality data, process capability measures, and quality control

circles. Researchers on quality have long prescribed these methods (see e.g. Deming 1982,

Juran & Gryna 1993, Hackman & Wageman 1995).

In the late 1980s, Bekaert realized that central R&D laboratories lacked the charac-

teristics of the dynamic production environment encountered in the factory. It therefore

re-located process optimization to the factory. In 1988, Bekaert established a "Model Line"

at Aalter (MLA) for an important, representative product, and gave a senior R&D manager

the responsibility to create fundamental process control knowledge without sacrificing the

production of saleable wire.

Its diligent efforts seem to have borne fruit, for in 1990, CEO Karel Vinck won the first

European Foundation for Quality Management Leadership Award. Many people believed

that the jury was favorably impressed by his vision for quality exemplified by the innovative

practices at Aalter. Bekaert's second major recognition for quality came in 1992, when its

Burgos plant (Spain) won a European Quality Prize.

In sum, Bekaert provided access, was a dynamic production environment with incomplete

technological knowledge, used common TQM methods, and achieved considerable success.

4.2 The Learning Process in Quality Improvement Projects

We had unlimited access to all of Aalter's records on improvement projects undertaken

throughout the 1980s. From these, we selected projects undertaken between 1982 and 1991

that (i) sought to improve product attributes or process control (as opposed to say, house-

keeping), (ii) had progressed (at least) past the testing stage, and (iii) had been adequately

documented. This selection left us with 62 projects.

We coded the projects on questions which dealt with their learning process and with their

performance. We structured these questions on the basis of the organizational learning lit-

erature and our experience with earlier exploratory research (Mukherjee & Van Wassenhove

1997). Questions were coded by giving a number or a response on a 5 point Likert scale.

Appendix A contains the questions relevant for the current paper. For each question, we in-

dicated how to determine the number or we gave typical examples for the extreme answers (1

or 5). To further improve reliability, we often developed typical answers for the intermediate

response categories as well. (See e.g. the question on the use of experiments in appendix A.)

These typical answers allowed one of us to objectively code the responses. For 55 projects,

he based his assessments primarily on the final project reports, which under a management

edict, contained fairly detailed information in a standardized format. Additionally, he re-

lied on any available interim reports. Finally, he resolved any ambiguity by discussing the

affected project with the plant's TQM coordinator, who had not only participated in all of

these projects, but had maintained his own notes about them. Data on 7 additional projects

came from a prior research effort (Mukherjee 1992).

We conducted a factor analysis on the questions which dealt with the learning process.

Two of the resulting factors are crucial for this paper; they mapped onto what Kim (1993)

calls conceptual learning and operational learning. The organizational learning literature has

proposed many different definitions for what Fiol & Lyles (1985) call cognitive development

and behavioral development. Figure 2 shows why we chose Kim's (1993) terminology. He

12

equates conceptual learning with assessing observed experiences and designing concepts in

response. The questions that loaded onto conceptual learning measured (i) the use of science

and statistical experiments to assess cause-and-effect relationships, and (ii) the level of detail

used in the designed solution. Kim equates operational learning with the implementation of

a concept and observation of the results. The questions that loaded onto operational learning

measured the modification of action variables and the follow-up of experimental results.

Figure 2 about here

While figure 2 helps to assess the content validity of conceptual and operational learning,

OLS regressions in Mukherjee et al. (in press) help to assess the criterion-related (predictive)

validity. We found that operational learning was a good predictor for achieving goals in TQM

projects, whereas conceptual learning was the key predictor of acquiring better technological

knowledge (measured on a scale that addressed Bohn's (1994) stages of knowledge).

Finally, we found that conceptual and operational learning during the project changed

organizational behavior measured by modifications of Standard Operating Procedures and

Statistical Process Control rules after the project. It is important to note, though, that in

our cross-sectional study (Mukherjee et al. in press), both learning and changed behavior

were measured in the limited boundaries of the project context (i.e., one production line,

one machine or an identified group of machines). Typically, a project did not even affect one

hundredth of the total plant waste.

Discussions with managers and engineers revealed that some of these local projects had

produced knowledge applicable to the entire plant. Some project insights were actually

transferred to other plants as well. Dissemination of local project knowledge, however,

was far from routine practice. The projects were scattered over time and space in the

factory. Often, project results did not even spread in the vicinity of the project context;

personnel in other parts of the factory would argue that a particular project finding would

not be transferable because of slight differences in machine types, product specifications, raw

materials, etc. Several managers and engineers commented on this notorious reluctance to

13

consider knowledge obtained elsewhere in the factory. This was particularly aggravated by

the lack of a scientific knowledge base of wire drawing. So, the Not Invented Here syndrome

Hayes & Clark (1985) observed between plants also exists within a plant.

Consequently, in this paper we study the impact of local project learning on global waste

reduction. Given the role of conceptual and operational learning in changing factory person-

nel's behavior locally, we want to investigate the impact of these two learning dimensions on

the plant's global learning curve. We now describe our longitudinal data to do so.

4.3 Longitudinal Data

In January 1984, the Aalter plant introduced a reporting system that provided standardized

waste and production data on a monthly basis.1

• Waste rate. For all major process stages, the plant reported the percentage of steel

wire scrapped because of irreparable defects. Let Wit denote this waste rate for process

stage i in month t. The yield rate for stage i is then 1 — Wit . Hence, the yield

rate for the entire factory is n ( 1 - Wit ) . (Cf. the probability that one ton of wire

passes flawlessly through all process stages.) Finally, we obtain the factory's waste

rate Wt 1 - - Wit ) . Figure 3 shows the waste evolution. We correct In

for seasonal patterns by subtracting the sample average for that month. Thus we

control for effects like increased waste levels due to start-ups in production months

that followed the holidays.

• Production. The factory's production volume x t was readily available in the re-

ports: the tonnage of wire produced in the last stage. The proven effective capacity is

constructed straightforwardly as zt = max,<tSee figure 4.

• Projects. For 50 projects the project reports provided the project completion dates.

The completion dates for the remaining 7 projects were recorded in Mukherjee (1992).

lAalter produces three product types. The waste and production data concern its most important product

type, tire cord. Five projects did not concern tire cord. Hence, we only retained the 57 projects that did.

14

For each project we had factor scores' for conceptual and operational learning at our

disposal from our previous study. A simple way of measuring knowledge accumulation

by projects would be to construct one variable Ct , the cumulative number of projects

completed up to month t. However, to investigate whether different ways of learning

had a differential impact, we compare the factor scores for conceptual and operational

learning with the median values in the sample of 57 projects, and classify each project

according to high (H C ) or low (LC ) conceptual learning and high (H°) or low (L°)

operational learning. This allows us to compute CL'14, the cumulative number of

projects with low conceptual learning and low operational learning completed up to

month t. Similarly, we compute CL'Il t̀', le14, and Cli cT. Thus, the projects split

up 14-14-14-15.

Figures 3 and 4 about here

Our sample includes 80 production months: January 1984 to March 1991 (there are 11

production months per year). By studying this time horizon, we control for learning through

both internal and external benchmarking as well as changes in technology. First, until 1991

Aalter pioneered Bekaert's TQM efforts, so it did not transfer knowledge from other Bekaert

plants. Second, the industry is very secretive, making it difficult to learn from competitors.

Moreover, during the 1980s Bekaert remained the leading, largest independent tire cord

producer. Third, the technology depicted in figure 1 remained stable during the 1980s.

It is important to point out the timing of events measured for each project. The items that

loaded onto conceptual learing measured assessments of cause-and-effect and detail in the

2 A factor analysis reduces a set of variables to a smaller set of new variables called factors. This set

accounts for the larger part of the variation in the original set. In standard factor analysis (principal

components followed by varimax rotation) the newly constructed factors are uncorrelated with one another.

Typically, each original variable is highly correlated with one factor, and relatively uncorrelated with the

other factors. For each observation factor scores can be constructed. A factor score for an observation on a

particular factor is a weighted average of the standardized responses to the original questions. The weights

are determined by the correlations between the original questions and the specific factor.

15

proposed design prior to final testing. The items that loaded onto operational learning dealt

with the final test results of the project. Finally, the date of project completion occurred

after the final test results of the project. The date of project completion denotes the time

when a project gets added to a cumulative number of projects variable, which subsequently

is linked to the waste evolution. This sequence of events is summarized in figure 5. The

longitudinal nature of the data will allow us to make inferences about the impact of the

learning constructs on waste reduction. We now turn to the econometric analysis.

Figure 5 about here

5 Quality Learning Curve Estimates

We first estimate the traditional learning curve with the various experience variables pro-

posed in the literature, time (t), cumulative volume (q), and proven effective capacity (z):

lnWt = a+blnt (7)

lnWt = a+blnqt (8)

ln Wt = a + b ln zt (9)

Next, we estimate the quality learning curve derived in section 3 with a learning rate that

is (i) constant, (ii) affected by projects assumed to be homogenous, and (iii) a function of

projects that differ in their learning processes:

In Wt = a + Oort (10)

In Wt = a + Po + OiCt)zt (11)

In Wt = a + (fio + Oi CLcift) + 02 CLcirt + )33CHcL7 + 04CHclinzt (12)

In each model, we allow for first order serial autocorrelation. The traditional learning curve

estimates confirm Mishina's findings (rows 1 to 3 in table 1). Maximum proven capacity

explains more variation. This particularly makes sense in Bekaert's production environment

16

where scale-up (increasing machine speeds and/or number of machines) forces factory per-

sonnel to learn how to produce good output in the changed operating conditions. As figure 4

shows, 50% of the scale-up during the 1980s occurred at the end of 1987 and the beginning of

1988. This scale-up coincided with a major reduction in the waste rate (figure 3). Nonethe-

less, for all traditional learning curve models the autocorrelation coefficients are statistically

significant indicating ill-specified models.

Table 1 about here

From rows 4 and 5 in table 1 one might conclude that the quality improvement projects

did not affect the learning rate. Both models show the same degree of autocorrelation and

explain the same amount of variation. The estimates for model (12) in row 6, however,

show that the implicit assumption in model (11) that all projects have the same impact on

waste reduction is erroneous. First, while model (11) suffers from autocorrelation indicating

an ill-specified model, model (12) does not. Second, the amount of variation explained is

significantly larger than either model (10) or (11). The corresponding F-statistics for testing

the 15% increments in R2 are 13.46 and 17.62 respectively. Both are highly statistically

significant. Third, two types of projects significantly affected the learning rate over time.

Consequently, even though model (12) requires estimating more parameters than the tradi-

tional models, it yields the insight that the learning rate is not constant, but dynamic in

nature affected by different ways of learning in TQM projects.3

We will argue that the impact of the different project types on the learning rate relates

to the transferability of project results. Before we do so, we need to rule out an alternative

explanation. Is it possible that the waste reduction resulted directly from local area efforts?

Four CI-IcII° projects were conducted on Aalter's model line spanning parts of all process

3Appendix B displays the table of correlations between the dependent and independent variables, and

discusses why the highly correlated cumulative project variables do not pose a multicollinearity problem.

This appendix also describes additional tests to demonstrate that the estimates in row 6 in table 1 are rather

robust.

17

stages. The other Cli cH° projects (which we defined as ClI cH°*) resembled the CL'Il°

projects in size—only a small part of one process stage. We re-estimated equation (12)

with the modified Cli c11°* variable. Again, Cli cH°* significantly enhanced waste reduction.

So, CI11-1°* and CI, c11° are two sets of local area efforts, which both yielded local results

(high operational learning). Yet, 14 of these combined (CUR') did not affect global waste

reduction, whereas 15 — 4 = 11 did. It seems therefore that the knowledge of the latter

11 projects must have been replicated in other areas of the plant to have had a significant

impact on global waste reduction. This also confirms discussions we had with the plant's

TQM coordinator who facilitated the local projects. His examples of projects that were

implemented elsewhere were all CHell°* projects. This result also rules out project size as

an alternative explanation for the significant estimate for 44 in equation (12). We will now

discuss each coefficient in turn.

• The insignificance of i3 indicates that projects with little conceptual and little opera-

tional learning (CUL') did not affect the learning rate. In these projects teams hardly

reflected on the causes of contingencies, implemented only minor changes, and con-

ducted little follow-up. We label these projects "firefighting." The lack of root cause

analysis and the lack of operational results lead other people to ignore these (local)

efforts.

• Projects with little conceptual learning but substantial operational learning (CLcil°)

generate know-how. Recall that operational learning basically means modifying pro-

cess variables and obtaining follow-up of the experimental results, i.e. these projects

generated successful solutions within the project context. However, the insignificance

of 42 suggests that they did not affect the learning rate. Due to the lack of conceptual

learning, the engendered know-how was not well understood. It was more art than

science. We therefore label these projects "artisan skills."

The lack of conceptual learning leaves a lot of room for ambiguity. Will the operational

learning apply to other areas of the factory as well? As mentioned in section 4.2,

18

personnel in other areas of the factory would argue that local project results would not

be applicable to their part of the factory because of slight differences in machine types,

product specifications, raw materials. As a result, the lack of conceptual understanding

of why something works makes artisan skills hard to transfer. Teece & Pisano (1994,

p.549) noted that "replication and transfer are often impossible absent the transfer

of people, though this can be minimized if investments are made to convert tacit

knowledge to codified knowledge." This was indeed the case at Bekaert; foremen

and supervisors who led these TQM projects were not transferred and the focus on

operational learning implied lack of codified knowledge.

• In projects with substantial conceptual learning and little operational learning (ClIcI,°)

teams assessed cause-and-effect relationships using science and statistics and came up

with a high level of detail in the designed solution. Yet, they failed to obtain opera-

tional results. Hence, we label these efforts "unvalidated theories" (we cannot distin-

guish between theories that were wrong and "correct" theories that were unsuccessfully

implemented). One would expect that the lack of local area operational results would

lead others to ignore these projects like the firefighting projects. However, rather sur-

prisingly, these projects actually had a significantly disruptive impact on the learning

rate (witnessed by the positive sign of g3). Unvalidated theories therefore must have

spread: production personnel elsewhere apparently took detrimental actions based on

conceptual ideas that had not been validated in a full scale manufacturing environment.

Why can the implementation of unvalidated theories actually slow down the learning

curve? Several of these projects used insights obtained at central R&D. Pisano (1997,

p.44) notes that even if "... the technology developed and tested in the laboratory is

replicated nearly perfectly in the plant ... specific elements of the plant environment

(such as equipment configurations) can cause a deterioration in process performance.

This is not a problem of technology transfer, but one of technology development."

Similar problems arise with technology transfers within a plant where areas differ in

19

elements like equipment configurations. At Aalter, instead of first fine tuning R&D

insights locally, plant personnel transferred unvalidated theories to other parts of the

factory. Often, this process was aided by centrally issued directions.

• The negative and significant sign of 444 implies that projects with both substantial

conceptual and operational learning (Cli c1-1°) enhanced the learning rate. In these

projects, teams used scientific models and statistical experiments to develop theories

that explained the occurrence of process interruptions. Based on these theories, the

teams implemented changes and obtained empirical evidence of performance. There-

fore, we label these projects "operationally validated theories."

Conceptual learning guides the team in determining the key variables to modify.

Instead of changing variables by trial-and-error, the team applies scientific principles

to develop falsifiable hypotheses and build models of cause-and-effect. Therefore, the

team is less likely to make erroneous links of cause-and-effect. According to Teece &

Pisano (1994) it is often difficult to replicate organizational processes (here operational

results), because all relevant knowledge is not fully understood and codified. In opera-

tionally validated theories, conceptual learning enhances the build-up of deep process

understanding and codification of the results obtained with operational learning. It is

particularly powerful in isolating the principles from specific contexts. I.e., notwith-

standing differences in machine types, product specifications, raw materials etc., con-

ceptual learning helps to identify the principles that are transferable. Consequently, it

becomes easier to replicate these project results.

So, in order for successful TQM solutions to spread and thus to accelerate the learning

rate both conceptual and operational learning are required. In the next section, we explore

whether plant management can consistently produce both.

20

6 Implications

"In too many TQM programs ... science is fading, the slogans are staying, and the implica-

tions are worrisome." (Hackman & Wageman 1995, p.338) Our findings show the limitations

of solely relying on operational learning. We do not dispute that artisan skills can yield use-

ful solutions for local problems. However, in a production environment characterized by

dynamic complexity and ambiguity such locally acquired know-how does not affect other

people's strongly held beliefs, or myths. It takes conceptual learning to challenge myths.

Our findings suggest that accelerating learning for quality improvement requires both know-

why and know-how. Operationally validated theories are capable of overturning received

wisdom. Furthermore, their codification makes them are easier to disseminate.

The question arises, however, whether the adoption of conceptual learning will generate

desired results. At Aalter some projects with high conceptual learning were able to yield

the corresponding operational results whereas others were not. Can one a priori avoid the

potentially disruptive unvalidated theories? Certainly, with a single learning curve study we

cannot give the ultimate answer. However, we do have some encouraging case evidence.

Five projects in our dataset were conducted on the "Model Line" at Aalter (MLA). The

MLA was a fundamentally new organizational structure at Bekaert. Typically, Bekaert steel

cord factories organized personnel and machines by functional departments corresponding

to the stages in the production process depicted in figure 1. The MLA, on the contrary, was

an integrated line corresponding to a product cutting across all stages. People and machines

were dedicated to a product, and managed by a team led by a senior R&D manager. The

MLA team set goals for productivity and quality improvement based on both R&D and

production experience. It collected data on any product/process variable deemed relevant.

Through its unique information processing system, it was able to link data across process

stages. The MLA team typically solved problems by building scientific models from which

it derived testable hypotheses. These hypotheses were tested with natural and controlled

21

experiments (i.e. conceptual learning). The resulting regression output was then used to

modify limits for Statistical Process Control (based on operational learning). The MLA was

essentially a learning laboratory in the factory (Leonard-Barton 1992).

Four out of five model line projects were operationally validated theories. (The fifth

was an artisan skill directly below the median.) So it seems that Bekaert had created an

organizational structure that facilitated the learning rate enhancing mix of conceptual and

operational learning. One key reason for the strength of the MLA cited by Bekaert's man-

agement was the dedication of resources to the MLA. Pisano (1997, p.46) confirms that

because of the "complex and sometimes chaotic environments" encountered in factories,

"... slight, but unavoidable batch-to-batch differences in raw materials, equipment work-

ers, equipment settings, and other process parameters will reduce the signal-to-noise ratio

of factory experiments." The organizational structure of the MLA, therefore, provided a

controlled environment that enhanced the signal-to-noise ratio of factory experiments. This

leads us to speculate that it is possible to avoid the disruptive unvalidated theories as well

as firefighting projects that do not accelerate the learning rate. By careful project selection

in a factory environment with controlled conditions, a plant can save a lot of resources.

To appreciate the magnitude of the estimates in table 1 one might ask, how much does

one additional project of type i change the initial learning rate 0 0 ? For each type of project

we calculated 132 d30 x 100% to measure this effect. Table 2 shows that one additional oper-

ationally validated theory would improve the learning rate by 3.6%, whereas one additional

unvalidated theory would worsen the learning rate by 3.2%.

Table 2 about here

With 14 unvalidated theories, 15 operationally validated theories, and the estimates in

table 2, one can compute that the learning rate in 1991 was not very different from what it

was in 1984. Do these numbers imply that TQM did not pay off?

First of all, there were times when the learning rate was accelerated because of oper-

ationally validated theories. This resulted in faster waste reduction. So, it is wrong to

22

conclude that TQM did not pay off. The plant, however, was unable to keep these gains in

the learning rate due to unvalidated theories. Clearly, the plant would have been better off

without these. Also, it could have saved the resources spent on firefighting projects. It is

key to carefully select the right projects, and provide the right environment (organizational

structure) to produce the right learning (which is both conceptual and operational).

7 Conclusion

Our contribution to the learning curve literature is the introduction of learning process vari-

ables into the learning curve. Instead of merely linking production and cost improvement,

we use proxies for the learning process to classify different types of knowledge acquisition.

Moreover, we provide empirical evidence for Dutton & Thomas's (1984) assertion: the learn-

ing rate is definitely not constant, in fact, it can be modelled as a dependent variable, and

autonomous and induced learning are important explanatory variables for the learning rate.

This paper also sheds new light on Adler & Clark's (1991) findings. They found that

induced learning can disrupt as well as facilitate the learning process. Our analysis confirms

this. However, we offer a systemic explanation based on the dimensions of the learning

process. Induced learning that yields both know-why and know-how enhances the learning

rate, whereas induced learning that only yields know-why can disrupt the learning process.

In Mukherjee et al. (in press) we suggested future research concerning the reproducibility

of our work and the development of more rigorous definitions of conceptual and opera-

tional learning. Here, having confirmed our project level findings at the factory level, we

re-emphasize the importance of researching the organizational systems for consistently pro-

ducing and supporting conceptual and operational learning. Further study of sophisticated

experimentation in the factory could build on Bohn's work on experimentation (1987).

This paper focuses on the autonomous/induced learning dimension from Dutton &

Thomas's (1984) framework. Future research should include their endogenous/exogenous

dimension. This requires the study of multiple production units. Which types of knowledge

23

are easier to transfer? Argote, Epple and others have studied the transfer of learning by

doing (e.g. Argote et al. 1990), yet the transfer of different types of knowledge has received

little attention in empirical learning curve research.

The current learning curve study focuses on a quality measure. Another important area

for future research would be the development and estimation of learning curve models for

productivity/cost measures with a theoretical footing. We believe that answers to these

questions will enhance firms' efforts to manage and measure their learning curve processes.4

Appendix A: Learning Process Questions

1. Use of statistical experimentation in the project

1 none, 2 moderate use of seven statistical tools, 3 extensive use of seven statistical

tools and/or ad hoc experiments, 4 natural experiments, 5 controlled experiments

2. Use of scientific models or R&D/engineering staff

1 none ... 5 extensive

3. Detailed design

Number of levels in the Ishikawa diagram used in the final solution

4. Degree of modification of action variables or Ishikawa diagram

1 no change ... 5 major change/primary variable dropped or added

5. Follow-up of experimental results, in the testing stage of the project

1 no sufficient follow-up nor improvement in the mean or the variance of the focal

variable ... 5 for sufficient follow-up (large period, or large sample size) a distinct

improvement in the mean and/or the variance of the focal variable

4We gratefully acknowledge the thoughtful comments made by Roger Bohn, Melissa Schilling, Ludo Van

der Heyden, an associate editor and three anonymous referees. We thank the management and employees

of N.V. Bekaert, S.A. for their unstinting cooperation. This research was supported by INSEAD, and (for

part of the field research) the Harvard Business School Division of Research. The work by Michael Lapre

was supported by Arthur D. Little, Inc., N.V. Bekaert, S.A., INSEAD and the Sasakawa Foundation.

24

Appendix B: Correlations and Robustness Tests

In W

In t

In q

hat

—0.779

lnq

—0.796

0.999

lnz

—0.898

0.881

0.899Correlations for the traditional learning curve models

z Cxz CUL° x z CL cir x z ClicL° x z Cli cli° x z

In W

z

Cx z

CUL° x z

CUFF x z

Cli cL° x z

—0.898 —0.848

0.932

—0.871

0.954

0.991

—0.854

0.937

0.997

0.990

—0.770

0.875

0.981

0.954

0.970

—0.859

0.921

0.993

0.981

0.987

0.964Correlations for the quality learning curve models

As can be expected, the cumulative project variables are highly correlated. This, however,

does not pose a multicollinearity problem if there are periods with faster project accumu-

lation of one type associated with accelerated waste reduction, and/or if for another type

faster project accumulation coincides with slower waste reduction. Two project types were

significant in model (12): Cli cli° and Cli c11°. These are graphed in figure 6. Cli cli° grew

faster in mid 87-mid 88 and early 90, two periods of accelerated waste reduction (figure 3).

ClicL° grew faster in mid 88-mid 89, during which waste flattened (figure 3). So even though

the independent variables are highly correlated, this does not prevent different patterns of

project knowledge accumulation from explaining changes in the rate of waste reduction.

Figure 6 about here

We also tested the robustness of row 6 in table 1. The projects were categorized in a 2 x 2

matrix defined by the median values for conceptual and operational learning. We lowered the

25

split up value for conceptual learning until one project jumped from one category to another.

We further lowered the split up value until the next project switched to another category.

This gave us two alternative categorizations. We obtained two more by increasing the split

up value instead of lowering it. Likewise, we obtained four alternative categorizations by

changing the split up value for operational learning. For all eight categorizations, we re-

defined the four cumulative project variables and re-estimated equation (12). The R es varied

between 0.777 and 0.792 (cf. 0.790 in table 1). /3, and 1j2 were never statistically significant

at 0.05. 43 was always positive and significant at 0.05, f34 was always negative and 7 out of

8 times significant at 0.05, only once did the p-value go up to 0.07. From these additional

tests it seems that the estimates in row 6 in table 1 are rather robust.

References

Adler, P.S, "When Knowledge is the Critical Resource, Knowledge Management is the

Critical Task," IEEE Transactions on Engineering Management, 36, 2 (1989), 87-94

Adler, P.S. and K.B. Clark "Behind the Learning Curve: A Sketch of the Learning Process,"

Management Science, 37 (1991), 267-281

Argote, L., S.L. Beckman and D. Epple, "The Persistence and Transfer of Learning in

Industrial Settings," Management Science, 36 (1990), 140-154

Bohn, R.E., Learning by Experimentation in Manufacturing, HBS Working Paper 88-001,

1987

Bohn, R.E., "Measuring and Managing Technological Knowledge," Sloan Management Re-

view, (Fall 1994), 61-73

Cangelosi, V.E. and W.R. Dill, "Organizational Learning: Observations Toward a Theory,"

Administrative Science Quarterly, 10 (1965), 175-203

Deming, W.E., Quality, Productivity, and Competitive Position, MIT Center for Advanced

Engineering Study, Cambridge, 1982

Duncan, R. and A. Weiss, "Organizational Learning: Implications for Organizational De-

26

sign," Research in Organizational Behavior, 1 (1979), 75-123

Dutton, J.M. and A. Thomas, "Treating Progress Functions as a Managerial Opportunity,"

Academy of Management Review, 9 (1984), 235-247

Epple, D., L. Argote and K. Murphy, "An Empirical Investigation of the Micro Struc-

ture of Knowledge Acquisition and Transfer Through Learning by Doing," Operations

Research, 44 (1996), 77-86

Fiol, C.M. and M.A. Lyles, "Organizational Learning," Academy of Management Review,

10 (1985), 803-813

Hackman, J.R. and R. Wageman, "Total Quality Management: Empirical, Conceptual, and

Practical Issues," Administrative Science Quarterly, 40 (1995), 309-342

Hayes, R.H. and K.B. Clark, "Exploring the Sources of Productivity Differences at the

Factory Level," in The Uneasy Alliance, K.B. Clark et al. (Eds.) HBS Press, 1985

Hax, A.C. and N.S. Majluf, "Competitive Cost Dynamics: The Experience Curve," Inter-

faces, 12, 5 (1982), 50-61

Jaikumar, R. and R.E. Bohn, "A Dynamic Approach to Operations Management: An

Alternative to Static Optimization," International Journal of Production Economics,

27 (1992), 265-282

Juran, J.M., and F.M. Gryna, Quality Planning and Analysis, McGraw-Hill, 1993

Kim, D.H., "The Link between Inidividual and Organizational Learning," Sloan Manage-

ment Review, (Fall 1993), 37-50

Leonard-Barton, D., "The Factory as a Learning Laboratory," Sloan Management Review,

(Fall 1992), 23-38

Levy, F.K., "Adaptation in the Production Process," Management Science, 11 (1965), B-

136-B-154

Lieberman, M.B., "The Learning Curve and Pricing in the Chemical Processing Industries,"

RAND Journal of Economics, (Summer 1984), 213-228

March, J.G. and J.P. Olsen, "The Uncertainty of the Past: Organizational Learning under

27

Ambiguity," European Journal of Political Research, 3 (1975), 147-171

March, J.G. and H.A. Simon, Organizations, Wiley, 1958

Mishina, K., "Learning by New Experiences," HBS Working Paper 92-084, 1992

Mukherjee, A.S., The Effective Management of Organizational Learning and Process Con-

trol in the Factory, Unpublished Doctoral Dissertation, HBS, Boston, 1992

Mukherjee, A.S., M.A. Lapre and L.N. Van Wassenhove, "Knowledge Driven Quality Im-

provement," Management Science, (in press)

Mukherjee, A.S. and L.N. Van Wassenhove, "The Impact of Knowledge on Quality," in

The Practice of Quality Management, P.J. Lederer and U.S. Karmarkar (Eds.), Kluwer

Academic Publishers, 1997

Muth, J.F., "Search Theory and the Manufacturing Progress Function," Management Sci-

ence, 32 (1986), 948-962

Pisano, G.P., "The Development Factory: Unlocking the Potential of Process Innovation,"

HBS Press, 1997

Schneiderman, A.M. "Optimum Quality Costs and Zero Defects: Are They Contradictory

Concepts?," Quality Progress, (November 1986), 28-31

Schneiderman, A.M., "Setting Quality Goals," Quality Progress, (April 1988), 51-57

Sterman, J.D., "Learning in and about Complex Systems," System Dynamics Review, 10

(1994), 291-330

Teece, D. and G. Pisano, "The Dynamic Capabilities of Firms: an Introduction," Industrial

and Corporate Change, 3, 3 (1994), 537-556

Von Hippel, E. and M.J. Tyre, "How Learning by Doing is Done: Problem Identification

in Novel Process Equipment," Research Policy, 24 (1995), 1-12

Wadsworth, H.M., K.S. Stephens and A.B. Godfrey, Modern Methods for Quality Control

and Improvement, Wiley, 1986

Yelle, L.E., "The Learning Curve: Historical Review and Comprehensive Survey," Decision

Sciences, 10 (1979), 302-328

28

Traditional Learning Curve Estimates

In t In q In z p R2

(1) -0.18* -0.66* 0.246

(-5.01) (-7.75)

(2) -0.18* -0.65* 0.279

(-5.46) (-7.49)

(3) -1.32* -0.43* 0.631

(-11.5) (-4.14)

Quality Learning Curve Estimates

z C x z CUL° x z CL c 1-1° x z CH cL° xz Clic li° xz p R2

(4) -323' -0.42* 0.636

(-11.6) (-4.05)

(5) -289* 0.27 -0.42* 0.639

(-3.96) (-0.51) (-4.01)

(6) -227* -0.93 -0.65 7.33* -8.23* -0.20 0.790

(-3.15) (-0.19) (-0.13) (2.71) (-2.26) (-1.70)Dependent variable In W. Sample size 80.

Coefficient estimates for independent variables in rows (4)-(6) x 10-6.

T-statistics in parentheses. * signifies significant at 0.05 in a 2-tail test.

Table 1: Learning Curve Estimates

29

Dry Drawing Heat treatment Wet Wire Drawing BunchingBrass coating

$41:111411102Process

0 q q q 0 0 q q q qq q q 0 q q q q q q

—k- q 0 q 0 q -0-00 q 00q q 0 q 0 q q q 0 q0 0 0 0 0 q 0 0 0 0

I 1

I I

unvalidated theories operationally validated

high (CI-I'L') theories (CI-1'11°)

+3.2% * —3.6%*

firefighting artisan skills

low (CUL') (CL'I-1°)

n.s. n.s.

low high

conceptual

learning

operational learning

Table 2: Impact of four types of induced learning: percentage change in the initial learning rate

by adding one project (a negative sign means faster waste reduction). * signifies significant at 0.05

in a 2-tail test. n.s. means not significant.

Factoryfloor

Figure 1: Simplified, schematic process flow of a Bekaert factory. In reality several steps are

repeated. An example of dynamic complexity is incomplete knowledge of the effects of upstream

process settings on downstream quality. Incomplete knowledge combined with heterogeneous inputs

and hundreds of process variables to control often gives rise to ambiguity concerning causes of

production problems.

30

Conceptual

use experiments (.81)

Learning use science (.70)

detail offinal design

(.65)

modify actionvariables (.66)

follow-upof results

(.91)

OperationalLearning

Figure 2: Mapping of factors onto Kim's (1993) conceptual and operational learning (factor

loadings in parentheses)

' _w 1

1984

1985

1986

1987

1988

1989

1990

1991

Figure 3: Waste evolution (W). To protect Bekaert's proprietary data, we do not report the scale

on the vertical axis, nor do we report the estimate of the intercept.

31

E

O

E

E

es

E

O

1984

1985

1986

1987

1988

1989

1990

1991

Figure 4: Production (x) and scale-up (z) history

Time

Conceptual Operational Project Link tolearning learning completion waste

Figure 5: Timeline of events measured for each project

16—

14 m

"2

12 —ca.

``a 10-Lm.ctE 8-c=

6 —

4 —E

2—

0

CHLCHH

1 1984 1985 1986 1987 1988 1989 1990 1991

Figure 6: Evolution of unvalidated theories (CIFL°) and operationally validated theories (C1-19.1°)

32