Working Paper Series

WP 2016-19

School of Economic Sciences

Supplier-induced demand, repeat visits and the

informal healthcare sector in rural north India

Richard A. Iles1†

December 2016

Supplier-induced demand, repeat visits and the informal

healthcare sector in rural north India

Richard A. Iles1†

Working Paper

December 2016

Abstract

The role of unqualified healthcare providers, who operate in the informal healthcare sector, is

an important policy issue as developing economies consider heightened demands for

universal healthcare. A dominant argument within the economics literature is that the

informal sectors decline as economies grow. Empirical literature from north India indicates

that the quality gap between government doctors and those of private unqualified providers is

not large. Recent theoretical work concerning the behaviours of healthcare providers in

developing economies indicates that providing government doctors with market incentives

will increase provider effort and, by extension, the quality gap. However, the thesis of this

paper is that the existence of widespread repeat visits for outpatient fever treatment indicates

that, under current market conditions, the informal healthcare sector is not likely to diminish

as developing economies grow. Empirical results indicate that formal and informal providers

systematically use repeat visits to treat mild - severe fever symptoms and that the health

seeking behaviour of social disadvantaged groups may also drive the practice of repeat visits.

Keywords:

India; utilisation; informal; fever; supplier-induced demand

JEL: D12, I11, I15, O17

1 The author thanks the support given by Saroja Selvanathan in commenting on earlier versions of this work.

The research assistance offered by Waheed Chaudhury is gratefully acknowledged. The logistical assistance

provided by Broadwell Christian Hospital, Harriet Benson Memorial Hospital and Prem Sewa Hospital

contributed to this research. Financial support for this research was provided by Griffith University (Brisbane,

Australia) through the provision of a PhD scholarship.

†School of Economic Sciences & The Paul G. Allen School for Global Animal Health, Washington State

University. richard.iles@wsu.edu

2

1. Introduction

The existence of the informal sector in developing economies is attributed, via standard

development policy analysis, to institutional weaknesses where the cost of regulatory

enforcement is large (Thurnham, 1993). However, the economics literature provides

additional explanations into the on-going existence of the informal sector. These include

consumer preferences of low quality goods (Banerji and Jain, 2007; BÖhme and Thiele,

2012), factor prices (Banerji and Jain, 2007) and different labour endowments (La Porte and

Schleifer, 2014). Irrespective of these different contributing factors, a general consensus

exists that the size of the sector will diminish as an economy develops, due to the

counter-cyclical nature of the sector causes it to shrink as an economy grows.

The informal healthcare sector is a common feature of many developing economies (Ahmed

et al., 2009; Amin et al., 2003; Lindelow and Serneels, 2006; Rao et al., 2011). The informal2

healthcare sector in north India, as well as in other developing economies, may not follow the

standard development economics conclusion, as proposed by La Porte and Schleifer (2014),

that the informal sector will shrink over the long run. Building on the existing literature, we

argue that analysis of repeat visits for a single episode of illness establishes a case for the

existence of supplier-induced demand (SID) within north India’s outpatient healthcare

market. The combined effects of consumer use of prices as a signal for provider quality

across all providers and the widespread use of repeat visits allows a ‘higher’ wage received

by unqualified private providers. This higher wage rate and the reported low diagnostic effort

2 The informal healthcare sector is defined as producers of goods and services that are not State authorised or

registered. This definition reflects that given by Bloom et al. (2008) …’degree to which an activity is recognised

in law or by legally recognised regulatory agencies…’ and the synonymous terms identified by Banerji and Jain

(2007) – secondary, hidden, underground, parallel and twilight.

3

of qualified health providers (Das et al. 2016) is hypothesised to provide continued life to the

informal healthcare sector in north India.

The importance of assumed informational asymmetries within healthcare markets gives rise

to concerns about the principal-agent relationship. Consumers, many of whom have low

levels of education and literacy in low-middle income countries, often lack information

relating to healthcare providers’ qualifications, appropriate prescribing behaviours, and

effective treatment (Das and Das, 2004). The sub-optimal performance of primary healthcare

markets is often attributed to informational asymmetries (Bloom et al. 2008; Hass-Wilson,

2001). This is often evidenced in developing economies by the over prescription of medicines

(Kamat and Nichter, 1998) and a reliance on healthcare information from social-networks

(Behrman et al., 2002), which is likely to be variable in quality.

The thesis of this paper is that the existence of the widespread pattern of repeat visits for

outpatient fever treatment is one indicator that under current market conditions the informal

healthcare sector is not likely to diminish as developing economies grow. Although the

casual factors driving repeat visits are not proven in this paper, the empirical results indicate

that formal and informal providers systematically use repeat visits to treat mild - severe fever

symptoms. Moreover, the behaviour of social disadvantaged groups in delaying treatment and

having a larger number of repeat visits, is associated with bimodal distribution of provider

visits.

This paper is structured as follows: Section 2 outlines the empirical framework, Section 3

presents the data, Section 4 contains the results and Section 5 concludes. Research ethics for

this research was obtain from the Griffith University Human Ethics Committee. Verbal

4

consent was requested and obtained from all interview and survey participants before data

collection.

2 Methods

2.1 Literature

Analysis of the private informal healthcare sector in developing economies has not been

extensively studied3. The growing body of empirical findings of India’s healthcare sector by

Das and colleagues (Das et al. 2016; Das et al. 2014; Das and Hammer 2007a,b) indicates

that the quality-gap between qualified and unqualified outpatient healthcare is not as large as

may be expected. Measuring providers’ adherence to clinical best-practice in diagnosing and

treating a range of conditions, via observational techniques (Standarised Patient), the quality

gap between qualified government Bachelor of Medicine and Bachelor of Surgery (MBBS)

providers and unqualified private sector counterparts, is small (Das et al. 2014). This is

particularly so when treatment measures are used. Further work argues that the sub-optimal

clinical behaviour by qualified private and government doctors is induced by patient

expectations of appropriate treatment (Das et al. 2007b). This sub-optimal clinical practice

helps to explain the relatively modest difference in treatment quality given by government

MBBS providers and unqualified, informal providers. Additional institutional weaknesses in

the government sector heightens the competitiveness of the informal sector (Banerjee et al,

2004).

3 The recent Lancet series (2016) into the private healthcare sector likely reflects the lack of attention

historically given to the topic within population health.

5

More recent theoretical work by Das et al. (2016) implicitly assumes that providers view

consultation and treatment of an illness episode as a single event. They conclude that market

incentives play an important role in increasing healthcare provider effort to diagnose more

optimally, under conditions when provider and patient have diverging utility functions. This

is due to a range of factors common in developing economies that include low levels of:

altruism, intrinsic motivation, training, professionalism, peer pressure and monitoring (Das et

al. 2016). The authors argue that the observable effort of government MBBS providers in

north India, as measured by length of consultation per patient or quality of diagnostic

questioning, relative to their private sector ‘moonlighting’ effort, is small.

The two stage model of consultation and treatment, proposed by Das et al. (2016), does not

consider that provider effort at an initial consultation may incorporate the expectation of

repeated pay-offs through repeat visits for a single episode of illness. Although SID is not

directly incorporated into the model, the authors acknowledge that if the rewards to treatment

dominate a provider’s incentive to exert effort, a provider may not increase effort even under

conditions of market incentives.

Theory concerning the principal-agent relationship between patients and ‘doctors’ offers an

important means of understanding the observed repeated healthcare visits for a single episode

of fever. Notions of supplier-induced demand (SID) have featured widely in the health

economic literature. However, only limited consensus exists regarding its importance and the

validity of empirical testing of its existence (Labelle et al., 1994; Wennberg et al., 1982). In

this research, SID is understood to exist when healthcare providers (i.e. agent) induced

consumer demand that is not in the financial and/or clinical best interests of the consumer

6

(i.e. principal) (Scott and Vicks, 1999). Therefore, the level of congruence between agent

behaviour and consumers’ financial and clinical best interest explains the existence of SID.

Modelling consumer healthcare utilisation based on initial provider choice for i) a single

consultation and ii) multiple visits for multiple illness episodes are standard approaches. With

respect to the single consultation choice modelling, in many health system contexts this

approach is sufficient to capture consumer preferences for healthcare attributes, given their

own socio-economic characteristics and health status. This approach may be deemed

sufficient in the context where healthcare providers diagnose an illness in one or two

consultations, with necessary referrals, and then prescribe a full course of medicines or other

relevant treatment. In the case of many common ailments, where broad-spectrum antibiotics

are suitable, the approximate diagnosis and prescribed treatment is often completed in a

single visit. Thus demand estimation based on a single consultation and initial choice is

generally sufficient.

Multiple visits for multiple episodes of illness is also commonly modelled using discrete

count data. This approach in developed countries is reasonable where healthcare provider

quality is broadly homogeneous (Bago d'Uva, 2005, 2006; Deb and Trivedi, 1997, 2002;

Jiménez-Martín et al., 2004; Santos Silva and Windmeijer, 2001; Vera-Hernández, 1999;

Windmeijer and Silva, 1997). The study by Chang and Trivedi (2003) using Vietnamese data

on ‘self-medication’ represents an important departure from previous studies in its focus on

utilisation4 for medical treatment issued by ‘unqualified’ distributors. Therefore, the

4 The health economics literature widely uses the term ‘demand’ instead of utilisation. However, the term

utilisation is used as the characteristics of alternative goods are not used in estimation.

7

modelling of repeat visits to a healthcare provider during a single illness episode represents a

departure from the existing literature.

2.2 Empirical framework

Count models have widely been used to analyse multiple visits for multiple illness episodes.

These studies have typically emphasized econometric models over fully specified economic

behavioural models (Bago d'Uva, 2005, 2006; Deb and Trivedi, 1997, 2002; Jiménez-Martín

et al., 2004; Santos Silva and Windmeijer, 2001; Vera-Hernández, 1999; Windmeijer and

Silva, 1997). Even though count models are commonly used, Cameron and Trivedi (1986)

argue that ordinal conceptions of doctor visits are also possible. The example given by the

authors, that three healthcare provider visits represents a greater supply of healthcare than

two visits, it does not proceed that the movement from two to three represents 50 per cent

more healthcare supplied, support the use of an ordinal framework. As such, the ordered

probit model proposed by McKelvey and Zavoina (1975) provides an appropriate modelling

framework for healthcare provider visits.

Application of standard cardinal count models to data used in this paper is not appropriate

due to its bimodal feature (see Figure 1). As a result, the ordered probit model is preferred

because it imposes no distributional assumptions over the discrete dependent variable. Also,

stochastic utility maximisation theory of consumer behaviour is maintained. The

development of a zero-inflated extension of the ordered probit model, gives rise to two

plausible models to estimate the determinants of healthcare consumers’ single fever episode

utilisation.

8

2.2.1 Household Utilisation

Life-cycle approaches to healthcare demand theory help to explain the complex interpersonal

dynamics of health capital decisions within households. Viewing intra-household

decision-making in a bargaining framework is common (Basu, 2006; Bolin et al., 2001, 2002;

Manser and Brown, 1980). The respective bargaining power between spouses influences the

outcome of household decision-making. Strategic behaviour between spouses regarding

health capital investment centres on the non-transferable nature of health capital with respect

to issues concerning divorce (Bolin et al., 2001, 2002), and gender discrimination (Klasen,

1994). However, the effect of India’s commonly practiced shared intergenerational household

on bargaining power and household health capital investment decisions is unclear. The

bargaining model’s assumption of a two spouse and one child household is not often

applicable in rural north India (e.g. mean household size of ~seven). In particular, the effects

of elderly family members, interaction between family sub-sets (i.e. mother and

daughter-in-law) and unmarried adult children are unclear in this limited bargaining

framework. Despite limitations to the current bargaining model it does help provide a

framework to better understand the dynamics of household health capital investments.

Moreover, the broader life-cycle human capital approach provides a useful frame to interpret

utilisation results in rural north India.

2.2.2 Estimation specifications

The base ordered probit (OP) model is a latent regression. The dependent variable g* is

unobserved. The latent regression for consumer q takes the form,

(1)

9

where xq is a vector of the explanatory variables, is a vector of unknown coefficients and

q is a standard normally distributed error term. The expected conditional mean of the

residual is zero and the conditional variance of the residual is 1. The observed counterpart of

*qg is gq, which takes discrete values gq = 0,1,…, P. The process through which gq is

observed takes the following form,

gq = 0 if gq 0,

= 1 if 0 < gq 1,

= 2 if 1 < gq 2,

….

= P if gq P-1 ,

where the boundary parameters, p, p = 0,1,2,…,P-1, are free parameters that take the

threshold values -1 = -, 0 = 0 and P = + (Greene, 2012).

The probabilities that enter the Maximum Likelihood estimator’s log-likelihood function are

estimated by,

(2)

where is the cumulative density function (cdf) of the assumed normally distributed

residual.

10

The zero-inflated ordered probit (ZIOP) model proposed by Harris and Zhao (2007)

incorporates a latent binary probit splitting regression that separates participants from

non-participants, into an ordered probit model. Accounting for two separate processes that

explain observed zero count observations, via latent classes, has the “effect of inflating the

number of observed zeros” (Greene, 2012, p.E22-75). Harris and Zhao (2007, p.1076)

explains the zero-inflated component of the model in this way: “the probability of a zero

observation has been inflated as it is a combination of the probability of ‘zero-consumption’

from the OP process plus the probability of ‘non-participation’ from the split probit model”.

The binary splitting regression in the ZIOP takes the form

(3)

The latent variable y* denotes the unobserved measure of fever sickness for each individual

(q). The observed variable y takes the value 1 when y* is greater than the normalised

threshold of zero.

Conditional on y = 1, utilisation of healthcare providers in rural north India for a single

episode of fever is represented by the discrete ordered variable, kq'. Equation (4) provides the

ZIOP probability estimate,

. (4)

11

The ZIOP model may also estimate a correlation coefficient between the residuals from

equation (1) and u from equation (3). The basic ZIOP model assumes zero correlation. When

the residuals are assumed correlated the estimate of the correlation coefficient ( r ) takes a

bivariate normal distribution. When this correlation is incorporated, the probability estimate

of equation (5) takes the form

(5)

where 2 is the cdf of the bivariate normal distribution.

The utilisation decisions of adults are modelled according to several deterministic random

utility model functions. The splitting equation is the binary model of self-reporting a fever

illness and is provided in equation (6). Equation (6) is provided below:

P(Fever12-months=1) = β0 + β1 Age + β2 Dur2+ β3 Dur3+ β4 Dur4 + β5 DistB + u,

(6)

where Age is the represents those over the age of 18 years, Dur2 – 4 dummy variables

accounting for fever duration intervals 4-6 days, 7-9 days and 10 or more days, and DistB is a

binary variable for geographical location.

Two sets of consumers’ repeat visits to a primary healthcare provider for a single episode of

fever demand estimates are carried-out. The first set of demand estimates relate to repeat

12

visits for individual healthcare provider types – unqualified (jhola chhaap5 - JC) provider,

government MBBS (GDr) and private MBBS (PDr). The second set compares estimates for

males and females attending an unqualified provider. Both conform to a standard reduced

form utility function, which is given below:

U =U(X,H,Z), (7)

where X, H and Z are vectors for consumer characteristics, health status and provider

characteristics.

The utilisation functions for unqualified providers, government MBBS providers and private

MBBS providers include log of household income (lnhinc), and the dummy variables female

and muslim. The gender dummy, female, variable is included because both the human capital

and bargaining model predict that females will invest less in health capital, than males (Bolin

et al., 2001; Grossman and Benham 1974; Lancaster et al., 2008). The religion dummy

variable, muslim, is included to test for the difference in behavioural pattern of Muslims

attending government health centres.

The presence of endogeneity biases and measurement error are usually expected when using

recall survey data. Uncontrolled endogeneity effects of several explanatory variables (price,

income and fever duration) in the model are expected to exert a downward bias on the

estimates. It should be noted, that while the presence of measurement error in the dependent

discrete count variable may be a source of estimation bias, the estimation of a Full

5 A colloquial hindi term used to refer to unqualified doctors, which carries negative overtones. May be likened

to the term ‘quack’. The definition of unqualified in this research follows the common linguistic understanding

of the term ‘jhola chhaap’ which is commonly understood across north India.

13

Information Maximum Likelihood sample selection ordered probit model did not provide a

good fit for the data. As a result, this sample selection specification was not employed

further.

3 Data

A sample size of 1174 is used, capturing respondents’ recall of health provider visits within

the previous 12-month. Appendices A and B provide details about the sampling of data and

choice of recall period. The bimodal nature of the distribution of consumer visits to outpatient

healthcare providers is further shown in the histograms in Figure 1. Frequency histograms for

JC, GDr and PDr providers are presented for males and females. The frequency histograms

for males and females, for each healthcare provider, show a consistent bimodal distribution of

visits. The very low number of two visit treatment consultations is evident for males and

females across all healthcare providers. The literature on the pattern of prescribing medicines

over the course of fever treatment for a single episode of fever is limited for India. Based on

preliminary qualitative observations made in preparation of the survey (see Appendix A), it

appeared a common practice among unqualified providers to dispense a small number of pills

at any one consultation. This has the effect of requiring repeat visits to the same healthcare

provider.

In a scenario where a consumer responds once to a provider’s instruction to return for a

repeat dose of medicines, then providers may believe that they can expect the returned patient

to make at least one or two more visits. If it is assumed that private sector providers make

money from the prescription of medicines, then these providers have an incentive to

maximise their revenues by manipulating their combination of prescribed quantity and

quality of medicines over an expected multi-visit treatment response. This scenario provides

14

one means of explaining the bimodal distribution of consumer utilisation of outpatient fever

services.

The top panel in Figure 1 shows that female consumers have a higher frequency of multiple

visits to unqualified providers, compared to males. This higher frequency is most evident at

the levels of three and four visits. The gender difference in the number of healthcare provider

visits are less pronounced among government and private MBBS providers.

The mean number of visits to a healthcare provider for a single episode of fever treatment is

relatively stable across provider types. Discrete observations greater than five are truncated.

This truncation is corrected in the estimates. The number of observations affected are small

for each provider type (see Table B1). Table 1 summarises the mean number of visits (),

which are truncated, by gender and across the four outpatient provider groups – JC, GDr, PDr

and Other (OT). On average, females consult unqualified healthcare providers 2.1 times per

episode of fever with unqualified providers and 1.0 time for Other non-MBBS providers. For

users of unqualified services, the mean 2.1 visits for females is greater than the male rate of

1.8 times. The qualified MBBS doctor rate of use is consistent at between 1.9 and 1.8 times,

irrespective of government or private sector ownership and gender of consumer.

Table 2 presents the mean and categorical variable percentages for the ≤ 12-months sample

used. The mean price for the ≤ 12-months recall group is INR 98.2. The standard error of the

price variable is noticeably lower at INR 253.6. The percentage of consumers travelling

either 0-4 km and 5-9 km were 41.9 per cent and 24.8 per cent. The values for the distance

intervals 10-14 km, 15-19 km and 20+ km for each recall group were 12.1, 4.2 and 2.5 per

15

cent. These descriptive statistics are consistent in comparison with those using a 14-days

recall period (Iles, 2014).

4 Results

Two sets of results are presented. The first is for pooled male and female estimates, by

provider type – Model 1. The parameter estimation from the OP and ZIOP estimators is

provided in Appendix C and account for village clustering in robust standard errors. This

clustering should control for expected village-based correlations between observed outcomes

due to environmental factors affecting fever incidence. These estimates are each carried out

using the software Limdep version 10. The parameter estimates for the OP and ZIOP models

have no direct interpretation. Instead, the marginal effects for each discrete dependent

variable values are presented below.

Table 3 presents estimates of boundary (threshold) parameters and goodness-of-fit measures

for the OP and ZIOP estimates for model 1 for JC, GDr and PDr specific utilisation data. This

presentation follows the work of (Harris and Zhao, 2007). The results, presented in Table 3,

show that within model 1 the ZIOP estimates provide better data specific estimates than those

of the OP. The coefficients for the four boundary parameters are all statistically significant at

the one per cent level for the OP and ZIOP models across the three data sets. However, the

lower log-likelihood, Akaike and Bayesian Informational Criterion favour the ZIOP model

over the standard OP.

16

The penalty imposed on additional model parameters by the Bayesian Information Criterion

(BIC) reduces the comparative advantage of the ZIOP. Using JC data, the BIC measures for

the ZIOP model is 2738.1 compared to the OP BIC measure of 2749.4.

Three test statistics comparing the model specifications of the ZIOP relative to the OP are

also provided in Table 3. The likelihood ratio (LR) test statistic is large and in favour of the

ZIOP specification. The Hausman test is another general specification test comparing

estimators. However, its power is uncertain due to the implicit use of an unspecified

alternative hypothesis. The Hausman test statistics are large for the JC and PDr data

estimates. The test statistic of 6.8 for the GDr data is below the χ2 critical value of 18.3 while

the other two test statistics are well above the critical value. The final test is the Vuong test of

non-nested model comparison. This test is widely used to test the presence of zero-inflation

in count data models (Cameron and Trivedi, 2013). The Vuong test statistic for the three data

sets, JC, GDr and PDr are 4.3, 3.7 and 3.6, respectively, and are all greater than the critical

value at the five per cent level of 1.96 indicating that zero-inflation is present.

The mean conditional marginal effects derived from the ZIOP model estimates are presented

in Table 4. These results reveal that gender, religion and price differences are instrumental in

determining the pattern of utilisation of outpatient fever treatment. The marginal effect of

price on the likelihood of consulting a government and private MBBS provider, at least once,

is consistently negative for government MBBS and positive for private MBBS. This is true at

each value greater than zero, except for two visits. The negative marginal effect of price on

the likelihood of consulting a government MBBS one or more times suggests that consumers

are price sensitive to the need to make informal payments. The positive price marginal effect

coefficient for one or more private MBBS provider visits indicates that for these doctors price

17

acts as a marker of quality. This price signalling relates to both the medicines prescribed and

the perceived quality of the consultation.

Consistent with the results of other chapters in this thesis, income has no marginal effect on

the pattern of outpatient healthcare provider utilisation. The non-significance of marginal

household income on healthcare provider utilisation is evident across each of the three

providers – JC, GDr and PDr. With respect to unqualified providers, this negligible effect of

income on utilisation supports the results of Chang and Trivedi (2003) for Vietnam.

The result of the model 1 ZIOP estimates in Table 4 show that the number of consumer visits

to outpatient healthcare providers indicates that patient behaviour is also important in

determining repeat visits. This is particularly true for the minority groups of females and

muslims. With respect to females, at the values three, four and five healthcare provider

consultations, females, relative to males, are more likely to utilise unqualified provider fever

treatment services. The marginal values indicate that muslim are less likely to seek fever

treatment from government MBBS providers. At the values one, three, four and five visits,

the dummy variable for the marginal effect of muslim identity on government MBBS

provider utilisation is negative and statistically significant at the five per cent level.

Model 2 marginal effects for males and females are separately presents in Table 5. The

district dummy variables (DistB and DistC) are removed and the dummy variables denoting

consultation for a second provider for the same episode of fever (JCGDr and JCPDr) are

included. The estimated coefficients for the ZIOP model presented in Table 5 are presented in

Appendix Table C2.

18

The results presented in Table 5 show that, for provider visits numbering three or more, the

fever duration variables, Dur2, Dur3 and Dur4, for males are positive and statistically

significant at the five per cent level. Only the variable Dur4 is positive and statistically

significant at the one per cent level for females. The increased likelihood of seeing an

unqualified provider for fevers lasting 10 days or more (Dur4) for females explains the

higher observed proportion of visits to these providers numbering three and four, relative to

males. Possible explanations for this delay in seeking fever treatment include: intra-

household resource constraints, and females, on average, having a higher perception

threshold of fever severity.

The inclusion of a dummy variable reflecting the likelihood of males and females consulting

a second provider reveal contrasting results. At the discrete values of three, four and five,

males have a negative marginal effect of consulting a government MBBS provider during the

same fever episode. These marginal coefficients are significant at the one per cent level.

Negative marginal effect coefficients are estimated for females, at the same discrete values of

visits, for consulting private MBBS providers as a second provider for the same fever

episode. One plausible explanation for the reduced likelihood of females seeking further

fever treatments from a private MBBS provider, in addition to intra-household resource

constraints, is the reduced opportunity to travel further distances to seek private qualified

doctor fever treatment.

The marginal effect coefficient for the muslim identity dummy variable shows that muslim

women are also more likely to consult unqualified providers than their male counterparts. At

both the discrete visit values zero and three, at which counts have the highest number of

observations, the muslim female coefficients are positive and statistically significant at the 10

19

per cent level. The magnitude of the coefficient and the t-ratio remains larger for females,

than males, at discrete visit values four and five. The reason why there exists a tendency for

Muslim consumers not to consult government MBBS providers, as evidenced in Table 4 and

Table 5, is unclear. The effect of some dimension of consumer trust, and lack of trust in

individual government doctors or the institution, is a possible explanation .

The comparison of the OP and ZIOP models favour the later, with its ability to better account

for the zero count observations. The combined use of the probit splitting model and the

ordered probit model provide a better method of accounting for zero observations due to

non-participation and zero recorded visits. Gender and religious differences are evident in the

utilisation patterns and choice of provider decisions. Women are more likely to consult an

unqualified provider, as their initial provider choice, during more sever fever episode, than

compared to men.

5 Discussion

Supply-side behaviour is a plausible explanation for the observed pattern of healthcare

provider utilisation in treating fever symptoms in rural north India. The modelling framework

of repeat visits, while controlling for the number of illness episodes, allows for insights into

single episode utilisation patterns. The uneven distribution of provider visits across all

healthcare provider types offers a prima facie case for supplier-induced demand (SID) with

negative health and financial outcomes. The absence of two-visit responses from consumers,

but the strong presence of one and three visit responses, is counter-intuitive. The absence of a

‘low grade’ ‘high grade’ explanation of the distribution offers more support to the SID

hypothesis. Moreover, the literature by Das and colleagues (2005, 2007a, 2007b) who argue

that qualified and unqualified healthcare providers over prescribe medicines, in part, due to

20

profit maximising motivations also supports the SID hypothesis. The presence of this

utilisation pattern in the informal and formal outpatient markets, with the presence of

informal patient payments for government MBBS services, suggests systematic supplier

behaviour potentially away from patients’ best clinical and financial best interests.

The apparent positive price signalling of quality healthcare in the private outpatient sector,

coupled with potential SID for fever treatment, offers insights into healthcare providers’

incentives to over-charge. The positive marginal effect price coefficient for private MBBS

providers, and the corresponding positive price coefficient for male consumers of unqualified

providers, for visit responses greater than two indicates that a segment of outpatient

consumers use price information to identify healthcare quality. In this scenario, healthcare

providers have an incentive to increase their prices. As a result, using the logic of Das et al.

(2016), the rewards to treatment are likely to outweigh the market incentives for providers to

commit more effort in their consultations.

The finding that females seek fever treatment for more severe fevers than males is supported

by the human capital interpretation of consumer demand for healthcare. This apparent

behaviour of not seeking fever treatment for less severe fevers supports the view that

households invest initially more in male fever treatments and delay investment in female

fever treatment. However, the more severe nature of female fevers forces females to make

repeat visits.

The existence of repeat visits for outpatient fever treatment, through SID and behaviours of

households in delaying care, leads one to believe that under current market conditions the

shrinkage of the informal sector is unlikely to occur as the developing economies grow. The

21

incentives of providers to increase their diagnostic efforts are outweighed by the rewards to

inducing repeat visits. This is believed to have major policy implications for developing

economies seeking to implement universal healthcare system. Without further knowledge of

the outpatient market in developing economies, such as in north India, it is highly likely that

policy initiatives aimed at promoting the formal healthcare sector will be eroded by the

continued existence of the informal healthcare market.

22

References

Ahmed SM, Hossain MA, Chowdhury MR., 2009. Informal sector providers in Bangladesh:

how equipped are they to provide rational health care? Health Policy and Planning 24,

467-478.

Amin AA, Marsh V, Noor AM, Ochola SA, Snow RW., 2003. The use of formal and

informal curative services in the management of paediatric fevers in four districts in Kenya.

Tropical Medicine and International Health 8, 1143-1152.

Bago d'Uva T., 2005. Latent class models for use of primary care: evidence from a British

panel. Health Economics 14, 873-892.

Bago d'Uva T., 2006. Latent class models for utilisation of health care. Health Economics 15,

329-343.

Banerjee A, Deaton A, Duflo E., 2004. Wealth, health, and health services in rural Rajasthan.

The American Economic Review 94, 326-330.

Banerji A, and Jain S., 2007. Quality dualism. Journal of Development Economics 84,

234-250.

Basu K., 2006. Gender and Say: a model of household behaviour with endogenously

determined balance of power. The Economic Journal 116, 558-580.

Behrman JR, Kohler H-P, Watkins SC., 2002. Social networks and changes in contraceptive

use over time: evidence from a longitudinal study in rural Kenya. Demography 39, 713-738.

Berman P, Ahuja R, Bhandari L., 2010. The impoverishing effect of healthcare payments in

India: new methodology and findings. Economic and Policial Weekly (April 17); 65-71.

Bloom G, Standing H, Lloyd R., 2008. Markets, information asymmetry and health care:

Towards new social contracts. Social Science and Medicine 66, 2076-2087.

23

Bolin K, Jacobson L, Lindgren B., 2001. The family as the health producer — when spouses

are Nash-bargainers. Journal of Health Economics 20, 349-362.

Bolin K, Jacobson L, Lindgren B., 2002. The family as the health producer—when spouses

act strategically. Journal of Health Economics 21, 475-495.

BÖhme M, and Thiele R., 2012. Is the Informal Sector Constrained from the Demand Side:

Evidence for Six West African Capitals. World Development 40, 1369-1381.

Cameron AC, Trivedi PK., 1986. Econometric models based on count data. Comparisons and

applications of some estimators and tests. Journal of Applied Econometrics 1, 29-53.

Cameron AC, Trivedi PK., 2013. Regression Analysis of Count Data. Cambridge University

Press: New York.

Chang F-R, Trivedi PK., 2003. Economics of self-medication: theory and evidence. Health

Economics 12, 721-739.

Das J, and Das S., 2003. Trust, learning, and vaccination: a case study of a North Indian

village. Social Science & Medicine 57, 97-112.

Das J, Hammer J., 2007a. Location, location, location: residence, wealth, and the quality of

medical care in Delhi, India. Health Affairs 26, w338-w351.

Das J, Hammer J., 2007b. Money for nothing: the dire straits of medical practice in Delhi,

India. Journal of Development Economics 83, 1-36.

Das J, Holla A, Das V, Mohanan M, Tabak D, Chan B., 2014. In urban and rural India, a

standardized patient study showed low levels of provider training and hugh quality gaps.

Health Affairs 31, 2774-2784.

Das J, Holla A, Mohpal A, Muralidharan K., 2016. Quality and Accountability in Healthcare

Delivery: Audit-Study Evidence from Primary Care in India. American Economic Review

24

106, 3765-3799.

Deb P, Trivedi PK., 1997. Demand for medical care by the elderly: a finite mixture approach.

Journal of Applied Econometrics 12; 313-336.

Deb P, Trivedi PK., 2002. The structure of demand for health care: latent class versus two-

part models. Journal of Health Economics 21, 601-625.

Greene WH., 2012. Limdep version 10 Econometric Modelling Guide. Econometric Software

Inc: New York.

Gopalakrishnan S, Ajitha K, Ganeshkumar P, Selvaraj I, Logarai M., 2012. Assessment of

patient care and health facility indicators among urban and rural private practitioners in

Kancheepuram district of Tamil Nardu, India. National Journal of Research in Community

Medicine 1, 178-241.

Grossman and Benham, L., 1974. ‘Health, Hours and Wages’, in The Economics of Health

and Medical Care, ed. by Perlman, M., London: Macmillan, 205-233.

Harris MN, Zhao X., 2007. A zero-inflated ordered probit model, with an application to

modelling tobacco consumption. Journal of Econometrics 141, 1073-1099.

Hass-Wilson D., 2001. Arrow and the information market failure in health care: the changing

content and sources of health care information. Journal of Health Politics and Law 26, 1031-

1044.

Hazra A, Tripathi SK, Alam MS., 2000. Prescribing and dispensing activities at the facilities

of a non-government organization. The National Medical Journal of India, 13, 177-182.

Iles, R., 2014. Demand for qualified and unqualified primary healthcare in rural north India.

Unpublished PhD thesis. Griffith University: Brisbane, Australia.

25

Jiménez-Martín S, Labeaga JM, Martínez-Granado M., 2004. An empirical analysis of the

demand for physician services across the European Union. The European Journal of Health

Economics 5, 150-165.

Kamat VR, Nichter M., 1998. Pharmacies, self-medication and pharmaceutical marketing in

Bombay, India. Social Science and Medicine 47, 779–794.

Klasen S., 1994. "Missing Women" Reconsidered. World Development 22, 1061-1071.

La Porta R, Shleifer A., 2014. Informality and Development. Journal of Economic

Perspectives 28, 109-126.

Labelle R, Stoddart G, Rice T., 1994. A re-examination of the meaning and importance of

supplier-induced demand. Journal of Health Economics 13, 347-368.

Lancaster G, Maitra, P, Ray, R., 2008. Household expenditure patterns and gender bias:

evidence from selected Indian states. Oxford Development Studies 36, 133-157.

Lindelow M, Serneels P., 2006. The performance of health workers in Ethiopia: results from

qualitative research. Social Science and Medicine 62, 2225-2235.

Manser M, Brown M., 1980. Marriage and household decision-making: a bargaining

analysis. International Economic Review 21, 31-44.

McKelvey RD, Zavoina W., 1975. A statistical model for the analysis of ordinal level

dependent variables. The Journal of Mathematical Sociology 4, 103-120.

Rao M, Rao KD, Kumar AKS, Chatterjee M, Sundararaman T., 2011. Human resources for

health in India. The Lancet 377, 587-598.

Santos Silva JMC, Windmeijer F., 2001. Two-part multiple spell models for health care

demand. Journal of Econometrics 104, 67-89.

26

Scott, A, Vick, S., 1999. Patients, doctors and contracts: an application of principal-agent

theory to the doctor-patient relationship. Scottish Journal of Political Economy 46, 111-134.

Turnham D., 1993. Employment and Development: A Review of Evidence. Development

Centre, OECD. Paris.

Wennberg JE, Barnes BA, Zubkoff M., 1982. Professional uncertainty and the problem of

supplier-induced-demand. Social Science and Medicine 16, 811-824.

Windmeijer FAG, Silva JMCS., 1997. Endogeneity in count data models: an application to

demand for health care. Journal of Applied Econometrics 12, 281-294.

27

Appendices

Appendix A

A multi-stage clustering sampling frame was used in Stages One and Two of the data

collection. Stage One included 48 surveys with consumers, healthcare providers and local

level key informants (Community Health Workers – ASHAs and Village Leaders –

pradhans). Stage Two encompassed the delivery of the Revealed Preference (RP) surveys.

An important constraint surrounding the data collection process was the need to control for

potential seasonal effects associated with fever treatment demand. The monsoon season is

expected to generate a high level of fever treatment demand in north India. Stage One data

was collected during April and May in the dry-season of 2012. Stage Two data was collected

during the north Indian monsoon season, which usually runs from mid-June to mid-

September. However, in 2012 the monsoon onset was delayed by three-four weeks in north

India. As a result, the average rainfall during the first 2-months of the season was below

average (Government of India, 2013). This contracted the duration of the monsoon and

limited the window of opportunity. However, Stage Two data was still collected during the

monsoon.

Sampling frame

The primary sampling units of the study were three districts from Uttar Pradesh (UP).

Purposeful sampling was employed in selecting the three districts. Districts were selected due

to their representative mean income profiles and for accessibility reasons (Government of

Uttar Pradesh, 2006). The district mean incomes of the three districts cover the interquartile

range of UP. In addition, one known NGO hospital from the Emmanuel Hospital Association

was located in each district. While the data were collected independently of the NGO

hospitals, logistical support was provided by these hospitals throughout Stages One and Two.

28

At the sub-district level the State administrative hierarchy continues to the Block and Gram

Panchayat6 levels before reaching individual villages. Gram Panchayats are a collection of

four - six villages under the administrative leadership of an elected leader and council. Blocks

within a district typically contain Block Development Offices that are responsible for local

level administration and program ‘oversight’.

The sampling units used in this study were district level development blocks. These were

selected at random from a list of development blocks available on district websites. At the

tertiary level, Gram Panchayats from the select block(s) were stratified according to the

Hindu and Muslim religious majority of the Gram Panchayats. Assistance in stratifying was

obtained from district Block Development Offices. In total, eight Gram Panchayats were

selected: four from Fatehpur and two each from Lalitpur and Balrampur. The location of

Fatehpur in UP’s central region allowed for proportional sampling of Hindu and Muslim

individuals (three Hindu majority Gram Panchayats and one Muslim majority). The small

proportion of Muslim residents in Lalitpur from the Bundelkhand region, limited the ability

to sample Muslim respondents. This was balanced by sampling an equal number of Gram

Panchayats in Balrampur, which had a strong Muslim representation.

Sampling of village households, within all villages except Village Two (Fatehpur),

systematically covered all geographic sections in a quasi-random process. Maps of the

villages were not available, so local knowledge of the village was drawn upon to ensure that

enumerators sampled evenly across the whole village. This approximate even sampling

across any given village was important as many villages were informally divided according

to religion and caste. Enumerators selected the households and individuals to survey. As a

check on the micro-level sampling by enumerators, sampling profiles for each enumerator of

6 A collection of several neighbouring villages - often between 2-4 villages.

29

their survey respondents was monitored during each day of data collection. This monitoring

of sampling profiles included consideration of mean age and gender proportions. In this light,

the sampling of individuals conformed to a quota method. Village Two was the first village

sampled. The village elected leader (i.e. pradhan) organised the recruitment of villagers

according to our representative sample request.

Choice set creation

A survey of providers was also completed as part of Stage One of the study to help verify the

range of provider characteristics available in the local outpatient market. Village level

healthcare providers were identified by local key informants. Table A1 provides a summary

of the providers surveys and their attributes.

Broad uniformity among surveyed unqualified – jhola chhaaps – is evident from Table A1:

average stated prices range between 40 – 60 rupees for a single visit, a margin of 10 rupees is

added to the cost of medicine by providers, word-of-mouth recommendation were widely

seen to be the main mechanism to grow one’s business and no direct advertising was

employed by unqualified – jhola chhaaps – providers. All providers cited that fever

symptoms were one of the leading complaints by patients. Based on the survey result of

healthcare providers it is assumed that in the treatment of mild-severe fevers the attributes of

unqualified providers were homogeneous.

30

Table A1: Summary of provider survey results

VillageNo. 1 1 1 2 2 3 3 4 nonsample 5 5 7 7 7 8 8

Description JholaChhaap Pharmacist JholaChhaap Ayurvedic JholaChhaap JholaChhaap JholaChhaap 'Wardboy' MBBSDr MBBSDr JholaChhaap Nurse JholaChhaap JholaChhaap Pharmacist Jhola Chhaap

Gov'tFacility 0 1 0 0 0 0 0 1 1 1 0 1 0 0 1 0

Nearestgov'tfacility <1km <1km <1km <1km 7km 7km 5km 10km 10km < 1 km

Gender male male male male male male male male male male male female male male male male

Age 50-54 45-49 25-29 65+ 45-49 30-34 20-24 45-49 35 25-29 36 50 22 50-54 45-49

Religion hindu hindu hindu hindu hindu hindu hindu hindu hindu hindu hindu hindu hindu hindu hindu hindu

Experience(years) 25 2 54 20 8 1 10 2 2 30 3 30 20 27

Typeofclinic Drugstore CHC informal informal drugstore informal informal CHC CHC PHC informal CHC shopfront shopfront PHC shop front

sizeofclinic 1 1 2 1 1 1 1 1

AveragePrice–fever denied 1 50-100costofinjection+

margin40 40 35 1 1 40 50-60 50 1 50

Margin/Fees denied 50% 10INR 15INR 5INR 10INR 10INR 10INR 10INR 10 INR

Validity 15 days 5 days 1visit 1visit 1visit 15days 15days 1visit 1visit 1visit 1 visit

Patients per day* 100 12 2 6 10 7 100 80 5 12 15 15

ConsultLength 4 mins 5mins 15mins+ 2mins^ 10mins 4mins 5mins 10mins 5mins^ 3 mins^

PatientTraveldistance–

radius10km 10km 2km 2km 1km upto40kms upto20kms 1km 1km 2 km

Fever(50%) Fever(80%) Fever-most Fever(50%) Fever(70%) Fever(70%) Fever(50%) Fever(50%) Fever(50%) Diarrhea(50%) Diarrhea(40%) Diarrhea(40%)

Diarrhea(10%) Diarrhea(20%) Diarrhea(25%) Diarrhea(25%) Diarrhea(20%) Diarrhea(15%)Women'shealth

(25%)Diarrhea(20%) Fever (40%) Fever(40%) Fever(40%)

SkinDiseases(4%) Other(25%) Other(5%) Other(10%) Lung(20%) Diarrhea(20%) Lung(20%) Lung(10%) Other(20%) Injury+Other(20%)

Other(36%)Women'shealth+

Other(15%)Lung(5%) Other(10%)

Advertising Word-of-mouth Word-of-mouth Word-of-mouth Word-of-mouthWord-of-mouthand

pastexperience

Word-of-mouth&

pastexperience

Word-of-mouth

&past

experience

Word-of-mouth&

pastexperience

Price Location Qualification Location PersonalExperinence Location Price Price Price Location Experience Price

Location AvailabilityofMedicine Location Price Price PersonalExperience TrustinHospital Recommendation Distance Price Location Location

Recommendation Price PersonalExperinece PersonalExperience Recommendation Price Medicines Personalexperience Personalexperience Recommendation Price PersonalExperience

Comments

Respondent claimed to

not prescribe medicine.

This was in opposition to

the claim ofvillage

residents.

Respondents indicated

that doctors lived 2

hours away (x 2 Drs)

and 30 minutes away (x

1) in district centres.

NoDrsavalable.

Respondentindicated

thatDrswereoftennot

present-living2+hours

away(x2)and45mins

away(x1).

NoDrsavailable.Claimed

thatDrsoftenbackdated

recordstogivethe

appearancethattheywere

present.Translatorwasan

oldacquanceofthenurse.

DrNotavailable.

Translatorwasanold

acquanceofpharmacist.

*patientsperdaydoesn'tnecessarilyrefertoathomevisits

^observed

TopAilments

Perception-Whypatients

attend?

31

Appendix B

The pattern of consumer visits to providers shows no signs of statistical difference across

shorter recall periods. Table B1 presents the percentage distribution of healthcare provider

visits by recall period. The table contains the following: column (1) lists the number of visits

recorded as a value between one – five, columns (2)-(5) list the percentage of observed

healthcare provider visits for a single fever episode for independent recall periods, 14 days,

15-30 days, 2-6 months, 7-12 months; column (6) is the weighted average of the percentage

in the preceding four columns; and column (7) gives the p-values for the null hypothesis that

values in the ≤ 14-day recall are the same as those in the 7-12 month recall period. The p-

values in column (7) indicate that there is no statistically significant difference between the

percentage of healthcare provider visits at each discrete interval, in the range one to five.

Table B1: Distribution of visits to healthcare providers across recall periods ≤ 14-days

to 7-12 months

The dependent variable in count and ordered probit models are discrete values. A plausible

explanation for the low frequency of two visits, as highlighted in Table 1, is the confounding

of the number of visits by severity of fever. Table B2 provides the frequency distribution of

visits to i) unqualified providers (JC), ii) government MBBS providers (GDr), and iii) private

MBBS provider (PDr) across fever duration intervals. Controlling for severity of fever, the

Number of Vis i ts ≤ 14days 15-30days 2-6months 7-12months Average*

χ2 test

p-values

(1) (2) (3) (4) (5) (6) (2) vs (5) = (7)

1 49.2 53.0 51.6 48.5 50.7 0.897

2 4.4 5.2 4.0 7.6 5.0 0.142

3 25.6 25.9 31.3 29.2 27.8 0.391

4 14.8 12.4 10.7 10.5 12.4 0.188

5 3.4 2.8 1.2 4.1 2.8 0.686

> 5 2.7 0.8 1.2 0.0 1.3 0.030

n 297 251 252 171 971

Note: * weighted average

32

frequency data in table 2 demonstrates that there is a uniform absence in the percentage of

respondents making two visits to all health provider types. The percentage distribution of

zero visits is the largest observed outcome across each provider type. Naturally, as the market

share of any provider type declines, within the 12-month recall period, the resulting

percentage of zero visits, all other things being equal, increases. This explains the first cause

of zero visits. Those consumers who have been sick in the past 12 months have initially

sought fever treatment from another provider. The second explanation for the high percentage

of zeros observed is that respondents did not consider themselves as having fever symptoms

of sufficient severity to warrant seeking treatment.

Table B2: Distribution of visits to healthcare providers in 12-month recall period, by

provider type and fever duration

Visits # JC Gdr Pdr JC Gdr Pdr JC Gdr Pdr JC Gdr Pdr JC Gdr Pdr

0 22.6 32.6 35.3 17.8 21.9 27.1 8.1 8.5 9.2 3.2 3.2 3.8 9.5 9.8 10.2

1 10.9 3.1 4.2 5.1 4.0 2.1 0.9 1.8 0.9 0.3 0.7 0.2 0.9 1.2 0.9

2 1.1 0.7 0.1 0.7 0.3 0.0 0.3 0.4 0.1 0.0 0.1 0.0 0.2 0.2 0.0

3 4.7 2.3 0.7 5.0 3.1 1.2 1.5 0.6 0.7 0.4 0.3 0.5 0.3 0.9 0.3

4 1.2 1.6 0.3 1.8 1.0 0.1 0.7 0.3 0.5 0.4 0.3 0.1 1.1 0.1 0.3

5-n* 0.0 0.3 0.1 0.3 0.3 0.2 0.2 0.2 0.3 0.3 0.2 0.1 0.4 0.3 0.6

* nmax values: JC=11, GDr=11 & PDr=13

1-3 days 4-6 days 7-9 days 10-12 days 13+ days

33

Appendix C

Table C1: Zero-inflated Ordered Probit coefficient estimates by provider type –

clustered and truncated

JC Gdr Pdr

Coeff. Coeff. Coeff.

Price <0.001 -0.005 *** <0.001 ***

Lnhinc 0.044 -0.099 -0.013

DUR2 d 0.166 * - -0.031

DUR3 d -0.023 - 0.420 ***

DUR4 d 0.007 - 0.177

D1 d 0.957 ** -0.712 ** -0.624 *

D2 d 0.786 *** -0.349 ** -0.432

D3 d 0.618 *** 0.256 -0.444

D4 d 0.196 0.470 * -0.234

Female d 0.258 *** -0.021 0.109

DB d 0.068 - -

DC d 0.143 - -

Muslim d 0.327 *** -0.377 *** -

JOB1 d 0.149 - -0.312 *

JOB2 d 0.337 ** - -0.068

JOB9 d -0.093 - -0.384 **

Constant -1.806 0.835 -0.537

34

Table C2: Zero-inflated Ordered Probit coefficient estimates for Unqualified – jhola

chhaap - provider by gender – clustered and truncated

Males Females

Coeff. Coeff.

Price 0.004 *** 0.000

Lnhinc -0.031 0.070

DUR2 d 0.675 *** 0.002

DUR3 d 1.017 *** 0.507

DUR4 d 1.464 *** 1.207 ***

D1 d 1.326 *** 1.197 ***

D2 d 1.216 *** 0.894 ***

D3 d 0.873 *** 0.832 ***

D4 d 0.425 0.046

Muslim d 0.191 0.298 *

JOB1 d -0.005 -0.004

JOB2 d 0.537 *** 0.315

JOB9 d -0.528 -0.298

JCGDr d -1.449 ** 0.298

JCPDr d 0.405 -0.618 ***

Constant -1.136 -0.974

Mu(1) 1.017 *** 0.679 ***

Mu(2) 1.176 *** 0.763 ***

Mu(3) 2.131 *** 1.536 ***

Mu(4) 2.948 *** 2.681 ***

35

Tables

Table 1: Mean number of visits to healthcare providers for a single episode of fever, by

gender

JC GDr PDr OT

Male Female Male Female Male Female Male Female

μ rate 1.8 2.1 1.9 1.8 1.9 1.9 1.3 1.0

N 261 263 179 166 93 105 53 53 Note: JC = unqualified ‘doctors’; GDR = Government MBBS doctor; PDR = private MBBS doctor and OT =

Other.

36

Table 2: Descriptive statistics of ≤ 12-month sample

≤ 12-months

Mean (se)/%

Age in years (18 years and over) 39.6 (15.6)

Percentage of female respondents 51.1

Mean price across a l l healthcare provider types in INR 98.2 (253.6)

Log of annual household Income 10.7 (0.7)

Number of fami ly members in joint household 6.9 (3.1)

Percentage of respondents with a fever las ting 1-3 days 42.1

Percentage of respondents with a fever las ting 4-6 days 31.6

Percentage of respondents with a fever las ting 7-9 days 11.6

Percentage of respondents with a fever las ting 10+ days 14.7

Percentage of respondents who accessed healthcare providers at home 1.0

Percentage of respondents who travel led within the vi l lage 13.7

Percentage of respondents who travel led 0-4 km 41.9

Percentage of respondents who travel led 5-9 km 24.8

Percentage of respondents who travel led 10-14 km 12.1

Percentage of respondents who travel led 15-19 km 4.2

Percentage of respondents who travel led 20+ km 2.5

Percentage of respondents who are Brahmin 12.1

Percentage of respondents who are Kshratrya 4.0

Percentage of respondents who are Vaisya 36.9

Percentage of respondents who are Shuda 22.0

Percentage of respondents who are Triba l 3.5

Percentage of respondents who are none of the above caste 21.3

Percentage of respondents who sel f-reported being i l l i terate 45.3

Percentage of respondents who sel f-reported having completed some

school ing less than completed High School43.7

Percentage of respondents who sel f-reported having completed at least

Senior High School .10.9

Percentage of respondents who are class i fied as a Farmer 25.4

Percentage of respondents who are class i fied as a labourer 24.5

Percentage of respondents who are class i fied as a Market sel ler 0.5

Percentage of respondents who are class i fied as a Shopkeeper 2.7

Percentage of respondents who are class i fied as a Tradesperson 2.2

Percentage of respondents who are class i fied as government employees 1.7

Percentage of respondents who are class i fied as running own bus iness 0.4

Percentage of respondents who are class i fied as Unemployed 7.9

Percentage of respondents who are class i fied as unpaid Domestic worker 29.7

Percentage of respondents who are class i fied as in the “Other” category 5.0

DistA 46.5

CHC 6.2

PHC1 15.9

DistB 30.0

Dis tC 23.6

PHC2 12.2

Note: Standard deviations are in parenthesis where appropriate.

Travel distance interval – healthcare provider (% of all providers)

Caste* (% of total sample, including Muslim respondents)

Literacy (%)

Employment (%)

District (%)

Percentage of respondents in Vi l lage 8 which has a Primary Health Centre

Percentage of respondents in Dis trict A

Percentage of respondents in Vi l lage 1 which has a Community Health Centre

Percentage of respondents in Vi l lage 2 which has a Primary Health Centre

Percentage of respondents in Dis trict B

Percentage of respondents in Dis trict C

Job 5

Job 6

Job 7

Job 8

Job 9

Job 10

High Li t

Job 1

Job 2

Job 3

Job 4

Caste4

Tribe

Other

Il l i t

Li t

Caste3

D0

D1

D2

D3

D4

D5

D6

Caste1

Caste2

Dur4

Variable Defini tion

Age

Female (%)

Price

lnhinc

Hhs ize

Dur1

Dur2

Dur3

Illness Duration (%)

37

Table 3: Goodness-of-fit measures for ordered probit and zero-inflated ordered probit

OP ZIOP OP ZIOP OP ZIOP

μ1 0.572*** (0.073) 0.823*** (0.084) 0.473*** (0.071) 0.583*** (0.056) 0.515*** (0.099) 0.635*** (0.078)

μ2 0.660*** (0.085) 0.940*** (0.889) 0.565*** (0.079) 0.689*** (0.061) 0.531*** (0.103) 0.656*** (0.079)

μ3 1.310*** (0.066) 1.781*** (0.126) 1.097*** (0.144) 1.284*** (0.088) 0.956*** (0.123) 1.190*** (0.126)

μ4 2.082*** (0.065) 2.774*** (0.181) 1.677*** (0.102) 1.905*** (0.136) 1.265*** (0.086) 1.577*** (0.156)

LL -1300.5 -1259.5 -923.9 -894.3 -640.6 -610.3

AIC 2643 2581 1873.8 1834.5 1317.3 1266.6

BIC 2749.4 2738.1 1939.6 1951.1 1408.5 1383.1

LR versus OP 81.9 812.5 60.7

Hausman versus OP 62.3 6.8 760.5

Vuong versus OP 4.3 3.7 3.6

JC Gdr Pdr

38

Table 4: Marginal effects of explanatory variable from ZIOP model for each provider

estimate

JC Gdr Pdr

Variables Coefficient t-ratio Coefficient t-ratio Coefficient t-ratio Variables Coefficient t-ratio Coefficient t-ratio Coefficient t-ratio

Price <-0.001 -0.41 0.002 7.68 -0.001 -8.56 Price 0.000 0.41 -0.001 -7.17 <0.001 7.15

Lnhinc -0.024 -0.79 0.031 1.20 <0.001 0.00 Lnhinc 0.012 0.79 -0.010 -1.21 <-0.001 0.00

DUR2 d-0.122 -2.43 - - 0.002 0.04 DUR2 d

0.064 2.42 - - <-0.001 -0.04

DUR3 d-0.252 -4.82 - - -0.156 -2.44 DUR3 d

0.125 6.47 - - 0.046 2.39

DUR4 d

-0.393 -7.95 - - -0.033 -0.72 DUR4 d

0.146 5.72 - - 0.009 0.73

D1 d

-0.312 -6.49 0.185 3.88 0.104 2.16 D1 d

0.142 8.81 -0.054 -4.08 -0.028 -2.31

D2 d-0.320 -5.41 0.065 1.07 0.066 1.31 D2 d

0.161 6.10 -0.020 -1.08 -0.018 -1.33

D3 d-0.249 -4.50 -0.111 -1.56 0.090 1.94 D3 d

0.127 5.16 0.036 1.53 -0.024 -1.96

D4 d

-0.067 -0.89 -0.131 -1.57 0.027 0.47 D4 d

0.035 0.88 0.043 1.52 -0.007 -0.48

Female d-0.117 -2.56 0.002 0.07 -0.036 -0.98 Female d

0.061 2.55 -0.001 -0.07 0.010 0.97

DistB d

-0.002 -0.03 - - - - Dis tB d

0.001 0.03 - - - -

Dis tC d

0.085 1.53 - - - - Dis tC d

-0.043 -1.52 - - - -

Mus l im d-0.082 -1.83 0.130 3.02 - - Mus l im d

0.043 1.84 -0.039 -3.17 - -

JOB1 d-0.059 -1.13 - - 0.073 2.03 JOB1 d

0.031 1.13 - - -0.020 -2.02

JOB2 d-0.182 -3.99 - - -0.038 -0.79 JOB2 d

0.095 4.06 - - 0.011 0.77

JOB9 d

0.076 1.25 - - 0.078 1.84 JOB9 d

-0.039 -1.27 - - -0.021 -1.83

Price <0.001 -0.02 -0.001 -8.68 <0.001 13.16 Price <0.001 0.40 <-0.001 -5.26 <0.001 3.44

Lnhinc <-0.001 -0.02 -0.013 -1.20 <-0.001 0.00 Lnhinc 0.009 0.78 -0.004 -1.21 <-0.001 0.00

DUR2 d

-0.007 -0.43 - - -0.001 -0.04 DUR2 d

0.049 2.09 - - <-0.001 -0.04

DUR3 d-0.073 -1.54 - - 0.073 2.98 DUR3 d

0.149 3.01 - - 0.019 1.85

DUR4 d-0.161 -3.41 - - 0.017 0.75 DUR4 d

0.276 6.97 - - 0.003 0.69

D1 d-0.108 -2.06 -0.091 -3.63 -0.060 -2.09 D1 d

0.201 4.09 -0.021 -3.78 -0.009 -2.07

D2 d-0.022 -0.57 -0.028 -1.06 -0.036 -1.32 D2 d

0.135 4.01 -0.009 -1.08 -0.006 -1.29

D3 d

-0.043 -1.16 0.044 1.72 -0.051 -1.92 D3 d

0.124 3.20 0.018 1.40 -0.008 -1.80

D4 d-0.004 -0.34 0.049 1.87 -0.015 -0.47 D4 d

0.027 0.80 0.022 1.34 -0.003 -0.49

Female d

-0.001 -0.04 -0.001 -0.07 0.020 0.99 Female d

0.043 2.33 0.000 -0.07 0.004 0.96

DistB d0.000 -0.02 - - - - Dis tB d

0.001 0.03 - - - -

Dis tC d

-0.004 -0.46 - - - - Dis tC d

-0.028 -1.52 - - - -

Mus l im d-0.005 -0.45 -0.060 -2.85 - - Mus l im d

0.033 1.65 -0.016 -3.08 - -

JOB1 d-0.002 -0.26 - - -0.041 -1.99 JOB1 d

0.023 1.05 - - -0.007 -1.95

JOB2 d-0.022 -0.81 - - 0.020 0.81 JOB2 d

0.082 2.86 - - 0.004 0.73

JOB9 d-0.003 -0.26 - - -0.043 -1.80 JOB9 d

-0.026 -1.32 - - -0.007 -1.71

Price <0.001 0.39 <-0.001 -1.77 <0.001 0.26 Price <0.001 0.40 <-0.001 -4.33 <0.001 3.37

Lnhinc 0.001 0.70 -0.002 -1.00 <-0.001 0.00 Lnhinc 0.002 0.77 -0.001 -1.16 <-0.001 0.00

DUR2 d0.006 1.40 - - <-0.001 -0.04 DUR2 d

0.010 1.91 - - <-0.001 -0.04

DUR3 d

0.005 1.52 - - 0.002 0.25 DUR3 d

0.046 2.05 - - 0.016 1.78

DUR4 d-0.002 -0.24 - - <0.001 0.22 DUR4 d

0.133 3.46 - - 0.003 0.70

D1 d

0.003 0.62 -0.014 -1.66 -0.001 -0.25 D1 d

0.073 2.33 -0.006 -3.56 -0.006 -2.16

D2 d0.014 1.71 -0.005 -0.95 -0.001 -0.25 D2 d

0.032 2.91 -0.003 -1.06 -0.005 -1.21

D3 d0.009 1.96 0.008 1.16 -0.001 -0.25 D3 d

0.032 2.33 0.006 1.25 -0.006 -1.74

D4 d0.003 0.90 0.009 1.17 0.000 -0.22 D4 d

0.006 0.73 0.008 1.17 -0.002 -0.50

Female d0.006 1.32 0.000 -0.07 0.001 0.25 Female d

0.008 2.16 0.000 -0.07 0.003 0.89

DistB d

0.000 0.03 - - - - Dis tB d

0.000 0.03 - - - -

Dis tC d-0.005 -0.93 - - - - Dis tC d

-0.005 -1.55 - - - -

Mus l im d

0.004 1.37 -0.010 -1.52 - - Mus l im d

0.007 1.49 -0.005 -3.16 - -

JOB1 d0.003 0.98 - - -0.001 -0.26 JOB1 d

0.004 0.99 - - -0.005 -1.75

JOB2 d0.008 1.71 - - 0.001 0.25 JOB2 d

0.019 2.27 - - 0.003 0.72

JOB9 d-0.004 -0.93 - - -0.001 -0.26 JOB9 d

-0.005 -1.33 - - -0.005 -1.69

Note: * Y=N equals the number of vis i ts ; d dummy variable

Y=3

Y=4

Y=5

JC Gdr Pdr

Y*=0

Y=1

Y=2

39

Table 5: Zero-inflated ordered probit marginal effects for unqualified (jhola chhaap)

providers, by gender

Variables Coefficient t-ratio Coefficient t-ratio Variables Coefficient t-ratio Coefficient t-ratio

Price -0.001 -4.64 <-0.001 -0.65 Price 0.001 4.38 <0.001 0.65

Lnhinc 0.010 0.25 -0.028 -0.60 Lnhinc -0.007 -0.25 0.012 0.60

DUR2 d

-0.192 -3.59 -0.001 -0.01 DUR2 d

0.133 3.79 <0.001 0.01

DUR3 d

-0.229 -4.27 -0.188 -1.60 DUR3 d

0.154 6.76 0.083 1.71

DUR4 d

-0.309 -5.26 -0.392 -5.81 DUR4 d

0.178 5.60 0.145 7.77

D1 d

-0.274 -4.82 -0.388 -5.35 D1 d

0.163 5.42 0.143 7.81

D2 d

-0.348 -4.73 -0.335 -3.61 D2 d

0.222 6.34 0.146 3.96

D3 d

-0.227 -3.52 -0.302 -3.38 D3 d

0.158 4.41 0.130 3.92

D4 d-0.119 -1.58 -0.018 -0.14 D4 d

0.085 1.54 0.008 0.14

Musl im d-0.058 -1.04 -0.115 -1.86 Musl im d

0.040 1.06 0.052 1.88

JOB1 d0.002 0.03 0.002 0.01 JOB1 d

-0.001 -0.03 -0.001 -0.01

JOB2 d

-0.155 -2.89 -0.121 -1.41 JOB2 d

0.109 2.90 0.055 1.42

JOB9 d0.191 0.57 0.116 1.45 JOB9 d

-0.108 -0.69 -0.052 -1.46

JCGDr d0.531 2.57 -0.117 -0.66 JCGDr d

-0.215 -4.85 0.053 0.66

JCPDr d-0.112 -1.58 0.243 2.73 JCPDr d

0.081 1.60 -0.110 -2.72

Price <-0.001 -0.84 <0.001 0.60 Price <0.001 3.05 <0.001 0.65

Lnhinc 0.002 0.24 0.003 0.53 Lnhinc -0.003 -0.25 0.010 0.60

DUR2 d-0.067 -1.31 <0.001 0.01 DUR2 d

0.085 2.49 <0.001 0.01

DUR3 d-0.159 -2.14 -0.002 -0.09 DUR3 d

0.155 2.66 0.087 1.20

DUR4 d-0.223 -2.96 -0.060 -1.32 DUR4 d

0.221 3.66 0.239 3.53

D1 d

-0.214 -2.86 -0.061 -1.22 D1 d

0.205 3.51 0.238 3.22

D2 d-0.098 -1.36 0.019 0.74 D2 d

0.148 3.21 0.137 2.87

D3 d-0.108 -1.58 -0.007 -0.23 D3 d

0.120 2.33 0.145 2.42

D4 d-0.047 -0.84 0.002 0.16 D4 d

0.055 1.14 0.007 0.14

Musl im d-0.015 -0.81 0.006 0.69 Musl im d

0.022 0.99 0.046 1.59

JOB1 d<0.001 0.03 <-0.001 -0.01 JOB1 d

-0.001 -0.03 -0.001 -0.01

JOB2 d-0.052 -1.16 0.006 0.68 JOB2 d

0.067 2.12 0.049 1.25

JOB9 d

-0.016 -0.15 -0.011 -1.13 JOB9 d

-0.041 -0.95 -0.043 -1.34

JCGDr d-0.196 -1.31 0.013 0.63 JCGDr d

-0.061 -4.37 0.041 0.65

JCPDr d-0.046 -0.96 -0.027 -1.31 JCPDr d

0.053 1.26 -0.086 -2.59

Price <0.001 1.33 <0.001 0.59 Price <0.001 2.68 <0.001 0.64

Lnhinc -0.001 -0.24 0.001 0.54 Lnhinc -0.001 -0.25 0.001 0.60

DUR2 d

0.011 1.56 <0.001 0.01 DUR2 d

0.030 2.01 <0.001 0.01

DUR3 d0.001 0.13 0.006 1.11 DUR3 d

0.077 1.86 0.014 0.92

DUR4 d-0.004 -0.23 0.005 1.00 DUR4 d

0.137 2.24 0.063 1.99

D1 d-0.005 -0.30 0.005 0.91 D1 d

0.125 2.00 0.063 1.63

D2 d

0.019 1.58 0.012 1.18 D2 d

0.057 2.33 0.022 1.92

D3 d0.009 1.50 0.008 1.22 D3 d

0.049 1.66 0.026 1.63

D4 d0.007 1.51 0.001 0.14 D4 d

0.019 0.95 0.001 0.13

Musl im d

0.004 1.04 0.004 1.02 Musl im d

0.007 0.90 0.006 1.38

JOB1 d<-0.001 -0.03 <-0.001 -0.01 JOB1 d

<-0.001 -0.03 <-0.001 -0.01

JOB2 d0.009 1.46 0.004 0.97 JOB2 d

0.022 1.69 0.007 1.10

JOB9 d-0.017 -0.48 -0.005 -0.89 JOB9 d

-0.009 -1.22 -0.005 -1.20

JCGDr d

-0.047 -1.07 0.005 0.59 JCGDr d

-0.012 -3.15 0.005 0.65

JCPDr d0.006 1.50 -0.010 -1.10 JCPDr d

0.018 1.00 -0.010 -2.13

Note: * Y=N equals the number of vis i ts ; d dummy variable

Y=1

Y=2

Y=3

Y=4

Y=5

Males FemalesMales Females

Y*=0

40

Figures

Figure 1: Frequency distribution of visits to healthcare providers for a single fever

episode, by gender