Incentivizing Demand for Supply-Constrained Care: Institutional Birth in India

IT is increasingly common for governments to incentivize individuals to use health care.¹ The rationale is clear; barriers, such as a lack of information or resources, might prevent individuals from seeking beneficial care. However, health systems in many low- and middle-income countries (LMICs) that use demand-side incentives are often poorly resourced and provide poor-quality care.² Given this, incentivizing demand could backfire. The additional demand may lead to congestion externality if, by causing fixed resources to be spread more thinly, it results in the quality of care falling for everyone, including those who would have used the care without any subsidy. Moreover, when health care quality is poor, the marginal group who takes up care due to the incentive may not benefit. Finally, to manage an increase in demand for the incentivized care, providers may pull effort and resources away from other types of care, which could risk a wider deterioration in quality.

This paper examines these issues in the context of the world's largest conditional cash transfer program: India's Janani Suraksha Yojana (JSY). Introduced in 2005, JSY pays substantial cash incentives—about 28 times the average rural daily wage for casual labor—to women who give birth in a health facility.³ The scheme does not provide increased funds to facilities to cover the cost of additional deliveries. JSY is widely credited with causing a rapid decline in home births, which fell from 80% in 2005 to 40% in 2011 in the states targeted by the program.⁴ However, rigorous research has found that it had no overall impact on birth outcomes (Powell-Jackson et al., 2015). This is puzzling as a set of simple medical procedures can mitigate the vast majority of mortality risks associated with childbirth for both mother and infant (Lawn et al., 2005). The pervasive poor quality and low capacity of much of India's government health system might explain why bringing births into medical facilities had no overall effect on health. At less than 1% of GDP, India's public spending on health is a third (relative to national income) of other major emerging economies, and the system has a huge shortage of qualified health professionals (Rao et al., 2011) compounded by high rates of absenteeism (Chaudhury et al., 2006).⁵

This study explores the effects of incentivizing the demand for health care in this highly supply-constrained system. We begin by creating a flexible general-equilibrium model of birth outcomes and families' decisions about whether to deliver at home or in a facility. Crucially, we allow for births to be heterogeneous in their complexity and in how much they benefit (in expectation) from taking place in a facility. We allow births to sort into facilities endogenously, whether partially or fully or not at all, on the basis of these benefits. Key to our model is a congestion, or overcrowding, externality that causes the expected riskiness of all births in medical facilities, and especially the most-complex facility births, to increase when the overall use of facilities rises relative to their capacity.

In our model, the introduction of a subsidy has two effects. First, it results in private benefits (or harms) to the marginal births that take place in a different location as a result of the subsidy. Second, it results in social costs to the births that would have already occurred in facilities, the inframarginal facility births, which now become riskier because of increased congestion. The degree to which births select into facility delivery based on their private gains plays an important role in shaping both effects. For example, in the extreme case of perfect selection on gains (a classic Roy model), the marginal births will be harmed by being incentivized into facilities, as these births were making a privately optimal choice to deliver at home before the subsidy. Moreover, in this case the inframarginal facility births that will be exposed to the additional congestion are the most-complex births, which are precisely the births that are most harmed by congestion. In this extreme case, then, a subsidy unambiguously worsens health outcomes.

On the other hand, our model highlights that there are circumstances in which a subsidy could improve health outcomes. The more complex the marginal births are that are brought into facilities due to the subsidy, the more they will benefit from the move. At the same time, the fewer complex births that occur in facilities to start with, the smaller the aggregate size of the congestion externality. Weak initial sorting on gains but a marginal utility of receiving the subsidy that is increasing in birth complexity could give rise to such conditions.⁶ While it is theoretically ambiguous, and ultimately an empirical question, how the impact of the subsidy varies with respect to preexisting capacity, we show that there are several channels through which higher preexisting capacity increases the chance that a subsidy will be useful.

We examine the impacts of JSY empirically, focusing on two dimensions of heterogeneity that our model suggests will be key: preexisting system capacity and birth complexity. We use geographic disparities in the preexisting capacity of district health systems—disparities caused by the central government's allocation of funds and variation in states' political priorities (Kumar et al., 2011)—to assess how the effects of stimulating demand depend on supply constraints. Specifically, exploiting exogenous variation in JSY's rollout, we examine whether the preexisting capacity (specifically the numbers of doctors, nurses, and beds relative to the size of the population served) affected its impacts.

We find that JSY led to a large average increase, of 7.86 percentage points, in the probability that births were delivered in health facilities. In aggregate, this doubled the number of institutional deliveries per day for which each government-run secondary care facility was responsible. The increase in deliveries per facility was particularly stark in districts with below-median capacity in the secondary care system. Here caseload increased by a factor of 2.5, from 1.92 deliveries per facility per day to 4.80. We show evidence that while there was some degree of sorting on gains—more-complex births were more likely to occur in facilities than less-complex births both before and after JSY—JSY did not disproportionately draw in complex births. More- and less-complex births were equally likely to be marginal to the subsidy, suggesting that JSY was not particularly effective at targeting the deliveries that would have privately gained the most from moving to a facility.

We show that in these districts with below-median capacity in the secondary health care system, the average risk of perinatal mortality increased as a result of JSY by 0.90 percentage points, or by 24.3% relative to the rate before JSY. Although JSY caused the same increase in the institutional delivery rate in areas with above-median preexisting capacity, it had no impact on perinatal mortality, neither positive nor negative, in these areas. We show further that this increase in perinatal mortality risk was entirely concentrated on complex births. This pattern of heterogeneity can, in our model, only be rationalized by the presence of a congestion externality. Moreover, we find that JSY reduced the probability that infants received any checkup between 2 and 10 days after a facility birth, which could indicate that facilities reacted to congestion by discharging women earlier, or by reducing checkups for babies still in inpatient care.

Our empirical analysis goes further by examining whether the above responses are short-lived and whether the effects of the subsidy spill over into other areas of care. We present more tentative evidence suggesting that harmful effects of JSY in low-capacity areas may have persisted up to 5–10 years after the start of JSY: over this period, low-capacity areas with higher compliance with JSY had higher mortality than low-capacity areas with lower compliance. Finally, we also find that stimulating demand for institutional delivery had adverse spillover effects on the proportion of children who were up to date with all vaccinations.

Our findings have clear implications for policy. They show that the capacity of a health system to deliver a rapidly increasing amount of incentivized care without reducing the quality of services should be a first-order consideration when deciding whether to adopt demand-side incentives. These considerations are likely to be particularly crucial in determining whether to incentivize the take-up of invasive, time-consuming, and potentially risky procedures such as institutional delivery.

Our paper contributes to several literatures. First, it contributes to the study of demand-side incentives in health, which have proved especially popular policies in LMICs (Kremer and Glennerster, 2011; Dupas and Miguel, 2017). We bring an explicit focus on congestion and offer a flexible general-equilibrium model of the health effects of incentives in the presence of congestion externalities. Demand incentives are generally effective at increasing uptake of health care (Lagarde et al., 2009; Giedion and Díaz, 2010), and therefore, unless there is spare or rising capacity in the system, they will increase congestion. The mixed evidence on the impact of such incentives on health is consistent with the idea that congestion externalities might be key to shaping their effects. For less invasive health interventions, such as the health checkups, vaccinations, or nutrition advice that early conditional cash transfers in Latin America typically incentivized, health benefits have sometimes, but not always, been found (see the review by Lagarde et al., 2009). However, when demand for more invasive procedures has been targeted with no change to the supply side, studies, including previous studies of the effect of JSY, have generally not found health benefits (Giedion and Díaz, 2010; Das and Hammer, 2014; Powell-Jackson et al., 2015; Okeke et al., 2020). Conversely, where the supply side has been improved simultaneously, health benefits have often followed (Miller and Vera-Hernández, 2013; Gruber et al., 2014; Cesur et al., 2017). Our paper contributes by directly examining the role of supply in explaining heterogeneity in how demand-inducing policies affect health. We do so within a single health system and focus on invasive procedures whose riskiness depends directly on quality. Finally, by exploring heterogeneity by case complexity, we provide evidence that congestion externalities, rather than only the private impacts of health care on the marginal cases, are important in limiting the effectiveness of demand-inducing policies in low-capacity areas.

Second, this paper adds to the evidence on the importance of health care quality. In many health systems in LMICs, including in rural India, inputs such as trained health professionals and equipment are often not effectively translated into evidence-based care (Chaudhury et al., 2006; Das et al., 2008, 2016, 2018; Mohanan et al., 2015). Nevertheless, we show that even if existing health professionals and equipment are not used as effectively as they might be, their availability can still be an important predictor of how increased use of health care translates into health outcomes.⁷ Interestingly, we find that capacity in secondary-care health facilities is a very important predictor of the impacts of increasing institutional deliveries on health.

Third, our paper contributes to the literature on congestion externalities in health systems. Overcrowding has been the focus of much attention within the medical literature (Hoot and Aronsky, 2008), and a growing number of papers have also considered these issues from an economic perspective (Freedman, 2016; Hoe, 2022; Marks and Choi, 2019). We focus on India's public health system, which, similarly to public health systems in many LMICs, operates with far fewer resources than the systems in the United States or the United Kingdom, where the empirical economic research on congestion externalities has focused to date. We might expect congestion externalities to play a more prominent role in such resource-constrained settings. Differently from much of the empirical work, which has focused on day-to-day fluctuations in demand (Freedman, 2016; Hoe, 2022; Marks and Choi, 2019), we focus on a context where a high level of congestion is a constant feature of the health system and thus we explicitly model how patients react endogenously to congestion in choosing whether or not to seek health care.

The final literature we contribute to considers the spillover effects from policies that target one aspect of health care onto another aspect, in line with Holmstrom and Milgrom (1991). Previous work has found that when health providers are incentivized in reimbursement contracts to carry out more of a certain procedure, this often comes at the expense of nonincentivized tasks (e.g., Dumont et al., 2008). Our paper is the first to assess whether incentivizing the demand for specific procedures leads to a deterioration in the use or quality of other procedures. We find that it does. Whether such a reallocation is optimal depends on the relative productivity of the different procedures. In this case, however, the reallocation of resources prompted by JSY provided no aggregate health benefits but caused a reduction in childhood vaccinations.

This paper proceeds as follows. Section II sets up our conceptual framework. Section III provides a basic overview of JSY, while section IV discusses our data, choice of sample, and construction of measures. Section V presents our methods and results for assessing the impact of JSY and discusses the different mechanisms. Section VI concludes.

We begin by creating a theoretical model to clarify the different mechanisms through which subsidizing facility delivery might affect birth outcomes. We develop a general equilibrium model of both health and facility choice that allows births of heterogeneous complexities to select into facility delivery endogenously. Crucially, the model allows for congestion externalities whereby an increase in the overall use of facilities increases the risk associated with facility delivery for all births.

Given the model's various mechanisms, it does not provide an unambiguous prediction of whether or not subsidizing demand will improve health outcomes, although it nests a classic Roy model that does give unambiguous predictions. Instead, our goal is to characterize the features of the underlying environment that will shape a subsidy's impact.

Births are indexed by $i$ and may result in either a good health outcome (⁠$Hi=1$⁠) or a bad health outcome (⁠$Hi=0$⁠). We allow births to be heterogeneous in their ex-ante risk, in how much they gain or lose from facility delivery, and in the degree to which they suffer from congestion. We summarize this heterogeneity by allowing births to differ along one underlying dimension: their “complexity,” denoted by $ηi$⁠. Without loss of generality, we assume that $ηi$ is uniformly distributed across the interval [0, 1], with 0 being the least-complex birth and 1 being the most complex.

Women may give birth either at home (⁠$Fi=0$⁠) or in a health facility (⁠$Fi=1$⁠). We let the ex-ante probability of a good health outcome for home birth $i$ be denoted $πh(ηi)$ while the ex-ante probability of a good health outcome from a facility birth is $πf(F/X,ηi)$⁠, where $X$ is capacity, $F≡∫iFidi$ is overall usage, and $F/X$ is congestion. We make the following assumptions over how complexity affects the private costs/benefits from facility birth:

• A1. More-complex births are always more risky than less-complex births that occur in the same location: $dπh(ηi)dηi<0$ and $dπf(F/X,ηi)dηi<0,∀F/X$⁠.

• A2. More-complex births have more to gain (or less to lose) from facility delivery: $∂πf(F/X,ηi)∂ηi-dπh(ηi)dηi>0,∀F/X$⁠.

Our model nests a model with no congestion externality. In that case, we have assumptions A1 and A2 and a “no externality” condition (i.e., $dπf(F/X,ηi)d(F/X)=0∀ηi$⁠). If a congestion externality is present, then we further assume the following:

• E1. Increased congestion increases the risk of facility births: $dπf(F/X,ηi)d(F/X)<0,∀ηi$⁠.

• E2. More-complex births are at least as harmed by congestion as less-complex births: $d2πf(F/X,ηi)d(F/X)dηi≤0$⁠.

• E3. The harmful effects of congestion on the health of facility births are increasing in current congestion: $d2πf(F/X,ηi)d(F/X)2<0$⁠.

Although not formally modeled, we note that a lack of responsiveness to ex-ante risk (i.e., a low $λ$⁠) could be driven by uncertainty over birth complexity rather than families not valuing health outcomes per se.

In choosing how many births to allocate to facilities, the planner will trade off gains to the marginal births with the marginal increase in the congestion externality (the “social marginal cost,” or the dashed line in figure 1), which is experienced by all births already taking place in facilities. The social planner chooses $F$ to equate the social marginal cost from moving an additional birth to the facility with the private marginal benefit to the birth that is moved. This social optimum is marked as point A in figure 1.

In appendix section B.2, we show that $dFds>0$⁠, that is, a subsidy will unambiguously increase the proportion of facility births. The degree to which births of complexity $ηi$ are marginal to the subsidy is described by $dF¯(ηi,F/X,s)ds$ [equation (B.8) in appendix B].

Appendix B.1 gives the full expressions for the total differentials.

In other words, effect (1) is the expected private health cost or benefit to marginal births of complexity $ηi$⁠. It is the difference in ex-ante risk between facility and home births for births of complexity $ηi$ at the current levels of crowding (⁠$F/X$⁠) weighted by the extent to which births of this complexity are marginal to the subsidy. Effect (2) is the congestion externality. It is the reduction in the probability of a good health outcome from facility births of complexity $ηi$ that occurs because of the overall increase in facility births due to the subsidy. This is scaled by the proportion of births of complexity $ηi$ that will be subject to this externality (i.e., the proportion currently taking place in facilities).

Overall, a positive subsidy will be optimal whenever there is a marginal gain in health from moving from a zero subsidy to a marginally positive subsidy, that is, whenever $dH(X,s)ds|s=0>0$⁠. This will depend on the sign and magnitude of the private marginal benefits accrued, and, if these are positive, then whether they are large enough to offset the congestion externality imposed on births currently taking place in facilities. We now discuss each effect in turn.

The sign and size of the marginal private costs or benefits will depend on which births are marginal to the subsidy. The more complex the set of marginal births is, the more the marginal births will gain from being moved into facilities.

As shown by the “high sorting” (blue dashed lines) example in figure 2, a high $λ$⁠, indicating a high degree of selection on gains, implies that the most-complex births already largely occur in facilities even without the subsidy (plot a), which means that they are not marginal to the subsidy (plot b). In this scenario, the subsidy will primarily draw in the next most-complex births out of the births currently taking place at home (plot b).

As $λ$ approaches 1, the model approaches a pure Roy model with complete selection on gains. In this case, without the subsidy, births select into facilities in the order of their complexity (from the most complex to the least complex) until the marginal private health gain to the last birth is zero (marked by point B in figure 1). Even without the congestion externality too many births occur in facilities relative to the social optimum (point A in figure 1). Subsidizing demand in this case unambiguously worsens health. Under perfect selection on gains, the marginal births to an infinitesimally small subsidy are unaffected by whether or not they occur in a facility. A discrete increase in the subsidy (point C in figure 1) brings in births that are harmed by the fact that they are now taking place in a facility. This is on top of any costs from the overcrowding externality.

If births respond only very weakly to ex-ante health risk in selecting into facilities (i.e., if $λ$ is low), then there is a pool of highly complex births that currently take place at home and could gain a lot from being moved into facilities. If the marginal utility of consumption $θ(ηi)$ is independent of $ηi$ (as in the solid line in figure 2), the subsidy will always draw in the births uniformly from across the complexity distribution. However, if the marginal utility of consumption $θ(ηi)$ is increasing in birth complexity (as demonstrated by the dotted line in figure 2), then more-complex births will be more likely to be marginal. Since more-complex births have more to gain from facility delivery, this would result in larger private gains from the subsidy. Such a scenario might arise if complex births are disproportionately from poorer families, who might, for example, have poorer access to prenatal care and a higher marginal utility of consumption.

When there is no sorting on gains, the initial equilibrium is independent of capacity, and higher preexisting capacity will unambiguously mean a subsidy is more beneficial (or less harmful) because higher capacity (1) increases the health gain for marginal births; (2) means the identical absolute increase in facility usage translates into a smaller increase in congestion $F/X$⁠; and (3) means the marginal increase in congestion will be less harmful because higher-capacity areas are starting from a lower initial level of congestion (assumption E3). When there is sorting on gains, including in the extreme case of the Roy model (⁠$λ=1$⁠), how capacity affects the impact of a marginal subsidy will be theoretically ambiguous, as we document in appendix B.3. Thus, this is ultimately an empirical question. Intuitively, the same three factors as listed above go in the direction of a subsidy being more useful, or less harmful, in higher-capacity areas. However, higher-capacity areas will have more births already occurring in facilities, and thus more births will suffer from the increase in congestion.

Our model illustrates that the impact of a demand subsidy for facility delivery is theoretically ambiguous, rendering it a key question for empirical research. It highlights that the impact of a subsidy will be composed of two key effects: first, the benefits/costs to the marginal births; and, second, social costs to all facility births due to the congestion externality.

The framework illustrates that the ways in which health impacts from the subsidy vary across the distribution of birth complexity are informative about whether or not a congestion externality is at play, or whether impacts can be explained by the aggregation of private gains/losses to the marginal births alone. Conditional on being marginal to the subsidy, private benefits are likely to be concentrated on more-complex births and any private costs concentrated on less-complex births. Conversely, any social costs due to a congestion externality will always be concentrated on complex births. This insight is key to how we interpret our findings in section VE.

The model shows that a positive subsidy is optimal if the marginal births benefit from being moved into facilities and the magnitude of this benefit outweighs the externality cost borne by births already occurring in facilities. This is more likely to be the case if facilities are less crowded (more capacity, less usage) to begin with, and if it is more-complex births (who have more to gain from facility delivery) who are marginal to the subsidy.

JSY was launched in April 2005 with the aim of reducing maternal and infant mortality through increasing rates of institutional delivery (Ministry of Health and Family Welfare Government of India, 2005).

Many life-saving procedures can only be carried out in health facilities or by a trained birth attendant (Lawn et al., 2005), which creates many reasons why facility delivery could improve birth outcomes, especially for more-complex births. However, evidence of poor quality in the form of few health professionals compounded by absenteeism, a lack of physical resources, and the presence of dangerous pathogens could all reduce the potential gains from facility delivery or even make it more risky than home birth (Chaudhury et al., 2006; Das et al., 2008; Mohanan et al., 2015; Das et al., 2016). We note, though, that under a congestion externality it could be the case that all births gain privately from facility delivery yet the aggregate impact of increasing facility delivery on birth outcomes is negative.

We focus on JSY's effects on rural households in states that the scheme designated as “Low Performing States” based on their low prior institutional delivery rates. In rural areas of these states, JSY is universal and cash incentives are more generous than in either “High Performing States” or in urban areas. Specifically, JSY provides a financial incentive of Rs.1400, or USD 32, to all pregnant women who give birth in a government health facility.⁹ JSY also provides an incentive of Rs. 600, or USD 14, to community health workers for every pregnant woman they bring to a facility. JSY did not provide new funding to help facilities expand obstetric services (Ministry of Health and Family Welfare Government of India, 2005).

We focus our analysis on the households primarily targeted by JSY: rural households in the nine states designated by JSY as “Low Performing States.”¹⁰ Together, the rural population of these states comprised 33.9% of India's population in the 2001 census and 48.7% of India's deaths of infants under seven days of age between 1990 and 2001.¹¹

Data on perinatal mortality, defined as stillbirth after 22 weeks of pregnancy or death within 7 days of birth, come from the 2007–2008 DLHS-3 pregnancy roster for ever-married women, which recorded all pregnancies since January 1, 2004. We focus on perinatal mortality given its particular sensitivity to the quality of care during the birth (World Health Organization, 2006). As our sample, we use all births (i.e., live births and stillbirths but excluding miscarriages and abortions) that occurred between January 1, 2004 and December 31, 2007 and within nine quarters of JSY's rollout in the district as our sample. We thus drop the 36 districts where JSY was not rolled out until after quarter 4 of 2007, which leaves us 256 districts. Table 1 provides sample descriptives. Place of birth is available for each respondent's most recent birth, and thus effects here should be interpreted as averages for this subpopulation.

We construct measures of preexisting capacity in each district using three inputs that are easy to measure and over which national guidelines exist: (i) beds, (ii) doctors/medical officers, and (iii) nurses/midwives. We distinguish between inputs in primary-care and secondary-care facilities.¹² We use the facility survey of the DLHS-2, which, in 2003, surveyed government health facilities in 182 of the 256 districts we focus on in this paper. These data cover all secondary-care health facilities and a random sample of primary-care health centers (PHCs). We estimate the number of beds, doctors, and nurses/midwives per 10,000 of the rural population in each district's primary- and secondary-care facilities using population figures from the 2001 census. Further details and density plots of these measures are given in figure A1.

We use exploratory factor analysis to assess the dimensionality of these six measures and to create summary indices. We first run a single factor analysis on all six measures, which estimates that there are two orthogonal factors with eigenvalues greater than 1 (the Kaiser criterion for retaining a factor) which implies that these measures can be well summarized by two underlying factors: one primarily reflecting secondary capacity and the other primary capacity (factor loadings in columns 1 and 2 of table A1). We next run two separate factor analyses for the primary and secondary capacity measures to create summary indices for each without imposing orthogonality. These are the indices that we use in our analysis, and columns 3 and 4 show the loadings. In our main analysis, we use an indicator of whether the district had above- or below-median preexisting capacity in the primary and/or secondary system based on these factor measures. We use the continuous measures in robustness analysis.

We estimate JSY's rollout from the DLHS-3. Each respondent was asked whether they had received a payment through JSY or another state-specific scheme for their last birth. Once JSY was rolled out in a district, all births in government institutions should have received payments. We define JSY as being active in a district from the first quarter (after its official launch) in which 25% of births that occurred in government institutions were reported to have received a JSY payment in both that quarter and the following year.¹³ There is strong variation in the timing of implementation within and between states (figure A2). In robustness analysis, we use an alternative, fractional measure of JSY intensity—the proportion of eligible births that received the payment (see figure A3).

As we hope to use the rollout of JSY to identify causal effects, we now examine how the rollout relates to longer-run trends in our outcomes of interest. A correlation between the rollout and already-existing trends could suggest that parallel-trends assumptions required for our differences-in-differences approach to be valid might not be plausible. We use the DLHS-2 to calculate district-level changes in rates of institutional delivery, 7-day infant mortality, and vaccinations between 1990 and 2001.¹⁴ We then run district-level regressions of JSY's start date on these changes interacted with the prior capacity of the district (table A2). Columns 1, 4, and 7 show that the start date of JSY is uncorrelated with long-run changes in these variables, confirming the finding of Powell-Jackson et al. (2015) that the rollout appears uncorrelated with prior trends. The other columns of table A2 show that the start date of JSY is correlated with capacity in the primary- and secondary-care sectors, but not with the interaction of preexisting capacity with the long-run changes. Figures A4, A5, and A6 in appendix A provide a more detailed analysis and find no evidence that districts where JSY rolled out at different points were on different trends before the program.

We do not have a large “always untreated” sample, as JSY had rolled out to almost all districts by the end of our sample period. Therefore, we must restrict two lags to identify the model (Borusyak et al., 2021); we thus set $β-9$⁠, the first event period, and $β-1$⁠, the period immediately before JSY's rollout, to zero. The specification allows us to test for differences in nonlinear pretrends—by testing the null hypothesis $H0:β-8=⋯β-2=0$—but not for differences in the linear component of pretrends (Borusyak et al., 2021). This test complements our earlier analysis, in section IVC, which suggested that the rollout was unrelated to longer-run trends in the outcomes of interest.

Our identifying assumption is that, conditional on district, quarter-of-birth, and birth order, birth-specific shocks are mean independent of JSY's rollout, that is, $E[νibdt|1{Kibdt=k},θb,θd,θt,∀k={-8,...,-2,0,...,9}]=0$⁠. This is a multiperiod parallel-trends assumption and rules out that the rollout of JSY was related to birth-specific shocks but does not rule out that the rollout was related to birth-specific gains from JSY. Under this assumption, and the assumption of no systematic heterogeneity in treatment effects across cohorts (Sun and Abraham, 2021), $βk$ is the average causal effect of JSY $k$ periods after it began.¹⁵

where $Cd$ is a vector containing measures of the capacity of district $d$ prior to the rollout and where we again impose effects at $k=-9$ and $-1$ to be zero. The identifying assumption here is $E[νibdt|1{Kibdt=k},Cd1{Kibdt=k},θb,θd,θt,∀k={-8,...,-2,0,...,9}]=0$⁠. This rules out that the rollout within districts with the same preexisting capacity level was related to birth-specific risks once conditioning on quarter of birth and birth order. In our main analysis, $Cd$ is simply a binary indicator for district $d$ having above-median preexisting capacity; in this case, $βk$ represents the effect of JSY in low-capacity districts $k$ periods after it began and $βk+γk$ represents JSY's effect in high-capacity districts.

where $JSYdt$ is an indicator for whether JSY had begun in district $d$ and quarter $t$⁠, that is, $JSYdt=1{Kibdt≥0}$⁠. We estimate the identical model allowing for an interaction between treatment terms and preexisting capacity to examine heterogeneity.

When treatment effects vary with time, this approach is preferable to a two-way fixed effects regression, which recovers a weighted average of treatment effects that places negative weight on longer-run effects (Borusyak et al., 2021). To increase precision, we estimate dynamic effects under the restriction that the coefficients on the lags of JSY's rollout, $β-8$ through $β-2$⁠, are zero after first testing this restriction.

Next, we estimate the average of these dynamic effects, weighted according to the event-time distribution of our sample [equation (11)]. Overall, JSY increased the probability of institutional delivery by 7.86 percentage points (⁠$p<0.001$⁠), or 35% of the mean before the policy's launch, across all sample districts (see table 2, panel A, column 1) and 7.62 percentage points across districts with capacity data available (column 2). These effect sizes are in line with those found by Powell-Jackson et al. (2015). The increase appears similar in districts with more and with less prior capacity (columns 3–5).

As we see in table A3 in appendix A, breaking this down into different types of institutional deliveries, we see that this increase was comprised of an increase of 6.30 percentage points (⁠$p<0.001$⁠) in the probability of births occurring in government-run secondary-care facilities, an increase of 3.14 percentage points (⁠$p<0.001$⁠) for government-run primary-care facilities, and a decrease of 1.58 percentage points (⁠$p=0.003$⁠) for private facilities (in most states, payments were not given for births in private facilities).

The magnitudes of these increases are large relative to the prior rate of institutional delivery and the capacity of the government health system. A back-of-the-envelope calculation suggests that an increase in the proportion of births occurring in government facilities (secondary and primary) of 9.44 percentage points (table A3) translates into an increase in the average number of births per day occurring in each government-run secondary-care facility, or in a government-run primary-care facility under the supervision of that secondary facility, from 1.73 before JSY to 3.59 afterwards. Even though the increase in institutional delivery was the same across districts with more and with less capacity in the secondary-care system, the difference in the number of facilities means the increase in caseload was more pronounced in districts with low preexisting secondary-care capacity (from 1.92 to 4.80) than with higher capacity (1.54 to 2.38).¹⁶

We now move on to assess impacts on perinatal mortality. In line with Powell-Jackson et al. (2015), we find that the average effect, across districts with high and low preexisting capacity, of JSY is close to zero and not statistically significant; this holds both in terms of the dynamic effects (figure A8a) and when averaged over event time (table 2, panel B, columns 1 and 2).

However, the results are markedly different in districts with below-median preexisting capacity in the secondary health care system from districts with above-median capacity. In the former, low-capacity districts, figure 4b shows that JSY increased the risk of perinatal mortality, with effects growing somewhat in magnitude over time. The average impact across event times (table 2, panel B, column 3) is an increase in the risk of perinatal mortality of 0.90 percentage points (⁠$p=0.020$⁠). In contrast, in districts with above-median secondary capacity, JSY had no impact at any event time or averaged across event times (⁠$p=0.811$⁠), with the difference in the average effect between these above-median capacity districts and the districts with low capacity being statistically significant (⁠$p=0.022$⁠). These magnitudes are substantial compared to the rate of perinatal mortality of 3.70% in the 15 months before JSY's launch.

Although point estimates for the effect of JSY of perinatal mortality in districts with low primary-care capacity are positive (an increase in mortality risk), these are not statistically significant (table 2, panel B, columns 4 and 5). Neither is the difference in effects between districts with high and low primary-care capacity.

In appendix A, we explore the robustness of our results. We show results using a static two-way-fixed-effects specification, that is, $Yibdt=α+β1{Kibdt≥0}+θb+θd+θt+νibdt$⁠. Effect sizes for perinatal mortality (table A4) are nearly identical to the weighted average of the dynamic effects, which is unsurprising as the dynamic effects do not vary substantially with time (figure 4). For institutional delivery, the static specification produces smaller, but still highly statistically significant, estimates (table A5), which is expected given that, as shown in figure 4, the dynamic effects increase with time (Borusyak et al., 2021). We further show that results are robust to using a probit model (tables A6 and A7) and to using a fractional indicator of JSY's rollout (table A8 and A9), to using continuous measures of preexisting capacity (tables A10 and A11), and to allowing for differential time trends by preexisting capacity (tables A12 and A13). In table A14, we show that we find no evidence of heterogeneity in the impact of JSY on perinatal mortality across districts that adopted JSY earlier and later.

Given that health system capacity was not randomly assigned, heterogeneity by capacity could, theoretically, be capturing heterogeneity across distinct but correlated characteristics of districts (see table A15). We mostly rely on our conceptual framework to motivate why system capacity is a natural dimension over which heterogeneity would arise. However, we also use LASSO penalized regression to analyze which district characteristics predict heterogeneity in the impact of JSY on perinatal mortality. We allow for heterogeneity to be predicted by district-level measures of poverty, men's education, women's education, distance from different health facilities, and our two measures of system capacity. Figure A9 shows that at all levels of penalization, the standardized coefficients on both measures of capacity are greater than all alternative predictors. Moreover, LASSO sets coefficients on all other predictors to zero well before setting either of the capacity coefficients to zero. This reassures us that the heterogeneity we have recovered is not simply standing in for heterogeneity across other dimensions.

We have shown that JSY increased the perinatal mortality risk in areas with low preexisting capacity in the secondary-care system. Our theoretical model demonstrates that two mechanisms could have led to this. Even in the absence of any congestion externality, JSY could harm health if the marginal births moved by the subsidy were riskier when taking place in a facility rather than at home. As discussed in section III, if facilities are poor quality, this is not implausible. However, our model also suggests a second mechanism: that the overall increase in facility delivery rates increased congestion in facilities, leading to a deterioration in the quality of medical care for all facility births, including those that would have taken place in facilities even in the absence of the subsidy.

In this section, we explore the mechanisms at play. We begin by looking at how the impacts of JSY differed by birth complexity as an attempt to shed light on the role of a congestion externality in driving the negative effects of JSY. To this end, we create a proxy for birth complexity by using variables containing information that would have been known to families at the time they were choosing where to deliver: complications experienced during pregnancy, the number of previous pregnancies, whether the woman had previously had a miscarriage, whether she had previously had a stillbirth, and whether she had previously given birth to a baby who subsequently died. We use these variables, and their interactions, to predict perinatal mortality in the full sample for which this information is available (the sample of last births). To do so, we used penalized LASSO regression to avoid overfitting. We then predict a risk index using the chosen predictors. See table A17 in appendix A for details. We divide births into those with a risk index above and below the median.

In column 1 of table 3, we estimate the impact of JSY on institutional delivery using methods identical to those we employed earlier but now interacting the secondary health system capacity with our proxy for birth complexity in addition to JSY's rollout. The first important thing to notice here is that even in the absence of JSY, in both low-capacity (line e) and high-capacity areas (line f), more-complex births were more likely to occur in facilities than less-complex births, suggesting that births select into health facilities at least partially on this basis of their expected complexity. While this suggests that, prior to JSY, births sorted into facilities somewhat based on gains, lines (a)–(d) of column 1 show that JSY drew in more-complex and less-complex births to the same degree, and this was true in both low- and high-capacity areas.¹⁷ We argued in section II that for JSY to stand the best chance of being effective at improving health outcomes, it would have to disproportionately induce the most-complex births into health facilities. We see here that this does not seem to have been the case.

Next, we turn to how JSY's impacts on perinatal mortality differed by births' complexity in column 2. In rows (a) and (b) we see that practically all the increase in perinatal mortality in low-capacity areas was driven by more-complex births. We estimate that the perinatal mortality risk of these complex births in low-capacity areas increased by 1.59 percentage points due to JSY, while less-complex births saw no change. The difference between the two is highly statistically significant. Interestingly, we even see tentative evidence that in higher-capacity areas too, more-complex births may have suffered due to JSY, although to a lesser extent than in low-capacity areas. The fact that the harms of JSY appear to be concentrated on more-complex births despite JSY having moved more-complex and less-complex births into facilities at the same rate suggests that a congestion externality is key to explaining JSY's perverse effects.

To see this, first imagine the special case of our model with no congestion externality. In this case, all of the aggregate harms due to the subsidy are made up of private harms to the marginal births. In our model, conditional on moving the same proportion of less-complex and more-complex births into facilities (which we see empirically), the more-complex births would privately gain more or be harmed less. Therefore, the fact that harms appear concentrated on complex births is inconsistent with the version of our model without a congestion externality.

However, our model predicts that under sorting on gains (which we see empirically), the overall cost of the externality will be disproportionately borne by complex births. This is the case for two reasons. First, a disproportionate number of complex births will already be occurring in facilities and thus will be subject to the externality. Second, we anticipate (as set out in assumption E2) that congestion will harm more-complex facility births the most. Thus, the fact that we find clear evidence that the harms associated with JSY are concentrated on complex births is evidence that a congestion externality is crucial to understanding our results.

Besides this indirect evidence on the importance of congestion externalities, the results in columns 3 and 4 provide more direct evidence on congestion. Here we look at whether respondents reported that their baby received a checkup after their facility birth. Note that we condition on births occurring in facilities here as it would be surprising if moving births into facilities did not result in some increase in the proportion reporting a check from a health professional;¹⁸ instead, we want to focus on the quality of care provided by facilities. In column 3, we look at whether or not respondents reported that their baby received a checkup within the first 24 hours. Overall, we do not see evidence that JSY changed the likelihood of this, which, given that even a very cursory checkup from a health worker would be counted here, is perhaps not surprising. In column 4 we look at whether babies received a checkup between days 2 and 10. We take this as an indicator of how long women remain in facilities after birth, and the intensity of the continued care they receive (both in facilities and in the community). Here we find that in low-capacity areas (but not in high-capacity areas) JSY reduced the probability of such checkups by around seven percentage points (lines a and b). We interpret this as evidence that in low-capacity areas, facilities reacted to congestion by either discharging women and babies earlier or by providing less continued care. Furthermore, the fact that this reduction was equal for more-complex (line b) and less-complex births (line a) suggests that facilities did not prioritize maintaining more-intensive care for more-complex births in reacting to this congestion.

We employ the same empirical strategy to examine spillover impacts of JSY on another service that the local health system performs: childhood vaccinations. Our outcome here is the proportion of vaccinations a child has received of those that a child of the same age should have received.¹⁹ Given that polio vaccinations are largely administered by a parallel system, the “Pulse Polio Initiative,” we focus on the six nonpolio childhood vaccinations that are delivered through the public health system.

In Table 4, we find that, averaged across event time, JSY reduced the proportion of the recommended vaccinations that children aged 9 months or older had received by 2.05 percentage points (⁠$p=0.007$⁠). Breaking this down by the six nonpolio recommended vaccinations, we find that JSY significantly reduced the probability that children received four of these recommended vaccinations. Interestingly, the negative effect of JSY on vaccination rates is smaller in vaccines given closer to birth than in those given later. This is consistent with JSY reallocating resources and reducing vaccination rates, but this is offset for vaccines given at an early age because JSY also led to an increase in institutional deliveries, which could have increased parents' attachment to health care providers right after birth.

Reductions in vaccination rates do not vary with preexisting capacity (figure A10, table A18).²⁰ In appendix A, we show these findings are robust to using a static specification (table A19), to using a probit model (A.20), and to using a fractional indicator of JSY intensity (A.21). In appendix C, we argue that these results are not driven by differential selection caused by JSY's impacts on mortality.

We have provided evidence that in the short run JSY caused an increase in perinatal mortality in areas with low preexisting rates of capacity in the secondary-care system. A crucial question is whether this was a short-term effect, in which case the impacts could have been due to transition difficulties, or whether JSY's negative impacts persisted. Here, we briefly assess whether there is any evidence of persistence. To do this, we draw on the birth record of the fourth National Family Health Survey (2014–2015). The NFHS-4 records detailed information, including JSY uptake and place of delivery, of all births since 2010 in addition to mortality data for a longer history of births.

In the short term, we used the staggered rollout of JSY across districts to estimate its effects. However, by 2010 JSY had been operational for upwards of 4 years in most districts, making this approach less plausible. Instead, we explore how variation in JSY compliance between districts correlates with variation in both institutional delivery rates and perinatal mortality rates over the medium run. Specifically, we create an indicator for whether or not 90% or more of births that should have been eligible for a JSY payment (i.e., births that occurred in a government facility) between 2010 and 2015 reported receiving one. We use the identically constructed measures of pre-JSY health system capacity.

Beginning with rates of institutional delivery in table 5, we find that, after controlling for quarter-of-birth fixed effects, districts where compliance was greater than 90% had institutional delivery rates of 4.5 percentage points greater than in districts with lower compliance (column 1). Note that because we have no historical data from this data set on institutional delivery, we cannot control for district fixed effects.

We then proceed to analyze the medium-run effects on perinatal mortality. Column 2 shows that districts with 90% or higher JSY compliance had perinatal mortality rates that were 0.71 percentage points higher than districts with lower compliance (controlling for time trends). Columns 3–5 show that, consistent with our short-term results, this result is driven by low-capacity areas, although whether the difference is statistically significant depends on the precise specification. Note that these results are subject to some important limitations. In columns 3 and 4, where we only use births from 2010 onwards, the sample size is much smaller than in table 3 and we cannot include district fixed effects. In column 5, we append the data with births between 2000 and 2005, which enables us to include district-level fixed effects, although this assumes that the unobserved district-level components remain constant over that long period, which is a strong assumption.

While the extent to which different districts complied with JSY over the medium run may not be exogenous for many reasons, these associations are certainly consistent with JSY continuing to have perverse effects on perinatal mortality up to 10 years after it was first introduced. The associations are perhaps most convincing at suggesting that it is unlikely that the short-term harmful impacts of JSY in low-capacity areas turned beneficial over the medium run. In many ways, this seems unsurprising given that since JSY's introduction in 2005, India's public spending on health care has remained unchanged at less than 1% of GDP, far less than other rapidly growing economies. Low and steady levels of public expenditure since JSY's introduction perhaps gave little scope for a rapid improvement in the capacity of the supply side over this period.

Conditional cash transfer programs are frequently used to increase the uptake of health services, but evidence on their impact on health is mixed (Lagarde et al., 2009). In this paper, we bring to the fore the role of congestion externalities in explaining the detrimental effects of JSY on perinatal mortality in areas of India that have a low-capacity health care system. The overall impact of JSY on health is composed of both its effect on inframarginal facility births (those that would have taken place in a facility regardless of JSY) and its effect on marginal births (those births induced to take place in the facility because of JSY). Under a congestion externality, the effect on the inframarginal births will always be detrimental due to increased congestion. Hence, for JSY to improve outcomes, any positive effect on marginal births must outweigh the congestion externality.

The theoretical model highlights interesting interactions between the size of the congestion externality and what drives selection into institutional delivery. Under full selection on gains, the inframarginal facility births are the riskiest and, consequently, they are the most harmed by the additional congestion caused by JSY. Moreover, under full selection on gains, the births that are marginal to JSY are harmed from switching from home to facility delivery. Hence, in this case, JSY will have detrimental effects on health. On the contrary, if the initial selection into health facilities is unrelated to health gains, then there could be a large pool of potentially marginal births that could gain a lot from facility delivery. Furthermore, if more complex births are more responsive to the subsidy, this would give JSY a better shot at attracting the births that would particularly gain more from facility delivery.

Unfortunately, our empirical analysis reveals that JSY was not successful at attracting disproportionately complex births; more- and less-complex births were similarly attracted Moreover, we find evidence that in low-capacity districts, JSY generated additional and harmful congestion. In these districts, JSY decreased the likelihood that babies born in facilities received a postnatal checkup after the first day, which suggests that they were discharged sooner or simply that checkups were overlooked. We show that the detrimental effects of JSY were concentrated on complex births, a result that in our model can only be explained by the presence of a congestion externality. Overall, JSY caused perinatal mortality to increase in low-capacity districts, and more tentative evidence suggests that these harmful effects may have persisted for up to 10 years.

Our general equilibrium framework points to alternative approaches on how to use public funds to improve health outcomes when congestion externalities are a concern. First, policies that help identify complex cases and direct those cases toward health facilities could be particularly useful. Our framework demonstrates that, for a given degree of capacity and usage, increasing the complexity of the cases that select into formal health care will improve overall health. While there are often important barriers to patients discerning the usefulness of a given health care intervention and this is a fundamental barrier to individuals seeking out the optimal amount of care (Arrow, 1963), good-quality prenatal care could help direct more complex births towards facilities.

Second, our framework highlights the potential of policies that effectively increase the capacity of health care systems and the quality of care they provide, including both policies that increase the resources available and those that ensure those resources are used efficiently (Dupas and Miguel, 2017; Kremer and Glennerster, 2011; Björkman and Svensson, 2009). Our framework suggests that such efforts are likely to improve outcomes for existing users of health care and to attract more individuals to take up care. Also, our empirical results highlight the stakes of such efforts to improve quality, especially as efforts intensify to ensure universal access to health care. We found that in areas with few of the most basic health care inputs—beds, nurses, and doctors—increased use of health care harmed overall health.