Risk Protection, Service Use, and Health Outcomes under Colombia’s Health Insurance Program for the Poor

Abstract

Unexpected medical care spending imposes considerable financial risk on developing country households. Based on managed care models of health insurance in wealthy countries, Colombia’s Régimen Subsidiado is a publicly financed insurance program targeted to the poor, aiming both to provide risk protection and to promote allocative efficiency in the use of medical care. Using a “fuzzy” regression discontinuity design, we find that the program has shielded the poor from some financial risk while increasing the use of traditionally under-utilized preventive services – with measurable health gains.

In developing countries, the inability to smooth consumption directly reduces welfare and leads to informal risk management strategies that stifle productive activity (Paxson 1993, Townsend 1994, Morduch 1995). Because unexpected illness is a leading source of economic risk, the expansion of health insurance is therefore a public policy priority in many parts of the developing world (Gertler and Gruber 2002, GTZ, WHO, and ILO 2005, WHO 2010, Mohanan 2012).¹ Such initiatives are often large, centrally-planned programs operated exclusively through the public sector – and they focus primarily on reducing the out-of-pocket price of medical care.²

An early exception is Colombia’s Régimen Subsidiado (or “Subsidized Regime,” henceforth “SR”). Introduced in 1993, the SR is a pluralistic, publicly-financed health insurance program targeted to the poor.³ Colombians meeting a proxy means-test (determined by the Sistema de Identificación de Beneficiarios, or SISBEN) are fully-subsidized to purchase insurance from private, government-approved insurers. In contrast to the classical ‘managed competition’ model of insurance (Enthoven 1978a and 1978b), participating insurers must offer standardized benefits packages and accept standardized premiums. Insurers can, however, form restrictive medical care networks, deny reimbursement for services deemed ‘unnecessary,’ and pay health care providers in ways that encourage higher quality and lower cost medical care (through capitated payment contracts, for example – fixed payments per enrollee per month). Overall, Colombia’s SR shares features of managed care models of health insurance that emphasize cost-containment and allocative efficiency traditionally found only in wealthy countries. It therefore represents important early experience in a markedly different institutional environment.

This paper studies the impact of the SR on financial risk protection, service use, and health outcomes among Colombia’s poor. Program eligibility is supposed to be determined according to a discrete threshold in the continuous SISBEN index, so in principle we could use a regression discontinuity design to do so. Because SISBEN scores are manipulated in practice (BDO and CCRP 2000, DNP 2001, DNP 2003a, and 2003b, Fresneda 2003, Camacho and Conover 2011), we instead use underlying index components collected through independent household surveys to generate our own (un-manipulated) SISBEN score calculations. We then instrument for SR enrollment with our re-constructed eligibility measure (Hahn, Todd, and Van der Klaauw 2001). In general, our estimates are robust across a variety of parametric and non-parametric specifications.

Despite our strategy for addressing manipulation of eligibility, our approach has limitations. First, because we use household surveys to implement a “fuzzy” regression discontinuity design, our samples are relatively small. Second, we measure SISBEN index components after official SISBEN classification occurred, presumably resulting in a degree of measurement error. Third, due to financial shortfalls, many of Colombia’s municipios (hereafter, “counties”) used eligibility thresholds that fell short of the official one. Following Chay, McEwan, and Urquiola (2005), we therefore estimate and use county-specific thresholds. These limitations introduce noise into the regression discontinuity design and generally bias us against finding behavioral responses to the SR.

We first find evidence that by the mid-2000s, the SR succeeded in protecting poor Colombians from financial risk associated with the medical costs of unexpected illness. In particular, SR enrollment appears to have reduced the variability of out-of-pocket spending for inpatient care. Despite this reduction in risk, however, we observe little evidence of meaningful portfolio choice effects (changes in the composition of household assets, human capital investments, or household consumption), perhaps because the SR falls short of providing full insurance.

Our results also suggest that SR enrollment is associated with large increases in the use of traditionally under-utilized preventive services – some of which nearly doubled. Moreover, we find evidence of health improvement under the SR as well – specifically, gains along margins sensitive to the increases in preventive care that we observe. There is more mixed evidence of changes in the use of curative services (although theoretical predictions about the use of curative care are ambiguous).

We conclude by discussing the underlying behavioral mechanisms that may explain our results. Because the SR is complex and multi-faceted, it is important to note that we cannot draw firm inferences about them; we emphasize this as an important direction for future research. Overall, we highlight two mechanisms that we suspect are important: high-powered supply-side incentives and the possibility that enrollees receive care from higher-quality private sector facilities.

I. Background and Policy Context

A. Public Sector Health Insurance for Colombia’s Poor Prior to the Reform

Prior to the introduction of the SR in 1993, roughly 25% of Colombians (a subset of those with formal sector jobs) had any form of explicit health insurance (Pinto 2008). However, Colombians lacking formal insurance also had a degree of implicit insurance provided through the public sector. Specifically, they could receive medical care from public sector hospitals and clinics for a fraction of the full cost of their services; out-of-pocket payments were generally progressive and loosely based on socio-economic status. Public sector facilities, in turn, covered their losses with direct transfers from national and local governments, and health care professionals were typically paid fixed salaries that did not reward productivity. Thus, poor Colombians effectively had a degree of implicit health insurance coupled with inefficient provider incentives – and the reform that we study aimed to expand coverage while improving efficiency.

B. Overview of Colombia’s Subsidized Health Insurance Regime for the Poor

Under Law 100 in 1993, Colombia introduced the SR, a novel form of publicly-financed health insurance for the poor (Gwatkin, Wagstaff and Yazbeck 2005, Escobar 2005). Primarily through SR expansion, formal health insurance coverage in Colombia grew from about 25% of the population in 1993 to 80% in 2007 (CENDEX 2008). The SR is organized as a variant of the classical ‘managed competition’ model (Enthoven 1978a and 1978b). Beneficiaries are fully subsidized to purchase health insurance from competing health plans. During our study period (the mid-2000s), subsidies were financed by a combination of public resources including payroll taxes and national and local general revenue. These resources are transferred to county governments, which in turn are responsible for eligibility determination, enrollment, and contracting with health plans. This role of local government has important implications for our empirical analysis.

Health plans charge county health authorities a government-regulated premium per enrollee and offer a standardized package of benefits (see Appendix A for the details of these benefits).⁴ Participating health plans then act as group purchasers of health services for their enrollees by contracting with a network of health facilities and clinicians. Because premiums and benefit packages are standardized by law (unlike the classical ‘managed competition’ model), health plans compete for enrollees on the basis of provider networks and service quality. In practice, however, very few cities had more than one insurer during the years that we study.

Insurers’ ability to contract with health care providers (hospitals and medical groups) for more efficient service provision is a central innovation of the SR. Insurers receive premiums (fixed payments per enrollee per unit time) for all covered services, giving them strong incentives to constrain total spending. Insurers then (partially) transmit these incentives to provider organizations through capitated payment contracts, and they have the authority to deny reimbursement for services deemed inefficient.⁵

In the next section we describe the major demand- and supply-side incentive changes under the SR, and Section II then provides predictions based on them for health service use, out-of-pocket spending on medical care, and health outcomes.

C. Changes in Demand- and Supply-Side Incentives under the SR

There were important changes in both demand- and supply-side incentives with the introduction of the SR. On the demand-side, out-of-pocket prices for covered curative services are 10% of the full price (as opposed to 30% for poor uninsured Colombians).⁶ Moreover, total out-of-pocket spending per episode of illness each year is capped at half of the monthly minimum wage for SR beneficiaries – while the same cap is six times as large for uninsured patients (Decreto 2357 and Acuerdo 260 of the Consejo Nacional de Seguridad Social en Salud). Notably, preventive services are essentially free regardless of insurance status.

On the supply-side, changes in provider payment contracts created substantially different incentives for the provision of health services. Specifically, there are two types of contracts between insurers and provider organizations under the SR: capitated primary care contracts and fee-for-service specialty care contracts. For primary care, insurers pay providers fixed amounts per month for all services used by enrollees (“capitation”). These contracts create strong incentives for providers to constrain total spending on primary care and have important implications for both preventive and curative health care (as discussed in the next section). Importantly, promotion of preventive care can be a central means of reducing total expenditures.⁷

For specialty care, insurers generally pay providers a pre-determined fee for each covered service that they supply (i.e., on a “fee-for-service” basis). These contracts encourage the provision of all reimbursable services (both efficient and inefficient). However, SR insurers also have the authority to deny reimbursement on a case-by-case basis for inefficient specialty care (termed “utilization review”), allowing them to limit wasteful service use.⁸

D. Eligibility for the SR

Eligibility for the SR is determined using a poverty-targeting index called SISBEN (or Sistema de Identificación de Beneficiarios). The original SISBEN index consisted of fourteen components measuring different aspects of household well-being (such as housing material, access to public utilities, ownership of durable assets, demographic composition, educational attainment, and labor force participation – for a complete description, see Appendix B).⁹ On each dimension, households are classified according to mutually exclusive, collectively exhaustive categories with varying weights assigned to each category; these weights vary between urban and rural areas. A household’s SISBEN score is then calculated by summing points across components. Possible scores range from 0 to 100 (with 0 being the most impoverished) and are divided into six strata. Households scoring in SISBEN strata 1 and 2 (the lowest strata) are eligible for the SR (below 48 in urban areas, below 31 in rural areas).¹⁰

E. Eligibility and Enrollment in Practice

Although eligibility for the SR increases the likelihood of enrollment, neither one necessarily implies the other for at least three reasons: misclassification or manipulation of SISBEN scores, shortfalls in local government revenue, and enrollment that preceded SISBEN enumeration.¹¹

First, both local governments and households have incentives to manipulate SISBEN scores. Local governments receive fixed transfers from the national government for each resident they enroll, creating incentives to maximize enrollment. The selective enrollment of key constituents can also provide political benefits (Camacho and Conover 2011). Households prefer enrollment over “uninsurance” as well because co-payments are lower for SR beneficiaries than for those lacking formal insurance. Consistent with both types of incentives, there is evidence of considerable SISBEN score manipulation between 1997 and 2003 (Camacho and Conover 2011).¹²

Second, most local governments lack sufficient revenue to finance the enrollment of all eligible residents. According to law, those with lower SISBEN scores and those belonging to specific targeted groups (such as children under five and pregnant women) are therefore prioritized for enrollment.¹³ This means that many counties use de facto eligibility thresholds that fall below the uniform national threshold.

Third, some counties began enrolling residents in the SR before all of their residents had been classified using SISBEN. These counties instead used other means-test criteria such as residents’ estrato, a neighbourhood-level measure of socio-economic status used to establish electricity prices paid by local households.

In general, these realities of the program’s implementation have two broad implications for our empirical analyses (which we discuss in greater detail in Section IIIB). One is the necessity of an empirical strategy that addresses manipulation of SISBEN scores. The other is that de facto county threshold must be estimated (because they are not officially reported).

F. Previous Studies

Several previous empirical papers have studied Colombia’s SR as well (see Giedion and Uribe (2009) for a comprehensive review). Trujillo, Portillo, and Vernon (2005) use propensity score matching to compare SR enrollees to observationally similar uninsured Colombians, finding that the SR is associated with greater medical care use. Giedion et al. (2009) also use propensity score matching and find that SR enrollment is associated with more visits to health care providers, higher vaccination coverage rates, and fewer reports of not seeking medical care for financial reasons. Gaviria, Medina, and Mejía (2007) instrument for SR enrollment using length of time a household head lived in his/her current place of residence and report that SR enrollment is associated with better self-reported health, more preventive and curative outpatient care, and fewer hospitalizations. Finally, matching official SISBEN score data with birth records, Camacho and Conover (2008) use a regression discontinuity design in two samples: a county lacking clear evidence of manipulation and selected counties observed shortly after SR implementation. They find that SR enrollment is associated with increased birth weight and better APGAR scores but not antenatal care use, medical supervision of deliveries, or probability of hospital delivery. Relative to previous studies of the SR, our paper employs a different strategy for addressing manipulation of program eligibility, uses an approach that disentangles the role of the SR from other public programs using SISBEN, and studies a broader range of behavioral responses to health insurance.¹⁴

II. Empirical Predictions on the Interaction of Supply- and Demand-Side Incentives

Although we emphasize the centrality of changes in supply-side incentives under the SR, out-of-pocket prices for covered services also fell. This section therefore considers the joint effect of supply- and demand-side changes in formulating predictions about changes in risk protection, the use of preventive and curative services, and health outcomes. Section IV then provides empirical evidence on each.

Risk Protection and Out-of-Pocket Medical Spending

Protecting households against medical care costs associated with unexpected illness is a central objective of health insurance. As Appendix A shows, the SR covers many of the most expensive services including trauma care, dialysis, and major joint replacement (hips and knees, for example). It also caps total out-of-pocket spending per episode of illness each year at half of the monthly minimum wage. The SR should therefore reduce the variability of out-of-pocket medical spending. Assuming that the price elasticity of demand for medical care is less than one, the SR should also reduce total out-of-pocket spending for medical care.

Preventive Health Services

Most preventive services in Colombia are free regardless of insurance status (i.e., free both for SR enrollees and the uninsured), so the main difference in incentives for preventive care originate on the supply-side. Primary care providers are paid on a capitated basis, so they have strong incentives to limit total primary care spending. Increasing preventive service use can reduce the need for more costly curative care – and can therefore be cost-saving – so capitated primary care incentives are likely to increase preventive care use.¹⁵

An opposing force at work is ex ante moral hazard: SR enrollees may have weaker incentives to use preventive services because they pay less for curative care (Nordquist and Wu 1976, Phelps 1978). Although we fail to find evidence of ex ante moral hazard (as shown in Section IVE) and there is little evidence of it in other studies (Kenkel 2000), our preventive care estimates reflect the net effect of these forces.¹⁶ Because preventive services generate important positive externalities (both pecuniary and infectious disease-related), increases in their use presumably improve welfare.¹⁷

Curative Medical Care

Three changes under the SR are likely to influence the use of curative medical care: reductions in out-of-pocket prices, increases in the use of preventive services, and supply-side incentives for limiting total medical spending. First, for negative price elasticities of demand, reductions in out-of-pocket prices will increase the use of curative care. This increase is inefficient under standard assumptions (ex post moral hazard); however, curative care in developing countries may produce positive externalities, and there may also be important credit constraints. Second, any increases in preventive service use may reduce the use of curative care (an efficient result). Third, both capitation and utilization review produce incentives for providers to limit the use of curative medical care. Some reductions may be efficient (counterbalancing demand-side incentives leading to ex post moral hazard) while others may be inefficient (leading to “stinting” – see for example Ellis and McGuire (1990), Holmstrom and Milgrom (1991), Ma and McGuire (1997), and McGuire (2000)). Predictions about changes in curative care under the SR as well as their welfare implications are therefore ambiguous.

Health Outcomes

Both preventive and curative services are inputs into health production. If at least one of these two types of care increases and neither of them decreases, health should presumably improve to some degree. Alternatively, the prediction is ambiguous if one type increases and the other decreases. In our empirical analyses, we are able to examine the prevalence of infectious diseases common among children (diarrhea, cough and fever) as well as the number of days lost to illness.

III. Data and Empirical Strategy

A. Data

Our empirical approach requires household survey data containing three types of information: (1) enrollment in the SR, (2) components of the SISBEN index (enabling us to simulate SR eligibility), and (3) potential behavioral responses and outcomes of interest (both welfare-improving and distortionary). There are two candidate Colombian household surveys that meet these criteria: the Encuestas de Calidad de Vida (ECV) and the Demographic and Health Surveys (DHS). The ECVs are nationally-representative household surveys designed to measure socioeconomic well-being and “quality of life,” broadly defined. The DHS data reports detailed fertility, health, and socio-economic information for nationally-representative samples of fertile age women (defined as ages 15–49) and their households. Because the de facto implementation of the SR occurred in 1996/1997, we use the 2003 ECV and the 2005 DHS for our analyses.¹⁸ Table 1 shows descriptive statistics by type of behavior/outcome for the full samples as well as those with and without SR coverage.

Table 1.

Descriptive Statistics

Variable	Total			Data Source	Not Enrolled in the Subsidized Regime			Enrolled in the Subsidized Regime			Data Source
	Mean	SD	N		Mean	SD	N	Mean	SD	N
	Outcome Variables
Risk Protection, Consumption Smoothing, and Portfolio Choice
Individual Inpatient Medical Spending	10,794	129,777	4219	ECV	11,360	129,227	2252	10,145	130,435	1967	ECV
Individual Outpatient Medical Spending	2,118	22,565	4218	ECV	2,513	20,170	2249	1,666	25,019	1969	ECV
Variability of Individual Inpatient Medical Spending	21,935	127,975	4219	ECV	26,338	126,615	2252	16,894	129,363	1967	ECV
Variability of Individual Outpatient Medical Spending	5,477	21,993	4218	ECV	7,343	19,017	2249	3,347	24,796	1969	ECV
Individual Education Spending	7,588	20,279	3567	ECV	7,501	23,234	1874	7,684	16,407	1693	ECV
Household Education Spending	35,145	48,468	4222	ECV	34,089	52,464	2253	36,352	43,428	1969	ECV
Total Spending on Food	274,881	217,392	4096	ECV	279,128	201,735	2171	270,091	233,756	1925	ECV
Total Monthly Expenditure	654,552	599,861	966	ECV	688,065	716,382	479	621,590	455,559	487	ECV
Has Car	0.03	0.16	3334	DHS	0.03	0.18	1758	0.02	0.15	1576	DHS
Has Radio	0.61	0.49	3334	DHS	0.60	0.49	1758	0.62	0.49	1576	DHS
Medical Care Use
Preventive Physician Visit	0.47	0.50	4222	ECV	0.39	0.49	2253	0.57	0.49	1969	ECV
Number of Growth Development Checks Last Year	1.16	1.81	1167	DHS	1.00	1.75	683	1.38	1.87	484	DHS
Curative Care Use (Not Conditional on Health Status)	0.06	0.24	4222	ECV	0.06	0.23	2253	0.07	0.26	1969	ECV
Primary Care	0.05	0.21	4222	ECV	0.04	0.20	2253	0.06	0.23	1969	ECV
Medical Visit - Specialist	0.02	0.12	4222	ECV	0.01	0.12	2253	0.02	0.13	1969	ECV
Hospital Stay	0.08	0.26	4222	ECV	0.07	0.25	2253	0.08	0.28	1969	ECV
Medical Visit for Chronic Disease	0.58	0.49	564	ECV	0.52	0.50	262	0.64	0.48	302	ECV
Curative Care Use among Children (Not Conditional on Health Status)	0.31	0.46	1161	DHS	0.29	0.45	680	0.34	0.48	481	DHS
Health Status
Child Days Lost to Illness	0.60	2.01	1161	DHS	0.65	2.14	680	0.53	1.82	481	DHS
Cough, Fever Diarrhea	0.56	0.50	1167	DHS	0.56	0.50	682	0.56	0.50	485	DHS
Any Health Problem	0.65	0.48	1161	DHS	0.64	0.48	680	0.66	0.47	481	DHS
Birthweight (KG)	3.26	0.57	897	DHS	3.25	0.53	534	3.27	0.63	363	DHS
Behavioral Distortions
Drank Alcohol during Pregnancy	0.11	0.31	998	DHS	0.12	0.32	504	0.10	0.30	494	DHS
Number of Drinks per Week during Pregnancy	2.75	9.76	110	DHS	3.15	10.85	60	2.28	8.36	50	DHS
Months Child Breastfed	12.58	10.87	946	DHS	11.15	10.18	578	14.83	11.53	368	DHS
Folic Acid During Pregnancy	0.54	0.50	988	DHS	0.56	0.50	500	0.53	0.50	488	DHS
Number Months Folic Acid during Pregnancy	3.96	2.42	515	DHS	4.05	2.39	266	3.86	2.45	249	DHS
Contributory Regime Enrollment	0.09	0.29	4222	ECV	0.17	0.38	2253	0.00	0.00	1969	ECV
Contributory Regime Enrollment	0.14	0.34	3334	DHS	0.26	0.44	1758	0.00	0.00	1576	DHS
Other Health Insurance	0.44	0.50	4222	ECV	0.82	0.39	2253	0.00	0.00	1969	ECV
Other Health Insurance	0.01	0.09	3334	DHS	0.01	0.12	1758	0.00	0.00	1576	DHS
Uninsured	0.00	0.06	4222	ECV	0.01	0.08	2253	0.00	0.00	1969	ECV
Uninsured	0.38	0.49	3334	DHS	0.73	0.45	1758	0.00	0.00	1576	DHS
	Baseline Demographic Variables
Balance
Household Head Age	46.81	14.81	3334	DHS	47.36	15.37	1758	46.18	14.14	1576	DHS
Household Head Age	46.58	14.79	4222	ECV	45.71	15.12	2253	47.57	14.35	1969	ECV
Completed Elementary School	0.18	0.39	3333	DHS	0.18	0.38	1757	0.19	0.39	1576	DHS
Completed Elementary School	0.19	0.39	3764	ECV	0.18	0.39	1985	0.19	0.39	1779	ECV
Completed Secondary School	0.20	0.40	3333	DHS	0.21	0.41	1757	0.19	0.39	1576	DHS
Completed Secondary School	0.07	0.26	3764	ECV	0.08	0.27	1985	0.07	0.25	1779	ECV
Household Head Completed Elementary School	0.30	0.46	3334	DHS	0.29	0.45	1758	0.31	0.46	1576	DHS
Household Head Completed Elementary School	0.28	0.45	4222	ECV	0.27	0.45	2253	0.28	0.45	1969	ECV
Household Head Completed Secondary School	0.02	0.15	3334	DHS	0.02	0.13	1758	0.03	0.18	1576	DHS
Household Head Completed Secondary School	0.02	0.13	4222	ECV	0.01	0.11	2253	0.02	0.15	1969	ECV
Student Received School Grant	0.08	0.26	1305	ECV	0.05	0.22	651	0.10	0.30	654	ECV
Benefits to Buy House	0.01	0.08	4222	ECV	0.00	0.06	2253	0.01	0.10	1969	ECV
Attended Training	0.05	0.21	3010	ECV	0.06	0.23	1593	0.04	0.19	1417	ECV
Household in Hogar Comunitario program	0.10	0.30	4222	ECV	0.09	0.29	2253	0.11	0.31	1969	ECV
Services from Bienstar Familiar	0.18	0.38	4222	ECV	0.16	0.37	2253	0.20	0.40	1969	ECV

Notes: All data summarized is from samples of “urban” individuals within two SISBEN index points of county-specific eligibility thresholds in either the 2003 Encuesta de Calidad de Vida (ECV) or the 2005 Demographic and Health Survey (DHS) (as indicated in the last column). SISBEN index calculations are described in detail in Appendix 2, and estimation of county-specific eligibility thresholds is presented in Section IIIB. The first group of three columns present summary statistics for the entire bandwidth of 2 sample, the second group for those not enrolled in the Subsidized Regime, and the third group for those enrolled in the Subsidized Regime.

As our empirical strategy requires, we calculate household-level SISBEN scores to simulate SR eligibility because simulated eligibility should not reflect manipulation of SISBEN scores (Camacho and Conover 2011).¹⁹ However, not all household surveys contain all necessary components of the SISBEN index. Appendix B provides a complete description of the SISBEN components present in each survey.²⁰ We impute values using ordered probit models for the few variables that are missing (firm size and per-capita income in the DHS and primary roof material in both surveys). Estrato is also missing for some households and we impute them on the basis of the most common in the household’s neighborhood.

B. Empirical Strategy

Instrumenting for Enrollment with Simulated Eligibility

In principle, the SISBEN index’s SR eligibility threshold can be used to study behavioral responses associated with SR enrollment. This discontinuity induces an abrupt shift in eligibility (and enrollment) along otherwise smooth distributions of household characteristics; coincident shifts in behaviors and outcomes can reasonably be linked to the program. However, adverse or propitious selection into eligibility – or manipulation according to unobserved household characteristics, as discussed in Section IE – is likely to bias the estimates of interest (McCrary 2008).

To circumvent this difficulty, we employ an instrumental variables strategy closely resembling one proposed by Hahn, Todd, and Van der Klaauw (2001). Conceptually, we seek to reconstruct ‘true’ SISBEN scores when both official SISBEN scores and observed SR enrollment reflect manipulation. To do so, we calculate SISBEN scores for each household in the ECV and DHS data and then use calculated scores to instrument for SR enrollment (for prominent examples of simulated instruments, see Currie and Gruber (1996a and 1996b), and Cutler and Gruber (1996), and Hoxby (2001)).²¹

Using urban households with simulated SISBEN scores near the urban eligibility threshold (we do not use rural households because of sufficiently inconsistent application of the rural scale)²², we could in principle begin by estimating the following first-stage equation for individuals i in household h:

$${\mathit{enroll}}_{ih}=\alpha +{\mathit{ybelow}}_h+\beta {\mathit{SISBEN}}_h+{\mathrm{\sum }}_k{\delta }_k{\mathit{estrato}}_{hk}+{\epsilon }_{ih},$$ (1)

where enroll is an indicator for whether or not individual i is enrolled in the SR, below is an indicator for simulated SISBEN score lying below the eligibility threshold, SISBEN is simulated SISBEN score, and estrato is a dummy variable for an estrato category (a neighborhood-level measure of socio-economic status used to establish electricity prices paid by local households). Using Two-Stage Least Squares (2SLS), we could then estimate the following second-stage equation:

$${\mathit{outcome}}_{ih}=\varphi +\lambda {\mathit{enroll}}_{ih}+\theta {\mathit{SISBEN}}_h+{\mathrm{\sum }}_k{\pi }_k{\mathit{estrato}}_{hk}+{\xi }_{ih},$$ (2)

instrumeting for enroll with below. The relationship between behavioral outcomes of interest (outcome) and SR enrollment would then be captured by estimates of the parameter λ.

Estimating County-Specific Eligibility Thresholds

As described in Section IE, financial shortfalls led many Colombian counties to use SR eligibility thresholds at SISBEN scores below the official national threshold. We therefore use county-specific eligibility thresholds. In addition to improving the strength of our first stage, this approach offers another key benefit: because some local governments use the official national threshold for other public benefits, changes in outcomes observed at county-specific SR thresholds will not reflect behavioral responses to other public programs. (Section IVA shows that participation in other public programs is not discontinuous at county-specific thresholds.)

Exact county-specific eligibility thresholds are unknown, so we estimate them following Chay, McEwan, and Urquiola (2005). Specifically, using our full samples, we establish county-specific breaks in SR eligibility at the SISBEN score that maximize the goodness-of-fit of a model of SR enrollment as a function of a dichotomous indicator for whether or not a household’s score falls below the threshold.²³ This approach establishes thresholds that maximize the percentage of individuals correctly classified as eligible in each county. Threshold means in the ECV and DHS are 45.03 and 45.83, respectively, and their interquantile ranges are 4.96 and 3.89. Intuitively, estimated thresholds are positively correlated with current county income (the sum of taxes, transfers from the state, and non-tax income) and negatively correlated with the share of households having unmet basic needs.²⁴

We then use county-specific thresholds to re-code the variable below for each individual i in households h and Colombian counties c and estimate the following first stage equation:

$${\mathit{enroll}}_{\mathit{ihc}}=\alpha +\gamma {\mathit{below}}_{hc}+\beta {\mathit{SISBEN}}_h+\varphi \mathit{SISBEN}\_{\mathit{diff}}_{hc}+{\mathrm{\sum }}_k{\delta }_k{\mathit{estrato}}_{hk}+{\mu }_c+{\epsilon }_{\mathit{ihc}},$$ (3)

where below is now an indicator for whether or not individual i’s simulated SISBEN score falls below the eligibility threshold in the individual’s county c, SISBEN_diff is the difference between an individual’s simulated SISBEN score and the estimated eligibility threshold in the individual’s county (i.e., relative SISBEN score), μ_c represents county fixed effects (allowing us to focus on within-county variation in simulated eligibility across county-specific thresholds), and all other variables are defined as in equation (1). To adhere transparently to the identifying assumption that individuals with simulated SISBEN scores very near the threshold are comparable with the exception of their eligibility, we conservatively focus on individuals whose calculated scores lie within two index points of the county-specific cutoff (our main estimates persist across various bandwidths, as shown in Section IVF).²⁵

Figures 1A and 1B use ECV and DHS data to show SR enrollment and “uninsurance” by simulated SISBEN score relative to county-specific eligibility thresholds. Each county’s threshold is normalized to zero, and the figure then shows means and 95% confidence intervals for each SISBEN index integer relative to the threshold as well as non-parametric kernel density plots on either side. The figure illustrates large discrete increases in the probability of enrollment and concomitant decreases in the probability of uninsurance at the threshold ranging between 25 to 30 percentage points. Figures 2A and 2B show the enrollment and “uninsurance” separately by regions of Colombia as well. It is worth noting than SR enrollment falls short of 100% at low (calculated) SISBEN scores. This may not only be due to mistargeting, but also to our use of household survey data (rather than official data) to construct SISBEN scores

Figure 1. Subsidized Regime Enrollment and ‘Uninsurance’ across SISBEN Eligibility Thresholds.

Notes: Source of Figure A is ECV 2003 and of Figure B is DHS 2005

Figure 2.

Subsidized Regime Enrollment and ‘Uninsurance’ across SISBEN Eligibility Thresholds by Region

Using our re-coded variable below to instrument for enroll, we then estimate the following equation by 2SLS:

$${\mathit{outcome}}_{\mathit{ihc}}=\varphi +\lambda {\mathit{enroll}}_{\mathit{ihc}}+\theta {\mathit{SISBEN}}_h+\psi \mathit{SISBEN}\_{\mathit{diff}}_{hc}+{\mathrm{\sum }}_k{\pi }_k{\mathit{estrato}}_{hk}+{\mu }_c+{\xi }_{\mathit{ihc}},$$ (4)

where the estimate of interest is the estimate of λ.²⁶ Section IVF shows that our results are robust across bandwidths to including higher-order relative SISBEN score polynomials, using interactions between relative SISBEN scores and simulated eligibility, and excluding county fixed effects. We also estimate λ using local linear regression without any covariates other than relative SISBEN score.²⁷

Limitations of the Empirical Strategy

Although our approach addresses key irregularities in SR implementation, it also has shortcomings. First, because the ECV and DHS household survey data that we use to calculate SISBEN scores was collected after official SISBEN classification, our calculations presumably include measurement error (even absent manipulation of official SISBEN status). Second, our estimates of county-specific eligibility thresholds presumably do not match those used in practice exactly, introducing additional noise into the relationship between our measure of simulated eligibility and actual enrollment. Third, manipulation of official SISBEN scores further weakens the strength of the first stage relationship. Fourth, our sample sizes are relatively small in the neighborhood of the eligibility thresholds. Overall, these limitations clearly compromise the power of our estimates (few are significant at significance thresholds below α=0.10) and suggest that our results should also be interpreted as lower bounds.

IV. Results

This section presents empirical evidence on a variety of important behavioral responses to the SR. We begin by demonstrating that both background characteristics not influenced by the program and participation in other social programs using SISBEN are balanced across estimated SR eligibility thresholds. We then investigate the effectiveness of health insurance in protecting households against financial risk (and possibly changing the optimal composition of household spending and assets). Next, we study changes in the use of traditionally under-utilized preventive services in light of the SR’s emphasis on allocative efficiency. In doing so, we also examine changes in health outcomes that are sensitive to the use of important preventive services. We then analyze how SR enrollment is related to changes in the use of curative medicine. To investigate possible behavioral distortions, we also test for reductions in private health investments (ex ante moral hazard) and insurance ‘crowd-out’ associated with SR enrollment. Finally, we assess the robustness of our estimates through a variety of additional parametric and non-parametric specifications.²⁸

A. Balance across Discontinuous Eligibility Thresholds

Our empirical approach assumes that no individual or household characteristics – other than SR enrollment – that could influence the outcomes of interest vary discontinuously across our estimated eligibility thresholds. To test this assumption, Table 2 shows results obtained by estimating equations (3) and (4) for individual attributes that could not reasonably change in response to SR enrollment (such as age or educational attainment among adults). Consistent with our assumption, the estimates are not generally distinguishable from zero. (Examining the distribution of simulated SISBEN scores across eligibility discontinuities, Appendix C shows evidence in support of this assumption as well.)

Table 2.

Balance Across Eligibility Thresholds

Panel A
Outcome	Household Head Age	Household Head Age	Completed Elementary School	Completed Elementary School	Completed Secondary School	Completed Secondary School	Household Head Completed Elementary School	Household Head Completed Elementary School
2SLS Estimate, Subsidized Regime Enrollment	4.38 [3.05]	3.05 [7.68]	0.01 [0.08]	−0.15 [0.10]	0.03 [0.07]	0.05 [0.05]	−0.09 [0.12]	−0.37 [0.24]
Intent to Treat Estimate	1.77 [1.22]	0.79 [1.84]	0.01 [0.03]	−0.04** [0.02]	0.01 [0.03]	0.01 [0.01]	−0.04 [0.05]	−0.10** [0.05]
First Stage Estimate, Below Eligibility Threshold	0.40*** [0.04]	0.26*** [0.07]	0.40*** [0.04]	0.25*** [0.06]	0.40*** [0.04]	0.25*** [0.06]	0.40*** [0.04]	0.26*** [0.07]
First Stage F-Statistic	125.75	14.08	125.91	17.02	125.91	17.02	125.75	14.08
OLS Estimate	−0.04 [0.66]	1.62*** [0.55]	0.00 [0.01]	0.00 [0.01]	−0.01 [0.02]	−0.01 [0.01]	0.02 [0.02]	−0.01 [0.02]
Mean for those Not Enrolled in the Subsidized Regime	47.36	45.71	0.18	0.18	0.21	0.08	0.29	0.27
Observations	3,334	4,222	3,333	3,764	3,333	3,764	3,334	4,222
Data Source	DHS	ECV	DHS	ECV	DHS	ECV	DHS	ECV

Panel B
Outcome	Household Head Completed Secondary School	Household Head Completed Secondary School	Student Received School Grant	Benefits to Buy House	Attended Training	Household in Hogar Comunitario Program	Services from Bienestar Familiar
2SLS Estimate, Subsidized Regime Enrollment	−0.02 [0.03]	−0.04 [0.07]	−0.06 [0.11]	0.02 [0.01]	0.01 [0.04]	0.03 [0.08]	−0.04 [0.13]
Intent to Treat Estimate	−0.01 [0.01]	−0.01 [0.02]	−0.01 [0.02]	0.00** [0.00]	0.00 [0.01]	0.01 [0.02]	−0.01 [0.03]
First Stage Estimate, Below Eligibility Threshold	0.40*** [0.04]	0.26*** [0.07]	0.21*** [0.09]	0.26*** [0.07]	0.27*** [0.06]	0.26*** [0.07]	0.26*** [0.07]
First Stage F-Statistic	125.75	14.08	5.17	14.08	23.45	14.08	14.08
OLS Estimate	0.01* [0.01]	0.01 [0.01]	0.03 [0.02]	0.00** [0.00]	−0.02*** [0.00]	0.01 [0.01]	0.02 [0.02]
Mean for those Not Enrolled in the Subsidized Regime	0.02	0.01	0.05	0.00	0.06	0.09	0.16
Observations	3,334	4,222	1,305	4,222	3,010	4,222	4,222
Data Source	DHS	ECV	ECV	ECV	ECV	ECV	ECV

Notes: Individual-level “urban” data for those within two SISBEN index points of county-specific eligibility thresholds from the 2003 ECV and 2005 DHS. Dependent variables are shown at the top of each column. The first row shows 2SLS estimates for enrollment in the Subsidized Regime (SR), instrumenting for SR enrollment using simulated eligibility. The second row shows reduced-form intent-to-treat estimates for simulated SISBEN score falling below the county-specific eligibility threshold. The third row shows first stage estimates from OLS regressions of Subsidized Regime enrollment on an indicator for falling below the county-specific eligibility threshold. All specifications

^***Significant at the 1 percent level.

^**Significant at the 5 percent level.

^*Significant at the 10 percent level.

We then consider whether or not our estimates from equation (4) could be attributed to participation in other public programs that also use the SISBEN index. Before investigating this possibility directly, we first note that it is unlikely because these programs use the uniform national eligibility threshold, while we estimate and utilize de facto county-specific thresholds for the SR – many of which fall below the uniform national threshold. To confirm this, we re-estimate equations (3) and (4) using a dichotomous indicator for participation in these other programs – including job training, home mortgage subsidies, education vouchers, Hogares Comunitarios (a large child care program), and services provided by the Instituto Colombiano de Bienestar Familiar (the largest social welfare agency in Colombia) – as the dependent variable in equation (4).²⁹ Table 2 presents these results, suggesting that participation in other programs is also balanced across county-specific SR eligibility thresholds.

B. Financial Risk Protection and Portfolio Choice

Although a key innovation of the SR is its emphasis on allocative efficiency, a central objective of any health insurance program is to provide protection against financial risk. We therefore begin by examining the relationship between SR enrollment and both level and variability of out-of-pocket medical spending. To construct our variability measure, we first calculate mean individual spending separately among those enrolled and those not enrolled in the SR. For each individual, we then measure the difference between each individual’s spending and the mean among those with the same enrollment status, constructing our variability measure as the absolute value of this difference. We analyze outpatient and inpatient out-of-pocket spending (within the past twelve months) separately and emphasize the latter because of idiosyncrasies in how outpatient spending is reported.³⁰

We first graphically examine shifts in the distribution of medical spending associated with simulated eligibility for the SR. Figure 3 shows cumulative density functions for inpatient medical spending in the preceding year separately for those falling above and below county-specific thresholds (using our sample of those within two index points of the cutoff).³¹ Both distributions are heavily right-skewed, but mass in the distribution for those who are eligible (those below the threshold) falls to the left of the distribution for those who are ineligible. No clearly comparable pattern is present for outpatient medical spending (Figure 4). Overall, these figures suggest that SR eligibility is associated with reductions in right-tail inpatient medical spending – a pattern consistent with risk protection.³²

Figure 3. Individual Inpatient Medical Spending (Among those within Two Points of County-Specific Eligibility Thresholds).

Notes: Source ECV 2003

Figure 4. Individual Outpatient Medical Spending (Among those within Two Points of County-Specific Eligibility Thresholds).

Notes: Source ECV 2003

The first four columns of Panel A in Table 3 then present econometric results obtained by estimating equations (3) and (4) for medical spending by type. The first row presents IV estimates for SR enrollment, and the second row reports intent-to-treat (ITT) estimates for simulated SR eligibility (estimates for a dummy variable coding whether or not an individual falls below the eligibility threshold obtained by OLS regressions of outcomes on this dummy and the other covariates in equation 4). The first column suggests that SR enrollment lowers mean inpatient spending by about 60,000 pesos (α=0.10), around a 30% reduction among those using any inpatient services. Perhaps more importantly for assessing protection against financial risk, the third column shows that SR enrollment is associated with reductions in the variability of inpatient medical spending (α=0.05).³³ There is no statistically significant association for outpatient care. Appendix D Figure 1 graphically shows all outcomes examined in Panel A across county-specific eligibility thresholds (essentially, graphical versions of our intent-to-treat analyses).

Table 3.

Risk Protection and Portfolio Choice

Panel A: Risk Protection
Outcome	Individual Inpatient Medical Spending	Individual Outpatient Medical Spending	Variability of Individual Inpatient Medical Spending	Variability of Individual Outpatient Medical Spending
2SLS Estimate, Subsidized Regime Enrollment	−58,870* [33,263]	3,562 [2,702]	−67,499.38** [32,906]	167.57 [2,417]
Intent to Treat Estimate	−15,108* [8,888]	918.23 [821]	−17322.90* [9,120]	43.20 [626]
First Stage Estimate, Below Eligibility Threshold	0.26*** [0.07]	0.26*** [0.07]	0.26*** [0.07]	0.26*** [0.07]
First Stage F-Statistic	13.91	14.01	13.91	14.01
OLS Estimate	−5,655 [3,898]	−1,204*** [342]	−13,888*** [3,893]	−4,387*** [357]
Mean for those Not Enrolled in the Subsidized Regime	11,359.86	2,512.98	26338.40	7342.59
Observations	4,219	4,218	4,219	4,218
Data Source	ECV	ECV	ECV	ECV

Panel B: Portfolio Choice
Outcome	Individual Education Spending	Household Education Spending	Total Spending on Food	Total Monthly Expenditure	Has Car	Has Radio
2SLS Estimate, Subsidized Regime Enrollment	−341.68 [3,781]	30,366 [25,055]	32,136 [103,540]	−33,826 [278,060]	0.01 [0.04]	0.17 [0.11]
Intent to Treat Estimate	−84.72 [945]	7,815 [4,880]	8,709 [28,491]	−14,036 [115,736]	0.01 [0.01]	0.07 [0.05]
First Stage Estimate, Below Eligibility Threshold	0.25*** [0.06]	0.26*** [0.07]	0.27*** [0.06]	0.41*** [0.12]	0.40*** [0.04]	0.40*** [0.04]
First Stage F-Statistic	19.28	14.08	18.80	12.18	125.75	125.75
OLS Estimate	122.82 [231]	2,952.32*** [902]	−12,036 [10,330]	−39,273 [58,730]	−0.01 [0.01]	0.03 [0.02]
Mean for those Not Enrolled in the Subsidized Regime	7,501	34,089	279,128	688,065	0.03	0.60
Observations	3,567	4,222	4,096	966	3,334	3,334
Data Source	ECV	ECV	ECV	ECV	DHS	DHS

Notes: Individual-level “urban” data for those within two SISBEN index points of county-specific eligibility thresholds from the 2003 ECV and 2005 DHS. Dependent variables are shown at the top of each column. The first row shows 2SLSestimates for enrollment in the Subsidized Regime (SR), instrumenting for SR enrollment using simulated eligibility. The second row shows reduced-form intent-to-treat estimates for simulated SISBEN score falling below the county-specific eligibility threshold. The third row shows first stage estimates from OLS regressions of Subsidized Regime enrollment on an indicator for falling below the county-specific eligibility threshold. All specifications also include SISBEN score, distance from the county-specific threshold, estrato dummy variables, and county fixed effects. Standard errors (clustered by county) are shown in brackets below each estimate.

^***Significant at the 1 percent level.

^**Significant at the 5 percent level.

^*Significant at the 10 percent level.

Overall, the results shown in Table 3 Panel A suggest that SR enrollment is associated with meaningful risk protection benefits. By reducing household exposure to financial risk, SR enrollment could also produce meaningful changes in the composition of household assets, human capital investments, and household consumption (i.e., portfolio choice effects). Specifically, it may increase investments not previously undertaken because of costly informal risk-management activities (such as precautionary saving). Panel B of Table 3 presents estimates for durable goods not used to construct the SISBEN index (car and radio ownership) as well as household education and consumption expenditures. In general, it implies that SR enrollment is not associated with discernible portfolio choice effects (perhaps because the SR falls short of providing full insurance).

C. Preventive Service Use and Health Status

As described earlier, a key innovation of the SR is its focus on improving allocative efficiency in medical care. In particular, increasing the use of highly beneficial preventive services – many of which produce large positive externalities – is likely to improve welfare. These externalities are due both to reduced rates of infectious disease transmission and to reduced curative care costs borne by others through risk pools.

The first two columns of Table 4 report estimates for different types of preventive care use obtained from equations (3) and (4) (Appendix D Figure 2 shows graphical versions of the intent-to-treat analyses). In general, they suggest substantial increases in the use of preventive health care services.³⁴ Specifically, SR enrollment is associated with a 29 percentage point increase in the probability of a preventive physician visit in the past year (a 75% increase; α=0.01). Importantly, the SR is also associated with 1.50 more growth-monitoring and well-care visits in the past year, a 50% increase relative to uninsured children (α=0.05). These well-care visits are a principal way of addressing important childhood health problems (parents receive nutritional advice, iron supplements, and de-worming medications, for example – which are important for strengthening children’s immune systems and making them less susceptible to major childhood illnesses such as severe diarrhea and acute respiratory infections/pneumonia).³⁵

Table 4.

Use of Preventive Medical Care and Health Status

	Use of Preventive Care		Health Status (Children)
Outcome	Preventive Physician Visit	Number of Growth Dev. Checks Last Year	Child Days Lost to Illness	Cough, Fever, Diarrhea	Any Health Problem	Birthweight (KG)
2SLS Estimate, Subsidized Regime Enrollment	0.29*** [0.11]	1.50** [0.69]	−1.40** [0.65]	−0.18 [0.17]	−0.06 [0.18]	0.26 [0.29]
Intent to Treat Estimate	0.08*** [0.03]	0.55** [0.25]	−0.49** [0.20]	−0.07 [0.06]	−0.02 [0.06]	0.11 [0.12]
First Stage Estimate, Below Eligibility Threshold	0.26*** [0.07]	0.36*** [0.07]	0.35*** [0.07]	0.37*** [0.07]	0.35*** [0.07]	0.41*** [0.09]
First Stage F-Statistic	14.08	25.24	23.46	25.19	23.46	19.10
OLS Estimate	0.17*** [0.01]	0.33*** [0.12]	−0.04 [0.17]	0.01 [0.04]	0.03 [0.04]	0.04 [0.05]
Mean for those Not Enrolled in the Subsidized Regime	0.39	1.00	0.65	0.56	0.64	3.25
Observations	4,222	1,167	1,161	1,167	1,161	897
Data Source	ECV	DHS	DHS	DHS	DHS	DHS

Notes: Individual-level “urban” data for those within two SISBEN index points of county-specific eligibility thresholds from the 2003 ECV and 2005 DHS. Dependent variables are shown at the top of each column. The first row shows 2SLS estimates for enrollment in the Subsidized Regime (SR), instrumenting for SR enrollment using simulated eligibility. The second row shows reduced-form intent-to-treat estimates for simulated SISBEN score falling below the county specific eligibility threshold. The third row shows first stage estimates from OLS regressions of Subsidized Regime enrollment on an indicator for falling below the county-specific eligibility threshold. All specifications also include SISBEN score, distance from the county-specific threshold, estrato dummy variables, and county fixed effects. Standard errors (clustered by county) are shown in brackets below each estimate.

^***Significant at the 1 percent level.

^**Significant at the 5 percent level.

^*Significant at the 10 percent level.

We then investigate whether or not health status has improved under the SR – in particular, dimensions of health that are sensitive to the use of preventive care. Although our preventive physician visit variable is not service-specific, our measure of childhood growth and development checks should correspond directly with infectious disease prevalence. Columns three through six of Table 4 show estimates for child health outcomes linked to preventive care use. SR enrollment is associated with 1.4 fewer child days absent from usual activities due to illness in the past month (α=0.05). Enrollment is also associated with an 18 percentage point reduction in the self-reported incidence of cough, fever, or diarrhea among children in the preceding two weeks (but it is not statistically significant at conventional levels). Appendix D Figure 2 shows graphical versions of the intent-to-treat analyses for these health outcomes as well.³⁶ Because these health gains can be linked to increases in preventive service use, they are also likely to improve welfare.

D. Use of Curative Medical Care

As Section II discusses, there are a variety of competing incentives and other forces influencing curative medical care under the SR. These include reductions in out-of-pocket prices, increases in the use of preventive services, and supply-side incentives for limiting total medical spending. Predictions about changes in curative service use under the SR – and their welfare implications – are therefore ambiguous, but they remain a central interest of policymakers and are important for assessing the costs of Colombia’s innovative health insurance reform.

Table 5 reports estimates for various categories of curative care obtained from equations (3) and (4) (Appendix D Figure 3 shows graphical versions of the intent-to-treat analyses). We find that SR enrollment is associated with a 13 percentage point increase in reported physician visits because of health problems within the past 30 days (α=0.05). There is no change in use of curative care among children, which may be due to improvements in child health associated with SR enrollment (as Table 4 shows). Finally, there is no meaningful relationship at conventional significance levels between participation in the SR and hospitalizations.

Table 5.

Use of Curative Medical Care

Outcome	Curative Use (Not Conditonal on Health Status)	Curative Use among Children (Not Conditional on Health Status)	Medical Visit for Chronic Disease	Hospital Stay
2SLS Estimate, Subsidized Regime Enrollment	0.13** [0.05]	−0.00 [0.16]	0.51* [0.28]	−0.04 [0.11]
Intent to Treat Estimate	0.03* [0.02]	0.00 [0.06]	0.18*** [0.06]	−0.01 [0.03]
First Stage Estimate, Below Eligibility Threshold	0.26*** [0.07]	0.35*** [0.07]	0.35*** [0.10]	0.26*** [0.07]
First Stage F-Statistic	14.08	23.46	13.49	14.08
OLS Estimate	0.03*** [0.01]	0.07* [0.04]	0.13*** [0.03]	0.02** [0.01]
Mean for those Not Enrolled in the Subsidized Regime	0.06	0.29	0.52	0.07
Observations	4,222	1,161	564	4,222
Data Source	ECV	DHS	ECV	ECV

Notes: Individual-level “urban” data for those within two SISBEN index points of county-specific eligibility thresholds from the 2003 ECV and 2005 DHS. Dependent variables are shown at the top of each column. The first row shows 2SLS estimates for enrollment in the Subsidized Regime (SR), instrumenting for SR enrollment using simulated eligibility. The second row shows reduced-form intent-to-treat estimates for simulated SISBEN score falling below the county specific eligibility threshold. The third row shows first stage estimates from OLS regressions of Subsidized Regime enrollment on an indicator for falling below the county-specific eligibility threshold. All specifications also include SISBEN score, distance from the county-specific threshold, estrato dummy variables, and county fixed effects. Standard errors (clustered by county) are shown in brackets below each estimate.

^***Significant at the 1 percent level.

^**Significant at the 5 percent level.

^*Significant at the 10 percent level.

E. Ex Ante Moral Hazard and Insurance Crowd-Out

Protection from financial risk associated with unexpected illness weakens private incentives for costly health protection (ex ante moral hazard) (Pauly 1968). Because we find evidence of greater risk protection, we investigate how protective private health behaviors not directly linked to medical care change with SR enrollment.³⁷ As Table 6 Panel A and Appendix D Figure 4 show, however, we find no meaningful change in breastfeeding or maternal investments in fetal health (alcohol, drug, or tobacco use during pregnancy; or prenatal dietary supplementation with iron, calcium, or folic acid), suggesting little ex ante moral hazard associated with SR enrollment.³⁸

Table 6.

Behavioral Distortion – Ex ante Moral Hazard and Insurance Crowd-out

Panel A: Ex-Ante Moral Hazard
Outcome	Drank Alcohol during Pregnancy	Months Breastfed as Child	Folic Acid During Pregnancy
2SLS Estimate, Subsidized Regime Enrollment	−0.10 [0.11]	−0.23 [5.36]	0.10 [0.18]
Intent to Treat Estimate	−0.04 [0.04]	−0.07 [1.64]	0.04 [0.07]
First Stage Estimate, Below Eligibility Threshold	0.40*** [0.07]	0.31*** [0.07]	0.41*** [0.07]
First Stage F-Statistic	28.83	19.00	30.12
OLS Estimate	−0.02 [0.02]	4.02*** [0.91]	0.02 [0.03]
Mean for those Not Enrolled in the Subsidized Regime	0.12	11.15	0.56
Observations	998	946	988
Data Source	DHS	DHS	DHS

Panel B: Insurance Crowd-Out
Outcome	Contributory Regime Enrollment	Uninsured	Other Health Insurance	Contributory Regime Enrollment	Uninsured	Other Health Insurance
Intent to Treat Estimate	−0.02 [0.02]	−0.23*** [0.06]	−0.00 [0.00]	−0.05** [0.02]	−0.35*** [0.04]	−0.01* [0.01]
Mean for those Not Enrolled in the Subsidized Regime	0.17	0.01	0.82	0.26	0.73	0.01
Observations	4,222	4,222	4,222	3,334	3,334	3,334
Data Source	ECV	ECV	ECV	DHS	DHS	DHS

Notes: Individual-level “urban” data for those within two SISBEN index points of county-specific eligibility thresholds from the 2003 ECV and 2005 DHS. Dependent variables are shown at the top of each column. The first row shows 2SLS estimates for enrollment in the Subsidized Regime (SR), instrumenting for SR enrollment using simulated eligibility. The second row shows reduced-form intent-to-treat estimates for simulated SISBEN score falling below the county-specific eligibility threshold. Intent-to-treat estimates only are reported for the crowd-out analyses of other insurance types (the last six columns). The third row shows first stage estimates from OLS regressions of Subsidized Regime enrollment on an indicator for falling below the county-specific eligibility threshold. All specifications also include SISBEN score, distance from the county-specific threshold, estrato dummy variables, and county fixed effects. Standard errors (clustered by household) are shown in brackets below each estimate.

^***Significant at the 1 percent level.

^**Significant at the 5 percent level.

^*Significant at the 10 percent level.

Manipulation of official SISBEN scores suggests that Colombians perceive benefits of SR enrollment, so we also investigate the possibility that SR enrollment displaces other forms of explicit health insurance.³⁹ Table 6 Panel B also presents intent-to-treat (ITT) estimates obtained by estimating equation (3) with dichotomous indicators for Regimen Contributivo enrollment, other forms of health insurance (those for the military, police officers, and certain industrial groups like oil industry workers, for example), and “uninsurance” as dependent variables. The Regimen Contributivo estimates are generally small (−0.02 in the ECV and minus;0.05 in the DHS) and the latter is statistically significant (α=0.05), suggesting that some modest crowding-out of formal insurance may have occurred.

F. Robustness

To investigate the robustness of our results, we estimate a variety of alternative parametric and non-parametric specifications based on our main estimating equations. First, we re-estimate equations (3) and (4) using SISBEN score bandwidths ranging between two and four. Second, at each bandwidth we estimate specifications that include higher order polynomials of relative SISBEN scores (including squared, cubic, and fourth power terms) as well as models that include interactions between relative SISBEN scores and simulated eligibility (although allowing SISBEN gradients to vary on either side of the eligibility threshold should matter little given our narrow bandwidth). Third, we re-estimate specifications that do not include county fixed effects (at each bandwidth), allowing us also to make cross-county comparisons among individuals with identical simulated SISBEN scores but that fall on opposite sides of county-specific eligibility thresholds. Finally, we also estimate models using non-parametric local linear regression.⁴⁰ As Tables 1–4 of Appendix E show that our results are generally robust (with some variation in precision) across these alternative bandwidths and specifications.

G. Heterogeneity and External Validity

While RD estimates can have good internal validity, their external validity is often limited because they typically utilize a sample only in the neighborhood of the discontinuity (Imbens and Lemieux 2008, Lee and Lemieux 2010). However, because SR eligibility thresholds vary by county, our empirical strategy allows us to investigate the external validity of our estimates by extending equation (4) to include interactions between absolute SISBEN score and SR enrollment.⁴¹ This allows us to test for heterogeneity by SISBEN score, a summary measure of socio-economic status. As Appendix F Tables 1–4 show, we find little evidence of heterogeneous effects, suggesting that our estimates are not sensitive to local variation in SISBEN scores and may therefore apply to a broader range of the Colombian population.⁴² We note, however, that the interaction terms’ standard errors are large, so our power to detect heterogeneity is limited.

V. Conclusion

This paper studies the role of Colombia’s Régimen Subsidiado in protecting the poor against financial risk, influencing their use of health services, and ultimately affecting their health status. Unusual among middle- and low-income countries at the time of its introduction, the SR is a publicly-financed health insurance program with features of managed care emphasizing allocative efficiency that are traditionally only found in wealthy countries. We find evidence that the SR provides some protection against financial risk and is associated with greater use of preventive services having positive externalities – as well as with measurable health gains among children.

Pinpointing the mechanisms through which complex insurance programs such as SR work is an important topic of future research. In our case, because preventive services are generally free regardless of insurance status, one plausible interpretation of the increase in prevention that we observe links them to the SR’s high-powered supply-side incentives (if providers with capitated contracts believe that prevention is cost-saving). Another potentially important channel is patient substitution from public to private sector providers contracting with insurance plans under the SR (if private providers were more likely to supply preventive care). Although we suspect that these two mechanisms may be most salient, we cannot rule out others (individual feeling more comfortable going to the doctor when insured, the provision of insurance making the universal entitlement to preventive care more salient, etc.).

We conclude by emphasizing that the welfare-improving potential of SR has yet to be fully realized. For example, a variety of political concessions followed the creation of the SR – including exemptions from the end of government subsidies as well as requirements that insurers contract with public facilities for a minimum share of the services that they finance. These concessions presumably limit the ability of health plans to contract with medical care providers in ways that best encourage higher quality and lower cost services.