Menu
Menu
Menu
Menu
The effect of voluntary health insurance on preventive health has received limited research attention in developing countries, even when they suffer immensely from easily preventable illnesses. This paper surveys households in rural south-western Uganda, which are geographically serviced by a voluntary Community-based health insurance scheme, and applied propensity score matching to assess the effect of enrolment on using mosquito nets and deworming under-five children. We find that enrolment in the scheme increased the probability of using a mosquito net by 26% and deworming by 18%. We postulate that these findings are partly mediated by information diffusion and social networks, financial protection, which gives households the capacity to save and use service more, especially curative services that are delivered alongside preventive services. This paper provides more insight into the broader effects of health insurance in developing countries, beyond financial protection and utilisation of hospital-based services.
The provision of preventive health services in Uganda is synchronised with the current policy of free access to all health services in public health facilities (Nabyonga Orem et al. 2005, 2011). Services in private-not-for-profit (PNFP) health facilities are subsidised by government subsidies for primary health (Amone et al. 2005; Okwero et al. 2010), making a majority of preventive health services almost universally freely available (MOH 2013). Moreover, products such as LLINs are highly subsidised or provided for free through donor-supported programmes (USAID 2015). Since preventive services are available and subsidised, their utilisation should, in principle, be high. In fact, utilisation of such services is low. For instance, while LLIN ownership rates have increased substantially, close to 10 percent of households that own a net do not regularly use it (UBOS and ICF 2018), contributing to close to 16 million annual malaria infections (MOH 2016). Only 59% of the population had a hand washing facility and 26% had an improved sanitation facility (UBOS and ICF, 2018). The Kisiizi CBHI scheme started in 1996 (Musau 1999) and currently covers above 45,000 individuals in 220 groups (Kisiizi Hospital 2020). At the time of data collection, households paid annual premiums ranging from Uganda shillings equivalent to US$ 3 (Uganda shillings 11,000) per person for households of 8–11 members to US$ 8 (Uganda shillings 28,000) per individual in a two-person household with additional coverage for private wards. Kisiizi CBHI scheme is a rural scheme with no sophisticated method of controlling moral hazard and adverse selection. Instead, three conditions are applied at enrolment. First, households enrol as a unit, such that all members enrol at once. Partial enrolment is therefore not permitted. Secondly, enrolment is group-based. Households are organised in groups rather than individual household enrolment. However, this is not typical group insurance since there is no join liability within groups. Burial groups are, therefore, only used for information diffusion and collection of premiums. Conducting enrolment at household and group level has been found to control moral hazard and adverse selection in other CBHI schemes such as in Pakistan (Fischer et al. 2018). It is important to note that group leaders are not incentivised or punished by the scheme in undertaking these roles. Some groups have therefore experienced leadership challenges such as corruption and misuse of groups’ money, which has led to some of them dropping out of CBHI. These groups have different leadership styles, some electing leaders every couple of years while others haven’t elected leaders in a long time. The scheme does not have any influence of groups affairs since such groups always have other areas of operation (such as funeral support, village saving and lending, agricultural labour support etc.) that are beyond the scope of CBHI. Of the 210 groups registered in CBHI at the time of our data collection, over 95% of them were primarily burial insurance groups though with additional community social support function. Funeral insurance groups are central in the promotion of health insurance across other developing countries (Dercon et al. 2006, 2014). Membership in funeral groups is based on kin or neighbourhood relationships and, therefore exogenous. Virtually every household belongs in one, and in sometimes, non-membership attracted communal sanctions (Katabarwa 1999). There is, therefore, a very important social network dimension. Finally, the scheme employs a substantially long waiting period. Newly enrolled households typically wait for about 12 months to be fully covered, in which time they are required to pay 90%t of medical costs in the instance of hospitalisation. This waiting time is significantly longer than other schemes, such as one in Nigeria (Bonfrer et al. 2015). The scheme covers outpatient and inpatient services, surgeries and emergences services. Investigative and imagining procedures such as X-rays, ultra-sounds and laboratory investigations are also covered up to the full cost of the treatment. Elective surgical conditions are covered up to 50% of the cost. However, the insurance does not cover dental, optical procedures and what it considers as self-inflicted injuries such as those arising from alcohol consumption and substance abuse. Care for chronic illnesses and all other services sought from health providers outside the network of the scheme’s health facilities are also not covered. The total ceiling for each illness episode is about US$600. It is important to note here that the preventive health outcomes of interest here are provided for free by all health facilities in the country, under the public health financing policy that provides free health care at public health facilities and subsidises private health facilities with grants to provide essential care for free. Therefore, the effect of interest in this study is mainly a behaviour change effect for health utilisation rather than the income effect of health insurance (Fig. 1). Coverage of the Kisiizi CBHI scheme. Source: Authors from scheme records The scheme operates in 5 districts in south-western Uganda. However, we conducted our study in areas within a 15-km radius from the main health provision facility, the Kisiizi hospital. This area was comprised of 3 sub-districts (sub counties) in Kabale (now Rukiga) and Rukugiri districts, which have the highest concentration of insured households. According to the 2014 national census, these three sub-districts had a combined population of 105,600 people (UBOS 2014). We used a multi-stage simple random sampling criterion to select 464 households in fourteen (14) villages scattered in the three sub-districts in the scheme catchment area. We invited community leaders from the three sub-districts and conducted a village listing exercise, which produced 174 villages in total. Going by a criterion of (1) having a market, (2) a school or health centre, and (3) a road in the village, the leaders categorised the villages into rich and poor villages. We then listed 104 poor villages and 70 rich villages. Seven villages were then selected from each category using a raffle draw. In the selected villages, all households with a child between 6 and 59 months were selected. Village lists were carefully cleaned after double-checking with leaders and selected households. Altogether, 464 households were selected, and all responded to the survey conducted between August 2015 and November 2015. The survey modules included a household demographic module, a child and maternal health module, and a nutrition module. Information on household social and economic welfare using durable assets holdings and other endowments in agriculture, water and sanitation, and housing was also collected to construct a wealth index, and social connectivity and perception modules were used to construct indices for social connectivity and perceptions. Village level information is also collected to account for village-level heterogeneity. The survey was administered on a computerised personal interviewing (CAPI) platform to enable cost efficiency in data transmission and avoid data losses (Caeyers et al. 2012). Research ethical clearance was obtained through the University of Bonn Center for Development Research ethics committee. Ethical reviews were further conducted by the Mengo Hospital Research Ethics Review Committee, and the Uganda National Council of Science and Technology provided a research clearance certificate (SS-3936). Informed consent was acquired from all participants. Our identification strategy is guided by a theoretical model of preventive health, advanced by Dupas (2011). In this model of health investments, Dupas (2011) shows that health insurance acts as both as a curative and preventive health investment. As a curative investment, it provides cover for the financial shock due to illness in the current period. As a preventive investment, it reduces the probability of illness in future periods if it contributes to the utilisation of preventive services in previous and current periods. To understand the relationships of interest, we apply propensity score matching (PSM), a robust quasi-experimental method that helps in accounting for possible endogeneity in differences between sub-samples exposed to the intervention and a sub-sample not exposed (Abadie and Imbens 2016; Jalan and Ravallion 2003; Smith and Todd 2005). The method is widely used in health evaluations, including those studying the effects of health insurance (Gustafsson-Wright et al. 2018; Trujillo et al. 2005; Woode 2017). With PSM, we are able to construct a control group that comprises of households that do not participate in CBHI but who have the same probability of participating based on a set on observable factors and compare them with those who participated in CBHI and estimate the effect of participation. PSM can reduce bias in observed differences between the treated and the control group if two conditions are met. The first is the conditional independence assumption or selection on observables assumption. For our case, this assumption requires that the determinants of participation in CBHI and those that determine the CBHI-related outcomes are observed. The second assumption is the common support or overlap assumption, which provides that the probability of participation for both treated and control groups should be similar between 0 and 1 (0<pTi=1|Xi<1). If these two conditions hold, then we can estimate the cross-sectional specification of the average treatment effect on the treated as follows. where ATET is the average treatment effect on the treated coefficient for outcome Y, which is either the use of an LLIN or a taking a deworming tablet in the previous six months, T denotes enrolment in CBHI while C denotes the control, not enrolled. P(X) is the probability of CBHI participation based on a vector of covariates X. To implement PSM, we use the Treatment Effects potential outcomes framework in Stata (Stata Corp 2015), implementing a PSM model with three nearest neighbours. We apply a calliper of 0.2 standard deviations of the propensity score, recommended by Austin (2011) and the standard errors are adjusted using the Abadie-Imbens method (Abadie and Imbens 2016). In this study, the main treatment is membership in the Kisiizi CBHI scheme, which is given as a dummy that takes the value of 1 if the household was a member of the CBHI scheme and 0 otherwise. We estimate the probability of CBHI participation using a set of child, parent, household and village controls. The child-specific variables include age (in months), gender, birthweight, and exclusive breastfeeding for a full six months. We include parent control such as mother’s age, education status of the mother and father, and father’s employment status. We then include household controls including household size proportion of under-5 children in the household, household assets shown by total livestock units, an index of access to water and sanitation facilities, ownership of radio, ownership of a mobile phone, and whether a household was catholic or not. We then include various variables for household social connectivity, which influence the decision to enrol in CBHI. These include using mobile at least once in the last 30 days, having a neighbour in CBHI, membership in a farmer self-help group, and having a household member on a village leadership committee. We then include variables for household use of health services such as attendance of a postnatal clinic, attendance of antenatal care for the recommended four times, hospital treatment visit after sickness, and satisfaction with hospital waiting times. Finally, we include in the model five village-level variables that control for environment and variation at the village level. First, we provide results of covariate balancing after propensity score estimation. For our total usable sample of 455 households, all households have at least one nearest neighbour to provide a match. We provide results of balancing covariates in supplementary tables. Figure 2 below shows the box plot of raw and matched samples. Balance box plots before and after matching Figure 3 below shows the kernel density plots for the distribution of the propensity score before and after matching. By assessing both the box plots and kernel density plots, we are relatively comfortable of the balance achieved. Kernel density plots for the distribution of the propensity score before and after matching We further implement a more flexible PSM framework using PSMATCH2, the Stata user-written comment (Leuven and Sianesi 2003). Essentially, PSMATCH2 and our preferred implementation (Stata treatment effects) conduct identical analysis and results. However, PSMATCH2 allows us to conduct two more important procedures to test for robustness and sensitivity or our results. First, we narrow the calliper from 0.2 standard deviations of the propensity score (Austin 2011) to 0.015 standard deviations because narrow calliper width attain more precision (Lunt 2013). Secondly, the PSMATCH2 framework enables us to conduct additional sensitivity analysis for hidden bias by assessing Rosenbaum bounds to test for the level of unobserved heterogeneity (Becker and Caliendo 2007). We therefore show that our results are not sensitive to hidden bias at close to doubling the odds of assignment to treatment due to unobserved factors for LLIN or increase by over 50% for deworming.
N/A
