Effects of social health insurance on access and utilization of obstetric health services: Results from HIV+ pregnant women in Kenya

Background: Reducing maternal morbidity and mortality remains a top global health agenda especially in high HIV/AIDS endemic locations where there is increased likelihood of mother to child transmission (MTCT) of HIV. Social health insurance (SHI) has emerged as a viable option to improve population access to health services, while improving outcomes for disenfranchised populations, particularly HIV+ women. However, the effect of SHI on healthcare access for HIV+ persons in limited resource settings is yet to undergo rigorous empirical evaluation. This study analyzes the effect of health insurance on obstetric healthcare access including institutional delivery and skilled birth attendants for HIV+ pregnant women in Kenya. Methods: We analyzed cross-sectional data from HIV+ pregnant women (ages 15-49 years) who had a delivery (full term, preterm, miscarriage) between 2008 and 2013 with their insurance enrollment status available in the electronic medical records database of a HIV healthcare system in Kenya. We estimated linear and logistic regression models and implemented matching and inverse probability weighting (IPW) to improve balance on observable individual characteristics. Additionally, we estimated heterogeneous effects stratified by HIV disease severity (CD4 < 350 as "Severe HIV disease", and CD4 > 350 otherwise). Findings: Health Insurance enrollment is associated with improved obstetric health services utilization among HIV+ pregnant women in Kenya. Specifically, HIV+ pregnant women covered by NHIF have greater access to institutional delivery (12.5-percentage points difference) and skilled birth attendants (19-percentage points difference) compared to uninsured. Notably, the effect of NHIF on obstetric health service use is much greater for those who are sicker (CD4 < 350) - 20 percentage points difference. Conclusion: This study confirms conceptual and practical considerations around health insurance and healthcare access for HIV+ persons. Further, it helps to inform relevant policy development for health insurance and HIV financing and delivery in Kenya and in similar countries in sub-Saharan Africa in the universal health coverage (UHC) era. We hypothesize that NHIF improves access and utilization of institutional delivery services, and maternal health and hence influences the well-being of HIV+ pregnant women. The hypothesis is based on economic theory suggesting that people purchase health insurance not only to avoid risk of financial loss, but also as a mechanism for gaining access to healthcare that would otherwise be unaffordable [33] [34]. We use data from the Academic Model Providing Access to Healthcare (AMPATH). AMPATH is one of the largest and most comprehensive HIV/AIDS healthcare systems in SSA providing care to more than 150,000 HIV+ individuals in Western Kenya, tests approximately 80,000 pregnant women annually for HIV, and has robust electronic medical records [35] [36]. AMPATH has also been at the forefront of helping the Kenya Ministry of Health (MOH) formulate and implement healthcare policy initiatives and changes [37]. The main data for this analysis comes from cross-sectional records of medical encounters for HIV+ individuals within the AMPATH system between 2008 and 2013. The data is stored in the AMPATH Medical Records System (AMRS) - an electronic database of clinical encounters spanning more than 500 healthcare facilities in Western Kenya with extensive socio-demographic, economic, and biological variables [35]. The study population includes HIV+ pregnant women (ages 15–49 years) who get their HIV care at AMPATH clinics and have had a delivery (full term, preterm, miscarriage) with their NHIF enrollment status available in the dataset. The study analytic samples comprise of HIV+ pregnant women whose information on their outcomes between 2008 and 2013 is complete. The Institutional Delivery and Skilled Birth Attendant samples are generated by intersecting the observed outcomes and reported insurance status leading to a cross-sectional dataset of the most recent observed outcomes. The independent variable is NHIF enrollment (Yes/No). The outcome or dependent variables are institutional delivery (birth at a hospital: Yes/No) and skilled birth attendant (SBA) -help by a nurse, doctor, or trained midwife at delivery: (Yes/No) - this is the WHO definition of SBA [38]. The dataset also includes the following covariates: age; number of children, access to electricity and piped water, education, Cluster of Differentiation antigen 4 (CD4) count, travel time to clinic, and clinic site (urban or rural). As detailed below, this high dimensional vector of covariates allows for appropriate regression adjustment and use of the observed characteristics of enrollment in NHIF to estimate the effect of insurance enrollment on outcomes based on matching methods. First, we conduct bivariate analysis to determine the nature and degree of the relationship between enrolling in NHIF and socio-economic and demographic variables. Next, we estimate the association between NHIF and obstetric healthcare access using unadjusted and adjusted linear and logistic models. Despite the outcomes being binary (0,1), we use both linear and logistic regression models given that unless the probabilities being modeled are extreme, then linear and logistic models fit equally well and the linear model in the econometric literature is favored for ease of interpretation [39] [40]. Additionally, if the probabilities are extreme i.e. closer to 0/1, then the logistic regression can suffer from complete separation, quasi-complete separation, and rare events bias especially in small samples [41]. To improve comparability as selection into insurance is not random, we implemented matching methods based on the conditional probability of enrolling in insurance given a set of observed covariates as defined by Rosenbaum and Rubin [42]. The matching estimation takes advantage of the covariates available in the dataset including their higher order terms (squares, cubes, and quadratics) and reduces bias due to differences in observed covariates thus balancing the covariates in the insured and uninsured groups [42]. After using the logit model in estimating the propensity scores and achieving balance of the propensity score between the insured and uninsured we estimate and report the ‘average treatment effect on the treated’ (ATT) i.e., the average response to treatment (insurance) for those pregnant women that enrolled in or were enrolled in health insurance. Specifically, we use three matching methods (stratification, kernel, and radius) to estimate ATT based on the propensity score [43]. The Stratification method consists of dividing the range of variation of the propensity score into intervals such that within each interval, treated and control units have on average the same propensity score [43]. Stratification method however discards observations in blocks where either treated or control units are absent [43]. In Radius matching, each treated unit is matched only with the control units whose propensity score falls into a predefined neighborhood of the propensity score of the treated unit – for this paper the radius is 0.01. While in Kernel Matching, those insured are matched with a weighted average of all uninsured with weights that are inversely proportional to the distance between the propensity scores of insured and uninsured [43]. Both Radius and Kernel matching help address the limitations of stratified matching. Additionally, we implemented inverse probability weighting (IPW). IPW weights subjects by the inverse probability of treatment received thus creating a synthetic sample in which treatment assignment is independent of measured baseline covariates and allows one to obtain unbiased estimates of average treatment effects [44]. IPW was implemented given the potential for unequal probabilities of NHIF enrollment [45]. The results from the IPW estimates are our preferred results in line with the literature [44]. Given that, this study is an observational cross-sectional study with a single treatment variable, as discussed by Bender and Lange 2001, multiple test adjustments were not performed [46]. All statistical analyses were implemented in Stata 13 with the program “pscore.ado” used for matching estimation and all standard errors bootstrapped. Further, there is also the potential that the average effects estimated from the different models are heterogeneous for those with and without NHIF, and thus differ from the estimated average effects. The potential for impact heterogeneity is addressed by further stratifying the analysis based on HIV disease severity. HIV disease severity is defined using CD4 counts with CD4 < 350 as “Severe HIV disease,” and CD4 > 350 otherwise [47].