The Association of Health Insurance with institutional delivery and access to skilled birth attendants: Evidence from the Kenya Demographic and health survey 2008-09

Background: Healthcare financing through health insurance is gaining traction as developing countries strive to achieve universal health coverage and address the limited access to critical health services for specific populations including pregnant women and their children. However, these reforms are taking place despite limited evaluation of impact of health insurance on maternal health in developing countries including Kenya. In this study we evaluate the association of health insurance with access and utilization of obstetric delivery health services for pregnant women in Kenya. Methods: Nationally representative data from the Kenya Demographic and Health Survey 2008-09 was used in this study. 4082 pregnant women with outcomes of interest-Institutional delivery (Yes/No-delivery at hospital, dispensary, maternity home, and clinic) and access to skilled birth attendants (help by a nurse, doctor, or trained midwife at delivery) were selected from 8444 women ages 15-49 years. Linear and logistic regression, and propensity score adjustment are used to estimate the causal association of enrollment in insurance on obstetric health outcomes. Results: Mothers with insurance are 23 percentage points (p < 0.01) more likely to deliver at an institution and 20 percentages points (p < 0.01) more likely have access to skilled birth attendants compared to those not insured. In addition mothers of lower socio-economic status benefit more from enrollment in insurance compared to mothers of higher socio-economic status. For both institutional delivery and access to skilled birth attendants, the average difference of the association of insurance enrollment compared to not enrolling for those of low SES is 23 percentage points (p < 0.01), and 6 percentage points (p < 0.01) for those of higher SES. Conclusions: Enrolling in health insurance is associated with increased access and utilization of obstetric delivery health services for pregnant women. Notably, those of lower socio-economic status seem to benefit the most from enrollment in insurance. We used data from the 2008–09 Kenya Demographic and Health Survey (KDHS) for this analysis. KDHS is a nationally representative survey that sampled 10,000 households [22] and collected detailed health and sociodemographic information. A total of 400 clusters—133 urban and 267 rural—were selected from the master frame [22]. This sample was constructed to allow for separate estimates for key indicators for each of the eight provinces in Kenya, as well as for urban and rural areas separately [22]. Urban areas were oversampled to get enough cases for analysis [22]. As a result the KDHS sample is not self-weighting at the national level; therefore, the empirical strategies implemented in this analysis are based on weighted data. The women’s sample comprises of 8444 women ages 15–49. The final analytic sample has 4082 women who report two outcomes of interest – institutional delivery (Yes/No – delivery at hospital, dispensary, maternity home, and clinic) and access to skilled birth attendants (help by a nurse, doctor, or trained midwife at delivery). This definition of skilled birth attendant is based on the WHO recommendations [38]. The final analytic dataset of 4082 takes into account three (3) mothers who were missing values on their outcomes as well as covariates. Given that this represents 0.04% missing values, the missingness was ignored and the three mothers were dropped from the analysis. The independent variable is insurance enrollment (Yes/No). We generated the variable by combining responses to enrollment in different kinds of health insurance – community based health insurance, insurance from employer, government or social security, privately purchased insurance and insurance from other source. The 2008–09 KDHS was the first survey to include questions on insurance enrollment. Insurance enrollment however, is not random as individuals can select whether or not to enroll in insurance and at what time during the year they actually enroll. Because insurance enrollment is not a random event, we use a selection of covariates in the analysis including age, marital status, education, total number of children, total number of household members, employment status, urban or rural residence, HIV test, frequency of reading newspapers, cooking fuel and whether or not they have electricity. This vector of covariates allows for appropriate regression adjustment and use of the observed characteristics to construct counterfactuals of enrollment in insurance based on propensity scores. We would require a counterfactual to estimate the causal effects of insurance status on access to care for pregnant women [39] -- i.e., what would have happened to the women in the absence of the intervention - in this case enrollment in insurance. The ideal way of achieving a counterfactual is through randomization. However, insurance enrollment is not randomized thus this observational study uses rigorous non-experimental methods. First, we estimate the association between health insurance and healthcare access using unadjusted and adjusted linear and logistic models. In the logistic regressions, we estimated the marginal effects – Table ​Table3.3. We estimated models of the general form: Linear and Logistic Regression Estimates of the Association of Insurance with Institutional Delivery & Skilled Birth Attendant Notes: In the table above models 1 & 5 are unadjusted linear models and 3 & 7 are unadjusted logistic models. Models 2 & 6 are linear models with controls and 4 & 8 are logistic with controls. The vector of controls includes age, household characteristics, education, pregnancy history, HIV test, and urban residence. Reported for models 3, 4, 7 & 8 are average marginal effects and the R squared is Pseudo R2. In parentheses are robust Std Errors. Significance levels: ***p < 0.01, **p < 0.05, *p < 0.1 where: the subscript i runs over observations i = 1 , … , n; y i is the outcome of interest (institutional delivery or skilled birth attendant); Xβ is the linear predictor. However, as selection into insurance is not random and in order to make any empirical estimations of the causal association with insurance, the adverse selection has to be accounted for. To reduce selection on observables, we implemented propensity score methods based on the conditional probability of enrolling in insurance given a set of observed covariates as defined by Rosenbaum and Rubin [40]. The propensity score estimation takes advantage of the covariates available in the KDHS and reduces bias due to differences in observed covariates thus balancing the covariates in the insured and uninsured groups. After using the logit model in estimating the propensity scores and achieving balance of the propensity score between the insured and uninsured, the goal was to estimate the Average Treatment Effects (ATE) or population effects of enrollment in insurance. ATE can be determined as the difference in average outcomes for insured and uninsured and can be written as shown in eq. 2: where n = the total number of pregnant women; y 1i is outcomes for the insured; and y 0i is outcomes for the uninsured. However, we cannot estimate eq. (2) as we cannot observe both y 1i and y oi (counterfactuals/potential outcomes) for every pregnant woman. And given that our study is observational, it is likely that the outcomes of interest (institutional delivery and access to skilled birth attendant) are dependent on treatment (insurance enrollment) leading to biased ATE. We therefore use the propensity scores for estimation of the causal association of enrollment in health insurance. Specifically 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. From the ATE equation above (equation 2), we estimate the ATT equation below: where X is a set of covariates to condition on and Z is the treatment (enrollment in health insurance). The ATT estimation is based on the following assumptions [39–43]: These assumptions allowed for the construction of matched insurance samples based on the balancing score – the propensity score [40] and estimation of causal association of enrollment in health insurance by stratification, kernel, and nearest neighbor matching. We also conducted inverse probability weighting (IPW). 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 [44]. We used sampling weights in all analysis to account for the complex sampling strategy in the KDHS discussed above, and all statistical analyses were implemented in Stata 13. The average effects estimated from the linear, logistic, and propensity score methods may be heterogeneous for those with and without insurance. We addressed the potential for impact heterogeneity by further stratifying the analysis based on socio-economic status (SES). The SES index is a binary variable based on having electricity at home and current employment status. Because about 75% of the study sample lives in rural areas, having electricity and or working are good proxies for higher SES status.