Randomization inference for stepped-wedge cluster-randomized trials: An application to community-based health insurance

Abstract

National health insurance schemes are generally impractical in low-income countries due to limited resources and low organizational capacity. In response to such obstacles, community-based health insurance (CBHI) schemes have emerged over the past 20 years. CBHIs are designed to reduce the financial burden generated by unanticipated treatment cost among individuals falling sick, and thus are expected to make health care more affordable. In this paper, we investigate whether CBHI schemes effectively protect individuals against large financial shocks using a stepped-wedge cluster-randomized design on data from a CBHI program rolled out in rural Burkina Faso. We investigate statistical properties of the stepped-wedge design following the parametric mixed model approach proposed by Hussey and Hughes in 2007. We find that testing for the treatment effect is generally sensitive to specification of the parametric model. For instance, if we fail to account for cluster-by-time interactions present in the data, the Type I error rate is severely inflated. We develop a more robust and efficient strategy—randomization inference. We demonstrate how to apply randomization inference to test for constant treatment effects and discuss test statistics suitable for the stepped-wedge design. Randomization inference guarantees a valid Type I error rate; simulation studies show that randomization inference test statistics also have power that is comparable to the currently used procedures that do not guarantee a valid Type I error rate. Finally, we apply our proposed method to the Burkina Faso CBHI dataset. We conclude that the insurance had limited effects on reducing the likelihood of low to moderate levels of catastrophic health expenditure, but substantially reduced the likelihood of extremely high health expenditure that exceeds half of a person’s monthly income.