A Bayesian framework to account for misclassification error and uncertainty in the estimation of abortion incidence

Obtaining reliable estimates of the incidence of induced abortion remains a significant challenge in abortion re- search. In recent years, one indirect, survey-based technique for measuring the incidence of induced abortion which has gained significant attention is the Confidante Method. The method has been employed in various legal and social contexts, however, its efficacy has not been uniformly established and many studies have found its success to be context-dependent. Greater focus has been placed recently on assessing the method’s key assumptions and quantifying and adjusting for biases that arise from violations in those assumptions. In this paper we propose a general statistical framework to conceptualize and quantify the impact of biases on measuring abortion incidence from surveys. Specifically, we define the relationship between observed and true abortion prevalence based on misclassification error related to the sensitivity and specificity of the survey instrument. We argue that this formulation leads naturally to a Bayesian modeling approach to estimate abortion prevalence. Such a modelling framework allows for different levels of uncertainty about the misclassification parameters to be incorporated in the modeling process, with that uncertainty being propagated through to the final estimates. We illustrate our framework and modelling approach on data from an application of the Confidante Method in Uganda in 2018, where we account for systematic differences in confidante abortion reports based on the self-reported abortion experiences of surveyed individuals.