Investigating the Impact of Missing Confounders in Propensity Score Analyses of Treatment Effects
Frangou E., Collins G.
The popularity of estimating treatment effects in settings other than randomised controlled trials, such as observational studies, has grown over recent years. However, balance between baseline characteristics in the treated and untreated groups is not guaranteed in the absence of random treatment allocation. This could, then, lead to bias caused by confounding by indication. Propensity scores offer a solution to this problem. They represent the probability of a patient receiving the experimental treatment conditional on pre-treatment characteristics at baseline. A plethora of matching algorithms has been implemented in various software packages, allowing propensity scores to be used to create well-balanced treatment and comparison groups[1]. Missing data is another, unavoidable, challenge in research. One solution is to analyse the complete (observed) data. However, this can generate misleading conclusions. Strategies such as overall mean replacement and multiple imputation have thus been developed to handle incomplete data. The aim of our research study is to address missing data and estimating treatment effects in a non-random setting when analysing binary outcomes. We explore the impact of missing confounders and missing data within confounders when estimating the treatment effect. We use simulations covering a wide range of outcome-generating mechanisms, sample sizes and proportions of missingness to demonstrate how different methods of propensity score matching and different methods of dealing with missingness affect treatment effects. References: 1 Austin, P. C. A comparison of 12 algorithms for matching on the propensity score. Statistics in Medicine 33, 1057-1069, doi:10.1002/sim.6004 (2014).