Masters candidate in Biostatistics, Joe Swintek, will present:

“A Comparison of Bias Reduction Between Analysis Methods using Propensity Scores”

Plan B Adviser: Chap Le

Abstract: This manuscript compares the bias reduction properties of propensity scores using propensity score matching, stratification, and regression applied to a single variable. Comparisons of methodology are performed using both a real-world clinical data set and simulated data sets containing similar properties. Analysis on the clinical data set shows that matching produced a bias reduction of 80%, stratification showed a reduction between 86% and 75%; depending on the number and type of strata used, and regression only achieved a reduction of 0.8%. Simulations are then used to show that for all but the most extreme cases matching produced an average bias reduction of 93% to 98% while stratification produced an average reduction of 88% to 96%, but with more variation then what was observed from matching. The simulations also showed that for extreme cases matching either failed to work outright or actually increased the amount of bias between treatment groups. Finally, it is concluded that the use of stratification is more desirable over matching because it is easier to implement, works for a larger variety of cases, and provides nearly the same bias reduction as matching.