The number of strata in propensity score stratification for a binary outcome

Non-interventional and other observational studies have become important in medical research. In such observational, non-randomized studies, groups usually differ in some baseline covariates. Propensity scores are increasingly being used in the statistical analysis of these studies. Stratification, also called subclassification, based on propensity scores is one of the possible methods. There is the quasi-standard of using five strata. In this paper we focus on a binary outcome and evaluate the above-mentioned standard of using five strata.

Bias and power for different numbers of strata are investigated with a simulation study. The methods are illustrated using data from a study where patients with diabetes mellitus and triple vessel disease undergoing coronary artery bypass surgery with and without previous percutaneous coronary intervention were compared.

We show that more than five strata can be more powerful and give less biased results. However, using more than ten strata hardly gives any further benefit.

When applying a stratification, more than five strata may be preferable, especially because of increased power. Our simulation study does not show a clear winner; hence a useful strategy could be to work with five as well as with ten strata.