The stratified analysis method is commonly used to provide appropriate analysis to avoid bias. Its applications not only apply to comparative studies in observational studies, but also in randomized clinical trials. In observational comparative studies, baseline characteristics for patients between the two comparison cohorts can be different due to multiple factors. The intention of stratification is to create buckets so that within each bucket the patients between the treatment and the control groups are similar and the comparison within a bucket is apple-to-apple. For randomized clinical trials, stratified randomization is typically employed to ensure balance in important baseline characteristics between the treatment and control groups. However, stratified randomization cannot accommodate many factors as the number of stratifications increases exponentially as the number of stratification factors increases. Therefore, tangible imbalance in important factors not used in stratified randomization could happen by chance. Stratified analyses adjusting for the above factors are necessary to avoid bias.  

As one can see, the stratified analyses possess nice properties as the appropriate analysis strategy to avoid bias. The paper ‘The utilities and pitfalls of stratified analysis in challenging situations’ published at Pharmaceutical Statistics illustrated situations where the stratified analyses could lead to efficiency loss against un-stratified analyses where the distributions of patient numbers between the treatment and control groups are imbalanced across strata while the distributions of outcomes between the two comparison groups are balanced across strata. Simulations and discussions in this paper help better understand this situation with suggestions provided to handle this dilemma.