From: Clinical data mining: challenges, opportunities, and recommendations for translational applications
Aspect | PS matching | PS hierarchy | PS correction | PS weighting |
---|---|---|---|---|
Principle | Matching one or more control cases with a propensity score almost equal to the PS for each treatment case | Stratifying the sample based on rank-ordered PSs and performing comparisons between groups within each stratum | Incorporating PS values as a covariate in regression analysis models | Utilizing the PS to develop weights and applying all outcomes of interest |
Ability to control confounding bias | Superior to the PS hierarchy and PS correction | Weaker than other methods in particular to survival analysis | Weaker than PS matching and PS weight | Superior to PS hierarchy and PS correction |
Data utilization | Removing data that does not match the study objectives | Retaining data from all study objectives | Retaining data from all study objectives | Retaining data from all study objectives |
Causal effect estimation | Matching can estimate only the ATT | Hierarchy can estimate only marginal effect but neither the ATT nor the ATE | Correction can estimate only marginal effect but neither the ATT nor the ATE | Weighting can estimate either effect (ATT or ATE) according to the way weights are defined |
Advantages | Addressing the confounding from multiple variables to guarantee equalization between individuals; The strength of the argument is strong and mirrors a closer randomized experiment | Achieving equilibrium of intergroup covariates within each stratum | Model-based analysis with a straightforward application | The strength of the argument is strong and mirrors a closer randomized experiment |
Disadvantages | As only areas of the domain that are mutually supported by PS values can be matched, the sample size is reduced | Inadequate covariate equalization, particularly for the uppermost and lowest tiers | A model-dependent approach that can sometimes be challenging to meet the assumptions of the model | The sample in the study is only theoretical; Excessive weight will have an impact on the effect estimates |