P128: Statistical Methodology for Safety Study using Real-World Data (RWD)
Poster Presenter
Li Huang
Principal Biostatistician
Phastar United States
Objectives
The objective of this study is to apply propensity score adjustment method for confounding control in statistical analysis of nominal and time-to-event variables in observational studies.
Method
In this study, we exploit applying logistical regression propensity score (PS) to analyze real-world data for the cofounding effect. The PS is estimated, checked between exposure cohorts and applied in the statistical analysis of nominal variables and time-to-event variables.
Results
The results of this procedure proved that it is efficient for the statistical analysis in RWD. In this study, adaptive propensity score procedure is applied to select outcome-specific lists of propensity score variables used for each outcome-specific propensity score model to balance the observed baseline covariates between treatment groups. Kernal-density plots of the estimated probability distribution are then compared with larger overlapping between two groups indicating better comparability. Next, 1:1 PS matching is used and the propensity score–based weight for each patient is calculated, normalized and examined to assess covariate balance before and after weighting using the absolute standardized difference (ASD) between exposure cohorts. ASD less than 0.20 means balance achieved for important confounders. Finally, the PS weighing is used in statistical analysis of incidence rates, cumulative incidence, time-specific risk ratios and risk differences, and hazard ratios. The algorithm to calculate the statistical results and related SAS code are included.
Conclusion
Currently Propensity Score Matching (PSM) is a popular method to analyze RWD to create a balanced covariate distribution between treatment groups. However detailed procedure to apply this approach is not clear. We went through a detailed procedure which demonstrates how to calculate propensity score, apply PSM to balance treatment groups, and use matching weights to adjust possible confounding effects via weighted estimation of incidence rates, cumulative incidence, time-specific risk ratios and risk differences, and hazard ratios. Using this procedure we can solve various problems in observation studies.
