Selection of Comparable Subjects from Different Treatment Groups When Randomization is Not Feasible

Abstract

Objective: The aim of this research is to find solutions for selecting comparable subjects in non-randomized situations across different treatments.
Method: A series of analysis were performed to find subjects with similar profiles. The Propensity Score (PS) was estimated using logistic regression with potential covariates and a Cartesian product of PS across the subjects and treatments was created. The absolute difference in PS was derived and the pair with the minimum difference in PS was obtained, finally dose effect was estimated from selected paired.
Results: There were 23 subjects in the low dose group and 186 in the high dose group under treatment to assess the effectiveness of the ADAS-cog score in terms of change from Baseline to Week 24. The unequal proportion of subjects between the two dose groups has raised questions about balance and bias. An optimum matching technique was used to find profiles similar to the 23 subjects in the high dose group. Although the result was non-significant, it was more reliable because it was based on similar profiles, inadequate sample size could be an issue. The re-sampling technique was utilized to generate 80 more samples from the original 23 subject data from low dose. The dose effect was again estimated and found to be significant at Week 16 and Week 24.
Conclusion: The PS method was able to find subjects with similar profiles across the dose groups and provide reliable results regarding dose differences. After re-sampling and making covariate adjustments, the dose difference was found significant.