Background: Unanchored indirect comparisons (IC) involve the use of data for treated and control cohorts that were generated from separate studies, such as two single-arm trials or an RWE comparator to a single-arm trial. When full individual patient data (IPD) are available (i.e., IPD available for both cohorts), a variety of methods may be used to balance covariates to mitigate bias; most commonly, balancing is achieved through propensity score matching or weighting. However, IPD are often available for only one arm (i.e., researchers have access to IPD for one cohort but have only cohort-level covariate distributions from a published ‘Table 1’ for the other); in such cases, traditional maximum likelihood estimation of propensity scores cannot be completed.
Objectives: To demonstrate how optimization-based stable balancing weights (SBW) can be used to flexibly target estimands for IC with IPD available for only one arm.
Methods: : Retrospective study of patients undergoing elective left-sided colorectal resection. IPD from the Premier Healthcare Database were available for 5,333 patients in whom manual circular staplers were used for anastomosis (IPD cohort); Table 1 data from the publication of a single-arm trial were available for 167 patients in whom powered circular staplers were used for anastomosis (Table 1 cohort). SBW (R software ) were used to re-weight the IPD cohort on prognostic covariates available for both cohorts: age, sex, Hispanic ethnicity, insurance type, diabetes, hypertension, surgical approach, indication for surgery, teaching vs. non-teaching hospital, and hospital bed size category (400–499 vs. 500+). A grid search was used to find the lowest standardized mean difference (SMD) at which an optimal solution (weights of minimum variance) could be attained for re-weighting the IPD cohort to mimic the marginal distribution of covariates for the Table 1 cohort.
Results: After weighting, the effective sample size (size of the weighted pseudo-population) of the IPD cohort in Premier data was N=1,882. SBW balanced the IPD cohort relative to the single-arm trial Table 1 cohort with SMD mean =0.001 (range 0.000 to 0.002) and SMD ≤0.002 for all covariates, with no covariate differing by more than 1 percentage point between the cohorts. To compare SBW’s performance, IPD were obtained for the Table 1 cohort and traditional propensity score weighting (PSW) was performed; PSW resulted in 13% lower effective sample size (N=1,633) and higher imbalance in covariates, with SMD mean=0.009 (range 0.00-0.029) and 13 of 18 covariates SMD >0.002.
Conclusions: SBW is an innovative optimization-based approach to covariate balancing that allows researchers to flexibly target estimands for IC with IPD to trials where only published tables on covariates are available.
