Table 1.
k is the degree of overlap, π is the censoring proportion, Xsmall = (X1, …, X7)T while Xlarge consists of Xsmall as well as all quadratic and two-way interaction terms computed from Xsmall.
Fig 1.
Relative bias of the MHRATT estimation methods for different DGPs.
No censoring (π = 0) and good overlap (k = 1).
Fig 2.
Experiment 1: Misspecification and overlap.
Relative bias of the MHRATT estimation methods for varying degrees of overlap and different DGPs. No censoring (π = 0). Outliers above 0.5 were left out to facilitate visual comparison between the experiments; a complete visualization can be seen in S3 Fig.
Fig 3.
Experiment 2: Misspecification and censoring.
Relative bias of the MHRATT estimation methods for varying censoring rates and different DGPs. Good overlap (k = 1).
Fig 4.
Experiment 3: Overspecification.
Relative bias of the MHRATT estimation methods for different DGPs when all models were overspecified. No censoring (π = 0) and good overlap (k = 1). Outliers above 0.5 have been left out to facilitate visual comparison between the experiments; a complete visualization can be seen in S5 Fig.
Table 2.
Characteristics of treated and untreated subjects in the original sample as well as balance between the two groups for each covariate.
Table 3.
Balance distributions, after weighting with GLM, LASSO, npCBPS, CAL-ET, and SBW, as well as the unweighted dataset (NAIVE), when considering balance on all untransformed covariates (Xsmall) and all covariates including quadratic terms and interactions (187 terms in total; Xlarge).
Fig 5.
Kaplan-Meier survival curves for the original sample and the samples weighted by SBW, LASSO, and GLM, respectively.
In this case, survival implies absence of an event (death or ischemic stroke).
Table 4.
Marginal hazard ratio (MHRATT) and absolute risk reduction (ARR) estimates of the anti-coagulant effect on the time to either a second ischemic stroke or death.
CIL and CIU are the lower and upper limits of 95% confidence intervals for MHRATT.