Statistics

The TACE and SBRT groups were compared at both patient and tumor levels. T tests, Wilcoxon Mann-Whitney tests, z tests, and χ² tests were used for normal variables, ordinal but nonnormal variables, 2-sample proportions, and nominal categorical variables, respectively. Local control and FFLP were calculated at the tumor level as time from treatment initiation to subsequent local progression, liver transplant, or last follow-up. Overall survival was calculated at the patient level as the time from first treatment (with TACE or SBRT) to death (from any cause) or last follow-up. The effect of treatment and other covariates on LC was modeled using a mixed-effects Cox model with patient-level random effects to adjust for correlation between tumors of the same patient (23). We applied inverse probability of treatment weighting (IPTW) to the Kaplan-Meier method and Cox models for LC to adjust for potential treatment assignment imbalances (24). The treatment probabilities (ie, propensity) were calculated, at the tumor level, from a logistic regression using a set of covariates likely to have affected original treatment decisions, including referring physician, age, number of liver tumors, tumor size, performance status, and number of prior treatments. All of these variables were included, regardless of statistical significance. To avoid extreme weights, we truncated the estimated propensities by the 5th and 95th percentiles. We conducted univariate analysis using either the treatment-only model or models that included 1 covariate and the treatment indicator. We also conducted multivariate analysis for treatment and covariates altogether. Analyses were performed using R (version 3.1.1; R Foundation for Statistical Computing, Vienna, Austria).