In the as treated analysis, we considered orlistat initiators to be exposed until 180 days after the discontinuation of orlistat or an additional prescription first for any other anti-obesity drug; we censored non-initiators for initiation of any anti-obesity drug. We chose 180 days to allow for a carry-over effect or latency period of cancer detection. In both analyses, the follow-up started at 180 days after the start date to account for an induction period of cancer pathogenesis in both cohorts and ended with the earliest of death, any type of cancer (except non-melanoma skin cancer), migration out of the healthcare system, or end of the study period. Statistical analysis We used a Cox proportional hazards model with robust variance to estimate the hazard ratios of colorectal cancer, overall and over time, and their corresponding 95% confidence intervals.

To control confounding, we first assembled cohorts by matching and then controlled for remaining imbalances by using propensity scores weighting.14 15 We used this two step approach because we intended to control tightly for age, sex, and body mass index, which are the most important confounders in this study, without the need to rely on correct specification of multivariable models. We did sensitivity analyses to assess the robustness of our assumptions of induction and latency periods. We repeated the main analysis with various lengths of time for induction and latency periods. Note that intention to treat analysis is an extreme form of (infinite) latency. We also did subgroup analyses. We used SAS software (version 9.

2) for all statistical analyses. Results This study included 33625 orlistat initiators and 160347 matched non-initiators. Of all non-initiators, 20664 started anti-obesity drugs during follow-up. Table 11 compares the characteristics of orlistat initiators and non-initiators. Compared with non-initiators, orlistat initiators had a slightly higher prevalence of diabetes and hypertension at baseline and were more likely to receive drugs, including oral anti-diabetes drugs, statins, and non-steroidal anti-inflammatory drugs. After propensity scores weighting, baseline characteristics were well balanced between orlistat initiators and non-initiators. Table 1 Matched cohort characteristics at baseline, before and after propensity score weighting Table 22 shows incidence rates and hazard ratios of colorectal cancer.

In the intention to treat analysis, we observed 57 colorectal cancers among orlistat initiators and 246 among non-initiators during 106708 and 488526 person years of follow-up. The incidence rate of colorectal cancer per 100000 person years was 53 (95% confidence interval 41 to 69) for orlistat initiators and 50 (44 to 57) for non-initiators. The hazard ratio Carfilzomib of colorectal cancer comparing orlistat initiators with non-initiators was 1.11 (95% confidence interval 0.84 to 1.47) after propensity scores weighting.