Reverse causation bias: A simulation study comparing first- and second-line treatments with an overlap of symptoms between treatment indication and studied outcome.

Abstract

Reverse causation is a challenge in many drug-cancer associations, where the cancer symptoms are potentially mistaken for drug indication symptoms. However, tools to assess the magnitude of this type of bias are currently lacking. We used a simulation-based approach to investigate the impact of reverse causation on the association between the use of topical tacrolimus and cutaneous T-cell lymphoma (CTCL) in a multinational, population-based study using topical corticosteroids (TCS) as comparator. We used a multistate model to simulate patients' use over time of a first- (TCS) and second-line treatment (topical tacrolimus), onset of atopic dermatitis (indication for drugs) and CTCL (the studied outcome). We simulated different scenarios to mimic real-life use of the two treatments. In all scenarios, it was assumed that there was no causal effect of the first- or second-line treatment on the occurrence of CTCL. Simulated data were analysed using Cox proportional hazards models. The simulated hazard ratios (HRs) of CTCL for patients treated with tacrolimus vs. TCS were consistently above 1 in all 9 settings in the main scenario. In our main analysis, we observed a median HR of 3.09 with 95% of the observed values between 2.11 and 4.69. We found substantial reverse causation bias in the simulated CTCL risk estimates for patients treated with tacrolimus vs. TCS. Reverse causation bias may result in a false positive association between the second-line treatment and the studied outcome, and this simulation-based framework can be adapted to quantify the potential reverse causation bias.