Improved two-stage estimation to adjust for treatment switching in randomised trials: g-estimation to address time-dependent confounding.

In oncology trials, control group patients often switch onto the experimental treatment during follow-up, usually after disease progression. In this case, an intention-to-treat analysis will not address the policy question of interest - that of whether the new treatment represents an effective and cost-effective use of health care resources, compared to the standard treatment. Rank preserving structural failure time models (RPSFTM), inverse probability of censoring weights (IPCW) and two-stage estimation (TSE) have often been used to adjust for switching to inform treatment reimbursement policy decisions. TSE has been applied using a simple approach (TSEsimp), assuming no time-dependent confounding between the time of disease progression and the time of switch. This is problematic if there is a delay between progression and switch. In this paper we introduce TSEgest, which uses structural nested models and g-estimation to account for time-dependent confounding, and compare it to TSEsimp, RPSFTM and IPCW. We simulated scenarios where control group patients could switch onto the experimental treatment with and without time-dependent confounding being present. We varied switching proportions, treatment effects and censoring proportions. We assessed adjustment methods according to their estimation of control group restricted mean survival times that would have been observed in the absence of switching. All methods performed well in scenarios with no time-dependent confounding. TSEgest and RPSFTM continued to perform well in scenarios with time-dependent confounding, but TSEsimp resulted in substantial bias. IPCW also performed well in scenarios with time-dependent confounding, except when inverse probability weights were high in relation to the size of the group being subjected to weighting, which occurred when there was a combination of modest sample size and high switching proportions. TSEgest represents a useful addition to the collection of methods that may be used to adjust for treatment switching in trials in order to address policy-relevant questions.

Two-stage estimation to adjust for treatment switching in randomised trials: a simulation study investigating the use of inverse probability weighting instead of re-censoring.

BACKGROUND: Treatment switching is common in randomised trials of oncology treatments, with control group patients switching onto the experimental treatment during follow-up. This distorts an intention-to-treat comparison of the treatments under investigation. Two-stage estimation (TSE) can be used to estimate counterfactual survival times for patients who switch treatments - that is, survival times that would have been observed in the absence of switching. However, when switchers do not die during the study, counterfactual censoring times are estimated, inducing informative censoring. Re-censoring is usually applied alongside TSE to resolve this problem, but results in lost longer-term information - a major concern if the objective is to estimate long-term treatment effects, as is usually the case in health technology assessment. Inverse probability of censoring weights (IPCW) represents an alternative technique for addressing informative censoring but has not before been combined with TSE. We aim to determine whether combining TSE with IPCW (TSEipcw) represents a valid alternative to re-censoring. METHODS: We conducted a simulation study to compare TSEipcw to TSE with and without re-censoring. We simulated 48 scenarios where control group patients could switch onto the experimental treatment, with switching affected by prognosis. We investigated various switching proportions, treatment effects, survival function shapes, disease severities and switcher prognoses. We assessed the alternative TSE applications according to their estimation of control group restricted mean survival (RMST) that would have been observed in the absence of switching up to the end of trial follow-up. RESULTS: TSEipcw performed well when its weights had a low coefficient of variation, but performed poorly when the coefficient of variation was high. Re-censored analyses usually under-estimated control group RMST, whereas non-re-censored analyses usually produced over-estimates, with bias more serious when the treatment effect was high. In scenarios where TSEipcw performed well, it produced low bias that was often between the two extremes associated with the re-censoring and non-recensoring options. CONCLUSIONS: Treatment switching adjustment analyses using TSE should be conducted with re-censoring, without re-censoring, and with IPCW to explore the sensitivity in results to these application options. This should allow analysts and decision-makers to better interpret the results of adjustment analyses.

Causal inference for long-term survival in randomised trials with treatment switching: Should re-censoring be applied when estimating counterfactual survival times?

Treatment switching often has a crucial impact on estimates of effectiveness and cost-effectiveness of new oncology treatments. Rank preserving structural failure time models (RPSFTM) and two-stage estimation (TSE) methods estimate 'counterfactual' (i.e. had there been no switching) survival times and incorporate re-censoring to guard against informative censoring in the counterfactual dataset. However, re-censoring causes a loss of longer term survival information which is problematic when estimates of long-term survival effects are required, as is often the case for health technology assessment decision making. We present a simulation study designed to investigate applications of the RPSFTM and TSE with and without re-censoring, to determine whether re-censoring should always be recommended within adjustment analyses. We investigate a context where switching is from the control group onto the experimental treatment in scenarios with varying switch proportions, treatment effect sizes, treatment effect changes over time, survival function shapes, disease severity and switcher prognosis. Methods were assessed according to their estimation of control group restricted mean survival that would be observed in the absence of switching, up to the end of trial follow-up. We found that analyses which re-censored usually produced negative bias (i.e. underestimating control group restricted mean survival and overestimating the treatment effect), whereas analyses that did not re-censor consistently produced positive bias which was often smaller in magnitude than the bias associated with re-censored analyses, particularly when the treatment effect was high and the switching proportion was low. The RPSFTM with re-censoring generally resulted in increased bias compared to the other methods. We believe that analyses should be conducted with and without re-censoring, as this may provide decision-makers with useful information on where the true treatment effect is likely to lie. Incorporating re-censoring should not always represent the default approach when the objective is to estimate long-term survival times and treatment effects.