Handling treatment changes in randomized trials

Treatment changes in randomized trials are common: for example, in a trial evaluating psychotherapy, individuals allocated to psychotherapy may attend only partially or not at all; or in a trial evaluating a drug treatment, individuals allocated to no drug treatment may nevertheless receive the treatment. The issue is especially important in drug trials for late stage cancer where control group members typically receive the active treatment on disease progression. This talk focuses on time-to-event outcomes. In some cases, it is important to estimate the effect of the treatment if some or all of these treatment changes had not occurred: for example, for a health economic model exploring whether a drug should be available on the NHS, we would need to compare survival of a treated group with survival of a completely untreated group. Twelve years ago, I published strbee, which implements in Stata the rank-preserving structural failure time model (RPSFTM) of Robins and Tsiatis (1991). This is a model that infers a comparison of the randomized groups if they had had different treatment experiences; estimation is based only on comparisons of randomized groups. Over the intervening years, the RPSFTM has been increasingly (though not widely) used, and various problems have been identified. First, it assumes treatment benefit is the same whenever treatment is received, and a sensitivity analysis to address possible departures from this assumption is needed. Second, the method's power is low and declines as follow-up extends: later times that contribute little information are given the same weight as earlier times. Third, a wider range of estimation procedures is required. I will review the RPSFTM method and its alternatives in a Stata context, and I will describe an update of strbee which addresses the above issues.