Evaluating the effects of treatment switching with randomization as an instrumental variable in a randomized controlled trial

Patients in some randomized controlled trials (RCTs) may switch from the treatment arm to which they were randomized. Different methods can be utilized to analyze RCT data which contains subjects that switch treatment groups. In an intention-to-treat (ITT) analysis, patients are analyzed according to the treatment arm to which they are randomized, regardless of any switches in treatment. An ITT analysis has been the gold-standard analysis in RCTs, since randomization protects it from bias due to the balanced selection of different types of patients in treatment groups (Bang 2007). However, an ITT analysis provides estimates of treatment effectiveness, rather than efficacy (Little 2009). Alternatives to an ITT analysis include an as-treated (AT) analysis and a per protocol (PP) analysis. AT and PP analyses estimate treatment efficacy but are prone to selection bias since subjects who comply with a treatment may differ from the sample of subjects randomized to that treatment (Little 2009). A more recent ITT analysis alternative involves randomization as an instrumental variable (IV). An IV is a variable that is associated with an exposure but not with the study outcome except through the exposure (Myers 2011). Analyses using randomization as an IV correct the ITT estimator for treatment switching and provide a direct estimate of treatment efficacy, all while being protected from bias by randomization (Little 2009). The purpose of this study was to utilize simulated data based on an ongoing RCT to evaluate the effect of treatment switching with randomization as an IV at differing levels of treatment group crossovers, for both continuous and binary outcomes, and for ideal and non-ideal data. Additional data, both ideal and non-ideal, were simulated to analyze some general RCT scenarios. All data were simulated with differing levels of treatment group switching and were analyzed using IV, ITT, and PP methods. Results from this study indicated that an IV analysis can provide the most unbiased point estimates of the three methods. Because of the method in which IV estimates were derived, the resulting standard errors and confidence interval widths can be higher than those obtained from the other two methods. The IV method was found to be the least biased of the three methods, since it had equal or higher power and higher coverage probabilities compared to the ITT estimates, and because a PP analysis, in many cases, can be biased because of its exclusion of non-compliant study subjects. Larger standard errors and CI widths can be seen as a tradeoff when trying to find the most accurate treatment effect estimate while retaining the integrity of randomization in an RCT that experiences some degree of treatment group switching