Is treatment A better than treatment B? This is a standard question in health economics. Typically, the answer is determined by a clinical trial. Treatment efficacy (as measured in a trial), however, may not match treatment effectiveness (as measured in the real world). For instance, patient adherence to treatment may be worse, if the treatment regimen is complicated real-world providers may have less expertise in administering the treatment in the trial, or there may be effect modifiers that are not captured precisely in the trial due to sample size issues. In addition, patients change treatments much more frequently in the real world compared to clinical trials.
So how does one treatment effect in the real world in the presence of switching? The two most basic approaches are to use an intent to treat (ITT) approach–which compares treatments based only on the initial treatment given–and censoring (i.e., dropping from the sample) patients who switch. The former assumes that switching occurs at random, which may not be the case if patients switch more often when the treatment is less effective. Further, patients may be assigned to a treatment arm of the study when they only took the drug once. Censoring has a similar problem whereby dropping patients who switch from the sample may create a bias if an unobserved factor influenced both treatment switching and patient outcomes.
A short paper by Aslop et al. (2020) provides a nice summary of the main assumptions needed for the following approaches to evaluating treatment effectiveness to be valid.
- Intention-to-treat analysis. Assumes switching occurs at random.
- Exclude/censor switches. Assumes that there are no confounders that affect both the reason for switching and the treatment outcome.
- Include treatment as time-varying covariate. Assumes that there are no confounders that affect both the reason for switching and the treatment outcome.
- Inverse probability of censoring weights. Assumes no unmeasured confounders
- Rank-preserving structural failure time models. Assumes groups are randomized groups with a common treatment effect.
- Two-stage model. Assumes that there are no unmeasured confounders and there exists a second baseline from which the effect of switching can be estimated
This is a helpful guide for those of you working with real-world evidence in the presence of treatment switching.