Parallel randomized and nonrandomized clinical trials

In a randomized clinical trial where one treatment is more invasive (e.g. surgery versus no surgery), a large proportion of patients eligible for randomization may decline consent to be randomized. Some investigators have responded to this situation by conducting, in addition to the randomized trial, a separate nonrandomized but otherwise identical trial consisting of those patients who are eligible for randomization, but instead choose their own treatment. Three such trials considered are those comparing coronary bypass surgery versus no surgery, adenoidectomy versus no surgery, and tonsillectomy versus no surgery. Various types of bias can arise in parallel randomized and nonrandomized trials. Patients who accept randomization may differ from those in the nonrandomized trial in ways that affect response. In addition, those in the nonrandomized trial who receive the treatment may differ from those who receive the control. Nonetheless, it has been asserted in the literature that the nonrandomized trial can add useful information. This paper evaluates the claims from the above studies that the addition of the nonrandomized trial enhances generalizability and that similar pretreatment characteristics and similar results from the randomized versus norandomized portions imply similarity between the randomized patients and the target population consisting of all eligible patients. We conclude that the addition of the nonrandomized trial gives valuable information concerning generalizability, however if several sources of bias mask each other, it is possible for the results of the randomized and nonrandomized trials to be similar, even though the results from the randomized trial are not generalizable to the target population. We also develop ways to adjust the estimate of treatment effect to reduce bias due to imbalance with respect to observed covariates, as well as tests to evaluate when adjustment is appropriate. Finally, the effect of hidden bias is discussed and methodology is developed to assess the sensitivity of the magnitude of hidden bias to various assumptions concerning an unobserved covariate