Estimating Treatment Effects in Longitudinal Surgical Clinical Trials with Partial Compliance Affecting Both Treatment Arms
Keywords: partial compliance, surgical trials, longitudinal data
Clinical trials for surgical interventions often are affected by non-adherence to assigned treatment groups. This paper deals with some of the special analytical issues encountered in such trials. In particular, we focus on recently published results from the Spine Patients Outcomes Research Trial (SPORT) which was designed to the treatment effects of surgery on longitudinally measured quality of life following either elective spine surgery or non-operative care. Following randomization, many (~40%) patients assigned to non-operative treatment elected to have surgery. Additionally, many (~40%) patients assigned to surgery delayed their operation or decided not to have the procedure performed. With the levels of crossover seen, the treatment effect in the intent-to-treat (ITT) analysis is likely to be biased toward the null. Our interest is in estimating the longitudinal response following surgery and comparing this to the response that would be observed in the absence of surgery. This leads to an approach we describe as a "moving baseline" model which "resets the clock" at the time of surgery. This model ascribes treatment effects as originating from the time of surgery or the beginning of nonoperative therapy, and allows for possibly nonlinear trends. Patients delaying surgery beyond their first follow-up(s) are counted as nonoperative until they have surgery. Adjustments are made for baseline variables shown to be related to crossover and missed visits. Because we are reassigning the visit times when we "reset the clock", we include a term to interpolate to the most recent visit. We show conditions under which this model may be fit with ordinary mixed model software. The statistical properties of alternative estimators are studied using simulation models allowing for adherence to depend on the outcome history and baseline covariates.