Re: Comparative Effectiveness of Prostate Cancer Treatments: Evaluating Statistical Adjustments for Confounding in Observational Data

Extract

Hadley et al. ( 1 ) use an analysis of Surveillance, Epidemiology, and End Results–Medicare prostate cancer data to argue in favor of instrumental variable analysis for controlling confounders in observational studies. They claim that “instrumental variable results … accounted for unobserved confounding” and “were more similar to results from the benchmark randomized controlled trial than … propensity score reweighted analyses.”

The key part of this claim is that of results being “more similar.” There are two possibilities why the authors may have drawn this conclusion. The first is that the central estimate from the instrumental variable analysis (a hazard ratio of 0.73) is similar to that reported in the randomized trial (a relative risk of 0.87). However, the hazard ratio appears to be similar only because the reference group has been reversed. The relative risk of 0.87 for surgery compared with conservative management from the trial would be equivalent to a hazard ratio for conservative management vs surgery of approximately 1.15. It is not at all apparent that the instrumental variables estimate of 0.73 is closer to the benchmark of 1.15 than is the propensity score–adjusted estimate of 1.59.