Reducing Bias in the Assessment of Treatment Effectiveness

Abstract

Indication bias is the major challenge in assessing treatment effectiveness in observational studies. We explored the potential advantages of using an instrumental variable approach in the context of primary androgen deprivation therapy (ADT) for prostate cancer. We identified 31,930 men in the linked Surveillance, Epidemiology, and End Results-Medicare database with a diagnosis of prostate cancer who were not treated definitively with radical prostatectomy or radiation in the years 1992 through 2002, with follow-up through 2005. The association between use of primary ADT and overall, prostate cancer-specific, and nonprostate cancer survival was assessed using multivariable regression and instrumental variable methods. Two instrumental variables, based on region and urologist prescribing preference, were constructed and analyzed using exogenous probit models. Prespecified subgroup analyses in patients with lower-risk and higher-risk prostate tumors were also carried out. In the overall cohort, standard multivariable regression analyses showed a significantly increased risk of prostate cancer-related death, whereas the instrumental variable approaches showed a protective effect of primary ADT, which was significant for the urologist preference instrument (hazard ratio: 0.74; 95% confidence interval, 0.60-0.93). In the high-risk subgroup, using urologist preference for primary ADT as the instrument, there was a significant reduction in overall mortality (hazard ratio: 0.75; 95% confidence interval, 0.57-0.99), driven by a large reduction in prostate cancer-specific mortality. Instrumental variable analysis appears to provide better control of bias when assessing the effectiveness of primary ADT for prostate cancer, although the results may be more applicable to policy rather than to clinical decisions.