Adjustments for confounding variables in observational population-based studies in oncology and comparison to results from randomized controlled trials.

Abstract

e18135 Background: Observational population-based studies (OBS) allow large scale analyses to be done, and are increasingly used to compare cancer treatment modalities. Given the immense heterogeneity in these cohorts and the high potential for bias, we investigate how often appropriate statistical strategies are used to address confounding factors and the agreement of these studies with results from randomized controlled trials (RCT). Methods: A systematic literature search was conducted in MEDLINE using controlled vocabulary to identify OBS between 2000 and 2011, identifying 1814 studies. Studies reporting on survival that compare local therapies were included. Study details were extracted including survival outcomes, whether adjustments were made for age, stage, comorbidities, and if propensity score analyses or sensitivity analyses were performed. Studies with robust adjustments were compared to representative RCT results. Results: A total of 155 treatment comparisons met inclusion criteria. Median follow-up was available in 37% and median/mean age in 64% of comparisons. Age and stage were not adjusted for in 12% and 8% of studies, respectively. Only 10% adjusted for comorbidities, and 15% (n = 24) used either propensity adjustment (n = 21) or sensitivity analyses (n = 8). Of these 24, 11 had corresponding RCTs, and six had discordant findings. In five of the six studies, the RCT showed no survival difference while the corresponding OBS suggested more treatment was better than less, suggesting potentially unadjusted confounding. Greater degree of statistical adjustments correlated with year of publication, and all studies that used propensity adjustment and sensitivity analysis were after 2007. Conclusions: During the decade of OBS analyzed, the majority did not report the median follow-up or age, adjust for comorbidities, or perform robust statistical analyses to correct for potential bias. Even among studies with rigorous statistical adjustments, over half had discordant findings from RCTs. Given these findings, comparative OBS in oncology should be seen as hypothesis generating and caution should be used when interpreting the results.