Purpose: Comparative efficacy research performed using population registries can be subject to significant bias. There is an absence of objective data demonstrating factors that can sufficiently reduce bias and provide accurate results.
Methods: MEDLINE was searched from January 2000 to October 2016 for observational studies comparing two treatment regimens for any diagnosis of cancer, using SEER, SEER-Medicare, or the National Cancer Database. Reporting quality and statistical methods were assessed using components of the STROBE criteria. Randomized trials comparing the same treatment regimens were identified. Primary outcome was correlation between survival hazard ratio (HR) estimates provided by the observational studies and randomized trials. Secondary outcomes included agreement between matched pairs and predictors of agreement.
Results: Of 3,657 studies reviewed, 350 treatment comparisons met eligibility criteria and were matched to 121 randomized trials. There was no significant correlation between the HR estimates reported by observational studies and randomized trials (concordance correlation coefficient, 0.083; 95% CI, -0.068 to 0.230). Forty percent of matched studies were in agreement regarding treatment effects (κ, 0.037; 95% CI, -0.027 to 0.1), and 62% of the observational study HRs fell within the 95% CIs of the randomized trials. Cancer type, data source, reporting quality, adjustment for age, stage, or comorbidities, use of propensity weighting, instrumental variable or sensitivity analysis, and well-matched study population did not predict agreement.
Conclusion: We were unable to identify any modifiable factor present in population-based observational studies that improved agreement with randomized trials. There was no agreement beyond what is expected by chance, regardless of reporting quality or statistical rigor of the observational study. Future work is needed to identify reliable methods for conducting population-based comparative efficacy research.