Big data, observational research and P-value: a recipe for false-positive findings? A study of simulated and real prospective cohorts.

BACKGROUND: An increasing number of observational studies combine large sample sizes with low participation rates, which could lead to standard inference failing to control the false-discovery rate. We investigated if the 'empirical calibration of P-value' method (EPCV), reliant on negative controls, can preserve type I error in the context of survival analysis. METHODS: We used simulated cohort studies with 50% participation rate and two different selection bias mechanisms, and a real-life application on predictors of cancer mortality using data from four population-based cohorts in Northern Italy (n = 6976 men and women aged 25-74 years at baseline and 17 years of median follow-up). RESULTS: Type I error for the standard Cox model was above the 5% nominal level in 15 out of 16 simulated settings; for n = 10 000, the chances of a null association with hazard ratio = 1.05 having a P-value < 0.05 were 42.5%. Conversely, EPCV with 10 negative controls preserved the 5% nominal level in all the simulation settings, reducing bias in the point estimate by 80-90% when its main assumption was verified. In the real case, 15 out of 21 (71%) blood markers with no association with cancer mortality according to literature had a P-value < 0.05 in age- and gender-adjusted Cox models. After calibration, only 1 (4.8%) remained statistically significant. CONCLUSIONS: In the analyses of large observational studies prone to selection bias, the use of empirical distribution to calibrate P-values can substantially reduce the number of trivial results needing further screening for relevance and external validity.