(PTFS06-03-23) Application of Empirical P-Value Calibration in Nutrition Epidemiology: A Simulation Study With Time-to-Event Competing-Risk Outcomes

Objectives: Negative controls are used in epidemiological studies to detect bias by assuming no causal relation between exposure and outcome. Empirical calibration, which uses a set of negative controls to remove biases in association estimation from observational data, has been established in some cases, but not in competing hazards models. This study simulates the application of empirical calibration in a model relevant to nutritional epidemiology where the effect of a food and specific cause of death are of interest.

Methods: Survival data were generated based on a competing risk model with nine causes of death as negative control outcomes and one as outcome of interest, all associated against a common exposure. To evaluate the empirical calibration for competing-risk time-to-event data, three models were fit. Model 1 was a typical Cox PH model of the association between exposure and outcome of interest, accounting for measured confounding. Model 2 was empirical calibration that adjusted Model 1 using similar Cox PH models for negative control outcomes. Model 3 was a Cox PH model of exposure on outcome of interest accounting for both measured and unmeasured confounding. The performance assessment of empirical p-value calibration was evaluated by computing type I error rates and power varying parameter values for confounding and true hazard ratios.

Results: The simulation demonstrated that Model 1 had substantial type I error (range 0.153 to 1), especially when the sample size and magnitude of unmeasured confounding were large. Model 2 controlled type I error (range 0.051 to 0.213) compared to Model 1 and sometimes to the level of Model 3 (range 0.027 to 0.051). Model 1 (range 0.257 to 1) had high power as a function of the tendency to over-reject the null hypothesis, while Model 2 (range 0.075 to 0.992) had lower power than Model 3 (range 0.035 to 0.560).

Conclusions: Empirical p-value calibration reduces the risk of wrongly claiming a significant association in nutrition intake and cause-specific mortality risk, compared to typical Cox PH approaches. However, type I error inflation and limited power remain compared to models that consider all confounding variables. So, empirical calibration is a methodological improvement, but controlling potential confounders remains essential.

Funding Sources: Funded by the Beef Checkoff. Content and editorial authority remain with authors.