A Quantitative Bias Analysis to Assess the Impact of Unmeasured Confounding on Associations between Diabetes and Periodontitis

AIM To investigate unmeasured confounding in bidirectional associations between periodontitis and diabetes using quantitative bias analysis. METHODS Subsamples from the Veterans Affairs Dental Longitudinal Study were selected. Adjusted for known confounders, we used Cox proportional hazards models to estimate associations between pre-existing clinical periodontitis and incident-Type II-Diabetes (n=672), and between pre-existing diabetes and incident severe periodontitis (n=521), respectively. Hypothetical confounders were simulated into the dataset using Bernoulli trials based on pre-specified distributions of confounders within categories of each exposure and outcome. We calculated corrected hazard ratios (HR) over 10,000 bootstrapped samples. RESULTS In models using periodontitis as the exposure and incident diabetes as the outcome, adjusted HR=1.21(95%CI:0.64-2.30). Further adjustment for simulated confounders positively associated with periodontitis and diabetes greatly attenuated the association or explained it away entirely (HR=1). In models using diabetes as the exposure and incident periodontitis as the outcome, adjusted HR=1.35(95%CI:0.79-2.32). After further adjustment for simulated confounders, the lower bound of the simulation interval never reached the null value (HR≥1.03). CONCLUSIONS Presence of unmeasured confounding does not explain observed associations between pre-existing diabetes and incident periodontitis. However, presence of weak unmeasured confounding eliminated observed associations between pre-existing periodontitis and incident diabetes. These results clarify the bidirectional periodontitis-diabetes association.