Interpreting Robustness in Observational Real-World Evidence: Hydrochlorothiazide and Reported Nonmelanoma Skin Cancer History as an Empirical Case

Abstract

Background: We used the association between hydrochlorothiazide (HCTZ) and nonmelanoma skin cancer (NMSC) to examine how adjustment, comparator choice, and assumptions about bias affect the interpretation of observational real-world evidence. Methods: We analyzed pooled pre-pandemic National Health and Nutrition Examination Survey data from 2009 through March 2020. Survey-weighted models estimated the cross-sectional prevalence-odds association between current/recent HCTZ use, defined from the past-30-day prescription medication inventory, and reported NMSC history. We adjusted for measured covariates, varied the adjustment set, compared HCTZ-only with angiotensin-converting enzyme inhibitor (ACEI)-only current users, and added a general healthcare-access measure. Differential outcome-ascertainment quantitative bias analysis (QBA) used the primary adjusted model. A separate post-hoc analysis examined generic unmeasured confounding of the crude association. Results: Among 30,413 adults, 2,839 were current/recent HCTZ users and 476 reported an NMSC history. The crude positive association was substantially attenuated after measured-covariate adjustment (odds ratio [OR], 1.733; 95% confidence interval [CI], 1.240–2.421 vs OR, 0.976; 95% CI, 0.686–1.389). Estimates were similar when hypertension was omitted (OR, 1.016), when HCTZ-only users were compared with ACEI-only users (OR, 1.053), and when a general healthcare-access measure was added (OR, 0.969). Differential outcome-ascertainment QBA yielded ORs of 0.880–1.038 across valid S1–S4 scenarios. The sensitivity and specificity assumptions were illustrative rather than empirically calibrated for NHANES-reported NMSC; the CIs excluded bias-parameter and measurement-structure uncertainty. Conclusions: The association changed markedly after measured adjustment. Additional analyses gave similar estimates but examined different model choices, populations, and potential biases. Current/recent use could not be shown to precede historical NMSC, so these data cannot answer the cumulative-dose causal safety question addressed by temporally ordered longitudinal studies.