Answers are grounded in this paper and cite their sources.
Background. Routinely-collected laboratory data are an increasingly attractive substrate for observational epidemiology, but selective test ordering, missing covariates and unmeasured confounding can substantially bias exposure–outcome estimates. Individual bias-analysis tools — inverse-probability-of-selection weighting, spline exposure–response models, E-values and negative-control outcomes — are well established, yet are rarely combined into a single reproducible pipeline with a pre-declared benchmark. We describe such a framework and demonstrate its behaviour in two worked exemplars. Methods. Single-centre 2014–2023 records from the University Medical Centre Ljubljana Institute of Clinical Chemistry and Biochemistry, all measured on the same Siemens Advia platform. 82,279 adult records met plausibility bounds. Two exposure–outcome cohorts were pre-specified: urate (22,059 patient-year records, 15,320 patients) and total calcium (34,924 patient-year records, 23,982 patients). Each was analysed with a five-step adjustment cascade (M0 crude → M4 primary [age spline, sex, meteorological season, era, serum albumin] → M5 [+eGFR-MDRD, same-sample]), stabilised inverse-probability-of-selection weights, natural-spline exposure–response contrasts, robust-Poisson prevalence ratios for clinical phenotypes, E-values for unmeasured confounding, and a five-analyte negative-control-outcome benchmark (chloride, potassium, sodium, total CO2, phosphate). Results. After full adjustment, 25(OH)D was associated with a change of −1.00 µmol/L (95% CI −1.60, −0.40) in serum urate and +4.12 µmol/L (3.55, 4.69) in total calcium per 10 nmol/L, with directionally consistent inverse-probability-weighted estimates. Prevalence ratios were near-null with confidence-limit E-values of 1.00 (hyperuricemia), 1.25 (hypocalcaemia) and 1.09 (hypercalcaemia). Standardised M4 effects fell within the range of five negative-control outcomes in both cohorts. Conclusions. The framework quantifies how much of an apparent exposure–outcome association survives conventional and non-conventional bias probes on a single dataset, and provides an implementable, reusable template for observational research built on routinely collected laboratory data. Full Python and mirror-R implementations are released with the paper.
No takes yet. Share an insight, caveat, or question.
Osredkar et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: