Quantifying unmeasured confounding in three epidemiological studies associating myelodysplastic syndromes (MDS) with cardiovascular disease (CVD).

Abstract

e19532 Background: Confounding occurs when an investigator studies the effect of an exposure on the occurrence of an outcome, but in the process measures the effect of another factor (a confounding variable). Knowledge of possible causal pathways assists an investigator to minimize confounding, but despite utmost attempts, errors are frequent. To assess role of a possible confounder (smoking), data from 3 studies (1. Herrera at al. 2019 2. Goldberg et al. 2010 and 3. Brunner et al. 2017) associating MDS (exposure) with CVD (outcome) was used to conduct sensitivity analysis. Methods: 2 x 2 tables were constructed with data from 3 studies (part 2 of table below). A vector defining the proportion of confounder (smoking) among exposed [prop (Smoking+|MDS+)], unexposed [prop (Smoking+|MDS-)] and risk ratio associating smoking status with CVD (part 1) using historical data was deduced. Extracted data was added to online bias analysis tool ( https://sites.google.com/site/biasanalysis/ ) to calculate crude as well as mantel-haenszel (MH) relative risk (RR) adjusted for smoking. Results: RRs unadjusted for smoking status for studies 1, 2 and 3 were 1.31, 1.34 and 2.21, respectively, signifying increased risk of CVD in MDS patients (part 3). Upon adjustment for smoking status, MH-RR for MDS-CVD relationship was 0.95, 0.97 for studies 1 and 2 respectively, signifying no increased CVD risk and 1.59 for study 3 signifying increased CVD risk but lower upon adjustment. Conclusions: Making inference of MDS-CVD association without accounting for smoking status can blur effects and at times result in false associations. Given ambiguity in results from these 3 studies, further evidence is needed to better assess MDS-CVD relationship. A simple online tool allows estimating nonrandom errors, assess magnitude and direction of biases, and help clinicians/researchers quantify these biases for better interpretation of observational data in a time efficient manner. Other bias analyses for these studies will be presented. [Table: see text]