A Framework to address potential bias in case-control studies: an application on Breast Cancer in Argentina.

DIAZ, MP; BECARIA COQUET J; MUÑOZ SE

BARCELONA

Congreso; XXIXTH INTERNATIONAL BIOMETRIC CONFERENCE; 2018

Case-control studies (CCS) are one of the epidemiologic study designs most used worldwide. Theserequire an adequate planning to avoid bias and obtain reliable risk estimates of the effect of an exposure on a healthevent. Bias derived from confounding, selection bias and information bias are the most common systematic errors inCCS. The present work provides a methodological framework to quantitatively assess systematic errors in a breastcancer CCS, carried out in Argentina (2008-2015). 844 subjects (318/526 cases/controls) were included. Dietarypattern was considered as exposure variable. Confounding was analyzed applying regression models approach forobserved and unobserved variables. In this former case simulated scenarios of different a priori variability?sdistributions imposed were proposed (standard deviation equal to 0, 1, 2 and 3). In addition, some a priori associationcoefficients between the unobserved variable and some observed variable were proposed, assuming they could havesimilar distributions (coefficients equal to 0.2; 0.3; 0.4; 0.5). Information bias, derived from missing data in covariates,was handled applying multiple imputation by chained equations (MICE), considering the MAR mechanism. 20datasets were generated, and the imputation method was performed when variables had more than 10% of missingvalues. Relative variance increase was chosen as the diagnostic measure. No confounding effect of 8 proposedvariables was found. Regarding the effect of not measured confounding variable, the modifications in the riskestimates were not relevant. A possible confounding effect of the unmeasured variable was only observed when agreater variability and coefficients were imposed. In respect of selection bias, differences regarding conventionalestimates were small and concentrated in the associations and the confidence intervals amplitudes. Analysis ofcomplete information (CC) was carried out in 32% of women included in the study and 83% of them were consideredwhen MICE was applied. Both approaches showed a promoting effect of the "Traditional" dietary pattern (CC OR:1.33;CI 95%:1.015-1.755; MICE OR:1.4; CI 95%: 1.184-1.657). Nevertheless, effects of other covariates were onlyobserved when MICE was applied. These covariates were BMI (OR:1.03; CI 95%:1.004-1.067) and breastfeedingpractice (OR:0.54; CI 95%:0.365-0.813).