How collider bias affects the relationship between skin color and heart attack using directed acyclic graphs, propensity scores, and stepwise approaches

Abstract

Abstract Background Statistical methods are essential in epidemiology research, but they can generate erroneous estimates when selecting variables based only on statistical criteria. The use of directed acyclic graphs (DAG) helps to understand the causal relationships between variables and to avoid biases. Objective Compare the estimate of the effect of skin color on heart attack obtained from three data analysis techniques: a stepwise approach based on statistical criteria, a propensity score technique, and a graphical approach based on causal criteria. Methods Population-based cross-sectional study using data from the second National Health and Nutrition Examination Survey (NHANES). The exposure variable was skin color (black or non-black) and the outcome was heart attack (yes or no). Multivariable logistic regressions were carried out using the stepwise, propensity score techniques and the DAG-based approach to identify the association between the variables. In the stepwise technique, all variables potentially related to the outcome were included in the model and a forward or backward algorithm was used. The propensity score was applied, estimating the probability of exposure based on the covariates and helping to create balanced groups for comparison. Different possible causal models were developed between the variables in the DAG-based approach, identifying confounding, mediation, and collision factors. The models were created considering self-rated health as a confounding or collider variable, and the modeling results were verified. Results A total of 10,351 adults were evaluated, the majority female (52.1%), aged 20 to 39 years (48.5%), and with non-black skin color (90.4%). The prevalence of heart attacks was 3.0%, and 17% rated their health as fair or poor. Using different modeling techniques, no association was found between skin color and heart attack (p > 0.05), except when self-rated health, a collider variable, was included in the stepwise models. In this case, there was an inverse and biased association between the two variables, indicating a collision bias (stepwise-backward-OR 0.48; 95%CI 0.33–0.70; stepwise-forward-OR 0.64; 95%CI 0.44–0.94). Conclusion Skin color was not associated with heart attack when controlling for appropriate confounding factors. However, when adjusting for self-rated health in stepwise techniques, a colliding variable, there was an inverse and distorted association between the two variables, indicating a collider bias. The DAG-based approach and propensity score can avoid this bias by correctly identifying confounding factors and colliders.