Abstract
In applied epidemiological studies, the term “risk factor” is used synonymously with both predictors and causal factors for disease, although causal aims of explanatory analyses are rarely stated explicitly. Consequently, the concepts of explaining and predicting are conflated in many risk factor analyses, where data-driven variable selection, not suitable for causal inference, is common. We review the current practice of evaluating risk factors with regression analysis in three medical journals, focusing on the definition of risk factors, covariate selection strategies, and interpretation of regression coefficients. In simulations, we compare common practices to a causal approach using a data learner that mimics data from a study investigating risk factors for COVID-19 severity among patients with Multiple Sclerosis. The implied meaning of the term’ risk factor’ varies across the reviewed articles, but many authors implicitly give a causal definition of the term. In the articles, logistic regression is the most frequently used model, from which the coefficients are taken as effect estimates of the risk factors under study. We identify three common covariate selection strategies: i) adjusting for a pre-specified set, ii) stepwise selection, and iii) univariable pre-filtering. The simulation study illustrates the limitations of the current practice compared to the estimation of a marginal causal odds ratio by highlighting the difference between the conditional odds ratio and the marginal odds ratio. When considering risk factors as causal factors, the effect of interest should be clearly defined, and variables should be selected based on the underlying causal structure. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12874-025-02704-0.