Logistic regression: a brief primer

Abstract

Regression techniques are versatile in their application to medical research because they can measure associations, predict outcomes, and control for confounding variable effects. As one such technique, logistic regression is an efficient and powerful way to analyze the effect of a group of independent variables on a binary outcome by quantifying each independent variable's unique contribution. Using components of linear regression reflected in the logit scale, logistic regression iteratively identifies the strongest linear combination of variables with the greatest probability of detecting the observed outcome. Important considerations when conducting logistic regression include selecting independent variables, ensuring that relevant assumptions are met, and choosing an appropriate model building strategy. For independent variable selection, one should be guided by such factors as accepted theory, previous empirical investigations, clinical considerations, and univariate statistical analyses, with acknowledgement of potential confounding variables that should be accounted for. Basic assumptions that must be met for logistic regression include independence of errors, linearity in the logit for continuous variables, absence of multicollinearity, and lack of strongly influential outliers. Additionally, there should be an adequate number of events per independent variable to avoid an overfit model, with commonly recommended minimum "rules of thumb" ranging from 10 to 20 events per covariate. Regarding model building strategies, the three general types are direct/standard, sequential/hierarchical, and stepwise/statistical, with each having a different emphasis and purpose. Before reaching definitive conclusions from the results of any of these methods, one should formally quantify the model's internal validity (i.e., replicability within the same data set) and external validity (i.e., generalizability beyond the current sample). The resulting logistic regression model's overall fit to the sample data is assessed using various goodness-of-fit measures, with better fit characterized by a smaller difference between observed and model-predicted values. Use of diagnostic statistics is also recommended to further assess the adequacy of the model. Finally, results for independent variables are typically reported as odds ratios (ORs) with 95% confidence intervals (CIs).