Regression analysis in clinical research
Bishoy Zakhary, Jeff Choi
Abstract
ABSTRACT: Regression modeling is a vital tool that develops correlations and associations between exposure and outcome. The outcome's characteristics and how it is captured in the data ultimately guides the decision to which model is selected. However, the interpretation and the statistics that go into model selection and study design dictate the validity of the model. Direct acyclic graphs and other study design techniques can be essential tools in determining the variables to include in the model and identify any potential shortcomings the software can miss. There are various types of regression models to select from depending on the hypothesis and study design, many of which fall under the tree of generalized linear models. Less commonly used models such as cox regression, negative binomial regression, and Poisson regression models can all provide potentially better alternatives in clinical research. In this review, we aim to examine regression techniques in beyond the usually reported multivariable linear and logistic regression models and focus on advanced statistical modeling and appropriate measures to account for variable selection and model behavior. LEVEL OF EVIDENCE: Biostatistical Review Article; Not Applicable.