Online Program

Statistical Methods for Bridging the Gap Between Interpretative and Predictive Analysis (302944)

 View Presentation

Keywords: Prediction, Interpretation, Ensemble methods, Fractional polynomials, Statistical engineering, Clinical data, Survey data

Those who routinely work with practitioners may have engaged in conversations in which the terms "predictors,” "explanatory variables,” and “covariates” are used interchangeably. Often, these conversations include a dual focus of interpretation and prediction. Interpretive models are constructed to gain insight into why outcomes occur, such as why a person contracts a particular disease or exhibits a specific behavior. Predictive models focus on the accuracy of identifying groups and determining future group inclusion. For example, can we anticipate how a person will react to an exposure? Interpretive and predictive model construction methodologies have different approaches for data use, covariate selection, and assessing model fit. Simply, a model built for interpretation may not be a model reasonable for prediction. Using clinical data, we present two approaches that bridge the gap between interpretive and predictive models: an ensemble method combining known models in an innovative manner and a single model method (fractional polynomials). Both approaches are based on principles for unstructured problems and represent small-scale statistical engineering within health sciences research.