Online Program

Effect Size Measures for Nonlinear Count Regression Models (303668)

 View Presentation

Keywords: effect size, Poisson regression, negative binomial regression, Cohen’s d, rate ratio, Monte Carlo

Measures of effect size provide much-needed context for statistical analyses by describing the magnitude of an observed relationship. For nonlinear models for count outcomes (such as Poisson regression; Cameron & Trivedi, 2013), effect size is typically presented as the rate ratio, which is the multiplicative change in the predicted outcome. This is different in form from the effect size measures commonly used with linear models, such as the squared correlation or the standardized mean difference (Cohen’s d). The standardized mean difference effect size can be informative for a nonlinear model, particularly when the predictor is a group variable, but is complicated to calculate due to the inherent non-constant variance. A Shiny app was developed to easily calculate several measures of effect size based on the regression coefficients and variance values. The app calculates nonlinear (rate ratio) and linear (Cohen’s d) effect size measures for count regression models (Poisson, quasi Poisson, negative binomial), as well as confidence intervals for the effect size using Monte Carlo simulation. Potential discrepancies between linear and nonlinear measures of effect size are discussed.