Online Program

Structural Equation Modeling with Count Variables (304133)

Alexandra Hanlon, Virginia Tech 
Robert Kirkpatrick, Virginia Commonwealth University 
*Kevin Landis McKee, Virginia Tech 
Joshua Pritikin, Virginia Commonwealth University 

Keywords: count variables, discrete distributions, structural equation modeling, latent variables, liability threshold, mediation, factor analysis, OpenMx, R, linear modeling

Count variables are common among health and social sciences but difficult to rigorously incorporate into linear models with complex structure, such as latent factors with many count indicators, mediation models, and interactions with variables of other types. Treating count variables as continuous risks introducing bias, reducing precision, or masking true effects. These issues are compounded if the count distribution is zero inflated. Zero-inflation is often handled by strategies that complicate post-hoc interpretation and inference, such as data transformations and mixture modeling. These issues can be resolved by treating count variables as indicators of a latent, normally distributed continuum that has been discretized into intervals matching any chosen discrete distribution, including zero-inflated variants. Effects are interpreted in terms of the latent continuum and thus follow convenient, linear path tracing rules. The model is demonstrated with simulations and an application to health data. Example R scripts are provided for convenient model setup, simulation, and power calculations.