Measurement error and outcome distributions: Methodological issues in regression analyses of behavioral coding data.

Abstract

Behavioral coding is increasingly used for studying mechanisms of change in psychosocial treatments for substance use disorders (SUDs). However, behavioral coding data typically include features that can be problematic in regression analyses, including measurement error in independent variables, non normal distributions of count outcome variables, and conflation of predictor and outcome variables with third variables, such as session length. Methodological research in econometrics has shown that these issues can lead to biased parameter estimates, inaccurate standard errors, and increased Type I and Type II error rates, yet these statistical issues are not widely known within SUD treatment research, or more generally, within psychotherapy coding research. Using minimally technical language intended for a broad audience of SUD treatment researchers, the present paper illustrates the nature in which these data issues are problematic. We draw on real-world data and simulation-based examples to illustrate how these data features can bias estimation of parameters and interpretation of models. A weighted negative binomial regression is introduced as an alternative to ordinary linear regression that appropriately addresses the data characteristics common to SUD treatment behavioral coding data. We conclude by demonstrating how to use and interpret these models with data from a study of motivational interviewing. SPSS and R syntax for weighted negative binomial regression models is included in online supplemental materials.