Heckman-Selection or Two-Part models for alcohol studies? Depends

TL;DR

This study demonstrates that the Heckman model effectively corrects for selection bias in alcohol research, outperforming two-part models when assumptions are validated, thus providing more accurate estimates of determinants of problem drinking.

Contribution

It re-introduces the Heckman model as a valid and superior alternative to two-part models in alcohol studies, emphasizing the importance of assumption validation and covariate restrictions.

Findings

Heckman model passes higher order multicollinearity tests.

Heckman model corrects for selection bias in alcohol data.

Two-part model shows biased coefficient estimates.

Abstract

Aims: To re-introduce the Heckman model as a valid empirical technique in alcohol studies. Design: To estimate the determinants of problem drinking using a Heckman and a two-part estimation model. Psychological and neuro-scientific studies justify my underlying estimation assumptions and covariate exclusion restrictions. Higher order tests checking for multicollinearity validate the use of Heckman over the use of two-part estimation models. I discuss the generalizability of the two models in applied research. Settings and Participants: Two pooled national population surveys from 2016 and 2017 were used: the Behavioral Risk Factor Surveillance Survey (BRFS), and the National Survey of Drug Use and Health (NSDUH). Measurements: Participation in problem drinking and meeting the criteria for problem drinking. Findings: Both U.S. national surveys perform well with the Heckman model and pass all higher order tests. The Heckman model corrects for selection bias and reveals the direction of bias, where the two-part model does not. For example, the coefficients on age are upward biased and unemployment is downward biased in the two-part where the Heckman model does not have a selection bias. Covariate exclusion restrictions are sensitive to survey conditions and are contextually generalizable. Conclusions: The Heckman model can be used for alcohol (smoking studies as well) if the underlying estimation specification passes higher order tests for multicollinearity and the exclusion restrictions are justified with integrity for the data used. Its use is merit-worthy because it corrects for and reveals the direction and the magnitude of selection bias where the two-part does not.