Multiple linear regression is widely used in empirically-based policy analysis. The central argument of the present paper is that much of this use is inappropriate, not because of the multiple linear regression methodology, but because of the nature of the data used. Too often analysts are carried beyond justified inferences into assertions for which there is essentially no sound defense, leading to policy recommendations of dubious provenance. Four alternative classes of policy interpretations are posited: mere description of data sets, simple prediction, causal models, and causal predictive models. Policy analysis finds statements from the last kind most useful, while multiple linear regression analysis of passively observed data is best suited to supporting statements of the first kind. The paper examines the inferential logic and technical issues that arise as one moves through the four classes. The paper then considers the role multiple linear regression of passively observed data can properly play in policy analysis and suggests alternative approaches.
ASJC Scopus subject areas
- Sociology and Political Science
- Social Sciences(all)
- Public Administration
- Management, Monitoring, Policy and Law