Regression Analysis and Causal Inference: Cause for Concern?

Abstract

Norman A. Constantine is senior scientist and director, Center for Research on Adolescent Health and Development, Public Health Institute, Oakland, CA, and clinical professor of community health and human development, School of Public Health, University of California, Berkeley. Regular readers of Perspectives on Sexual and Reproductive Health may have noticed a preference among published authors for regression analysis as their primary methodology. In fact, last year, a substantial majority of this journal’s articles employed some form of regression analysis, predominantly logistic regression. Such methodological dominance—which is by no means unique to this journal—supports the need for a critical review of the common uses and misuses of these types of analyses, and a careful examination of the validity threats and issues associated with the claims, conclusions and recommendations that typically result. When used appropriately, regression analysis can be a powerful tool: It allows one to statistically model the relationship between a dependent (outcome) variable and a set of independent (predictor) variables. Linear regression is used with continuous dependent variables, such as number of sex partners or infant birth weight, while logistic regression is used with dichotomous dependent variables, such as a history of pregnancy or STD infection. Both forms of regression allow for the assessment of whether an independent variable (such as age, attitudes, protective behaviors or services received) is associated with an outcome variable while controlling for (statistically removing) the outcome’s overlapping associations with other variables. These types of analyses are generally applied to correlational data, such as survey, census or administrative data, in so-called observational studies. Of course, regression analysis is not needed in every study employing quantitative data. For example, results of abortion surveillance over time and across locations can be usefully presented employing just percentage or rate distributions. The potential power and added complexity of regression analysis are best reserved for either predicting outcomes or explaining relationships. The prediction of outcomes on the basis of current characteristics is possible without regard to the causal relationships among variables. For instance, regression analyses have shown that Asian and college-educated parents are the least likely among all social and demographic subgroups to support human papillomavirus vaccination for their daughters; these fi ndings can be used to identify the need for educational campaigns for these two subgroups even without understanding why they are the least supportive of vaccination. However, to develop effective educational campaigns, it is also necessary to understand the factors that infl uence parents’ support for having their daughters vaccinated. When the goal is to understand (i.e., explain) the causal infl uences on a population outcome—a prerequisite for the design and development of any sexual health intervention—regression analysis can be a powerful tool, but it has some fundamental limitations. Its appropriate use requires substantial care and skill, as well as suffi cient inferential humility.