Using the Margins Command to Estimate and Interpret Adjusted Predictions and Marginal Effects

Abstract

Many researchers and journals place a strong emphasis on the sign and statistical significance of effects-but often there is very little emphasis on the substantive and practical significance of the findings. As Long and Freese (2006, Regression Models for Categorical Dependent Variables Using Stata [Stata Press]) show, results can often be made more tangible by computing predicted or expected values for hypothetical or prototypical cases. Stata 11 introduced new tools for making such calculations-factor variables and the margins command. These can do most of the things that were previously done by Stata's own adjust and mfx commands, and much more. Unfortunately, the complexity of the margins syntax, the daunting 50-page reference manual entry that describes it, and a lack of understanding about what margins offers over older commands that have been widely used for years may have dissuaded some researchers from examining how the margins command could benefit them. In this article, therefore, I explain what adjusted predictions and marginal effects are, and how they can contribute to the interpretation of results. I further explain why older commands, like adjust and mfx, can often produce incorrect results, and how factor variables and the margins command can avoid these errors. The relative merits of different methods for setting representative values for variables in the model (marginal effects at the means, average marginal effects, and marginal effects at representative values) are considered. I shows how the marginsplot command (introduced in Stata 12) provides a graphical and often much easier means for presenting and understanding the results from margins, and explain why margins does not present marginal effects for interaction terms.

Emergency readmissions are substantially determined by acute illness severity and chronic debilitating illness: A single centre cohort study

“…We used margins to estimate and interpret adjusted predictions for sub-groups, while controlling for other variables, using computations of average marginal effects [31]. Margins are statistics calculated from predictions of a previously fitted model at fixed values of some covariates and averaging or otherwise over the remaining covariates.…”

Does selection bias explain the obesity paradox among individuals with cardiovascular disease?

“…52). We used the -margins-routine in Stata 12 to estimate these marginal contrasts from logistic regression, with standard errors and confidence intervals estimated using the linearization method to account for the complex survey design and possible violations of distributional assumptions(52,53).…”

“…We used the -margins-routine in Stata 12 to estimate these marginal contrasts from logistic regression, with standard errors and confidence intervals estimated using the linearization method to account for the complex survey design and possible violations of distributional assumptions(52,53). Since AMEs are marginal risk differences, we can define the total and controlled direct effects on the linear scale:Total effect: E[mortality(obese=1)-mortality(obese=0)| C=c) Controlled direct effect among those with CVD: E[mortality(obese=1, CVD=cvd)-mortality(obese=0, CVD=cvd)| C=c]…”

An investigation into the credibility of currency board arrangements in Bosnia and Herzegovina and Bulgaria

“…These answers for the dependent variable may also be included by creating a separate category of the dependent variable and estimating the model with multinomial probit, since this estimator allows more than two non-ordinal categories of the dependent variable. However, since we have interaction terms in our preferred model, the interpretation of the results from this estimator is very complicated (Williams, 2012) and therefore we prefer the probit estimator. However, in the robustness check multinomial probit, with "do not know" answers included, is estimated.…”