MerQuR+Statistics Glossary

Statistics Glossary → Average marginal effect (AME)

Average marginal effect (AME)concept

In nonlinear models (logit, probit, Tobit), the effect of a change in a predictor on the probability or expected value of the outcome, averaged over all observations in the sample.

A logistic regression coefficient is on the log-odds scale, while its effect on the probability changes with a person's other characteristics. The marginal effect is the derivative of the probability with respect to the predictor for each observation (for a categorical predictor, the difference in probability between 1 and 0). The AME is the sample average of these individual effects and directly answers 'by how many points does the probability change, on average, when this variable increases by one unit'. Its standard error is usually computed with the delta method.

The marginal effect at the mean (MEM) is instead computed for a single 'average person' with every predictor fixed at its mean; because the mean of a dummy variable matches no real person, the AME is usually considered more meaningful. Marginal effects make results from different models or groups easier to compare than odds ratios, but they still depend on the model being correctly specified.

An everyday example

A bank models loan approval on income. The logistic coefficient on its own is not intuitive. An AME of 0.04 means that an extra 10 000 in income raises the approval probability by 4 percentage points on average; the effect is smaller for customers whose probability is already very high or very low and larger for borderline customers.

In MerQur

'Extra output (2): Average marginal effects' in Logistic Regression and 'Extra output: AME' in Probit Regression, with delta-method SEs; in Tobit Regression, 'Marginal effects (McDonald-Moffitt)' are evaluated at the mean of X (effects on the latent y*, the observed E[y], the expectation given uncensored and the probability of being uncensored).

Analyses where this concept appears

Logistic RegressionProbit RegressionTobit Regression (Censored Regression)

Related concepts

Delta methodOdds ratio (OR)