Akaike information criterion: what it means and how to read it

How to read it

For choosing between models of the same Y on the same rows: the lower AIC, the better the balance of fit and simplicity; a difference under about 2 is negligible, the value itself means nothing. Compare only converged models without lambda: with ridge or without convergence AIC is not the usual criterion.

In plain words

A teacher had two ways to guess who passes. The first used only study hours, the second used study hours and shoe size. She wanted to know whether shoe size was worth keeping. She counted AIC and got 62.8 for the first way and 64.5 for the second. Lower is better, so the first way wins, and a gap under 2 is tiny anyway. Now she can say shoe size brings nothing and the simpler way is enough. The number itself means nothing; only the gap between two ways does, and only on the same students with lambda at 0.

How it is computed

AIC = −2LL + 2k

Example

The cell with AIC is selected in each report; point at a label to see its note.

Where you see it

• STATS_LOGISTIC — Binary logistic regression with optional ridge regularization (lambda >= 0).

Related values

• Bayesian information criterion — −2 Log Likelihood plus a penalty of ln(n) for every coefficient.
• Omnibus test of the model — Whether the model as a whole predicts better than the intercept alone — the likelihood-ratio test. It is the Model row of the SPSS table of the same name: at a later step it tests the whole model, not only the predictors that step adds (SPSS's Step and Block rows are not reported).
• Intercept — The constant of the model: the prediction when every predictor is 0.
• Coefficient B — The coefficient of the predictor.
• Standard error of B — How much B would vary from sample to sample.

Get it in your sheet

One formula in a cell gives the whole report, Akaike information criterion included. ALS Stats is a free add-on for Google Sheets™.

Install ALS Stats from Google Workspace Marketplace (opens in a new tab)