MerQuR+Statistics Glossary

Statistics Glossary → Information criteria (AIC, AICc, BIC)

Information criteria (AIC, AICc, BIC)concept

Criteria for comparing models that weigh their fit to the data (likelihood) against their complexity (number of parameters); the model with the smaller value is preferred.

The Akaike information criterion is AIC = -2 ln L + 2k and the Bayesian (Schwarz) information criterion BIC = -2 ln L + k·ln(n), where L is the maximized likelihood, k the number of parameters and n the sample size. Better fit makes -2 ln L smaller, while each extra parameter adds a penalty. AICc is the small-sample correction of AIC: AICc = AIC + 2k(k + 1)/(n - k - 1), preferred when n/k is small. BIC's penalty is larger than AIC's once n ≥ 8, so BIC selects more parsimonious models.

The absolute value of an information criterion means nothing; only differences between models fitted to the same data are interpreted. For AIC, differences below 2 indicate similar support, and differences above 10 a strong advantage. AIC aims at the model that predicts best, while BIC aims at finding the true model (if it is among the candidates). Non-nested models can also be compared, but the outcome must be modelled in the same form (not logged in one and raw in the other) and on the same observations. Mixed models estimated by REML that differ in their fixed effects should not be compared with AIC.

An everyday example

A weather team tries two models for daily rainfall: one with temperature only (AIC = 512) and one that adds humidity and wind (AIC = 498). The 14-point difference shows that the gain in fit from the extra variables far outweighs the complexity penalty. When a third model also adds cloud type, AIC drops to 497; a 1-point difference suggests the extra complexity is not worth it.

In MerQur

Available as 'Selection criterion' (BIC / SABIC / AIC / CAIC) in latent class and latent profile analysis, 'Bandwidth criterion' (AICc / AIC / BIC / CV) in GWR, 'Extra output: AIC/BIC + form comparison' in nonlinear regression, the 'AIC/BIC table across distributions' in parametric survival and 'Model comparison (AIC/BIC)' in item response theory; ARIMA's automatic model selection uses AICc. 'Lag selection' (AIC / HQ / SC (BIC) / FPE) in VAR, lag selection in the time series, cointegration and panel unit root tests, the two-directional AIC variable selection in Cox regression and 'Span selection' (AICc) in LOESS can use these criteria; Bayesian ANOVA and hierarchical Bayesian regression give BIC-based Bayes factors. AIC and BIC are reported in the model fit output of the mixed models, factor analyses, structural equation and path models, count and survival models and GARCH.

Analyses where this concept appears

Poisson / Negative Binomial Regression (Count Regression)Multinomial Logistic RegressionLinear Mixed Models (LMM)Crossed Mixed Model (Crossed LMM, Crossed Random Effects Mixed Model)Panel Data RegressionLatent Class Analysis (LCA)Latent Profile Analysis (LPA)Exploratory Factor Analysis (EFA)Confirmatory Factor Analysis (CFA)Structural Equation Modeling (SEM)Item Response Theory (IRT / Rasch)Variance Components Analysis (VARCOMP)Generalized Additive Model (GAM, Nonlinear Regression)Nonlinear RegressionProbit RegressionBayesian ANOVAHierarchical Bayesian RegressionGWR (Geographically Weighted Regression)Path AnalysisCox Regression (Proportional Hazards Model)Parametric Survival (AFT)Time-dependent Cox RegressionInterval-Censored SurvivalTime SeriesARIMA ForecastExponential Smoothing (Holt-Winters)VAR (Vector Autoregression)Cointegration (Johansen / Engle-Granger)GARCH (Volatility Model)LOESS (Local Regression)

Related concepts

Likelihood ratio test (LRT)Stepwise variable selection