statsmodels.tools.eval_measures.aic_sigma¶
-
statsmodels.tools.eval_measures.aic_sigma(sigma2, nobs, df_modelwc, islog=False)[source]¶ Akaike information criterion
- Parameters
- sigma2float
estimate of the residual variance or determinant of Sigma_hat in the multivariate case. If islog is true, then it is assumed that sigma is already log-ed, for example logdetSigma.
- nobsint
number of observations
- df_modelwcint
number of parameters including constant
- Returns
- aicfloat
information criterion
Notes
A constant has been dropped in comparison to the loglikelihood base information criteria. The information criteria should be used to compare only comparable models.
For example, AIC is defined in terms of the loglikelihood as
\(-2 llf + 2 k\)
in terms of \(\hat{\sigma}^2\)
\(log(\hat{\sigma}^2) + 2 k / n\)
in terms of the determinant of \(\hat{\Sigma}\)
\(log(\|\hat{\Sigma}\|) + 2 k / n\)
Note: In our definition we do not divide by n in the log-likelihood version.
TODO: Latex math
reference for example lecture notes by Herman Bierens
References