Getting AICc using fitglme

I am using fitglme to get a general linear mixed model, then I use the Model.ModelCriterion to get the AIC and AICc. The strange thing is that although ModelCriterion uses the function modelcriterion, which calculates AIC, AICc, CAIC and BIC, the output does not include AICc or CAIC:

A1.ModelCriterion

ans =

    Model fit statistics

    AIC       BIC       LogLikelihood    Deviance
    1040.1    1057.3    -515.06          1030.1

I looked into the functions creating the general linear mixed model and found out that after getting all the model criterion, it creates a table as the example above, and doesn't save the AICc result anywhere.

I am looking for a way to get this AICc score, would appreciate any idea.

Thank you

Answer 1

[Julian Langowski]

Hi Ron,

I have the same question, but for the function fitlme. Is there an automatic way to get AICc? Thanks for any hints on how to solve this problem.

Best regards,

Julian

[Ernst Schwartz]

by copying out the relevant part of LinearMixedModel.m I think you can get the correct additional values using this function - however, I'm not entirely sure if model.covarianceParameters == model.slme.Psi.NumParametersExcludingSigma in all cases or if that needs to be adapted.

function crit = lmeModelFitInfos(model)

% (1) Get N and p.
N = model.NumObservations; % only contribution from subset.
p = length(model.Coefficients);

% (2) Add the fixed effects + residual variance to NumCoefficients.
stats.NumCoefficients = numel(model.covarianceParameters) + (p+1);

% (3) Get effective number of observations based on FitMethod.
switch lower(model.FitMethod)
    case {'ml'}
        stats.NumObservations = N;
    case {'reml'}
        stats.NumObservations = (N-p);
    otherwise
        % <entry key="BadFitMethod">''FitMethod'' must be either ''ML'' or ''REML''.</entry>
        error(message('stats:LinearMixedModel:BadFitMethod'));
end

% (4) Set maximized log likelihood or maximized restricted log
% likelihood. model.LogLikelihood also accounts for observation
% weights.
stats.LogLikelihood = model.LogLikelihood;

% (5) Call modelcriterion utility function.
crit = classreg.regr.modelutils.modelcriterion(stats,'all',true);


end