Models with more free parameters typically 'fit' the data better,
i.e., model error variance is reduced, but at a cost that has to
be considered.
The trade-off is that bias in parameter estimates increases, and
prediction uncertainty increases, as the likelihood of the model
improves (i.e., as the data are better explained by the model).
Too few parameters tend to underfit the data, too many overfit
the data - both extremes increase estimation bias.
You may know the principle of parsimony in non-statistical terms
if you are familiar with the concept of 'Occam's razor', it advocates
favouring the simplest hypothesis consistent with your knowledge.
Notwithstanding a failure of goodness-of-fit
diagnostics, the 'best' model is one that optimally balances the
competing stresses of better model fit and increased bias and parameter
uncertainty.
Fortunately, inference and ranking tools such as Akaike's Information
Criterion (AIC or AICc, QAIC, QAICc), Bayes Information Criterion
(BIC), and likelihood ratio tests assist investigators in deciding
upon the 'best' model.
However, the 'best' model may not be a good enough model if it
fails goodness-of-fit diagnostics or
retrospective analysis.
|