Statistical inference is, essentially, about
identifying the best model for a set of data. However, “best” entails a
tradeoff between model fit and model flexibility: A more flexible model
will generally fit better, but at the potential cost of fitting random
noise. Consequently, measuring model flexibility is a key problem. For
linear models, model flexibility can be measured by counting model
parameters, but what about nonlinear models? In the present talk, I
describe a technique for measuring the flexibility of nonlinear (and
constrained linear) models by finding the “effective” number of model
parameters, providing a common index of flexibility across different
model types.
.