Information theory offers a coherent perspective on model selection. As in Rissanen's original application of information theory to model selection, our perspective arises from viewing a model as a component of a compressed representation of data in a two-part code. The first part of such a code is an explicit representation of the model used to compress the data. Simpler models have shorter representations. The second part is the encoded data itself. Models which fit better compress the data into shorter sequences. The objective is to choose the model which produces the shortest total message length, requiring an explicit trade-off of model complexity (length of the first part) versus goodness-of-fit (length of second part). In adidtion to Rissanen's MDL criterion, this perspective illuminates the properties of numerous model selection criteria, including AIC, Cp, BIC, RIC, and EBIC. We show that each corresponds to a specific way of coding the model parameters. By selecting the model that minimizes the total message length, our representations of these criteria reproduce their more familiar definitions. Examples from wavelets illustrate the use of these methods.