Figure 3

The figure shows the morphological covariates (vertical axis) selected by the algorithm over a range of goodness-of-fit penalty levels for the functional domain WAIS-R (subtest from Wechsler Adult Intelligence Scale) and WCST (Wisconsin Card Sorting Test). The body of the plot shows, in black, which morphological features were selected for that cognitive test at the various penalty levels. The top half of the figure shows the main effects, and the bottom half (covariates with the prefix 'I') interaction with diagnosis. The penalty term is shown on the bottom axis, increasing from left to right. Covariates are often selected at several penalty levels, resulting in a black line stretching across the plot. As the penalty increases, a covariate needs more explanatory power to be selected. One should not infer that the growth in explanatory power is linear. The penalty at the left is 2, corresponding to the AIC. The middle of the graph represents a penalty of 4, the level suggested by the BIC criterion. Selections in this range represent choices that are supported by standard penalized error measures. At the right, the penalty is 6, a level stronger than called for by the common penalized error approaches.