How are the crossvalidation statistics defined in Partition?

[#1]

The k-fold crossvalidation method randomly divides all of the (non-excluded) rows in the datatable, D, into k subsets: D₁, D₂, . . . , D_k. Each row is randomly assigned to one of the k groups.

First, a model is fit using all of the data. Then, K different models are fit, using the same splits, with data from D - D_i, where D_i is the holdout fold. For continuous responses, the error for each observation in D_i is calculated using the model, evaluated from D - D_i. For nominal and ordinal responses, JMP^® calculates -2LogLikelihood for each observation in D_i using the model, trained from D - D_i. This is repeated for each of the k folds. The root mean square errors for each of the k folds are squared and summed (or in the case of nominal and ordinal responses, the -2loglikelihood values are summed) to construct the crossvalidation SSE (or crossvalidation -2loglikelihood in the case of nominal and ordinal responses). The resulting folded SSE is then used to calculate a folded R square value.

[Previously JMP Note 36388]