Nonparametric models can be fit via the GAM, GAMPL, LOESS, and TPSPLINE procedures. Nonparametric models cannot be represented with a simple equation as can regression and linear models from such parametric modeling procedures as REG, GLM, GENMOD, and others. As a result, there is no simple set of parameter estimates that can be saved in a data set and reused at a later time in a prediction equation to score new data.

Predictions for new data are done as the nonparametric model is fit via the SCORE statement in the GAM, LOESS, and TPSPLINE procedures. For example, these statements fit a spline model in PROC GAM, score the observations in the VALIDATE data set, and save them in the PREDS data set:

      proc gam data=training;
        model y = spline(x);
        score data=validate out=preds;
        run;

In PROC GAMPL, new observations (with response set to missing) can be added to DATA= data set and scores can be generated using the PREDICTED= option in the OUTPUT statement.

Operating System and Release Information