the sum used for prediction. The results of these studies Indicate that 015 and RIDGE , wi th one exception , always outperformed UNIT w ith respect to producing smaller errors of prediction and , what is more Important, RIDGE always did better than OLS. The one exception in which UNIT did better than OLS and RIDGE i s for the case i n whi ch all the " true" coefficien ts are pos iti ve , not too far apart , and the sample size Is relatively small (< 50). This is a very restricted class of

more »

... tions . The general conclusion is that UNIT weighting will be appropriate only in unusua l s ituations . Regress ion models are to be preferred as a way of generating differential weights . Also , the ridge method of estimation (RIDGE) always should be the preferred model over OLS. One practical implication of this is that if an Investi gator does not have the l uxury to do cross validation then RIDGE estimation can be used as a substitute for cross validati on .