An alternative approach to conventional soil–site regression modeling

Abstract

The conventional approach in site-quality studies has been to develop a multiple regression site index model with soil–site measurements from randomly selected plots. This approach has several weaknessess: (i) a potential prediction bias associated with most stepwise regression procedures; (ii) low precision of soil–site regression models developed in areas with diverse topography and geologic formations; and (iii) poor representation of rare prime sites by random sampling. An alternative approach, aimed at minimizing these problems, is presented. Prediction bias potential (due to overfitting a model with too many predictor variables) can be reduced by using cross validation during model development. Models that accurately predict prime sites can be more useful than imprecise soil–site regression models. This can be accomplished by stratified random sampling from prime and nonprime site areas. Classification-tree analysis was used to develop a model that predicts prime ponderosa pine (Pinusponderosa Laws.) sites on the basis of vegetation and soil variables. Forest habitat type, percent sand content, and soil pH were model predictor variables. Cross-validation was used to estimate the accuracy of the classification tree as 88%. A multiple regression model developed from randomly selected plots consistently underestimated site index when it was applied to plots randomly selected from prime site areas. The conventional regression model was also misleading because it contained a predictor variable that was not significantly different between prime and nonprime sites.