This paper discusses whether differences in the data structure of observational and experimental studies should lead to different strategies for variable selection. On the one hand, it is argued that outliers in the predictor variables have to be treated differently in the two kinds of studies. In experimental studies this results in philosophical problems with the applicability of cross validation. On the other hand, it is shown, however, that a well designed experiment might lead to a factor structure very appropriate for cross validation, namely a certain balance in the observations together with orthogonality of the factors. This might be the reason why in practice cross validation has proven to be a valuable tool for variable selection also in experimental studies. In contrast, however, it is shown that variables selection based on cross validation is not appropriate for saturated orthogonal designs. After this fundamental argumentation, we illustrate by a number of examples that the same methods for variable selection can be successfully applied in observational as well as experimental studies.
variables selection stepwise regression cross validation principal components screening optimization