Model selection from a very large number of effects is computationally demanding. For example, in analyzing microarray data, where each dot in the array corresponds to a regressor, having 35,000 such regressors is not uncommon. Another source of such large regression problems arises when you want to consider all possible two-way interactions of your main effects as candidates for inclusion in a selected model. See Foster and Stine (2004) for an example that uses this approach to build a predictive model for bankruptcy.

In recent years, there has been a resurgence of interest in combining variable selection methods with an initial screening step that reduces the large number of regressors to a much smaller subset from which the final model is chosen. You can find theoretical underpinnings of this approach in Fan and Lv (2008). See El Ghaoui, Viallon, and Rabbani (2012) and Tibshirani et al. (2012) for examples where screening has also been incorporated in the context of penalized regression methods (such as lasso) for performing model selection.

Screening uses a screening statistic that is inexpensive to compute in order to eliminate from consideration regressors that are unlikely to be selected if you included them in variable selection. For linear regression, you can use the magnitude of the correlation between each individual regressor and the response as such a screening statistic. The square of the correlation between a regressor that has one degree of freedom and the response is the R-square value for the univariate regression for the response with this regressor. Hence, screening by the magnitude of the pairwise correlations is equivalent to fitting univariate models to do the screening.

The first stage of the screening method chooses only the subset of regressors whose screening statistic is larger than a specified cutoff value or by choosing those regressors whose screening statistics are among a specified number or percentage of the largest screening statistic values. Then you perform model selection for the response from this screened subset of the original regressors.

One problem with this approach is that a regressor that is pairwise (marginally) uncorrelated or has very small correlation with the response can nevertheless be an important predictor, but it would be eliminated in the screening. You can address this problem by switching to a multistage approach. The first stage consists of screening the regressors and selecting the model for the response from the screened subset. The second stage repeats the first stage except that you use the residuals from the first stage as the response variable in this second stage. You can iterate this process by using the residuals from the previous stage as the response for the next stage. The final stage forms the union of all the screened regressors from the first stage with all the selected regressors at the subsequent stages and selects a model for the original response variable from this union.

Experimentation has shown that there is little benefit in practice in using more than one stage where the response is the residual from the previous stage. Hence, PROC HPREG implements a three-stage process by default. However, if you specify the SINGLESTAGE suboption in the SCREEN option in the SELECTION statement, then only the first screening stage is performed.