Pragmatische und zielgerichtete Variablenauswahl für lineare, logistische und Cox-Modelle
We present a new SAS macro %ABE that can be used for variable selection by combining backward elimination based on significance and the change-in-estimate criterion. By the latter criterion, non-significant variables are retained in a model if their exclusion leads to a relevant change in the parameter estimates of other variables in the model. The macro handles linear, logistic and Cox regression models. By standardizing the change-in-estimate criterion its application is independent from the scaling of the explanatory variables, and in linear models it is also independent from the outcome variable's scale. Furthermore, we calculate the change-in-estimate criterion on the effect size estimates usually reported and interpreted in a model, i.e., odds ratios in logistic regression or hazard ratios in Cox regression. A computationally efficient approximation to the change-in-estimate is used to decrease computational burden caused by evaluating many candidate models.
Augmented backward elimination extends the ideas of "purposeful variable selection" by Hosmer, Lemeshow and May (1999, Chapter 5), who proposed that the analyst should adapt variable selection to the specific modeling problem. The SAS macro %ABE is fully documented in a Technical Report.
Referenzen
Dunkler, D., Plischke, M., Leffondré, K., Heinze, G. (2014): "Augmented backward elimination: A pragmatic and purposeful way to develop statistical models", PLOS One, 9 (11):e113677 (doi:10.1371/journal.pone.0113677).
Hosmer D.W. Jr. , Lemeshow S., May S. "Applied survival analysis: Regression modeling of time to event data", Chapter 5 - Model developement. John Wiley & Sons: 1999.
Download
The SAS macro %ABE is available under a General Public License Version 2 (GPL-2).