We propose a general variable selection procedure to identify key input variables by applying elastic net regression to representative subdata in place of the full sample to select variables. We combine the lists of selected variables from each subdata through ensemble techniques, using the frequency of selecting the variable across different subdata as the final variable selection criteria. Using only variables that are frequently chosen (i.e., 90%), we are able to build a parsimonious model that optimizes predictive accuracy. We adapt this method to the rare event setting and show its application to bankruptcy data in Taiwan. In addition, we show how variable selection is affected by subdata size, , and sampling procedure.