Improving discrimination in data envelopment analysis: PCA–DEA or variable reduction