Correcting Sample Selection Bias of Historical Digital Trace Data