The implication of selection bias in clinical studies of late life depression: an empirical approach