Should citations be field-normalized in evaluative bibliometrics? An empirical analysis based on propensity score matching