My newly-discovered penchant for freezing up and forgetting rudimentary facts during interviews makes me question the sensitivity and specificity of a multi-round interview process. That also makes me question whether such thing as a ground truth for good employees exists (maybe performance on the job?)

Anyway, once we have a ground-truth, we can perform an ROC analysis at various performance “thresholds” (note that we need only interview a population of candidates once; the criteria are the stringency of selection). The area underneath the curve would yield a measure of the method’s accuracy, while the curve itself would illustrate the relationship between sensitivity and false positive rate (that is, how many good candidates are hired vs. how many bad ones are hired). Of course, what we’re really interested in is the Positive Predictive Value: how efficiently does the interview process select good employees?

I wonder if any studies were performed on this? Hopefully so; I’d hate to think that such a stressful process exists simply based on tradition.