I don't know if you have read Thinking Fast and Slow but your observations are in strong agreement with research.
From what I have read (mostly Kahneman) most interviews as performed are useless and suffer from the illusion of validity. Instead, structured interviews and scoring that are the same for all candidates are strongly recommended. Which is, draw up a short list ~= 6 independent factors which from experience are reflective of the job and whose answers you can easily judge objectively. For each factor come up with a small number of questions with a score, say from 1 - 5 and then simply add up the scores. No fancy weighting required - research shows that regression does not add much to this simple model. To avoid the "halo effect" do not jump question categories. This will vastly improve upon the more common unstructured interview and though still far from perfect you will be hard pressed to improve on that short of an IQ or equivalent test.
Interestingly, and reminiscent to bagging in machine learning, unstructured interviews can be brought nearly up to par by averaging multiple independently scoring interviewer opinions.
From what I have read (mostly Kahneman) most interviews as performed are useless and suffer from the illusion of validity. Instead, structured interviews and scoring that are the same for all candidates are strongly recommended. Which is, draw up a short list ~= 6 independent factors which from experience are reflective of the job and whose answers you can easily judge objectively. For each factor come up with a small number of questions with a score, say from 1 - 5 and then simply add up the scores. No fancy weighting required - research shows that regression does not add much to this simple model. To avoid the "halo effect" do not jump question categories. This will vastly improve upon the more common unstructured interview and though still far from perfect you will be hard pressed to improve on that short of an IQ or equivalent test.
Interestingly, and reminiscent to bagging in machine learning, unstructured interviews can be brought nearly up to par by averaging multiple independently scoring interviewer opinions.