This isn't a pharmaceutical trial but there's no reason why we can't follow along with the author's analogy of 'false negatives'.
Here are his words from 'The Harm of Google's Biases':
I strongly believe in gender and racial diversity, and
I think we should strive for more. However, to achieve
a more equal gender and race representation, Google has
created several discriminatory practices:
He goes on to list his examples, which include the following:
Hiring practices which can effectively lower the bar
for “diversity” candidates by decreasing the false
negative rate
I think you may have misread his position earlier; I noted that you misspoke when talking about this particular statement (i.e., whether the false-negative rate applied to 'diversity' vs 'non-diversity' candidates)
EDIT:
Didn't see your statistical comment. Unfortunately, I can't think of a case where 'adjusting a bar' to decrease the false-negative rate does not neceessarily increase the false-positive rate.
The fact that the author uses the phrase 'lowering the bar' suggests that he is also thinking of this statistical requirement.
> Unfortunately, I can't think of a case where 'adjusting a bar' to decrease the false-negative rate does not neceessarily increase the false-positive rate.
Here's one:
* Phone screens have a 50% false negative rate, and a 0% false positive rate.
* On-site interviews have a 0% false negative rate, and a 0% false positive rate.
Non-diverse candidates do one phone screen, and if passed, do an onsite. Diverse candidates do two phone screens, and if either passes, they go on to the onsite. For all candidates, if the onsite is passed, the candidate gets an offer.
* Unqualified candidates, regardless of diversity, never get an offer.
* Qualified, non-diverse candidates get an offer 50% of the time (the other 50% are erroneously eliminated at the phone interview stage).
* Qualified, diverse candidates get an offer 75% of the time (25% eliminated at the phone interview stage due to false negatives).
It seems to me that the author is advocating either doing two phone interviews for all candidates, or one phone interview for all candidates. Neither change would alter false positive rates, since false positive rates are 0% at all times.
If you're inclined to point out that all tests have a false positive rate >0%, substitute 0% for 0.00000001% and note that I stated the change in false positive rates could be trivially small to the point of omission in the comment above.
I think the quinessential example of this is reminding people about their bias against minorities. Several studies have shown this drastically reduces false negatives and raises false positives to a level that is no higher than the false positive rate in the non-minority population. This isn't enough to say the false positive rate doesn't raise, but it's more than enough to say increasing false positives doesn't imply unqualified.
I think we're drifting away from the author's article.
His statement about 'lowering the bar' with respect to decreasing false negatives directly implies a relationship between candidate quality and the false negative rate.
I can't comment on your hiring analogy, as you have more experience in this area than I do. My responses were only to highlight the subtle negative character of the author's statement.
This isn't a pharmaceutical trial but there's no reason why we can't follow along with the author's analogy of 'false negatives'.
Here are his words from 'The Harm of Google's Biases':
He goes on to list his examples, which include the following: I think you may have misread his position earlier; I noted that you misspoke when talking about this particular statement (i.e., whether the false-negative rate applied to 'diversity' vs 'non-diversity' candidates)EDIT:
Didn't see your statistical comment. Unfortunately, I can't think of a case where 'adjusting a bar' to decrease the false-negative rate does not neceessarily increase the false-positive rate.
The fact that the author uses the phrase 'lowering the bar' suggests that he is also thinking of this statistical requirement.