Noise, Kahneman, Sibony, and Sunstein

 

Noise: A flaw in human judgment by Daniel Kahneman, Olivier Sibony, and Cass R. Sunstein keeps showing up in my feeds, so I think I need to read it.

Started: 6/20/2021
Completed: 6/25/2021
Recommendation: Mild Recommendation
Recommended by: Social Media

Review:

It seems like the authors argue that it is best to clear the low-hanging fruit.  Attempting to correct bias is difficult in part because it is hard to identify.  Noise, or basically deviation from the average, however, is relatively straight-forward to identify and there are known ways to minimize noise.  Hence, rather than focusing on bias which is really hard to identify, focus on noise and a side-effect will be that bias (where it can be detected) will be much clearer.

The oft repeated example is shooting at a bull's eye.  If you look at the back of the target, you can see clustered results.  You may not be able to tell if the cluster is around the bull's eye (off by a bias) or not, but you can certainly see whether things are clustered.  The argument is that once you have the clustering complete, then it is easier to detect bias.

The benefits are manifest--ideally if you have no bias, you also want clustered results.  The risk, however, is that you can pull people who are constantly on-target, off-target in order to reduce "noise."  This actually induces bias.  The "reduction to the mean" means that there is an overriding assumption that doing the average thing consistently is better than doing nothing consistently.  This is especially true where it is difficult to determine what the right target is.

So, why not do a noise audit and then try to resolve the noise problem?  We are already working on the more difficult problem of eliminating bias which is a much harder problem, particularly in the presence of noise.  My thinking is as follows:

• Assume that there is a bias (say 3 out of 5 interviewers are white nationalists)
• Work to reduce the noise to the average (this will mean that 2 of the 5 interviewers will need to focus more on white nationalist issues and 3 of the 5 interviewers might have to focus a little less on white nationalist issues)
• Now, we have a cluster where white nationalism is more prevalent in employment
• Bias reduction is implemented and 2 of the 5 interviewers go back to doing what they were doing before, but 3 of the 5 interviewers have to change dramatically

Reducing noise brought no benefit in this contrived example (and actually brought harm, although it wasn't known for sure until the noise had been reduced) and forced 2 of the interviewers to espouse things they felt were wrong.  In the end, the minority position was the correct position and noise reduction moved the group off the correct position.  It is hard to know how often this happens (or if it even does).  It is hard to know how frequently this kind of thing perpetuates bias in the interest of "making things more fair."  I am just not convinced that low-hanging fruit is the best approach in this case.