The Azimuth Project
Goodhart's law

Goodhart’s law

Idea

Goodhart’s law is a phenomenological observation about attempting to use performance indicators to achieve ‘quality control’ in a market or sociological system. A slightly opaque formulation is:

Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes.

Details

Goodhart’s law is easiest to understand through an example:

Consider trying to improve the performance of some organisation, for example, the efficacy of a hospital ambulance service. Quantifying what makes a good ambulance service is complicated, but one might make a survey of existing ambulance services might find that the ones which one thinks of as good have teams who get personnel to a call site within 8 minutes, and that conversely those ambulance services which have problems tend to take significantly longer to reach the accident site.

So ‘to improve the ambulance service’ one could introduce either a reward for having personnel reaching a call site within 8 minutes, or a penalty for failing to do so. However, when this was actually tried in the UK it led to the introduction of additional ‘rapid response cars’ which had much greater speed and handling than an ambulance at the expense of only being able to carry a paramedic with very little medical equipment. The practical effect was that adding these vehicles to the fleet resulted in personnel arriving within 8 minutes at almost every call, but very often being unable to do more than observe and wait around for a conventional ambulance containing equipment to arrive before being able to start treatment.

In terms of Goodhart’s law, there was initially the statistic of response time which was a very good indicator of quality. When it was used to attempt to control the quality (by rewarding that criterion), those being controlled took actions which caused response time to become a very unreliable statistic.

Applicability

This is clearly an important consideration in any attempt to control behaviour using rewards/penalties based on some prespecified benchmarks.

References

Much more information can be found at

category: methodology