Goodhart's Law and Market Research
This weekend I came across Goodhart’s Law. Like a lot of “laws” it’s more Goodhart’s Observed Tendency: it basically says that when you start basing policy around an economic indicator, the information value of that indicator falls to zero. It’s a kind of decision-making equivalent of the observer effect, as seen in physics. Or, if you like, a fancy way of saying “a watched pot never boils”.
Now in its original form - talking about government policy - it’s a sly way of asserting that government intervention is useless and so economic agents should be left to their own devices. But I suspect Goodhart’s Law - or something like it - also applies at every other level of measurement activity, including the firms the majority of said agents operate in.
How might this actually work? We design metrics to simplify complex systems. But when a value designed to describe a system becomes a way of assessing success within that system, two things can happen:
- the system adapts to reflect (and therefore game) the metric
- the stuff not reflected in the metric goes unnoticed, becoming a big breeding ground for potential unintended consequences.
And the more important the metric is, the more it gets gamed.
Anyone who’s spent any time working with social media - looking at “buzz”, reputation systems, measuring “influence” etc. - won’t find these ideas particularly foreign: I wasn’t surprised to find good blog posts from a couple of years ago talking about it. You can see it happen on Twitter, say, with “follower counts”.
Over in the world of market research, though, these ideas aren’t quite as commonly expressed and I suspect many wouldn’t agree with them. In this business an aphorism I’ve heard quite a lot is “if you can’t measure it, you can’t manage it”. What Goodhart’s Law is implying is the inverse: “once you measure it, it manages you”
Let’s assume Goodhart’s Law is at least poking at a greater truth about the unreliability of metrics to assess complex systems. What can decision-makers do about it?
1. Abandon the metrics.
2. Double down on Goodhart’s Law and increase the importance of a metric until removal of the metric would crash the system.
3. Increase the opacity of the metric so gaming it becomes significantly more difficult.
All of these, of course, have their risks :)
