Economics and incentives
Incentives drive behaviour
People do more of what a system rewards and less of what it punishes, so to predict behaviour, look at what gets paid for.
A founding assumption of economics. Charlie Munger made it a mental model in "The Psychology of Human Misjudgment", the talk he revised in 2005 for Poor Charlie's Almanack. The line "show me the incentive and I will show you the outcome" is widely credited to him, but no dated transcript of it was found for this entry.
Whatever a system rewards, it gets more of, and whatever it penalises, it gets less of, whether or not anyone intended that. Munger opens his talk with Federal Express, which could not get its night shift to load the planes on time until it stopped paying by the hour and paid by the shift, letting people go home once the planes were loaded. His conclusion: "Never, ever, think about something else when you should be thinking about the power of incentives." The model is about the design of systems (salaries, bonuses, fines, targets, metrics), not about any one person's character. It says the outcome you will see is the one the rules make cheapest and most rewarding to produce.
It holds because people are not passive recipients of rules. They read how a rule changes what they gain and lose and adjust, usually faster than the rule's designer expects. Munger quotes Benjamin Franklin: "If you would persuade, appeal to interest and not to reason." In practice the test for any policy is three questions: who gets paid, for what, and what is the cheapest way to collect? The answer is the behaviour you will get. It explains why commissions, citation counts and quarterly targets so often produce the number without the thing the number stood for, and why Munger's Xerox example, where salesmen pushed the worse machine because it paid a higher commission, keeps repeating in new clothes.
Money is not the only incentive and adding it can backfire. People respond to the incentive they perceive, which includes status, approval and the wish to avoid blame, and a cash reward can crowd out a motive that was already doing the work, as the day-care fine below shows. The model is also misread as cynicism, as if everyone were only in it for the payout; the point is narrower, that at the margin behaviour bends toward the reward. So treat it as a question to ask before you attach money to anything: what do people care about now, and what will the new payment replace? For how incentives bend your own judgement without your noticing, see incentive-caused bias.
Example
In a field study in Haifa, Israel, Uri Gneezy and Aldo Rustichini watched 10 private day-care centres for 20 weeks. After four weeks of baseline, six of the centres began fining parents 10 new Israeli shekels (NIS), under three US dollars at the time, for collecting a child more than 10 minutes late. Late pick-ups did not fall. They rose steadily for three to four weeks and settled at almost twice the initial rate, and when the fine was removed at the start of week 17 they stayed at the higher level. The paper, "A Fine Is a Price", appeared in the Journal of Legal Studies in 2000.
Where it applies
Before you set a bonus, a sales target or a team metric, ask what the cheapest way to hit it is, because that is what you will get. In machine learning the same rule governs the model: a reward function is an incentive, and a reinforcement learning (RL) agent will find whatever behaviour scores highest, which is rarely exactly what you meant.
Think about it
What does your team actually get rewarded for this week, and where does that differ from what you say you want?
Applied it to something real? Log it.
Test yourselfA day-care centre starts fining parents who collect their children late. What did the Haifa study find, and why?
Late pick-ups nearly doubled and stayed high after the fine ended. A small fine turned a social obligation into a price, so the incentive you add can replace the one that was already working.
Related models
- Incentive-caused biasAn incentive bends your own judgement towards the conclusion that pays, and you will not feel it happening.
- Goodhart's lawWhen a measure becomes a target it stops being a good measure, because people optimise the number instead of the thing it stood for.
- The principal-agent problemWhen you pay someone to act for you, their interests differ from yours and you cannot see all they do, so expect them to serve themselves at the margin.
- Second-order thinkingKeep asking "and then what?" until you have looked at the consequences of the consequences.
- Moral hazardWhen someone else will bear the cost of your risks, you take more of them.