You Get What You Pay For

A colonial administration in Delhi is said to have grown concerned about cobras and offered a bounty for dead ones. People began breeding cobras. When the scheme was discovered and cancelled, the now-worthless snakes were released, and the city ended up with more cobras than it started with.

The story may well be embroidered. The pattern is documented beyond dispute: a French bounty on rat tails in Hanoi produced a population of tailless rats, which had been caught, docked and released to breed. Paying archaeologists per artefact recovered has produced artefacts broken into several pieces. Paying software teams per bug fixed has produced bugs.

The general form is simple enough to state and remarkably hard to remember: you get what you pay for, and what you pay for is never quite what you want.

The reason is that the thing you want is usually unmeasurable and the thing you pay for is a proxy. You want a healthy population; you can measure appointments. You want good teaching; you can measure test scores. You want a productive team; you can measure lines of code, tickets closed, hours logged. Each proxy is chosen because it correlates with the real objective, and each correlation exists because nobody was trying to move it.

Which is Goodhart's Law: when a measure becomes a target, it ceases to be a good measure. The correlation that made the indicator worth watching is a fact about behaviour that was not optimised for it, and applying pressure changes that behaviour. Ambulance response times in England improved after targets were set, partly through genuine improvement and partly through re-categorising calls. Hospital waiting lists shortened partly by changing the point at which the clock started. Nobody involved needs to be cynical: each local adjustment is defensible, and the aggregate is a number that has come loose from the thing it was standing for.

There is a deeper version of the problem, and it is the one that most surprises people who think of incentives as purely additive. In 1998 Uri Gneezy and Aldo Rustichini studied nurseries in Haifa where parents were sometimes late collecting their children. Some of the nurseries introduced a small fine for lateness.

Late collections roughly doubled.

The interpretation generally offered is that lateness had been governed by an obligation to the staff (you were imposing on a person who wanted to go home) and the fine replaced that obligation with a transaction. Ten shekels for twenty minutes is a fair price for childcare, and parents who would have been mortified to keep someone waiting were perfectly comfortable paying for the service. The finding that makes this more than a curiosity is what happened when the fine was removed: the rate did not return to where it had been. The norm had been displaced and did not come back.

So an incentive is not necessarily added to the motives already present. It can substitute for them, and the substitution can be one-way. This should make anyone designing a scheme nervous about domains currently running on professional pride, reciprocity or shame, because those are unpriced and are doing more work than the balance sheet shows.

It also cuts against the caricature of the economic view of human beings. The claim is that people respond to incentives, and incentives include status, convenience, inconvenience, guilt, habit, and the wish to see oneself as a decent person. Some of the most effective policy interventions ever measured cost nothing: switching organ donation or pension enrolment from opt-in to opt-out moves participation by tens of percentage points, purely by changing which option requires an action. Treating "incentive" as a synonym for "money" is the single commonest way of getting incentive analysis wrong, and it is a mistake made just as often by economics' critics as by its practitioners.

Behind most of this sits one recurring structure, and it is worth having the name. A principal-agent problem exists whenever someone acts on your behalf, knows things you cannot verify, and has interests that do not perfectly match yours. Shareholders and executives. Patients and doctors. Voters and ministers. You and the mechanic who has your car up on the ramp.

Almost every institution you deal with is an attempt to manage one of these. Share options exist to tie an executive's wealth to the shareholders'; they also create an incentive to raise the share price over a specific horizon by whatever means are available. Professional licensing exists so that patients need not evaluate a diagnosis; it also restricts entry and raises prices. Paying a hospital per procedure raises the number of procedures, including the ones that were not worth doing; paying per patient instead creates an incentive to do too little and to avoid difficult cases; paying for outcomes invites selecting patients likely to do well.

There is no scheme without a distortion. What there is, at best, is a choice about which distortion you would rather have and which you can monitor most cheaply, and the first sign that a scheme was designed by someone not thinking clearly is the claim that it aligns everybody's interests perfectly.