A “Confusion Matrix” for Incentives

Nobody gets everything right, especially not a social scientist with a computer. Armed with data and a model, it’s pretty common for someone to ask me to make predictions about a bunch of individual outcomes, or classify a bunch of people into one category and not another. A confusion matrix is a way to summarize just how good I am doing at making correct predictions or classifications. You can group the false positives, correct positives, false negatives, and true negatives into all kinds of statistics — accuracy, sensitivity, and all the rest — depending on which kinds of mistakes you want to avoid the most. The result is a confusion matrix.

I love the juxtaposition of words: hairy, messy “confusion” and logical, uptight “matrix.” Real meets ideal. Confusion matrix.


Another group of people who don’t get everything right includes leaders and managers. Say you are a leader who wants to create change in your organization, so you set out some clear incentives for units to work harder and make progress toward some important goals. And because you mean business, you are going to put some teeth on the incentives: allocate budget dollars, paid time off, or whatever stirs the passion of people in the different units. People go off and do their thing, time passes, and now it’s time to see how everyone did.

About those incentives: do you reward effort, or results? I remember being fascinated in graduate school by articles about the “principal-agent problem” by Bengt Holmstrom , and grinding through more difficult (to me) articles about incentive design by Rogerson, Maskin, Arrow, and others. Lots of econ Nobel prizes in those areas of research lately, so I suppose they solved all the solvable problems with the right Greek letters and numbered equations. But I don’t know if they solved the communication part.


A “Confusion Matrix” for Incentives; John L. King

No matter how much effort, creativity, and innovation the units apply to the important goals, some won’t get the results you are looking for. And some units won’t apply much effort at all, and somehow get results you are looking for (at least, by the metrics you have chosen). So whether you reward success, penalize failure, or both, it is possible to construct simultaneous positive messages and negative messages for each scenario. Maybe these will be said to you…or maybe these will be said about you.


Is there a way to avoid this problem of incentives? And if not, is there a way to create trust and goodwill so the positive message is spoken, received, and believed? What do you think?

How confusing.