The Conceit

The conceit feels like competence. The driver trusts his judgment, the manager trusts her plan, the forecaster trusts the narrow interval within which tomorrow is supposed to arrive. Each may have good reasons; each can also mistake the possession of reasons for the possession of enough of them. Overconfidence begins at that excess, but the excess has several objects. We can overestimate our own performance, overplace ourselves relative to others, or be too precise about what remains uncertain. A manager who expects a disappointing result may still exaggerate her lead over competitors; a forecaster who expects disaster may be unjustifiably certain of its date. The confidence that matters is a relation between a claim and its warrant, and the claim must first be made clear.¹

Even the arithmetic of being above average depends on which average we mean. Consider five invented scores: zero, eighty, eighty-five, ninety and ninety-five. Their mean is seventy. Four people stand above it, without vanity or mathematical offence; the low score has pulled the mean beneath most of the group. Their median is eighty-five, and only two stand strictly above that. More than half of a group cannot stand strictly above its own median on one common measure. Yet the common measure matters too. One driver may prize caution, another speed of response, another skill in difficult weather. If each silently chooses the dimension on which he excels, several favourable judgments need not describe the same contest. Before convicting the self of an impossible rank, we must establish the ranking it claims.²

A classic driving study made the comparison more definite. In 1981, Ola Svenson asked licensed student drivers in the United States and Sweden to locate their skill or safety among the other participants, using successive percentile bands. In both samples, majorities assigned themselves positions above the middle of their group. These judgments could not all be accurate rankings of the same group on a common criterion. The result gives the familiar conceit a particular setting and a measurable shape. It does not establish that every driver is overconfident, or that the people claiming high skill were bad drivers: the study elicited self-ratings without independently testing their driving. Nor does an aggregate discrepancy decide whether pride, private definitions of skill, or incomplete knowledge of other drivers supplied the mistake. The evidence is sharper when it is allowed to remain specific.³

The picture becomes stranger when confidence in a score and confidence in a rank move in opposite directions. In Moore and Healy’s trivia experiment, participants on easy quizzes underestimated their own scores while overplacing themselves relative to another participant; on hard quizzes, they overestimated their scores while underplacing themselves. Their account starts from an ordinary imbalance of information. We know more about our own performance than another person’s, so estimates of the other person may remain closer to an earlier expectation. When a task proves unexpectedly easy, my estimate of myself rises farther than my estimate of you; the comparison can flatter me even while I underestimate what I achieved. When it proves hard, the pattern can reverse. This is one account, supported by their experiment, with limits the authors make explicit. It gives the self’s apparent vanity a possible informational source and makes the task, as well as the person, part of the explanation.⁴

Overprecision survives as a different charge: the future has been given too little room to disagree. A forecast can put its centre in a plausible place and still draw its boundaries too tightly. In the same experiment, the ranges nearest participants’ ninety-percent confidence intervals contained the actual scores less often than the stated confidence warranted. Yet those who supplied broader intervals also tended to make larger errors in estimating scores; uncertainty had some information in it even when its amount was inadequate. The lesson cannot be reduced to making every forecast vaguer. Calibration asks whether the confidence assigned to comparable forecasts matches their record across enough cases. One event outside a ninety-percent interval is compatible with the interval’s promise. A persistent excess of such misses calls the promise into question. A curve purporting to describe everyone’s confidence, on every kind of question, would commit the very error it illustrates.⁵

The costs can grow with the authority of the person making the estimate. Imagine a project director whose narrow completion range determines the budget, the staffing and the promises made to customers. An omitted contingency becomes everyone else’s emergency. A defensible response begins before the result is known: preserve the forecast, its assumptions, its reference cases and the circumstances under which it should change. Otherwise a plan revised after each disappointment can always look more prescient than the plan that actually governed the decision. The same discipline separates bold action from inflated belief. If attempting a project costs little and success would be valuable, a modest probability may justify proceeding. There is no need to raise that probability merely to make the decision feel brave. The threshold for acting belongs to the payoffs; the estimate of success belongs to the evidence.

Confidence does work that accuracy alone cannot perform: it can sustain an attempt, secure cooperation and carry a project through ordinary doubt. Those possible benefits deserve examination without turning them into a story that every flattering belief must once have helped its owner survive. The earlier question therefore returns in a more exact form. Many of us may place ourselves too high in particular comparisons; we may also place ourselves too low, or understand our rank while claiming too much certainty about it. The corrective is a practice of asking what the assertion means, what comparison supports it, and what record could change it. Competence can bear those questions. Conceit borrows competence’s assurance while refusing its exposure to a result that might teach it something.