Paul Meehl was a strange and wonderful character, and one of the most versatile psychologists of the twentieth century. Among the departments in which he had faculty appointments at the University of Minnesota were psychology, law, psychiatry, neurology, and philosophy. He also wrote on religion, political science, and learning in rats. A statistically sophisticated researcher and a fierce critic of empty claims in clinical psychology, Meehl was also a practicing psychoanalyst. He wrote thoughtful essays on the philosophical foundations of psychological research that I almost memorized while I was a graduate student. I never met Meehl, but he was one of my heroes from the time I read his Clinical vs. Statistical Prediction: A Theoretical Analysis and a Review of the Evidence.
In the slim volume that he later called “my disturbing little book,” Meehl reviewed the results of 20 studies that had analyzed whether clinical predictions based on the subjective impressions of trained professionals were more accurate than statistical predictions made by combining a few scores or ratings according to a rule. In a typical study, trained counselors predicted the grades of freshmen at the end of the school year. The counselors interviewed each student for forty-five minutes. They also had access to high school grades, several aptitude tests, and a four-page personal statement. The statistical algorithm used only a fraction of this information: high school grades and one aptitude test. Nevertheless, the formula was more accurate than 11 of the 14 counselors. Meehl reported generally similar results across a variety of other forecast outcomes, including violations of parole, success in pilot training, and criminal
recidivism.
Not surprisingly, Meehl’s book provoked shock and disbelief among clinical psychologists, and the controversy it started has engendered a stream of research that is still flowing today, more than fifty yephy Љ diars after its publication. The number of studies reporting comparisons of clinical and statistical predictions has increased to roughly two hundred, but the score in the contest between algorithms and humans has not changed. About 60% of the studies have shown significantly better accuracy for the algorithms. The other comparisons scored a draw inaccuracy, but a tie is tantamount to a win for the statistical rules, which are normally much less expensive to use than expert judgment. No exception has been convincingly documented.
The range of predicted outcomes has expanded to cover medical variables such as the longevity of cancer patients, the length of hospital stays, the diagnosis of cardiac disease, and the susceptibility of babies to sudden infant death syndrome; economic measures such as the prospects of success for new businesses, the evaluation of credit risks by banks, and the future career satisfaction of workers; questions of interest to government agencies, including assessments of the suitability of foster parents, the odds of recidivism among juvenile offenders, and the likelihood of other forms of violent behavior; and miscellaneous outcomes such as the evaluation of scientific presentations, the winners of football games, and the future prices of Bordeaux wine. Each of these domains entails a significant degree of uncertainty and unpredictability. We describe them as “low-validity environments.” In every case, the accuracy
of experts was matched or exceeded by a simple algorithm.
As Meehl pointed out with justified pride thirty years after the publication of his book, “There is no controversy in social science which shows such a large body of qualitatively diverse studies coming out so uniformly in the same direction as this one.”
The Princeton economist and wine lover OrleyAshenfelter has offered a compelling demonstration of the power of simple statistics to outdo world-renowned experts. Ashenfelter wanted to predict the future value of fine Bordeaux wines from information available in the year they are made. The question is important because fine wines take years to reach their peak quality, and the prices of mature wines from the same vineyard vary dramatically across different vintages; bottles filled only twelve months apart can differ in value by a factor of 10 or more. An ability to forecast
future prices is of substantial value, because investors buy wine, like art, in the anticipation that its value will appreciate.
It is generally agreed that the effect of vintage can be due only to variations in the weather during the grape-growing season. The best wines are produced when the summer is warm and dry, which makes the Bordeaux wine industry a likely beneficiary of global warming. The industry is also helped by wet springs, which increase quantity without much effect on quality. Ashenfelter converted that conventional knowledge into a statistical formula that predicts the price of a wine—for a particular property and at a particular age—by three features of the weather: the average temperature over the summer growing season, the amount of rain at harvest-time, and the total rainfall during the previous winter. His formula provides accurate price forecasts years and even decades into the future. Indeed, his formula forecasts future prices much more accurately than the current prices of young wines do. This new example of a “Meehl pattern”
challenges the abilities of the experts whose opinions help shape the early price. It also challenges economic theory, according to which prices should reflect all the available information, including the weather. Ashenfelter’s formula is extremely accurate—the correlation between his predictions and actual prices is above .90.
Why are experts e yinferior to algorithms? One reason, which Meehl suspected, is that experts try to be clever, think outside the box, and consider complex combinations of features in making their predictions. Complexity may work in the odd case, but more often than not it reduces validity. Simple combinations of features are better. Several studies have shown that human decision makers are inferior to a prediction formula even when they are given the score suggested by the formula! They feel that they can overrule the formula because they have additional information
about the case, but they are wrong more often than not. According to Meehl, there are few circumstances under which it is a good idea to substitute judgment for a formula. In a famous thought experiment, he described a formula that predicts whether a particular person will go to the movies tonight and noted that it is proper to disregard the formula if information is received that the individual broke a leg today. The name “broken-leg rule” has stuck. The point, of course, is that broken legs are very rare—as well as decisive.
Another reason for the inferiority of expert judgment is that humans are incorrigibly inconsistent in making summary judgments of complex information. When asked to evaluate the same information twice, they frequently give different answers. The extent of the inconsistency is often a matter of real concern. Experienced radiologists who evaluate chest Xrays as “normal” or “abnormal” contradict themselves 20% of the time when they see the same picture on separate occasions. A study of 101 independent auditors who were asked to evaluate the reliability of internal corporate audits revealed a similar degree of inconsistency. A review of 41 separate studies of the reliability of judgments made by auditors, pathologists, psychologists, organizational managers, and other professionals suggests that this level of inconsistency is typical, even when a case is reevaluated within a few minutes. Unreliable judgments cannot be valid predictors of anything.
The widespread inconsistency is probably due to the extreme context dependency of System 1. We know from studies of priming that unnoticed stimuli in our environment have a substantial influence on our thoughts and actions. These influences fluctuate from moment to moment. The brief pleasure of a cool breeze on a hot day may make you slightly more positive and optimistic about whatever you are evaluating at the time. The prospects of a convict being granted parole may change significantly during the time that elapses between successive food breaks in the parole judges’ schedule. Because you have little direct knowledge of what goes on in your mind, you will never know that you might have made a different judgment or reached a different decision under very slightly different circumstances. Formulas do not suffer from such problems. Given the same input, they always return the same answer. When predictability is
poor—which it is in most of the studies reviewed by Meehl and his followers—inconsistency is destructive of any predictive validity.
The research suggests a surprising conclusion: to maximize predictive accuracy, final decisions should be left to formulas, especially in lowvalidity environments. In admission decisions for medical schools, for example, the final determination is often made by the faculty members who interview the candidate. The evidence is fragmentary, but there are solid grounds for a conjecture: conducting an interview is likely to diminish the accuracy of a selection procedure, if the interviewers also make the final admission decisions. Because interviewers are overconfident in their
intuitions, they will assign too much weight to their personal impressions and too little weight to other sources of information, lowering validity. Similarly, the experts who evaluate the quas plity of immature wine to predict its future have a source of information that almost certainly makes things worse rather than better: they can taste the wine. In addition, of course, even if they have a good understanding of the effects of the weather on wine quality, they will not be able to maintain the consistency of a formula.
The most important development in the field since Meehl’s original work is Robyn Dawes’s famous article “The Robust Beauty of Improper Linear Models in Decision Making.” The dominant statistical practice in the social sciences is to assign weights to the different predictors by following an algorithm, called multiple regression, that is now built into conventional software. The logic of multiple regression is unassailable: it finds the optimal formula for putting together a weighted combination of the predictors. However, Dawes observed that the complex statistical algorithm adds little or no value. One can do just as well by selecting a set of scores that have some validity for predicting the outcome and adjusting the values to make them comparable (by using standard scores or ranks). A formula that combines these predictors with equal weights is likely to be just as accurate in predicting new cases as the multiple-regression formula
that was optimal in the original sample. More recent research went further: formulas that assign equal weights to all the predictors are often superior, because they are not affected by accidents of sampling.
The surprising success of equal-weighting schemes has an important practical implication: it is possible to develop useful algorithms without any prior statistical research. Simple equally weighted formulas based on existing statistics or on common sense are often very good predictors of significant outcomes. In a memorable example, Dawes showed that marital stability is well predicted by a formula:
frequency of lovemaking minus frequency of quarrels.
You don’t want your result to be a negative number.
The important conclusion from this research is that an algorithm that is constructed on the back of an envelope is often good enough to compete with an optimally weighted formula, and certainly good enough to outdo expert judgment. This logic can be applied in many domains, ranging from the selection of stocks by portfolio managers to the choices of medical treatments by doctors or patients.
A classic application of this approach is a simple algorithm that has saved the lives of hundreds of thousands of infants. Obstetricians had always known that an infant who is not breathing normally within a few minutes of birth is at high risk of brain damage or death. Until the anesthesiologist Virginia Apgar intervened in 1953, physicians and midwives used their clinical judgment to determine whether a baby was in distress. Different practitioners focused on different cues. Some watched for breathing problems while others monitored how soon the baby cried. Without a standardized procedure, danger signs were often missed, and many newborn infants died.
One day over breakfast, a medical resident asked how Dr. Apgar would make a systematic assessment of a newborn. “That’s easy,” she replied. “You would do it like this.” Apgar jotted down five variables (heart rate, respiration, reflex, muscle tone, and color) and three scores (0, 1, or 2, depending on the robustness of each sign). Realizing that she might have made a breakequthrough that any delivery room could implement, Apgar began rating infants by this rule one minute after they were born. A baby with a total score of 8 or above was likely to be pink, squirming, crying,
grimacing, with a pulse of 100 or more—in good shape. A baby with a score of 4 or below was probably bluish, flaccid, passive, with a slow or weak pulse—in need of immediate intervention. Applying Apgar’s score, the staff in delivery rooms finally had consistent standards for determining which babies were in trouble, and the formula is credited for an important contribution to reducing infant mortality. The Apgar test is still used every day in every delivery room. Atul Gawande’s recent A Checklist Manifesto provides many other examples of the virtues of checklists and simple rules.