Simple Statistical Rules Often Beat Complex Expert Analysis

Robyn Dawes was a practical man, despite his training in philosophy and psychology. He liked to tell a story from his time working in a psychiatric ward in the late 1950s. A patient had developed a delusion—or so the psychiatrists believed. The man was convinced he was growing breasts. The doctors admitted him to a secure ward and spent weeks theorizing about the trauma that might have caused such a fixation, suspecting perhaps the recent death of a parent had triggered something deep in his psyche. Six weeks into his hospitalization, someone finally asked the patient to remove his shirt. He had a genetic condition. He was actually developing breasts. The delusion was real.

The episode stuck with Dawes as a lesson about expertise itself: even specialists, perhaps especially specialists, can become so enamored with elaborate explanations that they miss what is plainly visible. This observation would shape his fascination with the work of Ted Sarbin and Paul Meehl, two psychologists who had spent years documenting something counterintuitive—that simple statistical rules often predicted outcomes more accurately than expert judgment did.

Meehl had found the pattern everywhere. Sarbin used basic linear regression to forecast college grades based on high school rank and entrance exam scores. His formula beat the predictions of clinical psychologists who had access to the same information and much more besides. Meehl replicated the finding across domains: clinical diagnosis, academic performance, recidivism. The simpler the rule, the more often it won.

But how simple could these rules become? Dawes pushed the question further. Instead of optimizing weights mathematically to fit historical data as precisely as possible, why not choose them arbitrarily? Equal weight to each factor. Or random weights. Or weights picked out of thin air. He called this "improper" linear regression, and the results were startling: these crude formulas worked nearly as well as their mathematically perfected cousins.

Consider investment portfolios. Harry Markowitz won a Nobel Prize for showing how to allocate assets optimally—how to choose the precise weights that would maximize returns for any given level of risk. Yet when Markowitz himself had to decide how to invest his retirement contributions after publishing his theory, he split his money fifty-fifty between stocks and bonds. A crude, arbitrary allocation. A 2009 study by Victor DeMiguel and colleagues found that this simple equal-weight strategy was surprisingly effective—so much so that to prove the optimal approach was genuinely better, an analyst would need five centuries of data.

Dawes applied the logic to marriage. At a conference, a colleague challenged him: could he use one of his improper linear models to predict how well the colleague and his wife got along? Dawes thought he could. He had access to data on sex and relationships, and he proposed a predictor with two equally weighted variables: couples who reported being happy had sex more often than they argued, while unhappy couples argued more than they had sex. Two variables. Equal weight. Surely something more sophisticated existed? Yet the absurdly simple theory fit the evidence. Among twelve unhappy couples in one study, all argued more than they had sex. Of thirty happy couples, twenty-eight had sex more often than they fought. Subsequent small studies reached the same conclusion.

Why do these almost laughably simple models work? Part of the answer lies in the choice of variables themselves—there is expertise embedded in deciding what to measure, even if the weights are arbitrary. But another reason is that real-world outcomes often reflect straightforward combinations of factors. A criminal with multiple prior convictions is a bad sign, regardless of what else might be true. A couple having regular sex is probably a good sign, whatever the psychological complexities swirling around them. The third factor is noise. Marital happiness is hard to measure precisely. Risk is hard to measure precisely. Even the frequency of sex is trickier than it sounds—who is counting, and what counts? With all that noise in the data, an apparently perfect model can become overconfident as time passes and conditions shift. A cruder method, less ambitious in its claims, may prove more robust.

The point is not that simple analysis is always best, but that it is solid. It does not overreach. It can be sketched on a napkin in a bar or scrawled in a doctor's notebook. Before constructing an elaborate analytical edifice, sometimes it is worth asking someone to check under their shirt.