How We Err: Biases and Noise

[david_myers]

In their new book, Noise: A Flaw in Human Judgment, Daniel Kahneman, Olivier Sibony, and Cass Sunstein introduce two kinds of error.

Bias (systematic error) “is the star of the show.” The biases that skew our judgment captivate our science and our news.

Noise (variable judgments) plays off-Broadway, with no marquee. Yet its contribution to inaccuracy and misjudgment is scandalously big. Noise pervades

• medical judgments. In an ideal world, all physicians would correctly diagnose medical conditions. In the noisy real world, physicians presented with the same symptoms vary markedly in their diagnoses of cancer, heart disease, pneumonia, and especially mental disorders. Physicians also more often prescribe quick-fix opioids if tired and time-pressed at the day’s end, rather than at its beginning.
• political asylum decisions. Asylum seekers face a “refugee roulette,” their fates  determined by the luck of the draw as a judge becomes 19 percent less likely to approve asylum if two prior cases have been approved, or as one judge admits as few as five percent of applicants while another admits 88 percent.
• hiring and promotion decisions. Interviewers vary widely in their assessments of candidates, as do supervisors in their assessments of employees. There is more unreliability in assessment—more noise—than evaluators realize.
• sentencing and parole decisions. Criminal sentences for the same crime vary across judges. They also vary if a judge is hungry (sentencing before vs. after lunch), if the judge’s football team won or lost the day before, and if it is the defendant’s birthday.

To experience the bias/noise distinction—or to teach it to students—Kahneman et al. suggest a simple demonstration. Pull out your phone and open up its clock or stopwatch. Use its “lap” function to check your accuracy in estimating consecutive 10-second time intervals. After you hit “start” and look away, let your finger hover over the “lap” button. When you sense that 10 seconds has elapsed, touch “lap” and repeat, until you have five trials.

The difference between your averaged elapsed intervals and the actual 10.0 seconds reveals your bias—toward over- or underestimating the time interval. The variation among your estimates, represented by their range, is your judgmental noise. I launch my basketball free-throw shots with little bias—they average center of the net—but with more noise than I’d wish.

So, how might we reduce unwanted noise? Kahneman et al. offer suggestions:

• Compile the wisdom of the crowd. In 1907, Francis Galton invited 787 villagers at a country fair to guess the weight of an ox. Galton had little regard for individual people’s judgments, which displayed considerable noise (variation). Yet their average guess—1200 pounds—nearly hit the bull’s eye (the ox weighed 1,198 pounds). Likewise, a crowd’s average answer typically bests most individual judgments when estimating the number of jelly beans in a jar, the temperature one week hence, the distance between two cities, or future stock values. A large foundation that I serve recognizes the wisdom of the crowd when judging grant proposals: When much is at stake in a decision, they aggregate many expert opinions.
• Harvest the wisdom of the crowd within. When researchers Edward Vul and Harold Pashler asked individuals what percent of the world’s commercial airports are in the U.S., they hazarded a very rough guess. Three weeks later, the researchers asked them to guess again. The average of their two guesses (on this and other questions) tended to be closer to the truth (for airports, 32 percent). We are, after all, different people at different times. For example, our momentary moods affect what we notice, how we interpret it, what we recall, and even how gullible we are. Thus, just as it pays to combine the wisdom of multiple people, so it pays to combine the wisdom from across our varying states of mind. Sleep on it, and decide again.
• Harness the powers of statistical prediction and machine learning. Whether predicting suicide risk, GPA, criminal recidivism, employee success, or mental disorders, statistical models outperform noisy professional intuition (as I also explained in Intuition: Its Powers and Perils). Likewise, artificial intelligence now enables machines to excel by reducing judgmental noise when recognizing faces, generating driving directions, spotting breast cancers, and detecting impeding cardiac collapse.
• Assess job candidates with structured interviews and create behavioral scales for assessing employee performance. Replace informal hiring interviews with work sampling and structured interviews that assess candidates on each work dimension. When evaluating employees, break a complex judgment into specific, behaviorally-described components. To reduce noise from some raters being lenient graders and others being tough, ask each rater to rank those being assessed.
• Eliminate sequencing noise. First impressions often matter, coloring ensuing interpretations. Likewise, the first person to speak when assessing an idea or a candidate, and the first person to rate an online product, often gain added influence. Control for such sequencing noise by having people independently record their judgments.

Eliminating noise is usually, but not always, a good. The three-strikes life imprisonment policy reduced sentencing noise. But it did so with frequent injustice (when the offenses were minor, the life circumstances tragic, or the rehabilitation prospects promising). Algorithms that predict criminal risk may minimize noise yet be racially biased, as when the underlying crime data reflect over-policing in certain neighborhoods, over-reporting of certain offenses, or greater conviction rates for less affluent people.

But more often, noise entails unfairness. When decisions become arbitrary—a matter of who does the judging, or their mood, or the time of day, or who speaks first—Kahneman, Sibony, and Sunstein recommend decision hygiene. Aggregate multiple judgments. Think statistically; discount anecdotes. Structure decisions into independent tasks. Constrain premature hunches. Welcome dissent. In sum, think smarter, judge more reliably, and decide more wisely.

(For David Myers’ other essays on psychological science and everyday life, visit TalkPsych.com; follow him on Twitter: @DavidGMyers.)