Framing effect

Two managers, one freezer

A bakery’s walk-in freezer fails with $6,000 of stock inside. At one branch the manager is told: “Plan A salvages $2,000 for sure. Plan B has a one-in-three chance of salvaging all $6,000 and a two-in-three chance of salvaging nothing.” She picks Plan A. At another branch, in the same situation, the manager is told: “Plan C loses $4,000 for sure. Plan D has a one-in-three chance of losing nothing and a two-in-three chance of losing all $6,000.” He picks Plan D.

Plans A and C are the same plan, as are B and D. Either choice can be reasonable on its own. What fails is that the choice follows the description: described as money saved, the sure thing looks attractive; described as money lost, the gamble does. This is the pattern of Tversky and Kahneman’s classic problem (see Evidence), moved from lives to money.

The same statistic, two ways

A comparison website shows a robot vacuum’s reliability data in one of two page layouts, chosen at random. One says “92% still working after two years”; the other says “8% break down within two years.” Shoppers who see the first layout rate the vacuum as more reliable and are more willing to buy it.

This is attribute framing: one characteristic, stated positively or negatively. The two statements are the same fact, and because the layout was random, the choice of wording tells the shopper nothing about the product or the site’s opinion of it. The difference in ratings comes from the wording alone.

Discount or surcharge

A café charges $3.80 for cash and $4.00 for card. One week its sign says “Pay cash and save 20 cents.” The next week it says “20-cent surcharge for card payments.” In the second week, noticeably more customers switch to cash.

No percentages, no gambles, and no one would call it a “framing problem”, but it’s the same move. Paying by card costs 20 cents more under both signs. The only change is whether those 20 cents are presented as a gain you pass up or a charge you pay, and a charge feels worse.

Irwin Levin and colleagues sorted framing effects into three kinds, which work differently and shouldn’t be lumped together:

• Risky-choice framing: options involving risk are described as gains or as losses. The first example above. Gain descriptions tend to favor the sure option; loss descriptions, the gamble.
• Attribute framing: a single characteristic is described positively or negatively. In a 1988 study, students rated ground beef labeled “75% lean” more favorably than beef labeled “25% fat”; the difference was smaller when they had tasted the meat.
• Goal framing: the same action is urged by stating what you gain by doing it or what you lose by not doing it (“insulating saves you $300 a year” versus “not insulating costs you $300 a year”). This kind concerns how persuasive a message is.

• When the descriptions aren’t really equivalent. Percentages of different groups, a missing part of the outcome, or wording that can be read more than one way (“200 will be saved” may be heard as “at least 200”) can make two descriptions differ in what they say. Reacting to a real difference isn’t a framing effect.
• When the choice of frame carries information. People tend to describe things in terms of what has changed relative to their reference point, and a speaker’s chosen wording can signal what they think. Researchers call this information leakage. Drawing a sensible inference from how someone chose to put it is legitimate.
• When you’re told both descriptions and still prefer one option. A stable preference that survives restating the problem is a preference, not a frame at work.

The test: restate it in the other frame. Did any fact change, or only the wording?

The coworker who chose the wording

Two coworkers report on a trial of a new scheduling app. One says, “70% of the team liked it.” The other says, “30% of the team didn’t like it.” You conclude that the first coworker is more in favor of adopting it.

The two sentences state the same result, but they weren’t assigned at random: each speaker chose the frame. That choice reveals something about how each of them sees the result, which is what you inferred. This is the choice of frame carrying information. It would be a framing effect if the same sentence, chosen by a coin flip, changed your view of the app.

The choice that survives the other frame

A bakery manager facing the freezer problem above hears “Plan A salvages $2,000 for sure” and leans toward it. Before deciding, she rewrites both plans as losses: “A loses $4,000 for sure; B has a two-in-three chance of losing all $6,000.” She still picks A, because the branch couldn’t cover a total loss this month.

She made the same choice as the first manager in the examples, from the same gain-framed description, so it’s easy to call it a framing effect. But she checked it against the other frame and it held, and her reason (a total loss would do damage out of proportion to its size) doesn’t depend on the wording. That’s a stable preference, tested by restating the problem, not a frame at work.

Tversky and Kahneman explained risky-choice framing with two ideas from their prospect theory. People evaluate outcomes as gains or losses relative to a reference point, and a description sets that reference point: “salvages $2,000” counts up from nothing, “loses $4,000” counts down from everything. And people tend to avoid risk when choosing between gains but accept it when choosing between losses. A description that moves the reference point therefore moves the choice.

Attribute framing is usually explained more simply: a positive label brings positive associations to mind, and those color the evaluation.

Underneath both is a failure to translate. Kahneman counts framing effects among the consequences of “what you see is all there is”: people evaluate the description in front of them and rarely convert it into the equivalent one they weren’t shown.

Status: replicates robustly. The risky-choice framing effect has been reproduced in a multi-lab, preregistered project and survives a meta-analytic correction for publication bias, and a second multi-lab project replicated a related problem from the same paper. Whether it shows irrationality is disputed; whether it occurs is not.

• Tversky and Kahneman (1981) gave students the “Asian disease” problem, in which 600 lives are at stake. When options were described as lives saved, 72% of 152 respondents chose the sure option. When the same options were described as lives lost, 78% of 155 respondents chose the gamble.
• Kühberger (1998) meta-analyzed 136 papers (230 effect sizes, nearly 30,000 participants) on risky-choice framing. He found a small-to-moderate overall effect and concluded that framing is reliable, while noting that design features change its size considerably.
• Many Labs 1 (Klein et al., 2014) reran the problem in 36 samples (6,271 participants). The effect replicated, at about half the original size (d = 0.60, against 1.13 originally), and was significant in 86% of individual samples.
• Steiger and Kühberger (2018) re-analyzed the 1998 data with a method that corrects for publication bias, estimated d = 0.52, and found no evidence of intense p-hacking. That’s close to Many Labs 1, and a new analysis of studies published in 2016 gave a similar figure (d = 0.56).
• Many Labs 2 (Klein et al., 2018) reran a different problem from the 1981 paper, in which people decide whether to travel 20 minutes to save the same amount on a cheap or an expensive item. With 7,228 participants the effect replicated, again at less than half the original size.
• The rationality dispute. Sher and McKenzie (2006) argued that logically equivalent frames can still leak information, for instance about the speaker’s reference point, and that only frames equivalent in the information they convey show that a choice was irrational. Mandel (2014) reported that framing effects vanished when the numbers were explicitly stated as exact, and that most people spontaneously read quantities like “200 saved” as “at least 200”. Chick, Reyna and Corbin (2016) spelled out the unstated outcomes with instructions, examples and quizzes, and found that people who read the options as complete descriptions still showed robust framing effects.

The dispute concerns the interpretation of standard framing problems, not whether the pattern is real, so it doesn’t meet this site’s criteria for “contested”. What’s less settled is how much of the classic effect reflects people filling in unstated outcomes. Note also that the multi-lab and bias-corrected evidence above concerns risky-choice framing. Attribute and goal framing are reviewed with many examples by Levin, Schneider and Gaeth (1998), but the status here rests on the risky-choice record.