The framing effect has survived forty-five years of replication. It has also been generalized well beyond what the evidence supports.
A disease is expected to kill 600 people. Program A saves 200 for certain. Program B gives a one-in-three chance of saving all 600 and a two-in-three chance of saving nobody. Most people pick A.
Same disease, same 600, different wording. Under Program C, 400 people will die. Under Program D there is a one-in-three chance that nobody dies and a two-in-three chance that 600 die. Most people now pick D, the gamble.
“200 saved” and “400 die” are the same 600 people. When Amos Tversky and Daniel Kahneman put these problems to separate groups of students, 72 percent chose the sure thing in the first version and 78 percent chose the gamble in the second (Tversky & Kahneman, 1981). Nothing about the world changed. Only the description did.
That result has hardened into a parable: wording bends choice, so choose your words. The parable is not wrong, just far less useful than it sounds, because it skips the question that matters for anyone who communicates for a living. How much power should we attribute to framing, and what kind of power is it?
The effect is real, and it travels
Anton Kühberger’s 1998 meta-analysis pooled 136 papers, nearly 30,000 participants and 230 effect sizes, and concluded that framing in risky choice is a reliable phenomenon with an overall effect of small to moderate size (Kühberger, 1998). In 2022, Hohjin Im and Chuansheng Chen ran a variant of the disease problem past 102,830 respondents in 49 countries and found the effect in every one of them, with a pooled effect of h = 0.612 (Im & Chen, 2022).
It also appears in medically consequential scenarios, including among patients and physicians. In 1982, Barbara McNeil and colleagues asked 238 patients, 491 graduate students and 424 physicians to imagine they had lung cancer and to choose between surgery and radiation from the same data, expressed as either the probability of living or the probability of dying. In all three groups, surgery looked substantially more attractive when framed as survival (McNeil, Pauker, Sox & Tversky, 1982). Physicians were not immune, which does not make them irrational; it makes the format of information a clinical matter, as the authors argued.
Now the qualification, which is where the parable stops reading. Kühberger found not only that the effect exists but that its size swings hard with design: whether the manipulation shifts the reference point or merely makes one outcome more salient, whether people choose between options or rate them, and more. Im and Chen found the effect everywhere, but more collectivist countries tended to show smaller effects. Robust does not mean invariant.
Three different things are called “framing”
The parable also treats framing as one phenomenon. Irwin Levin, Sandra Schneider and Gary Gaeth argued in 1998 that the literature looked contradictory largely because different studies had manipulated different things under one label, probably tapping different processes (Levin, Schneider & Gaeth, 1998). They separated three kinds.
Risky-choice framing is the disease problem: a safe option and a gamble are described as gains or as losses, and what shifts is willingness to take the risk. Attribute framing is simpler: one feature of one thing is described positively or negatively, and what shifts is how favorably it is judged. Ground beef that is 75 percent lean is also 25 percent fat; no gamble, just an evaluation. Goal framing is the one communicators reach for most: a message stresses either the benefits of acting or the costs of not acting, and what shifts, if anything, is persuasion or behavior. The two versions emphasize different consequences of one action rather than redescribing one outcome.
These three measure different things: risk appetite, evaluation, compliance. They plausibly run on different machinery and differ enormously in how large and dependable their effects are. A finding about one does not license a claim about another.
Why would the same numbers produce different choices?
No single mechanism is proven. Several accounts are plausible and supported, and they may all be partly right.
The original account is prospect theory. Outcomes are evaluated as gains or losses relative to a reference point, and the frame sets that point. People tend to be risk averse for gains and risk seeking for losses, so the frame that codes the sure thing as a gain flatters it. Tversky and Kahneman themselves called this an approximate, simplified description and noted the pattern was not universal across individuals (Tversky & Kahneman, 1981). Prospect theory captures the shape of preferences well; it says less about what happens in a person’s head while the preference forms.
Fuzzy-trace theory answers at the level of representation. People encode both a precise, verbatim version of the options and a stripped-down gist, and choices often turn on the simplest gist available. “Some saved for sure” versus “maybe nobody saved” favors the sure thing; “some die for sure” versus “maybe nobody dies” favors the gamble. Same arithmetic, different bottom line. David Broniatowski and Valerie Reyna formalized this and reported that their model correctly predicted 82 of 88 gain-loss study pairs it reviewed (Broniatowski & Reyna, 2018).
Query theory answers at the level of attention. A frame changes the order in which you consult your reasons, and reasons retrieved first tend to crowd out those retrieved later. A 2025 meta-analysis of 27 papers found that frames do shift query order (d = 0.34), and that when researchers manipulated the order in which people considered reasons, the frame’s effect on choice fell from d = 0.92 to d = 0.39 (Composto, Duncan, Johnson & Weber, 2025). Fell, not vanished. The fourth account concerns the words themselves.
The complication: what the sentence leaves out
Look again at the original problem. “200 people will be saved” says nothing about the other 400. “400 people will die” says nothing about the 200 who live. The sure options are incomplete, in opposite directions, while the gambles spell out both outcomes. So the classic comparison pits a positively incomplete description against a negatively incomplete one. Earlier studies found that when each option states who lives and who dies, the effect tends to disappear (reviewed in DeKay, 2026), which fed a tempting conclusion: the famous effect is a wording artifact.
Michael DeKay and Shiyu Dou tested this properly in 2024, building 18 option descriptions, nine per frame, and comparing them in 81 combinations across 906 online participants and 521 students. The valence and gist of the descriptions explained a lot of the variation in risk preference; depending on what was stated or omitted, the effect could be amplified, eliminated or reversed. But in the balanced design, a considerable framing effect remained (DeKay & Dou, 2024). DeKay then replicated it in 2026 with 1,697 US adults matched to census data, 477 of whom chose between fully described options, and the effect persisted. When every option stated both the good and the bad outcome, the gamble was chosen in 25 percent of gain-framed choices and 35 percent of loss-framed choices, against 24 versus 48 percent in the standard version (DeKay, 2026). Smaller, but present and statistically unambiguous.
So two tidy stories fail. Prospect theory does not predict that the effect should swing with how complete the descriptions are, yet it does. The artifact story does not predict that it should survive complete, matched descriptions, yet it does. The phenomenon is real; its explanation has become more precise. That is what scientific refinement usually looks like.
Where words compete with evidence
The attribute case supplies a clean boundary condition. Levin and Gaeth asked people to evaluate ground beef labelled either “75% lean” or “25% fat.” The lean label won on ratings such as quality and greasiness. But when participants actually tasted the beef, the gap between the labels shrank. The authors describe this with an averaging model: a diagnostic experience is averaged in with the label and dilutes it (Levin & Gaeth, 1988).
That is the finding. The inference, a step beyond it, is that a description does the most work when it is the main input to someone’s representation of the thing. Once stronger, more direct evidence arrives, the label has to compete with it, and it loses ground. Notice what this does not say: that positive phrasing reliably improves anything downstream, or that experience erases the label. The effect was reduced, not eliminated.
Where communicators over-learn the lesson
If framing shifts risky choice across 49 countries, surely the right frame should move behavior too. Here the evidence stops cooperating.
Kristel Gallagher and John Updegraff pooled 189 effect sizes from 94 studies of gain- versus loss-framed health messages. Gain frames had a small edge for prevention behaviors; after a published correction, the figure is r = .075 (Gallagher & Updegraff, 2012, with the 2013 erratum). For attitudes and intentions, and for detection behaviors, they found no significant framing effect at all. Daniel O’Keefe and Jakob Jensen, reviewing 53 studies of disease-detection messages with 9,145 participants, found loss frames slightly more persuasive, r = −.039, an advantage that held for breast cancer detection and for no other category. Their conclusion: choosing loss over gain framing is unlikely to substantially improve persuasiveness (O’Keefe & Jensen, 2009).
These literatures use different framing manipulations, tasks and outcomes, so their effect sizes are not directly comparable. The safer conclusion is that robust risky-choice framing does not justify a general rule that gain- or loss-framed messages will substantially change behavior. Evaluation, choice, intention and behavior are four different outcomes, and “framing matters” is true of them to very different degrees.
What a responsible communicator should take from this
The common misuse of this literature is to read the disease problem, conclude that wording is a lever, and go looking for the frame that pulls hardest. The evidence supports something more demanding.
Identify what the audience is actually deciding and which outcome you are trying to move; the research on judgment, choice, intention and behavior is not interchangeable. Separate the facts from their representation and ask what the frame leaves unsaid. “200 saved” and “90 percent survival” are true, incomplete, and matched by equally true complements. Where stakes are high, state the complement. DeKay’s results suggest this will not reliably neutralize framing, but it removes an asymmetry in what people are told, which is the more defensible goal anyway.
Do not carry risky-choice results into persuasion problems; the strong effects live in equivalent descriptions of choices, and message-framing effects on behavior are small and inconsistent. If you hold direct, diagnostic evidence, give it; the beef study suggests it will compete with whatever label you attach. And test in the actual decision context, because Kühberger’s central finding is that effect sizes depend on task construction, which makes any imported framing rule a guess.
Underneath all of this is one idea. Words matter because they help build the representation from which a person decides. That power makes wording part of the information environment, with the same obligations as any other part of it: accuracy, completeness where it counts, and evidence proportionate to the claim.
Back to the 600
The disease problem endures because it looks like a trick: same facts, opposite majorities. Forty-five years of research have made it less a trick and more a diagnosis. The effect is real and it travels. It comes in several kinds that behave differently and runs on more than one mechanism. It is sensitive to what a sentence says and leaves out, yet survives even when the sentence says everything. And its practical effects become much less predictable once we move beyond risky-choice tasks into persuasion and behavior.
The lesson is not that audiences can be steered by wording. It is that presentation is part of the information people reason with, and anyone who controls the presentation is already inside the decision, whether they meant to be or not.
References
Broniatowski, D. A., & Reyna, V. F. (2018). A formal model of fuzzy-trace theory: Variations on framing effects and the Allais paradox. Decision, 5(4), 205–252. https://doi.org/10.1037/dec0000083
Composto, J. W., Duncan, S. M., Johnson, E. J., & Weber, E. U. (2025). A meta-analysis of query theory, a psychological process account of framing effects. Journal of Risk and Uncertainty, 71, 53–71. https://doi.org/10.1007/s11166-025-09458-6
DeKay, M. L. (2026). Risky-choice framing effects persist when option descriptions are matched and complete: A replication and extension of DeKay and Dou (2024). Psychonomic Bulletin & Review, 33, Article 139. https://doi.org/10.3758/s13423-025-02771-w
DeKay, M. L., & Dou, S. (2024). Risky-choice framing effects result partly from mismatched option descriptions in gains and losses. Psychological Science, 35(8), 918–932. https://doi.org/10.1177/09567976241249183
Gallagher, K. M., & Updegraff, J. A. (2012). Health message framing effects on attitudes, intentions, and behavior: A meta-analytic review. Annals of Behavioral Medicine, 43(1), 101–116. https://doi.org/10.1007/s12160-011-9308-7. Erratum (2013): Annals of Behavioral Medicine, 46, 127. https://doi.org/10.1007/s12160-012-9446-6
Im, H., & Chen, C. (2022). To save or lose? A cross-national examination of the disease risk framing effect and the influence of collectivism. Journal of Behavioral Decision Making, 35(4), e2276. https://doi.org/10.1002/bdm.2276
Kühberger, A. (1998). The influence of framing on risky decisions: A meta-analysis. Organizational Behavior and Human Decision Processes, 75(1), 23–55. https://doi.org/10.1006/obhd.1998.2781
Levin, I. P., & Gaeth, G. J. (1988). How consumers are affected by the framing of attribute information before and after consuming the product. Journal of Consumer Research, 15(3), 374–378. https://doi.org/10.1086/209174
Levin, I. P., Schneider, S. L., & Gaeth, G. J. (1998). All frames are not created equal: A typology and critical analysis of framing effects. Organizational Behavior and Human Decision Processes, 76(2), 149–188. https://doi.org/10.1006/obhd.1998.2804
McNeil, B. J., Pauker, S. G., Sox, H. C., Jr., & Tversky, A. (1982). On the elicitation of preferences for alternative therapies. New England Journal of Medicine, 306(21), 1259–1262. https://doi.org/10.1056/NEJM198205273062103
O’Keefe, D. J., & Jensen, J. D. (2009). The relative persuasiveness of gain-framed and loss-framed messages for encouraging disease detection behaviors: A meta-analytic review. Journal of Communication, 59(2), 296–316. https://doi.org/10.1111/j.1460-2466.2009.01417.x
Tversky, A., & Kahneman, D. (1981). The framing of decisions and the psychology of choice. Science, 211(4481), 453–458. https://doi.org/10.1126/science.7455683
Evidence note
The Gallagher and Updegraff figure above is the corrected value from the journal’s 2013 erratum, not the r = .083 in the original abstract. The DeKay (2026) percentages are shares of choices, not of participants; each participant made four choices.
Independent writing sample. Self-initiated and not commissioned by or affiliated with any client.