As a gain
90% survive
As a loss
10% die
Evidence
Tversky and Kahneman (1981) showed with the Asian disease problem that the choice between a sure and a risky option depends on the gain or loss frame: in the gain frame 72 percent of 152 respondents took the sure option, in the loss frame only 22 percent of 155. In the Many Labs project this effect replicated in 31 of the 36 samples, though at about half the strength of the original (d = 0.60 versus 1.13; Klein et al., 2014). The d measures how large an effect is: 0.2 counts as small, 0.5 as medium, 0.8 as large. Steiger and Kühberger (2018) re-analysed the research on the framing of risky decisions: 136 papers, over 30,000 participants, a mean of d = 0.31.
The re-appraisal is the more unusual finding, because the correction for publication bias went upwards: using p-curve the authors arrive at d = 0.52. Publication bias means that studies with a striking result are published more readily than studies with no finding; p-curve estimates from the distribution of results how much would remain if every study were available. The comparison is less direct than it sounds, though. The 0.31 averages over all 230 effect sizes including the null results, while the 0.52 rests on the 81 significant ones. Within those 81, p-curve does correct downwards — from d = 0.90 to 0.52.
Three independent estimates lie close together: 0.52 from the re-appraisal, 0.60 from the preregistered Many Labs project and 0.56 from a meta-analysis of papers published in 2016. The authors conclude that there is no replication crisis for the framing of risky decisions. Translated into people: at d = 0.52 the gain frame moves about 13 of 100 towards the sure option and the loss frame about 13 towards the risky one. The spread remains wide — depending on the experimental setup, between d = 0.16 (social topics) and d = 0.91 (repeated measures rather than separate groups). There is not the framing effect, the authors write.
The core of the effect was confirmed under pre-registration in 2025. Li and Feldman reproduced the gain–loss asymmetry with around 500 respondents: on gains 84 percent chose the sure option, on losses only 44 percent — the same reversal as in the original, at similar strength (Cramer’s V = .41 against .44 — a measure of association between two features that runs from 0 to 1). That evaluation hangs on the reference point rather than on final wealth has thus been independently tested.
A neighbouring finding from the same 1981 paper did not hold. The calculator-and-jacket example — the same five-dollar saving feels different depending on the total price — could not be repeated: 11 against 7 percent instead of 68 against 29 in the original. That concerns not the framing effect itself but topical account-keeping. Anyone treating the two as one transfers the good evidential standing of the first to the second.
The original paper is better secured than its reputation as a questionnaire study suggests. The ten problems were run on students at Stanford and Vancouver, but not all were hypothetical: on three of them one participant in ten actually played for money — the ball drawn in front of them, the winnings paid at once. A further version with real money and amounts divided by fifty produced the same pattern. Tversky and Kahneman state explicitly that the effects are “not restricted to hypothetical questions and are not eliminated by monetary incentives”.
Limits
The direction can reverse. What matters is how risk enters the task. In the re-appraisal the effect size is d = 0.56 where a concrete risky event lies behind it, as in the Asian disease problem. Where risk is used merely as a label (“the probability is 10 percent”), the mean turns in the opposite direction (d = −0.11). The p-curve on the significant opposite-direction results confirms it (d = −0.26). That is not noise but a distinct, reversed effect.
The authors name two qualifications themselves. The frequency of p-values rises just below 0.05, which points to some p-hacking. And p-curve also corrected the preregistered Many Labs data downwards, from 0.60 to 0.52, although there should be nothing there to correct; why, remains open.
Distinction: the classic framing effect concerns risky decisions. Discount framing (percent versus euro) is a different paradigm with its own, weaker evidence. The two should not be equated. Piñón and Gambara (2005) pooled 51 papers with 13,343 participants and averaged the classes separately: framing of risky choices comes to d = .44, attribute framing — the class discount framing belongs to — only to d = .26.
In attribute framing, more than one estimate in four points the wrong way. 27 percent of the 30 effect sizes are negative, the range runs from −.60 to 2.03, and across the four psychology papers the mean is −.25. What governs ad copy is therefore not only the weaker of the two effects but also the less consistent one.