Loss aversion

Evidence

Loss aversion is part of Kahneman and Tversky’s prospect theory (1979): the value function is steeper for losses than for gains. Three things about the paper belong in the assessment.

It gives no factor. All the 1979 paper says is that losses weigh more, derived from the observation that symmetrical bets of the form (x, 50%; −x, 50%) are found “distinctly unattractive”. There is no number.

The data are hypothetical choices. Students and faculty answered questionnaires with invented bets, usually 66 to 95 people per problem. The authors name this as a problem themselves: they are, they write, keenly aware of the questions about validity and generalisability, but hold the method to be the most workable for want of a better one.

Loss aversion is one property among several there. The larger part of the paper is taken up by the certainty, reflection and isolation effects, plus the weighting function for probabilities.

The factor of 2 comes from Tversky and Kahneman (1991), also checked. It rests on the mug experiments of Kahneman, Knetsch and Thaler (1990): sellers asked a median of $7.12 and $7.00, while choosers valued the same mug at $3.12 and $3.50. From this the authors derive a coefficient “slightly greater than two”. A bet with real money ($25 to win against $10 to lose) gave 2.5 to 1. The authors write expressly that the convergence of these estimates should be interpreted with caution.

Gal and Rucker (2018) contest the finding, and their objection is methodological rather than rhetorical: in the classic experiments “loss” coincides with “doing nothing” and “gain” with “acting”. Whoever keeps the mug is also staying passive.

The sharpest evidence comes from the same sample. In the classic endowment setup, 41 percent chose the mug when they owned it but only 18 percent when they did not. In a version where keeping it was an active choice too, the share was 26 percent — the same in both directions. Same stimuli, same population, one procedural change, and the effect disappeared.

The direct question also comes out against the theory: respondents rated the loss of a mug at −1.38 on a scale from −5 to +5, and gaining one at +2.71. For none of the objects examined was the loss rated as more impactful.

Mrkva, Johnson, Gächter and Herrmann (2020) reply, and their evidence base is far larger: five samples totalling 17,720 people, among them four random samples of US households. Two of their findings meet the objection head on.

The status-quo objection does not explain the finding. Gal and Rucker’s criticism applies to tasks where rejecting is also doing nothing. Mrkva and colleagues point to at least ten papers that offer a choice between two lotteries, neither of which is the status quo — and people are loss averse there too. In their own data, experienced drivers were in fact less loss averse, though their status quo should be more firmly rooted.

The effect is not confined to people short of money. Among the more than 3,000 respondents worth over a million dollars, only 22 percent accepted a bet of $100 to win against $50 to lose. Their median coefficient was 2.

Brown, Imai, Vieider and Camerer collected 607 estimates from 150 papers — across economics, psychology, neuroscience and other fields. Their preferred model gives a mean of 1.96 (95 percent interval 1.82 to 2.10); the median of the raw reported values is 1.69, and the range runs from 0.04 to 23.46. 93.9 percent of all estimates lie above 1.

The much-quoted 2.25, incidentally, comes from a third paper, Tversky and Kahneman (1992): 25 graduate students, unincentivised lottery choices, only the median reported. Brown and colleagues consider it somewhat too high.

So the existence of the effect is well supported. What remains contested is its size — and that varies so widely that any single number misleads.

Limits

Two different claims get equated. The research mostly measures how people value gambles or trades. The marketing rule “phrase it as a loss” is a claim about the effect of wording. That the second follows from the first is an assumption, not a finding. For framing effects in copy, the framing effect is the fitting page — and there the effect sizes for attribute framing are markedly lower.

The factor of 2 is not a constant. Brown and colleagues can explain only 14.4 percent of the variance between the 607 estimates from study characteristics — method, sample, reward type and continent taken together. The rest is unexplained. Carrying a textbook number onto a pricing page transfers a precision that does not exist.

At small stakes it nearly vanishes. In Mrkva’s German-speaking sample — 360 car buyers in Austria, Germany and Switzerland — the coefficient for bets of €6 averaged 1.12 (95 percent interval 1.07 to 1.17). For the same people, the trading tasks gave values between 1.8 and 2.7. Same people, different stakes, different effect.

Age and product knowledge shift it systematically. Older respondents were more loss averse (Cohen’s d between 0.11 and 0.20 across all five samples; 0.2 counts as a small effect), those who knew more about the specific product attribute distinctly less. The paper’s strongest single finding is that interaction: d = −0.53 for knowledge mattering only for the same attribute, not for others.

The criticism is older than 2018. Gal and Rucker’s own review traces the objections back to Gal (2006); 2018 is the year they were widely noticed in consumer psychology, not the year they arose.

Gal and Rucker do not claim the opposite. Their conclusion is not that gains weigh more, but that the relation depends on context: sometimes losses, sometimes gains, sometimes neither. Citing them as a refutation shortens them in the other direction.

The status-quo objection hits exactly the examples marketing likes. When customers stay on a plan, that may be loss aversion — or simply inertia and effort. The two look identical in the data but call for different measures.

Loss framing can backfire. Wording that creates pressure invites resistance; see reactance. For sales copy that link is not verified.

The figures from Brown and colleagues come from the working-paper version, CESifo Working Paper 8848 (January 2021); the citation is to the version published in the Journal of Economic Literature in 2024. Values may have shifted between the two versions.

The finding is broadly confirmed. Brown and colleagues pool 607 estimates from 150 papers; 93.9 percent lie above 1, with a mean of 1.96. Mrkva and colleagues confirm the finding across 17,720 respondents. The direction is thus as well supported as that of almost any pattern here. The magnitude is not: individual estimates run from 0.04 to 23.46, and only 14.4 percent of that spread is explained by study characteristics.