By Jason Hreha

Glossary

Loss aversion

Reviewed by Jason Hreha

Updated Published 9 min read

Losing a $20 bill seems to sting more than finding one pleases. If that asymmetry is real and large, it should show up wherever money can be framed as something taken away rather than something added: in taxes, fees, bonuses and prices. Whether it does, and by how much, is one of the more contested questions in behavioral economics.

What is loss aversion?

Loss aversion is the hypothesis that a loss relative to a reference point can receive more subjective weight than an equivalent gain. It is part of Kahneman and Tversky’s prospect theory (1979) and the reference-dependent model of Tversky and Kahneman (1991). The reference point matters: the same outcome can be described as a gain or a loss relative to different starting points.

Loss aversion is often listed among cognitive biases, although whether a particular case is a bias depends on the stakes and the reference point. Extra caution about a loss that could be ruinous is not obviously irrational.

Researchers summarize the size of the asymmetry with a loss aversion coefficient, usually written λ (lambda). It says how many times more a loss counts than a gain of the same size. A λ of 1 means no asymmetry. A λ of 2 means a $10 loss carries about as much weight as a $20 gain.

Tversky and Kahneman (1991) reported ratios “of just over 2:1” in several experiments and suggested that a coefficient of about two might explain both risky and riskless choices. Their 1992 estimate from risky choices, λ = 2.25, became the benchmark later studies compare against. Together these figures became the familiar claim that losses hurt “twice as much.” They were summaries of particular experiments, not a measured constant.

Loss aversion example

Imagine that losing $20 from your current balance matters more to you than gaining $20. That illustrates a loss-averse evaluation around the same reference point.

Now consider a coin flip: heads, you win some amount; tails, you lose $10. With λ of 2, the bet becomes worth taking only when the possible win is above $20. In Tversky and Kahneman’s report, a real-money bet to win $25 or lose $10 was “barely acceptable” to their participants. The example states the proposed asymmetry; it does not establish how strongly a particular person will react.

Ownership and buying versus selling

In Kahneman, Knetsch, and Thaler’s 1990 experiments, the classic demonstration of the endowment effect, some participants received coffee mugs and others became potential buyers. Participants stated whether they would trade at listed prices, with real transactions implemented for a selected market round. In the first experiment, median asking prices for the mugs ($5.25) were about twice median buying prices ($2.25 to $2.75), and repeated market rounds did not eliminate the gap. The endowment effect entry covers the later dispute over how the experiment was run.

A reference-dependent interpretation treats keeping an owned mug as the starting point: selling means giving something up, whereas buying means acquiring it. The buying-selling gap is an observed valuation difference. Calling it loss aversion is an explanation of that difference; the gap alone does not rule out other explanations involving the task or how people value ownership.

David Gal and Derek Rucker name some of those alternatives. Keeping what one already has requires no action, so a preference for the status quo can reflect inertia rather than a fear of losing. A buyer’s offer is also limited by budget in a way a seller’s asking price is not. These alternatives do not show that loss aversion is absent from the mug results. They show why the mug results alone cannot prove it.

Loss aversion in policy and business

Applied tests are clearest in settings where the same money can arrive as a gain or leave as a loss. Three kinds of example recur: taxes, small fees and pay.

Owing tax versus getting a refund

Alex Rees-Jones studied US tax returns to see whether the reference point of owing nothing changes how hard people work to reduce their tax bill. Taxpayers who would owe money at filing reduced their tax liability by about $34 more than taxpayers who were due a refund. The estimate comes from how returns cluster around a zero balance due, not from an experiment. It fits loss aversion around the zero point: owing feels like a loss, and a smaller refund does not feel like one to the same degree.

A bag tax versus a bag bonus

Tatiana Homonoff studied grocery shoppers before and after Montgomery County, Maryland, began charging 5 cents for each disposable bag in 2012. Compared with stores in areas without the change, the share of Montgomery County customers using at least one disposable bag fell by 42 percentage points, from a baseline of 82%. Some stores in the study already paid customers 5 cents for each reusable bag they brought. In those stores, 82% of customers used a disposable bag, against 84% in stores with no incentive, a difference that was not statistically significant.

The comparison is striking: the same nickel produced a large change as a charge and almost none as a reward. Homonoff interprets the contrast as consistent with loss aversion.

The two halves of the evidence are not equally strong, though. The tax estimate compares changes over time across areas; the bonus result compares different stores at one time. Shoppers also knew about the two policies to very different degrees: 98% of surveyed shoppers in the tax county knew about the tax, while only 52% of shoppers in bonus stores knew about the bonus. Homonoff concludes that awareness alone cannot account for the gap, but a new public charge may also carry a message that a store discount does not.

Pay framed as something to lose

Roland Fryer, Steven Levitt, John List and Sally Sadoff gave some teachers in nine Chicago Heights, Illinois, schools a $4,000 payment at the start of the year. They asked those teachers to repay any part not earned by their students’ test gains. Other teachers could earn the same bonus at the end of the year. In the first year, the loss-framed group’s students scored 0.234 standard deviations higher in math than controls; the gain-framed effect, 0.10 standard deviations, was not statistically significant. In the second year, neither form of incentive had an effect, and the pooled loss-framed estimate across both years was 0.124 standard deviations.

In a Nanjing, China, electronics factory, Tanjim Hossain and John List offered production teams a weekly bonus for meeting a target. For some workers the bonus was described as a reward; for others it was described as a provisional payment that would be reduced if the team missed the target. Framing the bonus as a potential loss raised total team productivity by about 1% beyond the effect of the incentive itself, and the effect did not fade over the study.

The pay studies show both the appeal and the risk of applying loss aversion. A change in framing, and for the teachers a change in when the money was paid, helped in one factory and one school year. It also failed to repeat in the second school year, and a clawback contract asks people to return money they have already received. For a business or agency, the lesson is to test a loss-framed design against a gain-framed one in its own setting, with follow-up, before adopting it.

How large is loss aversion? The evidence debate

Estimates across studies

A 2024 meta-analysis by Brown and colleagues combined 607 estimates from 150 articles published from 1992 to 2017 and reported an average loss-aversion coefficient of 1.955 (95% credible interval 1.82 to 2.10). Correcting for signs of publication bias lowered the estimate to about 1.74. Estimates were higher in field studies than in laboratory studies and higher in general-population samples than among students, but few study characteristics explained much of the variation. The analysis excluded studies of buying and selling prices, such as the mug experiments. This is evidence for a gain-loss asymmetry in the included studies, not a rule that every person experiences every loss as twice as painful as a gain is pleasurable.

Choice tasks and measurement conditions

A separate 2024 meta-analysis by Walasek, Mullett, and Stewart focused on fitting prospect theory to individual risky choices. It estimated a smaller coefficient of 1.31 (95% confidence interval 1.10 to 1.53) across 19 datasets and noted limited suitable data and imprecise estimates. In 12 of the 19 datasets, the confidence interval included 1, the value that means no loss aversion.

Yechiam and Zeif (2025) reanalyzed part of the broader literature and found that results depended on the symmetry and ordering of gains and losses. The subset with symmetric, unordered gains and losses did not show a reliable asymmetry, with a coefficient of about 1.07. An earlier review by Yechiam (2019) argued that the original studies showed very large losses being overweighted, while smaller losses often were not.

Questions about generality

Gal and Rucker’s critique (2018) challenged treating loss aversion as a general law, concluding that current evidence did not show losses to be, on balance, more impactful than gains. The journal published replies alongside it. Simonson and Kivetz rejected that conclusion as overstated while accepting that loss aversion is less robust and universal than often assumed. Higgins and Liberman agreed it is not as firmly established as typically assumed and emphasized the role of reference points.

The largest empirical reply came from Mrkva, Johnson, Gächter, and Herrmann (2020), who measured loss aversion in about 17,700 people, including four large surveys of US households. They found average loss aversion above 1 in every sample they examined, with a median coefficient of 2 in each of the four household surveys. Its size varied in consistent ways: older people were more loss averse, and people with more knowledge or experience in the relevant domain were less loss averse. Their smallest-stakes measure, a gamble to win €6 answered by 360 German-speaking adults, implied an average coefficient of only about 1.1, and only about half of those participants showed any loss aversion on it.

These studies differ in scope, data, and estimation. The evidence supports examining conditions and measurement rather than presenting the effect as either universal or wholly nonexistent.

Applied limits

Loss aversion is not the same as risk aversion, and a loss-framed message is not guaranteed to persuade. For a campaign, the practical question is whether the loss-framed version actually produces more of the desired action than the alternative. A coefficient from monetary choice tasks cannot simply be transferred to medical choices, relationships, or product marketing.

The examples above share one feature: each had a clear reference point (a zero balance due, a bonus already in hand, a bag that used to be free, an item already owned). Their size does not line up neatly with the stakes. In laboratory gambles, same-size gains and losses in no particular order show little asymmetry (Yechiam and Zeif), and Yechiam’s review finds smaller losses often were not overweighted in early studies. In the field, the 5-cent bag tax cut disposable-bag use by 42 percentage points, a large change for a tiny charge, possibly helped by features of a public charge beyond the money. A change in wording alone, as in the factory study, worked but was small.

Loss-framed messaging is a separate question with its own evidence. So is myopic loss aversion, which combines sensitivity to losses with evaluating outcomes over short intervals.

So does losing $20 hurt more than finding $20 pleases? The answer depends on how the question is measured. Averages across studies range from about 1.3 to 2 depending on the method. They fall to about 1 in studies that offer gains and losses of the same size in no particular order, and in some small-stakes gambles.

Defenders such as Mrkva and colleagues read this as a real effect with moderators. Critics read it as an effect confined to particular conditions: Gal and Rucker call for a contextual view of losses and gains, and Yechiam’s review points chiefly to large losses. The site’s essay The Death of Behavioral Economics argues the second view.

A team relying on the effect should define the reference point, compare a loss-framed design with a gain-framed one, and measure the behavior that matters. For the wider debate about which behavioral economics findings have held up, see the reading list of critical papers.

References