Loss aversion

The office coin flip

A coworker offers a friendly bet on a coin toss: heads, you win $60; tails, you pay $50. You have a steady income and some savings, and you enjoy small bets. “No thanks. Losing fifty would bother me more than winning sixty would please me.”

This is the pattern loss aversion was proposed to describe. On average the bet pays $5, and $50 is too small to change anything important in your finances. What decides the choice is that the tails outcome is framed as money going out of your pocket, which is weighted more heavily than the same amount coming in. Whether most people reliably behave this way outside such questions is part of the dispute (see Evidence).

Checking the fund every day

Someone saving for retirement in 30 years puts money in a stock index fund, then checks the balance each evening. After a month in which the fund was down on almost half the days, they move everything to a savings account: “I couldn’t stand watching it lose money.”

Over a single day a stock fund goes down often; over periods of decades, much less often. If each daily drop is felt as a loss and weighted more than each daily rise, then looking more often makes the same investment feel worse, without changing what it’s likely to be worth at retirement. The error is less obvious here, because each individual reaction is to a real drop. It lies in evaluating a 30-year decision one day at a time.

The price that went up

A piano teacher has charged $50 a lesson for years and raises the price to $55. Two long-time families quit, saying it’s too expensive. That same month, three new families sign up at $55 without hesitation.

Both groups face the same choice: lessons at $55. For the new families, $55 is simply the price. For the old ones, the familiar $50 is the reference point, so the extra $5 registers as a loss. If the families who quit would have gladly signed up at $55 as newcomers, the reference point, not the price, made the difference. This is the riskless form: no gamble is involved, only a trade-off measured from where you started.

• When the loss would do damage out of proportion to its size. If losing would mean missing rent, defaulting on a loan or closing a business, the loss really is worse than an equal gain is good. Weighting it more is an accurate reading of your situation.
• When losing carries extra costs of its own. Fees, penalties, borrowing costs or the hassle of replacing something can make a loss bigger than its sticker amount.
• When the offer itself is evidence. A bet or deal that seems favorable but comes from someone who profits if you accept may be less favorable than it looks.
• When the reference point reflects something real. A price you’ve paid for years or a salary you were promised can shape real plans and budgets, so falling below it can matter more than rising above it.

The test: would you make the same choice if the outcomes were described from a different starting point, and is the downside actually large relative to what you can absorb?

The contract that could sink the shop

A two-person design studio is offered a large fixed-price contract. If it goes well, the studio’s yearly profit doubles. If it runs over, which the owners judge about as likely, the losses would use up their cash and force them to close. They turn it down.

It looks like refusing an even bet because the loss looms larger. But here the loss really is larger: doubling a year’s profit is good, and closing the studio ends every future year. That’s damage out of proportion to its size, and weighting it more heavily is correct.

The stranger’s sure thing

At a street fair, a man offers passersby a card game: “Pick the red card and I pay you $20. Miss it and you pay me $10.” A visitor declines, although the odds he describes sound even.

The downside isn’t simply weighted more. The visitor has good reason to doubt the odds are as described, since the man running the game profits from it. That’s the offer itself as evidence, and it would justify declining even for someone who weighed losses and gains exactly equally.

Kahneman and Tversky’s prospect theory (1979) proposed that people evaluate outcomes as gains and losses relative to a reference point, and that the curve relating outcomes to value is steeper for losses than for gains. Tversky and Kahneman (1991) extended the idea to choices without risk: when trading one option for another, what you would give up counts as a loss and what you would get counts as a gain. On this account, loss aversion also helps explain the Endowment effect, the Status quo bias, and why describing the same options as losses or gains changes choices (see Framing effect).

Critics have proposed other mechanisms for the same observations:

• Inertia. David Gal (2006) argued that many demonstrations pit a loss against a gain and doing nothing against making a change, and that a preference for not changing is enough to explain them.
• The context of the experiment. Lukasz Walasek and Neil Stewart (2015) proposed that people judge each amount by comparing it with other amounts they’ve seen. In an influential design they discuss, gains ranged up to twice the size of losses, which on this account would by itself make losses seem about twice as large.
• Constructed preferences. Kellen Mrkva and colleagues (2020) found that people who knew more about a product feature showed less loss aversion for it, consistent with the idea that loss aversion is stronger when people have to work out on the spot what something is worth to them.

Loss aversion is often confused with negativity bias, the broader idea that bad events and information affect people more than good ones (see Negativity bias). Loss aversion is narrower: it concerns decisions, and compares losses and gains of equal size measured from a reference point.

These suggestions follow from the definition. They haven’t been tested as remedies in the studies cited here.

• Describe the outcomes as where you end up. With $5,000 in savings, “win $60 or lose $50” is the same bet as “finish with $5,060 or $4,950”, which has no loss in it. If the new wording changes your answer, the wording was doing the work.
• Ask how big the loss is next to what you can absorb. A loss that would really hurt deserves extra weight; a small one that you’d barely notice in a month probably doesn’t.
• Look at the decision over the time it actually covers. A long-term plan judged by daily swings will show many more losses than the plan itself is likely to produce.

The evidence on experience is correlational: in Mrkva and colleagues’ surveys, people with more knowledge, experience and education showed less loss aversion, though still some.

Status: contested. Measured the standard way, with gambles that mix a possible gain and a possible loss, loss aversion meets this site’s bar for “robust”: a large meta-analysis finds it, and the estimate stays well above “no loss aversion” after a correction for publication bias. But the entry’s claim is the general principle that losses loom larger than gains, and both its generality and how much of the standard measurement reflects the design of the experiments are seriously disputed in the literature. Where the evidence falls between two labels, the standard is to choose the less confident one.

The foundational work. Kahneman and Tversky (1979) proposed prospect theory as an alternative to expected utility theory, with a value function that is steeper for losses than for gains. Tversky and Kahneman (1991) applied the same asymmetry to riskless choice, with the central assumption that losses and disadvantages affect preferences more than gains and advantages do. The degree of loss aversion is usually summarized by a coefficient, lambda (λ): how many times more a loss weighs than an equal gain. Brown and colleagues (below) note that a widely used early estimate, λ = 2.25, came from 25 graduate students in a 1992 study by Tversky and Kahneman.

Support.

• Brown, Imai, Vieider and Camerer (2024) meta-analyzed 607 estimates of λ from 150 articles published between 1992 and 2017, across economics, psychology, neuroscience and other fields. Their preferred estimate of the mean is 1.955, with a 95% probability that the true value lies between 1.82 and 2.10. In the pre-publication version, a model correcting for publication bias gave roughly 1.74. They excluded studies that compare buying and selling prices, so these figures come mainly from choices involving risk.
• Mrkva, Johnson, Gächter and Herrmann (2020) studied five samples totaling 17,720 people, including car buyers. More knowledge, experience and education went with lower loss aversion, and older people showed more, but “people of all knowledge, experience, and education levels were loss averse.” The authors present their results as casting doubt on claims that loss aversion is a fallacy or fully explained by status quo bias.
• Gächter, Johnson and Herrmann (2022) measured loss aversion in 660 customers of a car manufacturer with both endowment-effect tasks and a simple lottery task, and found the two measures strongly correlated.
• Ruggeri and colleagues (2020) reran the problems from Kahneman and Tversky’s 1979 paper with 4,098 people in 19 countries and 13 languages. Results replicated for 94% of the items and 12 of 13 theoretical contrasts, usually with smaller effects than in 1979. The contrasts tested effects such as certainty, reflection and framing, not a loss aversion coefficient, and the authors note that the replication does not resolve the criticisms of loss aversion.

Challenges.

• Gal (2006) asked 133 students to split a hypothetical $100 between a safe option and an even bet that doubled or lost the money invested. Although fewer than 2% would accept a single even bet offered on its own, nearly 80% put at least some money into the even bet when it was one of two ways to invest. He argued that the usual refusal reflects a preference for the status quo rather than loss aversion.
• Walasek and Stewart (2015) varied the range of gains and losses offered in hypothetical 50/50 gambles across four experiments, and reported loss aversion, loss neutrality or its reverse depending on the ranges. André and de Langhe (2021) argued that the design estimated λ on different gambles in different conditions, and that analyzing only the gambles common to all conditions eliminates the pattern. In a reply circulated as a preprint, Walasek, Mullett and Stewart agreed that λ didn’t differ on the common gambles, but pointed to large differences in how often those gambles were accepted and argued that λ estimated from accept-or-reject choices is too unreliable to settle the question. Brown and colleagues also describe a narrower re-analysis by Walasek and colleagues, of raw data behind 19 published estimates, that found an average λ of 1.31.
• Gal and Rucker (2018), in a review, concluded that “current evidence does not support that losses, on balance, tend to be any more impactful than gains.” In published comments, Simonson and Kivetz (2018) disagreed with some of their evidence and called the conclusion overstated, but agreed that loss aversion “is less robust and universal than has been assumed” and that its best-known support, the endowment effect and status quo bias, is open to other explanations. Higgins and Liberman (2018) agreed it is less firmly established than assumed, while arguing that reference points do make people more sensitive to changes in value, and that depending on the person and situation they can produce either loss aversion or the reverse.

What remains uncertain. That people often turn down mixed gambles unless the possible gain is considerably larger than the possible loss is well documented. Still in dispute: whether this reflects a general asymmetry in how losses and gains are valued, rather than inertia or features of the experiments; how far it extends to choices without risk and to everyday decisions; and how large it is outside the lab.