Loss Aversion Is About 2 on Average, and Near 1 for Small Stakes

Textbooks often say a loss weighs 2.25 times as much as an equal gain. The largest synthesis of the evidence, a meta-analysis of 607 published estimates, puts the average loss aversion coefficient at 1.955, with a 95% interval of 1.820 to 2.102. Separate experiments with symmetric gambles find no loss aversion at all for losses up to about $20. Loss aversion appears reliably on average, yet its size shifts with the stakes and with how researchers measure it.

How a median from 25 graduate students became the default 2.25

Loss aversion is captured by one parameter, λ. In prospect theory's value function, gains and losses are measured from a reference point, and λ scales the loss side. A λ of 2 means a loss carries twice the weight of an equal gain, and a λ of 1 means the two are weighed equally.

The best-known estimate, λ = 2.25, comes from Tversky and Kahneman's 1992 experiments, in which 25 graduate students from elite American universities made unincentivized lottery choices. As the working-paper version of the meta-analysis points out, only a median was reported, with no mean and no measure of spread, and many behavioral finance simulations still use it as the typical value. Mixed gambles, with one possible gain and one possible loss, are required to identify λ, because it cancels out of prospects that contain only losses.

607 published estimates average 1.955, with a wide spread

Brown, Imai, Vieider and Camerer gathered estimates from 150 articles in economics, psychology, neuroscience and other fields, covering 1992 to 2017. Their preferred hierarchical model gives a mean coefficient of 1.955. Both 1, which would mean no loss aversion, and 2.25 fall outside the interval, and the authors say the original figure seems a bit too high.

The raw numbers are skewed. In the working-paper version, the median reported estimate is 1.69 and the simple mean is 1.97, and 93.9% of estimates exceed 1. Individual-level means averaged 2.95 in the raw data because a few very large values pull them upward. Study design explains little: the meta-regression accounts for 14.4% of the variation between observations, with field experiments, non-student samples and individual-level means associated with modestly higher estimates.

Other syntheses land lower. A meta-analysis of brand-choice studies reported 1.49 in its base model and 1.73 in an enhanced model, and an unpublished re-analysis by Walasek, Mullett and Stewart found a mean of 1.31 across 19 estimates. The gaps appear to come from different literatures, different tasks and different versions of the model.

The table turns each estimate into a concrete bet. With linear value and equal odds, a gamble that risks $100 breaks even in felt terms when the possible win equals λ times $100.

The last column is arithmetic on each reported λ, not additional data. The sources use different designs and definitions, so the column shows scale and does not rank the studies against one another.

Brokerage accounts realized 14.8% of gains and 9.8% of losses

Outside the laboratory, the closest match is the disposition effect, the tendency to sell winners and hold losers. Odean's analysis of 10,000 discount brokerage accounts from 1987 to 1993 compared how often investors sold a position with how often they could have. The chart shows the result for the full year, for January to November, and for December.

For the full year, the proportion of gains realized was 0.148 against 0.098 for losses, a ratio a little above 1.5. The monthly gain-to-loss ratio fell from 2.1 in January to 0.85 in December, which fits investors selling losers for tax reasons near year end.

The pattern had a cost. Winning stocks that investors sold beat the losing stocks they kept by 3.4 percentage points over the following 252 trading days, measured as excess return over the market index. A study of Taiwanese investors found they realized gains about 2.5 times as often as losses, against 1.5 times for the US investors in Odean's data.

Odean points out that prospect theory and a mistaken belief that prices revert to their averages both predict this behavior, and the data cannot separate them. The disposition effect is therefore evidence consistent with loss aversion, and it falls short of a direct measurement of λ. The sample comes from one brokerage and, by the author's own description, may lean toward successful investors because closed accounts were not replaced.

Below about $20, symmetric gambles show no loss aversion

Zeif and Yechiam ran five experiments with 2,001 participants in which gains and losses were equal in size and presented in random order. For average losses up to about $20, participants showed no loss aversion, even when real money was at stake. In an incentivized lottery with $6 gains and losses, the mean λ was 0.64, which points toward gain seeking. At an average loss of $40 the mean was 1.16, and only 51% of participants had a λ above 1. At $100 it was 1.54.

The same study reproduced stronger small-stakes loss aversion when it copied the design of earlier work, which used unequal gains and losses, losses that rose in order, and accept-or-reject wording. Randomizing item order alone cut the mean λ from 1.59 to 1.32. The authors conclude that λ varies with the amount at risk instead of staying fixed.

That finding sits inside a larger dispute. Gal and Rucker's 2018 review argued that the evidence does not show losses to be more impactful than gains on balance and called for a more contextual view. Mrkva and colleagues answered in 2020 that loss aversion has moderators yet persists even for small outcomes. A multinational replication of Kahneman and Tversky's 1979 prospect theory tests, with 4,098 participants in 19 countries and 13 languages, found that 12 of 13 theoretical contrasts held, with somewhat smaller effects.

What the 607 estimates leave out: stakes, samples and settings

The meta-analysis authors list where their data are thinnest: studies outside laboratory and field experiments, non-student populations, South America, Africa and Oceania, rewards that are not money, and elicitation methods other than sequential binary choice. Their dataset is also dominated by published studies. Unpublished papers ran about 0.3 points lower, roughly 1.7, a gap that is not statistically different at the 5% level.

The small-stakes evidence has its own limits. Zeif and Yechiam recruited experienced online participants, and they flag that a hypothetical or small online loss may not feel like losing money from one's own wallet. Their designs place the point where loss aversion becomes visible somewhere between $40 and $100, and no study in this evidence base pins the threshold down.

For anyone building a model or reading a claim, the usable range is wide. The averages above run from 1.31 to 2.25, and which value fits a given decision depends on the size of the stake and on how the coefficient was measured.