10 cognitive biases ranked by how well they replicate

Anchoring and the fundamental attribution error survive giant multi-lab re-tests with huge effects; sunk cost survives small; money priming and ego depletion vanish. Here is the scorecard, with Indian money examples.

The National Stock Exchange of India building in Mumbai
Photo: Ganesh Dhamodkar, CC BY-SA 4.0, via Wikimedia Commons

Ranked by how strongly they held up in large, preregistered multi-lab replications, the most reliable cognitive biases are anchoring (Cohen's d 1.2 to 2.4 across 36 samples), the correspondence bias (d 1.8) and the false consensus effect (d 0.95 to 1.18). Framing, the gambler's fallacy and loss aversion replicate at medium size, sunk cost replicates but small (d 0.27), and money priming and ego depletion failed completely (d near 0).

Key takeaways
  • Anchoring is the most reliable bias tested: d = 1.17 to 2.42, in all 36 Many Labs samples.
  • Correspondence bias (d 1.81) was just as strong outside the West as in it.
  • Framing replicates at about half its textbook size; sunk cost replicates but small (d 0.27).
  • Money priming (d -0.02) and ego depletion (d 0.02 across 24 labs) failed.
  • Across Many Labs 2, the median effect fell from 0.60 to 0.15: even replicated biases are weaker than books say.

The criterion is the pooled effect size from the biggest replication available, in this order of preference:

  1. a Many Labs project (thousands of participants, dozens of labs, protocols fixed in advance);
  2. another preregistered multi-site study;
  3. the largest meta-analysis.

Cohen's d measures the gap between conditions in standard deviations: 0.2 is small, 0.5 medium, 0.8 large. A bias that only shows up in its original 1980s lab study does not make this list, however often it appears in self-help books.

Each entry carries an Indian money example to show what the bias looks like in daily life. These are illustrations of behaviour, not investment advice, and a study showing a bias in a lab does not prove that bias explains any particular market outcome.

6,344participants in Many Labs 1, 36 samples, 13 effects
15,305participants in Many Labs 2, 125 samples, 28 effects
0.60 → 0.15median d, original vs replication, Many Labs 2
91%individual F&O traders with net losses in FY25 (SEBI)

Which cognitive biases replicate best?

Anchoring, the correspondence bias and the false consensus effect replicate best: in Many Labs 1 and 2 each produced a large effect in essentially every one of dozens of labs. The chart shows the pooled replication effect for every bias on this list, plus two famous failures.

Replication effect size by bias (pooled Cohen's d)
See the data as a table
Replication d (d)
Correspondence bias (ML2)1.81
False consensus, supermarket (ML2)1.18
Anchoring, lowest of 4 items (ML1)1.17
Less-is-better (ML2)0.78
Gambler's fallacy (ML1)0.61
Gain vs loss framing (ML1)0.6
Framing, Asian disease (ML2)0.4
Sunk cost (ML1)0.27
Flag priming (ML1)0.03
Ego depletion (RRR 2016)0.02
Currency priming (ML1)-0.02

Source: Klein et al. 2014 (Many Labs 1), Klein et al. 2018 (Many Labs 2), Hagger et al. 2016 ego-depletion registered replication · Weighted pooled effects. Anchoring items ranged from d 1.17 to 2.42; the Many Labs team later posted a correction to the anchoring figures on its OSF page.

RankBiasBest replicationOriginal dReplication dLabs showing it
1AnchoringMany Labs 1, 4 questions0.931.17 to 2.42100% of 36
2Correspondence biasMany Labs 22.471.81100%
3False consensusMany Labs 2, 2 scenarios0.80 to 0.990.95 to 1.1893 to 100%
4Less-is-betterMany Labs 20.690.7889%
5Gambler's fallacy (retrospective)Many Labs 10.690.6183%
6FramingMany Labs 1 and 21.08 to 1.130.40 to 0.6045 to 86%
7Loss aversionRuggeri et al. 2020, 19 countriesn/a94% of items replicatedn/a
8Default effectMeta-analysis, 58 datasetsn/a0.68not multi-lab
9Sunk costMany Labs 10.230.2750%
10Money (currency) primingMany Labs 10.80-0.023% (once, ns pattern)

1. Anchoring

Anchoring is the pull of an arbitrary first number on a later estimate, and it is the most reliable bias in the Many Labs data: all 36 samples showed it. The pooled effects by question were:

  • d = 2.42: babies born per day in the US;
  • d = 2.23: height of Mount Everest;
  • d = 1.79: population of Chicago;
  • d = 1.17: distance to New York.

Participants who first saw a high number guessed far higher than those who saw a low one.

The story

The classic demonstration used a wheel of fortune. In their 1974 paper in Science, Amos Tversky and Daniel Kahneman described asking people to estimate quantities given as percentages, such as the share of African countries in the United Nations. Before each estimate, a number between 0 and 100 was produced by spinning the wheel in front of the participants.

They first had to say whether the true value was higher or lower than the wheel's number, then give their own estimate. The number was plainly random, and it still dragged the answers. The group whose starting point was 10 gave a median estimate of 25%; the group that got 65 gave a median of 45%. Paying people for accuracy did not reduce the effect.

Tversky and Kahneman: Judgment under Uncertainty: Heuristics and Biases, Science 185 (1974)
The wheel-of-fortune anchor (1974) spin: 0 to 100 "Is the share of African countries in the UN higher or lower than this?" then: "What is your estimate?" Wheel stops at 10 median estimate 25% Wheel stops at 65 median estimate 45%
A random number moved the median estimate by 20 percentage points. Rewards for accuracy did not remove the pull (Tversky and Kahneman, Science, 1974).

Where anchoring weakens

The fine print matters. Many Labs 2 tested a subtler version, where an anchor sits incidentally in the environment (a jersey number), and found nothing (d = 0.04).

And a 2025 replication in Economic Inquiry by Li, Weigel, Ferraro and Messer re-ran a study in which anchors raised what people would pay by 31%; with four times the statistical power, they measured 3.4%, with a confidence interval from -3.4% to 10%. So anchoring on estimates is near-certain; anchoring on real money decisions with incentives is smaller and less certain than the classic studies suggested.

Indian example (illustrative price tag): the struck-out MRP on a sale tag. "₹2,999, now ₹999" anchors the value at ₹2,999 whether or not anyone ever paid that.

Psychologist Daniel Kahneman speaking on stage
Daniel Kahneman, who with Amos Tversky described anchoring, framing and loss aversion in the 1970s. Three of their findings survived the big replication projects.Photo: nrkbeta, CC BY-SA 2.0, via Wikimedia Commons

2. Correspondence bias (the fundamental attribution error)

The correspondence bias is the habit of explaining someone's behaviour by their character even when you know the situation forced it, and Many Labs 2 found it at d = 1.81 in Western samples and 1.84 in less Western ones, with 100% of samples significant. The test was a classic: people read an essay they knew the writer had been told to write, then still rated the writer as believing it.

Interestingly, the original 2002 study by Miyamoto and Kitayama had found the effect much weaker in Japanese participants (d = 0.74). The replication found it just as strong outside the West, which undercuts a popular claim that East Asian (and by extension Indian) thinkers are immune to it.

Indian example: a colleague's loss on a stock tip gets explained as greed or stupidity, while your own loss on the same tip gets blamed on "the market".

  • Replication: Many Labs 2
  • d = 1.81 (WEIRD), 1.84 (less WEIRD)
  • Original: 2.47 and 0.74

3. False consensus effect

The false consensus effect is the belief that most people would choose what you chose, and Many Labs 2 replicated both of Ross, Greene and House's 1977 scenarios with large effects: d = 1.18 for the supermarket scenario and 0.95 for the traffic-ticket scenario. Every sample showed the supermarket effect, and 93% showed the other.

Indian example: the trader who believes "everyone is doing options now". SEBI's own count shows the opposite of a crowd of winners: of 1.13 crore individual F&O traders between FY22 and FY24, 93% lost money.

  • d = 1.18 and 0.95
  • 100% and 93% of samples
  • Original: 0.99 and 0.80

4. Less-is-better effect

The less-is-better effect, first shown by Christopher Hsee in 1998, is the finding that a smaller option can be valued more when judged on its own, such as a $45 scarf from a shop where scarves run cheaper being seen as a more generous gift than a $55 coat from a shop where coats cost far more. Many Labs 2 found it larger than the original (d = 0.78 against 0.69), significant in 89% of samples.

Indian example (illustrative prices): a ₹500 gift that is the best in its category (a premium box of mithai) can feel more generous than a ₹700 one that is the cheapest in its category, when the receiver sees only one.

  • d = 0.78; 89% of samples
  • Original: 0.69

5. Gambler's fallacy

A roulette wheel with numbered red and black pockets
A roulette wheel has no memory. Tversky and Kahneman's 1974 paper used a long run of red as the textbook case of the gambler's fallacy: people feel black is "due".Photo: Oniff, CC BY-SA 4.0, via Wikimedia Commons

The gambler's fallacy is the belief that a random sequence must "correct itself", and the retrospective version replicated in Many Labs 1 at d = 0.61, in 83% of samples, close to the original 0.69. Participants who imagined seeing a man roll three sixes at a casino guessed he had been rolling longer than those who saw two sixes and a three.

Indian example: "Nifty has fallen five days in a row, so it must bounce tomorrow." Each day's move does not know about the previous five.

  • d = 0.61; 83% of samples
  • Original: 0.69
Net losses of individual equity F&O traders in India
See the data as a table
Net loss incl. transaction costs (₹ crore)
FY2240824
FY2365747
FY2474812
FY25105603
FY2691685

Source: SEBI studies of individual traders in equity derivatives (2024, 2025, 2026), via Angel One and IANS · FY25 as in SEBI's July 2025 study; the 2026 study restated FY25 at about ₹1.12 lakh crore.

6. Framing effect

The framing effect is a change in choice when the same facts are described as gains or as losses, and it replicates clearly but at about half its original size: d = 0.60 in Many Labs 1 (original 1.13) and 0.40 in Many Labs 2 (original 1.08). In Many Labs 2 it was significant in only 45% of individual samples, though the pooled effect was unmistakable.

The test is Tversky and Kahneman's "Asian disease" problem: people prefer a sure option when outcomes are described as lives saved, and a gamble when the same outcomes are described as deaths.

Indian example (illustrative amount): "₹200 cashback" and "pay ₹200 less" are the same money, but they can land differently.

  • d = 0.60 (ML1), 0.40 (ML2)
  • Original: 1.13 and 1.08
  • Real but about half the textbook size

7. Loss aversion

Loss aversion, the finding that a loss hurts more than an equal gain pleases, holds up both in replication and in meta-analysis. Ruggeri and 31 co-authors re-ran Kahneman and Tversky's 1979 prospect-theory problems with 4,098 participants in 19 countries and 13 languages and replicated 94% of the items, and 12 of 13 theoretical contrasts, with some shrinkage.

A 2024 meta-analysis in the Journal of Economic Literature by Brown, Imai, Vieider and Camerer pooled 607 estimates from 150 papers and put the average loss aversion coefficient at 1.955 (95% interval 1.82 to 2.10): losses weigh about twice as much as gains.

Losses weigh about twice as much as gains: λ = 1.955 across 607 estimates.

It ranks below framing only because it was not tested in a Many Labs project with a single standardised effect size; the evidence base is arguably as strong.

Indian example: the investor who will not sell a stock below the purchase price "until it comes back", because booking the loss makes it real.

  • Replication: 94% of items, 19 countries
  • Meta-analysis: λ = 1.955 (607 estimates)
Investigating Variation in Replicability: A "Many Labs" Replication ProjectKlein RA et al. · Social Psychology · 2014

Thirty-six samples (25 in the US, 11 elsewhere) ran the same 13 effects on 6,344 participants, online and in person. Ten effects replicated, one (imagined contact) was weak, and two priming effects (flag and currency) failed with pooled d near zero. Anchoring was enormous (d 1.17 to 2.42), gain-loss framing d = 0.60, sunk cost d = 0.27. Lab versus online and US versus international setting barely changed results. Limit: the anchoring original used a different design, so the comparison is loose, and the team later corrected the anchoring numbers.

Read the paper (free)

8. Default effect

The default effect is the tendency to stick with whatever option is pre-selected, and the best evidence is a 2019 meta-analysis by Jachimowicz and colleagues in Behavioural Public Policy: across 58 datasets with 73,675 people, defaults shifted choices by d = 0.68. Defaults worked in most studies but not all; their power varied with the domain and with how much effort switching took.

It ranks eighth because a meta-analysis of published studies is weaker evidence than a preregistered multi-lab test: it can be inflated by publication bias, which the authors examined.

Indian example: a travel-insurance or donation box that comes pre-ticked at checkout. Many people pay for it simply because unticking takes a step.

  • Meta-analysis: 58 datasets, n = 73,675
  • d = 0.68

9. Sunk cost fallacy

The sunk cost fallacy, continuing because of what you have already spent, replicated in Many Labs 1 but small: d = 0.27, significant in half of the 36 samples (original 0.23). Participants imagined holding a ticket to their favourite team's big game on a freezing day; those told they had paid for it were more likely to say they would go than those told it was free.

Indian example: SEBI found that 76.3% of individual traders who lost money in F&O in both FY22 and FY23 kept trading in FY24, and only 8.3% of those persisting made a profit in the third year. SEBI did not study why, so this is a pattern consistent with sunk-cost thinking, not proof of it. If trading losses are affecting your sleep or relationships, treat that as a reason to talk to a professional, not to trade harder.

  • d = 0.27; 50% of samples
  • Small but real
Many Labs 2: Investigating Variation in Replicability Across Samples and SettingsKlein RA et al. · Advances in Methods and Practices in Psychological Science · 2018

Twenty-eight classic and recent findings were each run on about half of 125 samples (15,305 participants, 36 countries and territories) with peer-reviewed protocols. Only 15 (54%) replicated at p < .05, and the median effect fell from d = 0.60 in the originals to 0.15. The bias findings did well: correspondence bias d = 1.81, false consensus 0.95 to 1.18, less-is-better 0.78, framing 0.40. Incidental anchoring failed (d = 0.04). Limit: one protocol per effect, so a failure means that version did not replicate, not that the idea is dead.

Read the paper (free)

10. Money priming (and other failures)

Money priming, the claim that merely seeing images of currency makes people endorse free-market inequality, failed in Many Labs 1: pooled d = -0.02 against an original 0.80, with no reliable effect in any of the 36 samples. Flag priming (seeing a US flag shifts political views) failed the same way, d = 0.03.

Ego depletion, the idea that willpower is a fuel tank that self-control empties, belongs beside them. A preregistered replication across 24 labs led by Hagger and Chatzisarantis found d = 0.02, with a confidence interval from -0.09 to 0.13, despite earlier meta-analyses reporting d = 0.62. More on why so many famous findings shrank is in the replication crisis scorecard.

  • Currency priming d = -0.02 (original 0.80)
  • Ego depletion d = 0.02 (24 labs)
Replicating patterns of prospect theory for decision under riskRuggeri K et al. · Nature Human Behaviour · 2020

The team re-ran the original 1979 prospect-theory problems, including those behind loss aversion and the certainty effect, with 4,098 participants in 19 countries and 13 languages. 94% of items replicated, and 12 of 13 theoretical contrasts held, with some effects smaller than in 1979 and full replication in some countries. Limit: these are hypothetical choices, not real-money decisions, and India was not among the 19 countries.

Read the paper
Phiroze Jeejeebhoy Towers, the Bombay Stock Exchange building on Dalal Street, Mumbai
The Bombay Stock Exchange on Dalal Street. Index options traded on Indian exchanges are where SEBI found most individual traders losing money.Photo: Niyantha Shekhar, CC BY 2.0, via Wikimedia Commons

Why do most F&O traders in India lose money?

Most individual F&O traders in India lose money because options are a leveraged, short-dated product in which professionals with algorithms take the other side and transaction costs add up, and SEBI's numbers make the result plain: 91% of individual traders lost money in FY25. SEBI's 2024 study found 93% of 1.13 crore individual traders lost a combined ₹1.81 lakh crore from FY22 to FY24, while proprietary traders and foreign portfolio investors made gross profits, 96% to 97% of it through algorithmic trading.

The 2026 study adds who loses: 43% of individual traders were under 30, 88.5% of them lost money, and 73% of individual traders earned under ₹5 lakh a year. Biases do not explain all of this, but several on this list point the same way: the gambler's fallacy after a run of losses, loss aversion that keeps losing positions open, and sunk cost that keeps people in the game.

Why do some biases replicate and others fail?

Biases that replicate tend to be judgments people make about numbers, probabilities and other people's motives in a single sitting; the failures tend to be subtle primes expected to shift attitudes through a fleeting cue. The Many Labs 1 team found that variation across 36 samples was driven far more by which effect was tested than by where (intra-class correlation 0.75 across effects against 0.005 across samples). In other words, a robust bias shows up almost everywhere it is tested; a fragile one shows up nowhere.

Two more patterns are worth remembering:

  • Replications shrink. Replication effects are usually smaller than originals (Many Labs 2's median fell from 0.60 to 0.15), so even a "replicated" bias is weaker than a textbook says.
  • Hypothetical beats real. Effects measured on estimates and hypothetical choices travel better than effects measured on real behaviour with money at stake, as the 2025 anchoring replication shows.
Original vs replication effect (Cohen's d) Original Replication dashed line: d = 0 2.471.130.800.600.23 1.81 correspondence bias 0.60 framing (ML1) 0.27 sunk cost 0.15 ML2 median, 28 effects -0.02 currency priming
Most lines slope down: replications usually find smaller effects. Sunk cost is the rare one that held its (small) size. Data from Many Labs 1 and 2.

Old Indian logic had a word for this caution: a reason (hetu) is only as good as its invariable connection to the conclusion, which is the subject of much of the Manthan section of this site.

Questions people ask

Which cognitive biases failed to replicate?

In the large preregistered tests, money (currency) priming (d = -0.02), flag priming (d = 0.03), incidental anchoring (d = 0.04) and ego depletion (d = 0.02 across 24 labs) failed. Many Labs 2 replicated only 54% of its 28 findings overall.

Is loss aversion real?

Yes, on current evidence. A 2020 replication in 19 countries reproduced 94% of the original prospect-theory items, and a 2024 meta-analysis of 607 estimates put the loss aversion coefficient at about 1.96, meaning losses weigh roughly twice as much as equal gains. Researchers still debate how much it varies by person and by the size of the stakes.

Is overconfidence on this list?

No, because I could not find a preregistered multi-lab replication of overconfidence with a pooled effect size comparable to the others. That does not mean it is fragile; it means the evidence is of a different kind, so ranking it against Many Labs numbers would be unfair.

Sources

  1. Klein RA et al.: Investigating variation in replicability: a "Many Labs" replication project, Social Psychology 45(3) (2014)
  2. Klein RA et al.: Many Labs 2: investigating variation in replicability across samples and settings, AMPPS 1(4) (2018)
  3. Hagger MS et al.: A multilab preregistered replication of the ego-depletion effect, Perspectives on Psychological Science (2016)
  4. Ruggeri K et al.: Replicating patterns of prospect theory for decision under risk, Nature Human Behaviour 4 (2020)
  5. Brown AL, Imai T, Vieider FM, Camerer CF: Meta-analysis of empirical estimates of loss aversion, Journal of Economic Literature 62(2) (2024)
  6. Jachimowicz JM et al.: When and why defaults influence decisions: a meta-analysis of default effects, Behavioural Public Policy (2019)
  7. Li T, Weigel C, Ferraro P, Messer KD: Underpowered studies and exaggerated effects: a replication of anchoring effects, Economic Inquiry 63(2) (2025)
  8. SEBI: Updated study reveals 93% of individual traders incurred losses in equity F&O between FY22 and FY24 (2024)
  9. ThePrint: Individual traders incurred cumulative net loss of Rs 1.8 lakh crore in 3 years (2024)
  10. Angel One: Retail F&O losses rose to over ₹1 lakh crore in FY25 (2025)
  11. IANS/Morung Express: 88.5% of traders under 30 lost money in F&O in FY26 (2026)
  12. Tversky A, Kahneman D: Judgment under uncertainty, heuristics and biases, Science 185(4157):1124-1131 (the wheel-of-fortune anchoring demonstration and the roulette example) (1974)
cognitive biasesMany Labsanchoringloss aversionframing effectsunk costreplicationSEBI F&O study
Sapta

Sapta (Saptarshi Sarkar) is a mechanical engineer turned marketer in Bengaluru and a solo touring rider. He writes the seven Saptaverse blogs. Work with him or read more of Mindblog.

Keep reading in Mindblog

From the other blogs