Cognitive-bias statistics do not describe one universal trait. They come from different tasks, scales, samples, and study designs. Across the studies summarized here, researchers measured framing, hindsight, overconfidence, anchoring, outcome bias, sunk-cost reasoning, belief bias, base-rate neglect, confirmation bias, and related statistical judgments. The results show why a person’s performance can vary substantially from one decision task to another.
On this page
- How susceptibility is measured
- Anchoring and numerical judgment
- Framing, outcomes, and sunk costs
- Overconfidence, hindsight, and belief
- Base rates and sample size
- Information search and confirmation bias
- What debiasing experiments found
How susceptibility is measured
The 2021 review The Measurement of Individual Differences in Cognitive Biases: A Review and Improvement used multiple studies of adults to examine how bias scores behave across tasks. Study 1 included 157 adults, Study 2 included 104 adults, and Study 3 included 102 adults. Scores cannot be compared directly when their scales differ, so the effect-size results are useful alongside the raw means.
In Study 1, the mean framing-bias score was 0.22 on a -5-to-5 scale, with Cohen’s d = 0.23. The mean hindsight-bias score was 3.38 on a -40-to-40 scale, with d = 0.93. The mean overconfidence-bias score was 7.22 on a -50-to-100 scale, with d = 0.68. Anchoring averaged 0.13 on a 0-to-1 scale, with d = 1.05. Outcome bias averaged 0.11 on a -5-to-5 scale, with d = 0.14.
The same Study 1 reported a mean sunk-cost-fallacy score of 3.73 on a 1-to-6 scale, with d = 2.76. Belief bias averaged 1.50 on a 0-to-4 scale, with d = 2.10. Base-rate neglect averaged 0.53 on a 0-to-1 scale, with d = 1.49. These effect sizes describe the relevant task results; they are not a ranking of people or a general measure of intelligence.
| Study 1 measure, 2021 | Mean score | Task scale | Cohen’s d |
|---|---|---|---|
| Framing bias | 0.22 | -5 to 5 | 0.23 |
| Hindsight bias | 3.38 | -40 to 40 | 0.93 |
| Overconfidence bias | 7.22 | -50 to 100 | 0.68 |
| Anchoring bias | 0.13 | 0 to 1 | 1.05 |
| Outcome bias | 0.11 | -5 to 5 | 0.14 |
| Sunk-cost fallacy | 3.73 | 1 to 6 | 2.76 |
| Belief bias | 1.50 | 0 to 4 | 2.10 |
| Base-rate neglect | 0.53 | 0 to 1 | 1.49 |
The observed ranges also show why a mean needs context. In Study 1, framing scores ranged from -1.75 to 2.50, hindsight scores from -4 to 15, overconfidence scores from -22 to 34, and belief-bias scores from 0 to 3. The anchoring score ranged from 0 to 0.57. The anchoring measure used 12 items and had split-half reliability of 0.68; the four-item base-rate-neglect measure had Cronbach’s alpha of 0.70. Source: The Measurement of Individual Differences in Cognitive Biases (2021, Study 1).
Results differed across the later studies. Study 2 had a mean framing-bias score of 0.94 on a -5-to-5 scale, d = 1.05; a mean overconfidence score of 6.10 on a -100-to-100 scale, d = 0.39; and a mean outcome-bias score of 0.77 on a -5-to-5 scale, d = 0.95. Its observed outcome-bias range was -1.50 to 4.17. Study 3 reported a mean hindsight score of 4.19 on a -60-to-60 scale, d = 0.73; a mean sunk-cost score of 3.61 on a 1-to-6 scale, d = 2.99; and a mean confirmation-bias score of 51.67 on a 0-to-100 scale, d = 3.64. All observed Study 3 sunk-cost scores fell between 1 and 6. Source: The Measurement of Individual Differences in Cognitive Biases (2021, Studies 2–3).
Anchoring and numerical judgment
Anchoring concerns the influence of an initial value or belief on a later estimate. In a 2015 study of Hungarian university students, 63.96% of pre-training responses were correct on the anchoring-bias task. That percentage is a task-specific result from the pre-training measurement, not a population estimate. Source: Is it time for studying real-life debiasing? (2015, pre-training).
The 2021 Study 1 anchoring scores ranged from 0 to 0.57 across observed participants. The measure contained 12 items and had split-half reliability of 0.68. In a separate 2007 retrospective analysis of 75 clinicians, prior belief significantly affected the post-search answer for anchoring, with P<0.001. A prospective student experiment in the same research also found a significant prior-belief effect, with P<0.001. Source: Do People Experience Cognitive Biases while Searching for Information? (2007, retrospective clinician analysis and prospective student experiment).
These figures measure different parts of numerical judgment: accuracy on a student task, score distribution in an adult measurement study, and the statistical significance of prior belief in information-search experiments. They should be read as complementary evidence rather than as interchangeable percentages.
Framing, outcomes, and sunk costs
Framing changes how equivalent choices are presented. In the Hungarian student sample, 50.00% of pre-training framing-effect responses were correct. The corresponding pre-training result for outcome bias was 44.48%, while 41.56% of sunk-cost-fallacy responses were correct. Source: Is it time for studying real-life debiasing? (2015, pre-training).
The 2021 measurements provide continuous score information alongside those percentages. Study 1’s observed framing-bias scores ran from -1.75 to 2.50. Study 2’s observed outcome-bias scores ran from -1.50 to 4.17. In Study 3, every observed sunk-cost-fallacy score was between 1 and 6. The Study 1 means were 0.22 for framing, 0.11 for outcome bias, and 3.73 for sunk-cost reasoning, but the task scales differ and the means should not be treated as a single common metric. Source: The Measurement of Individual Differences in Cognitive Biases (2021, Studies 1–3).
Overconfidence, hindsight, and belief
In the Hungarian university-student sample, 35.06% of pre-training overconfidence-bias responses were correct. This was the pre-training result for that particular task and sample. Source: Is it time for studying real-life debiasing? (2015, pre-training).
The 2021 Study 1 overconfidence scores ranged from -22 to 34, while hindsight scores ranged from -4 to 15. The belief-bias range was 0 to 3. Their Study 1 means were 7.22 on the -50-to-100 overconfidence scale, 3.38 on the -40-to-40 hindsight scale, and 1.50 on the 0-to-4 belief-bias scale. The associated Cohen’s d values were 0.68, 0.93, and 2.10, respectively. Source: The Measurement of Individual Differences in Cognitive Biases (2021, Study 1).
Study 2 reported a mean overconfidence score of 6.10 on a -100-to-100 scale, with d = 0.39. Study 3 reported a mean hindsight score of 4.19 on a -60-to-60 scale, with d = 0.73. Because the scale ranges changed between studies, the raw means do not establish that one study’s participants were more biased than another’s. Source: The Measurement of Individual Differences in Cognitive Biases (2021, Studies 2–3).
Base rates and sample size
Statistical-reasoning tasks produced some of the lowest pre-training correctness percentages in the Hungarian student study. Base-rate-neglect responses were correct 44.00% of the time. Covariation-detection responses were correct 27.27% of the time, regression-to-the-mean responses 20.78%, and insensitivity-to-sample-size responses 16.23%. These are measured percentages from the 2015 pre-training sample, not estimates for all university students. Source: Is it time for studying real-life debiasing? (2015, pre-training).
In the 2021 Study 1 measurement, the base-rate-neglect score ranged from 0 to 1, with a mean of 0.53. The measure used four items and had Cronbach’s alpha of 0.70. The study’s reported Cohen’s d for the base-rate-neglect task was 1.49. Source: The Measurement of Individual Differences in Cognitive Biases (2021, Study 1).
The difference between the 44.00% correctness result and the 0.53 mean score should not be converted into a new percentage without the scoring rules. One is a correctness percentage in a student task; the other is a score on a 0-to-1 measure in a separate 2021 study.
Information search and confirmation bias
The 2007 information-search research compared a retrospective clinician analysis with a prospective student experiment. In the retrospective clinician analysis, the search-session order effect was not significant (P=0.76), and the document-exposure effect was not significant (P=0.27). In the prospective student experiment, the search-session order effect was significant (P=0.026), and the document-exposure effect was significant (P=0.0081). Repeated exposure to the same document did not significantly change decision outcomes in either dataset: P=0.31 retrospectively and P=0.81 prospectively. Source: Do People Experience Cognitive Biases while Searching for Information? (2007, retrospective and prospective experiments).
The 2021 Study 3 confirmation-bias score had a mean of 51.67 on a 0-to-100 scale and Cohen’s d = 3.64. That result came from 102 adults in Study 3 and should not be merged with the national-risk experiment below. Source: The Measurement of Individual Differences in Cognitive Biases (2021, Study 3).
What debiasing experiments found
The article Debiasing training reduces confirmation bias in national risk analysts reports a 2022 experiment published in 2025. In that experiment, analyst confirmation-bias means were all above 0.351 at pretest and posttest; all t values were above 14.052 and all Cohen’s d values were above 1.895. Student confirmation-bias means were all above 0.453, with all t values above 21.245 and all Cohen’s d values above 1.973. These are reported experiment results, not forecasts.
The analyst-student comparison found a group main effect of F(1,169)=25.928, P<0.001, partial eta-squared=0.133. The debiasing-intervention main effect was F(1,169)=20.744, P<0.001, partial eta-squared=0.109. After the intervention, confirmation-bias scores were lower than at pretest in the general domain, F(1,168)=104.579, P<0.001, partial eta-squared=0.384, and in the risk domain, F(1,168)=45.999, P<0.001, partial eta-squared=0.215.
The same 2025 publication reported a bias-blind-spot domain effect of F(1,163)=4.547, P=0.034, partial eta-squared=0.027, and a bias-blind-spot intervention effect of F(1,163)=5.723, P=0.018, partial eta-squared=0.034. Together, these results indicate measured intervention and domain effects within that experiment. They do not show that one training format eliminates cognitive bias in every setting or population.