Conservatism bias
Updating beliefs too little when new evidence arrives.
What it means
Conservatism bias is the tendency to revise prior beliefs in the correct direction when new evidence arrives, but by less than Bayes' theorem warrants, so people cling too tightly to their starting view. The mechanism is sluggish belief updating: new data are underweighted relative to the prior, as though each piece of evidence carries less diagnostic force than it actually does. It is in a sense the mirror image of base-rate neglect — there the prior is ignored, here the prior dominates — and which failure occurs depends on how the evidence and the base rate are presented and how representative the data appear. A useful nuance is that the two biases can coexist across different tasks, so 'people under-update' and 'people over-update' are both true under the right conditions rather than contradictory. In practice, conservatism helps explain why financial markets under-react to earnings news and prices drift afterward, why forecasts lag turning points, and why organizations are slow to revise strategy in the face of new information.
The original experiments
The evidence came from a stripped-down task: participants watch draws from one of two bags whose exact compositions they know, then estimate the probability that a given bag is the source. Green, Halbert and Robinson (1965) and Phillips and Edwards (1966) found that revisions moved in the Bayesian direction but landed far short, extracting only about a third to a half of the certainty the evidence licensed. Edwards (1968) established 'conservatism' as a stable feature of intuitive inference. Because the signals were explicit, symmetric and numeric, the estimates could be scored against an exact Bayesian benchmark, a rarity in judgment research and part of why the finding anchored a decade of work on human information processing.
Why it happens
Early accounts blamed the response rather than the reasoning: people might infer correctly but compress their answers toward 50 percent, or mis-combine several independent draws (DuCharme, 1970). A more recent and better-supported account is perceptual. Augenblick, Lazarus and Thaler (2025) show that the size and even the direction of the error depend on how diagnostic each signal is: people over-infer from weak, barely informative signals and under-infer from strong ones. On this view conservatism is not a constant discount on evidence but the strong-signal tail of a broader distortion in how people read informativeness. They compress the perceived strength of signals toward the middle, much as the senses compress physical intensities into a narrower psychological range.
Where it shows up
In finance, conservatism is the leading behavioural explanation for underreaction: prices move on an earnings surprise but not far enough, so returns keep drifting the same way for weeks, the pattern known as post-earnings-announcement drift. Barberis, Shleifer and Vishny (1998) built a model in which conservative investors treat each announcement as noise until a run of them forces a jump, generating both underreaction and later overreaction. The same sluggishness surfaces wherever a standing estimate meets a stream of updates: forecasters who lag turning points in the economy, clinicians slow to abandon a working diagnosis, and managers who read early evidence against a strategy as temporary noise rather than the signal it is.
Related but distinct
Conservatism is easily confused with its apparent opposite, base-rate neglect, where people lean on the sample and forget the prior. Which one appears is not fixed. Grether (1980) showed that making evidence feel representative of a hypothesis pushes people toward neglecting the base rate, while abstract numeric tasks pull them toward conservatism. Benjamin's (2019) synthesis of hundreds of experiments finds under-inference the more common error overall, but reframes both as underweighting: of the signal in one case, of the prior in the other. Conservatism also differs from anchoring, which fixes on an arbitrary starting number, and from belief perseverance, where a belief survives even after the evidence that produced it has been discredited.
Examples
Investors under-react to an earnings surprise, so the price keeps drifting for weeks afterward.
A manager who rated a new hire outstanding at interview shifts only slightly after three missed deadlines, calling it settling-in noise rather than the early signal it plainly is.
Told a home test says positive, someone who felt fine that morning decides they are probably still fine, adjusting their sense of the odds far less than the test's accuracy justifies.
A growth team that pegged a landing page as the winner keeps favouring it after a fresh test shows fewer signups, dismissing the new data as a fluke rather than moving their estimate.
A juror who formed a strong early impression of guilt nudges it only slightly when the defence produces a solid alibi, updating far less than the new evidence warrants.
First described in Ward Edwards (1968).
Key references
- Augenblick, N., Lazarus, E., & Thaler, M. (2025). Overinference from weak signals and underinference from strong signals. The Quarterly Journal of Economics, 140(1), 335-401. doi.org/10.1093/qje/qjae032
- Benjamin, D. J. (2019). Errors in probabilistic reasoning and judgment biases. In Handbook of Behavioral Economics: Applications and Foundations 2 (pp. 69-186). North-Holland. doi.org/10.1016/bs.hesbe.2018.11.002
- Barberis, N., Shleifer, A., & Vishny, R. (1998). A model of investor sentiment. Journal of Financial Economics, 49(3), 307-343. doi.org/10.1016/S0304-405X(98)00027-0
- Grether, D. M. (1980). Bayes rule as a descriptive model: The representativeness heuristic. The Quarterly Journal of Economics, 95(3), 537-557. doi.org/10.2307/1885092
- Phillips, L. D., & Edwards, W. (1966). Conservatism in a simple probability inference task. Journal of Experimental Psychology, 72(3), 346-354. doi.org/10.1037/h0023653