Conservatism (belief revision)
Also known as: Conservatism in probability estimation
Given new data, we update our probabilities, but not far enough.
What it means
A bias in sequential probability judgment in which people revise their estimates in the correct direction as evidence accumulates but by less than Bayes' theorem prescribes, so they remain anchored toward their prior probabilities. Demonstrated in classic bookbag-and-poker-chip experiments, the underweighting of new evidence is the opposite failure to base-rate neglect, where priors are instead ignored — and reconciling the two has been a long-running puzzle, with the resolution turning on whether the prior or the likelihood is the more salient, evaluable input. The bias slows learning and helps explain phenomena like post-earnings-announcement drift, where markets digest news too gradually. It is distinct from belief perseverance, which concerns clinging to discredited beliefs rather than the rate of normative updating. It matters wherever beliefs should track incoming data — diagnosis, forecasting, finance — because sluggish updating leaves judgments persistently behind the evidence.
Why it happens
Edwards separated two possible sources. In misperception, people misjudge how diagnostic a single observation is, treating a red chip drawn from the mostly-red bag as weaker evidence than it really is. In misaggregation, they read each datum roughly right but fail to combine a run of them the way Bayes requires: the correct rule multiplies likelihood ratios, while subjective judgment tends to average or add them instead. Averaging drags the running estimate back toward the middle, and the gap from the Bayesian answer widens as the sample grows. A related pull is anchoring, since the .5 starting point acts as a reference the estimate never fully leaves. In practice the sluggishness usually reflects faulty aggregation more than faulty reading of any single clue.
Reconciling it with base-rate neglect
The same urns produce opposite errors depending on how the question is posed. Ask people to update a stated prior as chips accumulate and they cling to the prior; show them one description to match against a stereotype and they discard the prior entirely. Grether (1980) found base-rate neglect inside the same urn-updating paradigm once the task invited a representativeness judgment. With a single point prior and one datum the two errors are mathematically indistinguishable: a weak revision could mean a strong prior or a discounted likelihood. Howe and colleagues (2022) elicited full distributions to pull them apart and found no fixed trade-off within a person. What mattered was whether the prior was made explicit; implicit priors were largely ignored, explicit ones merely underused.
Real bias, or rational skepticism?
Not all of the shortfall need be irrational. Corner, Harris, and Hahn (2010) argued that laboratory participants may not take the stated reliability of the evidence at face value: if you privately doubt that the draws are truly independent or that the experimenter's numbers are exact, revising less than the textbook amount is a reasonable response, and modelling that skepticism shrinks the apparent bias. Hilbert (2012) offers a different deflation, deriving conservatism from noisy memory retrieval rather than a motivated reluctance to move, since random error in converting observations into estimates regresses judgments toward the middle. Both accounts leave the behavioral signature intact while questioning the label. The practical upshot is unchanged: whatever its source, revision runs slow.
Where it shows up
In markets, prices drift for weeks after an earnings surprise instead of jumping at once, the post-earnings-announcement pattern the definition names, and under-reaction to dividend changes and stock splits has been read the same way. In the clinic, a physician who has settled on a first diagnosis moves too little when a later test cuts against it, so the initial impression outlives the evidence. Intelligence and forecasting shops, where Edwards' work seeded early probabilistic information processing systems, build in structured updating precisely because analysts left to themselves lag the incoming reporting. The common thread is a cost of delay: sluggish updating is cheap when the world is stable and expensive exactly when it is changing fast.
Examples
Shown draws from one of two urns of known composition, people infer which urn far more cautiously than the probabilities justify, needing many draws to reach the confidence Bayes warrants after few.
A GP sure that a patient's cough is viral nudges her estimate only slightly when the chest X-ray comes back abnormal, still leaning on the first impression the evidence should have overturned.
A pundit who tipped a team for the title concedes only a little after four straight defeats, still rating them a top-four side when the season's results already point to relegation.
A juror who forms an early view of guilt from the opening statement inches toward acquittal as alibi evidence mounts, still leaning guilty when the testimony has undercut the case.
A product manager convinced a feature will lift signups nudges the forecast down only slightly after each week of flat A/B results, still projecting a win the data no longer supports.
First described in Ward Edwards (1968); Phillips & Edwards (1966).
Key references
- Howe, P. D. L., Perfors, A., Walker, B., Kashima, Y., & Fay, N. (2022). Base rate neglect and conservatism in probabilistic reasoning: Insights from eliciting full distributions. Judgment and Decision Making, 17(5), 962-987. doi.org/10.1017/S1930297500009281
- Hilbert, M. (2012). Toward a synthesis of cognitive biases: How noisy information processing can bias human decision making. Psychological Bulletin, 138(2), 211-237. doi.org/10.1037/a0025940
- Corner, A., Harris, A. J. L., & Hahn, U. (2010). Conservatism in belief revision and participant skepticism. In Proceedings of the 32nd Annual Conference of the Cognitive Science Society (pp. 1625-1630). escholarship.org/uc/item/79b7w6h3
- 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