Anchoring Bias: Definition, Examples, and Research Implications
Whatever number you saw first sticks. Tversky and Kahneman named it anchoring back in 1974 and it's held up across a few thousand replications since — in clinical diagnosis, survey design, peer review, and almost anywhere humans estimate things.
What is anchoring bias?
Anchoring bias — sometimes called the anchoring effect — is the pull an initial value exerts on any judgment you make next. The strange part: it works even when the anchor is obviously arbitrary. Random number, irrelevant context, doesn't matter. The estimate drifts toward it.
Tversky and Kahneman's wheel experiment is the canonical demo. Spin a wheel. Get a number. Now estimate what percentage of UN member states are African. Participants who saw a high number on the wheel guessed higher; those who saw a low number guessed lower. Everyone knew the wheel was random. It didn't matter.
Why does it happen? Two leading explanations. Insufficient adjustment says you start at the anchor and step away from it, but not far enough. Selective accessibility says the anchor primes anchor-consistent stuff in memory, so that's what comes to mind when you try to estimate. Probably both, depending on the task.
How anchoring bias occurs in research
It shows up at every stage of a project:
Hypothesis formation. You read the seminal paper. You see one preliminary result. Now every subsequent piece of data gets quietly calibrated against that first number rather than judged on its own. The literature you came in with set the rails.
Survey design. Scale endpoints and embedded numbers steer responses. "How many hours per week do you exercise — more or less than 10?" pulls answers up. The same question without the 10 produces a different distribution. The phrasing isn't neutral.
Peer review. Reviewers who see the authors' framing first — claimed effect size, sample size justification, choice of benchmark — anchor to it. If the author calls the effect "large," a borderline result reads as supportive rather than ambiguous.
Statistical estimation. Bayesian priors are anchors with a fancy name. So is using a previously published effect size to power your study; you've now told yourself what counts as a successful "replication" before collecting a single observation.
Concrete examples
In a classic series of experiments, Galinsky and Mussweiler (2001) found that the first offer made in a salary negotiation served as a powerful anchor: final agreed salaries were significantly higher when the opening offer was high and significantly lower when it was low, even when both parties knew the offer was unrealistic. This anchoring effect was reduced — but not eliminated — when negotiators were instructed to focus on their own goals rather than the counterpart's offer.
Medical research on diagnostic reasoning consistently demonstrates anchoring. Physicians who receive an initial (often incorrect) diagnosis from a referring clinician frequently anchor their subsequent reasoning to that label, seeking confirming evidence and discounting disconfirming symptoms. A 2017 review in Diagnosis found that anchoring was implicated in a substantial proportion of diagnostic errors in internal medicine and emergency care.
Strack and Mussweiler (1997) asked participants to estimate the age of Mahatma Gandhi at his death after first asking whether he died before or after either age 9 or age 140 — both obviously incorrect anchors. Estimates were still significantly influenced by the anchor direction, demonstrating that even implausible reference points shift numerical judgments.
How to detect anchoring bias
Spotting anchoring means tracing where the first numbers came from. A few questions worth asking:
- Were hypothesized effect sizes taken directly from a single prior study without independent justification?
- Do survey scales, vignettes, or experimental prompts embed numerical values that could orient participant responses?
- Did the research team see any preliminary data before finalizing the analysis plan?
- Are reviewers or raters given any reference values before making independent judgments?
Sensitivity analyses help. Re-run the analysis with different starting values, alternative priors, or other benchmarks and see how much the conclusion actually moves. If it doesn't, you're not anchored. If it does, you are.
Mitigation strategies
Pre-registration. Lock in your hypotheses, predicted effect sizes, and analysis plan before you collect data. The "anchor" is then your stated prediction, visible to everyone, instead of whatever pattern you happened to notice halfway through the spreadsheet.
Consider-the-opposite. Mussweiler, Strack, and Pfeiffer (2000) found that asking people to actively generate anchor-inconsistent arguments cut the effect substantially. Build it into your workflow as a devil's-advocate pass or a structured counter-hypothesis exercise.
Blinded analysis. Run the numbers without knowing which condition produced which data. Sometimes called masking. It blocks the obvious group difference from setting the frame before formal testing.
Independent estimates in replication. Replication teams should generate their own predicted effect sizes before peeking at the original. Otherwise the "replication" is calibrated to recover the original number, not to test it.
Questionnaire hygiene. Randomize scale endpoints. Don't seed instructions with example numbers. Pilot the instrument and look at the response distribution — anchor effects show up there.
Summary
Need academic sources on anchoring bias or cognitive biases in research methodology? Use CiteGenie to locate and cite peer-reviewed papers instantly.
Find Sources for Your Research