Confirmation Bias: Definition, Examples, and How to Avoid It in Research

Definition and etymology

Confirmation bias refers to the systematic tendency to favor information that is consistent with one's existing beliefs, hypotheses, or expectations — and to discount, ignore, or reinterpret information that contradicts them. The term was popularized by psychologist Peter Wason following his 1960 card-selection experiments, though the underlying concept was described much earlier by Francis Bacon in the Novum Organum (1620).

The word "confirmation" reflects the core mechanism: the mind seeks to confirm what it already suspects rather than to genuinely test it. This is distinct from motivated reasoning (which is driven by emotional investment) although the two often co-occur. Confirmation bias can operate entirely unconsciously even among experienced, well-intentioned researchers.

How confirmation bias occurs in research

In study design

Researchers with a strong prior belief may design studies that make confirmation of that belief more likely. This includes choosing outcome measures known to favor the expected result, selecting comparison groups that highlight the desired effect, or setting stopping rules that terminate data collection once a positive result appears.

In data collection

Interviewers who know a participant's group membership may probe more deeply when responses align with expectations. Observers conducting behavioral coding may rate ambiguous actions differently depending on whether a participant was assigned to treatment or control. Even automated systems can be affected if researchers selectively exclude data points labeled as "outliers" when they contradict the hypothesis.

In data analysis

The most dangerous form of confirmation bias in analysis is p-hacking: running multiple tests or re-specifying models until a statistically significant result appears, then reporting only that version. Researchers may also interpret the same effect size as "substantial" when it supports their hypothesis and "negligible" when it does not.

In reporting and publication

Positive results are more likely to be written up, submitted, and accepted for publication than null or negative results — a downstream manifestation of confirmation bias at the field level. Researchers may frame null findings as "preliminary" and shelve them, or introduce post-hoc explanations that rescue a favored theory from disconfirming data.

Real examples from research

Wason's 2-4-6 task (1960)

Peter Wason showed participants three numbers — 2, 4, 6 — and told them the sequence followed a rule. Participants had to discover the rule by proposing their own sequences and receiving yes/no feedback. Most participants quickly settled on "even numbers increasing by 2" and tested only sequences that confirmed that hypothesis (e.g., 8, 10, 12). The actual rule was simply "any ascending sequence." Participants failed to test disconfirming examples (e.g., 1, 2, 3) that would have revealed the broader rule — a textbook demonstration of confirmation bias in hypothesis testing.

Clinical interviewing and diagnostic bias

A study by Mendel et al. (2011) found that psychiatrists who were given a preliminary diagnosis before interviewing a patient were significantly more likely to confirm that diagnosis, even when the patient's actual presentation did not meet diagnostic criteria. The clinicians selectively elicited and weighted information consistent with the prior label.

The "file drawer problem" in psychology

Rosenthal (1979) documented that the published psychology literature was systematically skewed toward positive results because null findings remained unpublished in researchers' "file drawers." Meta-analyses built on this skewed literature overestimated effect sizes — a field-level consequence of confirmation bias in publication decisions.

How to detect confirmation bias

Detecting confirmation bias in one's own work requires deliberate effort because the bias operates largely below conscious awareness. Useful detection strategies include:

• Audit your literature search: Check whether your references predominantly support one position. A balanced review should engage seriously with disconfirming evidence.
• Pre-registration review: Compare your pre-registered analysis plan with what you actually ran. Deviations that consistently favor a positive result are a red flag.
• Adversarial collaboration: Ask a skeptical colleague to review your analysis plan and results before writing up. Their questions often surface assumptions you did not notice.
• Consider the alternative hypothesis seriously: Before finalizing your interpretation, write a paragraph explaining how a skeptic would interpret your data.

How to minimize confirmation bias

Pre-registration

Registering your hypotheses, sample size, and analysis plan on a platform such as OSF or AsPredicted before collecting data creates a public record that distinguishes confirmatory from exploratory analyses. It raises the cost of post-hoc rationalization.

Blinding

Where possible, keep data analysts blind to condition assignment until the analysis plan is finalized. Similarly, blind reviewers of outcomes (interviewers, raters, clinicians) to group membership to prevent expectancy effects from contaminating measurement.

Registered Reports

Some journals now accept Registered Reports — a format in which the introduction and methods are peer-reviewed and provisionally accepted before data collection. Acceptance is based on the quality of the design, not the outcome. This structurally decouples publication decisions from results.

Actively seek disconfirming evidence

Deliberately design tests that could falsify your hypothesis, not only confirm it — a practice Popper called falsificationism. Ask: "What result would change my mind?" and make sure your study can produce that result.

Quick summary

Related biases

• Publication bias: The field-level analog — journals preferentially publish positive results.
• Hindsight bias: After learning an outcome, believing it was predictable — can reinforce confirmation of one's theoretical framework.
• Availability heuristic: Overweighting easily recalled (often confirming) examples when judging frequency or likelihood.
• Observer bias: Expectations influencing what researchers record during data collection.