Participant Bias: Types, Example and Reporting

What is participant bias?

Participant bias is a systematic distortion in a study’s results that comes from the people who take part: who agrees to enter the study, who stays in it, and how they behave once they are in it. This guide splits it into two branches. The first is who ends up in the sample (volunteer bias and non-response bias). The second is how participants behave (the Hawthorne effect, social desirability and demand characteristics). The split matters because the two branches need different remedies: a selection problem cannot be fixed by wording a question more carefully, and a behaviour problem cannot be fixed by a higher response rate.

Use this page as a map. Each type has its own CASRAI guide, linked below, and the worked example on this page shows how to put a number on the selection branch.

Types of participant bias

Branch 1: who enters the sample

Volunteer bias. The Catalogue of Bias defines it by the fact that volunteers “intrinsically have different characteristics from the general population of interest”. The entry lists socio-demographics and attitudes to the trial or institutions as examples, and notes that it is “difficult to estimate the impact of volunteer bias and the direction of its effect.” Its prevention advice is that any technique that increases volunteer numbers is likely to reduce the bias.

Non-response bias. It arises when non-responders differ in a meaningful way from responders. A low response rate alone does not prove bias; what matters is whether the people who did not answer differ on the measure you care about. See the CASRAI guide to non-response bias, and sampling bias for the wider problem of a sample that does not represent its population.

Branch 2: how participants behave

Hawthorne effect. The Catalogue of Bias describes it as individuals modifying an aspect of their behaviour in response to their awareness of being observed. It cites Eckmanns (2006), where hand-hygiene (antiseptic hand rub) compliance was 55% greater when staff were being watched than when they were not. It also cites the McCambridge (2014) systematic review of 19 studies, which found a wide range of variation and no overall summary of the effect. The practical reading is that observation can change behaviour in some circumstances, not that it always does, and by a fixed amount. See the Hawthorne effect.

Social desirability. Participants answer in the way they expect to be viewed favourably. The CASRAI guide to social desirability bias covers it in detail.

Demand characteristics. Participants pick up cues about what the study is “for” and adjust their responses. See demand characteristics. The broader family of answer-distorting patterns is in the guide to response bias.

Formula: the bias from non-response, with every symbol explained

For the selection branch there is a simple identity (plain algebra, not a claim from a particular source). If you could compare responders with the full population, the bias in an observed mean is:

B = ȳ_R − ȳ = W × (ȳ_R − ȳ_NR)

B: bias of the observed mean, in the units of the measure.
ȳ_R: mean among responders (what you observe).
ȳ: mean in the whole target population (what you wanted).
ȳ_NR: mean among non-responders (usually unknown).
W: share of the population that did not respond (1 minus the response rate, applied to the population).

The identity shows two things: bias is zero if responders and non-responders have the same mean, and a high non-response share does not by itself produce bias.

Worked example with real numbers (illustrative)

Illustrative example with invented numbers, not data from any real study. A university wants the average weekly hours of exercise among its 10,000 staff. Suppose, for the sake of the example, that 30% of staff are the kind of person who readily volunteers for health surveys and average 8 hours a week, while the other 70% average 5 hours. Volunteer-type staff respond with probability 0.50 and the rest with probability 0.20.

True population mean: 0.30 × 8 + 0.70 × 5 = 2.4 + 3.5 = 5.9 hours.
Responding share of the population: volunteer-type 0.30 × 0.50 = 0.15; others 0.70 × 0.20 = 0.14; total response rate = 0.29 (29%).
Observed mean among responders: (0.15 × 8 + 0.14 × 5) / 0.29 = (1.20 + 0.70) / 0.29 = 1.90 / 0.29 = 6.55 hours.
Bias: 6.55 − 5.9 = 0.65 hours, about 11% too high.
Check with the formula: non-responders are 0.71 of the population; their mean is (0.15 × 8 + 0.56 × 5) / 0.71 = 4.00 / 0.71 = 5.63. So B = 0.71 × (6.55 − 5.63) = 0.71 × 0.92 = 0.65. The two routes agree.

Notice the direction: the survey overstates exercise because the people most willing to take part are also the most active. In real studies you rarely know the non-responder mean, which is why the Catalogue of Bias stresses that the size and direction of volunteer bias are hard to estimate.

Interpretation: how to read the result

What you see	What it suggests
Responders and non-responders (or volunteers and the target population) differ on measured characteristics	Selection into the sample is a plausible source of bias; report the differences.
High refusal or non-response rate	Greater room for bias (the Catalogue of Bias notes the risk rises as refusal rate climbs), but not proof of it.
Behaviour changes when observation is announced or removed	A Hawthorne-type effect is possible; the evidence base is varied, so avoid assuming a standard size.
Answers differ between anonymous and identified conditions	Socially desirable responding is a candidate explanation.

Assumptions and common mistakes

Equating a high response rate with no bias. Bias depends on how different responders are from non-responders, not only on how many there are.
Assuming the direction. The Catalogue of Bias states the direction of volunteer bias is hard to estimate; in the example above it inflates the mean, but in another topic it could deflate it.
Treating the two branches as one. Weighting or follow-up can address who responded; it does not remove the effect of being observed.
Using the example numbers as benchmarks. They are invented to show the arithmetic.
Confusing it with sampling design error. See sampling bias for frame and selection problems that arise before anyone is invited.

How to report it

Report the invited number, the number who took part, the response or refusal rate, and any comparison between responders and non-responders you could make. An example wording (illustrative figures): “Of 3,000 staff invited, 870 completed the survey (response rate 29.0%). Responders did not differ from the sampling frame by age or department, but responders may over-represent people who are more active, so the mean weekly exercise estimate may be biased upward.” State the likely direction only as a possibility, in line with the Catalogue of Bias caution that direction and size are hard to estimate.

How to calculate it in Excel, R or Python

These snippets reproduce the worked example (the only case where the true mean is known by construction).

Excel (one cell):

=(0.15*8+0.14*5)/0.29-(0.3*8+0.7*5)

resp_share <- c(0.15, 0.14); means <- c(8, 5)
observed <- sum(resp_share * means) / sum(resp_share)
truth <- 0.3 * 8 + 0.7 * 5
observed - truth   # 0.6517

Python:

resp_share = [0.15, 0.14]; means = [8, 5]
observed = sum(s * m for s, m in zip(resp_share, means)) / sum(resp_share)
truth = 0.3 * 8 + 0.7 * 5
print(round(observed - truth, 4))  # 0.6517

FAQ

Is participant bias the same as response bias?

Not in the way this guide uses the terms. Response bias is about how participants answer; participant bias here also covers who ends up in the sample. The two overlap, and the response bias guide covers the answering side.

Can participant bias be eliminated?

Usually it can only be reduced and described. For volunteer bias, the Catalogue of Bias says any technique that increases volunteer numbers is likely to reduce the bias, and that assuring anonymity and confidentiality helps. For observation effects, the Catalogue notes that merely knowing one is in a study may still change behaviour.

How do I know which direction the bias goes?

You often cannot be certain. Compare what you know about responders and non-responders on measured variables and report the likely direction as a limitation of interpretation, not as a fact.

Sources

Catalogue of Bias Collaboration, Brassey J, Mahtani KR, Spencer EA, Heneghan C. Volunteer bias. Catalogue of Bias, 2017. catalogofbias.org/biases/volunteer-bias (checked 2026-10-29).
Spencer EA, Mahtani K. Hawthorne effect. Catalogue of Bias, 2017. catalogofbias.org/biases/hawthorne-effect (checked 2026-10-29).
Turk A, Heneghan C, Nunan D. Non-response bias. Catalogue of Bias, 2019. catalogofbias.org/biases/non-response-bias (checked 2026-10-29).