Written and maintained by CASRAI Editorial Board
Last updated
Definition: what ascertainment means
Ascertainment is the process of determining what is happening in a population or study group, for example finding cases. When that process works unevenly, the result is ascertainment bias: a distortion in data, as presented or analysed, in which some members of the target population are more or less likely to be included in the final results than others, so the study population ends up differing from the target population (Catalogue of Bias, Oxford Centre for Evidence-Based Medicine; University of Bristol Medical School glossary).
The useful question for any study is not “was every case found?” (almost never) but “was the chance of being found the same in every group being compared?” This page answers it with a worked example, a short reading guide, the study-design fixes the Catalogue of Bias lists, and code to run the arithmetic yourself.
Two senses of ascertainment bias
The term is used for two related problems, and it helps to know which one a paper means.
Who gets included
The first sense is about inclusion in the data. The Catalogue’s own example is a Cochrane review of bicycle helmets (Thompson DC et al., Cochrane Database of Systematic Reviews 1999, Issue 4, CD001855). If information on helmet wearing is gathered differently for cases and controls, ascertainment bias can arise. The Catalogue notes that minor facial injuries can enter the case group only because cyclists seek care for head injuries, so the authors recommended limiting facial-injury cases to serious injuries that would lead to an emergency department visit regardless of head injury.
Who gets diagnosed or recorded
The second sense is about detection of the outcome. The Catalogue describes a database analysis (Zissimopoulos JM et al., JAMA Neurology 2017;74(2):225-232) that linked statin use to lower Alzheimer’s incidence. A letter (Freedman DM, Pfeiffer RM, JAMA Neurology 2017;74(7):868) argued ascertainment bias, because an Alzheimer’s diagnosis requires contact with a doctor and statin use may be related to doctor visits. Whoever sees a clinician more often has more chances to be diagnosed, whatever their true risk.
Formula: there is no standard one
Ascertainment bias has no single formula in the sources cited on this page; it is a design problem, not a statistic. What you can do is make the idea concrete with one simple quantity. In this page’s illustrative model (arithmetic by the author, not a published standard):
- Detection proportion = cases detected ÷ true cases in the group.
- Observed risk = cases detected ÷ group size.
- Observed risk ratio = observed risk in group A ÷ observed risk in group B.
The model assumes every detected case is real (no false positives) and that the detection proportion is known. In real studies it is usually not known, which is why the design measures in the prevention section matter.
Worked example with real numbers
This is an illustrative scenario, not data from any study. Suppose two groups of 1,000 people each. In truth, 50 people in each group develop the outcome, so the true risk is 50 ÷ 1,000 = 5.0% in both and the true risk ratio is 1.0. Group A has frequent contact with clinicians and 80% of its true cases are diagnosed. Group B has little contact and 50% are diagnosed.
| Quantity | Group A | Group B |
|---|---|---|
| People | 1,000 | 1,000 |
| True cases | 50 | 50 |
| Detection proportion | 0.80 | 0.50 |
| Cases detected (50 × detection proportion) | 40 | 25 |
| Observed risk (detected ÷ 1,000) | 4.0% | 2.5% |
Observed risk ratio = 4.0% ÷ 2.5% = 1.6, although the true ratio is 1.0. With equal true risk, the distortion equals the ratio of the detection proportions: 0.80 ÷ 0.50 = 1.6. Doubling both groups to 2,000 people gives 80 and 50 detected cases out of 2,000, the same 4.0% and 2.5% and the same ratio of 1.6. A larger sample sharpens the wrong answer; it does not correct it.
Interpretation: reading the result
| What you see | What it suggests (in the model above) |
|---|---|
| Detection proportions equal in both groups | Uneven detection does not distort the comparison, because both groups lose the same share of cases |
| Group A detection proportion higher than group B | Group A’s observed risk is inflated relative to group B by the ratio of the two proportions |
| Group A detection proportion lower than group B | Group A’s observed risk is deflated by the same ratio |
| Detection proportions unknown | The size of the distortion cannot be computed, so ask who has more contact with the system that records the outcome |
Assumptions and common mistakes
- Assuming a bigger sample fixes it. As the example shows, the ratio stays at 1.6 at any size, because the distortion comes from who is recorded, not from random error.
- Looking only at the outcome. Exposure status can change how hard anyone looks. The statin letter is an example of this concern: the exposure may be related to contact with doctors, and contact is what produces a diagnosis.
- Letting people who know the exposure screen for the outcome. The Catalogue’s prevention advice for cohort studies is to keep exposure knowledge separate from outcome screening, identification and recording.
- Mixing it up with other biases. Related ideas such as lead-time and length-time bias, participant bias and survivorship bias have their own pages; the broader map is in types of bias in statistics.
Preventing it by study design
| Design | Prevention listed by the Catalogue of Bias |
|---|---|
| Case-control | Appropriate inclusion of cases and controls (see case-control studies and case-control vs cohort) |
| Cohort | Keep knowledge of exposure separate from outcome screening, identification and recording |
| Trial | Allocation concealment and blinding of all study staff (see allocation concealment vs blinding) |
How to report it
Describe how cases were found in each group, and say plainly if the process differed. The STROBE checklist guide explains the item on case ascertainment and control selection for case-control studies. A reporting sentence for the worked example above, in the usual style of italic statistic symbols, might read:
Cases were identified through clinic diagnosis records in both groups. Observed risk was 4.0% in group A and 2.5% in group B (RR = 1.6); because group A had more clinician contact, this ratio may overstate the true difference.
Where the mistake is in recording rather than in finding cases, read differential vs non-differential misclassification next.
How to calculate it in Excel, R or Python
All three snippets reproduce the table above (true cases 50 per group, detection proportions 0.80 and 0.50, 1,000 people per group).
Excel
Put group sizes in B2:C2 (1000, 1000), true cases in B3:C3 (50, 50) and detection proportions in B4:C4 (0.8, 0.5). Then:
B5: =B3*B4 (cases detected, copy to C5)
B6: =B5/B2 (observed risk, copy to C6)
B7: =B6/C6 (observed risk ratio, A vs B)R
n = c(A = 1000, B = 1000)
true_cases = c(A = 50, B = 50)
det = c(A = 0.80, B = 0.50)
detected = true_cases * det # 40, 25
risk = detected / n # 0.04, 0.025
unname(risk["A"] / risk["B"]) # 1.6Python
n = {"A": 1000, "B": 1000}
true = {"A": 50, "B": 50}
det = {"A": 0.80, "B": 0.50}
detected = {g: true[g] * det[g] for g in n} # 40.0, 25.0
risk = {g: detected[g] / n[g] for g in n} # 0.04, 0.025
print(risk["A"] / risk["B"]) # 1.6FAQ
What is ascertainment in simple terms?
It is the process of finding out who has the condition or outcome you are studying. Ascertainment is fair when every group has the same chance of being found.
What is an example of ascertainment bias?
Two examples from the Catalogue of Bias: a helmet review where minor facial injuries can enter the case group only because cyclists seek care for head injuries (the authors recommend limiting facial-injury cases to serious injuries that would lead to an emergency department visit regardless of head injury), and a statin and Alzheimer’s analysis where a diagnosis needs doctor contact that may itself be related to statin use.
Does a larger sample size fix ascertainment bias?
No. In the worked example the observed risk ratio is 1.6 with 1,000 or 2,000 people per group, because the problem is unequal detection, not a small sample.
How do you reduce ascertainment bias?
Choose cases and controls appropriately in case-control studies, separate exposure knowledge from outcome screening in cohorts, and use allocation concealment and blinding in trials (Catalogue of Bias).
Sources
- Catalogue of Bias Collaboration, Spencer EA, Brassey J. Ascertainment bias. Catalogue of Bias, 2017. catalogofbias.org/biases/ascertainment-bias (date checked: 2026-10-29).
- University of Bristol Medical School, research glossary. bristol.ac.uk glossary (date checked: 2026-10-29).
- Thompson DC et al. Cochrane Database of Systematic Reviews 1999, Issue 4, CD001855; Zissimopoulos JM et al. JAMA Neurology 2017;74(2):225-232; Freedman DM, Pfeiffer RM. JAMA Neurology 2017;74(7):868; all as listed in the references of the Catalogue of Bias entry (date checked: 2026-10-29).
