Published  |  Last updated

Quick Answer: Spectrum bias arises when the people in a diagnostic accuracy study do not cover the range of disease severity and of look-alike conditions that the test meets in practice, so the sensitivity, specificity and likelihood ratios measured in the study do not transfer to the clinic.1,2 Case-control and two-gate designs, which sample cases and controls separately, are the classic cause.3 In 184 studies of 218 tests, a diseased group with a separate control group gave a relative diagnostic odds ratio of 3.0 (95% CI 2.0 to 4.5) against clinical-population studies, though the size varies between tests.4,5 STARD 2015 asks authors to report how patients were sampled and QUADAS-2 asks reviewers to judge it.6,7 Worked example (one CC BY 4.0 study of 53 ICU sepsis cases and 104 outpatient controls, recomputed in R by Editverse): the same sensitivity and specificity give a positive predictive value of 83.9% at the study prevalence of 33.8% and 34.9% at 5%.8

What Spectrum Bias Is, and Why Sensitivity and Specificity Are Not Fixed

In 1978 Ransohoff and Feinstein reviewed diagnostic investigations to see why many tests proved valueless after an optimistic start. They identified two problems. If the diseased and nondiseased patients studied do not span an appropriately broad spectrum, the test can receive falsely high rule-in and rule-out values. If the test result and the true diagnosis are not established independently, bias can falsely raise the apparent efficacy. Their examples were the carcinoembryonic antigen and nitro-blue tetrazolium tests.2

Later authors prefer the term spectrum effect, because performance varies inherently among population subgroups, and they note that spectrum bias is still used in the literature.1 The practical point is the same under either name: sensitivity, specificity and likelihood ratios vary with the prevalence and the distribution of disease in the population or sample, and are independent of it only in the rare case of a truly dichotomous disease or a perfect test. Positive predictive value is also well known to vary with prevalence.1

Swipe sideways to see every column.

MeasureDefinitionWhat it varies with
SensitivityProbability that a person with the disease has a positive result.Prevalence and distribution of disease in the population or sample.
SpecificityProbability that a person without the disease has a negative result.Prevalence and distribution of disease in the population or sample.
Likelihood ratiosPositive: sensitivity divided by (1 minus specificity). Negative: (1 minus sensitivity) divided by specificity.Prevalence and distribution of disease in the population or sample.
Positive predictive valueProbability that a person with a positive result has the disease.Prevalence, as is well known.

Definitions and variation are as given in the BMJ review of the spectrum effect.1

Which Study Designs Create Spectrum Bias

Rutjes and colleagues use the term two-gate for studies with separate sampling schemes for people with and without the disease. Restricted sampling of cases, controls or both changes the accuracy estimates. They conclude that these spectrum effects may be advantageous in the early investigation of a new test, but that for an overall evaluation of clinical performance a case-control study should closely mimic a cross-sectional diagnostic study.3

Swipe sideways to see every column.

Sampling approachSpectrum riskWhere it shows
Case-control or two-gate: cases from one source, controls from anotherCases and controls can sit at the two ends of the spectrum, which can inflate accuracy.STARD 2015 item 9; QUADAS-2 question: was a case-control design avoided?6,7
Nonconsecutive or convenience series (STARD item 9 lists convenience series as a category)Nonconsecutive inclusion was associated with a relative diagnostic odds ratio of 1.5 (95% CI 1.0 to 2.1).STARD 2015 item 9; QUADAS-2 question: was a consecutive or random sample enrolled?5,6,7
Exclusion of unclear or hard-to-classify samplesCase-control studies often select clearly positive or clearly negative samples and exclude samples with unclear results. Smallwood and colleagues note that this tends to overestimate accuracy and puts a study at risk of spectrum bias.QUADAS-2 question: did the study avoid inappropriate exclusions? The statement about unclear samples is from the HIV test review.9,7
Consecutive or random series from the intended-use populationClosest to the spectrum the test will meet, though accuracy can still differ between settings.STARD 2015 items 3, 7 and 8: intended use, and how and where participants were identified.1,6

STARD 2015 and QUADAS-2: What Each Asks About Spectrum

STARD 2015 is a list of 30 essential items for reporting diagnostic accuracy studies.6 QUADAS-2 is the tool used in systematic reviews to rate risk of bias and applicability. It has four domains (patient selection, index test, reference standard, and flow and timing), and the first three are also rated for applicability.10

Swipe sideways to see every column.

ToolItems or questions that bear on spectrum
STARD 2015Item 3: intended use and clinical role of the index test. Item 7: on what basis potentially eligible participants were identified. Item 8: where and when. Item 9: whether participants formed a consecutive, random or convenience series. Item 17: analyses of variability in diagnostic accuracy, pre-specified or exploratory. Item 19: flow diagram. Item 20: baseline demographic and clinical characteristics. Items 21a and 21b: distribution of disease severity, and of alternative diagnoses in those without the target condition.
QUADAS-2, patient selectionWas a consecutive or random sample of patients enrolled? Was a case-control design avoided? Did the study avoid inappropriate exclusions? Applicability: is there concern that the included patients do not match the review question?

Item numbers and wording are from the STARD 2015 checklist and the QUADAS-2 form.6,7

How Large Is Spectrum Bias? What Meta-Epidemiological Studies Found

Swipe sideways to see every column.

SourceScopeFinding
Lijmer 1999, JAMA184 original studies evaluating 218 diagnostic tests.4A diseased population with a separate control group overestimated diagnostic performance against a clinical population: relative diagnostic odds ratio (RDOR) 3.0 (95% CI 2.0 to 4.5). Different reference tests for positive and negative results: RDOR 2.2 (1.5 to 3.3).
Rutjes 2006, CMAJ31 meta-analyses with 487 primary studies.5Nonconsecutive inclusion RDOR 1.5 (1.0 to 2.1); retrospective data collection 1.6 (1.1 to 2.2); severe cases with healthy controls 4.9 (0.6 to 37.3); patients selected because they were referred for the index test rather than for symptoms 0.5 (0.3 to 0.9).
Dautzenberg 2020, MoCAOld-age psychiatry cohort, n = 710.11Cognitive impairment, healthy-control case-control design: area under the curve (AUC) 0.93 and specificity 73% below a score of 26. Referred patients without impairment as comparators: AUC 0.77 and specificity 37%.

The direction is usually overestimation, but the magnitude is not predictable. Rutjes and colleagues found that the effect of design (cohort versus case-control) varied from large to moderate between meta-analyses, and that the interval for severe cases with healthy controls (0.6 to 37.3) includes 1.5 The MoCA study analysed the same setting two ways, which is why it is a useful illustration. Its licence is CC BY-NC, so we cite its numbers and do not redraw them.11

A Real Example: ICU Sepsis Cases Versus Outpatient Controls

Agha and colleagues published a retrospective single-centre case-control study at Tawam Hospital, United Arab Emirates, under CC BY 4.0. It compared 53 adult ICU patients with sepsis (Sepsis-3 criteria) with 104 outpatient controls with no obvious evidence of infection, using a model built from extended inflammation parameters on routine blood counts. The reported apparent area under the curve was 0.941 (95% CI 0.902 to 0.980), with sensitivity 88.5% (77.0 to 95.8%) and specificity 91.3% (84.2 to 96.0%). The authors state that comparing ICU sepsis patients with outpatient controls introduces severe spectrum bias characteristic of two-gate studies, and that the findings are preliminary.8 We re-plotted the numbers and recomputed what could be derived in R.

Line chart of positive predictive value against prevalence for a sepsis model with sensitivity 88.5% and specificity 91.3%: 34.9% at 5% prevalence, 64.2% at 15%, 83.9% at the study sample prevalence of 33.8%, with a pale band showing the authors assumed lower sensitivity and specificity range
Positive predictive value of the sepsis model at different prevalences, computed in R from the reported sensitivity and specificity. The pale band is the authors’ assumed range, not a measurement.

Swipe sideways to see every column.

QuantityReported in the paperRecomputed by Editverse in R
Sensitivity88.5% (77.0 to 95.8%)47 of 53 is 88.7% (exact 95% CI 77.0 to 95.7%). The counts are inferred; we found no cross-tabulation in the paper, and the Figure 1 legend gives 156 participants with complete data against 157 enrolled, so 46 of 53 (88.5%) is equally possible.
Specificity91.3% (84.2 to 96.0%)95 of 104 is 91.3% (exact 95% CI 84.2 to 96.0%).
Likelihood ratiosPositive 10.2; negative 0.1310.2 and 0.13, as reported.
PPV and NPV at the study prevalence of 33.8%87.0% and 92.2%83.9% and 94.0% from the reported sensitivity (88.5%), specificity (91.3%) and prevalence (33.8%). We could not reproduce the paper values.
PPV at 2, 5, 15, 25 and 50% prevalence16.6, 35.2, 63.4, 75.9 and 91.0%17.2, 34.9, 64.2, 77.2 and 91.0%. Differences are up to 1.3 percentage points; the cause is unknown.
PPV at 5% prevalence if sensitivity were 58 to 74% and specificity 66 to 81%9 to 15%8.2 to 17.0% across the four corner combinations.

What this shows. With sensitivity 88.5% and specificity 91.3% held fixed, a positive result is right about 1 time in 3 (34.9%) when 5% of the people tested have sepsis, against about 5 times in 6 (83.9%) in a sample that is 33.8% sepsis. A case-control sample fixes prevalence by design, so its predictive values cannot be carried into another setting.8

What it does not show. The fall in predictive value is arithmetic and would occur with no spectrum effect at all. The study did not test the model in an emergency department cohort, so it cannot measure how far sensitivity and specificity would drop. The authors’ adjusted ranges (sensitivity 58 to 74%, specificity 66 to 81%, area under the curve 0.70 to 0.79) are their assumptions drawn from the literature, not results from their data.8

Does the Design Label Alone Predict the Direction? A Meta-Analysis Subgroup

Smallwood and colleagues pooled 17 studies of a point-of-care HIV antigen and antibody combination test, rated each study with QUADAS-2 and pre-specified a subgroup analysis by design: 10 studies were case-control and 7 were cross-sectional. Pooled sensitivity was 84.9% (95% credible interval 65.0 to 94.5) for case-control studies and 93.2% (83.9 to 96.9) for cross-sectional studies. The authors report no substantive differences and describe the estimates as imprecise. They also note that many studies restricted samples to p24 antigen or RNA positive specimens, which may have led to underestimation of overall performance.9

Interval plot of pooled sensitivity and specificity of an HIV point-of-care test in 10 case-control and 7 cross-sectional studies: sensitivity 84.9% (65.0 to 94.5) vs 93.2% (83.9 to 96.9); specificity 99.6% (79.1 to 100.0) vs 98.8% (37.2 to 100.0)
Pooled medians and 95% credible intervals by study design, as reported in Table 1 of the meta-analysis. The sensitivity odds ratio was computed in R from the intervals and is approximate.

From the published intervals we back-calculated standard errors on the logit scale and obtained an approximate sensitivity odds ratio of 2.4 (95% interval 0.58 to 10.2; p = 0.22) for cross-sectional against case-control studies. The specificity intervals run to 100% and cannot be compared this way. This is an Editverse approximation, not the authors’ test.9

What this shows. A case-control label did not produce higher sensitivity here, and the intervals overlap. Spectrum bias is a mechanism, not a design label: what matters is who was actually sampled and how severe, and how varied, their conditions were. What it does not show. That case-control designs are harmless. The meta-analysis is small and imprecise, and the authors rated many studies as low in methodological quality in the index test and patient selection domains (64.7% of studies).9

How to Reduce Spectrum Bias: Design, Reporting and Appraisal

  1. 1State the intended use and clinical role. STARD 2015 item 3 asks for it, and it fixes who the target population is.
  2. 2Enrol from the people who would be tested in practice. Use a consecutive or random series and report how, where and when participants were identified (STARD items 7, 8 and 9). QUADAS-2 asks whether a consecutive or random sample was enrolled.
  3. 3Avoid inappropriate exclusions and avoid separate control sources. Healthy volunteers can be reasonable in an early study, but they do not describe clinical performance.
  4. 4Report flow, severity and alternative diagnoses. STARD items 19, 20, 21a and 21b let readers see how wide the spectrum was.
  5. 5Pre-specify variability analyses. STARD item 17 asks authors to distinguish pre-specified from exploratory analyses of variability in accuracy.
  6. 6Report predictive values at the prevalence you expect. Compute them from sensitivity, specificity and the target prevalence, as in the sepsis example. The Diagnostic Accuracy Calculator below does this.
  7. 7Validate in the intended-use population. Usher-Smith and colleagues advise developing and evaluating tests with data from the population in which they are intended to be used. The sepsis authors call for prospective validation in consecutive emergency department patients.

Steps 1, 2, 4 and 5 follow the STARD 2015 and QUADAS-2 wording above.6,7 Step 3 reflects the point about early studies made by Rutjes and colleagues.3 Step 7 follows the BMJ review and the sepsis authors.1,8

Need the Analysis Run or Checked? Consult Editverse

Editverse subject-matter experts work to COPE, ICMJE and GPP 2022 guidance, with institutional and country-specific ethics followed.

Editverse logo

Editverse

Study Design

Planning a diagnostic accuracy study and unsure who to enrol?

You send: your research question, the target condition and the setting where the test will be used.

You get: three PhD methodologist sessions on research design, sampling strategy and validity, with a written methodology framework document.

When: You are at the proposal stage and the sampling plan is not settled.

Methodology Consultation Package · $400

Plan my study design →

Editverse logo

Editverse

Statistical Analysis

Data collected and the accuracy analysis is the gap?

You send: your dataset, research questions and preferred software.

You get: annotated code or syntax, complete output files, an interpretation guide and publication-ready tables, with a standard turnaround of 4 to 6 business days.

When: The data are ready and the analysis is not.

SPSS/R/STATA Statistical Analysis · $400

Analyse my data →

Editverse logo

Editverse

Manuscript Writing

Ready to turn the accuracy study into a submitted paper?

You send: your data, notes and target journal.

You get: a manuscript written and edited by PhD experts, with statistical review, journal selection and submission support.

When: Your study is finished and the paper is not.

Manuscript Writing Services · $1,500

Write my paper →

Free Research Tools from Editverse

Frequently Asked Questions

What is spectrum bias in a diagnostic accuracy study?

Spectrum bias arises when the people studied do not cover the range of disease severity and of look-alike conditions that the test meets in practice, so the measured sensitivity, specificity and likelihood ratios do not transfer to the clinic. Ransohoff and Feinstein described the problem in 1978. Some authors prefer the term spectrum effect because the variation is a property of the test and the population as well as a flaw in one study.

Is spectrum bias the same as spectrum effect?

They describe the same phenomenon. Usher-Smith and colleagues prefer spectrum effect because test performance inherently varies among population subgroups, and they note that the older term spectrum bias is still used in the literature.

Why do case-control designs overestimate accuracy?

They sample cases and controls separately, so cases can be clear-cut and controls can be healthy volunteers, with few borderline or look-alike patients in between. Rutjes and colleagues describe how this restricted sampling produces spectrum effects, which can be useful in the early investigation of a new test but do not describe its clinical performance.

Does spectrum bias change sensitivity and specificity or only predictive values?

Both can change. Predictive values depend on prevalence, which is well known. Usher-Smith and colleagues add that sensitivity, specificity and likelihood ratios also vary with the prevalence and distribution of disease in a population or sample. In our worked example only the prevalence effect on the positive predictive value could be computed, because the study did not measure the test in an emergency department cohort.

How large is the overestimation?

It varies. Across 184 studies of 218 tests, a diseased group with a separate control group gave a relative diagnostic odds ratio of 3.0 (95% CI 2.0 to 4.5) against studies of a clinical population. Across 487 studies from 31 meta-analyses, nonconsecutive inclusion gave 1.5 (1.0 to 2.1), and the authors reported that the effect of design varied between meta-analyses. One HIV test meta-analysis found no clear difference between case-control and cross-sectional studies.

Which STARD 2015 items address spectrum?

Item 3 asks for the intended use and clinical role of the index test. Items 7 and 8 ask how and where eligible participants were identified, and item 9 asks whether they formed a consecutive, random or convenience series. Item 17 asks for analyses of variability in accuracy, items 20, 21a and 21b ask for baseline characteristics, severity of disease and alternative diagnoses, and item 19 asks for a flow diagram. STARD 2015 has 30 items.

How does QUADAS-2 judge spectrum bias?

In the patient selection domain, signalling questions ask whether a consecutive or random sample of patients was enrolled, whether a case-control design was avoided and whether inappropriate exclusions were avoided. A separate applicability question asks whether the included patients match the review question. QUADAS-2 has four domains, and the first three are also rated for applicability.

How can I reduce spectrum bias in my own study?

Define the intended use first, enrol consecutive or random patients from the setting where the test will be used, avoid inappropriate exclusions, report severity and alternative diagnoses, pre-specify subgroup analyses and report predictive values at the prevalence you expect. Use a case-control design only as an early step, and validate the test in the intended population before any clinical use.

References

  1. Usher-Smith JA, Sharp SJ, Griffin SJ. The spectrum effect in tests for risk prediction, screening, and diagnosis. BMJ. 2016;353:i3139. Open access. pmc.ncbi.nlm.nih.gov/articles/PMC4916916/
  2. Ransohoff DF, Feinstein AR. Problems of spectrum and bias in evaluating the efficacy of diagnostic tests. N Engl J Med. 1978;299:926–30. pubmed.ncbi.nlm.nih.gov/692598/
  3. Rutjes AW, Reitsma JB, Vandenbroucke JP, Glas AS, Bossuyt PM. Case-control and two-gate designs in diagnostic accuracy studies. Clin Chem. 2005;51:1335–41. pubmed.ncbi.nlm.nih.gov/15961549/
  4. Lijmer JG, Mol BW, Heisterkamp S, et al. Empirical evidence of design-related bias in studies of diagnostic tests. JAMA. 1999;282:1061–6. pubmed.ncbi.nlm.nih.gov/10493205/
  5. Rutjes AW, Reitsma JB, Di Nisio M, Smidt N, van Rijn JC, Bossuyt PM. Evidence of bias and variation in diagnostic accuracy studies. CMAJ. 2006;174:469–76. pubmed.ncbi.nlm.nih.gov/16477057/
  6. Bossuyt PM, Reitsma JB, Bruns DE, et al. STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies. BMJ. 2015;351:h5527. Open access, CC BY 4.0. pmc.ncbi.nlm.nih.gov/articles/PMC4623764/
  7. University of Bristol. QUADAS-2 tool (quadas2.pdf), Domain 1: patient selection (signalling questions read on 4 November 2026). www.bristol.ac.uk/media-library/sites/quadas/migrated/documents/quadas2.pdf
  8. Agha A, Yasin J, AlShamsi F. Diagnostic potential of extended inflammation parameters for sepsis identification: a retrospective case-control study. Front Med (Lausanne). 2025;12:1673278. Open access, CC BY 4.0; counts, predictive values and ranges recomputed by Editverse in R. doi.org/10.3389/fmed.2025.1673278
  9. Smallwood M, Vijh R, Nauche B, Lebouché B, Joseph L, Pant Pai N. Evaluation of a rapid point of care test for detecting acute and established HIV infection, and examining the role of study quality on diagnostic accuracy: a Bayesian meta-analysis. PLoS One. 2016;11(2):e0149592. Open access, CC BY 4.0; the sensitivity odds ratio was computed by Editverse in R from the published intervals. doi.org/10.1371/journal.pone.0149592
  10. Whiting PF, Rutjes AW, Westwood ME, et al. QUADAS-2: a revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med. 2011;155:529–36. pubmed.ncbi.nlm.nih.gov/22007046/
  11. Dautzenberg G, Lijmer J, Beekman A. Diagnostic accuracy of the Montreal Cognitive Assessment (MoCA) for cognitive screening in old age psychiatry: determining cutoff scores in clinical practice. Avoiding spectrum bias caused by healthy controls. Int J Geriatr Psychiatry. 2020;35:261–9. Open access, CC BY-NC 4.0; numbers cited only. pmc.ncbi.nlm.nih.gov/articles/PMC7028034/

Editverse Research Team | Updated November 2026