Published | Last updated
A 2025 Cochrane review pooled 32 studies and 12,529 participants and found that Xpert Ultra, the rapid molecular test now recommended for pulmonary tuberculosis, correctly flags about 91 out of every 100 people who truly have the disease and correctly clears about 95 out of every 100 people who do not. Those two numbers, sensitivity and specificity, sound almost interchangeable to a first-time reader, yet mixing them up is one of the costlier mistakes a clinician or a researcher can make when reading a diagnostic test paper. A test can be outstanding at one and mediocre at the other, and a headline accuracy figure often hides which is which.
What follows works through what sensitivity and specificity actually measure, how positive and negative predictive values move with disease prevalence, why likelihood ratios solve a problem raw accuracy figures cannot, and how the STARD 2015 checklist keeps diagnostic accuracy papers honest enough to trust. A worked 2×2 table, built from the real Cochrane figures above, walks through the arithmetic by hand so the formulas stop being abstract.
In Simple Words
Sensitivity measures how well a test catches people who actually have a condition. Specificity measures how well it clears people who do not. Those two numbers stay fairly fixed properties of the test itself. Positive and negative predictive values then translate that fixed test performance into the number a patient actually wants, which is the chance that their own positive or negative result is correct. That chance depends heavily on how common the condition is in whoever is being tested, not only on how good the test is, and it can swing by a wide margin even when the test never changes.
Key Takeaways
- Sensitivity is the proportion of people with a condition that a test correctly identifies as positive. Specificity is the proportion without the condition that it correctly identifies as negative.
- Positive and negative predictive values answer the question a patient actually asks, and both move substantially as disease prevalence changes, even when sensitivity and specificity do not.
- Likelihood ratios combine sensitivity and specificity into a single number that updates pretest probability into posttest probability, and unlike predictive values they stay roughly stable across populations.
- ROC curves and their area under the curve (AUC) summarize how well a test discriminates across every possible cutoff, not just the one a lab happens to use.
- STARD 2015 sets 30 reporting items for diagnostic accuracy studies, and adherence still varies widely across published papers.
What Sensitivity and Specificity Actually Measure
Every diagnostic test comparison starts with a 2×2 table that cross-tabulates the test result against the true disease status, usually established by a reference standard such as culture, biopsy, or a validated clinical criterion. Four outcomes are possible. A true positive is a sick person the test correctly flags. A false negative is a sick person the test misses. A true negative is a healthy person the test correctly clears. A false positive is a healthy person the test wrongly flags.
| Condition Present | Condition Absent | |
|---|---|---|
| Test Positive | True Positive (TP) | False Positive (FP) |
| Test Negative | False Negative (FN) | True Negative (TN) |
From those four cells, sensitivity and specificity are defined as simple ratios within each disease-status column, which is exactly why they behave differently from predictive values later in this guide.
- Sensitivity = TP / (TP + FN), calculated only among people who truly have the condition.
- Specificity = TN / (TN + FP), calculated only among people who truly do not.

Because each measure is calculated within a single column of the table, neither one depends mathematically on how common the disease is in the sample. Trevethan’s widely cited review of the foundations and pitfalls of these metrics makes the same point, and also warns that sensitivity and specificity are not the fixed constants they are often treated as in practice, since spectrum effects and the mix of disease severity in a study population can shift both from one setting to the next.
A Worked Example: Sensitivity and Specificity From Real Data
Cochrane’s 2025 update on Xpert Ultra for pulmonary tuberculosis pooled data from 32 studies and 12,529 adults and adolescents, rating the evidence as high certainty. Against a culture reference standard, the review reported a pooled sensitivity of 90.7 percent (95 percent CI, 88.2 to 92.7) and a pooled specificity of 94.8 percent (95 percent CI, 92.8 to 96.3).
Turning that into an explicit table makes the arithmetic easy to follow. Suppose 1,000 symptomatic adults are tested, and suppose 10 percent of them, or 100 people, truly have pulmonary tuberculosis, a prevalence chosen here only for illustration.
| TB Present (100) | TB Absent (900) | Row Total | |
|---|---|---|---|
| Xpert Ultra Positive | 91 (TP) | 47 (FP) | 138 |
| Xpert Ultra Negative | 9 (FN) | 853 (TN) | 862 |
Working through each formula by hand shows sensitivity as 91 divided by 100, or 91.0 percent, which lands close to the review’s precise pooled figure of 90.7 percent, with the small gap explained by rounding to whole people in a table of 1,000. Specificity is 853 divided by 900, or 94.8 percent, matching the pooled estimate exactly at this rounding. Positive predictive value is 91 divided by 138, or 65.9 percent, the share of positive tests that are truly tuberculosis. Negative predictive value is 853 divided by 862, or 99.0 percent, the share of negative tests that are truly free of disease.
Notice how far PPV sits below sensitivity and specificity even though the test performs well by both measures. That gap is not a flaw in the arithmetic, but the direct, mechanical consequence of testing a population where most people do not have the disease, a subject the next section takes up directly.
Why Predictive Values Swing With Disease Prevalence
Positive and negative predictive values answer a different question than sensitivity and specificity do. Instead of asking how the test performs among people whose true status is already known, PPV and NPV ask how much a result should be trusted once it comes back, which is the question a patient or clinician is actually facing. Bayes’ rule underlies that shift. A test updates a starting, or prior, probability of disease into a revised, or posterior, probability, and the size of that update depends on both the test’s accuracy and how likely disease was before the test was ever run.
Applying those same Xpert Ultra accuracy figures to a much lower-prevalence group illustrates the point starkly. Picture screening 1,000 people from a lower-risk population, where only 1 percent, or 10 people, truly have tuberculosis.
| TB Present (10) | TB Absent (990) | Row Total | |
|---|---|---|---|
| Xpert Ultra Positive | 9 (TP) | 51 (FP) | 60 |
| Xpert Ultra Negative | 1 (FN) | 939 (TN) | 940 |
Sensitivity and specificity barely move, staying at roughly 90 and 95 percent as before. PPV, though, collapses to 9 divided by 60, or 15.0 percent. Most positive results in this lower-prevalence group would actually be false alarms, not because the test got worse, but because healthy people vastly outnumber sick people going in. NPV, by contrast, climbs to 939 divided by 940, or 99.9 percent, since a negative result is even more reassuring when disease was already unlikely. Bours’s 2021 review of Bayes’ rule in diagnosis walks through this same logic from the clinician’s chair, framing prevalence as the pretest probability that any test result then revises.
Spectrum bias becomes relevant here too. Sensitivity and specificity are described as properties of the test, but they are actually measured in a specific study sample, and a sample enriched with severe, textbook-obvious cases tends to inflate both figures relative to what a broader, more mixed patient population would produce. A researcher applying a published sensitivity or specificity figure to a different clinical setting is implicitly assuming the case mix travels well, which is not always a safe assumption.
Likelihood Ratios: A Prevalence-Independent Way to Read a Result
Predictive values are useful but population-specific, which limits how far a single published PPV or NPV travels to a new clinical setting. Likelihood ratios were built to solve that portability problem. The positive likelihood ratio, LR+, is sensitivity divided by 1 minus specificity, and it states how many times more likely a positive result is in someone with the disease compared with someone without it. Its negative counterpart, LR-, is 1 minus sensitivity divided by specificity, and it states how much a negative result should lower the odds of disease. Deeks and Altman’s 2004 BMJ statistics note remains one of the clearest walkthroughs of how these ratios connect to pretest and posttest odds through Bayes’ theorem.
Plugging the Xpert Ultra pooled figures into both formulas gives LR+ of 0.907 divided by 0.052, or about 17.4, and LR- of 0.093 divided by 0.948, or about 0.10. An LR+ near 17 means a positive result multiplies the odds of tuberculosis roughly seventeenfold, a large and often practice-changing shift. An LR- near 0.10 cuts the odds of disease by about ninety percent, which is why a negative result from a highly sensitive test is often trusted enough to stop a workup. Ratios well above 10 or below 0.1 are conventionally treated as producing large, frequently decisive shifts in probability, while ratios sitting closer to 1 add little practical information regardless of how statistically significant a study reports them to be.
Cardiology offers another, independently verified illustration outside tuberculosis. A systematic review and meta-analysis covering 56 studies and 67,945 patients found that accelerated high-sensitivity cardiac troponin algorithms, run at 0, 1, or 2 hours after presentation, reached a pooled sensitivity above 90 percent for ruling out acute myocardial infarction, with the troponin-I-based algorithms reaching a pooled specificity above 80 percent (Lee et al., 2020, American Journal of Emergency Medicine). Emergency departments lean on that high sensitivity specifically because missing a real heart attack is far costlier than a short-lived false alarm, which is the same logic that makes a highly sensitive screening test useful even when its specificity, and therefore its PPV, is modest.
ROC Curves, AUC, and Choosing a Cutoff
Sensitivity and specificity are calculated at a single cutoff, but most diagnostic tests produce a continuous number, whether that is a troponin concentration, a tumor marker level, or a risk score. Moving the cutoff up or down trades sensitivity against specificity in opposite directions, so no single pair of numbers fully describes how a continuous test performs. Plotting sensitivity against 1 minus specificity across every possible cutoff is how the receiver operating characteristic curve solves that problem, and the area under that curve, or AUC, condenses the whole curve into one summary statistic ranging from 0.5, no better than a coin flip, to 1.0, perfect separation between people with and without the condition (Nahm, 2022, Korean Journal of Anesthesiology).

An AUC tells a researcher how well a test discriminates in principle, but a clinic still has to pick one operating point to use. Among cutoff-selection rules, the Youden index is the most common, selecting the cutoff that maximizes sensitivity plus specificity minus 1, which geometrically is the point on the ROC curve farthest from the diagonal line of no discrimination. Our Youden index calculator and formula guide walks through that calculation on a worked dataset, and our companion piece on threshold effects on diagnostic accuracy covers what happens when a study reports only the cutoff that made its own results look best. For readers comparing multiple predictors at once rather than a single test, our guide to logistic regression in medical diagnostics shows how ROC curves and AUC extend naturally to multivariable prediction models.
Reporting Diagnostic Accuracy Studies: The STARD 2015 Checklist
Sound arithmetic still needs a study design and a written report that readers can trust and replicate. STARD 2015, published simultaneously across several journals including the BMJ, sets out 30 essential reporting items organized around the standard introduction, methods, results, and discussion structure of a scientific article, and it now expects every diagnostic accuracy study to include a participant flow diagram (Bossuyt et al., 2015, BMJ). That checklist exists because incomplete reporting has been identified as a major, avoidable source of waste in biomedical research, leaving readers unable to judge a study’s risk of bias or decide whether its findings apply to their own patients.
Adherence remains inconsistent even now. An assessment of medical imaging diagnostic accuracy studies published in 2023 scored a sample of 116 papers against the full 30-item STARD-2015 checklist and used regression modeling to identify which study and journal characteristics predicted more complete reporting, evidence that gaps in reporting are still common enough to study a decade after the guideline’s release (White et al., 2025, Journal of Clinical Epidemiology). Readers evaluating a diagnostic accuracy paper should specifically check whether the reference standard is described, whether the sample size calculation is reported, and whether the flow diagram accounts for every enrolled participant, since gaps in exactly those three items are common enough to change how a reported sensitivity or specificity should be interpreted.
STARD sits alongside a family of reporting guidelines built for related study types. Our guide to TRIPOD reporting for prediction model studies covers the closest relative, used when a study combines several predictors into a risk score rather than evaluating one test in isolation. When several diagnostic accuracy studies of the same test need to be pooled quantitatively, our meta-analysis masterclass covers the heterogeneity and publication-bias checks that a diagnostic accuracy meta-analysis needs on top of a standard STARD read.
Common Pitfalls That Distort Sensitivity and Specificity Estimates
A handful of recurring design problems inflate or deflate sensitivity and specificity independently of how good a test actually is.
- Spectrum bias. Testing only clear-cut, advanced cases against healthy volunteers, rather than the full mix of severity a test will meet in real practice, tends to push both sensitivity and specificity higher than they would be in routine use.
- Threshold and cutoff effects. A single sensitivity or specificity figure describes exactly one operating point on the ROC curve, and comparing two tests fairly usually means comparing full curves or AUC values, not cherry-picked cutoffs.
- Imperfect reference standards. When the reference test used to classify true disease status is itself imperfect, errors in the reference standard bias the estimated sensitivity and specificity of the index test being evaluated, and the direction of that bias is not always intuitive.
- Small, underpowered samples. A sensitivity or specificity estimated from a handful of diseased participants can carry a confidence interval wide enough to make the point estimate close to meaningless, which is a sample-size planning problem as much as an analysis problem.
Sample size deserves its own emphasis here, since a diagnostic accuracy study needs enough diseased and non-diseased participants separately, not just a large total, to pin down both sensitivity and specificity with reasonable precision. Our guide to power analysis and sample size calculations covers the planning math, and the free Editverse sample size calculator applies it directly to a proposed study design.
Applying Sensitivity and Specificity in Practice
Turning these concepts into a decision starts with naming what a wrong answer would cost, since that answer usually points straight at which property of a test matters most.
| If the goal is… | Prioritize… |
|---|---|
| Ruling out a dangerous condition safely | High sensitivity, so a negative result can be trusted |
| Confirming a diagnosis before risky treatment | High specificity, so a positive result can be trusted |
| Screening a low-prevalence population | PPV alongside sensitivity, since false positives will often outnumber true ones |
| Comparing a test across different clinical settings | Likelihood ratios, since they travel across prevalence better than predictive values do |
Beyond single-test comparisons, researchers building multivariable diagnostic or prognostic models often need related statistical tools working together. Our guide to discriminant analysis for biomarker-based disease classification covers an alternative classification approach worth comparing against logistic regression, and our explainer on risk ratios and odds ratios covers the related effect measures that often appear alongside sensitivity and specificity in the same results table. Choosing between a Bayesian and a frequentist framing of these questions is its own decision, and our Bayesian versus frequentist statistics guide lays out that trade-off in more depth. For a broader map of how diagnostic accuracy metrics fit into the wider statistical toolkit, our statistical methods for evaluating diagnostic test performance guide is the natural next stop.
Need Help Computing or Reporting These Metrics?
Editverse’s PhD statisticians support sensitivity, specificity, ROC/AUC, and predictive-value analysis in SPSS, R, and STATA, and can help structure the write-up around the STARD 2015 checklist.
FAQ
What is the difference between sensitivity and specificity?
Sensitivity is the proportion of people who truly have a condition that a test correctly identifies as positive. Specificity is the proportion of people who truly do not have the condition that a test correctly identifies as negative. A test can score very differently on the two measures, so reporting only one gives an incomplete picture of how it performs.
Why do positive and negative predictive values change even when the test stays the same?
PPV and NPV depend on disease prevalence as well as on sensitivity and specificity. When a condition is rare in the population being tested, positive results are more often false alarms even from an accurate test, which pulls PPV down while NPV stays high. Both worked examples above show the same Xpert Ultra test producing a PPV of 65.9 percent at 10 percent prevalence and just 15.0 percent at 1 percent prevalence.
What do likelihood ratios add that sensitivity and specificity do not?
Likelihood ratios combine sensitivity and specificity into a single number that shows how much a test result should shift the odds of disease, and unlike predictive values, they stay comparatively stable across populations with different disease prevalence. An LR+ above 10 or an LR- below 0.1 is generally considered to produce a large, often decisive shift in probability.
What does the area under the ROC curve actually mean?
AUC summarizes how well a test discriminates between people with and without a condition across every possible cutoff, on a scale from 0.5, equivalent to a coin flip, to 1.0, perfect separation. That answers a different question than sensitivity and specificity, which only describe performance at one chosen cutoff.
Why does the STARD 2015 checklist matter for a diagnostic accuracy manuscript?
STARD 2015 sets out 30 reporting items and a required participant flow diagram so readers can judge a diagnostic accuracy study’s risk of bias and decide whether its findings generalize to their own setting. Many journals now check submissions against it directly, and incomplete reporting on items like the reference standard or sample size justification is still common enough to be worth double-checking before submission.
Conclusion
Sensitivity and specificity describe a test’s behavior in two separate, fixed populations, people who have a condition and people who do not, which is exactly why neither one alone tells a patient what a personal result means. Predictive values fill that gap but travel poorly between settings with different disease prevalence, while likelihood ratios and the STARD 2015 reporting checklist give researchers and clinicians the tools to interpret a result honestly and to write up a study so someone else can trust it. Both tables above use one real, high-certainty accuracy estimate to show how far a positive predictive value can move without the underlying test changing at all, a habit of mind worth carrying out of this guide. Before trusting any diagnostic accuracy figure, ask what population it was measured in, and whether that population looks anything like the one in front of you.
Related Editverse Resources
→ Statistical Methods for Evaluating the Performance of Medical Diagnostic Tests
→ Youden Index Calculator and Formula
→ The Impact of Threshold Effects on Diagnostic Accuracy
→ The Impact of Spectrum Bias on Diagnostic Test Performance
→ Logistic Regression in Medical Diagnostics
→ Discriminant Analysis Visualized: Biomarker-Based Disease Classification
→ TRIPOD of Success: Prediction Model Studies
→ Meta-Analysis Masterclass: Step-by-Step Examples
→ Bayesian vs. Frequentist Statistics
→ Risk Ratios and Odds Ratios: Understanding and Calculating
→ Power Analysis Made Easy: Sample Size Calculations
References
- 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. https://doi.org/10.1136/bmj.h5527
- Trevethan R. Sensitivity, Specificity, and Predictive Values: Foundations, Pliabilities, and Pitfalls in Research and Practice. Frontiers in Public Health. 2017;5:307. https://doi.org/10.3389/fpubh.2017.00307
- Nahm FS. Receiver operating characteristic curve: overview and practical use for clinicians. Korean Journal of Anesthesiology. 2022;75(1):25-36. https://doi.org/10.4097/kja.21209
- Deeks JJ, Altman DG. Diagnostic tests 4: likelihood ratios. BMJ. 2004;329(7458):168-169. https://doi.org/10.1136/bmj.329.7458.168
- Bours MJL. Bayes’ rule in diagnosis. Journal of Clinical Epidemiology. 2021;131:158-160. https://doi.org/10.1016/j.jclinepi.2020.12.021
- Horne DJ, Zifodya JS, Shapiro AE, et al. Xpert MTB/RIF Ultra assay for pulmonary tuberculosis and rifampicin resistance in adults and adolescents. Cochrane Database of Systematic Reviews. 2025;7:CD009593. https://doi.org/10.1002/14651858.CD009593.pub6
- Lee CC, Huang SS, Yeo YH, et al. High-sensitivity-cardiac troponin for accelerated diagnosis of acute myocardial infarction: A systematic review and meta-analysis. American Journal of Emergency Medicine. 2020;38(7):1402-1407. https://doi.org/10.1016/j.ajem.2019.11.035
- White SJ, Chau M, Arruzza E, et al. Assessment of Standards for Reporting of Diagnostic Accuracy (STARD) 2015 guideline adherence in medical imaging diagnostic accuracy studies published in 2023. Journal of Clinical Epidemiology. 2025;179:111654. https://doi.org/10.1016/j.jclinepi.2024.111654
