Specificity

Specificity

Specificity is the proportion of people who truly do not have a specified target condition and receive a negative result from an index test. It measures how often the test correctly excludes the condition among reference-negative people at a stated threshold and in a stated setting. It is also called the true-negative rate. The term here refers to diagnostic classification, not whether a statistical association is narrowly specific to one outcome.

Definition and denominator

In a 2-by-2 comparison with an appropriate reference standard, true negatives (TN) do not have the condition and test negative; false positives (FP) do not have it but test positive. Estimated specificity is

$$\widehat{\mathrm{Sp}}=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}},$$

provided $\mathrm{TN}+\mathrm{FP}>0$. In probability notation, $\mathrm{Sp}=\Pr(T^-\mid D^-)$ for the stated test, threshold, condition and population. The denominator is all people without the condition according to the reference standard. A specificity of 80% means that approximately 80 of every 100 condition-negative people test negative in the evaluated setting. It does not mean that 80% of negative test results are correct.

The false-positive rate among condition-negative people is $1-\mathrm{Sp}=\mathrm{FP}/(\mathrm{TN}+\mathrm{FP})$. For a binary classification threshold, specificity and false-positive rate are complements for the same population and classification rules. Neither is the false-discovery proportion among positive results, whose denominator is instead all positive tests.

Worked 2-by-2 example

Suppose 1,000 people are evaluated with both an index test and credible reference standard for a specified condition. One hundred have the condition; 900 do not. Of those without the condition, 720 test negative and 180 test positive. Among those with it, 90 test positive and 10 negative. These invented counts match the paired Sensitivity teaching example and are not a real test evaluation.

Specificity is $720/(720+180)=720/900=0.80$, or 80%. The false-positive rate is $180/900=0.20$, or 20%. Sensitivity, the fraction of reference-positive people detected, is $90/(90+10)=0.90$. The positive predictive value in this 10%-prevalence sample is $90/(90+180)=1/3\approx33.3\%$; the negative predictive value is $720/(720+10)\approx98.6\%$. Do not substitute predictive values for sensitivity or specificity: each conditions on a different population.

With the same 80% specificity and 90% sensitivity, but a condition prevalence of 1% in a different population, Bayes' rule yields a negative predictive value of

$$\Pr(D^-\mid T^-)=\frac{0.80(0.99)}{0.80(0.99)+(1-0.90)(0.01)}\approx0.9987.$$

The corresponding positive predictive value would be about 4.35%, despite unchanged assumed sensitivity and specificity. These illustrative transport calculations require the accuracy parameters to remain valid in the new population. Disease spectrum, stage, sampling or test administration may change them; specificity is not guaranteed to be identical in all groups merely because prevalence does not appear in its conditional definition.

Test thresholds and clinical pathways

For an index measurement $X$ called positive when $X\geq c$, specificity at cutoff $c$ is $\Pr(X<c\mid D^-)$. Raising the cutoff typically reduces false positives and increases specificity, but can miss more people with the condition and reduce sensitivity. If low values indicate disease, reverse the inequality and interpret the trade-off accordingly. Show the chosen cutoff alongside both accuracy measures rather than describing either one as an intrinsic constant of a test.

A receiver operating characteristic plot places sensitivity against $1-\mathrm{Sp}$ across possible thresholds. Its area summarises ranking over thresholds, but it cannot tell a clinician the specificity at the single threshold actually selected for action. An apparently excellent overall area can coexist with unsuitable performance at the false-positive rate a service can tolerate. Thresholds should be chosen with the relative consequences of false positives and false negatives, prevalence and care capacity in view.

In a two-stage screen-and-confirm pathway, the specificity of the whole pathway is not generally equal to the confirmatory test's stand-alone specificity. It depends on joint test results among those without disease, whether only screen-positive people are confirmed, and how discordant or indeterminate results are handled. Report the complete pathway and use conditional probabilities or joint data; multiplying marginal specificities without the required independence assumptions can be wrong.

Estimation quality and uncertainty

The reference standard should classify absence of the condition credibly in the population of intended use. If its own classification is imperfect, an estimate conditional on it can differ from true clinical specificity. Blinding index and reference readings helps prevent interpretation bias, and applying the reference standard to only a selected subset risks verification bias. Specify the interval between tests, because status can change; avoid incorporation bias if the index result helps define reference status.

Sampling precision depends primarily on the number of reference-negative participants, not just the overall sample size. Here 720 of 900 condition-negative people test negative; report a confidence interval and method rather than only the 80% estimate. Subgroup, clustered-site, repeated-test and missing-data analyses may need methods beyond a simple independent binomial interval. A denominator with no reference-negative people leaves specificity undefined, not zero or 100%.

Specify whether a borderline, failed or indeterminate result is called positive, negative or separately classified, and count every eligible person in a transparent flow diagram. Excluding difficult-to-read results can make performance look better than what a care pathway delivers. In studies with healthy controls selected far from the clinical diagnostic boundary, specificity may be overstated for symptomatic people who need testing in practice. Evaluate relevant mimicking conditions and likely alternatives.

Consequences in health economics

At a fixed prevalence $\pi$ among $N$ eligible people, expected false positives are $N(1-\pi)(1-\mathrm{Sp})$ under the stated test model. For the worked population, $1{,}000(0.90)(0.20)=180$ false positives. Those results can lead to repeat tests, specialist visits, invasive confirmation, anxiety, inappropriate treatment or capacity pressure. Increasing specificity may reduce those burdens but can increase missed cases; a full economic evaluation models both branches and their downstream health and resource outcomes.

The monetary or health value of an avoided false positive varies by confirmatory pathway, harms, disease severity and available capacity. A test with high specificity is not automatically cost-effective if it is expensive, insensitive or used in a population where follow-up has little value. Express sensitivity, specificity and prevalence for the same target condition, threshold and time horizon in a decision tree. Include parameter uncertainty and dependence if the same studies or thresholds generated the estimates.

Reporting and source guidance

Report the intended-use population, target condition and reference standard, index test version and cutoff, complete 2-by-2 counts, treatment of missing and indeterminate results, both sensitivity and specificity with intervals, subgroup and site performance, and the proposed pathway after results. The FDA diagnostic test reporting guidance discusses the reference method and accuracy or agreement claims. STARD 2015 specifies transparent diagnostic accuracy reporting. The Cochrane Handbook for Diagnostic Test Accuracy Reviews supports critical appraisal and synthesis across settings. No illustrative accuracy value on this page should be used as a clinical input without a suitable empirical source.