Sensitivity, Specificity, PPV, NPV and ROC Curves: What Researchers Often Confuse

Start With the Direction of the Probability

Suppose disease status is denoted by D and the binary test result by T.

The four familiar measures ask four different conditional questions:

Rosner defines sensitivity as the probability that the test or symptom is present given disease, and specificity as the probability that it is absent given no disease. Predictive values reverse that conditioning: positive predictive value concerns disease given a positive result, while negative predictive value concerns absence of disease given a negative result (Rosner, 2016).

Altman's standard 2 × 2 representation makes the distinction especially clear:

From this table:

Sensitivity

a / (a + c)

Specificity

d / (b + d)

PPV

a / (a + b)

NPV

d / (c + d)

The denominators reveal the direction of conditioning. Sensitivity and specificity divide people according to true disease status. PPV and NPV divide them according to test result (Altman, 1991).

Sensitivity: Among People With Disease, How Often Is the Test Positive?

Sensitivity is the proportion—or probability—of diseased individuals who test positive:

P(T+ | D+)

A highly sensitive test has a relatively low proportion of false-negative results among diseased individuals. Its complement, 1 − sensitivity, is the false-negative fraction when that fraction is defined conditional on disease status (Altman, 1991; Rosner, 2016).

Common misinterpretation

A sensitivity of 90% does not mean:

“If the test is positive, there is a 90% probability that the patient has disease.”

That statement describes PPV, not sensitivity.

Sensitivity conditions on disease being present. It does not condition on a positive test result.

Specificity: Among People Without Disease, How Often Is the Test Negative?

Specificity is:

P(T− | D−)

It measures the proportion of non-diseased individuals correctly classified as negative. Its complement, 1 − specificity, is the false-positive fraction:

FPF = P(T+ | D−)

(Altman, 1991; Pepe, 2003).

Common misinterpretation

A specificity of 95% does not mean:

“A negative result means there is a 95% probability the patient does not have disease.”

That is an NPV interpretation.

Specificity begins with the non-diseased population and asks how frequently the test is negative.

PPV: Given a Positive Test, What Is the Probability of Disease?

Positive predictive value reverses the direction:

PPV = P(D+ | T+)

It therefore addresses a question much closer to what is often asked after a result becomes available: among people with a positive result, what proportion actually have the condition? Altman explicitly distinguishes this direction from sensitivity and specificity (Altman, 1991).

The crucial qualification is that PPV depends on the prevalence—or prior probability—of disease in the population to which the test is applied.

Using sensitivity (Se), specificity (Sp), and prevalence (p):

PPV = (Se × p) / [Se × p + (1 − Sp)(1 − p)]

(Altman, 1991).

This means PPV is not simply an intrinsic characteristic that can automatically be transported from one population to another.

NPV: Given a Negative Test, What Is the Probability of No Disease?

Negative predictive value is:

NPV = P(D− | T−)

and can be written as:

NPV = [Sp(1 − p)] / [(1 − Se)p + Sp(1 − p)]

where p is prevalence (Altman, 1991).

Like PPV, NPV depends on the prevalence of disease in the population being tested.

The Prevalence Effect Researchers Often Miss

One of the most important distinctions in sensitivity specificity PPV NPV interpretation is that predictive values change with prevalence.

Altman shows directly from Bayes' theorem that PPV and NPV depend on sensitivity, specificity, and disease prevalence. If prevalence is unknown, the predictive values applicable to a target population cannot be determined from sensitivity and specificity alone (Altman, 1991).

Holding sensitivity and specificity fixed, reducing disease prevalence generally reduces PPV and increases NPV. Altman's numerical illustration demonstrates that predictive values can change substantially even while sensitivity and specificity remain unchanged (Altman, 1991).

This becomes particularly important when the study design artificially determines the proportions of diseased and non-diseased participants.

Why Case-Control Diagnostic Studies Need Special Care

In a diagnostic case-control study, researchers deliberately select numbers of diseased cases and non-diseased controls. The resulting fraction of cases in the dataset therefore need not equal disease prevalence in the population.

Pepe explains that sensitivity/true-positive fraction and false-positive fraction can still be estimated from such sampling because these quantities condition on disease status. Diagnostic likelihood ratios can likewise be estimated. Naively estimating PPV or NPV from the observed case-control sample, however, is inappropriate when cases have been oversampled. Population prevalence is needed to recover population-relevant predictive values (Pepe, 2003).

Altman makes the same practical point for samples in which diseased and non-diseased groups have been selected independently: the proportion diseased in the assembled study may have no useful interpretation as population prevalence (Altman, 1991).

Diagnostic Likelihood Ratios: From Pre-Test to Post-Test Evidence

Likelihood ratios provide another way to express what a test result contributes.

For a positive binary test:

LR+ = P(T+ | D+) / P(T+ | D−)

or equivalently:

LR+ = Sensitivity / (1 − Specificity)

Altman describes the likelihood ratio as comparing the probability of obtaining a particular test result in diseased patients with the probability of obtaining that result in non-diseased patients. The positive likelihood ratio links pre-test and post-test odds:

Post-test odds = Pre-test odds × LR+

(Altman, 1991).

Pepe similarly defines positive and negative diagnostic likelihood ratios from the true- and false-positive fractions. For a binary test:

DLR+ = TPF / FPF

and

DLR− = (1 − TPF) / (1 − FPF)

(Pepe, 2003).

Continuous Tests Require a Threshold Before Sensitivity and Specificity Are Fixed

Many medical tests produce continuous measurements rather than an inherently positive or negative result.

To create a binary decision, a threshold c must be selected. If larger measurements indicate disease, the test might be defined as:

T+ = 1 if Y ≥ c

and negative otherwise.

Each possible threshold generates its own true-positive and false-positive performance. Pepe defines:

TPF(c) = P(Y ≥ c | D = 1)

and

FPF(c) = P(Y ≥ c | D = 0)

with the ROC curve consisting of the set of attainable (FPF(c), TPF(c)) pairs across thresholds (Pepe, 2003).

Altman similarly describes evaluating a continuous diagnostic measurement by considering each possible cutoff in turn and calculating sensitivity and specificity for each cutoff (Altman, 1991).

ROC Curve Explained: It Is a Map of Threshold Trade-Offs

An ROC curve plots:

• false-positive fraction on the horizontal axis; and
• true-positive fraction (sensitivity) on the vertical axis,

across possible thresholds.

Because specificity is 1 − FPF, the horizontal coordinate is commonly written as 1 − specificity.

As the threshold changes, the balance between identifying diseased individuals and falsely classifying non-diseased individuals changes. Pepe describes the ROC curve as the set of possible true-positive and false-positive fractions attainable by dichotomizing a continuous test at different thresholds (Pepe, 2003).

Those are different questions.

The Sensitivity–False-Positive Trade-Off

Suppose high test values indicate disease.

Lowering the threshold classifies more people as positive. That tends to increase the true-positive fraction—but also increases the false-positive fraction.

Raising the threshold does the opposite: fewer non-diseased people are classified positive, but more diseased people may also fall below the threshold.

Pepe emphasizes that threshold choice implicitly represents a trade-off between failing to detect disease and falsely identifying disease. The ROC curve describes the range of such trade-offs that the test can achieve (Pepe, 2003).

A point toward the upper-left region of ROC space combines high true-positive performance with low false-positive performance. A completely uninformative test has equal true- and false-positive fractions and therefore follows the diagonal line. A perfect test can achieve a true-positive fraction of 1 with a false-positive fraction of 0 (Pepe, 2003).

What AUC Means—and What It Does Not

The area under the ROC curve (AUC) summarizes separation between the distributions of test values in diseased and non-diseased individuals.

For a test oriented so that higher values indicate disease, the AUC has a probability/ranking interpretation related to the probability that a randomly selected diseased observation has a higher test value than a randomly selected non-diseased observation (Pepe, 2003).

For an uninformative test, AUC is 0.5. Perfect separation corresponds to an AUC of 1.

But AUC is a summary across thresholds. It can conceal differences in the particular region of the ROC curve that matters clinically. Pepe notes that substantial differences between ROC curves over a restricted false-positive range can translate into relatively small differences in overall AUC (Pepe, 2003).

For some applications, only a limited range of false-positive fractions is operationally acceptable. Pepe therefore discusses restricting attention to clinically relevant ROC regions or using measures such as partial AUC rather than assuming the entire ROC domain is equally important (Pepe, 2003).

Medical-test accuracy and prediction-model calibration are distinct evaluation problems.

Metric Comparison: What Does Each Measure Actually Condition On?

The first four probability directions are directly supported by Altman and Rosner; the ROC, AUC, and likelihood-ratio distinctions follow Pepe's diagnostic-accuracy framework (Altman, 1991; Pepe, 2003; Rosner, 2016).

Why “Sensitivity and Specificity Do Not Depend on Prevalence” Needs Qualification

There are two different ideas that researchers sometimes collapse into one statement.

First, if sensitivity and specificity are fixed and only the proportions of diseased and non-diseased observations are changed, their calculations do not change because each is conditional on disease status. PPV and NPV do change. Altman's prevalence examples demonstrate exactly this mathematical distinction (Altman, 1991).

Second, that does not imply sensitivity and specificity must be identical in every clinical population. Diagnostic accuracy can vary when the spectrum of diseased or non-diseased subjects changes.

Pepe identifies spectrum bias when cases or controls in a diagnostic study are not representative of the target diseased and non-diseased populations. Selecting unusually severe cases and unusually healthy controls, for example, can exaggerate apparent diagnostic accuracy (Pepe, 2003).

Diagnostic Accuracy Begins With Study Design

Correct formulas cannot rescue a poorly designed diagnostic-accuracy study.

Pepe identifies several potential sources of bias, including:

• verification bias, when definitive disease assessment is not performed randomly;
• imperfect-reference bias, when the reference standard itself misclassifies true disease status;
• spectrum bias, when cases or controls do not represent their target populations;
• test-interpretation bias, when information unavailable in intended practice influences interpretation;
• problems caused by excluding or mishandling unsatisfactory or uninterpretable tests; and
• extrapolation bias, when study conditions or populations differ from those in which the test will actually be used (Pepe, 2003).

The reference standard is particularly important because sensitivity, specificity, PPV, and NPV all rely on knowing—or estimating—the underlying disease classification against which the index test is evaluated.

A practical diagnostic-accuracy design check

Before interpreting any reported metric, ask:

1. Who was sampled? Do cases and non-cases represent the intended-use population?
2. How was disease status established? Is the reference assessment sufficiently credible?
3. Who received verification? Could the decision to perform the reference test depend on the index-test result?
4. Was test interpretation protected from inappropriate information?
5. How were uninterpretable results handled?
6. Was the study prevalence naturally observed or imposed by design?
7. Is the reported threshold prespecified, clinically established, or selected after examining the data?
8. Does the intended clinical setting resemble the study setting?

These questions determine whether the reported diagnostic test accuracy is relevant to the population in which researchers want to use it.

Threshold Selection Is a Clinical Decision as Well as a Statistical One

A common mistake is to calculate an ROC curve, identify a mathematically attractive point, and call it the “optimal” threshold without specifying what errors matter clinically.

Altman is explicit that choosing a cutoff is not simply a statistical decision. The preferred cutoff depends on the relative consequences of false-positive and false-negative results and on what clinical action follows a positive result (Altman, 1991).

Pepe likewise describes threshold selection in terms of the acceptable trade-off between missed disease and false identification. Different applications can justify focusing on different parts of the ROC curve. A screening setting may, for example, impose stringent requirements on false-positive performance, whereas another setting may prioritize maintaining a very high true-positive fraction (Pepe, 2003).

Therefore the clinically relevant threshold should be considered in relation to questions such as:

• What happens after a positive result?
• What is the consequence of missing a true case?
• What burden follows a false-positive result?
• Is the test being used for screening, diagnosis, triage, or another purpose?
• Are there practical constraints on the number of false-positive referrals or follow-up procedures?

Five Confusions to Eliminate From Diagnostic-Test Reports

1. Sensitivity is not PPV

Sensitivity asks:

Among people who have disease, how many test positive?

PPV asks:

Among people who test positive, how many have disease?

The conditioning is reversed.

2. Specificity is not NPV

Specificity conditions on being non-diseased.

NPV conditions on receiving a negative test.

Again, the direction is reversed.

3. PPV and NPV are not universally portable test characteristics

They depend on disease prevalence. Values obtained from a selected study sample may therefore be inappropriate for a different clinical population (Altman, 1991; Pepe, 2003).

4. ROC/AUC does not choose the clinical threshold

The ROC curve displays attainable TPF–FPF trade-offs. Choosing among those operating points requires consideration of the consequences of false-positive and false-negative decisions (Altman, 1991; Pepe, 2003).

5. AUC is not calibration

AUC summarizes separation/ranking of test values between disease groups. It does not establish that predicted probabilities from a risk model correspond to observed risks. Researchers evaluating probability predictions must treat calibration as a separate property rather than translating an AUC into a calibration claim.

Reporting Checklist for Researchers

When reporting sensitivity specificity PPV NPV and ROC analyses, make the statistical question visible.

• State the target condition and reference standard.
• Describe how diseased and non-diseased participants were selected.
• State the threshold used for every threshold-dependent estimate.
• Report sensitivity and specificity with uncertainty, rather than point estimates alone.
• Report PPV and NPV only with a clear description of the prevalence/population to which they apply.
• Do not estimate population PPV or NPV naively from an artificially assembled case-control sample.
• If reporting likelihood ratios, distinguish them from predictive probabilities.
• For a continuous test, explain how thresholds were evaluated or selected.
• If reporting an ROC curve, state that it summarizes the attainable true-positive/false-positive trade-off across thresholds.
• If reporting AUC, interpret it as a summary of separation/ranking rather than as a clinical threshold.
• Examine whether the clinically relevant region of the ROC curve matters more than the full AUC.
• Describe potential verification, reference-standard, spectrum, interpretation, and extrapolation biases.
• Connect threshold selection to the consequences of false-positive and false-negative decisions.
• Do not interpret diagnostic-test ROC/AUC as evidence of prediction-model calibration.

A Compact Decision Framework

When a researcher asks, “How accurate is this test?”, that question should be decomposed before selecting a metric.

Conclusion

The most important distinction in sensitivity specificity PPV NPV is not the formula—it is the direction of conditioning.

Sensitivity and specificity ask what the test does given disease status. PPV and NPV ask what disease status is likely given the test result. Because of that reversal, predictive values depend on prevalence and should not automatically be transported across populations.

For continuous tests, changing the threshold changes the balance between true-positive and false-positive performance. The ROC curve explained correctly is therefore a map of attainable threshold trade-offs, while AUC medical test interpretation concerns overall separation or ranking. Neither the ROC curve nor AUC determines which threshold has acceptable clinical consequences.

Finally, diagnostic accuracy is only as credible as the study used to estimate it. Subject selection, disease verification, the reference standard, test interpretation, intended-use population, and handling of thresholds all affect what the reported numbers mean.

FAQs

What is the difference between sensitivity and positive predictive value?

Sensitivity is P(T+ | D+): the probability of a positive test among people who have disease. Positive predictive value is P(D+ | T+): the probability of disease among people who have tested positive. They reverse the direction of conditioning (Altman, 1991; Rosner, 2016).

Does prevalence affect sensitivity and specificity?

Changing only the proportion of diseased versus non-diseased subjects does not mathematically change sensitivity or specificity because these measures condition on disease status. However, sensitivity and specificity can differ across populations when the disease or non-disease spectrum, testing procedures, or other relevant characteristics differ (Altman, 1991; Pepe, 2003).

Why does prevalence affect positive predictive value?

PPV is the probability of disease given a positive result. Bayes' theorem shows that this probability depends on sensitivity, specificity, and the prior probability or prevalence of disease. Consequently, the same sensitivity and specificity can produce very different PPVs in populations with different prevalence (Altman, 1991).

What happens to PPV when disease becomes rarer?

Holding sensitivity and specificity fixed, lower prevalence generally produces a lower PPV and a higher NPV. A larger share of positive results can then come from the much larger non-diseased population (Altman, 1991).

What is a diagnostic likelihood ratio?

A diagnostic likelihood ratio compares the probability of a test result in diseased people with its probability in non-diseased people. For a positive binary result, LR+ = Sensitivity / (1 − Specificity). It can be used to convert pre-test odds into post-test odds (Altman, 1991; Pepe, 2003).

What does an ROC curve show?

An ROC curve shows the combinations of true-positive fraction and false-positive fraction attainable as the threshold for a test changes. It therefore describes the trade-off between detecting disease and falsely classifying non-diseased people as positive (Pepe, 2003).

What does AUC mean for a medical test?

AUC summarizes the separation or ranking performance represented by the ROC curve. It has a probabilistic interpretation related to how often a randomly selected diseased observation has a higher test value than a randomly selected non-diseased observation when higher values indicate disease (Pepe, 2003).

Does a high AUC mean a test is clinically useful?

Not by itself. AUC summarizes discrimination across thresholds, whereas actual use requires an operating threshold and consideration of the consequences of false-positive and false-negative decisions. Performance in the clinically relevant part of the ROC curve can also matter more than overall AUC (Altman, 1991; Pepe, 2003).

How should researchers choose a diagnostic-test threshold?

Threshold selection should reflect the intended clinical use and the relative consequences of false-positive and false-negative decisions. It should not be treated as a purely statistical optimization problem (Altman, 1991; Pepe, 2003).

Can PPV and NPV be calculated directly from a case-control diagnostic study?

Not generally as population predictive values when the numbers of cases and controls were fixed or otherwise artificially sampled. Sensitivity/TPF and false-positive performance can be estimated because they condition on disease status, but population-relevant PPV and NPV require information about population prevalence (Pepe, 2003).

References

Altman, D. G. (1991). Practical statistics for medical research. Chapman & Hall.

Pepe, M. S. (2003). The statistical evaluation of medical tests for classification and prediction. Oxford University Press.

Rosner, B. (2016). Fundamentals of biostatistics (8th ed.). Cengage Learning.