Why a Test's Positive Predictive Value Depends on Prevalence

A result marked “positive” feels definitive. It is not. A test result changes a probability, and the size of that change depends on how likely the condition was before testing.

This is easiest to see with natural frequencies. Imagine one test used in two groups. Its sensitivity is 95%, so it detects 95 of every 100 people who truly have the condition; its specificity is also 95%, so 5 of every 100 people without the condition test positive. The test characteristics sound excellent in both groups. The meaning of a positive result still changes sharply when the condition is rare.

Start with the two-by-two table#

Every binary diagnostic study can be organized by actual condition status and test result:

• True positive: the condition is present and the test is positive.
• False positive: the condition is absent and the test is positive.
• False negative: the condition is present and the test is negative.
• True negative: the condition is absent and the test is negative.

Sensitivity looks down the column of people with the condition. Specificity looks down the column of people without it. PPV looks across the positive-result row. Negative predictive value, or NPV, looks across the negative-result row.

That change in denominator is the whole lesson. Sensitivity starts with known disease status. PPV starts with the result you actually received.

Constructed example: 1% prevalence#

Take an invented group of 10,000 people. If prevalence is 1%, then 100 have the condition and 9,900 do not. A 95%-sensitive test finds 95 of the 100 cases and misses 5; a 95%-specific test correctly returns 9,405 negatives among those without the condition, but it also produces 495 false positives.

There are 590 positive results: 95 true positives plus 495 false positives. PPV is 95 divided by 590, or about 16.1%. Most positive results in this example are false even though the test is 95% sensitive and 95% specific.

This is not a defect in arithmetic or proof that the test is useless. The false-positive rate is applied to a much larger group than the true-positive rate. Five percent of 9,900 is larger than 95% of 100.

Constructed example: 20% prevalence#

Now use the same test in another invented group of 10,000 where prevalence is 20%. Two thousand people have the condition and 8,000 do not. The test returns 1,900 true positives, 100 false negatives, 400 false positives, and 7,600 true negatives.

There are 2,300 positive results, of which 1,900 are true. PPV is 82.6%. Nothing about the stated sensitivity or specificity changed. The number of true cases available to be detected changed relative to the number of people who could generate false positives.

NPV moves in the opposite direction. Negative results are especially convincing in the 1%-prevalence group because almost everyone starts without the condition, and as prevalence rises, false negatives make up a larger share of all negative results.

The formulas, with the denominators visible#

The core measures are:

• Sensitivity = true positives / (true positives + false negatives)
• Specificity = true negatives / (true negatives + false positives)
• PPV = true positives / (true positives + false positives)
• NPV = true negatives / (true negatives + false negatives)

PPV can also be calculated from sensitivity, specificity, and prevalence:

PPV = sensitivity × prevalence / [(sensitivity × prevalence) + ((1 - specificity) × (1 - prevalence))]

The formula is useful for modeling, but natural frequencies are often safer for communication. A statement such as “95 of 590 positives are true in this constructed group” exposes the numerator and denominator and makes the false positives harder for you to overlook.

Prevalence is not exactly pretest probability#

Prevalence is the proportion with a condition in a defined population at a defined time. It might describe a community survey, an emergency department, a referral clinic, or people selected for testing. Those populations can have very different risk.

Pretest probability applies that background to an individual clinical question. It can incorporate symptoms, duration, known contact, age, physical findings, vaccination, local transmission, medications, prior imaging, or another test. A clinician's estimate can be imperfect, but it is more specific than a national average.

Selection into testing matters. If only people with classic symptoms are tested, prevalence among those tested may be high. If an employer tests everyone repeatedly regardless of symptoms, it may be low. Quoting a PPV from one pathway in the other can mislead you. That is why the analytic performance printed on a test report is not a personal posterior probability: the result still has to be combined with the setting in which it was ordered.

Bayes' theorem and likelihood ratios#

Bayes' theorem formalizes the update from pretest to post-test probability. Clinicians often use likelihood ratios because they separate the test's information from the starting odds.

The positive likelihood ratio is sensitivity divided by one minus specificity. With both at 95%, LR+ is 19. The negative likelihood ratio is one minus sensitivity divided by specificity, or about 0.053.

Convert pretest probability to odds, multiply by the likelihood ratio, then convert back to probability, and this yields the same result as the PPV formula when prevalence is the pretest probability.

An LR+ of 19 is a strong upward shift, but nineteen times very small odds can still leave meaningful uncertainty. A strong test cannot make the starting probability irrelevant. The related article on pretest probability and likelihood ratios develops this method further.

Sensitivity and specificity can change too#

Teaching examples often hold sensitivity and specificity constant. Real tests can behave differently across populations. This is called the spectrum effect.

People with severe, established disease may be easier to detect than people with early or mild disease, and healthy controls can be easier to classify than patients with similar competing diagnoses. Specimen timing, sample quality, age, treatment, immune status, instrument, operator, and threshold can all affect results.

A case-control study that compares obvious cases with healthy volunteers can therefore overestimate performance in the intended clinic. STARD 2015 asks diagnostic studies to report eligibility, setting, participant selection, index-test threshold, reference standard, and flow so that you can judge applicability.

Verification bias arises when only people with positive results receive the definitive reference test. Incorporation bias occurs when the test being evaluated contributes to the reference diagnosis. Missing or indeterminate results can also be excluded in ways that make performance look better.

Thresholds trade false positives for false negatives#

Many tests measure a continuous signal, and a cutoff turns that signal into a positive or a negative. Lowering the threshold generally raises sensitivity and creates more false positives; raising it generally raises specificity and creates more false negatives.

The best threshold depends on consequences. Missing a time-sensitive, treatable, serious condition may be much worse than performing a confirmatory test. In another setting, a false positive may lead to an invasive procedure, stigma, toxic treatment, or major anxiety, so specificity deserves more weight. There is no universally optimal cutoff independent of the purpose you have in mind, and screening, triage, confirmation, treatment selection, and disease monitoring can each require a different threshold on the same underlying measurement.

Why broad low-risk testing can cause harm#

When the target condition is rare, false positives can outnumber true positives. Each positive may lead to repeat tests, imaging, biopsy, isolation, medication, time away from work, cost, or worry. Some of those follow-up steps have their own false-positive and complication rates.

This does not mean never test low-risk people. Population screening can be worthwhile when disease burden, test strategy, confirmatory pathway, treatment benefit, and program harms support it. The evaluation must include the whole pathway, not just assay accuracy.

Incidental testing without a plan is different. A result is less useful when no one has decided what positive, negative, or indeterminate findings will change. Testing can create a diagnostic cascade without improving outcomes.

Confirmation helps, but independence cannot be assumed#

A second, more specific test can raise confidence after an initial positive. Repeating the same test can also help when random analytic error is important. The combined accuracy depends on the rule used, the tests' thresholds, and whether their errors are independent.

Two tests that use the same specimen, antigen, imaging feature, or flawed reference standard may fail together. Simply multiplying false-positive rates assumes independence and can exaggerate the value of repetition. Timing can also change the underlying condition between tests. The confirmation plan should be defined before testing, and it should state whether a different method is required, how discordant results are resolved, and what clinical follow-up serves as the reference.

How to communicate a positive result#

A useful explanation gives you four parts:

1. The result and what threshold produced it.
2. The estimated pretest probability and why.
3. The post-test probability or a natural-frequency example.
4. The next step, including confirmation, treatment, observation, or reassessment.

Use one denominator and one time horizon. “Sixteen percent of positives are true in a group like this” is clearer than alternating among sensitivity, relative risk, and odds without explanation. State the uncertainty when local prevalence or test performance is poorly known.

The FDA diagnostic-test reporting guidance recommends reporting the underlying two-by-two results and confidence intervals rather than using the ambiguous word “accuracy” alone.

References#

1. Altman and Bland, Diagnostic tests 2: Predictive values
2. Cochrane Handbook for diagnostic test accuracy reviews
3. STARD 2015
4. Spectrum bias and diagnostic test evaluation
5. CDC principles of interpreting test results
6. FDA statistical guidance for diagnostic test studies

For your own health, talk with your clinician.*