Bayes' Theorem and Diagnostic Testing

A diagnostic test sorts people into "positive" and "negative," but this sorting is imperfect in two independent ways. Sensitivity is the probability that a person who truly has the disease (D+) tests positive:

Specificity is the probability that a person who truly does not have the disease (D−) tests negative:

These are properties of the assay and the disease biology—how well an antibody binds, how a tumor's density differs from surrounding tissue, how a genetic marker segregates with a mutation. Critically, sensitivity and specificity do not depend on how common the disease is in the group being tested; they are conditional probabilities defined within the disease-status groups, measured against a gold-standard confirmatory test (biopsy, PCR, karyotype, autopsy). What clinicians and patients actually want is the reverse conditional probability: given a positive result, what is the probability of disease? This is the positive predictive value (PPV), $P(D+ \mid T+)$, and its negative counterpart the negative predictive value (NPV), $P(D- \mid T-)$. Unlike sensitivity and specificity, PPV and NPV depend heavily on prevalence, $P(D+)$, the baseline probability of disease in the population before any test is run. Prevalence plays the role of the prior probability in Bayesian language; the test result is the evidence; PPV is the posterior probability. Bayes' theorem is precisely the algebra that inverts a conditional probability:

where the denominator, the total probability of testing positive, expands over both disease states:

Substituting sensitivity, prevalence, and the false-positive rate $1 - \text{specificity}$ turns this into a fully computable formula. This bridges Conditional Probability—which defines $P(A\mid B)$ as a ratio of probabilities—into a genuinely predictive clinical tool, and it sets up the language of prior and posterior distributions that reappears once test results are treated as observations of a random variable.

Step 1: Start with a population and its true disease status. Imagine screening 10,000 people. Prevalence $p$ tells you how many truly have the disease: $10{,}000 \times p$ are D+, and $10{,}000 \times (1-p)$ are D−. This partition happens before any test is applied — it is nature's ground truth, established by a gold standard (biopsy, culture, genetic sequencing, or long-term clinical follow-up).

Step 2: Apply sensitivity to the diseased group. Of the D+ individuals, a fraction equal to sensitivity will test positive (true positives, TP); the remaining fraction, $1 - \text{sensitivity}$, are missed — these are false negatives (FN). Sensitivity is sometimes called the true positive rate.

Step 3: Apply specificity to the healthy group. Of the D− individuals, a fraction equal to specificity will correctly test negative (true negatives, TN); the remaining fraction, $1 - \text{specificity}$, will test positive despite being healthy — false positives (FP). $1 - \text{specificity}$ is the false positive rate.

Step 4: Build the 2×2 table. Four cells now exist: TP, FP, FN, TN. Every predictive value is just a ratio of these cell counts:

Step 5: Read off predictive values. PPV answers "of everyone who tested positive, what fraction are truly sick?" NPV answers "of everyone who tested negative, what fraction are truly healthy?" These are the numbers a clinician communicates to a patient, and they are the numbers that change dramatically as prevalence moves — even while sensitivity and specificity stay fixed, because they are architectural properties of the test itself.

Step 6: Recognize the amplifying role of low prevalence. When a disease is rare, the pool of healthy people is enormous compared to the pool of sick people. Even a small false-positive rate applied to a huge healthy population generates an absolute number of false positives that can dwarf the true positives drawn from a small diseased population. This is the mathematical heart of the "base-rate trap."

Notice in Figure 1 that despite 90% sensitivity and 95% specificity — numbers that sound excellent — the PPV is only about 15%. Roughly 5 out of every 6 people who test positive do not have the disease, purely because it is rare. This is not a failure of the test; it is an inescapable consequence of applying any imperfect test to a low-prevalence population.

Figure 2 shows the key qualitative fact: with sensitivity and specificity held perfectly constant, PPV climbs from about 15% to about 95% simply by moving prevalence from 1% to 50%. The test never changed. Only the population's baseline risk did. This is why the same test is used differently in general screening (low prevalence, expect many false positives, confirm before treating) versus confirmatory testing in a high-risk or symptomatic population (higher prevalence, PPV rises sharply).

Bayes' theorem itself is 18th-century mathematics (Thomas Bayes, published posthumously in 1763, generalized by Pierre-Simon Laplace). What establishes it as essential to medicine is a body of empirical psychology and clinical epidemiology showing that trained professionals reliably fail to apply it correctly — and that the failure has real diagnostic consequences.

Casscells, Schoenberger, and Grayboys (1978), published in the New England Journal of Medicine, posed this exact problem to 60 students, staff, and faculty at Harvard Medical School: "If a test to detect a disease whose prevalence is 1/1000 has a false positive rate of 5%, what is the chance that a person found to have a positive result actually has the disease, assuming you know nothing about the person's symptoms or signs?" The correct Bayesian answer is approximately 2% (essentially: 1 true positive per 1000 tested, against roughly 50 false positives from the 999 healthy people). The most common answer given by the Harvard-affiliated respondents was 95% — they anchored on the complement of the false-positive rate and ignored prevalence almost entirely. Only about 18% of respondents gave an answer close to the correct 2%. This was not a trick question aimed at laypeople; it was posed to people who order and interpret diagnostic tests professionally.

David Eddy (1982) extended this finding directly to breast cancer screening. He asked physicians to estimate the probability that a woman with a positive mammogram, and no other risk factors, actually had breast cancer, given realistic clinical parameters (prevalence around 1%, sensitivity around 79-92%, false-positive rate around 9-10% depending on the study cited). The correct Bayesian PPV works out to roughly 7-10%. Eddy found that a majority of physicians surveyed estimated the probability at 70-80%, overestimating by nearly an order of magnitude. Eddy's work, and later replications, directly influenced how radiology and primary care communicate mammography results and helped motivate formal Bayesian training in evidence-based medicine curricula.

Gigerenzer and Hoffrage (1995) showed that the errors documented by Casscells and Eddy were not really failures of logic but failures of representation. When the identical problem was reframed using natural frequencies — "10 out of 1000 women have breast cancer; of these 10, 8 will have a positive mammogram; of the 990 without breast cancer, about 95 will still have a positive mammogram" — rather than abstract conditional probability percentages, the proportion of physicians and lay subjects who computed the correct PPV rose sharply, in some studies more than doubling. This is exactly the "natural frequency" method embodied in the 2×2 table used in this article: converting probabilities to whole-number counts out of a fixed round population makes the Bayesian inversion almost mechanical rather than a source of algebraic dread. This finding reshaped how medical statistics is taught and is why counts-out-of-10,000 tables, not raw formulas, are now standard pedagogy. Together these three lines of evidence establish two things: (1) Bayes' theorem is the mathematically correct way to compute predictive values from sensitivity, specificity, and prevalence — this is a theorem, not an empirical claim, and follows deductively from the axioms of probability; and (2) unaided human intuition, including expert clinical intuition, systematically and severely violates it in the specific direction of over-trusting positive results when prevalence is low — this is the empirical, tested claim, replicated across decades and populations.

Example 1 — HIV screening in a low-risk population (asymmetric, low prevalence). A community clinic uses a rapid antibody test with:

Screen 100,000 people. True D+ = $100{,}000 \times 0.003 = 300$. True D− = $99{,}700$. TP $= 300 \times 0.997 = 299.1 \approx 299$ FN $= 300 - 299 = 1$ FP $= 99{,}700 \times (1 - 0.985) = 99{,}700 \times 0.015 = 1495.5 \approx 1496$ TN $= 99{,}700 - 1496 = 98{,}204$

Interpretation: a negative result is essentially conclusive (NPV ≈ 100%) — reassuring for ruling out infection in a low-risk person. But a positive rapid test in this population is correct only about 1 time in 6. This is precisely why CDC and WHO protocols require a confirmatory second test (e.g., a Western blot or a second, differently-targeted assay) before diagnosing HIV — a single positive screening result in a low-prevalence population is not enough on its own. Running a second, statistically independent confirmatory test effectively treats the first test's positive result as a new, sharply elevated "prevalence" (16.7% instead of 0.3%) fed into a second Bayesian update — the posterior from test one becomes the prior for test two — and PPV after two independent positives becomes very high.

Example 2 — Prenatal aneuploidy screening (moderate prevalence, weighing false negatives against false positives). A non-invasive prenatal test (cell-free fetal DNA, cfDNA) for Trisomy 21 is being evaluated for a 38-year-old patient, where age-adjusted prevalence of Trisomy 21 at this maternal age is about 1 in 150 pregnancies (≈0.67%). Suppose the test has:

Per 15,000 pregnancies at this age: True D+ $= 15{,}000 \times 0.00667 = 100$ True D− $= 14{,}900$ TP $= 100 \times 0.993 = 99.3 \approx 99$ FN $= 100 - 99 = 1$ FP $= 14{,}900 \times (1-0.998) = 14{,}900 \times 0.002 = 29.8 \approx 30$ TN $= 14{,}900 - 30 = 14{,}870$

Interpretation: because both sensitivity and specificity are very high and prevalence is not vanishingly small (compare to Example 1's 0.3%), PPV jumps to roughly 77% — high enough to counsel strongly toward confirmatory diagnostic testing (chorionic villus sampling or amniocentesis, which carry a small procedural miscarriage risk but give a karyotype-level definitive answer) rather than treating the screen alone as diagnostic. Compare this to the same test applied to a 25-year-old with baseline prevalence around 1/1200 (≈0.083%): repeating the arithmetic drops PPV to roughly 29% even though sensitivity and specificity are unchanged — the same screening technology, applied to a lower age-related prior probability, yields a starkly less trustworthy positive result. This is exactly why prenatal screening guidelines report age-specific or risk-adjusted predictive values rather than a single blanket number.

Thomas Bayes (c. 1701–1761) was an English Presbyterian minister and amateur mathematician, a Fellow of the Royal Society, who left almost no trace of mathematical publication during his lifetime. His now-famous essay, "An Essay towards solving a Problem in the Doctrine of Chances," was found among his papers after his death and communicated to the Royal Society in 1763 by his friend Richard Price, who edited and published it. Bayes's original problem concerned inferring the probability of an unknown event from repeated trials — closer to inverse probability for a binomial process than to medical testing — and it was Pierre-Simon Laplace, working independently and more thoroughly a few decades later, who developed the general form of the theorem and extended it into a systematic tool for inference. The application to diagnostic testing specifically — sensitivity, specificity, predictive values — is a 20th-century clinical epidemiology development, crystallized in the mid-1900s as biostatistics became formalized in medical training, long after Bayes's death.

Screening program design. Public health agencies use predictive-value calculations to decide whether mass screening (e.g., mammography, PSA testing, colorectal cancer screening) is net beneficial in a given age group, because low prevalence in younger, lower-risk groups can make PPV so low that screening causes more harm (biopsies, anxiety, overdiagnosis) than benefit — this is precisely why screening guidelines are age- and risk-stratified rather than universal.

COVID-19 and infectious disease testing. During periods of low community transmission, rapid antigen tests with high specificity but imperfect sensitivity produced very different predictive values than during outbreak surges; public health messaging around "test positivity" required exactly this Bayesian correction, and misunderstanding it fueled both under- and over-reaction to testing data.

Newborn and prenatal screening. Panels for metabolic disorders (e.g., phenylketonuria) and aneuploidy screening (as in Worked Example 2) are tuned with deliberately high sensitivity, accepting a lower PPV, because a false negative (a missed treatable metabolic disease) is far costlier than a false positive that gets ruled out on confirmatory testing.

Genetic counseling and hereditary cancer risk. BRCA1/2 and other hereditary cancer panel results are interpreted through family-history-adjusted prior probabilities before predictive values are quoted to patients — the same architecture as maternal-age-adjusted prenatal risk.

Forensic and legal contexts. DNA match probabilities, presented without base-rate context (the "prosecutor's fallacy"), are a direct legal analogue of the PPV trap and have led to documented wrongful convictions when courts conflated $P(\text{match} \mid \text{innocent})$ with $P(\text{innocent} \mid \text{match})$.

Machine learning and AI diagnostics. Algorithms trained on imbalanced medical imaging datasets face this exact problem: a classifier with 95% sensitivity and 95% specificity for a rare tumor subtype will still generate a flood of false positives when deployed at population scale, and engineers must report calibrated predictive values, not just headline accuracy, before clinical deployment.

Ask the tutor: If a screening test's sensitivity and specificity are both 90%, at what prevalence does PPV cross 50%, and why does that threshold matter for deciding whether to treat on a positive result alone?