Sensitivity vs. specificity: what they really mean

Worth the Test · Written by an emergency physician · Last reviewed September 2026

Two cancer-screening tests, the same job — find disease early — and opposite personalities. A PSA blood test at the usual 4 ng/mL cutoff, run on men who truly have prostate cancer, flags only about one in five of them; screening mammography catches nearly nine in ten women who truly have breast cancer. Yet PSA almost never cries wolf on a healthy man, while mammography raises a false alarm in more than one healthy woman in ten. Two numbers — sensitivity and specificity — capture that entire difference in temperament. And here is the catch that trips up patients and clinicians alike: neither number answers the question you actually care about once your own result comes back.

What each number measures

Every screening test sorts people into positive and negative, and the truth sorts them into has-the-condition and doesn't. Cross those and you get a 2×2 table: true positives, false negatives, false positives, true negatives. Sensitivity and specificity each read down one column of that table, conditioned on the truth.

Sensitivity = P(test positive | disease). Among the people who genuinely have the condition, it is the fraction the test catches — the catch rate. Its complement, 1 − sensitivity, is the miss rate: the share of real cases sent home falsely reassured (the false negatives).

Specificity = P(test negative | no disease). Among the people who genuinely do not have the condition, it is the fraction the test correctly clears — the clear rate. Its complement, 1 − specificity, is the false-alarm rate: healthy people flagged for a disease they don't have (the false positives).

Because both are computed within a truth column, they don't depend on how common the disease is. Move the same test from a high-risk clinic to a low-risk screening line and its sensitivity and specificity stay essentially put; what changes is who ends up in each column. (They can drift with the spectrum of disease a population carries, but they do not track prevalence the way predictive value does.) That column-wise view is exactly what the calculator's 2×2 confusion matrix draws.

SnNout and SpPin: what the mnemonics actually say

Two bedside sayings compress how to use these numbers, and they reward understanding over memorizing.

SnNout — a highly Sn (sensitive) test, when Negative, rules a diagnosis out. If a test catches almost everyone with the disease, it has very few false negatives, so a negative result is unlikely to be one of those rare misses. The reassurance of a negative rests entirely on that small miss rate.

SpPin — a highly Sp (specific) test, when Positive, rules a diagnosis in. If a test rarely flags healthy people, it has very few false positives, so a positive result is unlikely to be a false alarm. The confidence of a positive rests on that small false-alarm rate.

Both are heuristics, not guarantees. SpPin quietly assumes the true positives aren't swamped by false positives — an assumption that breaks when the disease is rare, because even a whisper of a false-alarm rate applied to a huge healthy group can outnumber every true case. SnNout leans on sensitivity sitting close to 100%: a merely 90%-sensitive test still misses one case in ten, which is not the same as ruling anything out. The rare-disease failure of SpPin is the base-rate fallacy, and it is why "the test is 94% specific" is not the same statement as "a positive means you have it."

A worked contrast: two real tests

Put concrete numbers to it. Screening mammography, in the Breast Cancer Surveillance Consortium benchmarks, runs about 86.9% sensitive and 88.9% specific (Lehman 2017). The PSA blood test at a 4 ng/mL cutoff runs about 21% sensitive and 94% specific (Prostate Cancer Prevention Trial; Thompson 2005). Read each as two separate thought experiments — one on a thousand people who have the disease, one on a thousand who don't.

For mammography, out of 1,000 women who truly have breast cancer, it flags about 869 (true positives) and misses about 131 (false negatives). Out of 1,000 women who are truly cancer-free, it correctly clears about 889 (true negatives) and false-alarms on about 111 (false positives).

For PSA at 4 ng/mL, out of 1,000 men who truly have prostate cancer, it flags only about 210 and misses 790. Out of 1,000 cancer-free men, it clears 940 and false-alarms on just 60.

The personalities are now explicit. PSA at this cutoff is the quiet test: it rarely disturbs a healthy man, but it sleeps through roughly four of every five real cancers. Mammography is more even-handed — far better at catching disease, at the price of more than one false alarm in ten healthy women. In a screening line, which is overwhelmingly made of people without the disease, specificity governs the sheer volume of false alarms (11% of a very large healthy group is a lot of callbacks), while sensitivity governs how many true cases slip through the much smaller diseased group. Those are different harms landing on different people — the subject of screening harms & biases.

The threshold trade-off and the ROC curve

Why can't a test simply be high on both? Because sensitivity and specificity usually sit on opposite ends of the same seesaw. Most "positive/negative" tests are really a continuous measurement — nanograms of PSA per milliliter, the optical density of an antibody assay, a radiologist's suspicion score — forced into a yes/no by a cutoff. Move the cutoff and both numbers move, in opposite directions.

PSA makes the trade-off vivid. At the 4 ng/mL cutoff it is about 21% sensitive and 94% specific. Ease the cutoff to 3.1 ng/mL and sensitivity rises to about 32% while specificity slips to roughly 87%; drop it near 1.1 ng/mL and sensitivity climbs above 80% — but specificity collapses to about 39%, flagging most healthy men (Thompson 2005). Raising the bar does the reverse. As the trial's investigators put it, no single cutoff delivers high sensitivity and high specificity at the same time.

Sweep the cutoff across every possible value and plot sensitivity (vertical) against the false-alarm rate, 1 − specificity (horizontal), and you trace the receiver operating characteristic (ROC) curve (Hajian-Tilaki 2013). Each point on it is one cutoff's sensitivity/specificity pair. A more discriminating test bows toward the top-left corner; a coin-flip of a test hugs the diagonal, where every gain in catch rate costs an equal rise in false alarms. The area under the curve sums up the test's discrimination across all cutoffs in one threshold-independent number. Choosing the actual cutoff, though, is a values decision, not a statistical one: a judgment about whether missing a cancer or over-investigating a healthy person is the error you least want to make.

The number the test can't give you

Neither sensitivity nor specificity answers the only question a patient truly asks: I tested positive — do I actually have it? That is the positive predictive value (PPV), and it needs one ingredient the test characteristics leave out — how common the condition is to begin with. Only now does prevalence enter, and with it the full 2×2 table fills in.

Take the four PSA cutoffs the trial's investigators reported — 1.1, 2.1, 3.1 and 4.1 ng/mL — and apply each one to the same 1,000 men at an illustrative 5.4% prevalence, the hypothetical input used on the PSA page rather than a measured rate for any particular group of men. At every cutoff, 54 men have prostate cancer and 946 do not; only the test's operating point changes.

Four PSA cutoffs applied to 1,000 men at an illustrative 5.4% prevalence (54 with prostate cancer, 946 without). Sensitivity and specificity for any prostate cancer from Thompson 2005. Counts rounded to whole people; PPV and NPV computed from the unrounded rates. On a narrow screen, scroll the table sideways.

Work the 2.1 ng/mL row by hand. Of the 54 men with cancer, 54 × 0.526 = 28.4 are flagged (true positives) and 25.6 are missed. Of the 946 without cancer, 946 × (1 − 0.725) = 946 × 0.275 = 260.2 are flagged anyway (false positives) and 685.9 are correctly cleared. That makes 28.4 + 260.2 = 288.6 positive results, of which 28.4 are real: PPV = 28.4 ÷ 288.6 ≈ 9.8%, about 1 in 10. On the other side, 685.9 of the 25.6 + 685.9 = 711.4 negative results are truly cancer-free: NPV ≈ 96.4%.

Now read down the table. Lowering the cutoff from 4.1 to 1.1 ng/mL flags about four times as many of the cancers (11 → 45), but the false positives grow nearly tenfold (59 → 578). The imbalance is built into the group sizes: each percentage point of specificity given up flags about 9.5 more of the 946 healthy men, while each point of sensitivity gained catches only about half a cancer more among the 54. So the positive column floods with false alarms — PPV sinks from about 16% to about 7% — while NPV edges up only from 95.4% to 97.6%. A lower cutoff buys fewer missed cancers with many more false alarms; which of those errors matters more is the values judgment from the section above. (The 4.1 ng/mL row is the operating point this site rounds to 21% / 94%; with those rounded figures the PPV is 16.7% rather than 15.9%.)

Notice the switch in denominator: sensitivity and specificity divide by the truth (true positives out of the 54 men with cancer, true negatives out of the 946 without), while PPV divides by the test result (true positives out of everyone flagged).

That gap between "how good the test is" and "what my result means" is the whole reason sensitivity and specificity are only the starting point. To turn them into the answer you want, follow the handoff to the base-rate fallacy and to pre-test to post-test probability, where the same positive result is filtered through prevalence — and see likelihood ratios for the one summary of a test that carries between populations unchanged.

Correction, September 2026. This guide described 5.4% as a measured first-round prostate-cancer prevalence; its source reports a detection rate, so the figure is now labelled illustrative. The PSA accuracy figures are now cited to the trial report (Thompson 2005), and the worked example now shows how the cutoff trades sensitivity against specificity. The corrections log lists every change.