The PSA test, by the numbers

Worth the Test · Educational analysis · Updated September 2026

A PSA result of 4.2 ng/mL lands on a lab report and feels like a verdict. The population arithmetic says otherwise. At the traditional 4.0 ng/mL cutoff, the test misses about four of every five prostate cancers that are actually present, yet when it does raise a flag it is right only about one time in six. Those two numbers fall out of the same short list of inputs and point in opposite directions — much of why PSA is the most argued-over test in cancer screening.

An unusual accuracy profile

Three figures drive everything downstream. Two describe the test; one describes the men being screened.

• Sensitivity ≈ 21% — of men who truly have prostate cancer, only about 21 in 100 have a PSA above the cutoff. The other ~79 are missed at this round.
• Specificity ≈ 94% — of men without cancer, about 94 in 100 are correctly cleared, so roughly 6 in 100 are flagged anyway.
• Illustrative prevalence = 5.4% — assume 54 of every 1,000 have prostate cancer. A study’s screen-detected yield is not the same as prevalence: it omits missed cancers and may use a different reference standard. Here 5.4% is a hypothetical input, not a validated combination with the PCPT accuracy estimates.

This is an unusual profile for a cancer screen. Most screens are tuned to miss as few cancers as possible and accept more false alarms in return; PSA at 4.0 ng/mL runs the other way, keeping false alarms modest but letting most cancers slip through. The misses are not confined to harmless disease — the Prostate Cancer Prevention Trial analysis behind these figures noted that a meaningful share of the tumors passing under the cutoff were high-grade, the aggressive cancers screening most wants to catch. Where a cutoff sits on that catch-versus-false-alarm curve is the subject of sensitivity vs. specificity; here we take the 4.0 ng/mL operating point as given and follow it through a population.

The 2×2, worked by hand

Screen 1,000 men at these three figures and every cell of the calculator's test view is fixed. Split the group first: 1,000 × 0.054 = 54 men with prostate cancer, and 946 without.

Now apply the test to each group. Of the 54 with cancer, 54 × 0.21 = 11.3, so about 11 are flagged (true positives) and about 43 are missed (false negatives). Of the 946 without cancer, 946 × 0.06 = 56.8, so about 57 are flagged anyway (false positives) and 889 are correctly cleared (true negatives).

Follow the positive column. About 68 men are told their PSA is elevated, and only about 11 of them actually have cancer. The share of positives that are real is the positive predictive value (PPV):

PPV = (0.21 × 0.054) ÷ [(0.21 × 0.054) + (0.06 × 0.946)] = 0.01134 ÷ 0.06810 ≈ 0.17 — about 1 in 6.

Computed from the exact rates before rounding to whole people — the way the calculator does it — that is 16.7%; dividing the rounded people, 11 ÷ 68, gives 16.2%. Same answer either way. Most elevated PSA results at this cutoff and prevalence are false alarms, which is why a raised PSA prompts a closer look — an MRI, a repeat measure, sometimes a biopsy — rather than a diagnosis.

A normal PSA is not an all-clear

The 21% sensitivity has a consequence that catches people off guard: a normal PSA barely changes the odds. Read the negative column. 932 men get a reassuring result, and 889 of them are truly cancer-free — a negative predictive value of 889 ÷ 932 ≈ 95%. That sounds strong until you compare it with where the men started. Before the test, a man in this group had a 5.4% chance of harboring cancer; after a normal PSA, he still has about a 4.6% chance (that is 1 − NPV). The 43 cancers sitting in that negative column — present but missed this round — are the whole difference.

The likelihood-ratio form makes the smallness explicit. A negative PSA carries a negative likelihood ratio of (1 − 0.21) ÷ 0.94 ≈ 0.84 — so close to 1 that it hardly budges the odds. As a rule of thumb, it takes a likelihood ratio below about 0.1 to confidently rule a condition out; 0.84 is nowhere near. So a normal PSA does not rule out prostate cancer, and because the missed tumors include high-grade ones, it does not even reliably rule out the dangerous kind — which is why PSA is one input to further evaluation, not a gatekeeper that clears men on its own.

Move the base rate, move the answer

The assay does not change from man to man, but the predictive value does — because the third number, prevalence, does most of the work. Take the same 21% / 94% PSA to a group with a 12% pre-test probability, used here only to show the mechanism, not as a figure for any specific age or risk band.

Now of 1,000 men, 120 have cancer and 880 do not. The test flags 120 × 0.21 ≈ 25 of the cancers and 880 × 0.06 ≈ 53 of the healthy — about 78 positives, of which 25 are real:

PPV = (0.21 × 0.12) ÷ [(0.21 × 0.12) + (0.06 × 0.88)] = 0.0252 ÷ 0.0780 ≈ 0.32 — about 1 in 3.

The identical elevated PSA that meant roughly 1-in-6 odds at illustrative prevalence now means roughly 1-in-3. Nothing about the test changed; only the population did. That single lever — pre-test probability in, post-test probability out — is the engine behind every screening result, and why a figure from a validation study cannot tell a man what his own positive means until his baseline risk is in the picture.

Benefit and harm on one page

None of the 2×2 arithmetic tells you whether screening is worth doing — that turns on what happens after the flag, and here the evidence is genuinely mixed. This site takes no side.

On the benefit side, the European Randomized Study of Screening for Prostate Cancer (ERSPC), at 16 years of follow-up, is summarized by the National Cancer Institute as needing roughly 18 cancers detected and treated to prevent one prostate-cancer death (a number-needed-to-diagnose). On the harm side sits the dominant, well-documented cost of PSA screening: overdiagnosis and overtreatment. Because many screen-detected prostate cancers grow too slowly to ever cause symptoms, a large share of the men who go on to surgery or radiation — carrying real risks of incontinence and impotence — could not have benefited, because their cancer would never have surfaced. The ERSPC Rotterdam modeling by Heijnsdijk and colleagues, a separate screening-program analysis, quantified this overtreatment directly.

The ERSPC number needed to diagnose is not a treatment NNT. It compares additional diagnoses and deaths prevented between screening arms over follow-up; it does not estimate benefit per treated true positive in this single-round model. The linked scenarios therefore use generic treatment inputs, and no trial-derived PSA treatment benefit or harm is claimed. A false positive, an overdiagnosed true cancer, and a treatment complication must also remain distinct.

Guideline bodies read this ledger differently for different men. The U.S. Preventive Services Task Force (2018) frames PSA screening for men aged 55 to 69 as an individual decision (Grade C) and recommends against it for men 70 and older (Grade D) — a split that is itself an admission that the benefit-harm balance depends on the numbers you put in.

What the numbers settle — and what they don't

The population arithmetic settles a few things cleanly: at a 4.0 ng/mL cutoff and illustrative prevalence, most elevated PSA results are false alarms, a normal PSA leaves meaningful residual risk, and the predictive value of any result rises and falls with prevalence rather than the assay. What it cannot settle is the value judgment underneath — how to weigh one prostate-cancer death deferred against many men worked up, biopsied, or treated for disease that might never have surfaced. That is a question about stakes and preferences, not a formula.

The mechanics here recur across every screen. Why a relatively uncommon condition makes most positives false is the base-rate fallacy; why a low-sensitivity threshold cannot rule disease out is the sensitivity-specificity trade-off; and the false positives, overdiagnosis, and overtreatment that follow a positive screen are quantified in screening harms and biases. PSA is not an unusually bad test — it is an unusually honest illustration of how much a single cutoff decides.