COVID rapid antigen tests, by the numbers

Written and reviewed by a practicing emergency physician · Updated August 2026

The COVID-19 rapid antigen test is the only diagnostic test most people have ever run on themselves, read themselves, and acted on themselves. That makes it the best teaching example in this library: everyone has stared at that little window wondering what one faint line — or its absence — actually proves. The honest answer depends on three numbers, and one of them is about you, not the test.

The accuracy numbers, and the twist

The definitive accuracy summary is a Cochrane systematic review by Dinnes and colleagues, pooling 152 evaluations across more than 100,000 samples. Its headline figures:

• Sensitivity ≈ 73% in people with symptoms — during symptomatic infection, roughly 27 in 100 infected people still test negative on a single rapid test.
• Sensitivity ≈ 55% in people without symptoms — used for asymptomatic screening, the same kind of test misses almost half of infections.
• Specificity ≈ 99% or better in both groups (99.1% symptomatic, 99.7% asymptomatic) — false positives are genuinely rare.

The twist is that first pair. "How accurate is a rapid test?" has no single answer, because sensitivity is not a fixed property of the plastic cassette — it depends on who is being tested. Symptomatic people tend to be tested at peak viral load, when antigen is abundant; asymptomatic people include many caught early or late in infection, with less antigen to find. This is spectrum effect: the same test, measured in two populations, earns two different report cards. It is the sharpest everyday example of a caveat that applies to every test on this site — a published sensitivity travels with the population it was measured in. (The sensitivity-and-specificity guide unpacks this.)

Worked example 1 — symptomatic, during a wave

Say you have a sore throat and fever during a winter wave, when a substantial share of similar symptomatic people — call it 20%, an illustrative figure, not a measured one — actually have COVID. Test 1,000 such people with a 73% / 99.1% test. First the truth: 200 infected, 800 not.

Of the 200 infected, 200 × 0.73 = 146 test positive; 54 test negative anyway. Of the 800 not infected, 800 × 0.009 ≈ 7 false positives; 793 correctly negative.

Two conclusions fall out, pointing in opposite directions:

A positive is near-decisive: PPV = 146 ÷ 153 ≈ 95%. With symptoms, in a wave, a line on the test settles it for practical purposes.

A negative settles much less: 54 of the 847 negatives — about 1 in 16 — are actually infected. Starting from 20%, one negative test only drops you to about 6%.

That residual 6% is why a single negative rapid test, taken while you're actively symptomatic, was never treated as an all-clear. The test's LR− is (1 − 0.73) ÷ 0.991 ≈ 0.27 — a real push, but nowhere near the below-0.1 territory that confidently rules things out. The standard remedies are the two classic moves of pre-test/post-test reasoning: apply a stronger test (a lab NAAT — the Cochrane authors frame antigen tests as triage for exactly this reason), or apply the same weak test again after the odds have had time to shift — antigen rises as an infection develops, so tomorrow's test is not just a coin re-flip.

Worked example 2 — no symptoms, quiet week

Now the same physical act — swab, swirl, wait — before visiting a relative, with no symptoms and no known exposure, in a quiet week where perhaps 1 in 200 similar people (0.5%, again illustrative) is infectious. The applicable numbers are now 55% sensitivity and 99.7% specificity, and the base rate has collapsed by a factor of forty. Run 1,000 such people: 5 infected, 995 not. The test catches 5 × 0.547 ≈ 3 of the infected and falsely flags 995 × 0.003 ≈ 3 of the healthy.

PPV = (0.547 × 0.005) ÷ [(0.547 × 0.005) + (0.003 × 0.995)] = 0.00274 ÷ 0.00572 ≈ 0.48 — a positive is a coin flip.

Sit with that: a test with 99.7% specificity — a false-positive rate of three per thousand — still produces positives that are only about half real, because at a 0.5% base rate the healthy crowd outnumbers the infected two hundred to one. This is the base-rate fallacy in its purest consumer form, and it is the same structure as the HIV screening paradox, where an even better test (LR+ ≈ 200) yields positives that are barely better than even odds at population prevalence. The rapid test's LR+ here is 0.547 ÷ 0.003 ≈ 180 — comparable pull — and the base rate eats it just the same.

The negative column, meanwhile, is quietly excellent in this scenario — residual risk about 0.2%, not because the test ruled infection out (at 55% sensitivity it barely argued) but because there was almost nothing to rule out. When the base rate is low enough, a negative result is mostly confirming what was already true.

One test, two verdicts

Put the two worked examples side by side and the deepest lesson on this site falls out: the same cassette, the same chemistry, the same faint pink line meant "you almost certainly have COVID" in one situation and "even odds" in the other. Nothing about the test changed. What changed was the pre-test probability — symptoms, exposure, and how much virus was circulating that week. A test result is not a verdict; it is an update to whatever the odds were before you swabbed. That is the entire content of Bayes' theorem, and the rapid test put a worked example of it in everyone's bathroom cabinet.

It also explains the era's practical rituals. Repeating a negative test a day or two later, confirming a surprising positive with a lab test, trusting a positive more in January than in June — each is an informal Bayesian move: stack a second likelihood ratio, or re-read the same result against a different base rate. The serial-testing guide works through what repetition does to both error rates, in both directions.