Paperonium

← Back to the plain-language case

All reported HPV prevalence confidence intervals are about 20 to 140 times narrower than the counts allow

BMC Med (2025) · doi:10.1186/s12916-025-04246-0 · PMC12211446

Summary

The Abstract, Results and Table 1 report a cervicovaginal HPV prevalence of 6.2% (51 of 822 women) with a 95% confidence interval of 6.18%–6.22%, and give intervals of ±0.01 to ±0.02 percentage points for every genotype-specific prevalence. From the paper's own counts, the exact binomial 95% interval for 51/822 is 4.65%–8.08%, 86 times as wide as the reported one, and the genotype intervals are 22 to 143 times as wide as reported; the point estimates themselves are correct. Four of the six reported intervals are reproduced to the printed decimal by one arithmetic step, a 95% half-width computed on the proportion scale (0.0165 for 51/822) added to a prevalence expressed in percent. Prevalence is the study's primary outcome, so its stated precision is overstated about 80-fold, which bears on the Discussion's comparison with 2009 estimates (the 2009 HPV16/18 value of 1.2% lies inside the corrected interval) and on any meta-analysis that uses the published interval.

What the paper reports

The prevalence of HPV infection was 6.2% (95% CI: 6.18%-6.22%), while the prevalence of high-risk genotypes 16 and 18 was 1.94% (95% CI: 1.93%-1.95%). The prevalence of 12 pooled high-risk HPV infection was 4.14% (95% CI:4.13%-4.15%).

Abstract, Results

The prevalence of cervicovaginal HPV infection among ever-married women of the 35-year-old age cohort was 6.2% (95% CI: 6.18%-6.22%), as 51, had tested positive for HPV infection out of 822 participants. The prevalence of high-risk carcinogenic HPV genotypes 16(n = 14) and 18(n = 2) was 1.94% (95% CI: 1.93%-1.95%), while the other 12 pooled high-risk HPV genotypes (n = 34) was 4.14% (95% CI: 4.13%-4.15%) (Table 1).

Results, paragraph 2

Table 1 ("Prevalence of cervico-vaginal HPV infection in the District") has a column headed "95% CI for the prevalence %". Its rows are: 12 pooled positive, 34, 4.14, 4.13-4.15; 16 positive, 14, 1.7, 1.69-1.71; 18 positive, 2, 0.24, 0.23-0.25; 16 & 12 pooled positive, 01, 0.12, 0.1-0.13; Negative, 771, 93.8; Total, 822.

On the design and analysis, the paper says:

In the district, there are 413 PHM areas.

Methods, paragraph 3 (each Public Health Midwife area is a cluster, and three women were selected from each)

the degree of accuracy desired specified was 0.016 (3.3/100 × 2). Therefore we needed 479 women.

Methods, paragraph 2 (the required sample size was then "multiplied by the design effect, which was taken as 1.1")

Data were entered by using SPSS version 20.

Methods, paragraph 8

Moreover, high-risk genotypes 16 and 18 prevalence were elevated from 1.2% to 1.94% within the same duration.

Discussion, paragraph 2

What the paper's own numbers show

Method. Each prevalence is a count out of n = 822. For each count I computed the 95% interval three ways: the Clopper–Pearson exact interval, the Wilson interval, and the Wald interval p ± 1.96·√(p(1−p)/n). I also computed the Wald interval with its variance inflated by the paper's stated design effect of 1.1. I then tested one candidate arithmetic route: take 1.96·SE on the proportion scale and add it, unconverted, to the prevalence in percent. All numbers come from recompute.py.txt.

The counts are internally consistent: 34 + 14 + 2 + 1 = 51 positives, 51 + 771 = 822, and 14 + 2 = 16 for HPV16 or 18. Every reported point estimate equals its count divided by 822 (for 16/822 = 1.946% the paper prints 1.94%, truncated rather than rounded, which does not matter here).

Quantity Location Reported Recomputed from the paper's counts
Any HPV, 51/822 Abstract; Results p2 6.2% (95% CI 6.18–6.22) 6.204%; exact 4.65–8.08; Wilson 4.75–8.07; Wald 4.56–7.85; Wald with design effect 1.1 4.47–7.93
HPV16 or 18, 16/822 Abstract; Results p2 1.94% (1.93–1.95) 1.946%; exact 1.12–3.14; Wilson 1.20–3.14
12 pooled high-risk types, 34/822 Abstract; Results p2; Table 1 4.14% (4.13–4.15) 4.136%; exact 2.88–5.73; Wilson 2.97–5.72
HPV16, 14/822 Table 1 1.7% (1.69–1.71) 1.703%; exact 0.93–2.84; Wilson 1.02–2.84
HPV18, 2/822 Table 1 0.24% (0.23–0.25) 0.243%; exact 0.03–0.88; Wilson 0.07–0.88
HPV16 & 12 pooled, 1/822 Table 1 0.12% (0.1–0.13) 0.122%; exact 0.00–0.68; Wilson 0.02–0.69
Half-width of the overall interval derived 0.02 points 1.65 points (Wald), 82 times the reported half-width; the exact interval is 86 times as wide as the reported one
How much too narrow, genotype rows (exact width ÷ reported width) derived — HPV16 or 18: 101×; 12 pooled: 143×; HPV16: 95×; HPV18: 42×; HPV16 & 12 pooled: 22×
Sample size a ±0.02-point interval would need at 6.2% derived — about 5.59 million
Population size at which a finite-population correction would give ±0.02 points derived — 822.1, i.e. the sample itself
Percent prevalence ± 1.96·SE in proportion units, centred on the reported rounded prevalence derived 6.18–6.22; 1.93–1.95; 4.13–4.15; 1.69–1.71; 0.23–0.25; 0.1–0.13 6.18–6.22; 1.93–1.95; 4.13–4.15; 1.69–1.71 (all four exact matches); 0.24–0.24; 0.12–0.12
Sample-size calculation Methods p2 479; 915 1.96²·0.033·0.967/0.016² = 478.9; 479·1.1/(1−0.424) = 914.8 (consistent)
Planned precision vs achieved Methods p2 (derived) ±1.6 points (d = 0.016) at an assumed 3.3% The 95% half-width at the observed 6.2% with n = 822 is 1.65 points, close to the planned ±1.6 points and about 80 times the reported 0.02

The paper's own sample-size calculation targeted a precision of ±1.6 percentage points (d = 0.016). The correct half-width at the observed prevalence, 1.65 points, is close to the planned ±1.6 points (the plan assumed 3.3% and n = 479, so the two need not coincide). The reported interval is roughly 80 times narrower than the precision the study was designed to achieve.

The arithmetic route above reproduces four of the six reported intervals exactly: 6.18–6.22, 1.93–1.95, 4.13–4.15 and 1.69–1.71. For the two smallest counts (2 and 1 positives), that route gives half-widths of 0.0034 and 0.0024 points. The paper reports 0.23–0.25 and 0.1–0.13, half-widths of 0.010 and 0.015 points, about 3 and 6 times what that route gives, so the route does not reproduce these two rows. They are still 42 and 22 times narrower than the exact intervals.

(An aside that does not affect the finding: the parenthetical "3.3/100 × 2" evaluates to 0.066, not 0.016. The stated 479 follows from d = 0.016, which is close to 3.3/100 ÷ 2 = 0.0165.)

Figure

Reported vs exact 95% intervals for each prevalence in the paper

Figure 1. Each row is one prevalence reported in the Abstract, the Results or Table 1, with its count out of 822. Orange diamonds show the paper's reported estimate and 95% CI; every reported interval is narrower than its marker, so its bar is hidden. Blue circles with capped bars show the exact (Clopper–Pearson) 95% CI computed from the same count. The labels to the right give each interval, marked "exact" or "reported". The dashed gray lines mark the 2009 values that the Discussion compares against (3.3%, 2.1% and 1.2%). The 1.2% value for HPV16 or 18 lies inside the exact interval for this study. The axis starts at zero and is linear.

Arithmetic chain for the overall interval

Figure 2. Arithmetic for the overall interval. Start from 51/822, compute the standard error and the 95% half-width (0.0165 on the proportion scale), then express the interval in percent. Multiplying the half-width by 100 gives 6.20% ± 1.65 points (blue, 4.56%–7.85% by the Wald method). Adding 0.0165 unconverted to the printed 6.2% gives 6.1835%–6.2165%, which rounds to 6.18%–6.22% (orange), the reported interval to the second decimal place. The same step reproduces the reported intervals for HPV16 or 18, the 12 pooled types and HPV16 (Table above).

Why it matters

Explanations considered

Status

As of 2026-10-23:

Suggested correction

Replace the prevalence intervals with intervals computed from the counts. Exact (Clopper–Pearson) 95% CIs, which a design effect of 1.1 would widen slightly, are:

Apply these in the Abstract, Results paragraph 2 and Table 1. If the authors prefer a design-based interval, the Wald interval with the stated design effect for the overall prevalence is 4.47%–7.93%. The Discussion's statement that HPV16/18 prevalence rose from 1.2% to 1.94% should note that 1.2% lies within the corrected interval for this study. The same intervals in the Abstract, Results and Table 4 of the 2021 BMC Public Health article from this survey (doi:10.1186/s12889-021-10190-4) would need the same correction.