Paperonium

← Back to the plain-language case

Abstract effect sizes for crowding, ageing and primary care coverage disagree with the paper's Table 2

Lancet Reg Health Am (2026) Β· doi:10.1016/j.lana.2026.101572 Β· PMC13382474

Summary

The abstract reports three effects from "the final spatial GAMLSS model", each per one percentage-point increase. It gives +3.25% expected tuberculosis (TB) mortality for household crowding, +5.92% for the population aged 60 or older, and βˆ’0.30% for primary health care (PHC) coverage. The paper's Table 2 reports that model's coefficients: 0.0211, 0.0432 and βˆ’0.0025. Converted with the formula the Results themselves use, 100Β·(exp(Ξ²)βˆ’1), these give +2.13%, +4.41% and βˆ’0.25%. The Results text prints exactly those three values. So the abstract states each effect larger than the paper's own estimate, by 52%, 34% and 20%. No coefficient in Table 2 or in the supplement's sensitivity models (Table S3) reproduces all three abstract figures. The direction and significance of each association are unchanged, but the abstract's headline magnitudes do not match the model they cite.

What the paper reports

The abstract:

In the final spatial GAMLSS model, proportions of household crowding (>2 residents/bedroom) and elderly population (>59 years) were associated with 3.25% and 5.92% increases in expected TB mortality for each one percentage-point increase, respectively. In contrast, each one percentage-point increase in primary health care coverage was associated with a 0.30% decrease in expected TB mortality.

β€” Abstract, Findings

The Results, describing the same model:

Table 2 presents the final spatial GAMLSS model associated with the outcome at p-value <0.01.

β€” Results, "Final association model", paragraph 1

Each additional month was associated with a 0.48% [ (exp (0.0048) -1) βˆ— 100] increase in the expected number of municipal-monthly TB deaths.

β€” Results, "Final association model", paragraph 2 (this states how coefficients are converted to percentages)

Among the contextual variables, each one percentage-point increase in the proportion of households with more than two residents per bedroom was associated with a 2.13% increase in expected TB deaths. Similarly, each one percentage-point increase in the proportion of residents aged 60 years or older was associated with a 4.41% increase,

β€” Results, "Final association model", paragraph 3

Each one percentage-point increase in coverage was associated with a 0.25% decrease in expected TB deaths.

β€” Results, "Final association model", paragraph 4

Table 2 ("Parametric covariate estimates from the spatial GAMLSS model for monthly TB deaths in SΓ£o Paulo, Brazil, 2020–2024"). The columns are variable, coded variable, estimate, standard error, exponential of the estimate and p-value:

Households with more than 2 residents per bedroom (%) | prop_mais2_dorm | 0.0211 | 0.0073 | 1.0213 | =0.0041

Population aged 60 years or older (%) | prop_mais59 | 0.0432 | 0.0077 | 1.0442 | <0.0001

Primary health care coverage (%) | cobertura_aps | -0.0025 | 0.0007 | 0.9975 | =0.0002

β€” Table 2

I checked each quotation against the article as archived in PubMed Central (src/show.py PMC13382474). Each occurs exactly once. recompute.py.txt repeats that check.

What the paper's own numbers show

Method. The model uses a negative binomial type I distribution with a log link and an offset of log municipal population. A one-unit increase in a covariate therefore multiplies expected deaths by exp(Ξ²), a change of 100Β·(exp(Ξ²)βˆ’1)%. The Results state this formula for the time trend. All three covariates are measured in percent, so a one-unit increase is one percentage point. I applied the formula to the printed Table 2 coefficients. To allow for rounding, I also used the full range each coefficient could have before it was rounded to four decimals (Ξ² Β± 0.00005). I inverted each abstract figure, ln(1 + p/100), to find the coefficient it would require. I formed 95% intervals as Ξ² Β± 1.96Β·SE from the printed standard errors; the paper reports SEs and p-values but no intervals. Finally, I read every coefficient in the supplement's Table S3, which has the final model and four models that each drop one covariate. I tested whether any of them gives an abstract figure, either exponentiated or as 100Β·Ξ².

Quantity Location Reported Recomputed from the paper's own numbers
Crowding (>2 residents/bedroom), % change per 1 pp Abstract +3.25% +2.13% = 100Β·(exp(0.0211)βˆ’1); rounding range +2.127 to +2.138%
Crowding, % change per 1 pp Results text; Table 2 +2.13%; Ξ² 0.0211, exp(Ξ²) 1.0213 +2.13% (consistent)
Coefficient needed for the abstract's +3.25% Abstract vs Table 2 Ξ² = 0.0211 ln(1.0325) = 0.0320
Population aged 60+, % change per 1 pp Abstract +5.92% +4.41% = 100Β·(exp(0.0432)βˆ’1); rounding range +4.409 to +4.420%
Population aged 60+, % change per 1 pp Results text; Table 2 +4.41%; Ξ² 0.0432, exp(Ξ²) 1.0442 +4.41% (+4.42% from the rounded 1.0442; consistent)
Coefficient needed for the abstract's +5.92% Abstract vs Table 2 Ξ² = 0.0432 ln(1.0592) = 0.0575
PHC coverage, % change per 1 pp Abstract βˆ’0.30% βˆ’0.25% = 100Β·(exp(βˆ’0.0025)βˆ’1); rounding range βˆ’0.255 to βˆ’0.245%
PHC coverage, % change per 1 pp Results text; Table 2 βˆ’0.25%; Ξ² βˆ’0.0025, exp(Ξ²) 0.9975 βˆ’0.25% (consistent)
Coefficient needed for the abstract's βˆ’0.30% Abstract vs Table 2 Ξ² = βˆ’0.0025 ln(0.9970) = βˆ’0.0030
Abstract Γ· Table 2 effect (crowding / age 60+ / PHC) Abstract vs Table 2 β€” 1.52 / 1.34 / 1.20
95% CI implied by Table 2, % per 1 pp (crowding / age 60+ / PHC) Table 2 (Ξ² Β± 1.96 SE) not reported +0.68 to +3.60 / +2.85 to +6.00 / βˆ’0.39 to βˆ’0.11
Other Table 2 rows: time trend, TB incidence, education, Black or brown population Results text +0.48%, +2.13%, βˆ’3.18%, +1.43% +0.48%, +2.13%, βˆ’3.19%, +1.43% (consistent within rounding)
Crowding coefficient in the Table S3 sensitivity models Supplement, Table S3 0.0040 / 0.0344 / 0.0381 +0.40% / +3.50% / +3.88%; none is +3.25%
Age 60+ coefficient in the Table S3 sensitivity models Supplement, Table S3 0.0357 / 0.0313 / 0.0259 / 0.0592 +3.63% / +3.18% / +2.62% / +6.10%; none is +5.92% after conversion
PHC coefficient in the Table S3 sensitivity models Supplement, Table S3 βˆ’0.0025 / βˆ’0.0037 / βˆ’0.0024 / βˆ’0.0033 βˆ’0.25% / βˆ’0.37% / βˆ’0.24% / βˆ’0.33%; none is βˆ’0.30%

Within the body of the paper the numbers agree with each other. Table 2's coefficients, its exponentiated column and the Results text all match, as do the four other covariates in the table. The abstract is the only place where the three values differ. Each abstract figure needs a coefficient roughly 20–50% larger in absolute value than the one printed, which is far outside any rounding range.

Figure

The abstract and the Results/Table 2 give different effect sizes for the same three covariates of the same model

Figure 1. The left panel shows the three per-percentage-point effects in the abstract's Findings. The right panel shows, for the same covariates of the same model, the Table 2 coefficient and its printed exponential, the percentage given in the Results text, and 100Β·(exp(Ξ²)βˆ’1) recomputed from the coefficient. Orange marks values as printed in the paper; blue marks values recomputed in recompute.py.txt. Each connector gives the ratio of the abstract value to the recomputed value. The Results text and the recomputation agree to two decimals in all three rows. The abstract does not agree with either in any row.

Abstract values plotted against the Table 2 estimates and their 95% confidence intervals

Figure 2. This figure adds the uncertainty. For each covariate, the blue dot is 100Β·(exp(Ξ²)βˆ’1) from Table 2 and the blue line is the 95% interval from Ξ² Β± 1.96Β·SE. The paper reports the SEs but not the intervals. The orange diamond is the value in the abstract, and the thin vertical line marks no change (0%). Each panel has its own horizontal scale because the effects differ by more than tenfold. All three abstract values lie inside the interval, so sampling uncertainty does not separate them from the estimate. Even so, none of them is the estimate the paper reports. The age 60+ value (+5.92%) sits near the upper limit (+6.00%), and the crowding value (+3.25%) sits in the upper quarter of its interval.

Why it matters

The abstract is where most readers, and most citing papers, get a study's effect sizes. Here every number in the abstract's Findings is overstated relative to the model the abstract names. Crowding is 1.52 times the Table 2 value, age 60+ is 1.34 times and PHC coverage is 1.20 times. The overstatement grows when the effects are applied to realistic differences between municipalities. Over a 10-percentage-point difference in crowding, Table 2 implies +23.5% expected TB deaths (100Β·(exp(10Β·0.0211)βˆ’1)), while the abstract's per-point figure compounds to +37.7%. For the population aged 60+ the corresponding values are +54.0% and +77.7%. The PHC effect is small either way: βˆ’2.5% versus βˆ’3.0% per 10 points. The abstract's Interpretation ends by calling to "strengthen primary health care", and that recommendation rests on this association.

The qualitative findings are not affected. Every association keeps its direction and its significance, and each abstract value lies inside the confidence interval implied by Table 2. The Interpretation, Discussion and conclusions describe the associations without these numbers, so they stand. The error is one of reported magnitude, confined to the abstract.

Explanations considered

No explanation considered accounts for the abstract's three figures. The values in the body of the paper are internally consistent and reproduce from the printed coefficients.

Status

As of 2026-10-23.

Suggested correction

In the abstract's Findings, replace the three effect sizes with the values the paper reports in its Results and Table 2. Adding the confidence intervals would also help:

In the final spatial GAMLSS model, proportions of household crowding (>2 residents/bedroom) and elderly population (>59 years) were associated with 2.13% and 4.41% increases in expected TB mortality for each one percentage-point increase, respectively. In contrast, each one percentage-point increase in primary health care coverage was associated with a 0.25% decrease in expected TB mortality.

If the authors meant the abstract to report a different model, the abstract should name that model and its coefficients should be published. As shown above, though, none of the reported models gives all three current figures. The other numbers in the Results that we checked need no change.