Paperonium

← Back to the plain-language case

Abstract transposes two of four unilateral and bilateral effusion mortality figures given correctly in Results

Eur Respir J (2015) Β· doi:10.1183/09031936.00217114 Β· PMC4857137

Summary

This cohort of 308 patients undergoing thoracentesis (247 unilateral, 61 bilateral) reports in its abstract that bilateral effusion, relative to unilateral, carried higher mortality, giving "17% versus 47%" with the 30-day hazard ratio and "36% versus 69%" with the 1-year hazard ratio. The Results instead give unilateral mortality as 17% (30 days) and 47% (1 year) and bilateral mortality as 36% and 69%, and only these figures reproduce the paper's reported totals of 65 and 158 deaths (43/247 + 22/61 and 116/247 + 42/61); no count out of 61 even rounds to 47%. The abstract, including the version of record indexed in PubMed, has therefore transposed 47% and 36%: it overstates bilateral 30-day mortality (47% instead of 36%), understates unilateral 1-year mortality (36% instead of 47%), and makes the absolute gap look 11 percentage points larger at both time points. The hazard ratios and the direction of the finding are unaffected, but this is the comparison the paper presents as its new contribution, and the incorrect 47% figure has since been cited.

What the paper reports

Abstract, Results (PMC author manuscript):

Patients with bilateral pleural effusion, relative to unilateral, were associated with higher risk of death at 30 days and 1 year (17% versus 47%; HR 2.58 CI [1.44–4.63] and 36% versus 69%; HR 2.32 CI [1.55–3.48])

Abstract, Results, as published in Eur Respir J 46(2):495–502 and indexed in PubMed (PMID 25837039):

Patients with bilateral, relative to unilateral, pleural effusion were associated with higher risk of death at 30 days and 1 year (17% versus 47% (hazard ratio (HR) 2.58, 95% CI 1.44-4.63) and 36% versus 69% (HR 2.32, 95% CI 1.55-3.48), respectively).

Results, first paragraph:

A total of 247 unilateral and 61 bilateral thoracenteses were performed. Overall mortality across all groups was 21% (65 of 308 patients) at 30-days and 51% (158 of 308 patients) at one year

Results, second paragraph:

As calculated from Figure 1, thirty-day and one-year mortality of patients with unilateral effusion was 17% and 47%, compared with 36% and 69% in patients with bilateral effusion.

Relative to patients with unilateral effusion, unadjusted mortality of patients with bilateral effusion was higher at thirty-days (p=0.0003) and at one-year (p=0.0001).

Results, fifth paragraph (Table 3, row "Bilateral", reads the same: 2.58 (1.44, 4.63), 0.002 | 2.32 (1.55, 3.48), <0.001):

was significantly associated with higher 30-day and 1-year mortality (HR 2.58 [1.44–4.63] p=0.002 and HR 2.32 [1.55–3.48] p<0.001, respectively)

Methods, Statistical Analysis, on what the percentages are:

Unadjusted 30-day and 1-year mortalities were calculated and diagrammed for subgroups according to the following hierarchy: laterality (bilateral versus unilateral) and malignancy (yes versus no).

Discussion, on the weight the paper gives the comparison:

The study is novel in evaluating mortality associated with bilateral effusions.

It is also the first study to report the association between bilateral pleural effusion and mortality.

Abstract, Conclusions:

Bilateral pleural effusion is distinctly associated with high mortality.

What the paper's own numbers show

Method. The group sizes (247 and 61) and the overall death counts (65 and 158 of 308) come from the Results. For each printed percentage I listed every integer number of deaths k for which k/n rounds to that percentage. I then asked which pairing of unilateral and bilateral percentages can reproduce the reported overall totals: the pairing in the Results (unilateral 17%/47%, bilateral 36%/69%), or the pairing in the abstract, where each "x% versus y%" is tied to one time point by the hazard ratio that follows it (30 days: 17% vs 47%; 1 year: 36% vs 69%). For each correct split I recomputed the risk difference and crude risk ratio. Every number below is printed by recompute.py.txt.

Quantity Location Reported Recomputed from the paper's numbers
Group sizes Abstract; Results ΒΆ1 247 unilateral, 61 bilateral 247 + 61 = 308
Overall deaths, 30 days Results ΒΆ1 65 of 308 (21%) 21.1%
Overall deaths, 1 year Results ΒΆ1 158 of 308 (51%) 51.3%
Unilateral, 30 days Results ΒΆ2; abstract 17% 43/247 = 17.4% (41–43 deaths round to 17%)
Bilateral, 30 days Results ΒΆ2 36% 22/61 = 36.1% (22 is the only count giving 36%)
Bilateral, 30 days Abstract (30-day pair, with HR 2.58) 47% not attainable: 28/61 = 45.9%, 29/61 = 47.5%; the value is the unilateral 1-year figure
Unilateral, 1 year Results ΒΆ2 47% 116/247 = 47.0%
Unilateral, 1 year Abstract (1-year pair, with HR 2.32) 36% would mean 88–90 deaths; the value is the bilateral 30-day figure
Bilateral, 1 year Results ΒΆ2; abstract 69% 42/61 = 68.9%
Total 30-day deaths implied, Results pairing derived 65 0.17 Γ— 247 + 0.36 Γ— 61 = 64.0 (63–65 with rounding); the only exact split is 43 + 22 = 65
Total 30-day deaths implied, abstract pairing derived 65 0.17 Γ— 247 + 0.47 Γ— 61 = 70.7 (about 69–72); no integer split gives 65
Total 1-year deaths implied, Results pairing derived 158 0.47 Γ— 247 + 0.69 Γ— 61 = 158.2 (157–159); the only exact split is 116 + 42 = 158
Total 1-year deaths implied, abstract pairing derived 158 0.36 Γ— 247 + 0.69 Γ— 61 = 131.0 (130–132); no integer split gives 158
Bilateral minus unilateral, 30 days Abstract vs Results 47 βˆ’ 17 = 30 pp 36.1 βˆ’ 17.4 = 18.7 pp (19 pp from the rounded Results figures); crude RR 2.07 (2.76 as printed in the abstract)
Bilateral minus unilateral, 1 year Abstract vs Results 69 βˆ’ 36 = 33 pp 68.9 βˆ’ 47.0 = 21.9 pp (22 pp from the rounded Results figures); crude RR 1.47 (1.92 as printed in the abstract)
Adjusted HR, bilateral vs unilateral Abstract; Results ΒΆ5; Table 3 2.58 (1.44–4.63) at 30 days; 2.32 (1.55–3.48) at 1 year identical in all three places; not affected

The four numbers in the abstract are the right four numbers, but 47% and 36% have swapped places. The Results figures, the death counts they imply and the reported overall totals all agree with one another. The abstract's pairing agrees with none of them.

One side observation does not form part of the finding. Pearson chi-square tests on the recomputed counts give unadjusted p = 0.0014 (30 days) and 0.0022 (1 year), and Fisher exact tests give 0.0026 at both time points. The Results report p = 0.0003 and 0.0001. The paper does not name the test, and a log-rank test on time to death could legitimately give different values, so I leave this unresolved.

Figure

The abstract's mortality pairs compared with the Results, with the transposed cells joined

Figure 1. The upper panel quotes the abstract's sentence and lays out its four percentages as the sentence assigns them: each "x% versus y%" pair is followed by its time-point hazard ratio (HR 2.58 is the 30-day HR and 2.32 the 1-year HR in Table 3), and in each pair bilateral is the higher value. The values as printed are shown in orange. The lower panel quotes the Results and shows, in blue, the death counts recomputed from the paper's own numbers: the only integer split that matches both the Results percentages and the reported totals of 65 and 158 deaths. The unilateral 30-day (17%) and bilateral 1-year (69%) cells agree. The bilateral 30-day and unilateral 1-year cells hold each other's values: the abstract's 47% is the unilateral 1-year mortality (116/247), and its 36% is the bilateral 30-day mortality (22/61).

Total deaths implied by each pairing compared with the reported totals

Figure 2. Total deaths implied by applying each pairing's percentages to the 247 unilateral and 61 bilateral patients, on one axis of deaths that starts at zero. Filled blue circles use the Results pairing and open blue circles the abstract's pairing. The ranges in the labels allow for rounding of the printed percentages. Orange lines mark the reported totals (65/308 at 30 days, 158/308 at 1 year). The Results pairing reproduces both totals. The abstract's pairing implies about 71 deaths at 30 days and 131 at 1 year, and neither range contains the reported total.

Why it matters

Explanations considered

Status

Suggested correction

In the Results sentence of the abstract, change

(17% versus 47%; HR 2.58 CI [1.44–4.63] and 36% versus 69%; HR 2.32 CI [1.55–3.48])

to

(36% versus 17%; HR 2.58 CI [1.44–4.63] and 69% versus 47%; HR 2.32 CI [1.55–3.48])

or, keeping the original order, "(17% versus 36% … and 47% versus 69% …)", stated explicitly as unilateral versus bilateral. The version of record needs the same change, with "respectively" kept. No change is needed to the main text, Table 3 or the conclusions. Because the incorrect figure has already been cited, a published correction notice linked to the article would be more useful than a silent update.