Paperonium

← Back to the plain-language case

Abstract's adjusted hazard ratio 1.30 lies outside its printed 95% CI; Table 2 gives 0.66–2.56

Stroke (2023) Β· doi:10.1161/STROKEAHA.123.043336 Β· PMC10519293

Summary

The abstract reports the multivariable hazard ratio for the primary outcome (any stroke or TIA, patients with new DWI lesions vs without) as "1.30 [95% CI, 0.10–1.02]". A point estimate cannot lie outside its own confidence interval, and no rounding closes the gap: the interval is centred on 0.32 on the log scale. The paper's own Table 2 gives the same estimate as 1.30 (0.66–2.56), P=0.454, and that interval is internally consistent: it is centred on 1.30 and implies Pβ‰ˆ0.45. The error is confined to the abstract's interval. The point estimate, its non-significance and the paper's conclusion are unchanged, but the abstract's headline adjusted result is impossible as printed and, read at face value, points in the protective direction.

What the paper reports

Abstract, Results:

Uni- and multivariable regression analyses did not show significant differences (hazard ratio, 1.50 [95% CI, 0.76–2.94] and hazard ratio, 1.30 [95% CI, 0.10–1.02], respectively).

Abstract, Methods (definition of the outcome these hazard ratios refer to):

The primary composite clinical outcome was the time to any stroke or transient ischemic attack during follow-up.

Methods, Statistical Analysis:

Cox proportional hazard regression models were used to calculate the hazard ratio (HR) with 95% CI for the association between DWI presence and cerebrovascular events during follow-up.

Models were adjusted for the demographic or clinical variables that showed an association of P<0.20 with the determinant (DWI presence) as well as with the outcome of interest (Table S1).

Results, first paragraph:

The binary uni- and multivariable cox regression analyses revealed no significant difference in recurrent cerebrovascular events between DWI+ patients and DWI- patients (Table 2).

Table 2, "HRs of Outcome Events During Follow-Up Depending on New Postoperative DWI Lesions", first column ("Stroke and TIA in any territory"), block "DWI+ (n=51) vs DWI– (n=111)". Table 2 is typeset as an image in the archived article (str-54-2562-g002.jpg), so these cells were read from that image rather than from a text conversion:

Unadjusted | 1.50 (0.76–2.94), P=0.245

Adjusted | 1.30 (0.66–2.56), P=0.454*

*Corrected for age and systolic blood pressure.

Abstract, Conclusions:

Based on our outcome analysis within the ICSS magnetic resonance imaging substudy, DWI lesions following carotid revascularization did not seem to have a relationship with long-term stroke risk.

What the paper's own numbers show

Method. Cox model confidence intervals are Wald intervals on the log-hazard scale, so a correctly printed 95% CI is symmetric around ln(HR). From each printed HR and CI I computed: whether the HR lies inside the interval; the geometric centre √(lower Γ— upper), which should equal the HR up to rounding; the log-asymmetry [ln(upper) βˆ’ ln(HR)] βˆ’ [ln(HR) βˆ’ ln(lower)], which should be about 0; the standard error (ln upper βˆ’ ln lower)/(2 Γ— 1.96); and the two-sided P implied by ln(HR)/SE, for comparison with the printed P. I applied the same checks to every DWI+ vs DWIβˆ’ cell of Table 2 and to the Fine-Gray sensitivity analysis in Supplemental Table S2, so that the abstract figure can be judged against the paper's other estimates. Every number below is printed by recompute.py.txt.

Quantity Location Reported Recomputed from the paper's numbers
Multivariable HR, any stroke/TIA Abstract, Results 1.30 [95% CI 0.10–1.02] 1.30 is outside the interval; geometric centre 0.32; log-asymmetry βˆ’2.81 (expected β‰ˆ 0)
Rounding margin derived β€” smallest HR that rounds to 1.30 is 1.295; largest upper limit that rounds to 1.02 is 1.025; gap 0.270
Adjusted HR, any stroke/TIA (age, systolic BP) Table 2 1.30 (0.66–2.56), P=0.454 centre 1.30; log-asymmetry 0.00; SE 0.346; P=0.448
95% CI implied by HR 1.30 with the Table 2 SE derived 0.10–1.02 (abstract) 0.66–2.56, identical to Table 2
Univariable HR, any stroke/TIA Abstract; Table 2 1.50 (0.76–2.94), P=0.245 centre 1.49; log-asymmetry βˆ’0.01; SE 0.345; P=0.240
Log-width of the CI Abstract vs Table 2 0.10–1.02 vs 0.76–2.94 and 0.66–2.56 2.32 vs 1.35 and 1.36: the abstract interval is 1.72 times as wide as the univariable one
Unadjusted / adjusted, ipsilateral stroke/TIA Table 2 1.91 (0.81–4.49), P=0.138; 1.25 (0.52–2.98), P=0.619 centres 1.91 and 1.24; P=0.139 and 0.616
Unadjusted, stroke in any territory Table 2 0.69 (0.26–1.84), P=0.456 centre 0.69; P=0.457
Unadjusted / adjusted, ipsilateral stroke Table 2 0.88 (0.24–3.25), P=0.845; 0.78 (0.21–2.93), P=0.713 centres 0.88 and 0.78; P=0.848 and 0.712
Adjusted, stroke in any territory (side observation) Table 2 0.70 (0.29–1.51), P=0.493 centre 0.66; log-asymmetry βˆ’0.11; P=0.397 from the CI. A smaller inconsistency, separate from this report
Fine-Gray unadjusted / adjusted, any stroke/TIA Supplemental Table S2 1.35 (0.68–2.66); 1.16 (0.57–2.36) centres 1.34 and 1.16; both consistent; neither is 0.10–1.02

Apart from the abstract's multivariable interval, every estimate of this outcome in the paper passes the checks to within rounding. Its adjusted interval in Table 2 has almost exactly the univariable width (1.36 vs 1.35 on the log scale), as expected when adjusting for two covariates in a cohort of 162.

Figure

The abstract's multivariable 95% CI (0.10–1.02) excludes its own HR of 1.30 and is centred on 0.32; the paper's Table 2 interval for the same estimate (0.66–2.56) is centred on 1.30.

Hazard ratios for the primary outcome (any stroke or TIA, DWI+ vs DWIβˆ’) on one log-scaled axis. Orange: the HR (dot) and 95% CI (bar) as printed in the paper. Blue diamonds: the geometric centre of each printed interval, recomputed in recompute.py.txt. For a Cox model's Wald interval it should coincide with the HR. The dashed line marks HR = 1 (no difference). The abstract's multivariable row is the only one where the centre (0.32) and the printed HR (1.30) disagree, and its HR sits to the right of the upper limit. The Table 2 adjusted row and the univariable row are self-consistent. The abstract's interval is also wider, not narrower, than the univariable one.

Why it matters

The abstract's multivariable hazard ratio is the adjusted result for the primary outcome, and most readers and indexing services see only the abstract. As printed, the pair is impossible. Taken at face value, the interval 0.10–1.02 describes an estimate near 0.32 (a protective association) whose upper limit only just reaches 1. The study's actual adjusted estimate is 1.30 (0.66–2.56): imprecise, not significant, and leaning towards a higher hazard in DWI+ patients. The upper limit of the correct interval, about a 2.6-fold hazard, also bears on how the conclusion "did not seem to have a relationship with long-term stroke risk" should be read: the data do not exclude a clinically relevant increase. That caveat applies whichever interval is printed, but the misprinted one hides it.

Neither the direction of the main conclusion nor its significance depends on the error. Both the printed and the correct intervals include 1, the point estimate is the same, and the correct value is printed in Table 2 of the same article. A reader who checks Table 2 can recover the right value. A reader of the abstract alone, or of a database record built from it, cannot.

Explanations considered

What remains is that the abstract's confidence limits were transcribed incorrectly. The paper does not show how this happened.

Status

As of 2026-10-24:

Suggested correction

In the Abstract, Results, replace "hazard ratio, 1.30 [95% CI, 0.10–1.02]" with "hazard ratio, 1.30 [95% CI, 0.66–2.56]", the value given for the same adjusted model in Table 2 (adjusted for age and systolic blood pressure; P=0.454). The authors may also wish to check the Table 2 cell for adjusted stroke in any territory, "0.70 (0.29–1.51), P=0.493". Its interval is centred on 0.66, and the P value it implies (0.40) differs from the one printed, which suggests that one of those numbers may also need correcting.