Paperonium

← Back to the plain-language case

Adjusted odds ratio for discharge to rehabilitation, 2.45, lies outside its printed 95% CI of 0.81–1.25

Stroke (2020) Β· doi:10.1161/STROKEAHA.120.031217 Β· PMC7678670

Summary

The abstract, the Results text and Table 3 all give the adjusted odds ratio for discharge to rehabilitation among patients with stroke, with versus without COVID-19, as 2.45 (95% CI 0.81–1.25), P = 0.103. The interval does not contain its own estimate: its geometric midpoint, where a logistic-regression Wald interval puts the estimate, is 1.006. The printed estimate, lower limit and P value are consistent with each other and imply an upper limit of about 7.2–7.4, matching the width of the paper's unadjusted interval (0.75–6.38). The result stays non-significant and in the same direction, and the primary and mortality results are unaffected, but the printed interval understates the standard error about fivefold.

What the paper reports

Abstract, Results:

In patients presenting with stroke, concurrent infection with severe acute respiratory syndrome-coronavirus 2 (SARS-CoV-2) was associated with higher case-fatality (odds ratio, 10.50 [95% CI, 3.54–31.18]) and a trend towards increased occurrence of discharge to rehabilitation (odds ratio, 2.45 [95% CI, 0.81–1.25]).

Results, section "Discharge to Rehabilitation for Ischemic Stroke":

Overall, 60% (1511/2513) of patients with stroke with COVID-19 were discharged to rehabilitation, compared with 40.7% (9074/22 295) of patients with stroke without SARS-CoV-2 infection. COVID-19 was associated with a trend towards an increased frequency of discharge to rehabilitation (OR, 2.19 [95% CI, 0.75–6.38]) in unadjusted analysis. Similarly, using a logistic regression with facility fixed effects, we identified that SARS-CoV-2 infection was associated with a trend towards an increased frequency of discharge to rehabilitation (OR, 2.45 [95% CI, 0.81–1.25]), compared with noninfected individuals, albeit nonsignificant (Table 3).

Table 3, "Models Examining the Association of COVID-19 With Unfavorable Outcomes in Patients With Acute Ischemic Stroke". PMC archives this table only as a typeset image, so the values below were transcribed from that image:

Unadjusted analysis: Mortality 9.68 (3.55–26.4), P <0.001; Discharge to rehabilitation 2.19 (0.75–6.38), P 0.143
Logistic regression†: Mortality 10.50 (3.54–31.18), P <0.001; Discharge to rehabilitation 2.45 (0.81–1.25), P 0.103
†Hospital ID was used as a fixed-effects variable.

Methods, "Statistical Analysis":

For our secondary outcomes, given the limited df (relatively small number of observations), we used a step-wise logistic regression model controlling for age, sex, hypertension, hypercholesterolemia, diabetes, and peripheral vascular disease.

All results are based on 2-sided tests, and the level of statistical significance was set at 0.05.

Discussion, first paragraph:

Patients with stroke with concurrent SARS-CoV-2 infection demonstrated increased case-fatality and a trend more discharge to rehabilitation.

Conclusions:

Patients with stroke with concurrent SARS-CoV-2 infection demonstrated increased case-fatality and discharge to rehabilitation.

What the paper's own numbers show

Method. A logistic-regression Wald interval is exp(Ξ² Β± 1.96Β·SE). It is symmetric on the log scale, so the odds ratio equals the geometric mean of the two bounds and ORΒ² = lower Γ— upper. I applied this check to all 20 odds ratios in Tables 2 and 3. I then rebuilt the adjusted rehabilitation interval in two ways. The first keeps the printed OR and lower bound. The second keeps the printed OR and P value. Every number below is printed by recompute.py.txt.

Quantity Location Reported Recomputed from the paper's own numbers
Does the adjusted rehabilitation CI contain its estimate? Abstract; Results; Table 3 OR 2.45, CI 0.81–1.25 No: 2.45 is 1.20 above the upper limit. An OR that rounds to 2.45 is at least 2.445, and an upper limit that rounds to 1.25 is below 1.255, so rounding cannot account for the gap
Geometric midpoint of the printed bounds same (should equal 2.45) √(0.81 Γ— 1.25) = 1.006
Upper limit implied by OR 2.45 and lower limit 0.81 Table 3 1.25 2.45Β² / 0.81 = 7.41; SE 0.565; implied P = 0.113
95% CI implied by OR 2.45 and printed P = 0.103 Table 3 0.81–1.25 SE 0.550 β†’ 0.83–7.19
95% CI if the adjusted model had the unadjusted model's SE (0.546) Table 3 0.81–1.25 0.84–7.15
Unadjusted rehabilitation OR Results; Table 3 2.19 (0.75–6.38), P = 0.143 midpoint 2.187; implied P 0.151 (consistent)
Unadjusted mortality OR Results; Table 3 9.68 (3.55–26.4) midpoint 9.681 (consistent)
Adjusted mortality OR Abstract; Results; Table 3 10.50 (3.54–31.18) midpoint 10.506 (consistent)
Stroke, ischemic stroke, TIA and MI ORs, 16 rows Results; Table 2 e.g. 0.25 (0.16–0.40); MI 0.87 (0.68–1.12), P 0.264 all contain their estimates; midpoints match the ORs to within rounding except TIA mixed-effects (below); MI implied P values 0.274, 0.379, 0.340 vs printed 0.264, 0.382, 0.341 (MI inverse-weighted: see below)
Crude stroke OR Results, "COVID-19 and Ischemic Stroke" 0.35 (0.23–0.55) 22/2491 vs 544/21 751 β†’ 0.353 (0.23–0.54)

The adjusted rehabilitation entry is the only one of the 20 checked odds ratios whose interval excludes its estimate. The two reconstructions agree closely (upper limit 7.19 vs 7.41; lower limit 0.83 vs printed 0.81), so the printed lower limit and P value fit OR 2.45, and the upper limit 1.25 is the number that does not. The reconstructed width also matches the unadjusted model's interval (0.75–6.38), as one would expect for two models fitted to the same 566 patients with few events. The exact correct upper limit cannot be fixed without the model output. A value of 7.25, which differs from 1.25 in one digit, would give a midpoint of 2.42 and an implied P of 0.109, but this is only one possibility.

Side observations (not part of the main finding).

Figure

The printed adjusted interval lies entirely below its own estimate; intervals rebuilt from the paper's other numbers are wide and include 1

Figure 1. Odds ratios for discharge to rehabilitation, COVID-19 vs no COVID-19, among patients with ischemic stroke, with 95% CIs on a log scale. The thin vertical line marks OR = 1. Orange: the values printed in the paper. The unadjusted model is 2.19 (0.75–6.38). The adjusted model is printed as 2.45 (0.81–1.25): the filled dot is the printed estimate, the bar is the printed interval, and the open circle is the interval's geometric midpoint (1.01), where a Wald interval puts its estimate. Blue: the adjusted interval rebuilt from the paper's own numbers. The first is rebuilt from OR 2.45 and the lower limit 0.81 by log symmetry (0.81–7.41, implied P = 0.113). The second is rebuilt from OR 2.45 and the printed P = 0.103 (0.83–7.19). Both rebuilt intervals have about the same width as the unadjusted one and both include 1.

The printed adjusted rehabilitation interval implies a standard error about one-fifth of every other Table 3 model's

Figure 2. The standard error of log(OR) implied by each 95% CI in Table 3, computed as (ln upper βˆ’ ln lower) / 3.92. Orange: intervals as printed. Blue: the two reconstructed adjusted rehabilitation intervals from Figure 1. Of the four models fitted to the same 566 stroke patients (22 with COVID-19), three imply SEs between 0.512 and 0.555; the printed adjusted rehabilitation interval implies 0.111. The reconstructed intervals imply 0.565 and 0.550, in line with the other three. The axis starts at zero.

Why it matters

Explanations considered

Status

Suggested correction

The adjusted odds ratio for discharge to rehabilitation should be reported with its correct 95% confidence interval in the abstract, in the Results ("Discharge to Rehabilitation for Ischemic Stroke") and in Table 3. If the printed estimate (2.45), lower limit (0.81) and P value (0.103) are right, the upper limit should be about 7.2–7.4, not 1.25. The exact value should be taken from the original model output. If another of the three numbers is the one in error, all three should be checked against that output. The Conclusions sentence "Patients with stroke with concurrent SARS-CoV-2 infection demonstrated increased case-fatality and discharge to rehabilitation" could be aligned with the abstract and Discussion, which call the rehabilitation result a non-significant trend. Separately, the denominators in the rehabilitation and case-fatality paragraphs (2513 and 22 295) and the abstract's "566 (0.2%)" (566/24 808 = 2.3%) could be checked.