Paperonium

← Back to the plain-language case

By the paper's own model, odds of death rise 2.7-fold per decade of age, not 11-fold

Brain (2022) Β· doi:10.1093/brain/awab466 Β· PMC9336587

Summary

The abstract and Results state that ageing by 10 years increased the odds of death in intensive care after traumatic brain injury 11-fold. The same Results sentence gives the model's estimate as an odds ratio of 1.10 per year of age (95% CI 1.05–1.16), and the authors' R output in the supplement gives the coefficient behind it: 0.09914 log-odds per year. In a logistic model, odds ratios compound over years, so 10 years corresponds to exp(10 Γ— 0.09914) = 2.70 (95% CI 1.59–4.57). The 11-fold figure, which lies far outside that interval, equals 10 Γ— 1.104: the per-year odds ratio multiplied by 10 rather than raised to the 10th power. The direction and significance of the age effect are unchanged, but the abstract overstates its size about fourfold, and Fig. 5D plots the same overstated value.

What the paper reports

Abstract:

A warmer mean brain temperature was associated with survival (P = 0.035), however, and ageing by 10 years increased the odds of death 11-fold (P = 0.0002).

Results, section "Daily brain temperature rhythms predict patient survival":

Applying a generalized linear mixed model (Fig. 5), we found that lack of a daily TBr rhythm, or an age increase of 10 years, increased the odds of death in intensive care 21-fold and 11-fold, respectively (odds ratio for death with rhythm 0.05; 95% CI 0.004–0.57, P = 0.016 and odds ratio for death with ageing by 1 year 1.10; 1.05–1.16, P = 0.0002).

Materials and methods, "Statistical analysis", the model behind both statements:

For the final analysis of patient data, a generalized linear mixed model for logit binomial distribution of patient outcome was chosen

where fixed effects include: Age (patient age in intensive care);

Fig. 5 legend:

(D) Odds of death in intensive care transformed from the data in A–C; the results for these three predictors are significant since the 95% CIs (double-ended error bars) do not include 1.

Fig. 5D itself (an image, read from the archived file awab466f5.jpg) has a row labelled "Age (per decade)" on a logarithmic "Odds of death" axis. Its marker sits at about 11, with a narrow interval.

Supplementary File016 (the authors' R code with its output) fits glmer(Outcome ~ Age +Sex + BrainMean +BrainRange +Daily + (1|Subject), ..., family=binomial) on 98 patients. Its data listing describes Age as "Patient age in years" (integer). It reports:

Age 0.09914 0.02696 3.677 0.000236 ***

exp(0.09914) # calculating Age odds for death [1] 1.104221

exp(c(0.0462984, 0.1519816)) [1] 1.047387 1.164139

What the paper's own numbers show

Method. In a logit-link model, the coefficient for age, Ξ², is the change in log-odds per year. The odds ratio for a k-year difference is therefore exp(k Γ— Ξ²) = (per-year odds ratio)^k, and its Wald 95% CI is exp(k Γ— (Ξ² Β± 1.96 Γ— SE)). Rescaling age from years to decades multiplies Ξ² and SE by the same factor, so z and P do not change. I used the authors' unrounded coefficient and SE from Supplementary File016. As a check, I also used the rounded values printed in the Results, and the two agree. To read Fig. 5D, I calibrated its log x-axis from the six major tick marks on the axis line (80.07 px per decade; largest residual 0.24 px; the minor ticks at 2 and 3 fall within 0.2 px of where a log scale puts them). I then took the centre of the Age ring marker and its two interval end caps. Every number below is printed by recompute.py.txt.

Quantity Location Reported Recomputed from the paper's own numbers
Odds ratio per 1 year of age Results 1.10 (1.05–1.16), P = 0.0002 exp(0.09914) = 1.104 (1.047–1.164); z = 3.677, P = 0.000236 (File016)
Internal consistency of the rounded per-year figures Results 1.10 (1.05–1.16), P = 0.0002 P implied by the CI = 0.00018: consistent
Odds ratio per 10 years of age Abstract; Results 11-fold exp(10 Γ— 0.09914) = 2.70
95% CI per 10 years not reported β€” 1.59–4.57 (from the rounded text: 1.63–4.41)
Per 10 years, from the rounded 1.10 Results β€” 1.10^10 = 2.59; any value from 1.095 to 1.105 gives 2.48–2.71
Per-year odds ratio that 11-fold over 10 years would need derived β€” 11^(1/10) = 1.271, i.e. Ξ² = 0.240, 5.2 SE above the fitted Ξ²
Per-year odds ratio multiplied by 10 derived β€” 1.10 Γ— 10 = 11.00; 10 Γ— 1.104 = 11.04
Per-year CI multiplied by 10 derived β€” 10.47–11.64
"Age (per decade)" marker and interval Fig. 5D plotted, no number given β‰ˆ11.1 (10.4–11.7) digitized; matches 10 Γ— the per-year values, not 2.70 (1.59–4.57)
P per decade Abstract 0.0002 unchanged by rescaling (z = 3.677): consistent
Stated / recomputed per 10 years β€” β€” 11 / 2.70 = 4.1 (4.2 using 1.10^10)
Lack of a daily rhythm, fold increase in odds Abstract; Results 21-fold (OR 0.05; 0.004–0.57; P = 0.016) 1/exp(βˆ’3.03024) = 20.7 (95% CI 1.76–243); P from the CI = 0.016: consistent
Mean brain temperature, OR per 1 Β°C Results 0.42 (0.18–0.94), P = 0.035 exp(βˆ’0.87808) = 0.416, P = 0.035 (File016): consistent

Only the 11-fold age figure, in the text and in Fig. 5D, disagrees with the model output. Every other odds ratio, interval and P value in this paragraph reproduces.

Figure

Odds ratio for death per 10 years of age: printed as 11, recomputed from the paper's own estimate as 2.70 (1.59 to 4.57)

Figure 1. Odds ratios for death in intensive care on one logarithmic axis; the thin vertical line marks no effect (1). Orange: values the paper reports. These are the per-year odds ratio in the Results (filled; its interval of 1.05–1.16 is narrower than the marker), the "11-fold" per 10 years in the Abstract and Results (filled; no interval is given), and the "Age (per decade)" marker in Fig. 5D as digitized from the archived image (open ring, β‰ˆ11 with interval 10.4–11.7). Blue: the 10-year odds ratio recomputed from the authors' own coefficient, exp(10 Γ— 0.09914) = 2.70, with its 95% CI of 1.59–4.57. The printed values lie well outside the recomputed interval.

Calculation flow from the model coefficient to the 10-year odds ratio, showing where compounding and multiplying diverge

Figure 2. The arithmetic chain. Both routes start from the same model output and the same per-year odds ratio of 1.104. Compounding over 10 years (blue, left) gives 2.70 (1.59–4.57). Multiplying the per-year odds ratio and its interval by 10 (orange, right) gives 11.04 (10.47–11.64), which matches both the printed "11-fold" and the position and narrow interval of the Fig. 5D marker. All values are printed by recompute.py.txt.

Why it matters

Explanations considered

Status

Suggested correction

In the Abstract, replace

ageing by 10 years increased the odds of death 11-fold (P = 0.0002)

with

ageing by 10 years increased the odds of death 2.7-fold (P = 0.0002)

In the Results, replace "increased the odds of death in intensive care 21-fold and 11-fold, respectively" with "increased the odds of death in intensive care 21-fold and 2.7-fold, respectively", and optionally add the 10-year interval (95% CI 1.6–4.6). In Fig. 5D, either plot the "Age (per decade)" row at 2.70 (1.59–4.57), or relabel the row "per year" and plot 1.10 (1.05–1.16). The per-year odds ratio, its confidence interval and all P values need no change.