Paperonium

← Back to the plain-language case

Two pooled pericarditis hazard ratios in Table 2 fall outside their own cohort estimates

JAMA Cardiol (2026) Ā· doi:10.1001/jamacardio.2026.0258 Ā· PMC13000738

Summary

The letter pools two biobank cohorts by inverse-variance meta-analysis and reports pooled pericarditis hazard ratios (HRs) of 1.66 (95% CI, 1.17-2.35) for large CHIP and 1.88 (1.45-2.43) for DNMT3A CHIP. Neither value can come from the cohort rows printed above it in Table 2: any weighted average of the large-CHIP cohort HRs (2.15 and 1.71) must lie between 1.71 and 2.15, and any weighted average of the DNMT3A cohort HRs (1.70 and 1.63) cannot exceed 1.70. Pooling the table's own cohort rows gives 1.86 (1.43-2.42) for large CHIP and 1.66 (1.17-2.35) for DNMT3A, so the printed pooled rows match the recomputed values with the two exposures exchanged, and the myocarditis block shows the same pattern. Two interpretive sentences in the Results rest on the printed values. The main finding is not affected: the any-CHIP point estimate recomputes (HR 1.65), with a printed CI (1.29-2.10) slightly narrower than the recomputed one (1.28-2.12).

What the paper reports

Methods, second paragraph:

Inverse variance-weighted meta-analysis (using fixed-effects models when I2 < 25% and random-effects models otherwise) was performed.

Results, first paragraph:

Any CHIP was associated with 1.65-fold increased pericarditis risk (95% CI, 1.29-2.10; I2 = 0). Large CHIP (variant allele fraction ≄10%) had a similar point estimate (hazard ratio [HR], 1.66 [95% CI, 1.17-2.35]; I2 = 0). DNMT3A CHIP accounted for the association (HR, 1.88 [95% CI, 1.45-2.43]; I2 = 0), while TET2 CHIP showed a directionally concordant association (1.35 [0.76-2.40]).

Results, second paragraph (myocarditis):

Point estimates for any CHIP (HR, 1.29 [95% CI, 0.84-1.97]), large CHIP (HR, 1.53 [0.92-2.54]), and DNMT3A CHIP (HR, 1.30 [0.81-2.09]) were positive but not significant.

There was moderate heterogeneity between cohorts (I2 = 37.1%-68.2%).

Table 2, "Multivariable-Adjusted Associations of CHIP and Gene-Specific CHIP Subtypes With Myocarditis and Pericarditis After Meta-Analysis of the VUMC BioVU and NIH All of Us Cohorts". These are the rows at issue (participants | events | adjusted HR (95% CI) | P):

Outcome, exposure Row Participants Events HR (95% CI) P
Pericarditis, Large CHIP All of Us cohort 6989 23 2.15 (1.40-3.30) <.001
BioVU cohort 7562 39 1.71 (1.23-2.38) .001
Meta-analysis 14 551 62 1.66 (1.17-2.35) .004
Pericarditis, DNMT3A CHIP All of Us cohort 4835 13 1.70 (0.98-2.97) .06
BioVU cohort 3754 20 1.63 (1.04-2.56) .03
Meta 8589 33 1.88 (1.45-2.43) <.001
Myocarditis, Large CHIP All of Us cohort 7013 3 0.54 (0.17-1.68) .28
BioVU cohort 7573 16 1.56 (0.93-2.61) .09
Meta-analysis 14 586 19 1.53 (0.92-2.54) .10
Myocarditis, DNMT3A CHIP All of Us cohort 4839 4 1.02 (0.38-2.75) .97
BioVU cohort 3756 10 1.87 (0.99-3.53) .05
Meta-analysis 8595 14 1.30 (0.81-2.09) .28

Discussion, first paragraph:

We and [ref 2] found that DNMT3A—but not TET2—CHIP was associated with pericarditis, although this may simply reflect statistical power.

(Bracketed text replaces the cited study's author name.)

What the paper's own numbers show

Method. I applied the method stated in Methods to every block of Table 2. Each cohort's log HR was taken with a standard error recovered from its 95% CI, SE = [ln(upper) āˆ’ ln(lower)] / (2 Ɨ 1.96). The two log HRs were pooled with inverse-variance weights, and I computed Cochran's Q and I². I used the fixed-effect result when I² < 25% and the DerSimonian-Laird random-effects result otherwise, as the paper specifies. There is also a check that needs no standard errors at all: a pooled estimate built from positive weights is a weighted average of the cohort log HRs, so it must lie between the two cohort HRs. This holds for fixed and random effects and for pooling on either the log or the HR scale. To test whether rounding to two decimals could matter, I varied every printed cohort HR and CI bound by ±0.005 (all 729 combinations of āˆ’0.005, 0 and +0.005, plus 20 000 random draws) in all eight blocks of Table 2, under both fixed and random effects. A printed two-decimal value is counted as fitting when its own rounding interval (±0.005) overlaps the recomputed range. Every number below is printed by recompute.py.txt.

Quantity Location Reported Recomputed from the paper's own rows
Range any weighted average can take, pericarditis, large CHIP Table 2 cohort rows 2.15 and 1.71 must lie in [1.71, 2.15]
Pooled HR, pericarditis, large CHIP Table 2; Results ¶1 1.66 (1.17-2.35), P = .004 1.86 (1.43-2.42), P = 3.2 Ɨ 10⁻⁶; Q = 0.69, I² = 0% (fixed and random effects identical)
Same, allowing for rounding of inputs Table 2 1.66 1.854 to 1.870
SE of pooled log HR, pericarditis, large CHIP Table 2 0.178 (implied by printed CI) 0.133; a pooled SE cannot exceed the smaller cohort SE (BioVU 0.168) when τ² = 0
Range any weighted average can take, pericarditis, DNMT3A Table 2 cohort rows 1.70 and 1.63 must lie in [1.63, 1.70]
Pooled HR, pericarditis, DNMT3A Table 2; Results ¶1 1.88 (1.45-2.43), P < .001 1.66 (1.17-2.35), P = .0046; Q = 0.01, I² = 0%
Same, allowing for rounding of inputs Table 2 1.88 1.652 to 1.663
SE of pooled log HR, pericarditis, DNMT3A Table 2 0.132 (implied by printed CI) 0.178
Participant and event totals, large CHIP pericarditis Table 2 14 551; 62 6989 + 7562 = 14 551; 23 + 39 = 62 (agree)
Participant and event totals, DNMT3A pericarditis Table 2 8589; 33 4835 + 3754 = 8589; 13 + 20 = 33 (agree)
Pooled HR, myocarditis, large CHIP Table 2; Results ¶2 1.53 (0.92-2.54), P = .10 fixed effect 1.30 (0.81-2.09), P = .27; I² = 63.5% (61.6% to 65.2% under rounding), so the stated rule gives random effects 1.04 (0.38-2.86), P = .93
Pooled HR, myocarditis, DNMT3A Table 2; Results ¶2 1.30 (0.81-2.09), P = .28 1.57 (0.92-2.67), P = .10; I² = 2.0%

The other four pooled rows. Re-pooled the same way, the two any-CHIP rows and the pericarditis TET2 row come close to the printed values, and the myocarditis TET2 row does not. Under the rounding test, none of the four is reproduced exactly:

Pooled row (Table 2) Reported Recomputed (rule-chosen model) Range allowing rounding of inputs Assessment
Pericarditis, any CHIP 1.65 (1.29-2.10) 1.65 (1.28-2.12), fixed effect, I² = 0% HR 1.639-1.651; lower 1.269-1.282; upper 2.110-2.134 HR fits. Printed CI lies inside every recomputed CI: implied SE 0.124 (0.123-0.126) vs 0.129-0.131 recomputed
Pericarditis, TET2 1.35 (0.76-2.40), P = .31 1.38 (0.78-2.44), P = .27, fixed effect, I² = 0% HR 1.371-1.387; lower 0.773-0.784; upper 2.426-2.463 Close (1.38 vs 1.35), just outside the rounding range 1.37-1.39; CI width fits (SE 0.293 vs 0.290-0.294)
Myocarditis, any CHIP 1.29 (0.84-1.97), P = .24 1.32 (0.86-2.01), P = .20, fixed effect, I² = 11.7% (7.0%-16.2%) HR 1.307-1.326; lower 0.857-0.871; upper 1.990-2.021 Close (1.32 vs 1.29), just outside the rounding range 1.31-1.33; the lower bound 0.84 is also outside 0.86-0.87
Myocarditis, TET2 0.69 (0.21-2.26), P = .54 0.90 (0.29-2.82), P = .86, fixed effect, I² = 13.8% (8.0%-19.0%) Fixed effect HR 0.880-0.925; random effects 0.848-0.904 Not reproducible (0.88-0.93 under rounding), but 0.69 lies between the cohort HRs 0.37 and 1.40, so it is not an impossible value

So the printed pooled rows are not, in general, an exact inverse-variance pool of the printed cohort rows, and differences of about 0.01 to 0.03 on the log scale appear even in rows with no other problem (any CHIP and TET2 for pericarditis, any CHIP for myocarditis). The large-CHIP and DNMT3A pericarditis rows differ by 0.115 and 0.126 in the HR alone and, unlike those rows, lie outside the range of their own cohort HRs. Myocarditis TET2 is a separate case: its printed HR is 0.268 on the log scale from the recomputed value, which is more than the large-CHIP and DNMT3A myocarditis rows (0.160 and 0.187). Both cohort CIs in that block are close to symmetric on the log scale (1.386 vs 1.397, and 2.001 vs 1.961), as Wald intervals are, so recovering SEs from the CIs does not explain the gap. The block has only 3 events (2 and 1), and I treat it as a side observation, not as part of the main finding.

The exchange pattern. The recomputed DNMT3A pericarditis pool, 1.66 (1.17-2.35) with P = .0046, reproduces the printed large-CHIP row, 1.66 (1.17-2.35) with P = .004. The largest log difference across HR and both bounds is 0.002, and all three printed values fall within the recomputed rounding ranges (HR 1.652-1.663, lower 1.162-1.175, upper 2.338-2.364). The recomputed large-CHIP pool, 1.86 (1.43-2.42), is within 0.011 on the log scale of the printed DNMT3A row, 1.88 (1.45-2.43), a gap similar in size to those in the any-CHIP and pericarditis TET2 rows. By contrast, each printed row differs from its own recomputed pool by 0.115 (large CHIP) and 0.126 (DNMT3A) on the log scale. Myocarditis shows the same pattern. The fixed-effect pool of the large-CHIP myocarditis cohorts, 1.30 (0.81-2.09), matches the printed DNMT3A row to two decimals (HR and both CI bounds; largest log difference 0.006; P = .27 recomputed vs .28 printed), and the DNMT3A pool, 1.57 (0.92-2.67) with P = .10, is close to the printed large-CHIP row, 1.53 (0.92-2.54) with P = .10 (largest log difference 0.052). In both outcome blocks the participant and event totals in the pooled rows agree with their own cohort rows. Only the HR, CI and P cells of the large-CHIP and DNMT3A pooled rows look exchanged. This is an inference from the numbers and has not been confirmed by the authors.

On its own, each printed myocarditis value is possible: 1.53 lies between 0.54 and 1.56, and 1.30 between 1.02 and 1.87. So the myocarditis evidence rests on the match with the exchanged values, not on an impossible value. Because the myocarditis TET2 row shows that a printed myocarditis pool can differ from re-pooling by more than these rows do, the myocarditis part of the finding is weaker than the pericarditis part.

Figure

Pericarditis: the printed pooled HRs for large CHIP and DNMT3A CHIP each lie outside the range of their own two cohort HRs, while the recomputed pools lie inside

Figure 1. Pericarditis rows of Table 2 for large CHIP (top) and DNMT3A CHIP (bottom), on one log-scale HR axis. Grey diamonds with lines are the two cohort estimates and 95% CIs as printed. Orange squares are the pooled meta-analysis rows as printed. Blue circles are the fixed-effect inverse-variance pools recomputed by recompute.py.txt from the two cohort rows directly above them (I² = 0% in both blocks, so random effects give the same values). The shaded band marks the interval between the two cohort point estimates, where any weighted average has to fall. Both printed pooled values lie outside their band, one below it and one above. Each printed value sits at, or close to, the recomputed value of the other exposure. The thin vertical line marks HR = 1.

Myocarditis: the printed pooled rows for large CHIP and DNMT3A CHIP match the recomputed pools of the other exposure

Figure 2. The same layout for the myocarditis rows of Table 2. In both blocks the printed values lie inside the band, but the printed DNMT3A pool matches the recomputed fixed-effect large-CHIP pool to two decimals, and the printed large-CHIP pool is close to the recomputed DNMT3A pool. For large CHIP the recomputed I² is 63.5%, so under the paper's stated rule the random-effects estimate, 1.04 (0.38-2.86), would apply. The figure shows the fixed-effect value because that is the one matching a printed row.

Why it matters

Explanations considered

Status

Suggested correction

In Table 2, replace the pooled (meta-analysis) rows for large CHIP and DNMT3A CHIP, under both pericarditis and myocarditis, with values recomputed from the cohort-level estimates using the stated inverse-variance method. From the rounded table values these are approximately:

Outcome, exposure Printed Recomputed from Table 2 cohort rows
Pericarditis, large CHIP 1.66 (1.17-2.35), P = .004 1.86 (1.43-2.42), P < .001
Pericarditis, DNMT3A CHIP 1.88 (1.45-2.43), P < .001 1.66 (1.17-2.35), P = .005
Myocarditis, large CHIP 1.53 (0.92-2.54), P = .10 fixed effect 1.30 (0.81-2.09), P = .27; or, since I² ā‰ˆ 64%, random effects 1.04 (0.38-2.86), P = .93
Myocarditis, DNMT3A CHIP 1.30 (0.81-2.09), P = .28 1.57 (0.92-2.67), P = .10

The authors should confirm these values from the unrounded cohort estimates. In the Results, the first paragraph would then read approximately:

Large CHIP (variant allele fraction ≄10%) had a higher point estimate (HR, 1.86 [95% CI, 1.43-2.42]; I2 = 0). DNMT3A CHIP was also associated with pericarditis (HR, 1.66 [95% CI, 1.17-2.35]; I2 = 0)

The myocarditis sentence would be updated in the same way. The authors may also wish to say which model (fixed or random effects) was applied to the large-CHIP myocarditis rows, given an I² of about 64%, and to check the reported myocarditis I² range.