Paperonium

← Back to the plain-language case

Abstract reports a fetal HLA-B*39:06 effect at P < 0.001; its own P-value is 0.00384

Int J Epidemiol (2023) Β· doi:10.1093/ije/dyad142 Β· PMC10859143

Summary

The abstract and the Results say that HLA-B39:06 has a fetal allelic effect on birthweight at P < 0.001, the paper's Bonferroni threshold (Ξ± = 0.05/50 = 0.001). The same Results paragraph, and Supplementary Tables S7 and S8, give that fetal P-value as Pf = 3.84 Γ— 10-3. The printed estimate and 95% CI, 0.090 (0.029, 0.151), imply P = 3.83 Γ— 10-3, and the supplement's estimate and SE imply 3.88 Γ— 10-3. So the reported Pf is internally consistent and the "P < 0.001" label is not: the effect is nominally significant but fails the threshold the paper set. This was the only fetal effect of a non-HLA-C allele that the abstract presents as passing the threshold, and no allele in Table S7 has an unconditional fetal P below 0.001. A second, smaller slip: the maternal 95% CI printed for B39:06 is not centred on its estimate.

What the paper reports

Quotations are verbatim from the PMC-archived text. Exponents are written here as superscripts, as they appear in the publisher's XML; the PMC plain text renders them as "10-3".

Abstract, Results:

Applying our SEM to imputed classical HLA alleles and own and offspring BW from the UKB replicated the previously reported association at the HLA-C locus and revealed strong evidence for maternal (HLA-A03:01, B35:01, B39:06, P <0.001) and fetal allelic effects (HLA-B39:06, P <0.001) of non-HLA-C alleles on BW.

Methods, "Application of SEM to BW data in the UK Biobank", on the tests used:

In addition to one degree of freedom tests where we evaluated the significance of the maternal and fetal components individually, we also compared the full model with a constrained model in which we fixed the maternal and fetal effects of the allele of interest to zero (i.e. a two degree of freedom test; Supplementary Note 6, Supplementary Table S6, available as Supplementary data at IJE online).

Results, "Empirical analysis in UK biobank", the significance threshold:

We further set a more stringent threshold for significance using Bonferroni correction for 50 tests (Ξ± = 0.05/50 = 0.001), on the rationale that the locus with the most coded alleles (HLA-B, 25 alleles) involved 50 statistical tests (i.e. one maternal and one fetal effect for each coded allele).

Same paragraph, next sentence:

Three of the alleles exhibited significant evidence for maternal effects, with P-values less than 0.001 (A03:01, B35:01, B39:06, labelled in Figure 4), and the HLA- B39:06 allele exhibited a significant fetal effect at P <0.001 (labelled in Figure 4).

Same paragraph, two sentences later:

The allele with the largest estimated maternal and fetal effects was B*39:06, which had a relatively low allele frequency of 0.69% and consequently wide confidence intervals for the effect sizes [maternal effect = -0.111 (-0.151, -0.029), Pm = 7.10 Γ— 10-4; fetal effect = 0.090 (0.029, 0.151), Pf = 3.84 Γ— 10-3].

Figure 4 legend:

Triangles indicate alleles where tests for both maternal and fetal effects reached the threshold; solid dots represent alleles which did not (P < 0.05). Alleles with P < 0.001 are explicitly labelled.

Discussion, fifth paragraph (first and last sentences):

The HLA-B*39:06 allele displayed the largest estimated maternal and fetal effects among all associated alleles.

These associations could be partly attributable to HLA-B*39:06, as our results indicate a positive fetal and a negative maternal effect of this allele on BW.

Supplementary Table S7, row B*3906 (columns: frequency; maternal effect, SE, P; fetal effect, SE, P):

B*3906 | 0.69% | -0.1108 | 0.0327 | 7.10E-04 | 0.0901 | 0.0312 | 3.84E-03

What the paper's own numbers show

Method. I treated the paper's 1-df maternal and fetal tests as Wald tests: two-sided P = 2Ξ¦(βˆ’|Ξ²/SE|), with the SE taken either from Table S7 or from the printed 95% CI as (upper βˆ’ lower)/(2 Γ— 1.96). Each P was compared with the paper's own threshold, 0.001. To check that this approximates the paper's SEM tests, I applied it to the other effects quoted in the Results; it reproduces their P-values to within about 3% (for example A03:01 maternal 8.98 Γ— 10-8 against a reported 9.21 Γ— 10-8, and DRB111:04 fetal 1.36 Γ— 10-3 against 1.34 Γ— 10-3). Every number below is printed by recompute.py.txt.

Quantity Location Reported Recomputed from the paper's own numbers
Bonferroni threshold Results Ξ± = 0.05/50 = 0.001 0.0010
B*39:06 fetal P, as labelled Abstract; Results "P <0.001" not supported (next rows)
B*39:06 fetal P, as printed Results; Tables S7, S8 Pf = 3.84 Γ— 10-3 3.84 times the threshold; Bonferroni-adjusted P = 50 Γ— Pf = 0.192
B*39:06 fetal P from the printed CI Results: 0.090 (0.029, 0.151) β€” SE 0.0311, z = 2.89, P = 3.83 Γ— 10-3
B*39:06 fetal P from Ξ² and SE Table S7: 0.0901, SE 0.0312 3.84E-03 z = 2.89, P = 3.88 Γ— 10-3; 95% CI (0.029, 0.151), matching the text
B*39:06 fetal P, one-sided derived from Table S7 β€” 1.94 Γ— 10-3, still above 0.001
Fetal estimate needed for P < 0.001 at SE 0.0312 derived β€” |z| β‰₯ 3.29, so Ξ² β‰₯ 0.103 (the estimate is 0.090)
B*39:06 fetal 99.9% CI (the interval matching Ξ± = 0.001) derived from Table S7 β€” (βˆ’0.013, 0.193), which includes zero
B*39:06 maternal P Results; Table S7: βˆ’0.1108, SE 0.0327 Pm = 7.10 Γ— 10-4 7.03 Γ— 10-4, below 0.001; the maternal claim holds
B*39:06 maternal 95% CI Results βˆ’0.111 (βˆ’0.151, βˆ’0.029) (βˆ’0.175, βˆ’0.047) from Table S7. The printed interval's midpoint is βˆ’0.090, not βˆ’0.111; it is the fetal interval (0.029, 0.151) with the signs reversed
Alleles with unconditional fetal P < 0.001 Table S7 (all 19 alleles with a maternal or fetal P < 0.05) the text names B*39:06 none; the smallest fetal P is DRB1*11:04, 1.34 Γ— 10-3
Alleles with maternal P < 0.001 Table S7 A03:01, B35:01, B*39:06 A03:01, B35:01, B*39:06 (agrees)
B*39:06 fetal P, conditional on rs9366778 and rs6911024 Table S8 6.04E-04 6.13 Γ— 10-4, below 0.001 (a different model; see Explanations)
B*39:06 maternal P, conditional Table S8 1.67E-02 1.66 Γ— 10-2, above 0.001
B*39:06 2-df joint test (maternal and fetal both zero) Table S6; Supplementary Note 7.65E-04 (Table S6); 7.56 Γ— 10-4 (Note) below 0.001, but a test of any effect, not of the fetal effect

Two independent routes, from the printed CI and from the supplement's estimate and SE, both give a fetal P of about 0.0038, matching the printed Pf = 3.84 Γ— 10-3. Only the label "P <0.001" in the abstract and the Results disagrees with it. A one-sided test would not close the gap. The maternal P for B*39:06 (7.10 Γ— 10-4) is below 0.001, so the abstract's list of maternal effects is correct. Because Table S7 contains every allele with a nominal maternal or fetal effect, no allele omitted from it could have a fetal P below 0.001.

Figure

The fetal P-value for HLA-B*39:06, as printed and as recomputed, lies to the right of the paper's 0.001 threshold, although the abstract labels it P < 0.001

Figure 1. P-values for HLA-B*39:06 on a log scale; the solid vertical line is the paper's Bonferroni threshold, Ξ± = 0.05/50 = 0.001, and the dashed line is the nominal 0.05. Values left of the solid line pass the threshold. The top row (orange arrow) is the abstract and Results label "P < 0.001" for the fetal effect. The next row (orange) is the fetal P printed in the Results and in Table S7, 3.84 Γ— 10-3. Blue rows are recomputed from the paper's own numbers: from the printed 95% CI (3.83 Γ— 10-3), from the Table S7 estimate and SE (3.88 Γ— 10-3), and as a one-sided test (1.94 Γ— 10-3). All fetal values lie right of the threshold. The two shaded rows show, for comparison, the maternal P (reported 7.10 Γ— 10-4, recomputed 7.03 Γ— 10-4), which does pass it. All values are from the unconditional analysis. The log scale is used because P-values span orders of magnitude; the gap between 0.001 and 0.0038 is a factor of 3.8.

The printed maternal 95% CI for B*39:06 is not centred on its estimate, and the fetal interval at the 0.001 level includes zero

Figure 2. Estimates for HLA-B*39:06 with confidence intervals, on the effect scale used in the paper; zero is the reference line. Orange: as printed in the Results. Blue: computed from the Table S7 estimates and SEs. The printed fetal 95% CI matches the recomputed one exactly. The dashed blue interval is the fetal 99.9% CI, the interval that corresponds to a two-sided test at Ξ± = 0.001; it runs from βˆ’0.013 to 0.193 and includes zero, which shows the same failure to reach the threshold on the effect scale. The printed maternal interval (βˆ’0.151, βˆ’0.029) is centred on βˆ’0.090 (black cross), not on its estimate of βˆ’0.111; the interval implied by Table S7 is (βˆ’0.175, βˆ’0.047).

Why it matters

Explanations considered

Status

Suggested correction

In the abstract, replace

revealed strong evidence for maternal (HLA-A03:01, B35:01, B39:06, P <0.001) and fetal allelic effects (HLA-B39:06, P <0.001) of non-HLA-C alleles on BW.

with wording such as

revealed strong evidence for maternal allelic effects (HLA-A03:01, B35:01, B39:06, P <0.001) and suggestive evidence for a fetal allelic effect (HLA-B39:06, P = 3.8 Γ— 10-3) of non-HLA-C alleles on BW.

In the Results, replace "and the HLA- B39:06 allele exhibited a significant fetal effect at P <0.001 (labelled in Figure 4)" with a statement that no fetal effect reached the Bonferroni threshold and that the B39:06 fetal effect was nominally significant (Pf = 3.84 Γ— 10-3). If the authors wish, they can add that it falls below 0.001 in the analysis conditioned on rs9366778 and rs6911024 (Pf = 6.04 Γ— 10-4, Table S8), identified as such. In the same paragraph, change the maternal interval for B*39:06 from "(-0.151, -0.029)" to "(-0.175, -0.047)", the interval implied by Table S7 (βˆ’0.1108, SE 0.0327), or to whatever interval the authors' fitted model gives.