Paperonium

โ† Back to the plain-language case

Abstract's adjusted odds ratios for education and prenatal care differ from the paper's own Results

Lancet Reg Health Am (2026) ยท doi:10.1016/j.lana.2026.101434 ยท PMC12995907

Summary

The Abstract reports that higher maternal education (aOR 2ยท41, 95% CI 2ยท35โ€“2ยท48) and receipt of prenatal care (aOR 2ยท64, 95% CI 2ยท56โ€“2ยท73) were associated with documented adequate partner treatment. The paper's national multivariable model, given in the Results text and in Fig. 2, has different values for the same two associations: 1ยท48 (1ยท39โ€“1ยท59) for higher vs primary education and 3ยท05 (2ยท92โ€“3ยท20) for prenatal care. In both cases the two intervals do not overlap. The race estimates in the same Abstract sentence match the Results exactly. Neither Abstract value appears anywhere else in the paper, and the Abstract's education interval is too narrow to be possible for a higher-vs-primary contrast given the Table 1 counts. Both associations keep their direction and statistical significance, but the Abstract overstates the education association, understates the prenatal-care association, and presents the two as about equal, while the Results show prenatal care as much the stronger of the two.

What the paper reports

Abstract, Findings:

Higher maternal education (aOR 2ยท41, 95% CI 2ยท35โ€“2ยท48) and receipt of prenatal care (aOR 2ยท64, 95% CI 2ยท56โ€“2ยท73) were associated with partner treatment, whereas odds were lower among Black (aOR 0ยท85, 95% CI 0ยท82โ€“0ยท89) and Mixed women (aOR 0ยท81, 95% CI 0ยท79โ€“0ยท84) and in the North region.

Abstract, Methods:

Multivariable logistic regression estimated adjusted odds ratios (aORs) and 95% CIs.

Results, paragraph on the national multivariable model:

Multivariable analysis included 214,259 participants, with 4069 exclusions due to missing maternal age. Prenatal care was the factor most strongly associated with adequate partner treatment (adjusted odds ratio [aOR] 3ยท05, 95% CI 2ยท92โ€“3ยท20; p < 0ยท001).

Higher maternal education remained positively associated with partner treatment (higher vs primary: aOR 1ยท48, 1ยท39โ€“1ยท59; secondary vs primary: aOR 1ยท23, 1ยท20โ€“1ยท27; both p < 0ยท001)

Compared with White women, lower odds persisted among Black (aOR 0ยท85, 0ยท82โ€“0ยท89; p < 0ยท001) and Mixed women (aOR 0ยท81, 0ยท79โ€“0ยท84; p < 0ยท001)

Fig. 2 caption:

Unadjusted ORs are from separate univariable logistic regression models; aORs are from the multivariable logistic regression model including maternal age, race, education, prenatal care, and macro-region.

Fig. 2 itself is an image. Read from the figure file (gr2.jpg): pre-natal care "Yes", unadjusted 3ยท33 (3ยท22 โ€“ 3ยท46) and adjusted 3ยท05 (2ยท92 โ€“ 3ยท20); "Higher education", unadjusted 1ยท85 (1ยท73 โ€“ 1ยท97) and adjusted 1ยท48 (1ยท39 โ€“ 1ยท59). Fig. 3 (gr3.jpg), the region-stratified models, gives prenatal care "Yes" aORs of 2ยท11, 2ยท58, 3ยท70, 4ยท50 and 4ยท87 and higher-education aORs of 1ยท39, 1ยท72, 1ยท48, 1ยท38 and 1ยท41 (North, Northeast, Central-West, Southeast, South).

Discussion, second paragraph:

Prenatal care emerged as the strongest predictor of partner treatment, with approximately threefold higher odds compared with women without prenatal follow-up.

Table 1 (columns: all cases, n = 218,328 | adequate partner treatment, n = 50,609 | no adequate partner treatment, n = 167,719 | p-value), education and prenatal-care rows:

Primary | 94,267 (43ยท2%) | 19,367 (38ยท3%) | 74,900 (44ยท7%) | <0ยท001

Higher | 4427 (2ยท03%) | 1431 (2ยท83%) | 2996 (1ยท79%) | <0ยท001

No | 32,162 (14ยท7%) | 2627 (5ยท2%) | 29,535 (17ยท6%) | <0ยท001

Yes | 180,009 (82ยท4%) | 46,943 (92ยท8%) | 133,066 (79ยท3%) | <0ยท001

What the paper's own numbers show

Method. I compared each Abstract estimate with every other estimate of the same association in the paper: the Results text, Fig. 2 (adjusted and unadjusted) and Fig. 3 (five regional models). From the Table 1 counts I computed crude odds ratios with Woolf 95% CIs. From each printed 95% CI I recovered the implied standard error of the log odds ratio, SE = ln(upper/lower)/(2 ร— 1ยท96), together with the largest SE the two-decimal rounding allows. I then compared those SEs with a lower bound that holds for any maximum-likelihood logistic model fitted to these cases with model-based standard errors. When the model contains the levels of a categorical variable, the fitted probabilities within each level sum to the observed number of events in that level. As a result, the variance of the log odds ratio between two levels can be no smaller than the crude Woolf variance 1/a + 1/b + 1/c + 1/d computed from the level counts, whatever other covariates are in the model. Any subset of the cases, such as the n = 214,259 model sample or a single region, has smaller counts, so the bound only increases. Finally I searched every merge of levels of the model variables in Table 1 for a crude OR near 2ยท41 or 2ยท64. Every number below is printed by recompute.py.txt.

Quantity Location Reported Recomputed from the paper's numbers
Higher education, adjusted Abstract 2ยท41 (2ยท35โ€“2ยท48) Appears nowhere else in the paper
Higher vs primary education, adjusted Results text; Fig. 2 1ยท48 (1ยท39โ€“1ยท59) Does not overlap the Abstract interval; the Abstract OR is 1ยท63 times larger (2ยท24 times on the log scale)
Higher vs primary education, crude Fig. 2 (unadjusted); Table 1 counts 1ยท85 (1ยท73โ€“1ยท97) 1ยท85 (1ยท73โ€“1ยท97) from 1431/2996 vs 19,367/74,900, matching Fig. 2, so the Abstract value is not the crude OR either
SE(log OR) implied by the education interval Abstract 2ยท35โ€“2ยท48 0ยท0137 (at most 0ยท0148 allowing for rounding)
SE(log OR) implied by the education interval Results; Fig. 2 1ยท39โ€“1ยท59 0ยท0343
Smallest SE(log OR) possible for higher vs primary Table 1 counts โ€” 0ยท0331. The Abstract interval is 0ยท41 times this. Even counting only the 1431 treated women with higher education, SE โ‰ฅ 0ยท0264; the Abstract interval would need at least 4570 such women
Prenatal care, adjusted Abstract 2ยท64 (2ยท56โ€“2ยท73) Appears nowhere else in the paper
Prenatal care yes vs no, adjusted Results text; Fig. 2 3ยท05 (2ยท92โ€“3ยท20) Does not overlap the Abstract interval; the Abstract OR is 0ยท87 times the Results value
Prenatal care yes vs no, crude Table 1 counts โ€” 3ยท97 (3ยท81โ€“4ยท13)
SE(log OR) implied by the prenatal interval Abstract 2ยท56โ€“2ยท73 0ยท0164 (at most 0ยท0174 allowing for rounding)
SE(log OR) implied by the prenatal interval Results; Fig. 2 2ยท92โ€“3ยท20 0ยท0234
Smallest SE(log OR) possible for prenatal yes vs no Table 1 counts โ€” 0ยท0211. The Abstract interval is 0ยท78 times this. Even pooling "Ignored/Unknown" with "No", the bound is 0ยท0182
Regional aORs, prenatal care Fig. 3 2ยท11, 2ยท58, 3ยท70, 4ยท50, 4ยท87 None is 2ยท64. The closest, Northeast 2ยท58 (2ยท37โ€“2ยท80), has SE 0ยท0425, about 2.6 times the Abstract's
Regional aORs, higher education Fig. 3 1ยท39, 1ยท72, 1ยท48, 1ยท38, 1ยท41 None is near 2ยท41
Black and Mixed vs White, adjusted Abstract vs Results 0ยท85 (0ยท82โ€“0ยท89) and 0ยท81 (0ยท79โ€“0ยท84) in both Identical, so the Abstract sentence uses the national model for race
Relative strength, log OR of prenatal care รท log OR of higher education Abstract vs Results 2ยท64 vs 2ยท41; 3ยท05 vs 1ยท48 1ยท10 in the Abstract, 2ยท84 in the Results
Crude OR near 2ยท41 or 2ยท64 from any merge of levels Table 1 (race, region, education, prenatal care) โ€” None within 0ยท06

Table 1 is internally consistent: every variable sums to 50,609 treated and 167,719 untreated cases. The Results and Fig. 2 intervals are compatible with the bound (1ยท04 and 1ยท11 times it), as they should be. The Abstract intervals are not.

One smaller observation is separate from the claim. The Fig. 2 unadjusted OR for prenatal care "Yes", 3ยท33 (3ยท22โ€“3ยท46), is not the Table 1 crude OR for yes vs no (3ยท97). It is reproduced exactly by yes vs "No" and "Ignored/Unknown" pooled: 3ยท33 (3ยท22โ€“3ยท46). The "Ignored/Unknown" row in Fig. 2, 2ยท28 (2ยท11โ€“2ยท47), is reproduced by ignored vs no alone. This suggests that the unadjusted "Yes" estimate may use a different reference group from the rest of the block. It does not affect the adjusted estimates.

Figure

The Abstract and the Results give different adjusted odds ratios for education and prenatal care but the same ones for race

Figure 1. Short excerpts from the Abstract (Findings, left) and from the Results text, whose values Fig. 2 repeats (right), for the four associations the Abstract quantifies. Orange arrows join two reported values of the same association that disagree: higher education 2ยท41 (2ยท35โ€“2ยท48) vs 1ยท48 (1ยท39โ€“1ยท59), and prenatal care 2ยท64 (2ยท56โ€“2ยท73) vs 3ยท05 (2ยท92โ€“3ยท20). In each case the intervals do not overlap. Dashed gray lines join values that are identical: Black women 0ยท85 (0ยท82โ€“0ยท89) and Mixed women 0ยท81 (0ยท79โ€“0ยท84). The excerpts are shortened; the full sentences are quoted above.

The Abstract confidence intervals are narrower than any model fitted to these cases allows

Figure 2. The standard error of the log odds ratio implied by each printed 95% CI, for higher vs primary education (top) and prenatal care yes vs no (bottom), on one linear axis that starts at zero. Orange diamonds are the Abstract intervals and hollow orange circles the Results / Fig. 2 intervals. The short orange bars show the range of SEs consistent with two-decimal rounding of the printed limits. The blue line in each group is the smallest SE any logistic model fitted to these cases can have, computed from the Table 1 counts (0ยท0331 and 0ยท0211). The shaded area to its left is unattainable. Both Results intervals lie just above the bound, and both Abstract intervals lie inside the unattainable area, the education interval at less than half the bound.

Why it matters

The Abstract is the most-read part of the paper and the only place, apart from the Results text and figures, where the paper puts numbers on its main associations. The title, the Research in context panel and the Conclusions describe these associations without numbers. In the Abstract:

  1. The education association is overstated. 2ยท41 is 63% larger than the Results' 1ยท48, and 2ยท24 times larger on the log-odds scale. The Abstract's interval (2ยท35โ€“2ยท48) also suggests a precision that the 4427 women with higher education cannot provide.
  2. The prenatal-care association is understated. 2ยท64 is 13% smaller than 3ยท05. The Discussion's statement of "approximately threefold higher odds" matches the Results, not the Abstract.
  3. The relative ranking is altered. The Abstract presents education and prenatal care as roughly equal (log-OR ratio 1ยท10). In the Results, prenatal care is the strongest predictor, with a log OR 2ยท84 times that of higher education. The Discussion states that "public health prioritization should be guided primarily by effect size", so the size and ranking of these two associations are what a reader would act on.

The direction and statistical significance of both associations are the same in the Abstract and the Results. So are the paper's qualitative conclusions: partner treatment is low, it is more likely with prenatal care and higher education, and it is less likely among Black and Mixed women. The race estimates in the Abstract are correct. The error is in the reported size of two of the four quantified associations in the Abstract.

Explanations considered

Status

Suggested correction

  1. Replace the Abstract's education and prenatal-care estimates with the values from the national multivariable model, stating the education contrast explicitly. For example: "Higher maternal education (higher vs primary: aOR 1ยท48, 95% CI 1ยท39โ€“1ยท59) and receipt of prenatal care (aOR 3ยท05, 95% CI 2ยท92โ€“3ยท20) were associated with partner treatment".
  2. If 2ยท41 (2ยท35โ€“2ยท48) and 2ยท64 (2ยท56โ€“2ยท73) come from another analysis, identify that analysis (model, sample, contrast and CI method) in the Methods and Results, and explain why the Abstract reports it instead of the main model.
  3. Minor, separate: check the reference group of the Fig. 2 unadjusted OR for prenatal care "Yes" (3ยท33), which matches yes vs "No" and "Ignored/Unknown" pooled rather than yes vs no (3ยท97), and state the reference in the figure if it differs from the adjusted estimate's.