Paperonium

โ† Back to the plain-language case

Abstract cites non-significant interaction odds ratios as evidence that correct fluid significantly reduced mortality

Lancet Child Adolesc Health (2018) ยท doi:10.1016/S2352-4642(18)30130-5 ยท PMC6004535

Summary

The abstract states that correct fluid prescription "significantly reduced the risk of early mortality (within 2 days) in all subgroups". It supports this with five adjusted odds ratios (AORs): 1ยท23, 0ยท95, 1ยท70, 1ยท50 and 0ยท86. These five values, with their CIs, are the "Multiplicative interaction" column of Table 2, which the table footnote defines as OR11/(OR10 ร— OR01), and they can be recomputed from the table's other cells. They measure how much the fluid association differs between children with and without a clinical sign. They are not the effect of correct fluid in either group. All five 95% CIs include 1 (approximate p = 0.50, 0.87, 0.07, 0.22, 0.58), and three of the five point estimates are above 1. So the statistics the abstract cites do not show a significant reduction in any subgroup. The same table, read for the correct-versus-wrong contrast within each stratum, does point to lower odds of early death with correct fluid in every subgroup. The headline direction may therefore hold, but the estimates, their label and the significance claim in the abstract's key fluid result are wrong as printed. Table 2 also contains two cell errors in its additive-interaction column.

What the paper reports

Abstract, Findings:

Correct fluid prescription significantly reduced the risk of early mortality (within 2 days) in all subgroups: abnormal respiratory signs (AOR 1ยท23, 0ยท68โ€“2ยท24), abnormal circulatory signs (0ยท95, 0ยท53โ€“1ยท73), pallor (1ยท70, 0ยท95โ€“3ยท02), dehydration signs only (1ยท50, 0ยท79โ€“2ยท88), and abnormal neurological signs (0ยท86, 0ยท51โ€“1ยท48).

Abstract, Interpretation:

the prescription of recommended rehydration guidelines reduces risk of death

Research in context, Added value of this study:

Our findings also show that comorbidities are common and that correct fluid prescription is associated with reduced risk of death in these patients.

Methods, Data analysis (third paragraph, on effect modification):

We calculated the relative excess odds due to interaction, the attributable proportion, and the multiplicative interaction odds ratio (OR).

Results (effect-modification paragraph), where the same numbers appear under their correct name:

In fact, positive multiplicative interactions (OR>1) for association of incorrect fluid management with death in children with dehydration signs (OR 1ยท50, 95% CI 0ยท79โ€“2ยท88), respiratory signs (OR 1ยท23, 0ยท68โ€“2ยท24), and pallor (OR 1ยท70, 0ยท95โ€“3ยท02) are evidence of interaction (figure 3).

Table 2, "Interactions of fluid management with risk factors for in-hospital early deaths" (columns: Multiplicative interaction | Fluid wrong and symptom present (OR10) | Fluid right and symptom present (OR11) | Fluid right and symptom absent (OR01) | RERI | Attributable proportion):

Pallor | 1ยท70 (0ยท95 to 3ยท02) | 2ยท16 (1ยท66 to 2ยท82) | 1ยท07 (0ยท69 to 1ยท67) | 0ยท29 (0ยท21 to 0ยท41) | -0ยท38 (-1ยท07 to 0ยท31) | -0ยท35 (-1ยท09 to 0ยท38)

Circulatory | 0ยท95 (0ยท53 to 1ยท73) | 1ยท92 (1ยท47 to 2ยท51) | 0ยท64 (0ยท40 to 1ยท03) | 0ยท35 (0ยท25 to 0ยท50) | -0ยท63 (-1ยท19 to -0ยท06) | -0ยท98 (-2ยท06 to 0ยท11)

Dehydration signs only | 1ยท50 (0ยท79 to 2ยท88) | 0ยท96 (0ยท70 to 1ยท33) | 0ยท38 (0ยท26 to 0ยท55) | 0ยท26 (0ยท15 to 0ยท45) | 0ยท25 (-0ยท18 to 0ยท48) | 0ยท40 (-0ยท53 to 1ยท33)

Neurological | 0ยท86 (0ยท51 to 1ยท48) | 3ยท57 (2ยท75 to 4ยท64) | 1ยท15 (0ยท77 to 1ยท72) | 0ยท37 (0ยท25 to 0ยท54) | -1ยท80 (-2ยท71 to -0ยท88) | -1ยท56 (-2ยท61 to -0ยท52)

Respiratory | 1ยท23 (0ยท68 to 2ยท24) | 3ยท18 (2ยท35 to 4ยท31) | 1ยท17 (0ยท80 to 1ยท73) | 0ยท30 (0ยท18 to 0ยท49) | -1ยท31 (-2ยท16 to -9ยท45) | -1ยท11 (-1ยท90 to -0ยท33)

Table 2 footnote (excerpts):

Interaction exists if RERI is not zero (RERI=OR11โ€ˆโ€“โ€ˆOR10โ€ˆโ€“โ€ˆOR01โ€ˆ+โ€ˆ1).

Multiplicative interaction=OR11/(OR10โ€ˆร—โ€ˆโ€ˆโ€ˆOR01).

The reference group was fluid wrong and symptom absent (OR00)=1ยท00.

Discussion, final paragraph:

Importantly, prescription of recommended rehydration guidelines reduces risk of death in this group of children.

What the paper's own numbers show

Method. I compared the five abstract values with each column of Table 2. Using the footnote formulas, I recomputed the multiplicative interaction OR, the RERI and the attributable proportion (AP) from the three cell ORs in each row (OR10, OR11, OR01; OR00 = 1). For each printed 95% CI, I took the standard error of log(OR) as (ln upper โˆ’ ln lower)/(2 ร— 1.96) and derived an approximate two-sided Wald p value. Within the "symptom present" stratum, the effect of correct versus wrong fluid is OR11/OR10, because both cells share the reference OR00. Within the "symptom absent" stratum it is OR01. The paper does not report a CI for OR11/OR10. I approximated one by adding the two variances on the log scale and ignoring their covariance. Both estimates are contrasts against the same reference group, so their covariance is usually positive, and in that case the approximate interval is wider than the true one. Every number below is printed by recompute.py.txt.

Quantity Location Reported Recomputed from the paper's numbers
Respiratory "AOR" for correct fluid Abstract; Table 2, multiplicative interaction 1ยท23 (0ยท68โ€“2ยท24) Identical to Table 2 interaction OR; OR11/(OR10ยทOR01) = 1.17/(3.18 ร— 0.30) = 1.23; CI includes 1; p โ‰ˆ 0.50
Circulatory "AOR" Abstract; Table 2 0ยท95 (0ยท53โ€“1ยท73) Identical to interaction OR; 0.64/(1.92 ร— 0.35) = 0.95; CI includes 1; p โ‰ˆ 0.87
Pallor "AOR" Abstract; Table 2 1ยท70 (0ยท95โ€“3ยท02) Identical to interaction OR; 1.07/(2.16 ร— 0.29) = 1.71; CI includes 1; p โ‰ˆ 0.07
Dehydration signs only "AOR" Abstract; Table 2 1ยท50 (0ยท79โ€“2ยท88) Identical to interaction OR; 0.38/(0.96 ร— 0.26) = 1.52 (rounding of the inputs allows 1.47โ€“1.58); CI includes 1; p โ‰ˆ 0.22
Neurological "AOR" Abstract; Table 2 0ยท86 (0ยท51โ€“1ยท48) Identical to interaction OR; 1.15/(3.57 ร— 0.37) = 0.87 (rounding allows 0.85โ€“0.89); CI includes 1; p โ‰ˆ 0.58
Cited CIs that exclude 1 Abstract "significantly reduced โ€ฆ in all subgroups" 0 of 5; 3 of 5 point estimates are above 1
Correct vs wrong fluid, symptom present (OR11/OR10) Derived from Table 2 not reported Respiratory 0.37 (โ‰ˆ0.23โ€“0.60); circulatory 0.33 (โ‰ˆ0.19โ€“0.57); pallor 0.50 (โ‰ˆ0.30โ€“0.83); dehydration only 0.40 (โ‰ˆ0.24โ€“0.65); neurological 0.32 (โ‰ˆ0.20โ€“0.52). Approximate CIs, covariance ignored
Correct vs wrong fluid, symptom absent (OR01) Table 2 0ยท30, 0ยท35, 0ยท29, 0ยท26, 0ยท37 Same values; all five printed CIs exclude 1
Identity linking the two Table 2 footnote interaction = OR11/(OR10ยทOR01) (OR11/OR10) / OR01 equals the interaction OR in every row: the interaction OR is the ratio of the two within-stratum fluid effects
Dehydration signs only, RERI Table 2 0ยท25 (โˆ’0ยท18 to 0ยท48) OR11 โˆ’ OR10 โˆ’ OR01 + 1 = 0.38 โˆ’ 0.96 โˆ’ 0.26 + 1 = 0.16; the printed CI's midpoint is 0.15
Dehydration signs only, AP Table 2 0ยท40 RERI/OR11 = 0.16/0.38 = 0.42 with the recomputed RERI; 0.25/0.38 = 0.66 with the printed RERI
RERI and AP, other four rows Table 2 e.g. โˆ’0ยท38, โˆ’0ยท63, โˆ’1ยท80, โˆ’1ยท31 Recompute to โˆ’0.38, โˆ’0.63, โˆ’1.79, โˆ’1.31; the AP CI midpoints in all four rows, and the RERI CI midpoints for circulatory, pallor and neurological, match the point estimates. The respiratory RERI CI midpoint (โˆ’5.81) does not (see next row)
Respiratory RERI CI Table 2 โˆ’1ยท31 (โˆ’2ยท16 to โˆ’9ยท45) Upper limit (โˆ’9.45) is below both the point estimate and the lower limit, which is impossible; a symmetric interval around โˆ’1.31 would end near โˆ’0.46

So the abstract's numbers are the Table 2 interaction ORs. The cell ORs in each row reproduce them, so the column labels are internally consistent. None of the five is significant. The contrast the abstract's sentence describes (correct versus wrong fluid within a subgroup) is a different quantity. Table 2 points to it being below 1 in all five subgroups, in both strata, but the paper does not report it.

Figure

Abstract-cited "AORs" compared with the correct-versus-wrong fluid odds ratio implied by Table 2, by subgroup

Figure 1. One log-scale odds-ratio axis; the vertical line marks OR = 1. Orange circles and lines are the five values the abstract cites as AORs for correct fluid prescription, with their printed 95% CIs. They are identical to Table 2's multiplicative-interaction column. Each crosses 1, and the approximate p value is printed beside it. Blue squares are the odds ratio for correct versus wrong fluid among children with the sign present, recomputed from the same Table 2 row as OR11/OR10. Their intervals are approximate because the covariance of the two estimates is not reported and is ignored here. All five blue estimates lie below 1. The abstract's sentence describes the blue quantity but cites the orange numbers.

RERI cells of Table 2: printed values and CIs compared with the footnote formula

Figure 2. Table 2's RERI column on a linear axis; the vertical line marks 0. Orange: RERI and 95% CI as printed. Blue squares: OR11 โˆ’ OR10 โˆ’ OR01 + 1 from the same row. Gray ticks: midpoint of the printed CI. In three rows (circulatory, pallor, neurological) the three agree. For dehydration signs only, the printed point (0.25) differs from both the formula (0.16) and the CI midpoint (0.15). For respiratory signs, the printed point (โˆ’1.31) matches the formula, but the printed interval runs from โˆ’2.16 to โˆ’9.45, so the point lies outside it and the CI midpoint (โˆ’5.81) is far from it.

Why it matters

The effect of correct fluid prescription is one of the paper's two stated aims and is named in its title. The abstract's Findings sentence is the only place in the summary that quantifies it, and the Interpretation ("reduces risk of death") and the closing Discussion sentence are the abstract-level and concluding statements of it. The body gives one other quantitative statement: "The predicted probability of early in-hospital death was reduced by 6% when derived from model II with correct fluid prescription compared with when fluid prescription is wrong" (Results, effect-modification paragraph). That is an overall model prediction, not a subgroup estimate, and it is printed without a CI or test, so it does not support "significantly โ€ฆ in all subgroups". A reader of the abstract alone would take the five AORs as subgroup treatment effects. Three are above 1 and none excludes 1, so the reader would see a contradiction or would misread interaction estimates as effects of fluid prescription. Meta-analysts and guideline reviewers who extract the abstract figures would record five non-significant ORs, some pointing to harm, for a result the paper calls a significant reduction. The Results paragraph also calls three non-significant interaction ORs "evidence of interaction". The data do not support that claim on the multiplicative scale. On the additive scale, the RERI CIs for pallor and dehydration signs only include 0, and the respiratory RERI CI is unusable as printed. The paper's own cells point to a consistent protective association of correct fluid. The defect is therefore in which estimates are reported and how they are described, not necessarily in the direction of the conclusion. The magnitude and precision of the subgroup effects cannot be checked from what is published.

Explanations considered

Status

As of 2026-10-23:

Suggested correction

  1. In the abstract Findings, replace the five interaction ORs with estimates of the quantity the sentence describes. One option is the adjusted OR for correct versus wrong fluid within each subgroup (OR11/OR10, with a CI from the fitted model including the covariance term). Another is the overall adjusted OR for correct fluid. Alternatively, keep the interaction ORs but describe them accurately, for example: "There was no evidence that the association of correct fluid prescription with early mortality differed by clinical sign on the multiplicative scale (interaction ORs 0ยท86โ€“1ยท70, all 95% CIs including 1)." Remove "significantly" and "in all subgroups" unless subgroup-specific estimates with CIs excluding 1 are reported.
  2. In the Results effect-modification paragraph, revise "are evidence of interaction" for the three interaction ORs whose CIs include 1.
  3. In Table 2, correct the dehydration-signs-only RERI. The footnote formula gives 0.16, consistent with the printed CI midpoint (0.15) and the printed AP (0.40, against 0.42). Correct the respiratory RERI CI, whose printed upper limit (โˆ’9ยท45) cannot be right; a symmetric interval would end near โˆ’0.46.
  4. Consider adding a column to Table 2 with the within-stratum correct-versus-wrong fluid OR and its CI for each sign, so readers can check the abstract's claim directly.