Research note

A Reversed Trial Entry Weakens a Heart-Failure Fluid-Restriction Meta-Analysis

An independent reproduction of a published mortality meta-analysis finds a reversed death count and a consequential sensitivity to the treatment of a crossover trial.

The Null Institute | Research Note | Updated 17 November 2026
Status: Independent, AI-assisted methodological audit, not externally peer-reviewed. A formal post-publication correction request was submitted to Oxford University Press on 27 October 2026. OUP acknowledged its receipt, but as of 17 November 2026 no substantive response has been received and no published correction has been verified. The publisher has not confirmed our findings. This note is not an official correction or clinical guidance.

The finding

A 2026 meta-analysis in ESC Heart Failure concluded that restricting fluid intake was associated with lower mortality among patients with heart failure: risk ratio (RR) 0.54, 95% confidence interval 0.31–0.94. Our reconstruction of Figure 2A reproduces the published estimate. However, its statistical significance does not survive three distinct diagnostic analyses: changing the apparent Holst death allocation, omitting the Holst crossover trial, or applying modified Hartung–Knapp inference to the original data. The first two test the treatment of one study; the third tests the choice of uncertainty estimator.

The forest plot assigns two deaths to restricted fluid intake and four to liberal intake in the Holst et al. (2008) trial. A separate 2026 systematic review in Heart reports four deaths during restriction and two during more liberal intake. The accessible original-trial abstract confirms that Holst used a randomised crossover design, but the complete original report was not independently checked line by line for the allocation and timing of all six deaths. We therefore describe the difference as an apparent extraction error pending full source-level verification.

What happens when the entry is checked?

Using the eight rows in Figure 2A (seven estimable; Aliti 2013 has zero deaths in both groups), we reconstructed inverse-variance log-risk-ratio pooling with REML between-study variance equal to zero and Hartung–Knapp inference. The original figure was reproduced to the published precision. Python calculations were re-executed on 17 November 2026; a separate R reproduction had also been recorded in the earlier technical audit.

Mortality analysis Pooled RR (95% CI) p-value
Published entries, unchanged 0.544 (0.315–0.940) 0.034
Holst death counts reversed to match its report 0.681 (0.331–1.402) 0.241
Holst excluded entirely 0.553 (0.278–1.098) 0.077

Both sensitivity analyses lose the originally reported threshold of statistical significance. This does not show that fluid restriction is ineffective or harmful. The pooled point estimates remain below one, but their uncertainty no longer excludes no difference.

Two additional checks

Leave-one-out. Removing each of the seven estimable studies in turn yields p > 0.05 in four cases: Albert 2013 (0.0743), Colin Ramirez 2010 (0.1172), Holst 2008 (0.0771), and Paterna 2009 (0.0593). This shows sensitivity of the significance threshold, not a reversal of the estimated effect direction.

Modified Hartung–Knapp. With the published counts, Cochran’s Q = 2.5739 on 6 degrees of freedom. The unmodified Hartung–Knapp variance factor Q/6 = 0.429 is smaller than one, narrowing uncertainty relative to an unscaled inverse-variance estimate. The established modified method floors this factor at one. The pooled RR remains 0.544, but its 95% CI becomes 0.236–1.254 and p = 0.1247. This is a method-choice sensitivity test, not evidence that using the unmodified method was an arithmetic mistake. Methodological reference.

Why simple correction is not enough

Holst was a crossover study: 74 individuals were randomised and 65 completed the trial. Its six deaths were described among non-completers in the earlier technical review. Treating two 65-person treatment periods as two independent randomised arms is therefore unsuitable for an ordinary mortality risk-ratio meta-analysis. Excluding this trial is a conservative diagnostic, not a corrected clinical effect estimate.

Counting participants. Table 1 of the 2026 meta-analysis lists Holst’s sample size as 130, precisely 2 × 65 treatment-period observations, although 74 distinct patients were randomised. The paper reports 1,271 participants across nine trials. Replacing only 130 with 74 gives 1,215, a numerical difference of 56 (4.4%), assuming its other study counts are unchanged. This is a counting audit, not a definitive corrected pooled participant count.

Several other contributing trials also changed salt intake alongside prescribed fluid limits. The pooled estimate should not be read as the isolated causal effect of drinking more or less plain water.

What the audit establishes

We identified a mortality event allocation in Figure 2A that conflicts with reports of the original trial and reproduced its numerical consequences. We have not established a treatment recommendation, a corrected crossover causal estimate, or a complete audit of every trial. The complete 2008 Holst full text was not available for a fresh line-by-line review; its public abstract confirms 74 randomised and 65 completed, while publisher-indexed text and a separate 2026 systematic review support the four-versus-two allocation. Individual timing and crossover sequence data remain unresolved. A formal correction request was submitted to OUP on 27 October 2026, and OUP acknowledged receipt through its submission system. As of 17 November 2026, we have received no substantive reply. No published correction has been independently verified. This procedural status is not evidence for or against the scientific finding.

Data and reproducibility

Download the reproducibility package: ZIP (5.4 KB): source-figure transcription, Python code, scenario results, leave-one-out results, README, source manifest and SHA-256 checks. SHA-256: 969edaba6eb80d684431ad801dbb745819a66854bba06a2a90e972ff18c98efc. The publisher’s original article PDF is not redistributed in the archive.

All Figure 2A rows are given below in the order restricted deaths/total, liberal deaths/total. These are transcribed aggregate entries, not patient-level records. The Aliti row has zero events in both columns and was reported as not estimable.

Albert 2013       1/20     4/26
Aliti 2013        0/38     0/37  (double zero; excluded)
Colin Ramirez     4/76    12/119
Herrmann 2025     2/250    1/254
Holst 2008        2/65     4/65
Machado 2018      2/30     2/23
Paterna 2009      1/52     5/52
Philipson 2013    1/49     1/48

The following self-contained Python code uses SciPy and prints the principal and leave-one-out sensitivity results; it does not fetch data or require access to any private dataset. A versioned analysis package additionally records the extraction table, complete calculations, source manifest, and SHA-256 hashes.

from math import log, exp, sqrt
from scipy.stats import t
D = [
 (1,20,4,26), (4,76,12,119), (2,250,1,254),
 (2,65,4,65), (2,30,2,23), (1,52,5,52), (1,49,1,48)
]
def pool(rows, modified=False):
 y = [log((a/n)/(b/m)) for a,n,b,m in rows]
 w = [1/(1/a-1/n+1/b-1/m) for a,n,b,m in rows]
 k = len(rows)
 mu = sum(a*b for a,b in zip(y,w))/sum(w)
 Q = sum(b*(a-mu)**2 for a,b in zip(y,w))
 q = Q/(k-1)
 se = sqrt((max(q,1) if modified else q)/sum(w))
 tc = t.ppf(.975,k-1)
 return (exp(mu), exp(mu-tc*se), exp(mu+tc*se),
         2*t.sf(abs(mu/se),k-1))
print("As published:", pool(D))
swapped = D.copy()
swapped[3] = (4,65,2,65)
print("Swapped:", pool(swapped))
print("Holst omitted:", pool(D[:3]+D[4:]))
print("Modified HK:", pool(D, True))
for i in range(len(D)):
 print("Leave out", i, pool(D[:i]+D[i+1:]))

Scope note: The swapped case deliberately retains the paper’s two 65-person denominators to isolate the published input-label sensitivity. It does not solve the crossover dependence issue or reconstruct an intention-to-treat mortality contrast.

Primary sources

Transparency and version: OpenAI AI tools assisted with data checking, coding and drafting. Computations were independently re-executed in Python on 17 November 2026; the earlier technical audit also recorded a separate R replication. The original trial full text was not freshly verified line by line. This public Research Note is not an OUP correction or a peer-reviewed clinical study. Findings and data provenance will be updated if new documents or corrections appear. No DOI has been assigned.