CRRT dose study correction: why marginal structural and Cox estimates should not be pooled

A correction notice published in Critical Care has clarified two swapped hazard ratio estimates from a cohort study on continuous renal replacement therapy (CRRT) dosing, and the authors have used the opportunity to make a broader statistical argument: that marginal structural model estimates and conventional Cox model estimates answer different clinical questions and should never be combined in a single meta-analysis. The letter, written by Maria Luiza Medeiros Faria and Alexandre Braga Libório, responds to two correspondence pieces that questioned both the reported numbers and the analytical approach behind them.

The numerical issue itself is straightforward. In the original article, the Results text listed two marginal structural model estimates in reverse order. The correct figures are those shown in Figure 3 of the paper. For the 13 mL/kg/h dose threshold, the marginal structural hazard ratio is 1.08, with a 95 percent confidence interval of 0.74 to 1.58. For the 20 mL/kg/h threshold, the hazard ratio is 0.87, with a 95 percent confidence interval of 0.66 to 1.14. Neither estimate is statistically significant, because both confidence intervals cross the null value of 1. The authors apologized for the error and the uncertainty it caused for review authors and readers, and stated that the conclusion of their study is unchanged. They will submit a formal correction to the journal Shock, where the original report appeared.

The more consequential part of the letter concerns what these estimates actually mean. Delivered CRRT dose, the authors argue, is not a fixed attribute assigned at the moment dialysis begins. Clinicians adjust the prescribed dose as treatment proceeds, and the delivered dose drifts in response to variables that change during therapy: acidosis, urea levels, potassium, fluid balance, urine output, hemodynamic instability, circuit clotting and interruptions in the circuit. These same variables predict death, and they are themselves influenced by the dose the patient received earlier. This creates what statisticians call time-dependent confounding affected by prior treatment, a classical setting in which conventional regression cannot provide an unbiased estimate even when covariates are updated over time.

The authors’ own data illustrate just how dynamic the exposure was. Between consecutive three-day blocks of treatment, only about half of patients remained in the same dose category. Patients who started below 13 or 20 mL/kg/h usually had their dose increased above 20 mL/kg/h within the first week. In that context, the authors contend, a whole-course mean dose with adjustment only for baseline characteristics mixes together exposures given under very different clinical conditions, blending measurements taken when patients were relatively stable with those taken during physiological deterioration.

This is where the two modeling traditions diverge. A conventional Cox model with baseline-only adjustment treats the dose as if it were a static exposure measured once at initiation. A marginal structural model, by contrast, uses inverse probability weighting to disentangle the causal pathway between time-varying treatment and time-varying confounders, targeting a different estimand altogether. The authors agree with Lumlertgul and Bagshaw, authors of one of the two letters, that a marginal structural hazard ratio and a conventional Cox hazard ratio target different estimands and should not be pooled together in a meta-analysis.

The authors then add a subtler point about heterogeneity statistics. The pooled estimate of 1.72 that drew criticism combined two estimates obtained with the same analytical approach, namely whole-course mean dose with baseline-only adjustment, and therefore with the same vulnerability to time-dependent confounding. Their agreement, reflected in an I² of 0 percent, shows only that the two estimates are similar to each other, not that they are unbiased. I² measures variability between studies, and a bias shared by both studies leaves it unchanged. Two flawed estimates can agree perfectly while both being wrong in the same direction, a phenomenon that heterogeneity statistics alone cannot detect.

On the clinical question of whether a minimum safe CRRT dose exists, the authors side with both letters in recommending caution. They agree that the 13 mL/kg/h threshold should be presented as an exploratory, estimate-sensitive observation and, in their view, should not be cited as evidence of a minimum safe CRRT dose. The corrected marginal structural estimates, both statistically non-significant, provide no support for a protective or harmful effect at either threshold examined.

The authors point to ongoing randomized trials as the path forward. Three trials, LIMIT (NCT06014801), KETZEREI (NCT06021288) and WISDOM (NCT06446739), are currently running and will, in the authors’ view, be informative on the question of CRRT dosing. Randomized designs, by breaking the link between a patient’s evolving clinical state and the treatment they receive, sidestep the time-dependent confounding that plagues observational analyses of this exposure.

The episode is a compact lesson in why the choice of statistical model is not a technical footnote but a substantive scientific decision. When an exposure changes over time in response to factors that also determine the outcome, the model must be built to handle that feedback loop, or the resulting estimate will absorb the very confounding it was meant to remove. The correction itself is minor, two numbers in reverse order, but the methodological clarification the authors attach to it is not: pooled estimates built on vulnerable analytical approaches inherit that vulnerability, no matter how tightly they cluster.

Subject of Research: Correction of swapped CRRT dose hazard ratios and the argument that marginal structural and conventional Cox model estimates should not be pooled

Article Title: Time-varying dose decisions in CRRT: why marginal structural and baseline Cox estimates should not be pooled

Article References: Faria, M. L. M., & Libório, A. B. (2026). Time-varying dose decisions in CRRT: why marginal structural and baseline Cox estimates should not be pooled. Critical Care, 30(1), Article 566. https://doi.org/10.1186/s13054-026-06440-9

Image Credits: AI Generated

DOI: 10.1186/s13054-026-06440-9

Keywords: CRRT, marginal structural model, Cox proportional hazards, time-dependent confounding, hazard ratio, meta-analysis, dose adjustment, continuous renal replacement therapy, I2 heterogeneity, randomized trials, observational cohort, correction