Erythropoiesis-Stimulating Agents in Dialysis

Why ESA Dose Comparisons Need a Longitudinal Approach

For a patient receiving dialysis, erythropoiesis-stimulating-agent (ESA) dose may be changed over follow-up in response to recent hemoglobin, comorbidity, and clinical status. Those same factors can predict later transfusion. Moreover, an earlier dose can affect later hemoglobin, which then influences the next dose. A comparison of patients who happened to receive high versus low doses can therefore mix treatment effects with differences in the histories that led clinicians to choose those doses.

The question here is not whether one dose is universally preferable. It is how to estimate the marginal risk of transfusion under specified, sustained dosing strategies in a defined dialysis cohort. The strategies must be clinically interpretable and sufficiently represented in the observed data. Dose categories below are illustrative; they are not clinical recommendations.

For example, a fixed strategy might assign the lower dose category at every interval. A dynamic strategy might specify a dose category using the most recent hemoglobin and comorbidity status. The rule and its timing must be defined before analysis: a dose decision cannot respond to a hemoglobin value measured after that decision. In this tutorial, transfusion by the end of follow-up is the endpoint; death before transfusion is treated as a competing event, not as ordinary loss to follow-up. A different handling of death defines a different estimand.

Assumptions and the Weighting Argument

Let \(H_t=(\bar L_t,\bar A_{t-1})\) be the measured information available just before the dose at time \(t\). For a deterministic strategy \(g\), define \(p_t(g_t(H_t)\mid H_t)\) as the conditional probability of receiving the dose that the strategy prescribes, among patients with history \(H_t\). For a fixed strategy, \(g_t(H_t)\) is simply its specified dose at interval \(t\).

The treatment weight for a patient whose observed doses follow \(g\) is the product of the reciprocals of the probabilities of receiving each prescribed dose. It upweights histories in which following the strategy was uncommon. A weighted mean among strategy-followers estimates the risk that would have been seen had the strategy been applied to the target population, subject to the assumptions above.

Here \(I(\cdot)\) is an indicator, and the sum is over the analytic cohort. This normalized, or Hájek, estimator is a finite-sample weighted risk; an unnormalized inverse-probability estimator uses the corresponding weighted outcome total divided by the original cohort size. Both target the same risk under correct identification and probability models, though their finite-sample behavior differs.

To see why, factor the observed-data distribution into the baseline distribution, the conditional distributions of covariates as they evolve, the conditional probabilities of treatment, and the outcome distribution. Restrict attention to records whose treatment at every interval equals the strategy's prescribed dose. Multiplying their contribution by \(W^g\) cancels the treatment-probability factors in that factorization. The resulting weighted distribution has the covariate histories that arise under the strategy, rather than histories selected by the observed dose decisions.

More explicitly, sequential exchangeability allows each observed conditional outcome distribution among patients receiving the prescribed dose, given \(H_t\), to stand in for the corresponding potential-outcome distribution under that dose. Consistency identifies the observed outcome with the potential outcome for strategy-followers. After cancellation of the treatment factors, summing (integrating for continuous covariates) over the baseline and subsequent covariate distributions gives \(E[Y^g]\). Since \(Y^g\) is binary, \(E[Y^g]=P(Y^g=1)\). Positivity ensures the factors being cancelled are nonzero. \(\square\)

This result also explains why prior hemoglobin must be handled with care. It can confound the next dose decision and transfusion, while also being affected by earlier doses. Weighting models the dose assignment conditional on that history; it does not simply remove hemoglobin from the analysis. The introductory material on marginal structural models gives the broader framework for this treatment-confounder feedback.

Worked Dose-Strategy Calculations

Diagnostics, Alternatives, and Limitations

A weight calculation is not evidence that its assumptions hold. Before interpreting a contrast, inspect the treatment probabilities and weights at every interval. Very small fitted probabilities yield large inverse weights; a few such patients can dominate the weighted risks. Examine the distribution of cumulative weights, their maximum and upper quantiles, the effective sample size, and whether weighted covariate histories are balanced across dose decisions. Assess overlap specifically within combinations of prior hemoglobin, comorbidity, and dose history—not just in the full cohort.

If histories with a very low probability of a prescribed dose are common, the comparison relies heavily on extrapolation. Truncating extreme weights may reduce variance but changes the estimator and can introduce bias; report the rule and compare results with the untruncated analysis. A better remedy may be to redefine the target population or strategies to avoid unsupported comparisons, if that revised question is scientifically useful. Correctly fitting a treatment model does not guarantee measured confounding control, and a good balance diagnostic cannot establish that important confounders were measured.

The parametric g-formula is a principal alternative: model the evolving covariates and outcome, then simulate outcomes under each strategy. It also requires adequate measurement of confounders and correct model specification, but it approaches the problem through outcome and covariate models rather than inverse treatment probabilities. Conventional regression that adjusts for current hemoglobin in an outcome model may be inadequate when hemoglobin is both affected by prior dose and a confounder of later dose. The appropriate alternative is not “adjust for nothing”; it is a longitudinal method whose assumptions and target estimand are clear.

An applied analysis also has to specify dose categories or dose rules, the time scale, the eligible population, the transfusion endpoint, and how missing covariates and loss to follow-up are handled. Coarse categories can hide meaningful dose variation; overly fine categories can create sparse histories. Results describe the strategies as defined and the population represented by the data, not an unrestricted biological effect of ESA dose.