Encoding and inference on separable effects for sustained treatments

Ignacio González-Pérez 1 , Kerollos Nashat Wanis 2 , Aaron Leor Sarvet 3 , Mats Julius Stensrud 1 · 2025

Read Encoding and inference on separable effects for sustained treatments by Ignacio González-Pérez 1 , Kerollos Nashat Wanis 2 , Aaron Leor Sarvet 3 , Mats Julius Stensrud 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-10702.txt.

Introduction

Ignacio González-Pérez and co-authors extend a method for pulling apart the effects of a treatment kept up over time. In medicine, what matters is often not just starting a drug, but staying on the plan through follow-up, and one part of that plan may drive a given harm or gain.

The team use a new way to encode long treatment plans, which gives short and clear forms for the claims they must make. Those forms lead to estimators that stay sound if one of two models is right. They apply the method to a large trial of blood pressure care, to weigh one plan's effect on kidney harm.

Catalog record

CREATORIgnacio González-Pérez 1 , Kerollos Nashat Wanis 2 , Aaron Leor Sarvet 3 , Mats Julius Stensrud 1
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE128 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-10702

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Encoding and inference on separable effects for sustained treatments Ignacio González-Pérez 1 , Kerollos Nashat Wanis 2 , Aaron Leor Sarvet 3 , Mats Julius Stensrud 1 Abstract . Sustained treatment strategies are common in many domains, particularly in medicine, where many treatment are delivered repeatedly over time. The effects of adherence to a treatment strategy throughout follow-up are often more relevant to decision-makers than effects of treatment assignment or initiation. Here we consider the separable effect of sustained use of a time-varying treatment. Despite the potential usefulness of this estimand, the theory of separable effects has yet to be extended to settings with sustained treatment strategies. To derive our results, we use an unconventional encoding of time-varying treatment strategies. This allows us to obtain concise formulations of identifying assumptions with better practical properties; for example, they admit frugal graphical representations and formulations of identifying functionals. These functionals are used to motivate doubly robust semiparametrically efficient estimators. The results are applied to the Systolic Blood Pressure Intervention Trial (SPRINT), where we estimate a separable effect of modified blood pressure treatments on the risk of acute kidney injury. Keywords: separable effects, sustained effects, sustained treatment encoding, adherence. 1. Introduction Treatments are often administered over time. Even in conventional randomized trials, where treatment strategies are assigned at baseline, the actual administration of treatment typically occurs sequentially. For example, the Systolic Blood Pressure Intervention Trial (SPRINT) [ 1 ] randomly assigned individuals to receive intensive or standard blood pressure therapy. While the assignment occurred at baseline, the therapy was meant to be taken daily in each arm. In trials like the SPRINT, where individuals’ long-term behaviours may deviate from their assigned protocols, it is debated whether the effect of assignment, usually termed the “intention-to-treat” effect, is the most relevant quantity for decision-making [ 2 , 3 , 4 , 5 ] . An alternative is a particular sustained effect: an effect defined by interventions where individuals continuously take a single treatment throughout follow-up. We will consider effects of sustained treatment use in settings with competing events. To be concrete, in the SPRINT example, consider the objective of estimating the effect of intense blood pressure therapy on the risk of acute kidney injury (AKI), one of the trial’s secondary outcomes. Because AKI only occurs in individuals that are alive, death acts as a competing event. In this trial, the investigators found that the cumulative incidence of AKI is higher in the intensive blood pressure therapy arm compared to the control arm. Interpreting this finding as an “effect” of intensive blood pressure therapy on AKI etiology requires reasoning about causal mechanisms; the problem is that the drug can effect AKI in ways that may or may not be of interest. On the one hand, intensive blood pressure therapy may prevent AKI by binding directly to receptors in the kidneys. On the other hand, intensive blood pressure therapy may cause AKI by preventing fatal events; when an individual survives such events, they might go on to experience an AKI that would otherwise be “prevented” by death. The confluence of these mechanisms complicates the interpretation of an overall effect of blood pressure therapy on AKI. Investigators have confronted these challenges by considering different estimands. These estimands, however, have their own interpretive complications: for example, the controlled direct effect considers unrealistic interventions on the the competing event, here death, [ 6 ] or the survivor average causal effect [ 7 ] condition on an unmeasurable sub-population defined by cross-world counterfactuals. Here we will consider separable effects [ 6 , 8 , 9 , 10 ] . The motivation for these effects is to formalize notions of causal mechanisms that are practically relevant. For example, investigators might be interested in the notion of “effects not through the competing event”. Separable effects formalize this notion by defining modified versions of the original treatment. These modified treatments might be assumed to affect outcomes through specific pathways, which we will represent in causal graphs; in this case, they represent hypothetical refined medications that aim to isolate certain beneficial effects of the treatment of interest. By requiring an explicit, possibly unrealized modification of the treatment, separable effects help clarify the causal question under investigation and may also stimulate new ideas on how treatments could be improved in practice [ 8 , 9 , 10 , 11 , 12 ] . Separable effects could, at least in principle, be directly computed in future clinical trials. Moreover, in observational data they usually rest on weaker assumptions for identification than previously proposed estimands in these settings [ 6 , 7 ] . So far separable effects have been considered in various settings with a single treatment assignment and failure time outcomes [ 6 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 ] . Alternatively, a sustained separable effect would enforce adherence to modified treatments continuously over time. As suggested in [ 12 , Sec. 3.5(a)] , extending existing results on separable effects requires a causal model that incorporates treatment decomposition at each time point. The complexity of directly applying this causal model has challenged theoretical development for these important estimands [ 12 ] . We overcome these challenges by reformulating sustained separable effects via an alternative encoding, which is similar to an encoding used in the literature on the sustained effects of dynamic strategies [ 18 , 19 , 20 , 21 ] . Therein, the actual treatments an individual takes are entirely summarized by a variable denoting adherence to a protocol of interest over time. This encoding provides a parsimonious representation of assumptions and identifying functionals, which is particularly useful in the separable effects setting. The resulting expressions closely resemble those from the simpler point-treatment setting [ 8 , 9 ] , which reduce to special cases of our formalism. The encodings are discussed in more detail in Section 2 . Then, in Section 3 we explicitly introduce the observed data structure for a longitudinal setting with competing events. In Section 4 we define the separable effects on the event of interest as contrasts of counterfactual hazards, and discuss their causal interpretation. Sufficient identification conditions are studied in Section 5 . This is followed by different estimation techniques in Section 6 . The properties of these estimators are analysed in Section 7 , where we perform a simulation study to assess coverage of bootstrap-based confidence intervals and robustness properties. In Section 8 , we apply our methods to the SPRINT, providing for the first time a sustained separable effects analysis of the results of this RCT. 2. A comparison of different treatment encodings Sustained treatment strategies are routinely encoded by random variables that specify the treatment taken by an individual at each point in time. Foundational work on time-varying treatments [ 7 ] introduced such encodings, which are now standard in causal inference textbooks [ 22 , 23 ] , and widely used in clinical trials [ 24 , 25 ] , mediation analyses [ 26 , 27 , 28 , 29 ] , observational studies [ 18 , 20 ] , and the dynamic treatment regime literature [ 30 , 31 ]

Read the complete text