On the limitations of causal inference with current-treatment Cox models

Mark B. Knudsen, Erin E. Gabriel, Torben Martinussen, Helene C. W. Rytgaard, Arvid Sjölander · 2025

Read On the limitations of causal inference with current-treatment Cox models by Mark B. Knudsen, Erin E. Gabriel, Torben Martinussen, Helene C. W. Rytgaard, Arvid Sjölander (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-2504-01524.txt.

Introduction

Mark Knudsen and four co-authors warn, in a 2025 paper, about a quiet flaw in a popular survival model. Cox models that track only a patient's current treatment are often read as if they showed what would happen under a set treatment plan. That reading needs a Markov property: the risk ahead must not depend on treatment history once the current level is known.

The team show, in continuous time, exactly when that property can hold for both the group a doctor sees and the true group, hidden frailty included. It demands no interplay between treatment and frailty on one scale, an assumption that can never be checked, since frailty is unmeasured by definition. They build a case where the model fits perfectly yet the plan reading still fails.

Catalog record

CREATORMark B. Knudsen, Erin E. Gabriel, Torben Martinussen, Helene C. W. Rytgaard, Arvid Sjölander
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE78 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-01524

Browse the full library · About the Tower of Babel

Full text

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

On the limitations of causal inference with current-treatment Cox models Mark B. Knudsen, Erin E. Gabriel, Torben Martinussen, Helene C. W. Rytgaard, Arvid Sjölander Abstract Cox models with time-varying treatments often include only the current treatment level. Translating such a model into causally meaningful intervention-specific survival probabilities relies on the Markov property: that the hazard is independent of treatment history conditional on current treatment. For the Markov property to not be population-specific, it needs to hold also conditional on any unmeasured prognostic heterogeneity (frailty). Using a discrete-time argument, earlier work concluded that the Markov property can hold both conditionally and marginally only in the absence of a treatment effect or an effect of the unmeasured heterogeneity. We broaden this argument by developing a continuous-time framework that encompasses both proposed extensions of the Kaplan-Meier curve and current-treatment marginal structural Cox models, and sharpen it by characterizing precisely the conditions under which the Markov property can hold both conditionally and marginally. Specifically, we show that it requires the absence of treatment-frailty interaction on the additive hazard scale. As frailty is inherently unmeasured, such a no-interaction assumption cannot be verified. Consequently, causal interpretation of a current-treatment marginal structural Cox model rests on a strong and unverifiable structural assumption. Illustrating this point, we construct a setting in which a Cox model is correctly specified conditionally on current treatment only, although the marginal Markov property fails. Transforming the fitted model to the survival scale does not recover the true intervention-specific survival probabilities. Thus, moving from the hazard scale to the survival scale is not in itself sufficient to obtain a causal interpretation in the time-varying treatment setting. Keywords: causal inference; frailty; Markov property; marginal structural Cox model; time-dependent treatment 1 Introduction Time-to-event methods are commonplace in medical research, and the Cox model ( Cox, 1972 ) is by far the most commonly used regression model among them. The Cox model parametrizes the hazard function, which is the instantaneous rate of events occurring among subjects who have not yet experienced the event. The hazard ratio (HR) comparing treated and untreated is often the primary effect measure reported in studies with a time-to-event outcome. In recent years, however, warnings have been issued about the subtle causal interpretation of the HR, as it mixes a (potentially) real treatment effect with an unavoidable selection effect if the treatment actually has a causal effect ( Hernán, 2010 ; Aalen et al., 2015 ; Martinussen et al., 2020 ) . Due to this, Hernán (2010) recommends supplementing HRs with estimated survival curves for easier causal interpretation. Many recent contributions to the causal time-to-event literature likewise focus on non-hazard estimands such as the probability of survival/death or the restricted mean survival time. In the Cox model with baseline binary treatment, the HR parameter equals the ratio of the log-survival probabilities for the treated and untreated, yielding a probability-based causal interpretation of the HR parameter in the absence of unmeasured confounding ( Martinussen et al., 2020 ) . This interpretation is, however, less straightforward. When treatment is time-varying, Cox models are commonly specified using only the current treatment level as a time-dependent covariate. In order to be a model for the hazard of a potential outcome under a hypothetical intervention on the treatment process, the Markov property needs to be fulfilled: conditional on current treatment, the hazard should be independent of past treatment. The scientific justification for such a marginal Markov property is less straightforward. A mechanistic explanation would naturally be formulated at the individual level, conditional on prognostic characteristics, including unmeasured frailty that affects the outcome but is independent of treatment. For this explanation to support a marginal Markov property, the conditional Markov property must be preserved after marginalizing over this heterogeneity. If conditional Markov fails, a marginal Markov property can only arise through cancellation after averaging over frailty. Its validity then depends on the distribution of frailty in the study population and cannot generally be transported to another population. The difficulty of reconciling the conditional and marginal Markov properties was highlighted by Sjölander (2020) who, in critically examining the causal interpretation of the extended Kaplan-Meier curve of Snapinn et al. (2005) ; Simon and Makuch (1984) , used a discrete-time approach to show that the conditional and marginal Markov property can hold simultaneously only in the absence of an effect of either the treatment or frailty. The former leaves no treatment effect to interpret, whereas the latter rules out any unmeasured prognostic heterogeneity and is therefore implausible in realistic applications. We revisit this problem in continuous time and broaden Sjölander ’s argument by developing a general framework that encompasses proposed Kaplan–Meier extensions, current-treatment Cox models, marginal structural Cox models (Cox MSMs), and multi-state models with an absorbing state. We sharpen the critique by characterizing precisely the conditions under which the Markov property can hold both conditionally and marginally in the continuous-time setting. Specifically, we show that it requires the absence of treatment-frailty interaction on the hazard scale: the frailty-conditional hazard must be additively separable into a frailty component and a treatment component. Frailty may be independent of the treatment process and therefore need not constitute unmeasured confounding. Nevertheless, because it is unmeasured, the required no-interaction condition cannot be verified from the observed data. The result has direct implications for Cox MSMs that include only current treatment ( Hernán et al., 2000 ; Hernán et al., 2001 ) , a specification commonly used in applied research ( Chirgwin et al., 2016 ; Choi et al., 2002 ; Li et al., 2016 ; Nahon et al., 2018 ; Tiihonen et al., 2009 ) . As in the baseline treatment setting, the resulting parameter has no simple causal interpretation as a hazard ratio. Translating the fitted model into intervention-specific survival probabilities does not solve the problem on its own, as it requires the hazard to be correctly specified and therefore relies, under a current-treatment specification, on the Markov property and the associated unverifiable no-interaction condition. The problem therefore concerns the structural assumptions required for causal interpretation, rather than merely the choice of effect scale. The idea that unmeasured frailty can induce dependence on the history of a process, and thus a violation of the Markov property, has been explored previously by Putter and family=Houwelingen (2015) in the context of multi-state models. Some works suggest including a shared multiplicative frailty term in all transition hazards of an illness-death model to account for residual correlation between transition times ( Haneuse and Lee, 2016 ) . However, these works focus on the feasibility and usefulness of frailty models in explaining how a hazard depends on the history. In this work, we view the presence of frailty as a more fundamental issue for the validity of the Markov property. If the Markov property is not satisfied, the current-treatment Cox model may still be correctly specified as a model for the rate, defined as the instantaneous risk of an event conditional on survival and current treatment only. However, this rate has no clear causal interpretation, except if the Markov property holds, in which case it equals the hazard conditional on the entire treatment history. We show that the extended Kaplan–Meier curve estimates the exponential of minus the cumulative rate,

Read the complete text