Causal treatment effect decompositions with time-to-event outcomes under competing events
Mikko Valtanen, Tommi Härkänen, Jenni Lehtisalo, Tiia Ngandu, Miia Kivipelto, Kari Auranen · 2026
Read Causal treatment effect decompositions with time-to-event outcomes under competing events by Mikko Valtanen, Tommi Härkänen, Jenni Lehtisalo, Tiia Ngandu, Miia Kivipelto, Kari Auranen (2026) 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-2605-19693.txt.
Introduction
Valtanen and team decompose treatment effects for time-to-event outcomes with rival events, where inference is obscured when it shapes the rival event. Their four-way decomposition shows how the treatment's effect on the target event arises through distinct mechanisms involving target and rival events. It uses cross-world measures where treatment affects the two events as if set differently. Exchangeability and consistency assumptions for estimating the split from data are given. The split reveals the rival event's role and grounds interpretable measures, shown on two randomised trials.
Catalog record
| CREATOR | Mikko Valtanen, Tommi Härkänen, Jenni Lehtisalo, Tiia Ngandu, Miia Kivipelto, Kari Auranen |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 64073 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.19693
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Causal treatment effect decompositions with time-to-event outcomes under competing events Mikko Valtanen, Tommi Härkänen, Jenni Lehtisalo, Tiia Ngandu, Miia Kivipelto, Kari Auranen Abstract Inference about treatment effects for time-to-event outcomes is often obscured by the presence of competing events. A particularly complex situation arises when the treatment influences the occurrence of the competing event. A comprehensive assessment should then account for different mechanisms by which the treatment and the competing event together produce the apparent treatment effect. Here, we propose a decomposition of the treatment’s effect on the event of interest (target), characterising how it arises due to four distinct mechanisms involving both the target and competing events. Based on a causal model, the decomposition relies on cross-world estimands reflecting counterfactual scenarios in which the treatment affects the two events as if set to conflicting levels. We specify exchangeability and consistency assumptions under which the decomposition can be estimated from observed data. We discuss how the new decomposition reveals the role of the competing event and serves as a basis for defining causally interpretable estimands in the presence of competing events. Finally, we demonstrate the use of the four-way decomposition with datasets from two randomised trials. 1 Introduction Analyses of time-to-event outcomes in health research are often complicated by competing events that prevent the event of interest (i.e. target event) from occurring, thereby obscuring interpretation of apparent treatment effects [ 16 , 9 , 2 ] . Since contrasts of hazards lack a coherent causal interpretation, risk-based effect measures are generally preferable in causal inference settings [ 8 , 1 , 19 , 12 , 13 ] . Two common risk-based measures include the net risk (i.e. marginal cumulative incidence), which treats the competing events as uninformative censorings, and the cause-specific cumulative risk [ 6 ] . A particular problem arises if the treatment also influences the competing event [ 7 ] . The treatment may appear protective against the event of interest simply by accelerating the competing event, thereby leaving less time for the former to occur. Conversely, if the treatment delays the competing events, it appears to increase the risk of target events. Thus, similar total effects may reflect different underlying mechanisms, which may point to different conclusions about the treatment effect. To characterise treatment effects comprehensively, it is therefore important to distinguish between different underlying mechanisms. Young et al. [ 24 ] framed the cause-specific and net risks as two causally interpretable estimands. Under certain exchangeability assumptions, the cause-specific risk difference was shown to estimate the average total effect (TE) while the net risk difference was interpreted as a controlled direct effect (CDE) reflecting the treatment’s effect under a hypothetical scenario where competing events were eliminated. CDE thus allows a two-part decomposition of the total effect, where the difference \text{TE}-\text{CDE} represents mechanisms in which the competing event intercepts the occurrence of target events. Stensrud et al. [ 20 ] proposed another two-way decomposition in the competing-risks setting. Their separable effects framework views the treatment as comprising two components: one affecting only the target event and the other affecting only the competing event. Separable direct and indirect effects are then defined as contrasts under modified treatments changing only one of the components at a time. A key strength of this framework is its testable real-world implications. Nonetheless, requiring a feasible treatment decomposition may be overly restrictive. For example, holistic treatments, such as lifestyle interventions, inevitably include aspects that influence the occurrence of both events. Another way to decompose the total effect includes path-specific effects, which assess the treatment’s effect through distinct pathways in a causal diagram [ 3 ] . Domingo-Relloso investigated certain path-specific effects involving competing risks under an additive hazards model [ 5 ] . Decomposing the treatment effect into its constituent parts is central to mediation analysis, in which a third variable, the mediator, transmits part of the treatment’s effect on the outcome [ 22 ] . In the mediation context, a widely used approach to decompose the total effect relies on cross-world estimands, i.e. potential outcomes under hypothetical interventions assigning the treatment to one level and the mediator to the level it would take under an alternative treatment level. The total treatment effect is partitioned into the direct part not operating through the mediator and the indirect part that is transmitted through the mediator [ 14 , 15 , 23 ] . VanderWeele [ 21 ] proposed a further decomposition of the total effect into four parts, separating mechanisms arising from treatment-mediator interactions. This paper proposes a four-way decomposition of the total treatment effect under competing risks. The decomposition comprises the controlled direct effect and three distinct interception effects, characterising how the apparent treatment effect arises due to the interplay of the two events. Based on a causal model, the decomposition relies on cross-world estimands through a process combining the indicators of the two events under elimination of the other. The new decomposition reveals the role of the competing event and serves as a basis for defining causally interpretable estimands in the presence of competing events. These components may be further combined to form different target estimands, including natural direct and indirect effects as special cases. The paper is organised as follows. Section 2 presents the data structure and notations. Section 3 introduces an empirical four-way decomposition of the total effect under competing events. Section 4 presents the causal definitions of the four components, based on appropriate cross-world estimands, and provides conditions under which their population-level averages are identifiable from empirical data. Section 5 illustrates the decomposition using simulation. In Section 6 we apply our framework to two real-world trials. Section 7 discusses the choice of causal effect estimands. Section 8 concludes the paper. 2 Observations and data structure Let A\in\{a,a^{\ast}\} denote a binary treatment, with a treated and a^{*} untreated. We consider a discrete-time model defined on a sequence of time points, t_{0}<t_{1}<\dots<t_{K} , at which events and state updates may occur. Let Y=\{Y_{k}\}_{k=0}^{K} and D=\{D_{k}\}_{k=0}^{K} denote the indicator processes associated with two competing events and C=\{C_{k}\}^{K}_{k=0} the indicator process for censoring. For each k , Y_{k} and D_{k} are binary indicators that equal one if the corresponding event has occurred by time t_{k} , and zero otherwise. The censoring indicator C_{k} equals one if an individual is lost to follow-up by time t_{k} without having experienced either Y or D . Let W=(W_{1},\ldots,W_{h}) denote the baseline covariates, and w a joint realisation. Overbars and underbars denote the histories and futures of the processes, respectively. For example, \bar{Y}_{k}=(Y_{0},\dots,Y_{k}) and \underline{Y}_{k}=(Y_{k},\dots,Y_{K}) . We take Y to represent the target event whose occurrence terminates the entire follow-up and D its competing event. We assume the ordering (C_{k} , D_{k} , Y_{k}) for all k . All individuals are assumed to be event-free and uncensored at time t_{0} , i.e. C_{0}=D_{0}=Y_{0}=0 . 3 Decomposing the total effect Let Y_{k}^{a^{\prime}} denote the potential outcome of Y_{k} when, possibly contrary to fact, the treatment is assigned to level a^{\prime}\in\{a,a^{*}\} . The average total effect (TE) on Y by time t_{k} is defined as the difference in mean potential outcomes between the treated and untreated levels [ 24 , 18 ] : \text{$\mathbb{E}$}\hskip-1.0pt\left[\hbox{TE}_{k}\right]=\text{$\mathbb{E}$}\hskip-1.0pt\left[Y_{k}^{a}-Y_{k}^{a^{*}}\right]=\text{$\mathbb{E}$}\hskip-1.0pt\left[Y_{k}^{a}\right]-\text{$\mathbb{E}$}\hskip-1.0pt\left[Y_{k}^{a^{*}}\right]. (1) Young et al. [ 24