WorksMathematics
Causality and order restricted inference in the Aalen additive hasard model
Abstract
We study a graphical model (DAG) with time to event as a response variable. We use the Aalen additive hazard model for modelling the effect on the response variable of covariates, combined with a multiple linear regression for earlier layers in a DAG. This enables modelling covariates affecting the event process both directly and indirectly. The indirect component effect is represented by a product , where describes the effect of the independent covariate on the dependent covariate and is the effect of the dependent covariate on the final time to response variable, in the additive hazard model. We estimate by the order-restricted estimator recently introduced in Anevski and Merai (arXiv:2606.23882v2), thereby enabling disease models with early onset or late onset covariate effects. Combining our previously obtained results with the DAG allows for modelling early onset or late onset covariate effects together with direct and indirect covariate effects, which should be of particular interest for studying causality for complex diseases. We present a nonparametric plug-in estimator of the indirect effect. We derive the limit distribution of the plug-in estimator of the indirect component, and show that it has pointwise asymptotic rate and a limit distribution which is proportional to the Chernoff distribution.
Introduction
We consider a survival analysis model, in which we treat right censored data, combined with a DAG. Thus we let be a positive continuous random variable with an unknown distribution function , where models the time to an event of interest. We assume that we observe the minimum of the time and a censoring time , together with an indicator for the time being exact, .
Introduce the individual counting processes for which one has the stochastic differential equation
for . Here is the individual hasard function, which exists when the distribution of is absolutely continuous, and is then given by , where is the left-continuous indicator process for the individual being at risk at time .
Also is the individual martingale, i.e. satisfying , and where the sigma algebras is a filtration storing the information available at times .
The -algebra generated by the information depends on the amount and type of information about the observed times that is available and the amount of information about the covariates that is available. Starting with the model (1) satisfying one needs to establish that where is the -algebra generated by the observed and available information at time . The concept of noninformative and independent censoring as well as the innovation theorem are used to establish this link. We do not make this assumptions explicit here, since they are of no relevance to us, and refer to reader to a standard reference such as [3], cf. also [5]. We will in the sequel assume that is a filtration, containing the information available at time , and that holds so that .
A basic inference problem in survival analysis is assessing the effect of group indicators or continuous covariate measurements on the distribution of the time to an event. It is then necessary to model the distribution function depending on covariates . This can be done using various models. One standard model which is particularly convenient in our framework is the Aalen additive hasard model
where are unknown functions. Thus is the unknown parameter vector of functions and is the parameter space, where is the set of integrable functions on .
The Aalen model is immediately generalisable to time dependent covariates , which we will allow for in the sequel, and they may even be stochastic processes; the only necessary demand is that is measurable, for every .
A principal difficulty in estimation in the Aalen model is that the coefficient functions are local functions and can not be estimated directly by nonparametric procedures. Instead, one turns to estimating the cumulative coefficient functions
The ordinary Aalen estimator, defined below in (10), provides an estimator of .
In [5] we proposed an order-restricted estimator for the coefficient functions under a monotonicity assumption. The estimator is obtained by projecting the local increments of the Aalen estimator , defined in (10), onto the cone of non-increasing sequences. Under some regularity conditions we showed in [5] that the resulting pointwise estimator has a nonstandard asymptotic rate and a Chernoff-type limit distribution.
One motivation for introduction order restricted inference of the functions is in order to model early onset covariate effects. In fact, we note that being decreasing means that the effect of (an increase in) the covariate value decreases with time or age.
A further motivation of the work in [5], was as a first step in the use of a graphical model in a survival analysis setting, thereby allowing for a categorising of the covariate effects as direct versus indirect effect. This was first suggested in [8], see also [9].
In this paper we incorporate our previously derived limit distribution results into such a graphical model setting and present estimators of momentanous direct and indirect effects, as well as limit distribution properties for the estimated effects. This has to our knowledge not been done before.
Thus the purpose of the present paper is to study a natural extension of a standard Aalen model, in which one of the covariates acts through an intermediate, possibly time dependent, covariate. This leads to a decomposition of the effect of a covariate into a direct component and an indirect component. For simplicity we will treat the simplest possible model, with two covariates. The generalisation to more complex models is straightforward.
More precisely, let denote an (independent) covariate and let denote another (dependent) covariate. We consider the structural representation
where is the event counting process, is the at-risk process, are -adapted processes and is a martingale with respect to the filtration .
The hazard model is
Substitution of the covariate equation (5) into (6) gives
Thus, at the level of this structural decomposition, the coefficient multiplying is
The two pathways represented in (7) are illustrated in Figure 1.
Figure 1. Dynamic path diagram illustrating the direct and indirect pathways from to the event process .
We thus define the instantaneous indirect effect of covariate as
Our main question is the point wise asymptotic behaviour of the plug-in estimator
of the instantaneous indirect effect , where is the order restricted estimator defined in [5], and where is a least squares estimator, to be defined in the sequel.
We will show below that the two factors and have different asymptotic rates. In fact, for each fixed , the estimator is obtained by ordinary least squares regression and satisfies a standard distributional limit, whereas has a nonstandard limit distribution behaviour that is common in order restricted inference. Since , estimation of does not contribute to the first order asymptotic distribution of the estimated indirect effect .
The main result therefore transfers the Chernoff-type limit of directly to the indirect component, with the limit multiplied by the fixed value . The proof is straightforward once the pointwise result for the order restricted estimator is available, but the formulation is useful because it identifies precisely which part of the estimation procedure determines the asymptotic behaviour.
The paper is organised as follows. Section 2 introduces the model and the estimators. Section 3 derives the limit distribution for the new estimator. Section 4 discusses the interpretation and limitations of the result.
Model and estimation
Assume we have data for for individuals, which are seen as observations in the counting process Aalen model (4), which satisfies the structural equations (5). Note that we allow the observed covariates to be time-dependent and arbitrary. The only requirement needed is that they are observations of -adapted processes, so that the martingale structure is maintained under stochastic integrations. We assume we have observations on a fixed compact time interval .
Now for each time point , let be the risk set, i.e. the index set of those individuals that have not yet experienced the event. Since we do not have left-truncated data, it is clear that is a set of indices that is decreasing with time, so that if . Furthermore the cardinality of decreases with one at each observation time , so that , assuming that we have no ties, which we hereby assume.
Therefore, for each fixed observed time point , for , the structural equations (5) give rise to a set of linear equations
for , and we note that there are many such equations. Therefore, for each fixed , we can estimate the parameters by ordinary least squares (ols) regression. We denote the ols estimates by . Since we can solve the system of equations (8) for every time point , we will obtain nonparametric estimates of the full parameter function and , and these are
In fact and are the ols estimate of the vectors and .
Note that even if the covariates would be time-constant, the ols estimates will typically be different for different , since the risk set changes with , and thus the set of linear equations (8) will change as well (the number of equations will change).
Note also that if the covariates would be time-dependent then we could formulate system of equation (8) for every fixed , with replaced by and the index set replaced by . We would then get ols estimates of and of and . Since we can do this for any fixed we actually get nonparametric estimates of the parameter functions, under very mild regularity conditions on the covariates, e.g. that they are piecewise continuous.
Then a multivariate central limit theorem will give a limit distribution result for the -length vector as
on where is a mean zero -dimensional Gaussian random vector, with a certain covariance structure. It is also straightforward to show that we have a process result
on , where is a mean zero Gaussian process in , with a certain covariance function, under mild regularity conditions on the covariates.
For the purpose of the results presented in this paper, only the limit distribution of at a fixed point interior is of interest. We state that limit distribution as an assumption, to allow for general statements.
Assumption 2.1. For the fixed interior point , the ols estimator satisfies
where is a mean zero random variable.
Note that this assumption is satisfied for the case that we treat above of independent data, with a mean zero Gaussian random variable, under mild regularity conditions.
Now, recall the definition of the cumulative functions , for , in (3). Then a standard result, cf. e.g. [3], is that the least squares estimator of is
where is a indicator for a certain design matrix in a matrix formulation of (1) and (2) has full rank, and denotes a generalized inverse of the corresponding design matrix; see [2] and [3] for more details and [5] for a review.
Recently, in [5], this estimator was used as an input into an order-restricted scheme, to obtain a monotone estimator of the functions themselves. Thus if we assume that is non-increasing, and denote the second coordinate of , then [5] defined an order-restricted estimator , namely the isotonic regression of the local increment estimates of onto the cone of non-increasing sequences. Furthermore, in [5] it was shown that a characterisation of the solution is
where the derivative is understood to be the appropriate one sided derivative. Furthermore, algorithms for calculating the isotonic regression are well known, and e.g. the minimum lower set algorithm works well, cf. e.g. [6].
Now, since we have obtained estimators of the factors and of the indirect effect , we can define the plug-in estimator of the indirect effect by
Limit distributions for the indirect effect estimator
We first recall and state the assumptions for the survival analysis setting that enable the limit distribution results of [5] for the order restricted estimator . In particular, these conditions ensure local regularity of the design and risk processes, nonsingularity of the limiting design matrix, and the smoothness and strict local monotonicity conditions on the coefficient function at the point under consideration. We emphasise that they are equivalent to the standard assumptions for deriving limit distributions of the Aalen estimator (10), cf. [3], and thus we do not need more than what is regularly assumed in a standard survival analysis setting.
Assumption 3.1. Assume that Assumptions 1,2 and 3 in [5] hold and that is differentiable with in a neighbourhood around the fixed interior point .
Under Assumption 3.1, Theorem 3 in [5] gives the limit distribution
where is a standard two-sided Brownian motion, and where
and, is specified in [5].
Next we state the main result on the limit distribution for the plug estimator , defined in (12), of the indirect effect .
Theorem 3.2. Suppose that Assumptions 2.1 and 3.1 hold. Then
We note that the limit distribution is proportional to the Chernoff random variable , common in order restricted inference.
Proof. We have that
We first see that the second term is
and by Assumption 2.1, since and the second factor is bounded in probability, the product converges to zero in probability, and therefore the second term is .
Next we treat the first term: Since , and since by (13)
we have that
Therefore
Thus
Multiplying by the fixed constant and applying the limit distribution result (13) gives
Conclusion and discussion
We have established the pointwise asymptotic distribution of a plug-in estimator for the indirect component in an DAG with the final stage modeled by an additive hazard model. When the coefficient modelling the independent to dependent covariate effect is estimated at the regular rate and the covariate coefficient in the final Aalen model is estimated by the newly developed order restricted Aalen procedure in [5], we show that the latter determines the asymptotic behaviour of the plug-in estimator of the indirect effect. In particular the resulting plug-in estimator has the same pointwise scale as the order-restricted coefficient, with a Chernoff-type limit multiplied by the exposure to covariate coefficient.
The result provides a theoretical basis for inference on an indirect component constructed from an order-restricted additive hazard estimator.
The suggested model and estimator is novel, and allows for studying both early onset covariate effects and indirect covariate effects, and should be highly interesting for modelling complex diseases with time-to-event as response variable.
References
- [2]O. O. Aalen. A model for nonparametric regression analysis of counting processes. In Mathematical Statistics and Probability Theory, Lecture Notes in Statistics, volume 2, pages 1–25. Springer, New York, 1980.DOI
- [3]P. K. Andersen, Ø. Borgan, R. D. Gill, and N. Keiding. Statistical Models Based on Counting Processes. Springer-Verlag, New York, 1993.
- [5]D. Anevski and E. M. Merai. Order restricted estimation of the parameter functions in an additive hazard model. arXiv preprint arXiv:2606.23882, 2026.arxiv.org/abs/2606.23882
- [6]T. Robertson, F. T. Wright, and R. L. Dykstra. Order Restricted Statistical Inference. Wiley, Chichester, 1988.
- [8]FOSEN, J., Ø. BORGAN, H. WEEDON-FEKJÆR, AND O. O. AALEN (2006). Dynamic analysis of recurrent event data using the additive hazard model. Biometrical Journal 48, 381–398.pubmed.ncbi.nlm.nih.gov/16845903
- [9]ODD AALEN, ORNULF BORGAN AND HAKON GJESSING (2008). Survival and Event History Analysis. Springer Verlag.DOI
