Asymptotic Normality of the Nelson-Aalen and the Kaplan-Meier Estimators in Competing Risks ()
1. Introduction and Background
The model of competing risks has been widely studied in the literature, see e.g., Heckman and Honoré [1] , Commenges [2] , Com-nougué [3] , Fine and Gray [4] , Crowder [5] , Fermanian [6] , Latouche, A. [7] , Geffray [8] , Belot [9] , Njamen and Ngatchou ( [10] , [11] ), Njamen ( [12] , [13] ). In most approaches, the competing risks are assumed to be either all independent or all dependent. Here, the independent component of the potential risks constitutes an independent censoring variable while the other risks are kept as possibly dependent. This approach is used by Geffray [8] . Namely, we consider a population in which each subject is exposed to m mutually exclusive competing risks which may be dependent. For
, the failure time from the jth cause is a non-negative random variable (r.v.)
. The competing risks model postulates that only the smallest failure time is observable, it is given by the r.v.
with distribution function (d.f.) denoted by F. The cause of failure associated to T is then indicated by a r.v.
which takes value j if the failure is due to the jth cause for a
i.e.
if
. The following modeling technique is extracted in Njamen and Ngatchou [10] : we assume that T is, in its turn, at risk of being independently right-censored by a non-negative r.v. C with d.f. G. Consequently, the observable r.v. are
where
and where
denotes the indicator function. As T and C are independent, the r.v. Z has d.f. H given by
. Let
denote the right-endpoint of H beyond which no observation is possible. The subdistribution functions
pertaining to the different risks or causes of failure are defined for
and
by
(1)
When the independence of the different competing risks may not be assumed, the functions
for
are the basic estimable quantities.
The Kaplan-Meier estimator was developed for situations in which only one cause of failure and the independent right-censoring are considered. Aalen and Johansen [14] were the first to extend the Kaplan-Meier estimator to several causes of failure in the presence of independent censoring. In the present situation, the d.f. F may be consistently estimated by the Kaplan-Meier estimator denoted by b
. For
, the subdistribution functions
may be consistently estimated by means of the Aalen-Johansen estimators denoted respectively by
, for
. Indeed, when the process of the states occupied by an individual in time is a time-inhomogeneous Markov process, Aalen and Johansen [14] introduced an estimator of the transition probabilities between states in presence of independent random right-censoring. The competing risks set-up corresponds to the case of a time-inhomogeneous Markov process with only one transient state and several absorbing states (that can be labeled
). Aalen and Johansen [14] obtained the joint consistency of
to
for
uniformly over fixed compact intervals
for
. They also obtained the joint weak convergence of the processes
on fixed compact intervals
for
.
The asymptotic properties of the Kaplan-Meier estimator on the distribution function have been studied by several authors (see Peterson [15] , Andersen and al. [16] , Shorack and Wellner [17] , Breslow and Crowley [18] ).
In this paper, in a region where there is at least one observation, we are interested in providing asymptotic properties of the Nelson-Aalen and Kaplan-Meier nonparametric estimators of the functions
and
. For
in the presence of independent right-wing censorship in the context of competitive risks set out in Njamen and Ngatchou ( [10] , [11] ).
The rest of the paper is organized as follows: Section 2 describes preliminary results and rappels used in the paper. In Section 3, we obtain two laws: In Section 3.1, we give limit law of Nelson-Aalen’s nonparametric estimator for competing risks as defined in Njamen and Ngatchou [10] and Njamen [12] . In Sect. 3.2, we give limit law of Kaplan-Meier’s nonparametric estimator in competing risks as defined in Njamen and Ngatchou [10] and Njamen [13] . In Section 4, we give the trust Bands, including the Hall-Wellner trust Bands and the Nair precision equal bands.
2. Preliminary and Rappels
For
, we introduce the following subdistribution functions
and
of H by:
and
and for
The relations
and
hold for
since the different risks are mutually exclusive. The relation
is also valid for
. The relations that connect the observable distribution functions
,
and
to the unobservable distributions F, G and
are given by:
and
The cumulative hazard function of T and the partial cumulative hazard function of T related to cause j for
are given for
respectively by the following expressions:
(2)
(3)
Let us set estimators for the different quantities. Let
be n independent copies of the random vector
. We define the empirical counterparts of
,
,
and H, for
by:
The relations
and
are valid for
. As T is independently randomly right-censored by C, a well-known estimator for F is the Kaplan-Meier estimator defined for
by:
where the left-continuous modification of any d.f. L is denoted by
. The Nelson-Aalen estimators of
and of
for
respectively are defined for
by:
(4)
(5)
The Aalen-Johansen estimator for
is defined for
by:
For all
, the following equalities hold:
where
, the Kaplan-Meier estimator of G, is defined for
by:
3. Results
In this section, we continue the works of Njamen and Ngatchou [10] , Njamen [12] and Njamen and Ngatchou [11] . In fact, Njamen and Ngatchou ( [10] , p. 9), studies the consistency of Nelson-Aalen’s non-parametric estimator in competing risks, while Njamen ( [12] , pp. 11-12) studies respectively the simple convergence and the uniform convergence in probability of Nelson-Aalen’s nonparametric estimator in competing risks; and Njamen and Ngatchou ( [11] , p. 13) study the bias and the uniform convergence of the non-parametric estimator survival function in a context of competing risks. It is also shown there that this estimator is asymptotically unbiased. For this purpose, we use the martingale approach as the authors mentioned above.
3.1. Limit Law of Nelson-Aalen’s Nonparametric Estimator for Competing Risks
In what follows, we study the asymptotic normality of Nelson-Aalen’s non-parametric estimator in competitive risks. For that, considering, for all
and
, one has the Nelson-Aalen type cumulative hazard function estimator (Nelson, [19] ; Aalen, [20] , Njamen and Ngatchou, [10] ) defined by
(6)
where
.
The cumulative risk in a region where there is at least one observation is given for all
, by (see Njamen, [12] . p. 9)
(7)
with
which indicates whether the individual i is still at risk just before time t (the individual has not yet undergone the event). Its estimator was defined in Njamen and Ngatchou ( [10] , p. 7).
The following theorem gives the limit law of the Neslson-Aalen estimator
in competing risks of Njamen (2017, p. 9). This is the first fundamental result of this article.
Theorem 1.
In a region where there is at least one observation, it is assumed that
for
and
. Then, for all
,
(8)
where
is a centered Gaussian martingale of variance such that:
(9)
where for all
,
(10)
with
standing for the distribution function of
and
the instant risk function.
To prove this theorem, we need the Robelledo theorem. In fact, the Rebolledo theorem below makes it possible to apply the central limit theorem for certain types of particular martingales.
Theorem 2. (Rebolledo’s Theorem)
Let
a sequence of martingales where
,
denotes a counting process and
its compensator. Consider the processes
, and for all
,
. Suppose that
and f are predictable and locally bounded
processes such that
Suppose also that the processes
are bounded. Let’s for all
,
. If
1)
;
2) for all
,
.
Then,
where
denotes the weak convergence in the space of continuous functions on the right, having a left-hand boundary with the topology of Skorokhod and where W is a Brownian motion.
To prove Theorem 1, it is sufficient to check whether the previous conditions of Rebolledo’s Theorem are satisfied:
Proof. For all
and
,
also decomposes into
which in turn can be written in terms of
by
which finally, can be rewritten as
where
can be seen as a random noise process. The martingale
above represents the difference between the number of failures due to a specific cause j observed in the time interval
, i.e.
(see Njamen, [12] , p.6), and the number of failures predicted by the model for the jth cause. This definition fulfills the Doob-Meyer decomposition.
This martingale is used in Fleming and Harrington ( [21] , p. 26) and in Breuils ( [22] , p. 25).
Now, to explain the asymptotic nature of the results, we defined, for all
,
, to pose:
In a subgroup
, where there is at least one observation, the survival function of
is defined for all
by:
Recall also that
is the distribution function of
,
is that of
’s and
that of the
’s. From the Glivenko-Cantelli theorem, one has:
(11)
Otherwise,
one has:
from which one obtains (see Theorem 3, p. 11 of Njamen, [12] ),
Differentiating the martingale
, one has:
and from
one obtains
Consequently, the increasing process of
is given by
Next, for all
and
, one has
Also, for all
and for all
, the process
is a martingale. We apply the central limit theorem for the martingales (Rebolledo’s Theorem). In this purpose, we show that the condition of this theorem is satisfied by
.
One has, for all
,
and also by the proof of the Theorem 3 of Njamen ( [12] , p. 11), we have:
So that, for all
, when
,
which is determinist. Thus, the first condition of Robelledo Theorem holds.
To check the second condition, for all
and
, define
where for all
,
.
We have to show that as
,
converges to 0 in probability.
One has, for all
,
because
Then
Thus, the second condition of Robelledo Theorem holds.
The conditions of the Rebolledo Theorem are verified and by consequently, for all
,
with
.
Finally, for all
,
This ends the proof of the Theorem 1.
The following subsection gives the asymptotic law of nonparametric Kaplan-Meier’s estimator of the survival function in the competing risks of Njamen and Ngatchou ( [10] , p. 13).
3.2. Limit Law of Kaplan-Meier’s Nonparametric Estimator in Competing Risks
The Kaplan-Meier estimator of the survival function (Kaplan and Meier, [23] ) is defined by
where
is the Nelson-Aalen estimator and where, for a process
continuous to the right with a left limit such that
For all
, an estimator of the variance of
, where
is the survival function associated with the subgroup
is given by
The variance of
approximated by that of
is:
(12)
The estimator of the corresponding variance of
is given by
(13)
The following result concerning the asymptotic law of nonparametric Kaplan-Meier estimator and constituted the second fundamental result of this paper:
Theorem 3.
In an area where there is at least one observation, if we assume that for all
and
,
1) for all
,
2) for all
,
3) for all
,
Then, for all
and
, the non-parametric estimator
checks
where
is the center Gaussian martingale and where
denotes the weak convergence in the space of continuous functions on the right, having a left-hand boundary with the topology of Skorokhod.
Proof. To prove this theorem, it suffices to show that it satisfies the conditions of the Rebolledo Theorem.
In an area where there is at least one observation, by posing, for all
,
,
where
.
For
and
, we have for all
and
,
By the proof of Theorem 3 of Njamen ( [12] , p.11), we deduce that
Hence the 1st condition of Robolledo’s Theorem.
For the second condition of Robolledo’s Theorem, condition B is similar to the proof of Theorem 1 above, we find that for all
,
So, for each
,
where
and where
Finally,
The fact that
, for all
and condition C implies:
As
when
, we deduce that:
It follows that:
This ends the proof of the theorem.
4. Confidence Bands of Survival Function
4.1. Confidence Intervals
For
, we wish to find two random functions
and
such that
,
Recall that from the previous sections, for all
,
converges in distribution to a Gaussian martingale centered (see Theorem 3 above). As a consequence,
is asymptotically Gaussian centered on
. Given the above results, the estimated standard deviation of
, noted
is given for all
by:
(14)
Therefore a threshold confidence level
can be built for all
and
, by:
(15)
Here
is the
percentile of a standard normal distribution.
A threshold confidence interval
can also be obtained for all
, by:
(16)
where
is the rank of fractile
of the standardized normal distribution.
A disadvantage of the construction of the confidence interval (CI) with the previous formula is that the bound can be obtained external to the interval
. A solution is to consider a
transform via a continuous function g, differentiable and invertible such that
belongs to a more wide space ideally unbounded and best approximate a Gaussian random variable. The delta method then allows for the estimation of
the standard deviation of the object created by
defined by
. The confidence interval associated with the risk threshold
is built as for all
,
The most common transformation is
, and in this case we have: for all
,
Remark 1. It is also possible to use log, square-root or logit-type transformations in most software defined respectively by for all
,
4.2. The Confidence Bands
The challenge now is to find an area containing the survival function with probability
, or a set of bounds
and
which, with probability
, contains
for all
and
. Among the proposed solutions, the two most commonly used are firstly Hall and Wellner ( [24] ) bands and secondly, strips Nair ( [25] ) (“equal precision bands”). If
is the maximum time event observed in the sample, then for the Nair bands, we have the following restrictions
, however, boter Hall-Wiener may authorize the nullity of
, let
. Technically obtaining these bands is complex, and their practical utility in relation to the point intervals is not obvious.
Remark 2. The starting point uses the fact that for all
,
converges to a centered Gaussian martingale. We then go through a transformation making appear a Brownian bridge
, weighted by
at Nair, to retrieve the suitable critical value.
In particular, because of the joined character, for a given t their extent is wider than that of the corresponding point IC. In what follows we give the expressions obtained in the absence of transformation.
4.2.1. The Hall-Wellner Confidence Bands
Under the assumption of continuity of survival functions
and
respectively related to the event time and the time of censorship, Hall and Wellner show that for every
, the IC joined the risk threshold
is given for all
and
by:
(17)
where
and
are given by
and
is bounds checking
4.2.2. The Nair Precision Equal Bands
Using a weighted Brownian bridge will notably modify the bounds to IC. For
,
and all
, they are then given by:
(18)
where
satisfies
If we compare (12) and (14), we see that the bounds relating to Nair ( [25] ) bands are proportional to the bounds IC and simply correspond to a risk adjustment threshold used in the past.
5. Conclusions and Perspectives
In this paper we have studied the asymptotic normality of Nelson-Aalen and Kaplan-Meier type estimators in the presence of independent right-censorship as defined in Njamen and Ngatchou ( [10] , [11] ) and Njamen [12] using Robelledo’s theorem that allows applying the central limit theorem to certain types of particular martingales. From the results obtained, confidence bounds for the hazard and the survival functions are provided.
As a perspective, obtaining actual data would allow us to perform numerical simulations to gauge the robustness of our obtained estimators.
Acknowledgements
We thank the publisher and the referees for their comments which allowed to raise considerably the level of this article.