Austin Biom and Biostat. 2016; 3(1): 1029.
Department of Biostatistics, Bioinformatics and
Biomathematics, Georgetown University, USA
Abstract
Multiple endpoints data are common in practice. There are various statistical
methods for the analysis of this type of data, however, genetic models for
familial observations with multiple endpoints data are relatively few, and the
existing methods are basically variations of the Elston-Stewart algorithm. Here
we consider several joint statistical models for such data with quantitative
measurements with a new algorithm, which is computationally more efficient
than the existing method. The proposed method is detailed in some commonly
used parametric, semi parametric and nonparametric settings for this type of
data. For un-genotyped data, the commonly used models are the mixture and
variance components models. We elaborate how these genetic models can be
extended for multiple endpoints data with the proposed method
Keywords: Censoring; Endpoints data; Familial structure; Genotype;
Missing observation
Introduction
Endpoints data are observed responses from patients of some
pre-specified clinical events of interests, such as death, loss of vision,
occurrences of certain diseases, or other symptomatic events. In
medical research, study participants are often followed for a long
time, during which some participants may drop out early, so that
random censorship may be present in the data. Such data have
missing observations, which may be inhomogeneous across the
patients. For example, in one patient we have observations on the
lung cancer and kidney disease, and on another we have observations
on lung cancer, diabetes and asthma. Analyses of such data largely
fall into two categories: hypothesis testing (usually non-model based)
and model inference. Here we concentrate on the model inference of
such data.
For censored data modeling there are extensive literatures [1-8],
just mention a few. For multiple endpoints data analyses, there are
various statistical methods [9-13], for example. Wei and Glidden [14]
provided an overview for some of the methods in this field.
Family genetic data differ from the ordinary data in that they are
collected in familial units, often with varying structures and sizes, and
with/without genotyping. These features make the models distinct.
The key in the modeling is the familial dependence structure and
the implementation of the genetic mechanism, existing methods
are basically variations of the Elston-Stewart algorithm, which is a
multi-level mixture model, and the computation is often challenging.
Genetic models for multiple endpoints data are relatively limited.
Here we consider some statistical joint models for such data with
quantitative measurements with a new algorithm, which is a one-level
mixture model, thus enhance the computation considerably. The
parametric method is used when one has some confidence about the
model specification. The semi parametric method can be used when
there is not enough information about the full parametric model
specification. The nonparametric method is used for the robustness of least model assumptions. We elaborate our methods for the
parametric, semi parametric and nonparametric cases. The methods
we describe below are valid for arbitrary pedigrees; however, in this
article we focus on the simpler case of nuclear family for illustration.
In genetic analysis, the data contains genotypes, partially
genotypes, or no genotypes. However, even if the data are genotyped,
it is still of interest to know whether there are some other unknown
gene(s) behind the response functioning. There are reports that
with added unknown gene locus, the likelihood Akaike information
reduced (e.g., [15], p.1091, [16], which makes sense, as correct
parameter(s) added to the model will reduce its AIC), or the
segregation analysis guided to some other gene(s) which deserve(s)
further investigation. So even for the genotyped data, a segregation
model is still of importance. It is also the general model including the
genotyped data case. In the following we derive some commonly used
regressive models for this case, including the parametric model, semiparametric
proportional hazards model, nonparametric least squares
model, variance components model and the competing risks model.
Also, hypothesis testing on parameters of interest can be conducted
using the likelihood ratio statistics based on the parametric models.
Our aim here is to present several new parametric, semi parametric
and nonparametric models for this familial data, and thus we focus
derivations of the basic forms of these models. Implementations of
these models and applications to real data will follow in our future
work.
Methods
Suppose there are d responses observed with some clinical events
of interests, along with r covariates, for each member in a family. We
concentrate on nuclear family structure for simplicity. In practice we
only observe a subset of the responses and covariates for each patient
in a family. Let yi=(yif,yim,yis)(i=1,…,n) be the vector of responses for
the i-th nuclear family, with corresponding covariates xi=(xif,xim,xis)
(i=1,…,n). Here yif a dif(<d) dimensional vector of responses of the father in the i-th family, which belongs to a dif dimensional subspace
of the d-dimensional space, with a response non-missing indicator
vector Iif and covariate non-missing indicator vector Jif. Similarly, yim
denotes a vector of responses of the mother in the i-th family and
yis=(yi1,…,yibi) is for offspring with each of the yij s has the same data
structure as that for yif. For example, there are three responses to be
observed, we have the first and third on the father, then Iif=(1,0,1),
and dif=|Iif|=2 is its dimension or cardinality. If there are total of five
covariates in the design, and we may only have the first, second,
fourth covariates for the father, then Jif=(1,1,0,1,0), and rif=|Jif|=3 is
its dimension. We assume random censorship. Let δif=(δif1,…, δifdif)
be the censoring indicator of yif, i.e. δifj=1 if yifj is uncensored, and
δifj=0 otherwise. Similar notations are used for the mother. For the
off springs, yis = (yi1,...,yki) denote the response vector, with yij be dij
dimensional observation for the jth sib, with response configuration
Iij and covariate configuration jif, (j=1,…,ki). Let di=(dif,dim,dis),
Ii=(Iif,Iim,Iis), Ji=(Jif,Jim,Jis), δi=(δif,δim,δis) and δi=(Cif,Cim,Cis). The
complete data information consists of Zi=(yi,xi,Ii,Ji,δi), (i=1,…,n). Let
χ(⋅) be the indicator function, i.e. χ(gir=s)=1 if the genotype of the rth
individual is s and zero otherwise. Let πs be the population proportion
of the S-th genotype, t(s|i,k) be the transmission probability of a off
string’s genotype s given the parents’ genotype (i,k), and θ be the
collection of all the parameters, including the as and βs in the mean
and the parameters in the within and between individual covariance
matrices and the genotype frequencies πks and transmission
probabilities t(s|i,k)s. With unobserved genotypes, the computation
is a serious challenge because of the mixture nature of the model.
Let f(yr|θ), F(yr|θ), and S(yr|θ)=1-F(yr|θ) be the density function,
distribution function and survival function of Yr, respectively. They
may be described by its genotype gr through the mean function
specification. To simplify model specification, we assume random
mating so that father and mother can be viewed as independent
in most cases. The case of non-random mating or within parent’s
dependence can be treated similarly with more involved notations.
Note that there is within family dependence but independence
among different families. We assume the genotypes of each patient
are unobserved; the case of observed genotypes is automatically
covered and simpler. Now we describe the methods in some common
settings below.
Parametric model
Let the genotypes at the locus of interest be coded as 1,…,k.
We first consider the case of no missing record and censoring. The
regressive model assumes yir=μ(gir)+∈ir, (i,=1,…,n; r=f,m,1,…,n),
where μ(gir)=μ0+αχ(gir)+βxir is the mean phenotypic value,
α=(α1,…,αk)', χ(gir)=(χ(gir=1),…,χ(gir=k), μ0 is the intercept vector,.
The residual error term ∈ir is a d dimensional random vector where
they are independent across i but dependent across r.
In the case of multivariate familial quantitative response data,
under the commonly used Elston-Stewart [15] algorithm or its
variants, the likelihood of the observation yi for the i-th nuclear family
is
where typically the density is assumed multivariate normal with
covariate matrix Σ, f(⋅|θ,k) is the density for residual ∈i with genotype
gi in the mean vector specification, f(yl|θ,i,j,r) is the density for residual
∈r with genotype gr and with adjusted mean and variance given by
μ(gil=r)-ΩΣ-1[(yif-μ(gf=k))+(ym-μ(gm=j))] and Σ-ΩΣ-1Ω
where Σ=Cov(Yl,Yl) and Ω= Cov(Yl,Ym); K(θ,I,j) is a quantity that
depends on the parents’ genotypes and the mean [18]. It is well
known that when the number of genotypes is relatively large, this
model is computationally inefficient [19]. Proposed a computational
more efficient model. In light of [19], let Gf=Gm=(π1,…,πI), then the
joint likelihood for the i-th family is written as
where
In the mean specification f(yl|G,θ,r) is the density of residual ∈l
with genotype gl=r with adjusted mean and variance given by
where yp=(yf,ym),μp=(μf,μm), similarly for ; μm
In comparison, model (1) has three layers of mixing (summation)
corresponding to bik3 function evaluations that grow exponentially
with the number of genotypes. On the other hand, model (2) has
only one layer of mixing in three factors each, or (bi+2)k function
evaluations that are linearly proportional the number of genotypes.
The reduction of computation will be more significant for multiple
loci case.
Here we extend this model in the case of censoring and partial
observation. In this case, the mean is modeled as
where the operation means the projection of μ0 onto the
subspace corresponding to the nonzero elements of Iir, similarly for
and . The corresponding error is now .
Recall that in the case of 1-dimensional observation without
genetic implementation, the likelihood for an observation yi with a
censoring indicator di is
To extend this to our situation, for any dimension indicator Ii, and
any d-variable function v(⋅), let be the marginal version of v(⋅)
with respect to the non-zero entry of Ii and
Let 1-di be the indicator with the same length of di but with 0 and 1
reversed. The full likelihood is
Here extra caution should be taken since the observation vector from each individual may vary in dimensions and sub-spaces. For a
d-dimensional vector v, let be its margin with respect to
; and for a d-dimensional matrix A, denote the submatrix
by the rows corresponding to the non-zero entry of dijIij and
columns corresponding to the non-zero entry of dilIil. In particular,
has adjusted mean given by
and adjusted variance matrix given by
where Iip=(Iif,Iim). The corresponding adjustment in
is made. The parameter θ is estimated by its MLE
under (3), along with the restriction
Semiparametric model
For censored data, a commonly used semi parametric regression
model is Cox’s proportional hazards model [20,21]. In the univariate
case, let y(3)<y(4)<…<y(n) be the ordered observations of y1,…,yn
(assume no ties for simplicity), x(i) and d(i) be the associated quantities,
for y(i), of the xi’s and di’s. Let R(i) be the i-th risk set, the set of all
individuals who are still under study at the ‘time’ just prior to y(i), U
be the set of all uncensored individuals, and
be the hazard function. The proportional hazards model has a form
of λ(y/x,θ)=h(β'x)λ0(y)-0.2cm for some known positive function h(⋅),
and unspecified baseline hazard rate λ0(⋅), which implies that the
distribution belongs to the Lehmann family [4] 1-F=(1-F0) for some
F0(⋅) and γ>0. Under these assumptions, the conditional likelihood
(partial likelihood [20,21]; marginal rank likelihood, [4]) is
where the estimate of θ is the MLE under Lc(y|θ). The optimality
property of is studied extensively. In the case of multivariate
observations, various extensions of this method have focused on
each marginal distribution and Markov chain Monte Carlo on the
margins [22]. Proposed a multivariate extension of the proportional
hazards model, or frailty model, which is equivalent to an exponential
specification of the joint survival function [23]. Proposed a class of
multivariate failure time distributions, including a multivariate
version of Cox’s proportional hazards model, in which the within
family dependence is modeled by a common latent variable with a
known parametric distribution given that all the family members
are independent. Then the joint distribution is obtained by taking
expectation of the conditional one. All these frailty models assume that
there is a shared common dependent latent variable. This assumption
basically requires that the distribution be interchangeable among the
involved individuals. This is reasonable for some familial data but
not generally true. Other existing multivariate proportional hazards
models [24-26] are similar in nature. Here we model the within family
dependence in a manifest way to be desirable for our genetic analysis.
We adopt a successive conditional version of the proportional
hazards model where we assume a special semi parametric form of the
survival function in order to evaluate the conditioning in closed form
easily. More specifically, in our multivariate proportional hazards model, we assume h(⋅) and λ0(⋅) are functions of d-variates each. Let
yi,(3),…,yi,(ni) be the ordered observations on the i-th variable (i=1,…,d),
define xi,(j),di,(j),Ri,(j), and Ui accordingly. Note there are structures in
h(⋅) through the dependent effects among the covariates. Recall β'x
is a d-vector, let
where Ω is the within individual covariance matrix. Then h(⋅) behaves
as a d-variate normal density, and its marginal and conditional
versions are well defined and in closed forms, although it is not
a proper density function. We need the successive ‘conditioning’
form of h(⋅) to apply the proportional hazards method. Specifically,
let wij be the j-th diagonal element of Ω where Ωj be the upperleft
j-dimensional sub-matrix of it, aj be the first j elements in the
j-th column; [β'x]j be the first j components of β'x,hj+1|j(⋅), be the
conditional version of covariates [β'x]j+1 given [β'x]j. Then hj+1|j(β'x)
is a univariate normal kernel with mean and variance, and
Thus, without mixing over gene, for singleton multivariate observations, the joint conditional likelihood is
Now for the case of nuclear family, inspired by (4), we assume
h(⋅) has the form
Treat h(⋅) as a ‘density’. Recall μi=(μif,μim,μi1,…,μibi). The
conditioning [μi]j+1|[μi]j can be applied component-wise, i.e.
[μi]j+1|[μi]j=([μif]j+1|[μif]j, [μim]j+1|[μim]j, [μi1]j+1|[μi1]j,…, [μibi]j+1|[μibi]j)
Now we have
In (7), [μ(gj,Iir)]1, means the first component of μ(gj,Iir) in Iir,
and |Iir| denote its cardinality (r=f,m,1,…,bi). Now, the conditional
likelihood is
where hj+1|j([μ(i)]j is given by (7). The MLE
of θ is obtained under
(8).
Nonparametric model
For univariate censored data, [27,28] considered a class of
estimators, including the weighted least squares estimators, for
censored data. Here the weights are determined by the ordered
statistics of the observations and the associated censoring indicators,
and are derived from the empirical survival function, i.e., the Kaplan-
Meier product limit estimator [29-32]. Formulated the multivariate
Kaplan-Meier estimator. Using the product integral, the mathematical expressions are quite involved. So instead of choosing the weights
according to the multivariate Kaplan-Meier estimator, we use the
nonparametric locally weighted least squares method, also called
locally linear regression smoothers [33,34]. Let Y and X be the d and
J-dimensional random vectors corresponding to the full observation
and the covariates for an individual. Let μ(x)=E(Y|X=x) denote the
regression function. In the univariate observation case, the locally
linear estimator of μ(x) is first to find and to minimize
Where K(⋅) is a kernel function, hn is the bandwidth, and
. In our case, keep the notations in section 1. We choose
the kernel to be the J-dimensional standard normal density f(⋅), and
(⋅) to be its distribution function. To simplify the expression of the
likelihood, let
and similarly for imr , , , , , and . To
estimate , inspired from the univariate locally linear estimator
and (4), we first find to minimize
Where Ω is the within individual variance matrix, ( and
are the adjusted quantities as those in (4). And I is
the sub-matrix of Ω with rows and columns corresponding to the
non-zero elements of Iim.
We estimate
We estimate Ω by with
where nrs is the total number of individuals with non-missing (r,s)-
th components, zrls are the rearrangement of the r-th component of
for which the (r,s)-th components
are non-missing.
Let be the minimize of (9), where the full
depends on the genotype r and the point value x. It has the intercept
term (recall (3)), and is approximated by setting
Direct computation of in (7) is not easy,
instead we use an iterative procedure as in the following steps.
Select starting values π(0) for π. With this π(0), compute Ω(0), and
T(0)(r)s. Let η=(μ0,α,β) be the full representation of the regression
parameters, Xir(r=f,m,1,…,ni) be the corresponding design matrix
for the r-th individual in the i-th family. In iterations 1-m do the
following
(i) Fix π(i) , Ω(i) and T(i)(r) s minimize (7) with respect to η to get
where for l=f,m, and for l=1,…,ni.
(ii) Fix (i), minimize (7) with respect to π, with the constraint to get
(r=1,…,k) (11)
and update Ω(i+1), and T(i+1(r) with (i+1).For some pre-specified ∈>0,
when the relative errors
we stop the process at the last step m, and take
For arbitrary kernel and reasonably chosen band width hn,
various asymptotic results are established in case of standard nonmixture
data. We conjecture that similar results will hold under some
regularity conditions.
Lastly, the bandwidth determines the smoothness of the estimate.
Interesting research that addresses the crucial problem of bandwidth
selection can be found in [35]. There are considerable literatures for
automatic methods that attempt to minimize a lack-of-fit criterion
such as an integrated squared error. But most of the methods
provide an optimal hn determined by some unknown quantities. For
simplicity, let k=|Jij| be the dimension of the observed covariate of the
j-th (j=f,m,1,…,ni) individual in the i-th family, for the corresponding
kernel, we choose hn=Cn-1/(k+1), for some constant C>0, and C can be
selected through numerical trial.
Variance components model
As an alternative to the mixture models considered above, the
Variance Components (VC) model [36,37] has received much
attention recently due to very efficient in computation as well as
relatively robustness to model misspecification [38-46].
Let yi be the trait vector of the i-th individual in the family, in
case without censoring and missing records, the commonly used VC
model describing the trait value is
yi=μ+gi+Gi+ηxi+ei
Where μ is the overall mean, gi is the unobserved random vector
of major gene effects at the trait locus with alleles A and B, Gi is the
unobserved polygenic effects vector, the ηj’s are effects associated
with the covariates xij’s, and ei is the residual random error vector.
The usual assumption is that gi,Gi and ei are uncorrelated and
E(gi)=E(Gi)=E(ei)=0. When missing records are present, the model is
modified as
In this model, the parameters of interests are specified in the
family variance matrix, thus computation can be carried out efficiently
without the multiple mixing. Let yk,πk and Ωk be the observation, its
mean and variance matrix of the k-th family. We can define Ik, dk and
Jk accordingly. The commonly used model for quantitative traits is the
multivariate normal distribution, thus the total likelihood is
Here f is the distribution function of the normal distribution
with mean 0 and variance Ω.
The key lies in the specification of the variance matrices Ωks,
which we illustrate in the following settings.
In the simplest case of Hardy-Weinberg equilibrium among locus
alleles without linkage to marker, and without censoring and missing
records, the covariance matrix between individuals i and j of a given
family can be found, for example, in [38]. Modified to our case, it is
where is the additive genetic variance matrix due to the locus,
is the dominant genetic variance matrix, is the
kinship coefficient between individuals i and j [47], and Δ7ij,Δ8ij,Δ9ij,
etc. are the condensed kinship coefficient of Jacquard [48], between
individuals i and j.
In the more general Hardy-Weinberg disequilibrium case, let f be
the within population inbreeding coefficient f at the trait locus [49-
51]. Introduced CV model in this case, which modified in our case is
where γl(f)s are matrices determined by 2
, and f etc., see there
for details.
In the case of linkage to marker with both Hardy-Weinberg and
linkage equilibrium, the covariance in our case can be specified based
on that of, for example [40], in the same way as above. In the case
of linkage to marker with either one or both Hardy-Weinberg and
linkage disequilibrium, the covariance in our case can be specified
based on that of [51], in the same way.
Competing risks
Now suppose that the response yi is the failure time and only the
failure for one of the d diseases is observed for each individual. For
the i-th family, the data have the form (yi,di,ji,xi), where yi=(yif,yim,yi1,…
,yibi), similarly for di,ji, and xi, where ji is the observed disease type
indicator. For example if the observed disease for the father is type
2, then jif=2. Given the data (yi,di,ji,xi)s, we like to investigate the
objective of interests for each of the d disease. This problem is that
of the competing risks. Note here the response for each individual
is one-dimensional, and hence the corresponding quantities have
simple notations. We are interested in the genetic regression analysis
for the competing risks. We use a variant of the proportional hazards
model. The mean of the j-th type, r-th member of the i-th family is
specified as
For a reasonably chosen function h(⋅), we specify
More convenient below is to use the notations
and
Let yj1<…<yjkj be the kj failures of type j(j=1,…,d), R(yji),be the risk
set at yji, the partial likelihood is
Asymptotic heuristic
For IID data, various asymptotic results can be obtained. The
results from the score function, the likelihood ratio statistic, and the
MLE are equivalent. These results can be used to establish confidence
intervals or hypothesis testing, etc. for θ. Here we are more interested in
using the MLE. For general dependence model, usually the treatment
is non-standard. But for our model, since the log-likelihood is in the
form of several additive pieces, standard method can be used to derive
the asymptotic distribution of the MLE. Let zi=(yi,di)(i=1,…,n). For
the IID data, it is well known that under mild regularity conditions,
the MLE nis strongly consistent and asymptotically distributed
normal with mean at the true parameter value θ0, and variance matrix
given by the inverse of the Fisher information. Here the observations
are unbalanced, the asymptotic variance is the Fisher information
times a weight matrix. To derive it, we need some notations, and
mainly concentrate on model (4).
Let Nip be the total number of parents with the i-th measurement
non-missing (i=1,…,d), Niss be those for the siblings, N be the total
number of individuals in the study, γNir=Nir/N(r=p,s). Assume
limN→8γNir=γir>0 exists (r=p,s;i=1,…,d). Let Yp and Yj be general random
vectors associated with a parent and sib respectively, and Δp and Δj be
the corresponding random vectors.
For model (4), let
here we use the notation Δp v(⋅) to represent the marginal version
of v(⋅) corresponding to the non-zero components of Δp. The Fisher
information matrix is
The above expectation is more involved than it looks, since
that involves summations of all possible combinations of non-zero
elements of it with respect to Δr, and also the unknown distribution of
it. Instead, an empirical version of it has a known form
where L(y\θ) is given by (4). At the true data generating parameter
θ0, IN is strongly consistent for I(⋅). To obtain the weight matrix, we
need to specify the parameter order in θ. We arrange the first k-1
entry to be π1,…,πk-1, next we arrange all the regression parameters
for the first response variable,..., all the regression parameters for the
last response variable, then all the independent parameters in the
variance matrix S and covariance matrix Ω in the similar order. It
is clear that, in the weight matrix W, for (i,j) corresponding to the
first k-1 components in θ, the weight should be γp=γ1p+…+γdp; for
(i,j) corresponding to the r-th and the l-th regression parameters,
the weight is ; for (i,j) corresponding to
the (a,b)-th and the (u,v)-th variance or covariance, the weight is [(γap+γas) (γbp+γbs) (γup+γus) (γvp+γvs)]1/4. Let θ0 be the true unknown
data generating parameter, d→
stands for convergence in distribution.
Then, we have
where A⊗B stands for the Kronecker product of matrices A and B.
Since I(⋅) involves unknown quantities, equivalently
where WN is W with the γs replaced by the γNs.
The above ideal applies to the other models in this paper, but the
results will be more involved, and we only discuss them briefly.
For the proportional hazards model, even for the IID data case, the
conditional likelihood looks much different from the full likelihood.
Interestingly, the MLE from this model (under the assumption of
correct model specification and some regularity conditions) has the
same asymptotic distribution as that from the full likelihood [21,52].
For the proportional hazards model, it is noted [4] that the survival
function can be written as
S(y\x,β)=-S0(y)h(β'x),
and f(y\x,β)=-dS(y\x,β)/dy. So to get the full log-likelihood (14),
we need the estimate of S0(y) for the d-dimensional case [29,30].
Proposed the multi-dimensional generalization of the Kaplan-Meyer
nonparametric estimator of S(y), similar technique can be used here
for the construction of S0(y). Then (18) continues to hold in this case.
Due to technical involvement, we will not pursue the details here.
For the least squares estimator, since the weight involves the
kernel and hn, the treatment is different from those above, and in the
case of full observation generally the asymptotic result is of the form
for some constant C and matrix Ω determined by the kernel and
the true (unknown) data and censoring distributions [53,54]. In our
case of partial observation, the above result holds with Ω replaced by
Ω⊗W.
For the competing risks model, the structure is similar to that of
the proportional hazards model. Here the response is one dimensional
so that the survival function can be estimated by the Kaplan-Meyer
estimator and the weight matrix W is the identity.
Discussion
We have considered several statistical methods, parametric, semi
parametric and nonparametric models, for the genetic regression
analysis of familial multiple endpoints data, with possible missing
records. Here we only considered the case of nuclear families and the
parameters are independent of time. The cases of arbitrary pedigrees
and/or the time dependent parameter can be treated similarly. The
variance components method can also be applied to the proportional
hazards model and in the analysis of competing risks. There are some
marginal models for the multiple endpoints data, which work well
in practice. But we think the joint model is more appropriate when
the within responses structure is important in the analysis. Another
commonly used method to deal with the missing data is the EM
algorithm [55], which can be implemented into the models considered here. But for the multiple endpoints data, the proportion of missing
part is usually large; the EM algorithm may not be efficient. When the
missing pattern is non-ignorable, more complicated approaches need
to be considered to reduce potential biases. Hypothesis testing for
parameters of interests can be conducted using the likelihood ratio
statistics based on the parametric models. We only derived the basic
forms of these models; more features can be implemented to them in
particular applications.
References
- Kaplan EL, Meier P. Nonparametric estimation from incomplete observations. Journal of the American Statistical Association. 1958; 53: 457-481.
- Breslow N. Covariance analysis of censored survival data. Biometrics. 1974; 30: 89-99.
- Fleming TR, Lin DY. Survival analysis in clinical trials: past developments and future directions. Biometrics. 2000; 56: 971-983.
- Kalbfleisch JD, Prentice RL. The Statistical Analysis of Failure Time Data, Wiley. 1980.
- Andersen PK, Gill RD. Cox’s regression model for counting processes: A large sample study. Annals of Statistics. 1982; 10: 1100-1120.
- Arjas E. Survival models and martingale dynamics (with discussion). Scandinavian Journal of Statistics. 1989; 16: 177-225.
- Flemming TR, Harrington DP. Counting Processes and Survival Analysis. Wiley, New York. 1991.
- Andersen PK, Borgan O, Gill RD, Keiding N. Statistical Models Based on Counting Processes. Springer-Verlag. 1991
- Printice R, Williams BJ, Peterson AV. On the regression analysis of multiple failure time data. Biometrika. 1981; 68: 373-379.
- Wei LJ, Lin DY, Weissfeld L. Regression analysis of multivariate incomplete failure time data by modeling marginal distributions. Journal of the American Statistical Association. 1989; 84: 1065-1073.
- Freedman LS, Graubard BI, Schatzkin A. Statistical validation of intermediate endpoints for chronic diseases. Stat Med. 1992; 11: 167-178.
- Fleming TR, DeMets DL. Surrogate end points in clinical trials: are we being misled? Ann Intern Med. 1996; 125: 605-613.
- Zhang J, Quan H, Ng J, Stepanavage ME. Some statistical methods for multiple endpoints in clinical trials. Controlled Clinical Trials. 1997; 18: 204-221.
- Wei LJ, Glidden DV. An overview of statistical methods for multiple failure time data in clinical trials. Stat Med. 1997; 16: 833-839.
- Remington DL. Effects of genetic and environmental factors on trait network predictions from quantitative trait locus data. Genetics. 2009; 181: 1087-1099.
- Knüppel S, Rohde K, Meidtner K, Drogan D, Holzhütter HG, Boeing H, et al. Evaluation of 41 candidate gene variants for obesity in the EPIC-Potsdam cohort by multi-locus stepwise regression. PLoS One. 2013; 8: e68941.
- Elston RC, Stewart J. A general model for the genetic analysis of pedigree data. Hum Hered. 1971; 21: 523-542.
- Bonney GE. Compound regressive models for family data. Hum Hered. 1992; 42: 28-41.
- Yuan A, Bonney GE. Two new recursive likelihood calculation methods for genetic analysis. Hum Hered. 2002; 54: 82-98.
- Cox DR. Regression models and life-tables. Journal of the Royal Statistical Society. 1972; 34: 187-202.
- Cox DR. Partial likelihood. Biometrika. 1975; 62: 269-276.
- Clayton DG. A Monte Carlo method for Bayesian inference in frailty models. Biometrics. 1991; 47: 467-485.
- Hougaard P. A class of multivariate failure time distributions. Biometrika. 1986; 73: 671-678.
- Holt JD, Prentice RL. Survival analysis in twin studies and matched pairs experiments. Biometrika. 1974; 61: 17-30.
- Oakes D. A model for association in bivariate survival data. Journal of the Royal Statistical Society. 1982; 44: 414-422.
- Wild CJ. Failure time models with matched data. Biometrika. 1983; 70: 633-641.
- Stute W. Consistent estimation under random censorship when covariable is present. Journal of Multivariate Analysis. 1993; 45: 89-103.
- Stute W. Distributional convergence under random censorship when covariables are present.Scandinavian Journal of Statistics. 1996; 23: 462-471.
- Dabrowska DM. Kaplan Meier estimate on the plane. Annals of Statistics. 1988; 16: 1475-1489.
- Dabrowska DM. Kaplan Meier estimate on the plane: weak convergence, LIL and the bootstrap. Journal of Multivariate Analysis. 1988; 29: 308-325.
- Gill R. Multivariate survival analysis. Theory of Probability and Its Applications. 1992a; 37: 18-31.
- Gill R. Multivariate survival analysis: Part 2. Theory of Probability and Its Applications. 1992b; 37: 284-301.
- Fan J. Locally linear regression smoothers and their minimax efficiencies. Annals of Statistics. 1993; 21: 196-216.
- Ruppert D, Wand MP. Multivariate locally weighted least squares regression. Annals of Statistics. 1994; 22: 1346-1370.
- Fan J, Gijbels I. Variable bandwidth and local linear regression smoothers. Annals of Statistics. 1992; 20: 2008-2036.
- Fisher RA. The correlation between relatives on the supposition of Mendel inheritance. Trans Roy Soc. Edinb. 1918; 52: 399-433.
- Harris DL. Genotypic Covariances Between Inbred Relatives. Genetics. 1964; 50: 1319-1348.
- Lange K, Boehnke M. Extensions to pedigree analysis IV. Covariance components models for multivariate traits. Am J Med Genet. 1983; 14: 513-524.
- Goldgar DE. Multipoint analysis of human quantitative genetic variation. Am J Hum Genet. 1990; 47: 957-967.
- Amos CI. Robust variance-components approach for assessing genetic linkage in pedigrees. Am J Hum Genet. 1994; 54: 535-543.
- Almasy L, Blangero J. Multipoint quantitative-trait linkage analysis in general pedigrees. Am J Hum Genet. 1998; 62: 1198-1211.
- Fulker DW, Cherny SS, Sham PC, Hewitt JK. Combined linkage and association sib-pair analysis for quantitative traits. Am J Hum Genet. 1999; 64: 259-267.
- Amos CI, Gu X, Chen J, Davis BR. Least squares estimation of variance components for linkage. Genet Epidemiol. 2000; 19: S1-7.
- Blangero J, Williams JT, Almasy L. Robust LOD scores for variance component-based linkage analysis. Genet Epidemiol. 2000; 19: S8-14.
- Sham PC, Purcell S. Equivalence between Haseman-Elston and variance-components linkage analyses for sib pairs. Am J Hum Genet. 2001; 68: 1527-1532.
- de Andrade M, Guéguen R, Visvikis S, Sass C, Siest G, Amos CI. Extension of variance components approach to incorporate temporal trends and longitudinal pedigree data analysis. Genet Epidemiol. 2002; 22: 221-232.
- Lange K. Statistics for Biology and Health. Springer. 1997.
- Jacquard A. The genetic structure of populations. 1974.
- Wright S. The genetical structure of populations. Ann Eugen. 1951; 15: 323-354.
- Cockerham CC. Variance of gene frequencies. Evolution. 1969; 23: 72-84.
- Yuan A, Chen G, Yang Q, Rotimi C, Bonney G. Variance components model with disequilibria. Eur J Hum Genet. 2006; 14: 941-952.
- Tsiatis A. A large sample study of the estimate for the integrated hazard function in Cox’s regression model for survival data. Technical Report No. 562. 1978; 172-180.
- Fan J, Hu I, Truong Y. Robust nonparametric function estimation. Scandinavian Journal of Statistics. 1994; 21: 433-446.
- Cai Z. Weighted local linear approach to censored nonparametric regression. Akritas MG, Politis DM. In: Recent Advances and Trends in Nonparametric Statistics. 2003; 217-231.
- Dempster AP, Laird NM, Rubin DB. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society. 1977; 39: 1-38.