Biostatistics (2006), 7, 4, pp.

599–614
doi:10.1093/biostatistics/kxj028
Advance Access publication on April 5, 2006

Parametric survival models for interval-censored data

with time-dependent covariates
YVONNE H. SPARLING, NAJI YOUNES, JOHN M. LACHIN∗
The Biostatistics Center, Department of Biostatistics and Epidemiology,
School of Public Health and Health Services, The George Washington University,
6110 Executive Boulevard, Suite 750, Rockville, MD 20852, USA
jml@biostat.bsc.gwu.edu

OLIVER M. BAUTISTA
Merck and Company, Blue Bell, PA 19422, USA

S UMMARY
We present a parametric family of regression models for interval-censored event-time (survival) data
that accomodates both fixed (e.g. baseline) and time-dependent covariates. The model employs a three-
parameter family of survival distributions that includes the Weibull, negative binomial, and log-logistic
distributions as special cases, and can be applied to data with left, right, interval, or non-censored event
times. Standard methods, such as Newton–Raphson, can be employed to estimate the model and the result-
ing estimates have an asymptotically normal distribution about the true values with a covariance matrix
that is consistently estimated by the information function. The deviance function is described to assess
model fit and a robust sandwich estimate of the covariance may also be employed to provide asymptoti-
cally robust inferences when the model assumptions do not apply. Spline functions may also be employed
to allow for non-linear covariates. The model is applied to data from a long-term study of type 1 diabetes
to describe the effects of longitudinal measures of glycemia (HbA1c ) over time (the time-dependent co-
variate) on the risk of progression of diabetic retinopathy (eye disease), an interval-censored event-time
outcome.

Keywords: Interval-censored data; Parametric models; Time-dependent covariate.

1. I NTRODUCTION
In medical and biological research, the analysis of event-time or survival data aims to describe the risk
(hazard) function of event times in a population, the associated survival or cumulative incidence functions,
and the effects of covariates on risk. When event times are not observed exactly, these times are censored.
The event time is “right censored” when follow-up is curtailed without observing the event. “Left censor-
ing” arises when the event occurs at some unknown time prior to an individual’s inclusion in a cohort.

∗ To whom correspondence should be addressed.


c The Author 2006. Published by Oxford University Press. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org.
600 Y. H. S PARLING AND OTHERS

The event time is considered “interval censored” when an event occurs within some interval of time but
the exact time of the event is unknown (cf. Kalbfleisch and Prentice, 2002).
A variety of models have been developed for interval-censored data. Finkelstein and Wolfe (1985)
present a semiparametric model, which is based on factoring the joint likelihood function for a random
interval, and they consider a set of distinct endpoints that comprise the interval as nuisance parameters.
Finkelstein (1986) also develops a method for fitting the proportional hazards model for interval-censored
data where the baseline survival function quantities at the distinct endpoints are considered nuisance
parameters. Seaman and Bird (2001) also extend the interval-censored proportional hazards model to
accommodate time-dependent covariates. The baseline hazard function is defined as piecewise constant
between a specified finite set of time points, and the function is estimated over each interval using the EM
algorithm. Betensky and others (2002) describe a proportional hazards model for interval-censored data
using local likelihood estimation, and they allow an arbitrarily smooth covariate function to describe the
covariate vector’s effect on the hazard function.
Goetghebeur and Ryan (2000) develop a semiparametric regression model for interval-censored data
that employs the EM algorithm. Here, the E-step requires estimating the risk set sizes and number of
events that occurred at each set of possible event times and the M-step estimates the regression coeffi-
cients. Rabinowitz and others (2000) use conditional logistic regression to fit proportional odds models
to interval-censored data, and they assume that the conditional distribution of the interval endpoints given
the covariates follows a semiparametric proportional odds regression model.
Parametric models have also been proposed. Odell and others (1992) describe a Weibull regression
model for interval-censored data with fixed (e.g. baseline) covariates. Rabinowitz and others (1995) ex-
tend the accelerated failure model to the interval-censored case. They present a class of score statistics for
estimating the regression coefficients without specifying the distribution function of the residuals or the
joint distribution of the covariates and the interval times. Moreover, Younes and Lachin (1997) present
a link-based model that can be applied to the interval-censored case. The model employs a link function
to describe the manner by which the covariates act upon the survival times, and it uses B-splines to ap-
proximate the background hazard function. Kooperberg and Clarkson (1997) apply the hazard regression
methodology of Kooperberg and others (1995) to interval-censored data and time-dependent covariates,
and they estimate the logarithm of the conditional hazard function using splines and tensor products.
All these methods for interval-censored data that allow for time-dependent covariates are either com-
putationally intensive or of high dimension due to the many nuisance parameters. As an alternative, we
present a family of parametric survival models for left, right, and interval-censored data with fixed and
time-dependent covariates. This approach provides a direct computational solution with only a few model
parameters in addition to the covariate effects. Furthermore, the proposed family of parametric models in-
cludes a proportional hazard model and a proportional odds model as special cases. As noted by Lindsey
(1998), parametric regression models in the presence of heavily interval-censored data are robust and are
generally more informative than the corresponding non-parametric models.
In Section 2, we present a three-parameter family of event-time distributions that includes the Weibull,
negative binomial, and log-logistic, among others, as special cases, and the associated hazard and cumula-
tive hazard functions. We then describe the likelihood function for the family in terms of linear functions
of fixed and time-dependent covariates, model estimation, and inferences. In Section 3, we present model
diagnostics to examine the event time distribution, the functional form of the covariates, and overall fit of
the model. An example is provided in Section 4, followed by caveats and discussion in Section 5.

2. M ODEL SPECIFICATION
Odell and others (1992), among others, describe a general likelihood function that allows for left, right,
and interval censoring. For the ith subject (i = 1, 2, . . . , n), let xi be the event time, ti be the observed
 Parametric survival models for interval-censored data 601

event, left- or right-censoring time, and for interval-censored observations, let tLi be the left-censoring
time and tRi the right-censoring time. Indicator functions for the ith observation are defined as follows:
δRi = 1 if right censored at time ti < xi , 0 otherwise;
δLi = 1 if left censored at time ti > xi , 0 otherwise;
(2.1)
δIi = 1 if interval censored with tLi < xi
tRi , 0 otherwise;
δEi = 1 if has an event observed exactly at ti = xi , 0 otherwise.
Note that δEi = 1 − (δRi + δLi + δIi ).
A subject is right censored (δRi = 1) when the subject was last known to be event free at time ti .
A subject is left censored (δLi = 1) when the subject is known to have had the event sometime prior to
ti but it is not known when the subject was previously event free or when the event occurred. A subject
is interval censored (δIi = 1) when it is known that the subject was event free at time tLi and to have
had the event sometime up to tRi > tLi . When time is defined with a definite start time at t = 0, then a
left-censored observation can be viewed as an interval-censored observation with tLi = 0 and tRi = ti . A
subject is known to have had the event exactly (δEi = 1) at time ti = xi , when the subject is known to
have been event free immediately prior to ti , i.e. tLi = tR−i = (tRi − t), ti = tRi .
Let f (t) denote the probability density function of the event times and F(t) denote the distribution
function. Under the assumption of independent censoring, the likelihood function for a sample of n inde-
pendent observations is
n 

 δ   δ     δ 
L= f i (ti )δEi Fi tLi Li 1 − Fi tRi Ri Fi tRi − Fi tLi Ii (2.2)
i=1

(Odell and others, 1992). To accomodate covariate effects, both quantitative and qualitative, let zi =
(z i1 · · · z i p )T be a vector of p fixed (e.g. baseline) covariates for the ith subject. Then also assume
that for the ith subject, additional time-dependent covariates are updated at a sequence of update times
τi0 , . . . , τiki , where τi0 is the time at which a subject enters follow-up (usually zero as in a clinical
trial). The set of update times {τi j } may differ among subjects. At the jth update time τi j of the ith
subject, let yi j = (yi j1 . . . yi jq )T denote a vector of q time-dependent covariate values that are up-
dated at that time. The covariate vector at the jth update time can also be denoted as yi (τi j ). Then let
Yi[τiki ] = (yi0 , . . . , yiki ) denote the complete sequence of time-dependent covariate values over time for
the ith subject, and let Yi[t] denote the sequence up to time t. Note that under this model, the covariates
are updated at discrete points in time. Different expressions are required if covariate values are updated
continuously, such as a function of time itself (see Section 5).
The likelihood then becomes

n 
 δ  δ
L= f i ti |zi , Yi[(ti )] Ei Fi ti |zi , Yi[(ti )] Li
i=1
 
δRi    
δI
Fi tRi |zi , Yi tR  − Fi tLi |zi , Yi tL 
i
× 1 − Fi ti |zi , Yi[(ti )] , (2.3)
i i

where f i (ti |zi , Yi[(ti )] ) is the event density for a specified event-time distribution conditional on the fixed
covariate values zi and the sequence of time-dependent covariate values Yi up to time ti . The function
Fi (ti |zi , Yi[(ti )] ) is the corresponding cumulative distribution function. Specification of the hazard func-
tion in terms of the covariates leads to a specification of the cumulative hazard and survival function prob-
abilities, conditional on the fixed covariates and the history of the time-dependent covariate processes for
a given subject.
602 Y. H. S PARLING AND OTHERS

Let γ and η be the coefficient vectors for the fixed and time-dependent covariates zi and yi j , respec-
tively, so that

γ  zi = γ1 z i1 + · · · + γ p z i p
η  yi j = η1 yi j1 + · · · + ηq yi jq . (2.4)

Let βi j = exp(θ + γ  zi + η  yi j ) designate a rate parameter conditional on the covariate values at update
time τi j . Note that with no time-dependent covariates, βi = exp(θ + γ  zi ).
A left-censored observation is assumed to have an initial value of the time-dependent covariate at time
τi j0 , assuming that the subject is known to be event free at that time. Otherwise, left-censored observations
must be excluded from the analysis. However, left-censored observations could be employed in a model
with only fixed covariate values, assuming that those values were determined prior to the event (e.g.
gender).
We now introduce a general form for the hazard function that with an additional parameter can span a
family of distributions such as the Weibull and log-logistic distributions, among others, as special cases.
For subject i, conditional on covariates measured at baseline and at time τi j , this hazard can be expressed
as
αβi j τiα−1
j
λ(τi j |zi , yi j ) = , (2.5)
[1 + βi j τiαj ]κ

where α > 0 and κ are general hazard function parameters. By construction, the rate parameter βi j is
assumed to be constant in the interval (τi j , τi( j+1) ], j = 0, 1, . . . , (ki − 1), for the ith subject.
Specific values for α and κ yield a specific distribution. In particular, α = 1 and κ = 0 yield a negative
binomial distribution for event times. More generally, κ = 0 yields a Weibull hazard, and the parameter
vectors γ and η are the change in the log relative risk per unit increase in z and yij , respectively. The hazard
function will be decreasing for 0 < α
1, constant for α = 1, and increasing for α > 1. Selecting κ = 1
yields a log-logistic hazard and γ and η are the change in the log odds ratio of cumulative incidences.
This hazard is decreasing for 0 < α
1. For α > 1 and fixed β, the hazard increases to a maximum
at time [(α − 1)/β]1/α , then decreases to zero as time approaches infinity. For α = 1.5 and κ = 0.5,
the hazard increases rapidly, plateaus, and then begins to slowly decline, similar to the hazard for the log
normal distribution. Thus, this family of hazards indexed by the additional parameter κ encompasses a
wide range of survival distributions.
For any time u ∈ (τi j , τi( j+1) ], let

βij u α 1
πi j (u) = , 1 − πi j (u) = (2.6)
1 + βij u α 1 + βij u α

for which the hazard function can be conveniently expressed as

α
λ(u|zi , yi j ) = πi j (u) [1 − πi j (u) ]κ−1 . (2.7)
u
This simplifies the expressions for the score equations and Hessian.
To describe the expression for the cumulative hazard, we impose the condition that the set of update
times {τi j } for the ith subject includes the event or censoring time (ti , tLi , tRi ) for that subject. If in fact,
as will most often be the case, the time-dependent covariate values are not updated exactly at an event or
censoring time, then the interval between two update times can be split into two intervals with an added
update time equal to the event or censoring time.
 Parametric survival models for interval-censored data 603

Let I{S} denote the indicator function where I{S} = 1 if S is true, 0 otherwise. Then, the cumulative
hazard at time t ∈ {τi j } for the ith subject is
  
i −1
k τi( j+1) αβi j u α−1
(t|zi , Yi(t) ) = I{τi( j+1) =t} du
τi j [1 + βi j u α ]κ
j=0
(2.8)
i −1
k
 
= I{τi( j+1) =t} µi j[τi( j+1) ] − µi j[τi j ] ,
j=0

where at time u
[1 + βi j u α ]1−κ (1 − πi j (u) )κ−1
µi j (u) = = . (2.9)
1−κ 1−κ
For κ = 1 the latter term is undefined. However, in this case, the antiderivative of the hazard is

µi j (u)|κ=1 = ln[1 + βi j u α ] (2.10)

and using l’Hospital’s rule with implicit differentiation it follows that limκ→1 µi j (u) = µi j (u)|κ=1 . The
resulting expressions for the hazard and survivor function equal those for the log-logistic model. Thus, the
expression in (2.10) should be employed to compute the gradient and Hessian in cases where κ is specified
to be 1, or the interim estimate in an iterative computation yields a value close to 1, i.e. 
κ ∈ (1 ± ε) for
some small ε.
For a subject with an observed event time or a right- or left-censored event time, the cumulative hazard
is evaluated at t = ti ; and for an interval-censored observation (δIi = 1) at both t = tLi and at t = tRi . In
the simple case with no time-dependent covariates, the term βi = exp(θ + γ  zi ) is constant over time for
the ith individual. The cumulative hazard function for the ith subject evaluated at time t is then expressed
as
 t
αβi u α−1 [1 + βi t α ]1−κ − 1
(t|zi ) = α κ
du = . (2.11)
0 [1 + βi u ] 1−κ
From the general expression for the likelihood in (2.3) and that for the hazard function with fixed and
time-dependent covariates in (2.5), the score equations and Hessian matrix can be derived. The expressions
are presented in the Appendix. These can then be used to provide the maximum likelihood estimates of
the model parameters and the estimated variance of the estimates using an iterative procedure such as
the Newton–Raphson algorithm or variations thereof. Alternatively, a derivative-free iterative procedure
may be employed to fit the model and estimate the Hessian. The program available from the authors
obtains the model estimates using the Newton–Raphson ridge optimization method (cf. Press and others,
1992) to maximize the log-likelihood function through the SAS IML function NLPNRR (SAS, 1999).
Initial values are obtained by fitting a Weibull accelerated failure time model using SAS PROC LIFEREG
and then transforming the time acceleration parameter estimates to Weibull risk model estimates (cf.
Lachin, 2000). At each iteration, the Hessian is estimated by the the SAS IML function NLPFDD that
uses the algorithm of Gill and others (1983) based on finite difference equations with central difference
approximations. The final estimate of the Hessian when the model has converged is then used to estimate
the observed information matrix and the covariance matrix of the coefficient estimates.
Using the theorem from Lehmann (1983, pp. 429–30), Sparling (2002) provides a proof that the result-
ing estimates are asymptotically normally distributed about the true values with a covariance matrix that
can be consistently estimated from the estimated observed information matrix. The resulting estimates
then provide a basis for confidence interval estimates and Wald or likelihood ratio tests of significance.
604 Y. H. S PARLING AND OTHERS

3. M ODEL DIAGNOSTICS
The shape of the assumed hazard function is determined by the values of α and κ. Specific hypotheses
such as H0 : κ = 0 (i.e. a Weibull model) or H0 : κ = 1 (i.e. a log-logistic model) can be tested using
a Wald or likelihood ratio test. If, for example, it is desired to employ a Weibull model, then that model
could be fit if the corresponding test is not significant by setting κ = 0 in all the equations.
Alternatively, the model can be fit for other values of κ, and the adequacy of those values can be
assessed by examining the values of the log likelihood function, or the optimal value κ̂ could be esti-
mated from the model. In this case, however, the coefficient estimates γ̂ or η̂ no longer have a convenient
interpretation as the log relative hazards (κ = 0) or as the log cumulative incidence odds ratios (κ = 1).
The functional form of the covariate effects can be explored using spline functions (Smith, 1979;
Ramsay, 1988).
Following Therneau and others (1990), the deviance for the family of models herein can be described
as
D = 2 sup{ln[L(saturated)] − ln[L(fitted)]}, (3.1)
h

where h is the set of subject-specific implied parameters. Let λ̃(t), S̃(t), and (t) ˜ denote the hazard, sur-
vival, and cumulative hazard functions, respectively, from the saturated model; and let λ̂(t), Ŝ(t), and (t) ˆ
denote the estimates from the fitted model. Let h i be the individual per-subject estimates of the parameter
vector  = (α, θ, κ, γ , η ).
The Appendix shows that D = −2 ln[L(fitted)] for left-censored, right-censored and interval-censored
observations. For non-censored observations where δEi = 1 and ti is the event time for individual i, then
    
D = 2 sup ln λ̃ ti |zi , Yi[(ti )] −  ˜ ti |zi , Yi[(ti )]
h
   
− ln λ̂ ti |zi , Yi[(ti )] + ˆ ti |zi , Yi[(ti )] . (3.2)

The first two terms are not readily obtained because the subject-specific estimate of the cumulative hazard
function at time ti is a function of the subject-specific estimate of the hazard function at time ti . However,
an approach similar to that in Therneau and others (1990) can be used. For the non-censored case in the
current model, the derivatives of the log-likelihood function with respect to each element of the parameter
vector  = (α, θ, κ, γ , η ) can be solved under the constraint that the second derivatives given are negative.
This approach involves solving a set of simultaneous equations per subject by numerical methods such as
Newton–Raphson ridge optimization.
The deviance statistic may indicate that the model does not fit the data well because important co-
variates have been omitted, or components of the model are mis-specified such as the variance of the
responses as a function of the conditional expectation. In the latter case, the “information sandwich” can
be used to provide an estimate of the covariance matrix that is robust to mis-specification (Royall, 1986).

4. G LUCOSE EXPOSURE AND THE RISK OF RETINOPATHY IN DIABETES

In longitudinal studies or clinical trials, subjects may undergo an examination or procedure at regularly
scheduled follow-up visits to determine whether disease progression has occurred. In this case, the time to
specific outcomes is interval censored by the schedule of follow-up assessments. Furthermore, covariates
related to the outcome may also be assessed periodically over time.
For example, the study of the Epidemiology of Diabetes Interventions and Complications (EDIC) is a
follow-up observational study of the subjects who had previously participated in the Diabetes Control and
Complications Trial (DCCT). Men and women aged 13–39 years with type 1 diabetes mellitus (T1DM)
 Parametric survival models for interval-censored data 605

were enrolled in the DCCT between 1983–1989. Patients were recruited into two cohorts—a primary
prevention cohort with no pre-existing complications and a secondary intervention cohort with minimal
complications present.
Patients were randomized to either intensive or conventional treatment and were followed for an av-
erage of 6.5 years. Intensive therapy was aimed at maintaining near normal levels of blood glucose while
conventional therapy had no such glucose target. The DCCT Research Group (1993) showed that inten-
sive therapy markedly reduced the risk of progression of diabetes complications, principally retinopathy
(diabetic retinal abnormalities, potentially leading to blindness) that was assessed from a retinal evaluation
every 6 months.
The level of glucose exposure (glycemia) over the preceding 6–8 weeks is provided by the hemoglobin
A1c (HbA1c ), expressed as the percentage of all hemoglobin (red cells) that have been glycosylated
through exposure to glucose molecules in blood, the half-life of hemoglobin being 6–8 weeks. The his-
tory of glycemia before study entry was represented by the level of HbA1c on initial screening and the
pre-existing duration of diabetes, and the history of glycemia during the study by the mean level of HbA1c
during the study and the duration of follow-up. The DCCT Research Group (1995) showed that the life-
time history of glycemia represented by all four factors was the dominant determinant of the risk of com-
plications, and that the group differences in the updated current mean HbA1c (a time-dependent covariate)
during the study explained virtually all the effect of DCCT treatment group on complications.
At the close of the DCCT in 1993, all subjects were referred to their personal physicians for care and
followed annually during EDIC at which time HbA1c was assessed. During EDIC, the levels of HbA1c
were approximately equal in the two former DCCT treatment groups. Retinopathy was assessed in about
one-quarter of the patients at years 1–3 timed in relation to the original date of entry (i.e. 4, 8, or 12
years since entry), and in all subjects at year 4. Thus, the times of progression of retinopathy are interval
censored with staggered intervals. One objective of EDIC is to assess the long-term effects of levels of
glycemia during the DCCT and EDIC on risk of further progression of retinopathy from the levels present
at the end of the DCCT. The DCCT/EDIC Research Group (2000) showed that former intensive ther-
apy reduced the risks of further progression of retinopathy during EDIC. The remaining question is the
extent to which glycemic levels during DCCT and EDIC are associated with risk of further retinopathy
progression during EDIC.
Fixed (baseline) covariates include primary versus secondary cohort at EDIC baseline (1 if in the
primary cohort, 0 if secondary), the duration of T1DM in months and HbA1c level (%) on initial screening
that represent the pre-DCCT level of glycemia, and the mean HbA1c (%) during the DCCT and months
duration of follow-up in the DCCT that represent the level of glycemia during the DCCT. The time-
dependent covariate is the updated current mean HbA1c during the EDIC, i.e. the value at year 1, the
mean of years 1 and 2, then years 1–3, updated at the time of each successive measurement. In those
cases where the EDIC annual HbA1c is missing (not measured), the mean value from the previous visit
is carried forward. Nine patients are deleted because they are missing all HbA1c values. Of the remaining
1316 subjects, 1085 are right censored and 231 have interval-censored times of progression.
Approximately half the subjects within each treatment group were enrolled (by design) into the pri-
mary and secondary cohorts. On entry, the mean duration of diabetes was 69 ± 49 (SD) months and the
mean level of HbA1c was 9.0 ± 1.6%. The mean duration of follow-up during the DCCT was 73 ± 20
months and the mean level of HbA1c was 8.1 ± 1.4% during the DCCT. The mean level of HbA1c during
EDIC was 8.2 ± 1.3%.
Prior knowledge about the distribution of progression of retinopathy suggests that the event times
might follow a Weibull distribution. Figure 1 presents the cumulative incidence of retinopathy progres-
sion estimated using the Turnbull (1976) empirical estimate and using the Weibull model-based estimate
separately within each treatment group and shows that the two estimates are superimposable. Under a
Weibull distribution assumption, the regression model coefficients have a log hazard ratio or log relative
606 Y. H. S PARLING AND OTHERS

Fig. 1. Cumulative incidence of retinopathy progression over 4 years of EDIC within the intensive and conventional
groups estimated using the Turnbull (1976) empirical estimate and using a Weibull model.

Table 1. Weibull model estimates and tests

MLE SE Wald Likelihood ratio relative risk (95% CI)
χ12 p χ12 p
α̂ Intercept −19.841 2.4860 63.70 <0.0001 — — —
γ̂1 retinopathy at baseline −0.2782 0.1965 2.003 0.1569 2.006 0.1566 0.75715 (0.515, 1.113)
primary v. secondary
γ̂2 entry diabetes duration 0.0041 0.0020 4.142 0.0418 4.436 0.0352 1.004 (1.0002, 1.008)
per month increase
γ̂3 entry HbA1c −0.1714 0.0777 4.867 0.0273 9.639 0.0020 0.842 (0.723, 0.981)
per unit (%) increase
γ̂4 DCCT follow-up 0.0062 0.0036 3.008 0.0828 2.989 0.0839 1.006 (0.999, 1.013)
per month increase
γ̂5 DCCT mean HbA1c 0.7141 0.1111 41.353 <0.0001 172.608 <0.0001 2.042 (1.643, 2.540)
per unit (%) increase
η̂1 EDIC HbA1c 0.1715 0.0762 5.070 0.0243 8.437 0.0037 1.187 (1.022, 1.378)
per unit (%) increase
MLE, maximum likelihood estimates.

risk interpretation. Analyses using regression splines to represent the effect of the EDIC HbA1c showed
that a simple linear effect was satisfactory (see Sparling, 2002, for details).
Table 1 presents the maximum likelihood estimates of the parameters, their variances, Wald tests, and
likelihood ratio tests for the Weibull model. The deviance for the fitted Weibull model is D = 1292.3
with df = 1316 − 6 − 1 = 1309 and p = 0.624 that indicates adequate fit of the model. Further, the
deviance/df = 0.99 does not indicate any overdispersion. Thus, a robust estimate of the variance–covariance
matrix is not employed.
Primary versus secondary cohort and duration of follow-up in the DCCT do not have any meaningful
effects on progression of retinopathy adjusted for the other covariates. T1DM duration and the HbA1c on
entry have significant effects whereas the effect of the duration of follow-up in the DCCT is not significant.
By far the greatest effect is contributed by the mean level of HbA1c during the DCCT with a 2.04-fold
increase in the risk of progression of retinopathy per unit increase in DCCT mean HbA1c percent (95%
CI: 1.64, 2.54, p < 0.0001). Interestingly, the time-dependent EDIC mean HbA1c over 4 years following
the DCCT has a significant effect on risk, but with a much smaller 1.19-fold increase in risk per unit
increase in HbA1c percent (95% CI: 1.02, 1.38, p = 0.0243). The finding that the glycemic exposure
during the DCCT persists and outweighs the level of HbA1c during the first 4 years of EDIC has lead to
the hypothesis of metabolic memory that effects of hyperglycemia are long lasting.
 Parametric survival models for interval-censored data 607

An unrestricted model provides a shape parameter estimate  κ = 0.3247 with SE = 0.7326 and
95% confidence limits (−1.11, 1.76). The Wald test of the hypothesis H0κ : κ = 0 yields p = 0.6576,
the likelihood ratio test yields p = 0.6878. These results indicate that the distribution of event times
conditional on covariates does not significantly deviate from a Weibull distribution. In this model,  α =
−20.5795 with SE = 2.2437, close to the values in the Weibull model. Further, Corr( κ,α ) = 0.80.
Because the observed cumulative incidence at the close of follow-up in the cohort is still low (∼0.20),
there is little information to reliably estimate the shape parameter κ.
Similar covariate effects were obtained in a logistic regression analysis of the subset of subjects who
were assessed at 4 years (The DCCT/EDIC Research group, 2000). That analysis described the cross-
sectional association of the mean HbA1c over 4 years with the prevalence of progression at 4 years,
whereas the analysis herein shows the more desirable prospective association between the time-dependent
HbA1c levels and the incidence (risk) of progression using all visits in all patients.

5. D ISCUSSION
We describe a family of parametric regression models for survival data that allows for fixed (e.g. baseline)
and/or time-dependent covariates with mixtures of left, right, and interval censoring. The model is fit
using standard maximum likelihood estimation from the full likelihood for which all the conditions for
convergence have been rigorously proven in Sparling (2002). Thus, the inferences based on the model are
based on large sample approximations. The model should be used with caution in instances with a small
sample size or a small total number of non-censored observations (events).
The family of models is characterized by an additional parameter (κ) that allows fitting a Weibull
model (κ = 0) or a log-logistic model (κ = 1) as special cases, or allows for the optimal value of
this parameter to also be estimated from the observed data. In the latter case, however, the ability to
differentiate a Weibull from a log-logistic model, or to accurately estimate the optimal value, will be
roughly proportional to the observed cumulative incidence. If the observed cumulative incidence is low,
as the example herein (Figure 1), there is inadequate information about the true shape of the hazard
function to allow precice estimation of κ.
The family of models presented herein allows for incorporation of time-dependent covariates for
which the values are updated or change at discrete points in time. For a time-dependent covariate that
changes continuously over time, such as a function of time itself, the integrals in the cumulative hazard
function may not be expressable in closed form. For example, rather than having a vector of discrete time-
dependent covariates yi j with distinct values at each of the update times {τi j }, a covariate that is a function
of time itself would have continuously changing time-dependent values specified as yi (u), u > 0. Then,
the hazard function is no longer constant over intervals of time. Consequently, cumulative hazard at time
t for the ith subject is
 t  
α e(θ +γγ zi +ηη yi (u)) u α−1
(t) = (θ +γγ  zi +ηη  yi (u)) u α ]κ
du. (5.1)
0 [1 + e

While such an expression may not have a closed form, the expression could be evaluated numerically.
Many covariates will in fact be a function of time, in theory, including biochemical or biological
measures that change from day to day, such as the HbA1c measured in the EDIC example above. However,
in practice, as in the EDIC, all that is known are the time-dependent covariate’s values for a finite set of
update times. In this situation, the time-dependent covariate effect must be interpreted in the context
of the study design and the follow-up schedule of assessments at which changes in the covariate are
observed. For the EDIC example, the estimated relative risk per unit increase in HbA1c is that associated
with differences between values updated approximately annually, the specified schedule during EDIC.
608 Y. H. S PARLING AND OTHERS

The relative risks (or hazard ratios) then have a prospective interpretation when applied to other settings
with the same (or a similar) schedule of update times.
As with any model that employs fixed and time-dependent covariates, caution should be taken in the
interpretation of a fixed (e.g. baseline) effect when a time-dependent covariate is influenced by that fixed
effect, such as treatment group (see, for example, Kalbfleisch and Prentice, 2002). In this situation, the
model describes the risks of the event given the time-dependent covariate values and treatment group.
If the effect of the treatment group is predominantly reflected in the time-dependent covariate process,
this type of analysis would show a minimal treatment group effect. However, an analysis not including
the time-dependent covariate might show a substantial treatment group effect. In this case, the proper
interpretation is that treatment group has an effect on risk, and also an effect on the time-dependent
covariate Y , but that after adjusting for Y, group has little further effect. This suggests that factors related
to Y reflect the underlying mechanism by which treatment has an effect on risk. In fact, such analyses and
findings are useful to illustrate the mechanism by which fixed baseline covariates such as treatment group
have an effect on risk.
A SAS IML macro written by Oliver Bautista and Yvonne Sparling is available from www.bsc.
gwu.edu under the link to downloadable software.

ACKNOWLEDGMENTS
This work was supported by a grant from the National Cancer Institute and by a contract from the National
Institute of Diabetes, Digestive, and Kidney Diseases for the study of the EDIC. Conflict of Interest: None
declared.

APPENDIX
A.1 Fitting the model
From (2.3), the log likelihood is
n 
       
ln(L) = δEi ln f i ti |zi , Yi[(ti )] + δLi ln Fi ti |zi , Yi[(ti )] + δRi ln 1 − Fi ti |zi , Yi[(ti )]
i=1
   

+ δIi ln Fi tRi |zi , Yi tR  − Fi tLi |zi , Yi tL 
i i

n 
     
= δEi ln λ ti |zi , Yi[(ti )] −  ti |zi , Yi[(ti )]
i=1
      
+ δLi ln 1 − exp − ti |zi , Yi[(ti )] − δRi  ti |zi , Yi[(ti )]
     

+δIi ln exp − tLi |zi , Yi tL  − exp − tRi |zi , Yi tR  . (A.1)
i i

In order to simplify the expressions, the conditioning on zi , Yi[(ti )] is omitted for the expressions for λ
and . For any parameter φ, where φ ∈ (α, κ, γ , η ), the score equation is
n   
∂ ln L  ∂ ln λ(ti ) ∂
U (φ) = = δEi − (ti )
∂φ ∂φ ∂φ
i=1
 Parametric survival models for interval-censored data 609

 ∂
  
∂φ (1 − exp{−(ti )}) ∂
+ δLi − δRi (ti )
1 − exp{−(ti )} ∂φ

 ∂
  
∂φ exp{−(t L i )} − exp{−(tRi )}
+ δIi . (A.2)
exp{−(tLi )} − exp{−(tRi )}

The derivatives of πi j (u) , of the log hazard and the increments in cumulative hazard with respect to
each parameter φ ∈ (α, κ, γ , η ) at covariate update time u, are

φ ∂πi j (u) /∂φ ∂ ln λ(u|zi , yi j )/∂φ ∂µi j (u) /∂φ

α πi j (u) [1 − πi j (u) ] log(u) 1
α + [1 − κπi j (u) ] log(u) [(1 − κ)πi j (u) ln τi j ]µi j (u)

βi j 1
βi j πi j (u) [1 − πi j (u) ] 1
βi j [1 − κπi j (u) ] 1
βi j (1 − κ)πi j (u) µi j (u)

θ πi j (u) [1 − πi j (u) ] [1 − κπi j (u) ] (1 − κ)πi j (u) µi j (u) (A.3)

γ πi j (u) [1 − πi j (u) ]z i [1 − κπi j (u) ]z i (1 − κ)πi j (u) µi j (u) z i

η πi j (u) [1 − πi j (u) ]yi(u) [1 − κπi j (u) ]yi(u) (1 − κ)πi j (u) µi j (u) yi(u)

κ 0 ln[1 − πi j (u) ] ln[1 − πi j (u) ] + 1

1−κ µi j (u)

where yi(u) is the value for the th time-dependent covariate value for the ith subject at update time u. The
derivatives of terms involving the cumulative hazard at event or censoring time t are then provided by


ki  
∂(t) ∂µi j[τi( j+1) ] ∂µi j[τi j ]
= I{τi( j+1) =t} −
∂φ ∂φ ∂φ
j=0

 
∂ e−(t) −(t) ∂(t)
= −e
∂φ ∂φ
 
∂[1 − e−(t) ] −(t) ∂(t)
=e . (A.4)
∂φ ∂φ
The score equation for parameter φ is then obtained by evaluating the respective derivatives for each
subject. The maximum likelihood estimates of the model parameters are those values ( γ ,
κ, 
α , η ) such
that the joint set of score equations are equal to zero when all parameters are fixed at these values.
The Hessian for any parameters φ ∈  = (α, κ, γ , η ) and ϕ ∈  has elements
  
∂ 2 ln L 
n
∂ 2 ln λ(ti ) ∂2
H (φ, ϕ) = = δEi − (ti )
∂φ∂ϕ ∂φ∂ϕ ∂φ∂ϕ
i=1

⎡   ⎤
∂2 ∂ −(ti ) ] ∂ −(ti ) ]
∂φ [1 − e ∂ϕ [1 − e
−(ti ) ]
∂φ∂ϕ [1 − e
+ δLi ⎣ − ⎦
1 − e−(ti ) (1 − e−(ti ) )2
610 Y. H. S PARLING AND OTHERS

  ⎡    

∂2
∂2
∂φ∂ϕ e− tLi
− e− tRi
− δRi (ti ) + δIi ⎣    
∂φ∂ϕ e− tLi
− e− tRi

    
     
 ⎤⎫
∂
e− tLi
− e− tRi ∂
e− tLi
− e − tRi ⎪
⎬
∂φ ∂ϕ ⎥
−    
2 ⎦ . (A.5)
⎪
⎭
e− tLi − e− tRi

The partial second derivatives with respect to πi j (u) , ln λ(ti |zi , , yi j ), and µi j (u) are presented in Tables
2–4, respectively.
The estimated Hessian matrix then has elements H( ) based on the vector of parameter estimates

 = ( α ,κ , γ, 
η ). The observed information is i( ) = −H( ). If φ is the th term in the parameter

vector  , then the estimated variance of φ is obtained as
(φ
V ),
) = i ( (A.6)
) is the (, ) term of i(
where i ( )−1 .

A.2 The deviance

To simplify the derivation, we use S̃(ti ) to refer to S̃(ti |zi , Yi[(ti )] ). In the left-censored case (δLi = 1 and
ti is the left-censoring time for individual i),

D = 2 sup(ln[1 − S̃(ti )]) − ln[1 − Ŝ(ti )]
h
   
= 2 ln 1 − inf S̃(ti ) − ln[1 − Ŝ(ti )]
h

= 2 ln{1 − 0} − ln{1 − Ŝ(ti )}

= − 2 ln[L(fitted)]. (A.7)

Table 2. The second derivatives of the term πi j (u) with respect to the elements of the parameter vector,
∂ 2 πi j (u) /(∂φ∂ψ)

α βi j θ
α πi j (u) (1 − πi j (u) )(1 − 2πi j (u) )(ln[u])2

βi j 1
βi j πi j (u) (1 − πi j (u) )(1 − 2πi j (u) ) − 2
πi2j (u) (1 − πi j (u) )
βi2j

θ πi j (u) (1 − πi j (u) )(1 − 2πi j (u) ) − β2i j πi2j (u) (1 − πi j (u) ) πi j (u) (1 − πi j (u) )(1 − 2πi j (u) )

γ z i πi j (u) (1 − πi j (u) )(1 − 2πi j (u) ) − 2z

βi j πi j (u) (1 − πi j (u) )
i 2 z i πi j (u) (1 − πi j (u) )(1 − 2πi j (u) )

η yi(u) πi j (u) (1 − πi j (u) )(1 − 2πi j (u) ) − 2yβii (u)

j
πi2j (u) (1 − πi j (u) ) yi(u) πi j (u) (1 − πi j (u) )(1 − 2πi j (u) )

κ 0 0 0
 Parametric survival models for interval-censored data 611

Table 2. Continued
γ η κ
γ z i πi j (u) (1 − πi j (u) )(1 − 2πi j (u) )
2 0

η z i yi(u) πi j (u) (1 − πi j (u) )(1 − 2πi j (u) ) 2

yi(u) πi j (u) (1 − πi j (u) )(1 − 2πi j (u) ) 0

κ 0 0 0
and in addition

∂ 2 πi j
= z i z im πi j (1 − πi j )(1 − 2πi j ),  = m
∂γ ∂γm

∂ 2 πi j
= yi(u) yim(u) πi j (1 − πi j )(1 − 2πi j ),  = m
∂η ∂ηm

∂ 2 πi j
= z i yim(u) πi j (1 − πi j )(1 − 2πi j ),  = m
∂γ ∂ηm

Table 3. The second derivatives of the log hazard function with respect to the elements of the parameter
vector, ∂ 2 ln λ(u|z i , yi j )/(∂φ∂ψ)

α βi j θ
α − 12 − κπi j (u) (1 − πi j (u) )(ln[u])2
α
βi j − βκ πi j (u) (1 − πi j (u) ) ln[u] − β1 (1 − κπi2j (u) )
ij ij

θ −κπi j (u) (1 − πi j (u) ) ln[u] − βκ πi j (u) (1 − πi j (u) ) −κπi j (u) (1 − πi j (u) )

ij

γ −κz i πi j (u) (1 − πi j (u) ) ln[u] − βκ z i πi j (u) (1 − πi j (u) ) −κz i πi j (u) (1 − πi j (u) )

ij

η −κ yi (τi j )πi j (u) (1 − πi j (u) ) ln[u] − βκ yi(u) πi j (u) (1 − πi j (u) ) −κ yi(u) πi j (u) (1 − πi j (u) )
ij

κ −πi j (u) ln[u] −πi j (u) /βi j −πi j (u)

γ η κ
γ −κz i πi j (u) (1 − πi j (u) )
2

η −κz i yil(u) πi j (u) (1 − πi j (u) ) −κ yi(u)

2 πi j (u) (1 − πi j (u) )

κ −πi j (u) z i −πi j (u) yi(u) 0

and in addition

∂ 2 ln λ(u|zi , yi j )
= −κz i z im πi j (u) (1 − πi j (u) ),  = m
∂γ ∂γm

∂ 2 ln λ(u|zi , yi j )
= −κ yi(u) yim(u) πi j (u) (1 − πi j (u) ),  = m
∂η ∂ηm

∂ 2 ln λ(u|zi , yi j )
= −κz il yim(u) πi j (u) (1 − πi j (u) ),  = m
∂γl ∂ηm
 612
Table 4. The second derivatives of the terms µi j (u) required for the second derivative of the cumulative hazard function with respect to the
elements of the parameter vector, ∂ 2 µi j (u) /(∂φ∂ψ)

α βi j θ
α [(1 − κ)πi j (u) (1 − κπi j (u) )(ln[u])2 ]µi j (u)

1−κ π −κ(1−κ) 2
βi j βi j i j (u) (1 − κπi j (u) )(ln[u]) µi j (u) πi j (u) µi j (u)
βi2j
(1−κ)
θ [(1 − κ)πi j (u) (1 − κπi j (u) )(ln[u])]µi j (u) βi j πi j (u) (1 − πi j (u) κ)µi j (u) (1 − κ)πi j (u) (1 − κπi j (u) )µi j (u)

(1−κ)
γ [z i (1 − κ)πi j (u) (1 − κπi j (u) )(ln[u])]µi j (u) βi j πi j (u) (1 − πi j (u) κ)µi j (u) z i z i (1 − κ)πi j (u) (1 − κπi j (u) )µi j (u)

(1−κ)
η [yi (τi j )(1 − κ)πi j (u) (1 − κπi j (u) )(ln[u])]µi j (u) βi j πi j (u) (1 − πi j (u) κ)µi j (u) yi(u) yi(u) (1 − κ)πi j (u) (1 − κπi j (u) )µi j (u)



κ [(1 − κ)πi j (u) (ln[u]) ln(1 − πi j (u) )]µi j (u) (1−κ) 1 (1 − κ) ln(1 − πi j (u) ) + 1−κ 1 − 1 πi j (u) µi j (u)
βi j πi j (u) ln(1 − πi j (u) ) + 1−κ µi j (u)

γ η κ
γ 2 (1 − κ)π
z i i j (u) (1 − κπi j (u) )µi j (u)

ij
η z i yil (τi j )(1 − κ)πi j (u) (1 − κπi j (u) )µi j (u) 2 (τ )(1 − κ)π
yi i j (u) (1 − κπi j (u) )µi j (u)


 
  2 
1 1 1 1
κ z i (1 − κ) ln(1 − πi j (u) ) + 1−κ − 1 πi j (u) µi j (u) yi (τi j ) (1 − κ) ln(1 − πi j (u) ) + 1−κ − 1 πi j (u) µi j (u) ln(1 − πi j (u) ) + 1−κ + (1−κ)2
µi j (u)
Y. H. S PARLING AND OTHERS

and in addition

∂ 2 µi j (u)
= z i z im (1 − κ)πi j (u) (1 − κπi j (u) )µi j (u) ,  = m
∂γ ∂γm

∂ 2 µi j (u)
= yi yim (τi j )(1 − κ)πi j (u) (1 − κπi j (u) )µi j (u) ,  = m
∂η ∂ηm

∂ 2 µi j (u)
= z i yim (τi j )(1 − κ)πi j (u) (1 − κπi j (u) )µi j (u) ,  = m
∂γl ∂ηm
 Parametric survival models for interval-censored data 613

For the right-censored case (δRi = 1 and ti is the right-censoring time for individual i),

D = 2 sup(ln[ S̃(ti )]) − ln[ Ŝ(ti )] = 2{ln(1) − ln[ Ŝ(ti )]} = −2[L(fitted)]. (A.8)
h

In the interval-censored case (δIi = 1 and tLi is the left endpoint of the censoring interval and tRi is the
right endpoint of the censoring interval for individual i),
   
 
    

D = 2 sup ln S̃ tLi − S̃ tRi − ln Ŝ tLi − Ŝ tRi . (A.9)

h
         !
Since 0
S̃ tLi − S̃ tRi
1, then suph ln S̃ tLi − S̃ tRi = 0 and
    

D = 2 0 − ln Ŝ tLi − Ŝ tRi = −2 ln[L(fitted)]. (A.10)

R EFERENCES
B ETENSKY, R. A., L INDSEY, J. C., RYAN , L. M. AND WAND , M. P. (2002). A local likelihood proportional hazards
model for interval censored data. Statistics in Medicine 21, 263–75.
F INKELSTEIN , D. M. (1986). A proportional hazards model for interval-censored failure time data. Biometrics 42,
845–54.
F INKELSTEIN , D. M. AND W OLFE , R. A. (1985). A semiparametric model for regression analysis of interval-
censored failure time data. Biometrics 41, 933–45.
G ILL , E. P., M URRAY, W., S AUNDERS , M. A. AND W RIGHT, M. H. (1983). Computing forward-difference inter-
vals for numerical optimization. SIAM Journal on Scientific Computing 4, 310–21.
G OETGHEBEUR , E. AND RYAN , L. (2000). Semiparametric regression analysis of interval-censored data. Biometrics
56, 1139–44.
K ALBFLEISCH , J. D. AND P RENTICE , R. L. (2002). The Statistical Analysis of Failure Time Data, 2nd edition.
New York: John Wiley and Sons, Inc.
KOOPERBERG , C. AND C LARKSON , D. B. (1997). Hazard regression with interval-censored data. Biometrics 53,
1485–94.
KOOPERBERG , C., S TONE , C. J. AND T RUONG , Y. K. (1995). Hazard regression. Journal of the American Statistical
Association 90, 78–94.
L ACHIN , J. M. (2000). Biostatistical Methods. The Assessment of Relative Risk. New York: John Wiley and Sons, Inc.
L EHMANN , E. L. (1983). Theory of Point Estimation. New York: John Wiley and Sons, Inc.
L INDSEY, J. K. (1998). A study of interval censoring in parametric regression models. Lifetime Data Analysis 4,
329–54.
O DELL , P. M., A NDERSON , K. M. AND D’AGOSTINO , R. B. (1992). Maximum likelihood estimation for interval-
censored data using a Weibull-based accelerated failure time model. Biometrics 48, 951–9.
P RESS , W. H., V ETTERLING , W. T., T EUKOLSKY, S. A. AND F LANNERY, B. P. (1992). Numerical Recipies in C.
London: Cambridge University Press.
R ABINOWITZ , D., B ETENSKY, R. AND T SIATIS , A. A. (2000). Using conditional logistic regression to fit propor-
tional odds models to interval censored data. Biometrics 56, 511–8.
R ABINOWITZ , D., T SIATIS , A. AND A RAGON , J. (1995). Regression with interval-censored data. Biometrika 82,
501–13.
614 Y. H. S PARLING AND OTHERS

R AMSAY, J. O. (1988). Monotone regression splines in action (with discussion). Statistical Science 3, 425–61.
ROYALL , R. M. (1986). Model robust inference using maximum likelihood estimators. International Statistical
Review 54, 221–6.
SAS I NSTITUTE I NC . (1999) SAS/IML Users’ Guide, Version 8. Cary, NC: SAS Institute Inc.
S EAMAN , S. R. AND B IRD , S. M. (2001). Proportional hazards model for interval-censored failure times and
time-dependent covariates: application to hazard of HIV infection of injecting drug users in prison. Statistics
in Medicine 20, 1855–70.
S MITH , P. L. (1979). Splines as a useful and convenient statistical tool. The American Statistician 33, 57–62.
S PARLING , Y. H. (2002). Parametric survival models for interval-censored data with time-dependent covariates,
[Doctoral Dissertation]. The George Washington University, Washington, DC.
T HE DCCT/EDIC R ESEARCH G ROUP. (2000). Retinopathy and nephropathy in patients with type 1 diabetes four
years after a trial of intensive therapy. The New England Journal of Medicine 342, 381–9.
T HE DCCT R ESEARCH G ROUP. (1993). The effect of intensive treatment of diabetes on the development and
progression of long-term complications in insulin-dependent diabetes mellitus. The New England Journal of
Medicine 329, 977–86.
T HE DCCT R ESEARCH G ROUP. (1995). The relationship of glycemic exposure (HbA1c) to the risk of development
and progression of retinopathy in the diabetes control and complications trial. Diabetes 44, 968–83.
T HERNEAU , T. M., G RAMBSCH , P. M. AND F LEMING , T. R. (1990). Martingale-based residuals for survival
models. Biometrika 77, 147–60.
T URNBULL , B. W. (1976). The empirical distribution function with arbitrarily grouped, censored and truncated data.
Journal of the Royal Statistical Society, Series B 38, 290–5.
YOUNES , N. AND L ACHIN , J. (1997). Link-based models for survival data with interval and continuous time censor-
ing. Biometrics 53, 1199–211.

[Received March 31, 2005; revised January 27, 2006; accepted for publication March 8, 2006]