Research Article  Open Access
Magdalena Murawska, Dimitris Rizopoulos, Emmanuel Lesaffre, "A TwoStage Joint Model for Nonlinear Longitudinal Response and a TimetoEvent with Application in Transplantation Studies", Journal of Probability and Statistics, vol. 2012, Article ID 194194, 18 pages, 2012. https://doi.org/10.1155/2012/194194
A TwoStage Joint Model for Nonlinear Longitudinal Response and a TimetoEvent with Application in Transplantation Studies
Abstract
In transplantation studies, often longitudinal measurements are collected for important markers prior to the actual transplantation. Using only the last available measurement as a baseline covariate in a survival model for the time to graft failure discards the whole longitudinal evolution. We propose a twostage approach to handle this type of data sets using all available information. At the first stage, we summarize the longitudinal information with nonlinear mixedeffects model, and at the second stage, we include the Empirical Bayes estimates of the subjectspecific parameters as predictors in the Cox model for the time to allograft failure. To take into account that the estimated subjectspecific parameters are included in the model, we use a Monte Carlo approach and sample from the posterior distribution of the random effects given the observed data. Our proposal is exemplified on a study of the impact of renal resistance evolution on the graft survival.
1. Introduction
Many medical studies involve analyzing responses together with event history data collected for each patient. A wellknown and broadly studied example can be found in AIDS research, where CD4 cell counts taken at different time points are related to the time to death. These data need to be analyzed using a joint modeling approach in order to properly take into account the association between the longitudinal data and the event times. The requirement for a joint modeling approach is twofold. Namely, when focus is on the longitudinal outcome, events cause nonrandom dropout that needs to be accounted for in order to obtain valid inferences. When focus is on the event times, the longitudinal responses cannot be simply included in a relative risk model because they represent the output of an internal timedependent covariate [1].
In this paper, we focus on a setting that shares some similarities with the standard joint modeling framework described above, but also has important differences. In particular, we are interested in the association between longitudinal responses taken before the actual followup for the timetoevent has been initiated. This setting is frequently encountered in transplantation studies, where patients in the waiting list provide a series of longitudinal outcomes that may be related to events occurring after transplantation. A standard analysis in transplantation studies is to ignore the longitudinal information and use only the last available measurement as a baseline covariate in a model for the allograft survival. It is, however, evident that such an approach discards valuable information. An alternative straightforward approach is to put all longitudinal measurements as covariates in the survival model. Nevertheless, there are several disadvantages with this approach. First, it would require spending many additional degrees of freedom, one for each of the longitudinal measurements. Second, patients with at least one missing longitudinal response need to be discarded, resulting in a great loss of efficiency. Finally, we may encounter multicollinearity problems due to the possibly high correlation between the longitudinal measurements at different time points.
Nowadays, when it comes to measuring the association between a longitudinal marker and an eventtime outcome, a standard approach is to use the joint model postulated by Faucett and Thomas [2] and Wulfsohn and Tsiatis [3]. In this setting, the longitudinal responses are considered realizations of an endogenous timedependent covariate (Kabfleish and Prentice [1]), which is measured with error and for which we do not have the complete history of past values available. To account for these features we estimate in the joint modeling framework the joint distribution of the survival and longitudinal processes. Unlike in the multivariate approach, where we have to interpret the estimates for each longitudinal measurement separately, the joint modeling approach allows to get more insight in the nature of the relation between the two analyzed processes since it assumes some underlying process for the longitudinal measures.
However, in contrast with the standard joint modeling setting, in our case (i.e., transplantation studies) the longitudinal responses do not constitute an endogenous timedependent variable measured at the same period as the time to event. In particular, since the longitudinal measurements are collected prior to transplantation, occurrence of an event (i.e., graft failure after transplantation) does not cause nonrandom dropout in the longitudinal outcome. Nevertheless, the problem of measurement error still remains. Ignoring the measurement error affects not only the standard errors of the estimates of interest, but also it can cause attenuation of the coefficients towards zero [4]. To overcome this, we propose a twostage modeling approach that appropriately summarizes the longitudinal information before the start of followup by means of a mixed effects model and then uses this information to model the time to event by including the Empirical Bayes estimates of the subjectspecific parameters as predictors in the Cox model. To account for the fact that we include the estimates and not the true values of the parameters, we use a Monte Carlo approach and sample from the posterior distribution of the random effects. The proposed method does not require joint maximization neither fitting the random effects model for each event time as in the twostage approach of Tsiatis et al. [5]. We compare this approach with the βnaiveβ one when the uncertainty about the estimates from the first step is not taken into account, as well as with the full Bayesian approach. Our approach shares similarities with the twostage approach of Albert and Shih [6]. They considered a model, in which a discrete event time distribution is modeled as a linear function of the random slope of the longitudinal process estimated from the linearmixed model. The bias from informative dropout was reduced by using the conditional distribution of the longitudinal process given the dropout time to construct the complete data set. To account for the measurement error in the mean of the posterior distribution of the random effects, the variance, that incorporates the error in estimating the fixed effects in the longitudinal model, was used. However, we use sampling not to impute missing values and correct for nonrandom dropout but in order to account for the variability in the predicted longitudinal covariates that are then used in survival model. A method of adjusting for measurement error in covariates, which was used by Albert and Shih, does not apply in our case since it requires the discrete timetoevent and linear model for longitudinal data. The timetoevent could be discretized but still we have a nonlinear model for longitudinal data.
Our research is motivated by data from an international prospective trial on kidneytransplant patients. The study has two arms, where in the first arm donorsβ kidneys were administered to cold storage and in the second arm they were administered to machine perfusion (MP). The advantage of machine perfusion is the possibility of measuring different kidneyβs parameters reflecting the state of the organ. One of the parameters of interest is renal resistance level (RR), which has been measured at 10 minutes, 30 minutes, 1 hour, 2 hours, 4 hours, and just before transplantation. Our aim here is to study the association of the renal resistance evolution profile with the risk of graft failure. The time of last measurement was different for different patients and often unknown exactly. However, based on the medical consult and visual inspection of the individual profiles, the last measurement was chosen to be taken at 6 hours for each patient.
The rest of the paper is organized as follows. Section 2 provides the general modeling framework with the definition of the two submodels for the longitudinal and survival data, respectively. Section 3 describes the estimation methods for the full likelihood and the proposed twostage approach. In Section 4, we apply the twostage approach to the renal data. Section 5 contains the setup and the results for the simulation study. Finally, in Section 6 we discuss the proposed methodology.
2. Joint Modeling Framework
Let denote the longitudinal profiles for individual , . We assume that are collected for the th individual prior to the specified time , . Let denote the time of the first longitudinal measurement and βthe time of the last collected measurement. can be different for different individuals, and we denote by the number of longitudinal measurements for subject collected until time and by the time of th measurement. Denote by the true survival time for individual . Since the survival time is right censored, we observe only , where is the censoring time with the failure indicator , which equals to 1 if the failure is observed and 0 otherwise, that is, with denoting the indicator function. We will assume that censoring is independent of all other survival and covariate information. In addition, we assume that the observed longitudinal responses are measured with error (i.e., biological variation) around the true longitudinal profile , that is, We will consider the longitudinal response that exhibits nonlinear profiles in time. Therefore, we approximate by means of a nonlinear mixed effects model: where is a nonlinear function, parameterized by the vector . The parameters control the shape of the nonlinear function, and subscript denotes that each subject may have its own nonlinear evolution in time in the family . For the subjectspecific parameter , we assume a standard mixed model structure with denoting the fixed effects design matrix with corresponding regression coefficients , the random effects design matrix, and the random effects. The random effects are assumed to be independent and normally distributed with mean zero and variancecovariance matrix .
For the event process, we postulate the standard relative risk model of the form: where is the hazard function and is the baseline hazard, which can be modeled parametrically or left completely unspecified. The subjectspecific parameters summarize the longitudinal evolutions of the response for each subject, and therefore coefficients measure the strength of the association between the different characteristics of the underlying subjectspecific nonlinear evolution of the longitudinal profiles and the risk for an event. Within the formulation of the two submodels (2.2) and (2.3), the same random effects now account for both the association between the longitudinal and event outcomes, and the correlation between the repeated measurements in the longitudinal process.
In the particular transplantation setting that will be analyzed in the following study, are the renal resistance level measurements collected for the th donor prior to the transplantation time and the same index is used to denote the allograft transplanted to the th patient. Time represents the time that the kidney is removed from the donor and put in cold storage or in a perfusion machine.
3. Estimation
3.1. Full Likelihood Framework: Bayesian Approach
In the standard joint modeling framework, the estimation is typically based on maximum likelihood or Bayesian methods (MCMC). This proceeds under the following set of conditional independence assumptions: In particular, we assume that given the random effects the longitudinal process is independent from the event times, and moreover, the longitudinal measurements are independent from each other.
Maximum likelihood methods use the joint likelihood and maximize the loglikelihood function . This requires numerical integration and optimization, which makes the fit difficult, especially in highdimensional random effects settings. Standard options for numerical integration are Gaussian quadrature, Laplace approximation, or Monte Carlo sampling [7, 8]. Maximization of the approximated loglikelihood is based on an EM algorithm [3, 5, 9β11]. Several authors proposed a Bayesian approach (MCMC) [2, 12, 13]. Bayesian estimation, that generalizes a joint model for the case with multivariate longitudinal data, has been discussed by Ibrahim et al. [14]. Brown and Ibrahim [15] considered semiparametric model relaxing the distributional assumption for the random effects. In most papers, the longitudinal submodel is a linear mixedeffects model. Joint models with nonlinear mixedeffects submodels have been less studied in the literature [16]. Nonlinear mixed models are more common in pharmacokinetics and pharmacodynamics, where they are jointly modeled with nonrandom dropout using NONMEM software. Several authors considered a Bayesian approach with a nonlinear mixed model and informative missingness [17, 18].
Here we will proceed under the Bayesian paradigm to estimate the model parameter. Under the conditional independence assumption (3.1), the posterior distribution of the parameters and the latent terms, conditional on the observed data, are derived as where is a vector of parameters from the longitudinal and survival models and the vector of the random effects, respectively, and denotes the appropriate probability density function. The likelihood contribution for the th subject conditionally on the random terms is given by The baseline hazard can be assumed of a specific parametric form, for example, the Weibull hazard. For the priors of the model parameters, we make standard assumptions following Ibrahim et al. [14]. In particular, for the regression coefficients of the longitudinal submodel and for the coefficients of survival submodel, we used multivariate normal priors. For variancecovariance matrices, we assumed an inverse Wishart distribution and for the variancecovariance parameters we took as a prior an inverse gamma. For all parameters, the vague priors have been chosen.
The implementation of the Cox and piecewise constant hazard models is typically based on the counting process notation introduced by Andersen and Gill [19] and formulated by Clayton [20]. In particular, we treat the counting process increments in the time interval as independent Poisson random variables with means : where is an observed process taking the value 1 if subject is observed at time and 0 otherwise and is the increment in the integrated baseline hazard function occurring during the time interval . Since the conjugate prior for the Poisson mean is the gamma distribution, we assume the conjugateindependent increments prior suggested by Kalbfleisch [21], namely, where is a prior mean hazard with being a scaling parameter representing the βstrength" of our prior beliefs. The gamma prior was also chosen for the baseline risk parameter of the Weibull distribution in parametric survival model. Alternatively to implement the Cox model in a fully Bayesian approach, one may use the βmultinomialPoisson trickβ described in the OpenBUGS manual that is equivalent to assuming independent increments in the cumulative hazard function. The increments are treated as failure times, and noninformative priors are given for their logarithms. Analogically to the Cox model, a piecewise constant hazard model was implemented. We have modeled baseline hazard using a step function with 3 quantiles , , and as changing points assuring the same number of events in between. Let denote the start of the followup, the maximum censoring time, and the increment in the integrated baseline hazard function occurring during the time interval . Then for different intervals, we specify a separate prior hazard mean and Similarly as for the Cox model, the results were not sensitive with respect to the choice of the hyperparameters as long as the priors were sufficiently diffuse. The above nonparametric approach can be criticized as having the independent priors for the hazard distribution. However, as suggested by Kalbfleisch [21] a mixture of gamma priors can be considered as an alternative. Moreover, in a piecewise constant model one can also include change points as unknown parameters in the model as proposed in a Bayesian context by Patra and Dey [22] and applied by Casellas [23].
In order to assess convergence for the full Bayesian model, standard MCMC diagnostic plots were used. The burnin size was set to 10000 iterations, which was chosen based on the visual inspection of the trace plots and confirmed by the the Raftery and Lewis diagnostics. The same number of iterations were used for constructing the summary statistics. Based on the autocorrelation plots, we have chosen every 30th iteration. Therefore, in total to obtain 10000 iterations for the final inferenc 300000 iterations were required after the burnin part. Additionally, we run a second parallel chain and used Gelman and Rubin diagnostic plots to assess the convergence.
3.2. TwoStage Approach
As mentioned in Section 1, the longitudinal measurements in our setting do not constitute an internal timedependent covariate, since the events took place after the last longitudinal measurement was collected. In particular, since events do not cause nonrandom dropout, the event process does not carry any information for the longitudinal outcome. Mathematically, this means that information for the random effects is actually only coming from the longitudinal responses, that is, Thus, we can avoid the computational complexity of the full likelihood framework presented in Section 3.1 by employing a twostage approach. More specifically, at Stage I, we obtain by maximizing the loglikelihood: This requires numerical integration, and we use a Gaussian quadrature for that purpose. Then we obtain the corresponding empirical Bayes estimates: and compute the predictions: At Stage II, we fit the relative risk model: However, a potential problem in the above is that is not the true subjectspecific parameters but rather an estimate with a standard error. If we ignore this measurement error, the regression coefficients will be possibly attenuated. To overcome this problem, we propose here a sampling approach to account for the variability in , very close in spirit to the Bayesian approach of Section 3.1. In particular, we use the following sampling scheme.
Step 1. simulate .
Step 2. simulate .
Step 3. calculate and fit the relative risk model from which and are kept.
Steps 1β3 are repeated times.
Step 1 takes into account the variability of the MLEs, and Step 2 the variability of . Moreover, because the distribution in Step 2 is not of a standard form, we use a independent MetropolisHastings algorithm to sample from it with multivariate proposal density centered at an Empirical Bayes estimates , covariance matrix , and . The low number of degrees of freedom was chosen to ensure that the proposal density has heavy tails to provide sufficient coverage of the target density . The variancecovariance matrix estimated from the nonlinear mixed model was additionally scaled by some parameter . The tuning parameter allows to control the acceptance rate through the range of the proposed distribution. If the range is too narrow, the proposed values will be close to the current ones leading to low rejection rate. On the contrary, if the range is too large, the proposed values are far away from the current ones leading to high rejection rate. We chose the acceptance rate to be 0.5 following Carlin and Louis [24] that suggests a desirable acceptance rates of MetropolisHastings algorithms to be around 1/4 for the dependence (random walk) MH version and 1/2 for the independent MH. Roberts et al. [25] recommended to use the acceptance rate close to 1/4 for high dimensions and 1/2 for the models with dimensions 1 or 2. They discussed this issue in the context of the random walk proposal density. The authors showed that if the target and proposal densities are normal, then the scale of the latter should be tuned so that the acceptance rate is approximately 0.45 in onedimensional problems and approximately 0.23 as the number of dimensions approaches infinity, with the optimal acceptance rate being around 0.25 in as low as six dimensions. In our case, an independent version of MetropolisHastings algorithm is applied. The proposal density in the algorithm does not depend on the current point as in the randomwalk Metropolis algorithm. Therefore, for this sampler to work well, we want to have a proposal density that mimics the target distribution and have the acceptance rate be as high as possible. In order to obtain a desirable acceptance rate one needs to run a sampling algorithm for a number of iterations for a given data set and compute an acceptance rate and then repeat the procedure changing the tuning parameter until the chosen acceptance rate, is obtained. Usually a small number of iterations (100β500) is sufficient for the purpose of calibration. More details about the MetropolisHastings acceptancerejection procedure can be found in the supplementary material (available online at http://dx.doi.org/10.1155/2012/194194). A final estimate of is obtained using the mean of the estimates from all iterations: To obtain the SE of , we use the variancecovariance matrix : where is the average withiniteration variance and is the betweeniteration variance, that is, represents a variancecovariance matrix estimated in iteration for .
4. Analysis of the RR Data
4.1. Modelsβ Specification
We apply the proposed twostage procedure and a fully Bayesian approach to the transplantation study introduced in Section 1. The data was taken from an international prospective trial on 337 kidney pairs, which aimed to compare two different types of storage, namely, cold storage and machine perfusion (MP). Here we focus on the second arm. Our main outcome of interest is graft survival time censored after 1 year. At the end of the study, only 26 graft failures were observed. The renal resistance level (RR) was expected to be an important risk factor for graft failure. It was measured using the perfusion machine at the moment of taking the organ out from a donor , and thereafter at 10 minutes, 30 minutes, 1 hour, 2 hours, 4 hours, and just before transplantation. As mentioned in the Section 1, the time of last measurement was different for different patients and sometimes unknown. However, there was a clear asymptote visible from the individual profiles that was reached after about 5 hours by each patient. Based on that behavior and after the medical consult, we chose the last measurement to be taken at 6 hours for each patient. Other variables of interest include the age of the donor, donorβs region (3 countries considered), and donorβs type (heart beating or nonheartbeating).
The individual profiles of 50 randomly selected kidney donors are presented in Figure 1. This plot confirms the biological expectation that allografts exhibit their highest renal resistance levels just after being extracted from the donor. Thereafter, they show a smooth decrease in RR until they reach an asymptote above zero indication that there is no βperfect flow" through the kidney. Furthermore, we observe that the initial RR level, the rate of decrease, and the final RR level differ from donor to donor. Additional descriptive plots for our data are presented in the supplementary material.
In the first step of our analysis, we aim to describe the evolution of the renal resistance level in time. Motivated by both biological expectation and Figure 1, we postulate the following nonlinear function: where is a lower asymptote, is an initial value for , and is the rate of decrease from to (see also Figureββ2 in supplementary material).
To accommodate for the shapes of RR evolutions observed in Figure 1, we need to constraint , , and to be positive. Moreover, in order to allow for individual donor effects, we use the following formulation: where and with and . In the second step, the predicted parameters () summarizing the RR evolution of the nonlinear mixed model are included in the graft survival model. The final model for graft survival was of the form: To investigate the impact of ignoring that the covariate is measured with error, we compared the naive approach in which we ignored this measurement error and our proposal that accounts for the uncertainty in via Monte Carlo sampling. We used MetropolisHastings algorithm with independent proposal and acceptance rate around 50 for the reason given in Section 3.2. We simulated samples with additional initial step of the scaling parameter calibration in order to achieve the desirable acceptance rate. The final estimates (and SE) of the parameters associated with RR covariates were calculated using the formulas described in Section 3.2. We compared the results from the twostage procedure with the estimates obtained from the fully Bayesian joint model fitted for the data using the priors specified in Section 3.1.
The analysis was performed using Statistical Software. Packages survival and nlme were used for the two submodels fit, and a separate code was written by the first author for the sampling part. The fully Bayesian model was fitted using OpenBUGS software with the priors specified in Section 3.1. In particular, for the variancecovariance matrices of multivariate normal priors, we used inverse Wishart distribution with degrees of freedom. For the variancecovariance parameter of the normal longitudinal response, we took as a prior an inverseGamma (). For the baseline risk parameter of the Weibull distribution in survival submodel, a Gamma() prior was used. To analyze the data using the fully Bayesian Cox model described in Section 3.1, we chose the scaling parameter in a gamma prior for the independent increments to be equal 0.001 and a prior mean . We did not observe any substantial difference for the different values of parameter as long as was small enough to keep the prior noninformative. We do not recommend too small values of the scaling parameter as they can lead to the computation problems. Analogically we have chosen gamma priors for the piecewise constant hazard model. The code for the Bayesian full joint model, as well the R codes for the sampling twostage procedure, is available from the authors on request.
4.2. Results
The results for the nonlinearmixed model are presented in Table 1, for the twostage approach and in supplementary material, for the full Bayesian approach with Weibull survival model. The results for the longitudinal part for the full Bayesian approach with Cox and piecewise constant hazard models were similar (not presented). It can be observed, based on twostage model results, that only Donor Age had a significant impact on the RR asymptote. Donor Type and Region had a significant impact on the steepness parameter. However, we keep all the covariates in the model for the purpose of prediction for the second stage. The mean RR profiles for HeartBeating and NonHeartBeating donors from different regions together with fitted values based on the obtained nonlinear mixed model are given in the supplementary material.

In the second step of the analysis, we applied at first the naive approach and used the estimates , and from the nonlinear mixed model as fixed covariates in the final Cox models for graft survival. Table 2 presents the results for the survival submodel for the all approaches, namely, the plugin method, twostage approach, and the fully Bayesian model. For the fully Bayesian approach, the results for the parametric Weibull model together with Cox and piecewise constant hazard models are listed. The results from Table 2 indicate that only the RR asymptote could have a significant impact on graft survival.
(a) Graft survival, plugin  
 
(b) Graft survival, sampling twostage  
 
(c) Graft survival, fully Bayesian, Weibull  
 
(d) Graft survival, fully Bayesian, Cox  
 
(e) Graft survival, fully Bayesian, piecewise constant hazard  

We observe that the estimates are close or almost identical as in plugin model. SE of the Cox regression coefficients for the model with sampling are greater than SE from the plugin model in Table 2(a), especially for the parameter . The increase in SE is somewhat the expected and is caused by the additional variability in covariates captured by the sampling approach. The fully Bayesian model produces similar results to our semiBayesian sampling model with somewhat lower SE. We do not observe substantial difference between fully parametric and nonparametric models in a fully Bayesian approach. Since in the analyzed real data the number of events is small, the fully Bayesian Cox and piecewise constant hazard Bayesian models were expected to produce similar results. We did not observe any substantial difference for the different values of hyperparameters.
Even though it is hard to compare exactly the computational time for the two approaches, the rough estimation of the total computational time needed to estimate and assess the convergence (2 chains) of the full Bayesian model was about 21.6 hours and depended on the implemented survival model. A similar computational time was needed for the full Bayesian model with the Cox survival model and piecewise constant hazard model with a slightly more time needed for the parametric Weibull model. For the twostage approach, the total computational time was about 10 hours using the Intel(R) Core(TM)2 Duo T9300 2.5βGHz and 3.5βGB RAM.
5. Simulations
5.1. Design
We have conducted a number of simulations to investigate the performance of our proposed twostage method. In particular, we compared the plugin method that uses the Empirical Bayes estimates from the nonlinear mixed model and ignores the measurement error, the twostage Monte Carlo sampling approach that accounts for the variability in , and the fully Bayesian approach. In the fully Bayesian approach, only the parametric Weibull model was considered.
For the longitudinal part, the data were simulated for 500 patients from model (5.1) with , and , , . The variancecovariance matrix of the random effects was chosen to be . We kept 7 measurement points as in the original analyzed data set as well as the nonequal distances between them. The variance of the measurement error was chosen to be 0.25, 1, and 25. Survival times were simulated for each patient using the exponential model with the rate parameter equal , where We considered scenarios with fixed coefficients , and . The censoring mechanism was simulated independently using an exponential distribution . Parameter was changed in order to control proportion of censored observations. Different scenarios with 40 and 20 of censoring were examined. For each simulated data set we fitted four survival models, namely, the gold standard model that uses the true simulated values , the plugin model, the sampling model, and fully Bayesian joint model. Neither nonparametric Cox nor piecewise constant hazard model were considered in a fully Bayesian approach since we have simulated the data from the parametric exponential model and wanted to compare the proposed twostage approach with the βbest" fully Bayesian model. All the prior settings, size of burnin, number of iterations, and so forth, for the fully Bayesian model were the same as for the real data analysis.
5.2. Results
In Table 3, we present the average results for 200 simulations of different scenarios are presented. The results from our sampling model were very close to the results obtained for the fully Bayesian model with slightly smaller bias and RMSE for the fully Bayesian approach. That was due to the better estimation of random effects variability in fully Bayesian approach. The plugin method produced the biggest bias that sometimes with combination with the small variability of the estimates around the biased mean resulted in larger RMSE than in sampling approach. As the measurement error or the percentage of censored observations increased, the estimates of survival submodel were more biased with larger RMSE for all approaches. The increase in bias was more severe when the measurement error variance was increased rather than when the percentage of to censoring was higher. This bias was, however, decreased when the number of repeated measures per individual was increased. This has to do with the amount of information that is available in the data for the estimation of . As it is known from the standard mixed models literature [26], the degree of shrinkage in the subjectspecific predicted values is proportional to and inversely proportional to and . To compare the relation between variance of the random effects and variance of the measurement error, one can use intraclass correlation (ICC) defined as the proportion of the total variability that is explained by the clustering with a given random effect. ICC was similar for the simulated and the real data for the biggest and increased in a simulation data as decreased.

Since the calculations for the simulation study were highly computationally intensive, we have used the cluster with about 20 nodes with AMD QuadCore Opteron 835X, 4 Γ 2βGHz, and 16βGB RAM per node. The analysis for the the 200 simulated data sets for a single scenario took about 65.5 hours using the Bayesian approach and 31.2 hours using the twostage approach.
6. Discussion
We have proposed a twostage method that can be used in a joint analysis of longitudinal and time to event data when the longitudinal data are collected before the start of followup for survival, and the interest is in estimation of the impact of longitudinal profiles on survival. The modeling strategy is based on specification of two separate submodels for the longitudinal and time to event data. First, the longitudinal outcome is modeled using a random effects model. Then the survival outcome is modeled using the Empirical Bayes estimates of the subjectspecific effects from the first stage. The variability of the estimates from the first stage is properly taken into account using a Monte Carlo approach by sampling from the posterior distribution of the random effects given the data.
As it was demonstrated, ignoring the additional variability of the subjectspecific estimates when modeling survival leads to some bias, and in particular, attenuates the regression coefficients towards zero [4]. That was also confirmed by our simulation study. In comparison with the fully Bayesian approach, the proposed partially Bayesian method produced similar results with substantially less number of iterations required. This is due to the fact that sampling was conducted already around the EB estimates, and there is no needed for a burnin part as in a standard fully Bayesian approach. We used 10000 iterations per subject, which was about the size of burnin needed in the fully Bayesian models. No thinning was used in our approach, based on the visual inspection of the trace plots. Though it is hard to compare the fully Bayesian approach and the twostage approach with respect to the computational time precisely, the rough approximation of the total computational time required for the twostage approach was about half in comparison with the fully Bayesian approach. The fully Bayesian approach provided similar results with the twostage approach for the special setting we have considered here. However, fitting a fully Bayesian model was a bit of βoverdone" in the sense that by design the longitudinal data could not be affected by the survival. Since in many transplantation studies, the longitudinal data are collected before the start of followup for survival; therefore, using our method in that cases seems to be more appropriate than using a fully Bayesian approach. We recommend the proposed approach not only for the particular transplantation studies but in any setting that shares the similarity of the separated followup periods for the two analyzed endpoints. That is, for example, when the event process does not carry any information for the longitudinal outcome and the condition (3.7) from Section 3.2 holds. The simulation results indicate that even if the data come from the real joint setting in which (3.7) may not hold, the proposed twostage procedure can be comparable to the fully Bayesian approach.
Since the sampling in the proposed method relies strongly on the results of the first part, the accurate estimation of all parameters of nonlinear mixed model is a key feature and should be performed carefully. This can be problematic when the deviation from normality of the random effects, is suspected. However, it was shown that even for the nonnormal random effects one can still use a standard software such as nlmixed in SAS with just a small change in a standard program code. In such cases, the probability integral transformation (PIT) proposed by Nelson et al. [27] can be used or the reformulation of the likelihood proposed by Liu and Yu [28]. An alternative is fitting a Bayesian model only to estimate the longitudinal submodel in the first stage, instead of the likelihood methods. Fitting nonlinear mixed models using Bayesian standard software can be, however, problematic due to the high nonlinearity in random effects that is caused both by the nonlinear function of the longitudinal profiles and by the possible restrictions on parameters [29].
In comparison with the twostage approach proposed by Tsiatis et al. [5], our method is less computationally intensive since it does not require fitting as many mixed models as there are event times in the data. An alternative, that is somewhat simpler to implement and does not require any assumption about the distribution on the underlying random effects, is the conditional score approach proposed by Tsiatis and Davidian [11]. However, this method is less efficient than the methods based on likelihood. The focus in the discussed approaches is on the association between the longitudinal and event time processes. However, in transplantation studies when the two followup periods for longitudinal and survival outcomes are often separated, the interest is rather in making an inference on the marginal eventtime distribution. This is similar to the Bayesian approach proposed by Xu and Zeger [12], that uses the longitudinal data as auxiliary information or surrogate for timetoevent data. Our approach is particularly useful in this setting. Since each of the two submodels is fitted separately, a standard software can be used to implement our method with just a small part of additional programming for Monte Carlo sampling.
Another advantage of the proposed twostage method is that it can be easily generalized from survival to other types of models as it was applied for the binary Delayed Graft Failure (DGF) indicator (results not shown) in the analysis of the renal data. For that purpose in the second step of the twostage procedure, the survival model was replaced by the logistic regression model for the DGF indicator. The first stage of the proposed approach could be also modified allowing for other types of longitudinal response and other types of mixed models. Therefore, instead of using a nonlinear mixed model a linear mixed model or generalized linear mixed model (GLMMs) can be considered depending on the type and the shape of the longitudinal response. In the presented real data example, we have chosen the three parameters that described the evolution of the longitudinal response. However, for the particular question of interest, one can easily choose the most convenient parametrization for the longitudinal model and use the selected parameters to analyze the nonlongitudinal response in the second stage.
Acknowledgment
The authors thank J. M. Smits from Eurotransplant International Foundation in Leiden for providing the data set analyzed in the paper and for the medical consult regarding the application of the proposed method.
Supplementary Materials
The following file consists a supplementary material for the article βA TwoStage Joint Model for Nonlinear Longitudinal Response and a TimetoEvent with Application in Transplantation Studiesβ published in Journal of Probability and Statistics. Joint Models and Their Applications. It provides technical details of the implemented MetropolisHastings algorithm. More descriptive statistics for the analyzed real data set are presented as well as the additional tables and figures regarding the final results.
References
 J. D. Kalbfleisch and R. L. Prentice, The Statistical Analysis of Failure Time Data, Wiley Series in Probability and Statistics, John Wiley & Sons, Hoboken, NJ, USA, Second edition, 2002. View at: Zentralblatt MATH
 C. L. Faucett and D. C. Thomas, βSimultaneously modelling censored survival data and repeatedly measured covariates: a Gibbs sampling approach,β Statistics in Medicine, vol. 15, no. 15, pp. 1663β1685, 1996. View at: Publisher Site  Google Scholar
 M. S. Wulfsohn and A. A. Tsiatis, βA joint model for survival and longitudinal data measured with error,β Biometrics. Journal of the International Biometric Society, vol. 53, no. 1, pp. 330β339, 1997. View at: Publisher Site  Google Scholar  Zentralblatt MATH  MathSciNet
 R. L. Prentice, βCovariate measurement errors and parameter estimation in a failure time regression model,β Biometrika, vol. 69, no. 2, pp. 331β342, 1982. View at: Publisher Site  Google Scholar  Zentralblatt MATH
 A. Tsiatis, V. DeGruttola, and M. Wulfsohn, βModeling the relationship of survival to longitudinal data measured with error: applications to survival and CD4 counts in patients with AIDS,β Journal of the American Statistical Association, vol. 90, pp. 27β37, 1995. View at: Google Scholar
 P. S. Albert and J. H. Shih, βOn estimating the relationship between longitudinal measurments and timetoevent data using a simple twostage procedure,β Biometrics, vol. 66, pp. 983β991, 2009. View at: Google Scholar
 D. Rizopoulos, G. Verbeke, and E. Lesaffre, βFully exponential Laplace approximations for the joint modelling of survival and longitudinal data,β Journal of the Royal Statistical Society B, vol. 71, no. 3, pp. 637β654, 2009. View at: Publisher Site  Google Scholar
 R. Henderson, P. Diggle, and A. Dobson, βJoint modelling of longitudinal measurements and event time data,β Biostatistics, vol. 4, pp. 465β480, 2000. View at: Google Scholar
 V. DeGruttola and X. Tu, βModeling progression of CD4 lymphocyte count and its relationship to survival time,β Biometrics, vol. 50, pp. 1003β1014, 1994. View at: Google Scholar
 S. Self and Y. Pawitan, βModeling a marker of disease progression and onset of disease,β in AIDS Epidemiology: Methodological Issues, N. P. Jewell, K. Dietz, and V.T. Farewell, Eds., Birkhauser, Boston, Mass, USA, 1992. View at: Google Scholar
 A. A. Tsiatis and M. Davidian, βA semiparametric estimator for the proportional hazards model with longitudinal covariates measured with error,β Biometrika, vol. 88, no. 2, pp. 447β458, 2001. View at: Publisher Site  Google Scholar  Zentralblatt MATH  MathSciNet
 J. Xu and S. L. Zeger, βJoint analysis of longitudinal data comprising repeated measures and times to events,β Journal of the Royal Statistical Society C, vol. 50, no. 3, pp. 375β387, 2001. View at: Publisher Site  Google Scholar  Zentralblatt MATH  MathSciNet
 Y. Wang and J. M. G. Taylor, βJointly modeling longitudinal and event time data with application to acquired immunodeficiency syndrome,β Journal of the American Statistical Association, vol. 96, no. 455, pp. 895β905, 2001. View at: Publisher Site  Google Scholar  Zentralblatt MATH  MathSciNet
 J. G. Ibrahim, M.H. Chen, and D. Sinha, βBayesian methods for joint modeling of longitudinal and survival data with applications to cancer vaccine trials,β Statistica Sinica, vol. 14, no. 3, pp. 863β883, 2004. View at: Google Scholar  Zentralblatt MATH
 E. R. Brown and J. G. Ibrahim, βA Bayesian semiparametric joint hierarchical model for longitudinal and survival data,β Biometrics, vol. 59, no. 2, pp. 221β228, 2003. View at: Publisher Site  Google Scholar  Zentralblatt MATH  MathSciNet
 L. Wu, X. J. Hu, and H. Wu, βJoint inference for nonlinear mixedeffects models and time to event at the presence of missing data,β Biostatistics, vol. 9, pp. 308β320, 2008. View at: Google Scholar
 C. Hu and M. E. Sale, βA joint model for nonlinear longitudinal data with informative dropout,β Journal of Pharmacokinetics and Pharmacodynamics, vol. 30, no. 1, pp. 83β103, 2003. View at: Publisher Site  Google Scholar
 N. A. Kaciroti, T. E. Raghunathan, M. A. Schork, and N. M. Clark, βA Bayesian model for longitudinal count data with nonignorable dropout,β Journal of the Royal Statistical Society C, vol. 57, no. 5, pp. 521β534, 2008. View at: Publisher Site  Google Scholar
 P. K. Andersen and R. D. Gill, βCox's regression model for counting processes: a large sample study,β The Annals of Statistics, vol. 10, no. 4, pp. 1100β1120, 1982. View at: Publisher Site  Google Scholar  Zentralblatt MATH
 D. Clayton, βBayesian analysis of frailty models,β Tech. Rep., Medical Research Council Biostatistics Unit., Cambridge, UK, 1994. View at: Google Scholar
 J. D. Kalbfleisch, βNonparametric Bayesian analysis of survival time data,β Journal of the Royal Statistical Society B, vol. 40, no. 2, pp. 214β221, 1978. View at: Google Scholar  Zentralblatt MATH
 K. Patra and D. K. Dey, βA general class of change point and change curve modeling for life time data,β Annals of the Institute of Statistical Mathematics, vol. 54, no. 3, pp. 517β530, 2002. View at: Publisher Site  Google Scholar  MathSciNet
 J. Casellas, βBayesian inference in a piecewise Weibull proportional hazards model with unknown change points,β Journal of Animal Breeding and Genetics, vol. 124, no. 4, pp. 176β184, 2007. View at: Publisher Site  Google Scholar
 B. P. Carlin and T. A. Louis, Bayes and Empirical Bayes Methods for Data Analysis, Chapman & Hall, New York, NY, USA, 2000.
 G. O. Roberts, A. Gelman, and W. R. Gilks, βWeak convergence and optimal scaling of random walk Metropolis algorithms,β The Annals of Applied Probability, vol. 7, no. 1, pp. 110β120, 1997. View at: Publisher Site  Google Scholar  Zentralblatt MATH  MathSciNet
 G. Verbeke and G. Molenberghs, Linear mixed models for longitudinal data, Springer Series in Statistics, Springer, New York, NY, USA, 2000.
 K. P. Nelson, S. R. Lipsitz, G. M. Fitzmaurice, J. Ibrahim, M. Parzen, and R. Strawderman, βUsing the probability integral transformation for nonnormal random effects in nonlinear mixed models,β Journal of Computational and Graphical Statistics, vol. 15: 3957, 2004. View at: Google Scholar
 L. Liu and Z. Yu, βA likelihood reformulation method in nonnormal random effects models,β Statistics in Medicine, vol. 27, no. 16, pp. 3105β3124, 2008. View at: Publisher Site  Google Scholar
 M. Davidian and D. M. Giltinan, Nonlinear Models for Repeated Measurment Data, Chapman and Hall, New York, NY, USA, 1998.
Copyright
Copyright © 2012 Magdalena Murawska et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.