A SAS macro for the joint modeling of longitudinal outcomes and multiple competing risk dropouts
Introduction
Long-term longitudinal studies often encounter data attrition because subjects may drop out or die before the study ends. Informative dropout due to competing risks is a critical aspect that should be taken into account when analyzing these data. For example, in the Atherosclerosis Risk in Communities Neurocognitive Study (ARIC-NCS), studying risk factors for cognitive decline is a primary interest [1], [2]. Cognitive function assessments were repeatedly measured on available participants over a 20 year study period, but many participants' cognitive function measurements were unavailable due to preceding dementia and/or death. This resulted in dropout mechanisms potentially highly correlated with participants' cognitive declines. Longitudinal analyses of cognitive function that ignore potential informative dropouts may lead to biased association estimates; determining the degree of this potential bias is of substantive import.
In the past two decades many statistical models and inference approaches have been proposed to accommodate the close relationship of repeated measurements coupled with multiple survival outcomes. The purpose of such analyses is to assess the impact on both repeated measurements and survival outcomes jointly, to improve inference for the longitudinal outcome taking into consideration the survival outcome (e.g. trajectory change assessment of longitudinal outcome in the presence of informative dropout) [3], [4], or to improve inference for the survival outcome in the presence of observed time varying covariates (e.g. effects of longitudinal biomarker measurements on the event times of interest) [5], [6]. Tsiatis and Davidian provided an excellent overview of early work on joint models [7], and Wu et al. reviewed recent progress and summarized the commonly used methods [8]. In this paper, we focus on sensitivity of the longitudinal submodel (repeated measurements) in conjunction with other researchers' recent work [9], [10], [11], [12]. Elashoff et al. proposed using a linear mixed model for the longitudinal outcome and semiparametric competing risk submodels, linked by common latent variables. The EM-based algorithm is derived to obtain the parameter estimates, and a profile likelihood method is used to estimate the standard errors [9]. Li et al. developed a joint model adopting a t- instead of normal distribution for measurement errors in the linear mixed effects submodel which is robust in the presence of outlying longitudinal observations during follow-up [10]. Ghosh et al. proposed a fully Bayesian approach to jointly model the viral RNA response and informative dropout in the presence of multiple changepoints [11], and Gueorguieva et al. constructed a joint model for the longitudinal outcome and cause-specific dropouts that allows for interval-censored dropout times [12]. Simulation studies in these references all showed that the joint analysis of longitudinal measurements and survival events gave nearly unbiased results in the presence of informative dropout, whereas standard linear mixed effects models alone did not.
Some statistical software packages have begun to appear recently to implement the aforementioned joint modeling methods (see the R package JM [13], JMbayes [14], JoineR [15] as well as Stata module STJM [16]), however, the majority of these available software packages allow only a single time-to-event outcome. We adopted the joint model proposed by Gueorguieva et al. [12], which assumes conditional independence between longitudinal and cause-specific survival outcomes given a set of underlying random effects, and we developed a SAS macro to accommodate the analysis of longitudinal outcomes in the presence of multiple competing survival/dropout events. Our proposed SAS macro allows up to three different competing events and we believe it is the first SAS program to facilitate this kind of problem.
The paper is organized as follows. Section 2 presents the shared parameter model specification. In Section 3, we introduce a SAS macro to analyze a longitudinal outcome in the presence of multiple informative competing risk dropouts. We demonstrate how to use this macro by analyzing one real ARIC-NCS data set in Section 4 and a simulation study is shown in Section 5. Section 6 provides discussion and concluding remarks.
Section snippets
The shared parameter model
Suppose there are n subjects in the study. Let Yit be the measurement of a response variable for subject i at time t, where i = 1, 2, …, n, and t = 1, 2, …, ni, with ni denoting the number of repeated observations on the ith subject. During follow-up, each subject may drop out from one of K different causes or could be right censored. Let Ci = (Ti, Ki) be the dropout data on subject i, where Ti is the time to dropout or censoring, and Ki takes value from {0, 1, 2, …, K}, with 0 indicating a
The SAS macro
We have written a SAS macro, called SPM, to apply the shared parameter model given in Section 2 to provide the inference on longitudinal as well as survival outcomes in the presence of potential informative dropouts. In addition, analysis results from separate models (linear mixed models for the longitudinal outcome and parametric survival models for survival outcomes) are also provided for comparison and sensitivity analysis purposes. In this section, we describe the syntax to prepare the
The data example
The ARIC-NCS study was designed to investigate the role of midlife vascular risks in predicting dementia and cognitive decline [1], [2]; here, we randomly chose subset of 1000 non-Hispanic white female participants to assess potential informative dropout effects on associations between midlife hypertension and cognitive decline using the SPM macro for illustration purposes. Three cognitive function tests, the Delayed Word Recall Test (DWRT), Digit Symbol Substitution Test (DSST) and Word
Monte-Carlo simulation
We performed a Monte-Carlo simulation study to further confirm our joint modeling approach. In this study, data are generated from a joint model using linear mixed and Weibull survival specifications,where k = 1, 2, indicating 1 or 2 different dropout reasons. Both longitudinal and survival submodels include a binary exposure indicator di (1 if exposed, 0, otherwise) and a normally distributed covariate w
Concluding remarks
We have created a SAS macro to jointly model a longitudinal outcome and multiple competing risk dropouts where the longitudinal and survival submodels are linked via shared random effects. The macro is highly useful for examining the sensitivity of standard mixed model estimates to potential informative missingness effects. The program allows up to three different dropouts, random intercept/slope specifications in the longitudinal outcome model, and provides fit criteria AIC and BIC to assist
Conflict of interest
None declared.
Acknowledgments
The Atherosclerosis Risk in Communities Study is carried out as a collaborative study supported by the National Heart, Lung, and Blood Institute contracts (HHSN268201100005C, HHSN268201100006C, HHSN268201100007C, HHSN268201100008C, HHSN268201100009C, HHSN268201100010C, HHSN268201100011C, and HHSN268201100012C). Neurocognitive data are collected by U01 HL096812, HL096814, HL096899, HL096902, HL096917 with previous brain MRI examinations funded by R01-HL70825 (NHLBI NIH). The authors thank the
References (21)
- et al.
GFREG: a computer program for maximum likelihood regression using the generalized F distribution
Comput. Methods Programs Biomed
(1985) - et al.
Impact of differential attrition on the association of education with cognitive change over 20 years of follow-up: the ARIC neurocognitive study
Am. J. Epidemiol
(2014) - et al.
Midlife hypertension and 20-year cognitive change: the atherosclerosis risk in communities neurocognitive study
JAMA Neurol
(2014) Methods for the analysis of informatively censored longitudinal data
Stat. Med
(1992)Modeling the drop-out mechanism in repeated-measures studies
J. Am. Statist. Ass
(1995)- et al.
Jointly modeling longitudinal and event time data with application to acquired immunodeficiency syndrome
J. Am. Statist. Ass
(2001) - et al.
A Bayesian semiparametric joint hierarchical model for longitudinal and survival data
Biometrics
(2003) - et al.
Joint modeling of longitudinal and time-to-event data: an overview
Stat. Sin
(2004) - et al.
Analysis of longitudinal and survival data: joint modeling, inference methods, and issues
J. Prob. Stat
(2012) - et al.
An approach to joint analysis of longitudinal measurements and competing risks failure time data
Stat. Med
(2007)
Cited by (11)
Concurrent hearing and vision impairment and 8-year memory decline in community-dwelling older adults
2023, Alzheimer's and DementiaJoint modeling of the longitudinal student mark and the competing events of degree completion and academic dropout
2023, Communications in Statistics - Theory and MethodsA MILP facility location model with distance value adjustments for demand fulfillment using Google Maps
2022, Journal of Engineering Research (Kuwait)Hip fracture in patients with non-dialysis chronic kidney disease stage 5
2021, Scientific Reports