2016-Advances in Longitudinal Data Analysis

Advances in Longitudinal Data Analysis 
Chair: Yogendra Chaubey (Concordia University)
Organizer: Sanjoy Sinha (Carleton University) 

KARELYN DAVIS, Health Canada
Longitudinal Analysis in Health Studies with Censored Observations  [PDF]
In performing statistical analysis of national health studies, statisticians often encounter data that is correlated either due to a natural clustering or is collected longitudinally. In particular, when concerned with contaminant levels in humans (e.g. mercury, lead), laboratory data may be collected in such a fashion so as to determine significant patterns in vulnerable populations such as pregnant women and children. However, such contaminant levels may be so low as to be below the laboratory limit of detection and are thus classified as non-detects. In this presentation, analyses encountered from a Health Canada study are discussed which extend linear mixed models to incorporate parameter constraints and analyses of left-censored observations. 
ABDUS SATTAR, Case Western Reserve University
Joint Modeling of Longitudinal and Survival Data With a Covariate Subject to Limit of Detection [PDF]
We develop and study innovative methods for jointly modeling longitudinal and time-to-event data with a covariate or biomarker subject to the limit of detection (LOD). We investigate the gain in efficiency from joint models over separate models when we attempt to estimate the effect of an endogenous time-dependent covariate on survival times. We also investigate the effects of misspecified random effects distributions on the likelihood inference. Our extensive simulation study indicates that if the assumed random effects distributions deviate from the true distributions to a large extent, the maximum likelihood method can produce systematically biased estimators. We present an application of the proposed method using a large clinical trial data from the Genetic and Inflammatory Markers of Sepsis (GenIMS) study. 
SANJOY SINHA, Carleton University
Methods for Longitudinal Data With Nonignorable Missing Responses and Outliers  [PDF]
We encounter missing data in many longitudinal studies. When data are nonignorably missing, it is important to analyze the data by incorporating a missing data model into the observed data likelihood function. It is well-known that the maximum likelihood (ML) estimators are generally sensitive to potential outliers in the data. I propose and explore a robust method, which is developed in the framework of the ML estimation and is useful for downweighting any influential observations when estimating the model parameters. The empirical properties of the robust estimators are studied in simulations. The proposed method is also illustrated in an example using actual longitudinal data on CD4 counts obtained from clinical trials of HIV-infected patients.