Browsing by Author "de Leon, Alexander"

Now showing 1 - 8 of 8
  • Thumbnail Image
    ItemOpen Access
    Analysis of Serially Dependent Multivariate Longitudinal Non-Gaussian Continuous Data
    (2015-09-29) Yeasmin, Fahmida; de Leon, Alexander
    Serially dependent multivariate longitudinal non-Gaussian outcome data are commonly encountered in many fields of study, especially in biomedical sciences, finance, and so on. However, flexible methodologies for joint analysis of these outcomes are not well developed. Recently, Wu and de Leon (2014) and Withanage and de Leon (2015) introduced the class of Gaussian copula mixed models (GCMMs) for joint analysis of non-Gaussian outcomes. We adapt and extend the GCMM to settings that involve conditional as well as serial dependencies among longitudinal observations on the same or on different outcomes. We investigate the impact of failing to account for these dependencies via simulations. We illustrate our methodology on two datasets: one on data obtained from primary biliary cirrhosis patients, and the other on data from the Iowa Youth and Families Project.
  • Thumbnail Image
    ItemOpen Access
    Binocular Sensitivity and Specificity of Screening Tests in Prospective Studies of Paired Organs
    (2013-01-28) Perera, Singappuli; de Leon, Alexander
    Diagnostic and screening studies in ophthalmology frequently involve binocular data where pairs of eyes are evaluated, through some diagnostic procedure, for the presence of certain diseases or pathologies. It is usually sufficient in practice that at least one eye is positively diagnosed for the patient to be sent for further and more extensive eye examination. More relevant diagnostic accuracy measures in these cases are therefore the probability of at least one correct positive diagnosis in patients with one or both eyes truly diseased and the probability of two correct negative diagnoses for patients with both eyes truly un-diseased. The former is analogous to sensitivity and the latter to specificity. Predictive values may be similarly re-defined. The thesis proposes these new sensitivity and specificity measures as alternatives to conventional ones for paired binocular binary diagnostic data arising from screening studies with cross-sectional sampling. The measures are defined for flexible models based on copulas and extensions of existing models for correlated binary data. The proposed methodology is illustrated with data from a study on diabetic retinopathy.
  • Thumbnail Image
    ItemOpen Access
    Causal Inference: Additive Hazard Model for Mediation Analysis with Measurement Error and Marginal Structural Models
    (2017) Shen, Lingzhu; Chen, Gemai; Yan, Ying; Lu, Xuewen; de Leon, Alexander
    In epidemiologic and social science studies, researchers are often interested in understanding the causal effect from an exposure variable to an outcome variable. In this thesis, we develop two different models and methods to conduct causal inference: (1). causal mediation analysis under the additive hazards model with exposure-mediator interaction; (2). marginal structural additive hazards model. The existing literature requires accurate measurements of the mediator and the confounders, which could be infeasible in biomedical studies. Furthermore, the current identification results of causal effects under the additive hazards model do not allow for exposure-mediator interaction. In this thesis, we derive identification results of causal effects under the additive hazards model with exposure-mediator interaction. Furthermore, we propose consistent measurement error correction methods in the absence/presence of exposure-mediator interaction. In the second part of the thesis, we propose a marginal structural additive hazards model. We develop an estimation method for the marginal structural additive hazards model and apply the simulation-extrapolation (SIMEX) method to correct for the bias resulting from measurement error.
  • Thumbnail Image
    ItemOpen Access
    Cluster Aanlysis of Gene Expression Profiles via Flexible Count Models for RNA-seq Data
    (2015-06-10) Ruan, Ji; de Leon, Alexander
    Clustering RNA-seq data is used to characterize environment-induced (e.g., treatment) differences in gene expression profiles by separating genes into clusters based on their expression patterns. Wang et al. [2013] recently adopted the bi-Poisson distribution, obtained via the trivariate reduction method, as a model for clustering bivariate RNA-seq data. We discuss the inadequacy of the bi-Poisson distribution in modelling the correlation between dependent Poisson counts, and its impact on clustering such data. We introduce an alternative Gaussian copula model that incorporates a flexible dependence structure for the counts, report simulation results to compare the performance of the Gaussian copula and bi-Poisson models, and investigate the impact on clustering of Poisson counts of misspecified dependence structures. We illustrate our methodology on a lung cancer RNA-seq data.
  • Thumbnail Image
    ItemOpen Access
    Exploring the impact of polygenes on genetic inheritance model identification, with application to Familial Colorectal Cancer Type X (FCCTX)
    (2017) Scory, Tayler; Kopciuk, Karen; Lu, Xuewen; de Leon, Alexander; Long, Quan
    Although a genetic inheritance pattern has not yet been identified, there seems to be a hereditary component for some types of cancer. The focus of this thesis is on identifying the factors that enable correct identification of genetic inheritance models. Exploring this topic involved complex segregation analysis on real FCCTX cancer registry data, then on simulated data (based on the real data characteristics) to determine what caused the model to be identified. If a strong polygenic effect is present, finding evidence for the correct genetic model is more likely. However, the correct model was identified roughly 50% of the time, so more factors should be explored. If the genetic inheritance pattern of a disease is identified, this would facilitate identifying the gene mutation in question, especially with rapidly advancing genomic technology. This work can be applied to other cancers, and can encourage exploration of non-Mendelian genetic inheritance.
  • Thumbnail Image
    ItemOpen Access
    Likelihood Analysis of Gaussian Copula Distributions for Mixed Data via a Parameter-Expanded Monte Carlo EM (PX-MCEM) Algorithm
    (2016) Ren, Mingchen; de Leon, Alexander; Yan, Ying; Lu, Xuewen; Ambagaspitiya, Rohana
    Mixed discrete and continuous data arise in a variety of settings. In this thesis, we adopt so-called Gaussian copula distributions (GCDs) as a general model for binary and continuous variables. The attractive feature of GCDs is their use of Gaussian copulas to separately model dependencies between variables, thereby preserving the variables' distinct marginal properties. We employ an efficient approach to maximum likelihood estimation for the model via a parameter-expanded Monte Carlo EM (MCEM) algorithm. By doing so, we not only avoid the direct evaluation of the likelihood function, which involves computing multivariate normal probabilities, but also improve the computational efficiency of the algorithm. Another advantage of the PX-MCEM algorithm is that it has an analytically tractable M-step, and hence does not require numerical optimization techniques. Based on simulations and an application to a breast cancer dataset, we show that the estimates are reasonably unbiased and their sampling variabilities can be accurately estimated by their bootstrapped standard errors.
  • Thumbnail Image
    ItemOpen Access
    Mean-Variance mixture model for calibrating loss given default (LGD) of a credit portfolio
    (2021-11) Sam, Charles; Ambagaspitiya, Rohana; Ambagaspitiya, Rohana; Scollnik, David; de Leon, Alexander; Fapojuwo, Abraham; Bégin, Jean-François
    We study the sensitivity of Value-at-Risk (VaR) and Tail-Value-at-Risk (TVaR) of credit portfolio of defaultable obligors to the tail fatness of the loss given default latent variable distribution. We consider a static structural model where obligors default and loss given default (LGD) latent variables have a common systematic risk factor. We propose the use of the Normal-Variance mixture model to model the LGD latent variable to account for certain random external risks, such as the collapse of Lehman Brothers Holdings Inc in 2008 which resulted in instability in the financial sector. We derive an analytical expression for finding the asymptotic portfolio loss rate. We also propose two importance sampling algorithms for finding conditional tail probabilities for the portfolio loss. Our approach is unique in two aspects. First, we capture the dependence between default and LGD. Second, we make LGD values to be random and between zero and one. We also show that our importance sampling algorithms are asymptotically optimal.
  • No Thumbnail Available
    ItemOpen Access
    Spiked-in Data Set for BMC Notes paper
    (2016) Kopciuk, Karen; Bathe, Oliver; McConnell, Yarrow; Welje, Aalim; de Leon, Alexander