Profiling medical sites is an important activity in both clinical research and practice. Many organizations provide public report cards comparing outcomes across hospitals. An analogous concept applied in multicenter clinical trials, such “report cards” guide sponsors to choose sites while designing a study, help identify areas of improvement for sites, and motivate sites to perform better. Sponsors include comparative performance of sites, a concept to perform risk-based monitoring and central statistical monitoring. In clinical research, report cards are powerful tools for relating site performance to treatment benefits. This study evaluates approaches to estimating the proportion of adverse events at the site-level in a multicenter clinical trial setting and also methods in detecting outlying sites. We address three topics. First we assess the performance of different models for obtaining estimates of adverse events rates utilizing Bayesian beta-binomial and binomial logit-normal models with MCMC estimation and fixed effects maximum likelihood estimation (MLE) methods. We - vary sample sizes, number of medical sites, overall adverse event rates, and intraclass correlation within sites. Second, we compare the performance of these methods in identifying outlier sites, contrasting MLE and Bayesian approaches. A fixed threshold method detects sites as outliers under a Bayesian approach, while in the fixed effects assumption, a 95% interval-based approach is applied. Third, we extend this approach in estimating multiple outcomes at the site level and detecting outlier sites. A standard bivariate normal MLE method is compared to a Bayesian bivariate binomial logit-normal MCMC. These are examined using simulation studies. Results show for single outcomes, Bayesian beta-binomial MCMC method perform well under certain parametric conditions for estimation and detecting outlier sites. For multiple outcomes with higher adverse event rate and larger difference between outliers and non-outliers, for detecting outlier sites, both methods – Bayesian MCMC and MLE work well, irrespective of the correlation between outcomes. / 2020-02-14T00:00:00Z

The CDC Wide-ranging ONline Data for Epidemiologic Research (CDC WONDER) makes many health-related datasets available to the public health community through web applications. One such available dataset is The Multiple Cause of Death data which displays county-level national mortality and population data. One of the main issues with this particular dataset is that the death counts within the age groups can be very small or equal to zero for various counties which can cause the conditional probability of death to be small or even zero. This issue causes the estimates for life expectancy within the abridged life table to be unreliable. This research utilizes the data provided by CDC WONDER, distance measures (Euclidean and discrete Hellinger distances), Metric Multidimensional Scaling, and Partitioning Around Medoids to identify patterns of life table elements among the "stable" counties within the dataset. The identification of these patterns is then used to classify the patterns which the "unstable" counties fall into. Future work will aim at borrowing from the "stable" counties, geographic and demographic information, which the "unstable" counties most closely resemble in order to better predict their life table elements, particularly life expectancies.

Due to the rapid development and growing need for information technologies, more and more researchers start to focus on high-dimensional data. Much work has been done on problems like point estimation possessing oracle inequalities, coefficient estimation, variable selection in high-dimensional regression models. However, with respect to the statistical inference for the regression coefficients, there have been few studies. Therefore, we propose a regularized efficient score estimation and testing (RESET) approach for treatment effects in the presence of nuisance parameters, either low-dimensional or high-dimensional, in generalized linear models (GLMs). Based on the RESET method, we are also able to develop another two-step approach related to the same problem. The RESET approach is based on estimating the efficient score function of the treatment parameters. This means we are trying to remove the influence of nuisance parameters on the treatment parameters and construct an efficient score function which could be used for estimating and testing for the treatment effect. The RESET approach can be used in both low-dimensional and high-dimensional settings. As the simulation results show, it is comparable with the commonly used maximum likelihood estimators in most low-dimensional cases. We will prove that the RESET estimator is consistent under some regularity conditions, either in the low-dimensional or the high-dimensional linear models. Also, it is shown that the efficient score function of the treatment parameters follows a chi-square distribution, based on which the regularized efficient score tests are constructed to test for the treatment effect, in both low-dimensional and high-dimensional GLMs. The two-step approach is mainly used for high-dimensional inference. It combines the RESET approach with a first step of selecting "promising" variables for the purpose of reducing the dimension of the regression model. The minimax concave penalty is adopted for its oracle property, which means it tends to choose "correct" variables asymptotically. The simulation results show that some improvement is still required for this approach, which will be part of our future research direction. Finally, both the RESET and the two-step approaches are implemented with a real data example to demonstrate their application, followed by a conclusion for all the problems investigated here and a discussion for the directions of future research.

Measurement error is a type of non-sampling error that could attenuate the effect of a risk factor on an outcome variable if no correction is made. Therefore, an effect might not be detectable, even if there is one. If a classical error type is present, then the power of the analysis will be lowered or a bigger sample size will be needed in order to maintain the desirable power. Thus, a correction should be made before drawing any conclusions from the analysis. The regression calibration and simulation extrapolation methods are some of the available methods developed to deal with this kind of problem. This dissertation proposes a Bayesian method that uses a hierarchical approach to jointly model true radon exposure (measurement error model) and its effect on lung cancer (excess odds model). This method takes subject-specific characteristics into account when making the correction, and uses random effects when missing data are present. We carried out a simulation study in order to compare this method to the regression calibration and simulation extrapolation (SIMEX). Different scenarios were simulated and the simulated data were analyzed with the three methods. This is the first time that these three methods have been compared in the context of radon risk assessment. The simulation results showed that the proposed Bayesian method had a consistent coverage through out the scenarios. However, the SIMEX method had the lowest bias and mean squared error and, most of the time, its coverage was the closest to the nominal coverage of 95%. The regression calibration was the fastest method to be implemented, but it was outperformed by the other methods. The dissertation finalizes by performing individual and pooled analyses using data from five case-control North America radon studies (Iowa, Missouri, Winnipeg, Connecticut, and Utah/South Idaho). The data from each study were analyzed individually, first without making any correction, and then using the three correction methods. Finally, the data were combined and the methods were applied to this bigger sample. To the best of our knowledge, regression calibration and SIMEX have not been implemented using this combined dataset.

Our interest in spatio-temporal models focuses on how a disease spreads within a body region. We use independent replications across individuals to better understand population level dynamics of disease spread. Our Bayesian hierarchical model incorporates independent spatio-temporal datasets to estimate population level parameters. A dimension reduction propagator matrix is used to identify the most variable spatial regions, which are then related to a set of latent variables and covariates. Posterior estimates of parameters allow us to create a predicted estimate of the overall disease evolution process for each individual. In addition, individual level rates of deterioration can be estimated and predictions of future spread are made. The motivating example for this model stems from a study of visual loss in participants with glaucoma. Participants’ vision was recorded across a grid covering the central part of the eye at baseline plus eight follow-up visits every 6 months. We use these spatio-temporal replications of independent participants to determine how human characteristics and demographics collectively affect the spread and progression of glaucoma. Our introduced model is available in the DROIIDS R package. We account for missing data through our model with a Bayesian imputation method.

The objectives of this thesis are 1. to provide a comprehensive guide to using Pocock's group sequential method for clinical trials and 2. to show by computer simulations that the group sequential method is appropriate when using the logistic regression model. In section 1.2, clinical trials are defined with an emphasis on Phase III clinical trials. The primary intent of sections 1.2 and 1.3 is to describe clinical trial characteristics which suggest that interim analyses are desirable. In section 1.4, it is shown that interim analyses should not consist of performing the usual significance tests. Chapter 1, then, describes the basis for the development of the group sequential method. Chapter 2 describes the group sequential method in detail (objective #1) and chapter 3 includes the results of the Monte Carlo study (objective #2).

Repeated measurement data or longitudinal data occur often in statistical applications. For example, in a clinical trial comparing the efficacy of a new treatment with that of a standard treatment, rather than measuring the main response variable only once on each patient, or subject, we can take several measurements over time on each subject. A Repeated measurement study differs from a longitudinal study. The latter generally refers to any study in which one or more response variables are repeatedly measured over time. The former usually imposes some restrictions on the data. One common restriction is that each response variable must be measured at the same time points. In this thesis, the discussion will be restricted to a repeated measurement study, which is defined as follows: a repeated measurement study is a study in which a univariate response variable is repeatedly measured at the same time points on each subject. It should be pointed out, however, that many of the methods discussed here can also be applied to more general longitudinal studies. The analysis of repeated measurement data involves two major difficulties. The first problem is the dependence among successive observations made on the same subject. Multivariate methods modeling the joint distribution of the repeated measures over time have been developed to solve this difficulty. The other, probably the more severe problem is missing data. In repeated measurement studies, the data are collected over a period of time, which in some studies could be many years. Therefore, complete control over the circumstances under which measurements are obtained is not possible. The occurrence of missing data is more likely in repeated than in non-repeated measure studies, and is sometimes unavoidable. In recent years, many methods for coping with the missing data problem in repeated measurement studies have emerged from various applications. The purpose of this thesis is to review and summarize these methods, apply some of them to a practical problem, and identify the needs of further research.

Nurses at the Medical College of Virginia Hospitals (MCVH) in Richmond, Virginia, use the Newborn Maturity Rating and Classification Tool to identify high risk infants. An estimate of gestational age is made and using this estimate, weight, length, and head circumference measurements are plotted on graphs on the tool to determine if the infant achieves intrauterine growth smaller, larger or equal to gestational age. The data used to generate the graphs on the Newborn Maturity Rating and Classification Tool were collected in Colorado during the 1950's. Two nurses at MCVH questioned the use of these graphs. They wanted to know if graphs produced from their population would be different from the graphs they now use because of population and time differences. An initial pilot study was done to examine any problems with measurement reliability. There were no problems with interrater reliability for the length and head circumference measurements. Examination of the chest circumference measurements revealed that one rater had consistently larger measurements than the other. Data from 98 infants were collected and graphs of weight, length, and head circumference produced. There were differences between the Richmond and Colorado graphs. The 10th percentile for weight for Richmond infants is higher than the 10th percentile for the Colorado infants for 35-42 weeks of gestation. At 40 and 41 weeks of gestation the 90th percentile for the Richmond infants is larger than the 90th percentile for the Colorado infants. These differences result in fewer Richmond infants being identified as small for gestational age and more Richmond infants being classified as large for gestational age than when the Colorado graphs are used.

The visual world paradigm is a tool that is widely used in the field of psycholinguistics to help investigate how people listen and understand words and sentences. Proportions of fixations to several different objects are recorded for a number of subjects, over a specific time period. Researchers have found it difficult to find models that can incorporate multiple random effects, account for the correlated nature of the data, and simultaneously fit multiple fixation curves/groups. We have taken a Bayesian hierarchical modeling approach for this multivariate non-linear longitudinal data. Within in this framework, we look at both parametric and nonparametric approaches in simultaneously modeling multiple curves. Finally, we will look at different comparison techniques to compare these curves under a Bayesian framework.

Time series data involving counts are frequently encountered in many biomedical and public health applications. For example, in disease surveillance, the occurrence of rare infections over time is often monitored by public health officials, and the time series data collected can be used for the purpose of monitoring changes in disease activity. For rare diseases with low infection rates, the observed counts typically contain a high frequency of zeros (zero-inflated), but the counts can also be very large during an outbreak period. Failure to account for zero-inflation in the data may result in misleading inference and the detection of spurious associations. In this thesis, we develop two classes of statistical models for zero-inflated time series. The first part of the thesis introduces a class of observation-driven models in a partial likelihood framework. The expectation-maximization (EM) algorithm is applied to obtain the maximum partial likelihood estimator (MPLE). We establish the asymptotic theory of the MPLE under certain regularity conditions. The performances of different partial-likelihood based model selection criteria are compared under model misspecification. In the second part of the thesis, we introduce a class of parameter-driven models in a state-space framework. To estimate the model parameters, we devise a Monte Carlo EM algorithm, where particle filtering and particle smoothing methods are employed to approximate the high-dimensional integrals in the E-step of the algorithm. Upon convergence, Louis' formula is used to find the observed information matrix. The proposed models are illustrated with simulated data and an application based on public health surveillance for syphilis, a sexually transmitted disease (STD) that remains a major public health challenge in the United States. An R package, called ZIM (Zero-Inflated Models), has been developed to fit both observation-driven models and parameter-driven models.