Standard parametric modelling approaches by applied researchers assume that a single distribution representative of population being sampled and modifications to that distribution are driven covariates in the regression model. In many cases, the implicit restrictions of the model structure (e.g., a linear function can describe the data) and underlying distribution assumption (e.g., the data has a normal distribution) does not impede reasonable analysis of the data at hand. However, when the data is difficult to fit, finite mixture models may provide useful alternative approach. An overview of various mixture distribution will be presented as well as a description of how EM-algorithm is used to estimate parameters for finite mixture. The local maxima problem associated with finite mixture models and few possible solutions will be presented. An example of a how a finite mixture model was used to estimate risk of exposure to ticks infected with Borrelia burgdorferi (the etiologic agent of Lyme Disease). The example will illustrate how the flexible nature of the mixture model allows the relationship between ecological values and risk of exposure can change through space and time. A second example of using mixture models to fit difficult to fit survival data will be presented. The example will illustrate how loosening the assumption of a single underlying distribution can allow for important improvements to fit of the data.

Bio: Paul Cislo received a BS in Biology from UCONN (Class 1992), an MS in Marine Science from University of South Carolina (Class of 1994) and a PhD in Biostatistics from Yale University (Class 2011). His dissertation work focused on development of spatial mixture models and applying those models to estimate risk of exposure to Borrelia burgdorferi (the etiologic agent of Lyme Disease) in the lower 48 states. After leaving an academic setting, Paul has primarily focused on health economics and outcomes research within the pharmaceutical industry gaining experience at Bayer, BMS, and Pfizer. Paul is a member of the Health Economics & Outcomes Research Statistics Group at Pfizer, which part of its Statistical Research & Data Science Center.

DATE: Wednesday, 3/1/23
TIME: 4:00 PM
PLACE: AUST 163
Webex link: https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=m3fbdb2de3b70213da9f4e2e4035005f5

Coffee will be served at 3:30 pm in the Noether Lounge (AUST 326)

Jingjing Yin, PhD.
Associate Professor of Biostatistics
Jiann-Ping Hsu College of Public Health
Georgia Southern University

Joint inference of ROC measures for diagnostic biomarker/test evaluation

It is common to use the summary ROC measures such as the area under the ROC curve (AUC) for diagnostic test accuracy evaluations and comparisons. We propose to use the AUC and the Youden index jointly for making inferences about diagnostic tests, as the two indices describe different aspects of the ROC curve. This can be done by first estimating the joint confidence region of the AUC and the Youden index. For deciding if a marker is achieving the targeted values, or comparing two biomarkers in a paired design, in terms of both the AUC and the Youden index, we can perform joint testing for such order-restricted hypotheses for which the traditional likelihood ratio test or its variant cannot apply. We propose and compare three testing procedures: 1) the intersection-union test; 2) the conditional test; and 3) the joint test. The performance of the proposed inference methods was evaluated and compared through simulations. The simulation results demonstrate that the proposed joint confidence region maintains the desired confidence level, and all three tests maintain the type I error under the null. Furthermore, among the three proposed testing methods, the conditional test is the preferred approach with markedly larger power consistently than the other two competing methods. In conclusion, estimating and testing jointly on the AUC and the Youden index gives more reliable results for biomarker evaluations than using a single index.

Bio: Dr. Jingjing Yin has a background in statistical methods in medical diagnostics, parametric and nonparametric inference, order-restricted inference, and sampling design. Meanwhile, she has diverse collaborative research in the field of medical science, Epidemiology, Environmental health, and health communication and management. Beyond traditional biomedical statistical research, she has also applied her expertise to the emerging field of big data analytics, with a focus on text-mining social media data. She has published more than 80 peer-reviewed articles in various statistical or medical journals including Statistics in Medicine, Statistical Methods in Medical Research, Biometrical Journal, Pharmaceutical
Statistics, and Journal of Biopharmaceutical Statistics.

Dr. Jingjing Yin is an Associate Professor of Biostatistics at Jiann-Ping Hsu College of Public Health, Georgia Southern University. She earned her Ph.D. (2014) and M.S. (2011) in Biostatistics from the University at Buffalo, and her B.A. (2009) from West China University of Medical Science in Chengdu, China.
DATE: Wednesday, 3/8/23
TIME: 4:00 PM
PLACE: Virtual
Webex link: https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=m85220c30543a9b4fa0a235575e52ba46

Shouhao Zhou
Department of Public Health Services
Division of Biostatistics and Bioinformatics
Penn State Cancer Institute
Pennsylvania State University

Predictive Bayes factors with an application model-assisted design

Bayes factors approach has been widely used as an applied model selection tool by comparing posterior model probabilities. However, it is often criticized for counterintuitive results in statistical hypothesis testing. We propose a new approach, predictive Bayes factors, for predictive model selection of fitted Bayesian models via posterior model probability. In theory, the approach asymptotically estimates the log predictive density ratio. In practice, it dramatically reduces sensitivity to variations in the prior and totally avoids the Lindley’s paradox in testing point null hypothesis, with decent small sample performance. We also develop a novel Bayesian model-assist trial design using this new approach, to determine the escalation and de-escalation boundaries for dose-finding trials. It overcomes the limitations of the previous model-assisted designs and serves as the first model-assisted design to guarantee global optimality and the asymptotic convergence to true MTD. Simulation results also demonstrate the superior operating characteristics based on a large number of scenarios.

Bio: Shouhao Zhou is an Assistant Professor of Biostatistics at Pennsylvania State University. He got his PhD in Statistics from Columbia University and previously was a junior faculty at MD Anderson Cancer Center, where he learned the state-of-the-art Bayesian adaptive designs. He is also interested in Bayesian hypothesis testing and model selection, meta-analysis, and statistical modeling of preclinical data.

DATE: Wednesday, 3/22/23
TIME: 4:00 PM
PLACE: AUST 163

Coffee will be served at 3:30 pm in the Noether Lounge (AUST 326)

Jing Qin, PhD
National Institute of Allergy and Infectious Diseases

A shape restricted propensity score matching method in casual inference
(joint work with Yukun Liu, East China Normal University)


Propensity score matching (PSM) is a pseudo-experimental method that uses statistical techniques to construct an artificial control group by matching each treated unit with one or more untreated units of similar characteristics. To date, the problem of determining the optimal number of matches per unit, which plays an important role in PSM, has not been adequately addressed. We propose a tuning-parameter-free PSM method based on the nonparametric maximum-likelihood estimation of the propensity score under the monotonicity constraint. The estimated propensity score is piecewise constant, and therefore automatically groups data. Hence, our proposal is free of tuning parameters. The proposed estimator is asymptotically semiparametric efficient for the univariate case, and achieves this level of efficiency in the multivariate case when the outcome and the propensity score depend on the covariate in the same direction.

We conclude that matching methods based on the propensity score alone cannot, in general, be efficient.

Bio: Jing Qin is a Mathematical Statistician at Biostatistics Research Branch in National Institute of Allergy and Infectious Diseases. He earned his Ph.D. from University of Waterloo (1992) and then became an assistant Professor at University of Maryland, College Park. Before moving to National Institute of Health (2004), he worked at Memorial Sloan-Kettering Cancer Center for 5 years. Dr. Qin’s research interests include empirical likelihood method, case-control study, biased sampling problems (now called covariate shift or prior probability shift problems in machine learning literature), survival analysis, missing data, causal inference, genetic mixture models, generalized linear models, survey sampling and microarray data analysis. He was elected as an American Statistical Society Fellow, 2006. Dr. Qin published a monograph on "Biased sampling, over-identified parametric problems and beyond" in 2017 (Springer, ICSA book series in statistics).

DATE: Wednesday, 3/29/23
TIME: 4:00 PM
PLACE: Virtual
Webex link: https://uconn-cmr.webex.com/uconn-cmr/j.php?MTID=m636152ee63a668032e47201fadcb0a90