Recent Past Seminars

Abstract: To address the appropriate management of individuals under an ever-changing screening landscape (i.e., the introduction of new screening technology, electronic health records providing greater access to patient histories, and HPV vaccination), representatives from 19 professional organizations agreed to change from issuing recommendations based on test results to recommendations based on precancer risk and a pre-agreed set of clinical action risk thresholds. Using electronic-health records from nearly 2 million women undergoing routine screening from 2003 to 2017, we estimated precancer risk for combinations of screening test results and relevant past histories. Because there can be precancers prevalent at the initial screen and precancer status is intermittently observed, resulting in left-, interval-, and right-censored time of precancer onset, we fit the data using prevalence-incidence mixture models (i.e., jointly estimated logistic regression and proportional hazards models). To inform the consensus risk thresholds, we provided to the working groups estimates of delayed diagnosis vs. colposcopic efficiency trade-offs. The new risk-based management recommendations were then externally validated using data from 2 trials, the New Mexico HPV-precancer registry, and a CDC program that provided screening for underinsured and uninsured individuals.

Location: Online via Zoom

Abstract: Drug development is a complex and expensive scientific endeavor. Statisticians play a unique role in the pharmaceutical industry with their quantitative training. Their work impacts business decisions that drive the success of a drug development. They are involved in all stages of a clinical trial, from trial design, protocol writing, data collection, to data analysis and results interpretation. Their contributions do not stop there. They play a pivotal role in the entire clinical development program and its lifecycle management. They have the opportunities to innovate clinical trials, to advance science, and to create new drugs that benefit millions of patients. In this presentation, we will provide an overview of how statisticians contribute to the drug development and what they do in their daily life from an early phase biometrics point of view. Particularly, we will cover some of the exciting innovations that we are working on at AstraZeneca.

Location: Building D, Warwick Evans Conference Room

Abstract: In this study, we consider the problem of regressing covariance matrices on covariates of interest. The goal is to use covariates to explain variation in covariance matrices across units. Building upon our previous work, the Covariate Assisted Principal (CAP) regression, an optimization-based method for identifying components associated with the covariates using a generalized linear model, two approaches for high-dimensional covariance matrix outcomes will be discussed.Our studies are motivated by resting-state functional magnetic resonance imaging (fMRI) studies. In the studies, resting-state functional connectivity is an important and widely used measure of individual and group differences. Yet, extant statistical methods are limited to linking covariates with variations in functional connectivity across subjects, especially at the voxel-wise level of the whole brain. Our work introduces modeling approaches that regress whole-brain functional connectivity on covariates and enable identification of brain sub-networks. The first approach identifies subnetworks that are composite of spatially independent components discovered by a dimension reduction approach (such as whole-brain group ICA) and covariate-related projections determined by the CAP regression. The second approach directly performs generalized linear regression by introducing a well-conditioned linear shrinkage estimator of the high-dimensional covariance matrix outcomes, where the shrinkage coefficients are proposed to be common across matrices. The superior performance of the proposed approaches over existing methods are illustrated through simulation studies and resting-state fMRI data applications.

Location: Online via Zoom

Abstract: Social network analysis has created a productive framework for the analysis of the histories of patient-physician interactions and physician collaboration. Notable is the construction of networks based on the data of “referral paths” — sequences of patient-specific temporally linked physician visits — in this case, culled from a large set of Medicare claims data in the United States. Network constructions depend on a range of choices regarding the underlying data. In this talk we first introduce the use of a five-factor experiment that produces 80 distinct projections of the bipartite patient-physician mixing matrix to a unipartite physician network derived from referral path data, which is further analyzed at the level of the 2,219 hospitals in the final analytic sample. We summarize the networks of physicians within a given hospital using a range of directed and undirected network features (quantities that summarize structural properties of the network such as its size, density, and reciprocity). The different projections and their underlying factors are evaluated in terms of the heterogeneity of the network features across the hospitals and association with a hospital-level outcome. In the second part of the talk, we use the findings from the first part to construct a shared-patient network of 10,661 physicians who delivered care to Medicare patients in Ohio that we examine for associations with the physicians’ risky prescribing behaviors. Risky-prescribing is the excessive or inappropriate prescription of drugs (e.g., Opioids) that singly or in combination pose significant risks of adverse health outcomes. This enables the novel decomposition of peer-effects of risky-prescribing into directional (outbound and inbound) and bidirectional (mutual) relationship components. Using this framework, we develop models of peer-effects for contagion in risky-prescribing behavior. Estimated peer-associations were strongest when the patient-sharing relationship was mutual as opposed to directional. Using simulations, we confirmed that our modeling and estimation strategies accurately and precisely estimate each type of peer-effect (mutual and directional). We also show that failing to account for these distinct mechanisms, a form of model misspecification, produces misleading results, demonstrating the importance of retaining directional information in the construction of physician shared-patient networks. 

Location: Online via Zoom

Abstract: For nearly a hundred years, the concept of “statistical significance” has been fundamental to statistics and to science. And for nearly that long, it has been controversial and misused as well. In a completely non-technical (and generally humorous) way, ASA Executive Director Ron Wasserstein will explain this controversy, and say why he and others have called for an end to the use of statistical significance as means of determining the worth of scientific results. He will talk about why this change is so hard for the scientific community to make, but why it is good for science and for statistics, and will point to alternate approaches.

Location: Warwick Evans Conference Room, Building D

Abstract: This session will provide you with an understanding of what a biosimilar product is and how it compares to the reference biologic product.  The session will include a review of how a company establishes that the proposed biosimilar is highly similar to the reference product.  In addition, the session will provide an overview of how FDA has handled interchangeability of biosimilar insulin products for the reference product.  The session will also review how someone can find information about biosimilar and reference products in FDA’s Purple Book as well as information on small molecules in FDA’s Orange Book.

Location: Online via Zoom

Abstract: In recent years, wearable and digital devices are gaining popularity in clinical trials and public health research due to technology advances, thanks to their reduced size, prolonged battery life, larger storage, and faster data transmission. In public health, countless research has been conducted to better understand the complex signals collected by wearable and digital devices and their relationship with human health. In clinical trials, emphasis has been given to the derivation and validation of novel digital endpoints and proper statistical methods to model such digital endpoints to objectively quantify disease progression and treatment effects. Such devices provide us a unique opportunity to reconsider the way data can be collected, from snapshot in-clinic measurements to continuous recording of behaviors of daily life. This presentation will highlight some representative examples of statistical and machine learning methodologies and applications of wearable and digital devices in health research.

Location: Online via Zoom

Abstract: Current precision medicine of cancer matches therapies to patients based on average molecular properties of the tumour, resulting in significant patient benefit. However, despite the success of this approach, resistance to drugs develops leading to variability in the duration of response. The approach is based on static molecular patterns observed at diagnosis whereas cancers are constantly evolving. We, therefore, focus on Dynamic Precision Medicine, an evolutionary-guided precision medicine strategy that explicitly considers intra-tumour heterogeneity and subclonal evolution and plans ahead in order to delay or prevent resistance. Clinical validation of such an evolutionary strategy poses challenges and requires bespoke development of clinical trial designs. In this talk, I will present preliminary results on the construction of such trial designs. The work is a joint collaboration with Georgetown (Robert Beckman, Matthew McCoy).

Location: Online via Zoom

Abstract: In this presentation, I’ll talk about recent work in my group that looked into a strange finding:  when training neural networks, periodically resetting some or all parameters can be helpful in promoting better solutions.  Beginning with a discussion of our findings in supervised learning where we relate this strategy to Iterated Learning, a method for promoting compositionality in emergent languages. We then show how parameter resets appear to offset a common flaw of deep reinforcement learning (RL) algorithms: a tendency to rely on early interactions and ignore useful evidence encountered later. We apply this reset mechanism to algorithms in both discrete (Atari 100k) and continuous action (DeepMind Control Suite) domains and observe consistently improving performance. I conclude with some recent finding and speculation about the underlying causes behind the observed effects of parameter resets. 

Location: Online via Zoom

Abstract: This talk introduces changepoint issues in time-ordered data sequences and discusses their uses in resolving climate problems.  An asymptotic description of the single mean shift changepoint case is first given.  Next, a penalized likelihood method is developed for the multiple changepoint case from minimum description length information theory principles.  Optimizing the objective function yields estimates of the changepoint numbers and location time(s). The audience is then walked through an example of a climate precipitation homogenization. The talk closes by addressing the climate hurricane controversy: are North Atlantic Basin hurricanes becoming more numerous and/or stronger? 

Location: Online via Zoom

 Abstract: In this paper, we consider the Bayesian design of a randomized, double-blind, placebo-controlled superiority clinical trial. To leverage multiple historical data sets to augment the placebo-controlled arm, we develop three conditional borrowing approaches built upon the borrowing-by-parts prior, the hierarchical prior, and the robust mixture prior.  The operating characteristics of the conditional borrowing approaches are examined. Extensive simulation studies are carried out to empirically demonstrate the superiority of the conditional borrowing approaches over the unconditional borrowing or no-borrowing approaches in terms of controlling type I error, maintaining good power, having a large “sweet-spot” region, minimizing bias, and reducing the mean squared error of the posterior estimate of the mean parameter of the placebo-controlled arm.  Computational algorithms are also developed for calculating the Bayesian type I error and power as well as the corresponding simulation errors. This is a joint work with Wenlin Yuan and John Zhong.

Location: Online via Zoom

Abstract: Causal inference with observational longitudinal data and time-varying exposures is complicated due to the potential for time-dependent confounding and unmeasured confounding. Most causal inference methods that handle time-dependent confounding rely on either the assumption of no unmeasured confounders or the availability of an unconfounded variable that is associated with the exposure (e.g., an instrumental variable). Furthermore, when data are incomplete, validity of many methods often depends on the assumption of missing at random. The proposed approach combines a parametric joint mixed-effects model for the study outcome and the exposure with g-computation to identify and estimate causal effects in the presence of time-dependent confounding and unmeasured confounding. G-computation can estimate participant-specific or population-average causal effects using parameters of the joint model. The joint model is a type of shared parameter model where the outcome and exposure-selection models share common random effect(s). The joint model is also extended to handle missing data and truncation by death when missingness is possibly not at random. Performance of the proposed method is evaluated using simulation studies, and the method is compared to both linear mixed- and fixed-effects models combined with g-computation as well as to targeted maximum likelihood estimation. The method is applied to an epidemiologic study of vitamin D and depressive symptoms in older adults and can be implemented using SAS PROC NLMIXED software, which enhances accessibility of the method to applied researchers.

Location: Online via Zoom

Abstract: Randomized clinical trials (RCT) are the gold standard for approvals by regulatory agencies. However, RCT’s are increasingly time consuming, expensive, and laborious with a multitude of bottlenecks involving volunteer recruitment, patient truancy, and adverse events. An alternative that fast tracks clinical trials without compromising quality of scientific results is desirable to more rapidly bring therapies to consumers. We propose a model-based approach using nonparametric Bayesian common atoms models for patient baseline covariates. This specific class of models has two critical advantages in this context: First, the models have full prior support, i.e., allow to approximate arbitrary distributions without unreasonable restrictions or shrinkage in specific parametric families; and second, inference naturally facilitates a re-weighting scheme to achieve equivalent populations. We prove equivalence of the synthetic and other patient cohorts using an independent separate verification. Failure to classify a merged data set using a flexible statistical learning method such as random forests, support vector machines etc. proves equivalence. We implement the proposed approach in two motivating case studies.

Location: Online via Zoom

Abstract: We present a general theoretical and algorithmic analysis of the problem of multiple-source adaptation, a key learning problem in applications such as medical diagnosis, sentiment analysis, speech recognition, and object recognition. We will also report the results of several experiments demonstrating the effectiveness of our algorithms and showing that they outperform all previously known baselines.

Location: Online via Zoom

Abstract: Science applied in a medicines regulatory context (“regulatory science”) sometimes has features that differs from traditional “academic science”. At the centre of regulatory science is that analyses must be decision-relevant, timely, and occasionally must rely on poor data. This seminar will present examples of biostatistical analyses that can support regulatory decision-making in the post-authorisation surveillance phase (e.g. pharmacoepidemiology). Examples will include analyses relying on data from the network of population-based Nordic healthcare databases. Lastly, the challenges posed with the advent of Advanced Therapeutic Medicinal Products (e.g. gene and cell therapies) will be discussed. These therapies have characteristics that are different from traditional medicinal products thereby having implications for approaches to pre- and post-authorisation evaluation.

Location: Online via Zoom

Abstract: Early or late phase clinical trials that aim to enroll a biomarker selected patient population are often initiated using a clinical trial assay (CTA) which may differ in assay components compared to the final companion diagnostic assay (CDx), potentially necessitating further clinical analysis to bridge study conclusions to results based on the CDx.  There may be substantial missing data due to the retrospective nature of CDx sample testing.  Key elements of the ideas behind bridging will be reviewed using a case study conducted for a randomized trial of pembrolizumab in second line metastatic non-small cell lung cancer.  Emphasis is on methods aimed at constructing an imputation model to (1) confirm the robustness of clinical trial conclusions via the Bayesian posterior predictive distribution of the study’s intention-to-treat testing results and (2) conduct sensitivity analysis for estimands of the intention-to-diagnose population, while capturing all sources of variability. 

Location: Online via Zoom

Abstract: Since 2014, synthetic controls, defined as a statistical method used to evaluate the comparative effectiveness of an intervention using a weighted combination of external controls in successful regulatory submissions in rare diseases and oncology is growing. In the next few years, the utilization of synthetic controls is likely to increase within the regulatory space owing to concurrent improvements in medical record collection, statistical methodologies, and sheer demand.  In this talk, I will focus on existing and new strategies from the applications of synthetic controls in the framework of augmented control designs. This will include (1) matching strategies and use of entrophy balancing; (2) distinction of causal estimands in augmented designs; (3) Bayesian methodologies for incorporating external information; (4) novel adaptive designs incorporating external information. 

Location: Online via Zoom

Abstract: Big data is best viewed not in terms of its size but in terms of its purpose:  if we can measure it all, maybe we can describe it all. In a molecular context, big data challenges the biomedical research community to describing all the levels of variation based on hundreds of thousands simultaneous measures and data types of DNA, RNA, and protein function alongside patient and tumor features. The bigger and more molecular data, the bigger the promise of advances and the greater the opportunities and challenges posed to the biomedical research community in our ability to harness the power of such data to obtain quick and reliable insights from it. Cancer research has been witnessed to many opportunities on the analytical front from the use of big data for big advances to usher in the era of precision medicine. We posit that the use of big data for big advances in cancer starts at the DNA-level and requires the synergy with, not replacement of, classical hypothesis-driven methods. We describe several methods, data types and metrics for detection of DNA variation at both the gene-and sample-level within the context of pancreatic cancer and their fit for purpose to improve our understanding of DNA level heterogeneity associated with phenotype diversity.   

Location: Online via Zoom

Abstract: Empirical data are seldom complete. Missing data pose a threat to the validity of statistical inferences when missingness is not a completely random process. Model based multiple imputation (MI) can make uses of all available information in the data file to predict missing information and can produce valid statistical inferences in many scenarios. In this talk, I give an introduction to MI, discuss pros and cons of MI and demonstrate how to use the popular mice package in R to create model based multiple imputations of missing values. Finally, I also show how to specify more advanced imputation models (using further add-ons to the mice package) for example for longitudinal count data based on piecewise growth curve models assuming a zero-inflated Poisson or negative Binomial data generating process.

Location: Online via Zoom

Abstract: In this talk, new computer vision algorithms to learn complex morphologies and phenotypes that are important for human diseases will be presented. I will illustrate this approach with examples that capture physical scales from macro to micro: 1) video-based AI to assess heart function (Ouyang et al Nature 2020), 2) generating spatial transcriptomics from histology images (He et al Nature BME 2020), 3) and learning morphodynamics of immune cells. Throughout the talk I’ll illustrate new design principles/tools for human-compatible and robust AI that we developed to enable these technologies (Ghorbani et al. ICML 2020, Abid et al. Nature MI 2020). 

Location: Online via Zoom

Abstract: Smoothing spline mixed-effects density models are proposed for the nonparametric estimation of density and conditional density functions with clustered data. The random effects in a density model introduce within-cluster correlation and allow us to borrow strength across clusters by shrinking cluster specific density function to the population average, where the amount of shrinkage is decided automatically by data. Estimation is carried out using the penalized likelihood and computed using a Markov chain Monte Carlo stochastic approximation algorithm.  We apply our methods to investigate evolution of hemoglobin density functions over time in response to guideline changes on anemia management for dialysis patients. Smoothing spline mixed-effects density models are proposed for the nonparametric estimation of density and conditional density functions with clustered data. The random effects in a density model introduce within-cluster correlation and allow us to borrow strength across clusters by shrinking cluster specific density function to the population average, where the amount of shrinkage is decided automatically by data. Estimation is carried out using the penalized likelihood and computed using a Markov chain Monte Carlo stochastic approximation algorithm. We apply our methods to investigate evolution of hemoglobin density functions over time in response to guideline changes on anemia management  for dialysis patients.

Location: Online via Zoom

Abstract: Phase I oncology trials aim to identify the optimal dose that will be recommended for phase II trials, but the standard designs are inefficient and inflexible. In this talk, I will introduce two innovative Bayesian adaptive designs we developed to improve the efficiency of the trial or incorporate the complex features of the trial. In the first example, we propose PA-CRM, a design for pediatric phase I trials when concurrent adult trials are being conducted. The design automatically and adaptively incorporate information from the concurrent adult trial into the ongoing pediatric trial, and thus greatly improves the efficiency of the pediatric trial. In the second example, we propose a design for early phase platform trials, where multiple doses of multiple targeted therapies are evaluated in patients with different biomarkers, with an objective to identify the best drug at an efficacious and safe dose for the subpopulation defined by a given biomarker. This design addresses complex issues of such platform trials, incorporates information about toxicity and response outcomes from multiple tested doses and biomarkers, and maximizes potential patients benefit from the targeted therapies.

Location: Online via Zoom

Abstract: A fundamental assumption usually made in causal inference is that of no interference between individuals (or units), i.e., the potential outcomes of one individual are assumed to be unaffected by the treatment assignment of other individuals. However, in many settings, this assumption obviously does not hold. For example, in infectious diseases, whether one person becomes infected may depend on who else in the population is vaccinated. In this talk we will discuss recent approaches to assessing treatment effects in the presence of interference.

Location: Online via Zoom

Abstract: Literature on change point analysis mostly requires a sudden change in the data distribution, either in a few parameters or the distribution as a whole. We are interested in the scenario where the variance of data may make a significant jump while the mean changes in a smooth fashion. The motivation is a liver procurement experiment monitoring organ surface temperature. Blindly applying the existing methods to the example can yield erroneous change point estimates since the smoothly-changing mean violates the sudden-change assumption. We propose a penalized weighted least squares approach with an iterative estimation procedure that integrates variance change point detection and smooth mean function estimation. The procedure starts with a consistent initial mean estimate ignoring the variance heterogeneity. Given the variance components the mean function is estimated by smoothing splines as the minimizer of the penalized weighted least squares. Given the mean function, we propose a likelihood ratio test statistic for identifying the variance change point. The null distribution of the test statistic is derived together with the rates of convergence of all the parameter estimates. Simulations show excellent performance of the proposed method. Application analysis offers numerical support to non-invasive organ viability assessment by surface temperature monitoring. Extension to functional variance change point detection will also be presented if time allows.

Location: Online via Zoom

Abstract: In this talk, we exploit the complementing features of randomized clinical trials (RCT) and real world evidence (RWE) to estimate the average treatment effect of the target population. We will review existing methods in conducting integrated analysis of the data from RCTs and RWEs. We will then discuss in detail new calibration weighting estimators that are able to calibrate the covariate information between RCTs and RWEs. We will briefly review asymptotic results under mild regularity conditions, and confirm the finite sample performances of the proposed estimators by simulation experiments. In a comparison of existing methods, we illustrate our proposed methods to estimate the effect of adjuvant chemotherapy in early-stage resected non–small-cell lung cancer integrating data from a RCT and a sample from the National Cancer Database.

Location: Online via Zoom

Abstract: In 2018, there were 52.4 million Americans aged 65 years and older and represented 16% of the population.  By 2040, it is projected that there will be about 80.8 million people in this age group.  The 85 and older population is projected to more than double from 6.5 million in 2018 to 14.4 million in 2040.  Although people of all ages may need post-acute or long-term care services, the risk of needing these services increases with age.  

The National Post-acute and Long-term Care study monitors trends in the supply, provision, and use of the major sectors of paid, regulated long-term care services. NPALS uses survey data on the residential care community and adult day services sectors, and administrative data on the home health, nursing home, hospice, inpatient rehabilitation, and long-term care hospital sectors.  This presentation will describe the study, methodological issues, some findings from the study, and a brief discussion of future possibilities.

Location: Online via Zoom

Abstract: We have proposed a novel, accurate low-cost method to estimate the incubation-period distribution of COVID-19 by conducting a cross-sectional and forward follow-up study. We identified those pre-symptomatic individuals at their time of departure from Wuhan and followed them until the development of symptoms. The renewal process was adopted by considering the incubation period as a renewal and the duration between departure and symptoms onset as a forward time. Such a method enhances the accuracy of estimation by reducing recall bias and using the readily available data. The estimated median incubation period was 7.76 days [95% confidence interval (CI): 7.02 to 8.53], and the 90th percentile was 14.28 days (95% CI: 13.64 to 14.90). By including the possibility that a small portion of patients may contract the disease on their way out of Wuhan, the estimated probability that the incubation period is longer than 14 days was between 5 and 10%.

This is a joint work with Dr. Andrew Zhou’s group at the Biostatistics Department in Peking University.

Location: Online via Zoom

Abstract: Cluster randomized trials (CRTs) are often used in public health research in which the intervention is naturally applied to clusters, or groups, of individuals. Common examples include schools, medical care practices, and even entire communities. In this talk, we introduce possible study design and statistical analysis options for cluster randomized trial. Utilizing real-life examples, we focus on stepped wedge designs and the use of generalized estimating equations. Furthermore, because CRTs often involve only a small number of clusters, small-sample adjustments with the GEE approach are explored.

Location: Online via Zoom

Abstract:  As nowadays confirmatory clinical trials become increasingly complex pursuing several goals simultaneously, the graphical testing approach is often used to address testing of multiple hypotheses of interest.  This approach for a strong control of the familywise error rate (FWER) has several features that are appealing to clinical teams. For example, it allows for a structured visualized evaluation while developing strategies based on importance of hypotheses and corresponding success likelihoods.

At this presentation we consider recent extensions of the graphical testing procedure to group sequential designs using a case study involving “dual” primary endpoints (pCR and MFS) along with several key secondary endpoints. This constitutes potential multiple paths to establish efficacy based on results at interim and final analyses that needs to be statistically addressed. Furthermore, hypothesis tests might have different numbers of analysis points and different rates of accumulation of statistical information.  Study design optimization  and practical implementations will be explored. 

Location: Online via Zoom

Abstract: Evaluation of medical products needs to consider its totality of evidence for benefits and harms that are measured through multiple relevant endpoints. The importance of these endpoints can vary for different clinical settings, clinicians, and patients. Evans and Follmann (2016) advocated the use the desirability of outcome ranking (DOOR) to integrate multiple outcomes according to their importance for the design and analysis of clinical trials. Wang and Chen (2019) proposed testing procedures for trend in benefit-risk analysis based on the importance of multiple outcomes.  In these approaches, the priority levels can be determined by clinicians or based on survey of patients. In this paper, we proposed a stratified randomization design and a composite win ratio endpoint (Pocock et al. 2012) to evaluate the treatment benefits based on patient individually reported outcomes and importance. We introduce motivation examples to illustrate importance of patient heterogeneity in endpoint importance, discuss mathematical properties and difference of our proposed approach with others, and demonstrate the differences and advantages via design examples. We will also discuss our experience using this approach in the development of a shared decision making trial for AFib stroke prevention. 

This is a joint work with Drs. Qian Zhao, Shiyan Yan, Lu TIan, Min M. Tu, and Jie Chen.

Location: Online via Zoom

Abstract: We propose a generalized empirical likelihood (GEL) method for causal inference based on a moment condition model that balances the treatment groups with respect to the potential confounders (e.g., as indicated by a directed acyclic graph). Allowing the number of moment conditions grow with the sample size, we show the asymptotic normality of the causal effects estimates without relying on regression models for the counterfactuals nor for the propensity scores. In a simulation study, we assess the finite sample properties of the proposed estimators in terms of the mean squared error and coverage of confidence intervals. We illustrate an application of the proposed method to data analysis of a training program study.

Authors: Pierre Chausse, Mihai Giurcanu, and George Luta

Location: Online via Zoom

Abstract:  The talk will concern the role of statistical thinking in the Search for Truth using data. This will bring us to a discussion of P-values, which has been, and still is, a central concept in Statistics and Biostatistics and has become a critical and highly controversial tool of scientific research. Recently it has elicited much, sometimes heated, debates and discussions. The American Statistical Association (ASA) was even compelled to release an official statement in early March 2016 regarding this issue, a psychology journal has gone to the extreme of banning the use of P-values in articles appearing in its journal, and a special issue of The American Statistician this year was fully devoted to this issue. The concerns about P-values has also been in relation to important issues of reproducibility and replicability in scientific research. This issue goes to the core of inductive inference and the different schools of thought (Fisher’s null hypothesis significance testing (NHST) approach, Neyman-Pearson paradigm, Bayesian approach, etc.) on how inductive inference (that is, the process of discovering the Truth through sample data) should be done. Some new perspectives on the use of P-values and on the search for truth through data will be discussed. In particular, we will discuss the representation of knowledge and its updating based on observations, and ask the question: “When given the P-value, what does it provide in the context of the updated knowledge of the phenomenon under consideration, and what additional information should accompany it?” This talk is based on the preprint in the Math ArXiV. (http://arxiv.org/abs/1910.05486 (new window))

Location: Warwick Evans Conference Room, Building D
4000 Reservoir Rd. NW, Washington, DC 20057

Abstract: Applied researchers often need to construct confidence sets for policy effects based on functions of estimated parameters, and often these functions are such that the delta method cannot be used. In this case, researchers currently use one of two bootstrap procedures, but we show that these procedures can produce confidence sets that are too small. We provide two alternatives, both of which produce consistent confidence sets under reasonable assumptions that are likely to be met in empirical work. Our second, more efficient method produces consistent confidence sets in cases when the delta method cannot be used, but is asymptotically equivalent to the delta method if the use of the latter is valid. Finally, we show that our second procedure works well when we conduct counterfactual policy analyses using a dynamic model for employment.

Location: Warwick Evans Conference Room, Building D
4000 Reservoir Rd. NW, Washington, DC 20057

Abstract: Genetic analyses have identified etiological factors for several complex disease but understanding how genetic factors interact to increase risk is challenging. In this talk I describe the development of novel methods for large-scale identification of population-level attributes that may affect genome-wide association studies. Because studies we conduct are planned for hundreds of thousands of samples, we had to develop machine-learning based procedures that could be generalized for identifying ethnic variability. We also evaluated tools for jointly estimating the effects of multiple genetic factors by using machine learning tools including random forests and classification trees. These particular analytical schemes identified novel interactions among alleles at multiple loci influencing risk for a rare autoimmune disease. We are now studying the efficacy of modeling with classification tree based analysis versus more traditional approaches such as logistic regression with lasso for high dimensional SNP studies.

Location: Warwick Evans Conference Room, Building D
4000 Reservoir Rd. NW, Washington, DC 20057

Abstract: For oncology clinical trials, time-to-event endpoints, such as overall survival or progression-free survival, are used widely as key endpoints and of great interest. While the classic log-rank test and Cox proportional hazards model have been considered as the “default” analysis methods for statistical inference and for quantifying treatment benefit, we have witnessed many challenges and issues when they cannot be readily or properly applied. Furthermore, in cancer treatment trials we also concurrently observe disease-related longitudinal processes or multiple outcomes that are dependently censored by some “terminal” event like death, and there is increasing need and interest in developing and applying alternative and statistically valid metrics and inference tools to assess the overall disease prognosis and benefit-risk profile more efficiently and in a timely fashion. In recent years, restricted mean survival time (RMST) has gained growing interests as an important alternative metric to evaluate survival data. In this talk we will review and discuss these recent methodological developments, as well as their potential use and implications in oncology clinical trials and drug development.

Location: Warwick Evans Conference Room, Building D
4000 Reservoir Rd. NW, Washington, DC 20057

Abstract: In this talk, we highlight the use of causal inference and machine learning methods to reduce bias when assessing disease burden and to improve precision when estimating intervention effects in randomized trials. We illustrate with data from the SEARCH Study, a large (>320,000-person) cluster randomized trial for HIV prevention and improved community health in rural Kenya and Uganda (NCT01864603).

Location: Zoom only

Abstract: Observational medical databases increasingly find uses for comparative effectiveness and safety research. However, the lack of analytic methods that simultaneously handle the issues of missing data and confounding bias along with the onus of model specification, limit the use of these valuable data sources. We derived a novel machine-learning approach based on trees to estimate the average treatment effect. In order to evaluate causal estimation by model-free machine-learning methods in data with incomplete observations, we conducted a simulation study with data generated from known models of exposure, outcome and missing mechanisms. Thus the true causal effect was known and used as the benchmark for evaluations. Two settings were studied. We compared the bias and standard error of causal estimates from our method to a multiply robust parametric method, the complete case analysis (CC) and a regression analysis after multiple imputations (MI). The proposed methods were applied to a real observational data set of implantable cardioverter defibrillator use.

Location: Proctor Harvey Amphitheater, Med-Dent Building C104
3900 Reservoir Rd. NW, Washington, DC 20057

Abstract: This seminar presents applications of statistics and data science in the design and analysis of biomedical research studies including drug combinations, adaptive and/or targeted clinical trials designs, statistical learning and predictive modeling in head and neck cancer with a series of validation in three large randomized studies, and search for patient subgroups in precision medicine. Part I will focus on drug combinations which are the hallmark of therapies for complex diseases such as cancer. A statistical approach for evaluating the joint effect of the combination is necessary because even in vitro experiments often demonstrate significant variation in dose-effect. I will present a novel methodology for efficient design and analysis of preclinical drug combination studies with applications in the combination of Vorinostat and Ara-C and etoposide. Further discussion will be on the utilization of the preclinical data to guide early phase clinical trial design and a novel adaptive phase I design for drug combinations.

Location: Proctor Harvey Amphitheater, Med-Dent Building C104
3900 Reservoir Rd. NW, Washington, DC 20057

Abstract:  Multi-regional clinical trials (MRCTs) help to synchronize drug development globally, whereby the time lag in marketing authorization among different countries is minimized. However, there are statistical concerns associated with analysis and interpretation of MRCTs. The results of certain countries/regions could vary significantly from the overall results. In this case, controversy exists regarding the extent to which country-specific result should be minimized/ignored and medical scientists/regulators should defer to the overall global treatment effect. Rather than analyzing data separately in each region, our discussion today focuses on developing a Bayesian framework for assessing heterogeneity of regional treatment effects that leverages data collected in other regions. The goal is to make scientifically valid judgments about the play of chance versus real regional differences when comparing results to the overall trial outcome.

Location: New Research Building Auditorium
3970 Reservoir Rd. NW, Washington, DC 20057

Abstract: 

Pathological evaluations are frequently reported on an ordinal scale. Moreover, diseases may progress from less to more advanced stages. For example, HCV infection progresses from normal liver tissue to cirrhosis, dysplastic nodules, and potentially to hepatocellular carcinoma. To elucidate molecular mechanisms associated with disease progression, genomic characteristics from samples procured from these different tissue types are often assayed using a high-throughput platform. Such assays result in a high-dimensional feature space where the number of predictors (P) greatly exceeds the available sample size (N). In this talk, various approaches to modeling an ordinal response when P>N will be described including algorithmic and statistical methods.

Location: New Research Building Auditorium
3970 Reservoir Rd. NW, Washington, DC 20057

Abstract: Based on best practices from federal law enforcement officials, this training program is designed to increase awareness of the “run, hide, fight” protocols in case of an active shooter or violent incident. The course, which uses the “See Something, Say Something” concept, as well as detailed steps in an active shooter incident, is taught by the Georgetown University Police Department (GUPD).

Location: New Research Building Auditorium
3970 Reservoir Rd. NW, Washington, DC 20057

Abstract: Modern clinical trials need to be designed with much flexibility and efficiency. Flexibility includes adaptation based on new information while trial is ongoing and being monitored dynamically with cumulative data. Efficiency includes adequate study power and timely decision making regarding the future course of the trial. Both aspects require the protection of trial integrity and validity. Adaptive sequential designs have been proposed for clinical trials for the last decades. However, traditional computing environment could not accommodate trial monitoring at any time in a timely fashion. In this high-speed, AI – (artificial intelligence) – everywhere era, we introduce Dynamic Data Monitoring concept in which the cumulative data are monitored continuously, and the treatment effect is examined timely over the trial duration. Changing frequency and schedules of the interim analyses can be monitored at real time. The accumulating data can be viewed whenever new data is available and the timing of efficacy and futility assessments and sample size adaptation is made very flexibly, and the type I error rate is always controlled. Numerical and real study examples and simulations are presented. (Collaborated with Drs. Gordon Lan, Joe Shih, and Peng Zhang)

Location: Pre-Clinical Science Building, LA4
3900 Reservoir Rd. NW, Washington, DC 20057

Abstract:  We have proposed new algorithms to estimate the rarefaction curve and accordingly the number of species. The key idea in developing the algorithms is based on an interpolated rarefaction curve through quadratic optimization with linear constraints. Our proposed algorithms are easily implemented and show better performances than existing methods in terms of computational speed and accuracy. We also provide a criterion of model selection to choose tuning parameters and the idea of confidence interval. A broad range of numerical studies including simulations and real date examples are also conducted, and the gain that it produces has been compared to existing methods.

Location: Proctor Harvey Amphitheater, Med-Dent Building C104
3900 Reservoir Rd. NW, Washington, DC 20057

Abstract: The stochastic gradient Markov chain Monte Carlo (SGMCMC) algorithms, such as stochastic gradient Langevin dynamics and stochastic gradient Hamilton Monte Carlo, have recently received much attention in Bayesian computing for large-scale data for which the sample size can be very large, or the dimension can be very high, or both. However, these algorithms can only be applied to a small class of problems for which the parameter space has a fixed dimension and the log-posterior density is differentiable with respect to the parameters. We propose a class of extended SGMCMC algorithms which, by introducing appropriate latent variables and utilizing Fisher’s identity, can be applied to more general large-scale Bayesian computing problems, such as those involving dimension jumping and missing data. For a large-scale dataset with sample size N and dimension p, the proposed algorithms can achieve a computational complexity of O(N^{1+\epsilon} p^{1-\epsilon’}) for some small constants \epsilon and \epsilon’, which is quite comparable with the computational complexity O(N p^{1-\epsilon’}) achieved in general by the stochastic gradient descent (SGD) algorithm. The proposed algorithms are illustrated using high-dimensional variable selection, sparse deep learning with large-scale data, and a large-scale missing data problem. The numerical results show that the proposed algorithms have a significant computational advantage over traditional MCMC algorithms and can be highly scalable when mini-batch samples are used in simulations. Compared to frequentist methods, they can produce more accurate variable selection and prediction results, while exhibiting similar CPU costs when the dataset contains a large number of samples. The proposed algorithms have much alleviated the pain of Bayesian methods in large-scale computing.

Location: Warwick Evans Conference Room, Building D
4000 Reservoir Rd. NW, Washington, DC 20057

Abstract: Extensive research has been conducted in the Multi-Regional Clinical Trial (MRCT) area. To effectively apply an appropriate approach to a MRCT, we need to synthesize and understand the features of different approaches. In this presentation, numerical and real data examples are used to illustrate considerations regarding design, conduct, analysis and interpretation of result of MRCTs. We compare different models as well as their corresponding interpretations of the trial results. We highlight the importance of paying special attention to trial monitoring and conduct to prevent potential issues associated with the final trial results. Besides evaluating the overall treatment effect for the entire MRCT, we also consider other key analyses including quantification of regional treatment effects within a MRCT and the assessment of consistency of these regional treatment effects.

Location: Warwick Evans Conference Room, Building D
4000 Reservoir Rd. NW, Washington, DC 20057

Abstract: Developing targeted therapies based on patients’ baseline characteristics and genomic profiles such as biomarkers has gained growing interests in recent years. Depending on patients’ clinical characteristics, the expression of specific biomarkers or their combinations, different patient subgroups could respond differently to the same treatment. An ideal design, especially at the proof of concept stage, should search for such subgroups and make dynamic adaptation as the trial goes on. When no prior knowledge is available on whether the treatment works on the all-comer population or only works on the subgroup defined by one biomarker or several biomarkers, it’s necessary to estimate the subgroup effect adaptively based on the response outcomes and biomarker profiles from all the treated subjects at interim analyses. To address this problem, we propose an Adaptive Subgroup-Identification Enrichment Design, ASIED, to simultaneously search for predictive biomarkers, identify the subgroups with differential treatment effects, and modify study entry criteria at interim analyses when justified. More importantly, we construct robust quantitative decision-making rules for population enrichment when the interim outcomes are heterogeneous. Through extensive simulations, the ASIED is demonstrated to achieve desirable operating char- acteristics and compare favorably against the alternatives.  

Location: Warwick Evans Conference Room, Building D
4000 Reservoir Rd. NW, Washington, DC 20057

Abstract: Neuroscience has a strong research tradition that employs experimental and observational studies in laboratory settings and controlled testing and evaluation in both clinical, educational and volunteer populations. In the past two decades, there has been increasing interest in conducting population-scale epidemiological studies of early age brain development and functioning as well as later age neurological functioning including cognitive impairment, dementias and Alzheimer’s disease. The data collected in these population-based studies is not restricted to observations on neurological systems and functioning but is collected in parallel with a wide array of information on participants’ life events, medical history, social and environmental exposures, genetics and genomics. This rich array of observational data has the potential to greatly advance our understanding of how complex neurological systems develop, are modified by internal or external factors or otherwise change over the life course. The growing field of epidemiological research also presents many challenging problems in design, measurement, data integration and analysis that those of us trained in biostatistics, bioinformatics and biomathematics will be called on to help to solve.

This presentation will use two cases studies to illustrate the nature of the statistical challenges in conducting population-scale neuroscientific research, describe current best practices and outline opportunities for future research. The first case study will be the Adolescent Brain Cognitive Development project (ABCD, https://abcdstudy.org), a 12-year longitudinal investigation of brain morphology and functional development in U.S. adolescents and teens. The second case study will focus on the challenges in design, measurement and analysis faced in special supplemental investigations of dementia and Alzheimer’s disease conducted under the auspices of the larger Health and Retirement Study (HRS, https://hrs.isr.umich.edu/about). Each case study review will include a description of the specific study challenges and current solutions. The major aim of this presentation is to increase awareness of these emerging lines of research and to promote interest on the part of the next generation of statisticians and data scientists who will be called upon to advance the various methodologies that will be required to better understand complex neurological systems and how they relate to our individual attributes and the world around us.

Location: Warwick Evans Conference Room, Building D
4000 Reservoir Rd. NW, Washington, DC 20057

Abstract:  Similar to its univariate counterpart, multivariate meta-analysis is a method to synthesize multiple outcome effects by taking in to account the available variance-covariance structure. It can improve efficiency over separate univariate syntheses and enables joint inferences across the outcomes. Multivariate meta-analysis is required to address the complexity of the research questions. Multivariate data can arise in meta-analysis due to several reasons.  The primary studies can be multivariate in nature by measuring multiple outcomes for each subject, typically known as multiple-endpoint studies, or it may arise when primary studies involve several comparisons among groups based on a single outcome or measures several parameters. Although it possesses many advantages over the more established univariate counterpart, multivariate meta-analysis has some challenges including modelling and estimating the parameter of interests. Under random-effects model assumption, we discuss the methods of estimating the heterogeneity parameters and effect sizes of the multivariate data and its application by using illustrative example and simulation results. 

Location: Warwick Evans Conference Room, Building D
4000 Reservoir Rd. NW, Washington, DC 20057

Abstract:  “I see yellow; therefore, there is colocalization.” Is it really so simple when it comes to colocalization studies? Unfortunately, and fortunately, no. Colocalization is in fact a supremely powerful technique for scientists who want to take full advantage of what optical microscopy has to offer: quantitative, correlative information together with spatial resolution. Yet, methods for colocalization have been put into doubt now that images are no longer considered simple visual representations. Colocalization studies have notoriously been subject to misinterpretation due to difficulties in robust quantification and, more importantly, reproducibility, which results in a constant source of confusion, frustration, and error. In this talk, I will share some of our effort and progress to ease such challenges using novel statistical and computational tools.

Location: Warwick Evans Conference Room, Building D
4000 Reservoir Rd. NW, Washington, DC 20057

Abstract: Causal inference is a key component for comparative effectiveness research in observational studies. The inverse-propensity weighting (IPW) technique and augmented inverse-propensity weighting (AIPW) technique, which is known as a double-robust method, are the common methods for making causal inference in observational studies. However, these methods are known not stable, particularly when the models for propensity score and the study outcome are wrongly specified.  In this work, we propose a model-free approach for causal inference. While possessing standard asymptotic properties, this method also enjoys excellent finite sample performance and robustness. Simulation studies were conducted to compare with the well-known IPW and AIPW methods for causal inference. A real-life example from an ongoing Juvenile Idiopathic Arthritis Study was applied for the illustration of the proposed method.

Location: New Research Building Auditorium
3970 Reservoir Rd. NW, Washington, DC 20057

Abstract: Joint modeling of longitudinal and survival data is an active area of statistics research and becoming increasingly essential in most epidemiological and clinical studies. As a result, a considerable number of statistical models and analysis methods have been suggested for analyzing such longitudinal-survival data. However, the following issues may standout. (i) a common assumption for longitudinal variables is that model errors are normally distributed due to mathematical tractability and computational convenience. This requires the variables to be “symmetrically” distributed. A violation of this assumption could lead to misleading inferences; (ii) in practice, many studies are often to collect multiple longitudinal exposures which may be significantly correlated, ignoring their correlations may lead to bias and reduce efficiency in estimation; (iii) the longitudinal responses may encounter nonignorable missing; (iv) repeatedly measured observations in time are often interrelated with a time-to-event of interest. Inferential procedures may complicate dramatically when one analyzes data with these features together. Under the umbrella of Bayesian inference, this talk explores a multivariate mixed-effects joint models with skewed distributions for longitudinal measures with an attempt to mediate correlation from multiple responses, adjust departure from normality, and tailor accuracy from nonignorable missingness as well as overcome shortage of confidence in specifying a time-to-event model. A data set arising from diabetes study is analyzed to demonstrate the methodology. Simulation studies are conducted to assess the performance of the proposed joint models and method under various scenarios.

Location: Proctor Harvey Amphitheater, Med-Dent Building C105 
3900 Reservoir Rd. NW, Washington, DC 20057

Abstract: In many social science experiments, subjects often interact with each other and as a result one unit’s treatment influences the outcome of another unit. Over the last decade, a significant progress has been made towards causal inference in the presence of such interference between units. Researchers have shown that the two-stage randomization of treatment assignment enables the identification of average direct and spillover effects. However, much of the literature has assumed perfect compliance with treatment assignment. In this paper, we establish the nonparametric identification of the complier average direct and spillover effects in two-stage randomized experiments with interference and noncompliance. In particular, we consider the spillover effect of the treatment assignment on the treatment receipt as well as the spillover effect of the treatment receipt on the outcome. We propose consistent estimators and derive their randomization-based variances under the stratified interference assumption. We also prove the exact relationship between the proposed randomization-based estimators and the popular two-stage least squares estimators. Our methodology is motivated by and applied to the randomized evaluation of the India’s National Health Insurance Program (RSBY), where we find some evidence of spillover effects on both treatment receipt and outcome. The proposed methods are implemented via an open-source software package.  

Location: New Research Building Auditorium
3970 Reservoir Rd. NW, Washington, DC 20057

Abstract: Physicians frequently use N-of-1 (single-patient) trial designs in an informal manner to identify an optimal treatment for an individual patient. An N-of-1 clinical trial that focuses exclusively on optimizing the primary outcome for a specific patient clearly may not be very useful for generalizability to a population of patients. A series or collection of N-of-1 clinical trials, however, could be generalizable. We review current literature on the design and analysis of N-of-1 trials, and this includes Bayesian approaches as well. We next describe the “Best African American Response to Asthma Drugs (BARD)” trial, which invokes a four-way crossover design and has the flavor of a series of N-of-1 trials. We propose a nonlinear mixed-effects model with a quadrinomial logistic regression for the analysis of the BARD data that constructs six pairwise comparisons of the four asthma treatments to (1) assess the optimal treatment for each study participant and (2) estimate population-level treatment comparisons.

Location: New Research Building Auditorium
3970 Reservoir Rd. NW, Washington, DC 20057

Abstract: Probability books sometimes present “cooked” counterexamples to warn students of the lurking dangers of compromising rigor.  Can such anomalies actually occur in practice?  This talk is proof that they can!  I present actual examples from my clinical trials experience in which I either fell into, or almost fell into, probability traps.  These experiences were a major motivation for our book, “Essentials of Probability Theory for Statisticians” (Proschan and Shaw, 2016, CRC Press, Taylor & Francis Group).   

Location: New Research Building Auditorium
3970 Reservoir Rd. NW, Washington, DC 20057