All Times EDT
Interest in causal inference as a statistical field has grown tremendously over the past decade. While this is partially motivated by the growing availability of non-randomized observational data, it has also been realized that ideas from the field of causal inference provide a useful framework for inference in randomized studies. For example the ICH E9 addendum on estimands uses the counterfactual viewpoint from causal inference to define treatment effects in clinical trials.
In this training we will outline the basic ideas of causal inference (causal effect, potential outcomes, standardization and inverse probability weighting). We will also illustrate how it relates to questions and concepts encountered in randomized clinical trials (intent-to-treat, per-protocol analyses, covariate adjustments in regression analyses, ...) and the strategies defined in the ICH E9 addendum on estimands. While the causal inference framework is in many aspects aligned with pharmaceutical statistics traditions, there are also areas where the framework sheds new light on established traditions, which we will outline in this training.
Outline:
1) Introduction to causal inference and potential outcomes (~45min)
2) Causal inference in randomized clinical trials and drug development (~45min)
3) Estimation methods for causal inference: Standardization and inverse probability weighting, including connection to treatment effect parameters in regression models (~60min)
4) Application example: Treatment switching (~30min optional)
Instructor(s) background: All three instructors have extensive experience with implementation of the ICH E9 addendum over the past few years (including contribution to health authority interactions), and how causal inference ideas and techniques can be utilized to define estimands as well as estimation techniques reflecting clinical questions in randomized clinical trials. The instructors have presented trainings on this topic successfully on multiple occasions.
References
Akacha, M., Bretz, F., Ohlssen, D., Rosenkranz, G., and Schmidli, H. (2017). Estimands and Their Role in Clinical Trials, Statistics in Biopharmaceutical Research, 9, 268-271.
Degtyarev, E., ..., Bornkamp, B., et al. (2019) Estimands and the Patient Journey: Addressing the Right Question in Oncology Clinical Trials. JCO Precision Oncology 3, 1-10.
Magnusson, B., Schmidli, H., Rouyrre, N., Scharfstein D. (2019) Bayesian inference for a principal stratum estimand to assess the treatment effect in a subgroup characterized by postrandomization event occurrence, Statistics in Medicine, 38, 4761-4771.
Bornkamp, B., Bermann G. (2019) Estimating the treatment effect in a subgroup defined by an early post-baseline biomarker measurement in randomized clinical trials with time-to-event endpoint, Statistics in Biopharmaceutical Statistics, to appear, https://doi.org/10.1080/19466315.2019.1575280
Artificial intelligence (AI) or machine learning (ML) has been used in drug discovery in biopharmaceutical companies for nearly 20 years. More recently AI has also been used for the disease diagnosis and prognosis in healthcare. In analysis of clinical trial data, predicted individual patient outcomes for precision medicine, similarity based machine learning (SBML) has recently been proposed for clinical trials for oncology and rare disease without the requirement of big data. The course will focus on supervised learning, including similarity-based learning and deep learning neural networks. We will also introduce unsupervised, reinforcement, and evolutionary learning methods. In addition, initiatives and innovative thinking of AI addressing key challenges in pharmaceutical industry will be discussed. The short course aims at conceptual clarity and mathematical simplicity. R code will be provided with examples for implementation. The course materials are based on instructor’s upcoming book in Feb, 2020: Artificial Intelligence in Drug Development, Precision Medicine, and Healthcare.
The course will cover:
(1) Introduction to AI: Classic Statistics versus AI; Past, Current, and Future of AI in Drug Development, Medicine, and HealthCare
(2) Deep Learning Neural Networks: Convolutional Neural Network (CNN); Recurrent Neural Network (RNN); Long Short-term Memory Networks (LSTMs); Deep Belief Network (DBN)
(3) Similarity Based Method: Similarity-Based Machine Learning; Kernel Method; Nearest-Neighbors Method; Support Vector Machine
(4) Overview of unsupervised, Reinforcement, Collective Intelligence, and Evolutionary Learning Methods
Goals: attendees will learn common AI methods in drug development and medical research, and will be able to use the AI methods with R to analyze clinical trial and other data, and interpret the results.
This one-day short course will cover a variety of sequentially adaptive phase I-II clinical trial designs that use both efficacy and toxicity to optimize dose, the dose pair of a two-agent combination, two doses given in sequence, or dose and schedule. The course will begin with an explanation of fundamental flaws with the conventional paradigm that separates phase I and phase II, followed by an overview of phase I-II designs. The remainder of the course will cover specific designs, with each illustrated by a practical application, including basic design structure, establishing numerical values of design parameters and prior parameters, and computer simulations to establish operating characteristics. Examples will include designs based on either efficacy-toxicity probability trade-offs, elicited joint utilities of efficacy and toxicity, methods for dealing with late onset outcomes, personalized (precision) dose-finding, optimizing molecularly targeted agents, two-agent combination trials, and methods for dealing with drop-outs.
Outline
Morning Lectures (Professor Thall)
• Flaws with the Conventional Phase I ? Phase II Paradigm
• Phase I-II Trials: Using Both Efficacy and Toxicity
• Efficacy-Toxicity Trade-Off Based Designs
• Utility Based Designs
Afternoon Lectures (Professor Yuan)
• Model-assisted Designs
• Designs with Late Onset Outcomes
• Optimizing Molecularly Targeted and Immunotherapy Agents
• Personalized Dose Finding
Learning Objectives
The two over-arching objectives of this short course are (1) to show the attendees the many serious flaws with the conventional approach that separates phase I from phase II, so that they may avoid this approach when possible, and (2) to present practical alternative phase I-II designs and methods.
Text Book
Yuan Y, Nguyen HQ, Thall PF. Bayesian Designs for Phase I-II Clinical Trials. Chapman & Hall/CRC Biostatistics Series. 2016.
Under the 21st Century Cures Act, the FDA is directed to develop a program to evaluate how real-world evidence (RWE) can potentially be used to support approval of new indications for approved drugs or to support/satisfy post-approval study requirements. This brings new opportunities to utilize statistical innovations and advances that are critical to assess and address data quality as well as establish causal inference based on real world data (RWD) for regulatory decision-making. However, designing a valid RWD-based study and draw inferences about RWE face with numerous challenges such as confounding, treatment switching, missing information, etc. Propensity score methods offer powerful and flexible approaches to design and analyze RWE studies that can efficiently address aforementioned statistical challenges. This course will start with a general overview on causal inference and related methods including matching and inverse probability weighting (IPW). We will discuss how to design and analyze an RWD-based study using these methods under a question of interest that is coherent to ICH-E9 regulatory definition of estimand. This course will provide hands-on training based on simulated data and implementation codes in R. During the training, participants will be able to learn (1) efficient way to examine balance of covariate distributions in the target population before/after matching or weighting, how to (2) fit the different causal models, (3) deal with tails of the PS distribution, and (4) interpret study findings based on selected causal model. Then we will illustrate a case-study example that demonstrates a practical use of IPW for drug safety evaluation. A mock SAS code will be provided during the case-study illustration. This course will wrap up with introduction to some recently developed methodologies such as overlap weights that places emphasis on clinical equipoise and targeted maximum likelihood estimation that utilizes machine learning algorithms.
Immunotherapy has emerged as a promising treatment option for cancer in recent years. A major challenge in immune-oncology development is the delayed onset of treatment effects due to the mechanism of immunotherapy which violates the proportional hazard (PH) assumption. It is often referred as the non-proportional hazard (NPH) problem. In contrast to the PH assumption, NPH constitutes a broad class of alternative hypotheses. A suitable design for time to event data with potential NPH needs to be flexible enough to incorporate the uncertainty of NPH type and provide a robust inference. Different alternative design and analysis approaches for immune-oncology trials will be discussed. These include piecewise log-rank test, weighted log-rank test, combination tests and Kaplan-Meier based methods (e.g. restricted mean survival time). We will introduce a new MaxCombo test which provides robust power under different NPH scenarios. The short course will provide the analysis methodology, general design framework and strategies, sample size calculation, strategies of interim analysis, evaluation of operating characteristics and necessary steps for protocol implementation. All methodologies will be illustrated with real life examples and implementation with the available R package Simtrial. Outline: 1. Introduction 2. Alternative methods for design and analysis 3. Introduction of the R package Simtrial 4. Designing a trial with potential NPH 5. Design using Simtrial 6. Group sequential design with potential NPH 7. Summary and discussion Instructor(s): Dr. Satrajit Roychoudhury is a Senior Director and a member of Statistical Research and Innovation group in Pfizer Inc. His areas of research include the use of survival analysis and Bayesian methods. Dr. Keaven M Anderson is a Distinguished Scientist and head of Methodology Research biostatistics group at Merck. He has interest in group sequential design and survival analysis and applications of multiplicity control.
Ask a group of statisticians, “What do effective leaders do?” and you’ll hear a sweep of answers: Leaders set strategy; they demonstrate technical competence; they influence and collaborate across functions; they possess strong communication and negotiation skills. Then ask: “What should leaders do?” If the group is seasoned, you’ll likely hear one response: The leader’s singular job is to get results.
The recently endorsed Leadership- in- Practice Committee (LiPCom) of the Biopharmaceutical Section of the ASA seeks to present an interactive practical leadership workshop. The workshop will introduce real-life scenarios common to statistical practice that require leadership skills to drive results.
Most statisticians don’t enter their careers intending to become leaders and statistical leadership does not always accompany titles. A statistician becomes a leader by recognizing a problem that matters and can turn bold scientific, strategic, or organizational objectives into reality. Statisticians who can do this are increasingly valued in today’s data-driven organizations. As stated by 2012 ASA President Bob Rodriguez, “Leadership ability is a prerequisite for the growth of our field because statistics is an interdisciplinary endeavor and our success ultimately depends on getting others to understand and act on our work.”
This half-day workshop will provide a forum to explore multiple dimensions of leadership and expose ways that statisticians can leverage these leadership skills in driving results.
Much has been written and hyped about Bayesian, adaptive, and complex clinical trials. This short course focuses on the practical details of each of these innovations. We will discuss what each of these concepts are, why they may improve a clinical trial, and the practical ramifications of each of them. A focus will be on providing multiple examples of each.
The What This section will work to explain each of these three innovations. What are adaptive designs, what are not adaptive designs? What are the ramifications of using a Bayesian analysis in a clinical trial? Finally, either of these concepts can create a situation in which calculating the operating characteristics is impossible without clinical trial simulation, and thus we label these trials as complex innovative designs. Examples of clinical trials labeled as adaptive, Bayesian, and complex will be presented.
The Why The second focus of the course will be on why these innovations may be preferred in a clinical trial. Each of these innovations can improve a trial to provide better answers, more answers, and more efficiently. In addition, there can be an impact on the patients in the clinical trial. All of these ramifications will be discussed. Again, examples of trials will be presented to demonstrate the potential promise of these innovations.
The How Perhaps most importantly is how do we utilize these innovations? How does one go about creating an adaptive design? What are the steps, what are the potential missteps of creating an adaptive design? Practical advice will be presented on good practices in the construction of adaptive, Bayesian, and complex clinical trials. Examples of trials will be presented where the focus will be on how they were constructed.
Instructor background:
Scott Berry, Ph.D., is President and a Senior Statistical Scientist at Berry Consultants, LLC and has done numerous short courses with awards.
All statisticians at some point in their careers are frustrated by the rejection of their ideas or resistance when they try to drive change by implementing a new method or process. Through time and experience, some statisticians learn to overcome these issues and successfully influence their collaborators and business partners. How do they do it? What skills do they practice? The simple answer is leadership - skills that you can learn to get people to listen to your perspective, trust your proposals, and invest in your ideas. With leadership skills, your ideas and innovation have a much greater chance of being adopted and of impacting the drugs you develop and the patients you serve. This short course will focus on the leadership skills necessary to influence and to take new ideas and innovative thinking into practice. The course will provide: - A brief introduction of leadership, why it is important, and its role in creating impact through innovation - A discussion of the critical skills required to influence and get collaborators to buy into your ideas, including communication, business acumen, and networking. - What it takes to develop and exhibit a professional presence that will allow you to establish and maintain the ability to influence decisions & strategy and impact your organization. Although this course will not turn you into an instant leader, it will provide you with knowledge of what it takes to improve as a leader and an initial direction & focus to get you started on your leadership journey.
Additional Information
This course will include select topics and concepts from two successful ASA leadership courses: Preparing Statisticians for Leadership (taught annually at JSM since 2014) and Leading with Executive Presence (developed and taught as part of Lisa LaVange’s ASA presidential leadership initiative in 2018). Dr. Gary Sullivan will be the course instructor.
Both Statistics and Machine Learning (ML), which is loosely called Artificial Intelligence in the media, are the fields of learning from data. The fact that they share many underlying math principles and theories overshadow the fact that they are indeed based on different philosophies. Ignoring the differences caused confusion among some statisticians and prevented them from effectively using some ML technologies. This short course is uniquely designed as an introduction to ML for statisticians. It avoids wasting time on topics that statisticians are already familiar. Instead, it puts emphasis on the areas unique in ML, and draws connections between the two fields for those with superficial similarity. This course was recently taught 6 times in Merck. Its success shows its effectiveness for statisticians to learn ML. The course has 5 sections: (1) What ML is: to show similarities and differences between ML and Statistics, and to provide an overview of the subareas within ML; (2) Supervised learning workflow and methods: to explain the workflow, along with key concepts, related to supervised learning tasks, and to introduce those popular ML methods, e.g. SVM, boosting machine, random forests, etc.; (3) Model inference: to use the trained models to predict, to select variables, and to gain insights into the data; (4) Unsupervised learning methods; (5) An introduction to deep learning.
After obtained his Ph.D. in 2001 with research focusing on modern ML, Junshui Ma has been working on many ML topics and projects across academia and industry, which produced a dozen of ML journal papers, covering diverse topics. Andy Liaw has been working in related areas after he obtained his Ph.D. in 1997. He is the author of the first popular R package for random forests, and the associated paper has been cited 10K+ times. One of the papers coauthored by them for predicting compound activities using deep learning methods became one of the most cited papers in that area.
Following the tremendous success in oncology drug development in the last decade (e.g., emergence of immune checkpoint inhibitors as a new backbone therapy in multiple tumor indications and cell-based therapies), recent years have witnessed an explosive growth in number of new drugs and vaccines in cancer trials. While the expectation is high, it is unrealistic to expect all of them to have the same success, as the improved SOC has raised the hurdle for them to demonstrate effectiveness. It is imperative to apply cost-effective designs to the development of these new agents.
In this short course, I will focus on three issues in contemporary oncology development: 1) efficacy screening post dose finding; 2) transition of a new program from early stage to late stage; 3) biomarker hypotheses in Phase 3 confirmatory trials. On efficacy screening, I will present a variety of optimal basket designs including an extension of Simon’s optimal designs for single-arm trials to multi-arm trials. I will also make a distinction between assessing whether any of the test drugs is effective and whether the test drug is effective in any of the tumor indications, and show its impact on design strategies. On transition from early stage to late strategy, I will extend the 2-in-1 design that has quickly become popular since its publication and then talk about advanced utilization of intermediate endpoints for making transitional decisions in an operationally seamless design with dose-selection and a statistically seamless 2-in-1 design. The use of a predictive biomarker can substantially complicate a Phase 3 trial design. I will discuss the various program level and trial options for test drugs with a predictive biomarker hypothesis including adaptive population expansion designs.