The clinical trial community has long understood the utility of historical information in the context of clinical trial design. From just "being aware of the literature", through to formal meta-analytic approaches, statisticians appreciate the role of evidence synthesis when tasked with designing a new study.
But why should that historical information not also play a role in the analysis of the new study, and contribute to inference about the parameter(s) of interest?
Straightforward application of Bayesian methods is one obvious way to bring the historical data (as informative prior) into that inference. But there is understandable scepticism about such a naïve approach (patients in new trials are unlikely to be exactly “the same” as patients in the past) combined with regulatory concerns over the impact on decision-making risks. Yet the feeling remains that to ignore that information is to render sub-optimal the learning process about parameters of interest.
What clinical trial practitioners would like is to be able to use an informative prior distribution that “looks like” the historical data, if the historical data are relevant – suitably modulated to incorporate the knowledge that the future is unlikely to be an exact extrapolation of the past – and which appropriately down-weights the impact of historical data should it not be commensurate with the new data.
This workshop will introduce several methods that have been discussed in the literature, such as power, commensurate and robust mixture priors, with a focus on re-use of control arm data - although other applications will also be highlighted. It is based on a successful course developed by GSK and the UK MRC Biostatistics Unit.
Realistic examples for a pharmaceutical R&D audience will be used, and as each method is introduced, participants will learn how to apply it to real data and to investigate its impact on the trial’s operating characteristics and the posterior inference they would make.
In the era of precision medicine, drug development has become increasingly complex when a companion diagnostic is developed in parallel. Drug/device developers and regulatory agencies have been facing challenges in syncing up the timelines of the two development processes, as well as associated statistical issues. In July 2016, FDA released the draft guidance: “principles for codevelopment of an in vitro companion diagnostic device with a therapeutic product”. This guidance laid out principles, recommendations, and methods for co-developing drug and companion diagnostic device. In addition, numerous statistical methods, especially in the field of trial design, have been proposed to maximize the probability of success of clinical development program. This short course consists of three parts. In Part 1, we will provide an overview of the FDA guidance and general considerations. In Part 2, statistical methods for addressing key issues such as biomarker threshold determination, bridging study, prescreening issues, are discussed. In Part 3, we will review innovative trial designs proposed to include biomarker hypothesis testing that may require a companion diagnostic device. These include platform, basket, umbrella trials for learning phase trials, various statistical testing strategies in biomarker stratified designs and adaptive biomarker designs for confirmatory trials. Basic knowledge of clinical trials statistics is required.
Multiple comparisons is for making multiple decisions. The Partitioning Principle is a fundamental approach to multiple comparisons. This principle guides the formulation of null hypotheses such that controlling the error rate familywise controls the probability of making incorrect decisions. When there is a pre-specified decision path (e.g., testing for efficacy in low dose only after efficacy in high dose has been shown), the Partitioning Principle formulates the null hypotheses in a way that guarantees decision-making will follow the pre-specified path, proving besides that multiplicity adjustment is not needed when testing follows a single path. In the absence of pre-specified decision paths, this Principle covers common step-down methods (e.g., Holm’s) and step-up methods (e.g., Hochberg’s) as special cases.
In settings with multiple paths (e.g., testing involves multiple doses and multiple endpoints), the Partitioning Principle more straightforwardly stays on paths than closed testing. In addition, this Principle crucially proves that if all the null hypotheses are rejected along a path, then the significance level allocated to that path can be re-allocated to the other paths. This helps building the connection to the recently popularized graphical approaches to multiple comparisons. Using graphical approaches, one can easily construct and explore different test strategies and thus tailor the test procedure to the given study objectives. The resulting multiple test procedures are represented by directed, weighted graphs, where each node corresponds to an elementary hypothesis, together with a simple algorithm to generate such graphs while sequentially testing the individual hypotheses.
Examples with testing multiple doses and multiple endpoints in combination, as well as testing targeted therapies with subgroups, will be given to illustrate the above approaches and principles.
In statistical methodology research, simulations are among the ways to show operating characteristics of the proposed method against the existing methods. Depending on the response variables of interest in such simulations, univariate or multivariate, iterative or non-iterative, simulation designs must be considered very carefully to produce generalizable and repeatable conclusions in any given platform and this task is much more difficult and under-recognized than it is supposed to be. In this short course, we will introduce simple to more complex simulation designs and the importance of simulation size; we will describe potential pitfalls that may not be easily recognizable and suggest what metadata to be captured for a clear description of the simulation results. We plan to carry out examples both in SAS and R to show similarities and differences between two platforms. In doing so, we will also utilize Graphical Analytics techniques, which are indispensable components of statistical learning and practice and must be made part of any simulation plans as well. We will introduce graphical approaches for simulation diagnostics and descriptive graphics to summarize the simulations.
Oncology immunotherapies have the potential to be effective in more tumor indications than a non-immunotherapy, as demonstrated by PD-1 (or PD-L1) immune checkpoint inhibitors such as pembrolizumab and nivolumab in recent years. Following the success of PD-1 (or PD-L1) inhibitors, a flood of next generation immunotherapies with different mechanisms of action (e.g., LAG3, CD40, ICOS, TIM-3, IDO1, GITR, STING, OX40, TIGIT) are being developed. While the expectation is high for these new immunotherapies, it is unrealistic to expect all of them to have the same success as the immune checkpoint inhibitors. In reality, they may be effective in a wide range of tumor indications, may be effective only in a few tumor indications, or may not be effective at all. Even if they are indeed as effective as the immune checkpoint inhibitors, it will be challenging to demonstrate their clinical benefit given the improved standard-of-care. It is imperative to apply innovative and cost-effective statistical strategies to the development of these new immunotherapies.
This tutorial will present the lessons and experiences learned from the development of the checkpoint inhibitors. It will also present statistical strategies on efficacy screening, design adaptation between Phase 2 and Phase 3, and adaptive biomarker enrichment design and platform/basket design for Phase 3 trials. Familiarity with multivariate normal distribution and an open mind are the only prerequisite for this tutorial.
Proper addressing missing data in clinical trial analysis remains complex and challenging, despite a great amount of research that has been devoted to this topic. Conventionally, under the missing at random (MAR) assumption, we often use maximum likelihood or multiple imputation based methods for inferences. However, the MAR assumption is unverifiable from data. More critically, the estimand under MAR is hypothetical as indicated in the recent ICH E9 addendum and has been considered as overly-simplistic and unrealistic. Both regulatory agencies and industry sponsors have been seeking alternative approaches to handle missing data in clinical trials under missing not at random (MNAR) assumption.
This half-day tutorial is intended to cover various methods that have been advocated in dealing with missing data and illustrates how to carry out the analyses using SAS software. The tutorial begins with an overview of conventional missing data handling methods such as maximum likelihood methods, multiple imputation, generalized estimation equation approaches, and Bayesian methods. The rest of the course is devoted to more recently-developed methods, such as sensitivity analysis to assess robustness, control-based imputation, and tipping point analysis. Real clinical trial examples will be presented for illustration with implementation of the analysis using SAS/STAT software, including the MIXED, MI, MIANALYZE, GEE, and MCMC procedures.
Outline: 1. Review missing data and conventional methods * Restricted maximum likelihood (REML) * Mixed-effects Model Repeated Measure (MMRM), * Constrained Longitudinal Data Analysis (cLDA) * GEE and wGEE * Multiple imputation (MI) * Bayesian approach
2. Recently-developed methods for missing data * Review common estimands in clinical trials * Alternative and sensitivity analysis models * Control-based imputation * Tipping point analysis
Believe or not, we are in the era of the fourth industrial revolution. It is critical for statisticians to understand or even master the center piece of this revolution, artificial intelligence (AI). Google, Amazon, Facebook, IBM and many other companies have broken new ground in AI including self-driving cars, cashier-free convenience stores, smart hospital care, personal assistants and precision medicine. The course will cover the concept of AI, breakthroughs of AI in drug development including drug discovery, patient recruitment, patient compliance and prediction of patient and clinical trial outcomes, and a tutorial on deep learning methods, a set of novel tools for generating artificial intelligence, which were developed based on a class of almost forgotten old algorithms, neural networks that revived to become a mainstream in big data analytics thanks to advancement of big data and computer processing power. The utility of deep learning in analyzing electronic medical records will be illustrated in terms of predicting patient and clinical trial outcomes for clinical trial optimization. Lastly, an overview will be given on Python, commonly used software for deep learning, and the very necessary computing environment, cloud.
Syllabus: 1. AI: fourth industrial revolution, AI's central role, its impact to required working skills and why statisticians need to understand or even master AI. Impact of AI to clinical development. What is AI and how to build AI?
2. Deep Learning: Why deep learning? i.Successful examples of deep learning, ii. The power of deep learning in analyzing big data Why deep learning works? What is deep learning? i. Neural network and deep learning ii. Big data and deep learning. How to train, validate and interpret a neural network?
3. EMR: Why? i.Real world evidence ii.Competitive intelligence. iii. Clinical trial results prediction. What are EMR? How to analyze EMR? Case Study
4. Python and cloud computing
This short course focuses on adaptive enrichment designs, that is, designs with preplanned rules for modifying enrollment criteria based on data accrued in an ongoing trial. For example, enrollment of a subpopulation where there is sufficient evidence of treatment efficacy, futility, or harm could be stopped, while enrollment for the complementary subpopulation is continued. Such designs may be useful when it’s suspected that a subpopulation may benefit more than the overall population. The subpopulation could be defined by a risk score or biomarker measured at baseline. Adaptive enrichment designs have potential to provide stronger evidence than standard designs about treatment benefits for the subpopulation, its complement, and the combined population. However, there are tradeoffs in using such designs, which typically require greater sample size than designs that focus only on the combined population.
We present new statistical methods for adaptive enrichment designs (part 1 of the course), simulation-based case studies in Stroke, Cardiac Resynchronization Therapy, Alzheimer’s Disease, and HIV (part 2 of the course), and open-source, trial optimization software (part 3 of the course). The tradeoffs involved in using adaptive enrichment designs, compared to standard designs, will be presented. Our software searches over hundreds of candidate adaptive designs with the aim of finding one that satisfies the user’s requirements for power and Type I error at the minimum expected sample size, which is then compared to simpler designs in terms of sample size, duration, power, Type I error, and bias in an automatically generated report.
Target Audience: statisticians at master's level or beyond.