It was announced in May, 2012 that Alexander Blocker is the inaugural Arthur P. Dempster award winner. Here is the abstract for his research presentation.
Preprocessing forms an oft-neglected foundation for a wide range of statistical analyses. However, it is rife with subtleties and pitfalls. Decisions made in preprocessing constrain all later analyses and are typically irreversible. Hence, data analysis becomes a collaborative endeavor by all parties involved in data collection, preprocessing and curation, and downstream inference. Even if each party has done its best given the information and resources available to them, the final result may still fall short of the best possible when evaluated in the traditional single-phase inference framework. This is particularly relevant as we enter the era of "big data". The technologies driving this data explosion are subject to complex new forms of measurement error. Simultaneously, we are accumulating increasingly massive databases of scientific analyses. As a result, preprocessing has become more vital (and potentially more dangerous) than ever before. In this talk, we propose a theoretical framework for the analysis of preprocessing under the banner of multiphase inference. We provide some initial theoretical foundations for this area, building upon previous work in multiple imputation. We motivate this foundation with two problems from biology and astrophysics, illustrating multiphase pitfalls and potential solutions. These examples also serve to emphasize the practical motivations behind multiphase analyses --- both technical and statistical. This work suggests several rich directions for further research into the statistical principles underlying preprocessing
It was announced in May, 2013 that Bo Jiang is the second Arthur P. Dempster award winner. Here is the abstract for his paper.
Variable selection methods play important roles in modeling high dimensional data and are keys to data-driven scientific discoveries. In this paper, we consider the problem of variable selection with interaction detection under the sliced inverse index modeling framework, in which the response is influenced by predictors through an unknown function of both linear combinations of predictors and interactions among them. Instead of building a predictive model of the response given combinations of predictors, we start by modeling the conditional distribution of predictors given responses. This inverse modeling perspective motivates us to propose a stepwise procedure based on likelihood-ratio tests that is effective and computationally efficient in detecting interaction with little assumptions on its parametric form. The proposed procedure is able to detect pairwise interactions among p predictors with a computational time of O(p) instead of O(p 2 ) under moderate conditions. Consistency of the procedure in variable selection under a diverging number of predictors and sample size is established. Its excellent empirical performance in comparison with some existing methods is demonstrated through simulation studies as well as real data examples.
It was announced in March, 2014 that Peng Ding is the third Arthur P. Dempster award winner. Here is the abstract for his paper.
Under the potential outcomes framework, causal effects are defined as comparisons between the potential outcomes under treatment and control. Based on the treatment assignment mechanism in randomized experiments, Neyman and Fisher proposed two different approaches to test the null hypothesis of zero average causal effect (Neyman's null) and the null hypothesis of zero individual causal effects (Fisher's null), respectively. Apparently, Fisher's null implies Neyman's null by logic. It is for this reason surprising that, in actual completely randomized experiments, rejection of Neyman's null does not imply rejection of Fisher's null in many realistic situations including the case with constant causal effect. Both numerical examples and asymptotic analysis support this surprising phenomenon. Although the connection between Neymanian approach and the Wald test under the linear model has been established in the literature, we provide a new connection between the Fisher Randomization Test and Rao's score test, which offers a new perspective on this paradox. Further, we show that the paradox also exists in other commonly used experiments, such as stratified experiments, matched-pair experiments and factorial experiments. (https://arxiv.org/abs/1402.0142)
It was announced in March, 2015 that Panagiotis Toulis is the fourth Arthur P. Dempster award winner. Here is the abstract for his paper.
The ideal estimation method needs to fulfill three requirements: (i) efficient computation, (ii) statistical efficiency, and (iii) numerical stability. The classical stochastic approximation of Robbins & Monro (1951) is an iterative estimation method, where the current iterate (parameter estimate) is updated according to some discrepancy between what is observed and what is expected, assuming the current iterate has the true parameter value. Classical stochastic approximation undoubtedly meets the computation requirement, which explains its popularity, for example, in modern applications of machine learning with large datasets, but cannot effectively combine it with efficiency and stability. Surprisingly, the stability issue can be improved substantially, if the aforementioned discrepancy is computed not using the current iterate, but using the conditional expectation of the next iterate given the current one. The computational overhead of the resulting implicit update is minimal for many statistical models, whereas statistical efficiency can be achieved through simple averaging of the iterates, as in classical stochastic approximation (Ruppert, 1988). Thus, implicit stochastic approximation is fast and principled, fulfills requirements (i-iii) for a number of popular statistical models including GLMs, GAMs, and proportional hazards, and it is poised to become the workhorse of estimation with large datasets in statistical practice
Statistical modeling using sets of probabilities offer a low-resolution alternative to precise probabilities. They alleviate the need to make unwarranted modeling assumptions, and can help reduce irreplicable findings. However, sets of probabilities pose a novel challenge on how to properly handle model conditioning in light of new information. Different conditioning rules may lead to different posterior inference from the same model, and may exhibit dilation, contraction and sure loss, paradoxical phenomena never to be seen in precise probability conditioning.
In this talk, I reaffirm the indispensability of sets of probabilities in expressing uncertain inference, through demystifying a collection of famous statistical ``paradoxes’’ within a common framework. I show that a logical fallacy stems from a set of marginally plausible yet jointly incommensurable assumptions, well-captured by a set of probabilities. We revisit the three prisoners/Monty Hall problem and Simpson’s paradox, and establish equivalence between each problem with a set-of-probabilities model equipped with a paradox-inducing conditioning rule. I also discuss theoretical posterior discrepancies between the generalized Bayes rule, Dempster's rule and the Geometric rule as alternative conditioning rules for Choquet capacities of order 2. Our findings highlight the invaluable role of judicious judgment in the handling of low-resolution statistical information.
Joint work with Xiao-Li Meng (arXiv:1712.08946).
Model-X knockoffs is a wrapper that transforms essentially any feature importance measure into a variable selection algorithm, which discovers true effects while rigorously controlling the expected fraction of false positives. A frequently discussed challenge to apply this method is to construct knockoff variables, which are synthetic variables obeying a crucial exchangeability property with the explanatory variables under study. This paper introduces techniques for knockoff generation in great generality: we provide a sequential characterization of all possible knockoff distributions, which leads to a Metropolis–Hastings formulation of an exact knockoff sampler. We further show how to use conditional independence structure to speed up computations. Combining these two threads, we introduce an explicit set of sequential algorithms and empirically demonstrate their effectiveness. Our theoretical analysis proves that our algorithms achieve near-optimal computational complexity in certain cases. The techniques we develop are sufficiently rich to enable knockoff sampling in challenging models including cases where the covariates are continuous and heavy-tailed, and follow a graphical model such as the Ising model.
The 2010 Pickard Award for Teaching and Mentoring went to Joseph Blitzstein, Professor of the Practice in Statistics. The award citation reads:
A WONDERFULLY CLEAR TEACHER AND COMMUNICATOR
A CLEARLY WONDERFUL MENTOR AND SCHOLAR
Three doctoral students were named 2010 Pickard Teaching Fellows.
Cassandra Pattanayak, who is now lecturing at Wellesley, was commended for:
... being a devoted, effective teaching fellow for five courses covering statistics, economics and law; for contributing substantially and innovatively to graduate and undergraduate education as a member of the "Happy Team," including creating the highly innovative and entertaining course trailer for a general education course designed and taught by the Happy Team.
Kevin Rader, who presently is an instructor in the department, was commended for:
... being an outstanding preceptor and Head Teaching Fellow; for exceptional skills in training teaching fellows and facilitators for the Study Network; for exemplary teaching in Harvard summer school, receiving the highest ratings of any other summer school course taught to a similar audience; and for generosity in working with students in class and in distance learning.
Xianchao Xie (a.k.a "Double X"), who currently works at Two Sigma Investments, was commended for:
... being a dedicated, energetic teaching fellow for a wide range of courses, from introductory undergraduate courses through advanced graduate level courses; for achieving the rare distinction of a perfect 5 course evaluation, and for being an extraordinary problem-solver and problem-creator, patiently helping his sections work through many interesting examples.
Congratulations and much appreciation to all!
The Department of Statistics awarded the 2021 Dempster Prize to Ambarish Chattopadhyay for his paper (with Jose Zubizarreta) "On the implied weights of linear regression for causal inference".
Abstract: A basic principle in the design of observational studies is to approximate the randomized experiment that would have been conducted under controlled circumstances. Now, linear regression models are commonly used to analyze observational data and estimate causal effects. How do linear regression adjustments in observational studies emulate key features of randomized experiments, such as covariate balance, self-weighted sampling, and study representativeness? In this paper, we provide answers to this and related questions by analyzing the implied (individual-level data) weights of linear regression methods. We derive new closed-form expressions of the weights and examine their properties in both finite-sample and asymptotic regimes. We show that the implied weights of general regression problems can be equivalently obtained by solving a convex optimization problem. This equivalence allows us to bridge ideas from the regression modeling and causal inference literature. As a result, we propose novel regression diagnostics for causal inference that are part of the design stage of an observational study. As special cases, we analyze the implied weights in common settings such as multi-valued treatments, regression adjustment after matching, and two-stage least squares regression with instrumental variables.
