Harvard Statistics Department Presents First Dempster Prize
Matthew McClellan, Harvard Staff Writer
Editor’s Note: This article originally appeared in the Harvard Gazette on May 23. We have reprinted it with permission.
“I’m not very good with numbers,” said Arthur Dempster, professor emeritus of theoretical statistics, while pondering the founding year of the American Statistical Association.
On May 11, generations of students and faculty celebrated the inauguration of the Arthur P. Dempster Award and the 55th anniversary of the Harvard Statistics Department. Of course, the large turnout gave lie to Dempster’s claim; he’s been quite good with numbers, shown by the many who attended.
Stephen Blyth, professor of the practice in statistics and managing director at the Harvard Management Co., established the Dempster fund to recognize promising graduate students in the department, especially those working in theoretical and foundational statistics. “I’m happy for the award to support ‘deep thinking’ about uncertainty, broadly defined, rather than simply rote or ‘procedural’ application of statistical machinery,” said Blyth, who hosted the event at the Federal Reserve, where the fund is located.
Blyth, who studied under Dempster and earned his PhD in 1992, sees the award as a commemoration of his experience as a student and recognition of the support he received from the faculty and the rest of the department. Dempster’s influence has only grown with time. “I know my argument is not strong enough if I cannot convince Art of its validity,” said Blyth.
The practical applications of statistics have always driven Dempster, so it is no surprise that he spent a year working at Bell Telephone Laboratories before he joined the department in 1958. Or that he has inspired students who work not only in academia, but in finance, technology, and elsewhere.
Much of Dempster’s work has been on integrating logical and probabilistic reasoning beyond the established Bayesian framework, which requires quantification of prior information. Dempster’s rule of combination established a method to combine evidence from different sources and Dempster-Schafer Theory, or the theory of belief functions, which provides a framework for quantifying ignorance.
“I am not pessimistic about the future,” said Dempster, “but I am pessimistic about what we can do with limited data.”
And this is where inaugural award winner Alexander Blocker picks up. His research deals with the problems of preprocessing data: To make reams of data comprehensible to an investigator, the raw data is typically calibrated, smoothed, or otherwise simplified before being analyzed. Information is necessarily reduced, or even destroyed. Retaining more information through preprocessing would allow for stronger conclusions, but scientists want to focus on processes of interest without getting stuck in low-level details of observation.
Blocker’s work finds inspiration and application in biology. Expression microarrays measure the level of gene expression in cell samples, allowing analyses of thousands of genes in parallel. They can be used to study changes in gene expression in response to pathogens in comparison with uninfected samples. “Millions of dollars are now invested in building huge collections of genomic data,” said Blocker. “For example, some hospitals are now routinely collecting microarray data on incoming cancer patients.”
As hospitals build massive databases to catalog the genetic profiles of tumors, improving preprocessing techniques could allow researchers to make better use of data they are already collecting. It also could identify relevant supplementary information that would allow investigators to reinterpret data later. “Handled improperly, preprocessing could compromise the utility of these endeavors, wasting effort and delaying scientific progress,” said Blocker.
Blocker’s ultimate goal is to establish principled guidelines for data preprocessing. The ambition was betrayed by a few nervous laughs and one sincere “good luck” from the crowd.
From the example of Dempster and the founding faculty, this kind of ambition has become a departmental tradition. Blocker also enjoys the department’s tradition of strong advising. Even the title of his presentation betrayed its influence: “The Potential and Perils of Preprocessing: A Multiphase Investigation.” “By the amount of alliteration,” said Blocker, “you all know one thing: I am Xiao-Li’s student.”
Xiao-Li Meng, chair of the department, said that when Blocker first came to the department (after studying economics at Boston University), he quickly produced research results that had eluded some of Meng’s existing PhD students and continues to produce strong research results. But, said Meng, Blocker and his research, very much like the department, are still great works in progress.
“When I first started, Xiao-Li would give me advice, and I would figure out how right he was two months later,” said Blocker. “Four years later, I’ve shortened that cycle to two weeks. If I get it down to one week, I think I’ll be ready to graduate.”