2011 ASA president, Nancy Geller, addressed members of the association on August 2, 2011, at the Joint Statistical Meetings in Miami Beach, Florida. The theme of the annual meeting was “Statistics: An All-Encompassing Discipline.” Geller’s remarks cover the importance of collaboration in allowing statistics to influence many other disciplines. A written version of the address is the lead article in the December issue of the Journal of the American Statistical Association (JASA). Geller draws on her 30 years of experience collaborating with biomedical scientists to provide advice about what it takes to have successful interdisciplinary projects. She emphasizes the critical role of communication for effective collaboration.

Applications and Case Studies

The Applications and Case Studies section includes numerous articles that represent the outcome of successful collaborations of the type described by Geller in her address. A collaboration between statisticians Roee Gutman and Jun Liu; neuropsychologist David Caplan; and speech, language, and hearing expert Gayle DeDe explores the nature of aphasia, the loss of the ability to produce and/or comprehend language due to brain injury. Their article, “Rasch Model and Its Extensions for Analysis of Aphasic Deficits in Syntactic Comprehension,” finds that the standard approach to analyzing performance on sentence comprehension, the Rasch model, does not capture scientifically important features of their data.

The study being analyzed asks subjects to perform two comprehension tasks on a range of sentence types. The Rasch model generally assumes independence of responses and thus fails to capture, for example, the correlation of respondents’ performance across the two tasks. Gutman and colleagues introduce a mixture model that clusters patients into similar response patterns and abilities. The mixture model better describes the experimental results and produces additional scientific insights into how patients use different cognitive resources in different comprehension tasks.

A second collaborative project uses statistical methods to improve quantitative precipitation forecasting. In “Geostatistical Model Averaging for Locally Calibrated Probabilistic Quantitative Precipitation Forecasting,” authors William Kleiber, Adrian Raftery, and Tilmann Gneiting develop an approach to combining information from an ensemble of weather prediction models that has good performance when examined globally (over a large region) or locally (over a smaller region)

The most basic approach to forecasting precipitation uses an ensemble of numerical weather prediction models to obtain probabilistic forecasts. These, however, are uncalibrated and biased in the sense that the reported probabilities do not match empirical results. A variety of authors have developed statistical post-processing techniques for these forecasts that reduce the bias and produce appropriate calibration when evaluated over a range of locations. Kleiber and coauthors introduce a technique they call “geostatistical model averaging” (GMA) that allows parameters from the precipitation model to vary from locality to locality and consequently achieve better local calibration. They demonstrate the approach on 48-hour-ahead forecasts of daily precipitation in the North American Pacific Northwest, where GMA is shown to have better local and global calibration than existing methods.

Theory and Methods

The notion of “agreement” permeates statistical thinking. A casual survey courtesy of the Web of Science reveals on the order of 200 papers associated with the topic keyword “agreement,” including one paper titled “Agreeing to Disagree.” In biomedical studies “agreement” is often used synonymously with “reliability” to refer to examining the similarity among different diagnostic tests or raters, or to validate a new diagnostic instrument by comparing it to a gold standard. Agreement studies of this type are generally designed to address the reproducibility of measures of the same kind and are therefore limited to comparing measurements on the same scale.

In “A Framework for Assessing Broad Sense Agreement Between Ordinal and Continuous Measurements,” Limin Peng, Ruosha Li, Ying Guo, and Amita Manatunga introduce a concept they call “broad sense agreement” (BSA). BSA extends the classical notion of “agreement” to allow one to assess the correspondence of a continuous measurement and an ordinal scale. The authors propose and study a robust nonparametric estimator of broad sense agreement. Inference, including variance estimation and confidence interval construction, is carried out via large-sample asymptotic approximations. The authors conclude with an application to a study of antidepressants in malignant melanoma patients. The new estimator of broad sense agreement is used to assess the agreement between two measures of depression, the Hamilton Depression Scale (HAM-D) and a self-report dimensional scale (Carroll-D).

Regression tree methods and related recursive-partitioning–based methods have been widely studied since the publication of Leo Breiman and coauthors’ Classification and Regression Trees. Although recursive-partitioning methods are commonly applied to survey data collected via complex sample designs, the sampling properties and large-sample convergence of trees and partitioning methods has been little studied for such data. Daniell Toth and John Eltinge address this gap in their article, “Building Consistent Regression Trees from Complex Sample Data.”

The authors propose and study a method for incorporating complex sample design information when constructing regression trees via recursive partitioning algorithms. Toth and Eltinge demonstrate sufficient conditions on the population, sample design, and partitioning algorithm for their proposed estimator to be asymptotically design unbiased (ADU) and asymptotically design consistent (ADC) with respect to a super-population model. That is, the authors define a nonparametric regression estimator based on a class of recursive partitioning algorithms that is both ADU and ADC for the super-population regression function.

Toth and Eltinge illustrate their recursive partitioning algorithm with an analysis of Occupational Employment Statistics Survey establishment data linked to Quarterly Census of Employment and Wage payroll data of the Bureau of Labor Statistics. The example illustrates the effect of complex sample designs on regression tree modeling and the benefits of the new method.

There are many other informative articles in both sections of the December issue, as well as a set of book reviews. The full list of articles and a list of the books under review can viewed on the JASA website.