A great many tools have been developed for supervised classification, ranging from early methods such as linear discriminant analysis through to modern developments such as neural networks and support vector machines. A large number of comparative studies have been conducted in attempts to establish the relative superiority of these methods. This paper argues that these comparisons often fail to take into account important aspects of real problems, so that the apparent superiority of more sophisticated methods may be something of an illusion. In particular, simple methods typically yield performance almost as good as more sophisticated methods, to the extent that the difference in performance may be swamped by other sources of uncertainty that generally are not considered in the classical supervised classification paradigm.

Linear mixed models are able to handle an extraordinary range of complications in regression-type analyses. Their most common use is to account for within-subject correlation in longitudinal data analysis. They are also the standard vehicle for smoothing spatial count data. However, when treated in full generality, mixed models can also handle spline-type smoothing and closely approximate kriging. This allows for nonparametric regression models (e.g., additive models and varying coefficient models) to be handled within the mixed model framework. The key is to allow the random effects design matrix to have general structure; hence our label general design. For continuous response data, particularly when Gaussianity of the response is reasonably assumed, computation is now quite mature and supported by the R, SAS and S-PLUS packages. Such is not the case for binary and count responses, where generalized linear mixed models (GLMMs) are required, but are hindered by the presence of intractable multivariate integrals. Software known to us supports special cases of the GLMM (e.g., PROC NLMIXED in SAS or glmmML in R) or relies on the sometimes crude Laplace-type approximation of integrals (e.g., the SAS macro glimmix or glmmPQL in R). This paper describes the fitting of general design generalized linear mixed models. A Bayesian approach is taken and Markov chain Monte Carlo (MCMC) is used for estimation and inference. In this generalized setting, MCMC requires sampling from nonstandard distributions. In this article, we demonstrate that the MCMC package WinBUGS facilitates sound fitting of general design Bayesian generalized linear mixed models in practice.

Commonly used methods to analyze incomplete longitudinal clinical trial data include complete case analysis (CC) and last observation carried forward (LOCF). However, such methods rest on strong assumptions, including missing completely at random (MCAR) for CC and unchanging profile after dropout for LOCF. Such assumptions are too strong to generally hold. Over the last decades, a number of full longitudinal data analysis methods have become available, such as the linear mixed model for Gaussian outcomes, that are valid under the much weaker missing at random (MAR) assumption. Such a method is useful, even if the scientific question is in terms of a single time point, for example, the last planned measurement occasion, and it is generally consistent with the intention-to-treat principle. The validity of such a method rests on the use of maximum likelihood, under which the missing data mechanism is ignorable as soon as it is MAR. In this paper, we will focus on non-Gaussian outcomes, such as binary, categorical or count data. This setting is less straightforward since there is no unambiguous counterpart to the linear mixed model. We first provide an overview of the various modeling frameworks for non-Gaussian longitudinal data, and subsequently focus on generalized linear mixed-effects models, on the one hand, of which the parameters can be estimated using full likelihood, and on generalized estimating equations, on the other hand, which is a nonlikelihood method and hence requires a modification to be valid under MAR. We briefly comment on the position of models that assume missingness not at random and argue they are most useful to perform sensitivity analysis. Our developments are underscored using data from two studies. While the case studies feature binary outcomes, the methodology applies equally well to other discrete-data settings, hence the qualifier “discrete” in the title.

Andrei Kolmogorov’s Grundbegriffe der Wahrscheinlichkeits-rechnung put probability’s modern mathematical formalism in place. It also provided a philosophy of probability—an explanation of how the formalism can be connected to the world of experience. In this article, we examine the sources of these two aspects of the Grundbegriffe—the work of the earlier scholars whose ideas Kolmogorov synthesized.

Yuan Shih Chow was born in Hubei province in China, on September 1, 1924. The eldest child of a local militia and political leader, he grew up in war and turmoil. His hometown was on the front line during most of the Japanese invasion and occupation of China. When he was 16, Y. S. Chow journeyed, mostly on foot, to Chongqing (Chung-King), the wartime Chinese capital, to finish his high school education. When the Communist party gained power in China, Y. S. Chow had already followed his university job to Taiwan. In Taiwan, he taught mathematics as an assistant at National Taiwan University until he came to the United States in 1954. At the University of Illinois, he studied under J. L. Doob and received his Ph.D. in 1958. He served as a staff mathematician and adjunct faculty at the IBM Watson Research Laboratory and Columbia University from 1959 to 1962. He was a member of the Statistics Department at Purdue University from 1962 to 1968. From 1968 until his retirement in 1993, Y. S. Chow served as Professor of Mathematical Statistics at Columbia University. At different times, he was a visiting professor at the University of California at Berkeley, University of Heidelberg (Germany) and the National Central University, Taiwan. He served as Director of the Institute of Mathematics of Academia Sinica, Taiwan, and Director of the Center of Applied Statistics at Nankai University, Tianjin, China. He was instrumental in establishing the Institute of Statistics of Academia Sinica in Taiwan. He is currently Professor Emeritus at Columbia University.

Y. S. Chow is a fellow of the Institute of Mathematical Statistics, a member of the International Statistical Institute and a member of Taiwan’s Academia Sinica. He has numerous publications, including Great Expectations: The Theory of Optimal Stopping (1971), in collaboration with Herbert Robbins and David Siegmund, and Probability Theory (1978), in collaboration with Henry Teicher. Y. S. Chow has a strong interest in mathematics education. He taught high school mathematics for one year in 1947 and wrote a book on high school algebra in collaboration with J. H. Teng and M. L. Chu. In 1992, Y. S. Chow, together with I. S. Chang and W. C. Ho, established the Chinese Institute of Probability and Statistics in Taiwan.

This conversation took place in the fall of 2003 in Dobbs Ferry, New York.