Alan Hubbard

Assistant Professor of Biostatistics

School of Public Health
University of California, Berkeley
140 Earl Warren Hall, #7360
Berkeley, CA 94720-7360

PHONE: 510-643-6160
FAX: 510-643-5163
OFFICE: 113B Haviland
EMAIL: hubbard@stat.berkeley.edu

Teaching

PH 242C: Longitudinal Data Analysis

This course covers the statistical issues surrounding estimation of effects using data on subjects followed through time. The course emphasizes a regression model approach and discusses disease incidence modeling and both continuous outcome data/linear models and longitudinal extentions to nonlinear models (e.g., logistic and Poisson). The primary focus is from the analysis side, but we will also discuss the mathematical intuition behind the procedures. The statistical/mathematical material include some survival analysis, normal linear models, logistic and Poisson regression and matrix algebra for statistics. Next taught in Spring, 2005.

Causal Inference

The course covers both the basic issues regarding the estimation of causal effects using observational data and also specific, recently developed models designed to estimate such effects. Topics to be discussed include confounding, counterfactuals, graphical models, direct and indirect effects, the G-computation algorithm, propensity scores and marginal structural models for both point treatment and longitudinal studies. Time permitting, additional topics include instrumental variables, dynamic treatment regimes, structural nested models and structural equation models.

Research

Clustering Functions

This research has revolved around the apparently simple question: How many different kinds of patients does many data set contain? It was motivated by a data set from San Francisco General Hospital (SFGH) on several hundred HIV subjects followed after initialization of HAART. Subjects were followed irregularly over time and both CD4 counts and viral loads were recorded. The basic method involves an ad hoc part (smoothing and prediction at grid points, clustering) and a rigorous part (choosing the parameters at each step by cross-validation). The result is a set of clusters defined by the longitudinal profiles of patients.

Dynamic Models of Infectious Disease

More of my work has been focused on infectious diseases and the unique statistical issues that arise when outcome data among subjects is inherently related (correlated). Part of the work involves using mathematical infectious disease models to investigate the potential bias of ignoring the feedback inherent in infectious diseases. (Eisenberg, et al., 2003)

In addition, a recently submitted paper on analyzing the different contributions (person-to-person, person-to-environment-to-person) to the Cryptosporidium outbreak in Milwaukee, we used a novel technique to find the posterior distribution (the estimation distribution) of the relevant parameters in the model. This involved a combination of profile likelihood methods and a modified MCMC algorithm. (similar to Hubbard, et al., 2002)

Risk Assessment

With Prof. Mark Nicas on assessing risk from respiratory infections, also incorporating previous work on dose-response. This work is inspired by characterizing risk of infection (and the efficacy of preventive measures) from bioterrorism or infection of hospital workers in an outbreak. (Nicas and Hubbard, 2002 and Nicas and Hubbard, 2003)

Computational Biology

I have recently completed an initial analysis on Affymetrix data and workers exposed to benzene. The data (from Prof. Martyn Smith’s lab) consists of 40,000+ gene expressions measured on 12 workers (6 exposed and 6 unexposed matched pairs) in

China

. In addition, we are examining a very similar data set on dioxin exposure.

Locally Efficient Estimation

Work on (treatment specific) locally efficient estimation in the presence of potentially informative censoring and confounding. (van der Laan, Hubbard and Robins, 2002, Hubbard, et al., 2000, and van der Laan and Hubbard, 1998)

Other Activities

Links

Curiculum vitae

Group in Biostatistics