September 1998-May 1999 Meetings
Wednesday, September 16, 1998
Wanzhu Tu, Ph.D., IU Biostatistics
Pairwise comparison of the means of independent log-normal populations
Two bootstrap procedures for comparison of the log-normal means were proposed to approximate the sampling distribution of the pivotal statistic generating confidence sets for pairwise differences. Operating characteristics of the proposed methods were compared with the widely used Bonferroni pairwise procedure in a simulation study. A large randomized clinical trial for assessing the effects of prospective drug utilization (DUR) intervention on health care charges was analyzed as an illustration.
Wednesday, October 21, 1998
Craig Mallinckrodt, Ph.D., Eli Lilly & Co.
Accounting for dropout bias using mixed-models
Early discontinuation (dropout) of patients in controlled clinical trials can make valid analyses of longitudinal data problematic. Dropouts reduce information, pose technical difficulties, and can bias estimates of treatment effects, especially if the reason for dropout is related to outcome. Analyzing data from only those patients that completed the study (completers analysis) and last observations carried forward to endpoint (LOCF) analyses can yield biased estimates of treatment effects. Recent advances in software and methods have made general linear mixed-models easily available for a wide variety of applications. The objective of this study was to assess and compare LOCF ANOVA and mixed-model repeated measures (MMRM) ANOVA of longitudinal data with dropout bias. Estimates of treatment group differences in mean change from baseline to endpoint from the two methods were compared in data simulated under several scenarios and in data from a randomized clinical trial. Estimates of treatment effects from MMRM were, on average, markedly closer to true values than estimates from LOCF in every scenario simulated. Standard errors and confidence intervals from MMRM accurately reflected the uncertainty of the estimates, whereas standard errors and confidence intervals from LOCF underestimated uncertainty.
Wednesday, November 18, 1998
Sujuan Gao, Ph.D., Division of Biostatistics, I.U. School of Medicine.
Estimating the Incidence of Dementia from Two-Phase Sampling with Nonignorable Missing Data
In a longitudinal study of dementia, the repeated clinical diagnosis of the disease is conducted several years after the baseline wave. Therefore, some study subjects may die before the follow up wave. Hence disease status prior to death for these subject is missing in the follow up. There are reasons to suggest that the missing due to death is nonignorable. Estimation of disease incidence from longitudinal dementia study has to appropriately adjust for data missing due to death as well as the sampling design used at each study wave. In this talk we adopt a selection model approach to model the missing data by death and use a likelihood based approach to derive incidence estimates. A modified EM algorithm is used to deal with data from sampling selection. The non-parametric jackknife variance estimator is used to obtain variance estimates for the model parameters and the incidence estimates. The proposed approaches are applied to data from the Indianapolis/Ibadan Dementia study.
Wednesday, January 20, 1999
Viswanath Devanarayan, Ph.D., Eli Lilly & Co.
Resampling-Like Methods for Handling Covariate Measurement Error in Generalized Linear and Nonlinear Models
Cook & Stefanski(1994, JASA) introduced a novel, computer-intensive estimation procedure for measurement error models, called SIMEX, for simulation and extrapolation. We refer to this method as parametric SIMEX because data with additional amounts of measurement error (pseudo data) are generated via simulation from a parametric family of distributions. We propose a different version of the SIMEX method, called empirical SIMEX, where the pseudo data are generated without the aid of a parametric model. The empirical SIMEX method automatically handles the inherent homogeneity or heterogeneity of the measurement error variances and therefore does not require any assumptions on the measurement error variance structure. We call the parametric and empirical SIMEX methods remeasurement methods; they are similar in flavor to resampling methods such as the jackknife and bootstrap.
In his talk, Devan gave a brief introduction to measurement error models and described the SIMEX algorithm. He then outlined a few important results and illustrated the application of these methods to data from the Framingham Heart Study. Finally, he presented an interesting perspective on the jackknife and bootstrap which elucidated the link between resampling and remeasurement methods.
Wednesday, February 17, 1999
Clement J. McDonald, M.D., Regenstrief Institute
The Regenstrief Medical Record System
The mission of the Regenstrief Institute for Health Care is to conduct research aimed at improving health care by improving the capture, analysis, content, and delivery of the information needed by patients, their health care providers, and policy makers. This includes carrying out intervention studies designed to measure the effect of the application of this research on the efficiency and quality of health care. To a great extent, achieving these goals involves or depends upon the Regenstrief Medical Record System (RMRS), a system that consists of millions of observations and links to other sources of clinical knowledge. Physicians actively interact with the RMRS when making decisions about the best course of patient care. In this presentation, I discuss the structure and use of the RMRS, and describe some of the challenges involved in the development and expansion of the system city-wide. This includes issues such as the standardization of medical terms, security, and confidentiality issues.
Wednesday, March 17, 1999
Barry Katz, Ph.D. IU Biostatistics,
Chairman of Local Assistance Committee
Central Indiana Chapter Planning for JSM 2000
For JSM 2000, we will need to staff two main committees, a Local Assistance Committee (LAC) chaired by Barry Katz and a Program Committee (PC) chaired by Bob Obenchain. These committees work primarily with the national ASA Meetings staff and the JSM 2000 Program Committee, respectively. Our LAC responsibilities will include such things as providing articles for the AMSTAT News, manning information booths at both JSM1999 and JSM2000, and organizing information for local events during the meetings. The Program Committee organizes one invited session and possibly special contributed session(s) for the JSM2000 meetings. We discussed the general organization of and asked for volunteers for both committees.
Wednesday, April 21, 1999
Siu Hui, Ph.D. IU Biostatistics
Spacing of Follow-up Waves in Incidence Studies
Longitudinal studies of dementia are often undertaken to estimate the incidence of dementia and to identify incident cases for the study of risk factors measured on the entire cohort at baseline. The power of these analyses is determined primarily by the number of demented cases identified. Increasing the number of waves of evaluation of the cohort in a given time period increases the yield of demented cases, but it also increases the cost of the study. This work provides a method for estimating dementia incidence in the presence of mortality and loss to follow-up. It also assesses the trade-off between the cost (as measured by the number of reevaluations) and the yield of incident dementia cases so that longitudinal studies of dementia can be designed optimally in the future.
Wednesday, May 19, 1999
Chunming Li, Ph.D. - IU Biostatistics
A Sequential Classification Model with Applications to Change Detection and Imaging
In object classification, the object can be any kind of entity or description (e.g. signals or images), and is represented by its attributes, which can be abstract features or elementary primitives, either numeric or symbolic. Active testing provides an efficient method for approximating standard classifiers (MLE, MAP, etc.) in problems (very high-dimensional or costly, for example) in which computing those is impossible or wasteful. In this context, testing refers to the process of evaluating the attributes or say, the realization of the features. I will discuss issues regarding active testing strategies studied by researchers in such fields as machine learning and computer vision. A new measure for testing strategies will be introduced. We investigate adaptive strategies for sequential testing, especially those driven by maximizing information gain when the conditional distribution of features given states of nature is Gaussian. We implement a classification algorithm in which tests are selected recursively and adaptively on-line. We show that such information-based strategies are statistically sensible and computationally efficient, and accommodate testing at multiple resolutions. Applications are made to change point detection and medical image classification.