Sponsoring Section/Society: ENAR
Session Slot: 10:30-12:20 Tuesday
Estimated Audience Size: 125-175
AudioVisual Request: Two Overheads
Session Title: Statistical Methods for Nonignorable Missing Data
Description of the session: Policy makers rely on accurate analysis
of data to make appropriate decisions. However, when we
report statistical results to policy makers, we are seldom
telling them what the assumptions behind the results are and what
the effects of violating these assumptions on the conclusions are. One
example is statistical analysis of missing data. Most
commonly used methods for analyzing missing-data require the
missing-at-random (or ignorable) assumption. However, if the
missing-data mechanism is non-ignorable, what are the effects
of assuming ignorable missing-data mechanism
on the conclusions and what are some
alternative approaches when the missing-data mechanism is
non-ignorable? In this session, we will have three talks. The
first talk, by Dr. Rotnitzky, will focus on semiparametric
sensitivity analysis with non-ignorable non-response data; the
second talk, by Dr. Baker, will focus on The design and
analysis of a randomized trial with a missing binary outcome:
the role of an auxiliary variable; the third talk, by
Dr. Zhou, will focus on statistical methods for evaluation of
diagnostic accuracy when verification is subject to
non-ignorable missing-data mechanism. The discussant in the
session will be Dr. Daniel Heitjan at Columbia University.
Theme Session: No
Applied Session: No
Session Organizer: Zhou, Xiao-Hua (Andrew) Indiana University School of Medicine
Address: Division of Biostatistics Indiana University School of Medicine Riley Research Wing, 135 699 West Drive Indianapolis IN 46202-5119
Phone: 317-274-2696
Fax: 317-274-2678
Email: zhou@mako.biostat.iupui.edu
Session Timing: 110 minutes total (Sorry about format):
Opening Remarks by Chair - 5 minutes First Speaker - 30 minutes Second Speaker - 30 minutes Third Speaker - 30 minutes Discussant - 10 minutes Floor Discussion - 5 minutes
Session Chair: Rubin, Donald B. Harvard University
Address: Department of Statistics Harvard University 1 Oxford Street Cambridge MA 02138-2901
Phone: 617-495-5498
Fax: 617-496-8057
Email: rubin@stat.harvard.edu
1. A Note About Multiple Imputation
Wang, Naisyin, Texas A&M University
Address: Department of Statistics Texas A&M University College Station TX 77843-3143
Phone: 409-845-3141
Fax: ( 409-845-3144
Email: nwang@stat.tamu.edu
Robins, Jamie, Harvard University
Abstract: We consider the asymptotic behaviour of various parametric multiple imputation procedures. The asymptotic variance structure of the resulting estimators is provided. This result is used to compare the relative efficiencies of different imputation procedures and it applies to non-ignorabl missing mechanism provided that certain conditions are fullfilled. Some examples are used to illustrate the results.
2. The Design and Analysis of a Cancer Prevention Trial: Adjusting for a Missing Binary Outcome, an Auxiliary Variable, and All-or-None Compliance
Baker, Stuart G., National Cancer Institute
Address: Biometry Branch Division of Cancer Prevention National Cancer Institute EPN 344 9000 Rockville Pike Bethesda MD 20892
Phone: 301-496-7708
Fax: 301-402-0816
Email: sb16i@nih.gov
Abstract: The National Cancer Institute is recruiting subjects into a randomized trial of daily finasteride versus placebo. The outcome is prostate cancer incidence determined by biopsy, either after seven years or following an elevated PSA (prostate-specific antigen) on yearly screening. There are two complications in the analysis. First, missing a biopsy will likely depend on whether or not a subject has an elevated PSA. The indicator of elevated PSA is called an auxiliary variable because it is observed after randomization and prior to endpoint. Second, there will likely be contamination in the placebo arm and nonadherence in the finasteride arm, starting soon after randomization. Compliance in this situation is called all-or-none. To adjust for these complications we formulate the appropriate likelihoods and obtain closed-form maximum likelihood estimates and variances. Some of these likelihoods involve nonignorable missing-data mechanisms.
3. Statistical Methods for Evaluation of Diagnostic Accuracy When Verification is Subject to Non-Ignorable Missing Data Mechanism
Zhou, Xiao-Hua (Andrew), Indiana University School of Medicine
Address: Division of Biostatistics Indiana University School of Medicine Riley Research Wing, 135 699 West Drive Indianapolis IN 46202-5119
Phone: 317-274-2696
Fax: 317-274-2678
Email: zhou@mako.biostat.iupui.edu
Abstract: The accuracy of a diagnostic test is often measured by its sensitivity, specificity, positive predictive and negative predictive values. More generally, a receiver operating characteristic (ROC) curve may be used to represent the accuracy of a diagnostic test. To calculate these measures, we need to determine the disease status of each patient in the sample. The procedure that establishes the patient's disease status is referred to as a gold standard. However, for some studies, only a subset of the patients with diagnostic test results are chosen to receive the gold standard assessment. For example, in some clinical circumstances, the disease status is obtained by surgery or biopsy, usually when a diagnostic test has provided strong evidence of the presence of disease. If the study population consists of only verified cases, the estimated accuracy of the diagnostic test may be biased. This type of bias is called verification bias, and may be treated as a missing-data problem. In this talk, I will first review available verification bias correction methods under the ignorable assumption. Then, I will discuss the the effects of non-ignorable verification bias on estimates, derived under the ignorable assumptions, and discuss a test procedure for the ignorable assumption. Finally, I will illustrate the methods with datasets from several clinical studies.
Discussant: Heitjan, Daniel Columbia University
Address: Division of Biostatistics Columbia University School of Public Health New York NY 10032-3702
Phone: 212-305-9557
Fax: 212-305-9408
Email: dfh5@columbia.edu
List of speakers who are nonmembers: None