IMS
Session Slot: 4:00- 5:50 Sunday
Estimated Audience Size: 150
AudioVisual Request: None
Session Title: Observational Studies
Theme Session: No
Applied Session: No
Session Organizer: Wasserman, Larry Carnegie Mellon
Address: Dept. of Stat, Carnegie Mellon
Phone: 412-268-8727
Fax: 412-268-7828
Email: larry@stat.cmu.edu
Session Timing: 110 minutes total (Sorry about format):
Opening Remarks by Chair - 0 minutes First Speaker - 30 minutes Second Speaker - 30 minutes Third Speaker - 30 minutes Discussant - 0 minutes Floor Discusion - 15 minutes
Session Chair: Wasserman, Larry Carnegie Mellon
Address: Dept. of Stat, Carnegie Mellon
Phone: 412-268-8727
Fax: 412-268-7828
Email: larry@stat.cmu.edu
1. Sensitivity Analysis in Observational Studies: Two Simple Examples
Rosenbaum, Paul, Univ. of Pennsylvania
Address: Paul R. Rosenbaum
The Wharton School
University Of Pennsylvania
Philadelphia, PA 19104-6302
Phone: 215-898-3120
Fax: 215-898-1280
Email: rosenbap@wharton.upenn.edu
Abstract: In an observational study, treatments are not randomly assigned to experimental subjects, so treated and control subjects may differ prior to treatment in ways that matter for the outcomes under study. Even when treated and control subjects are matched on recorded covariates, they may nonetheless differ in ways that have not been recorded. A sensitivity analysis asks how hidden biases of various magnitudes might alter the conclusions of a study. Two simple examples of sensitivity analysis will be presented, one from epidemiology and one from economics. In the economics example, the sensitivity of an instrumental variable estimator will also be examined.
2. Experiments, Randomized Trials, Observational Studies, and Causal Inference
Scheines, Richard, Carnegie Mellon University
Address: Dept. of Philosophy, CMU, Pittsburgh, PA 15213
Phone: 412-268-8571
Fax: 412-268-1440
Email: R.Scheines@andrew.cmu.edu
Abstract: Causal Inference from controlled experiments and large randomized clinical trials is relatively uncontroversial, whereas inference from uncontrolled observational studies is almost taboo. In this paper I use graphical causal models to show how the structure of inference is similar in all of these contexts. I show the assumptions needed to infer that a relationship is causal, how controlled experiments and randomized trials give us reason to endorse these assumptions, and how patterns of independence can do the same job in observational studies.
3. A New Approach to Observational Studies of Disease-specific Mortality
Joffe, Marshall, University of Pennsylvania
Address:
Department of Biostatistics and Epidemiology
602 Blockley Hall
423 Guardian Drive
University of Pennsylvania School of Medicine
Philadelphia, PA 19104-6021
Phone: 215- 573-7395
Fax: 215- 573-4865
Email: joffe@cceb.med.upenn.edu
Abstract: In observational research, investigation of possible causal links between a treatment or exposure and outcome often rely on subjects' self-report regarding exposure or treatment received; obtaining this information can be expensive. When the outcome event is rare, it is more efficient to obtain such self-report from all subjects who experience the event and a randomly chosen subset of the rest of the study cohort. Such case-control designs are problematic for mortality outcomes, because dead subjects cannot report information on previous exposures. For cause-specific mortality, a feasible hybrid design (Weiss and Lazovich, 1996) can often preserve much of the cost efficiency of case-control design. In this design, exposure is assessed at the time of disease diagnosis for all subjects diagnosed with the disease of interest, and for a subset of the remaining subjects. Unfortunately, analyzing data gathered this way as if it were standard case-control data can give biased results. When, as is usually the case with cancer outcomes, only a small proportion of potential subjects will be diagnosed with the condition of interest, an alternate method of analysis of such data is possible. This approach, which is based on formal definitions of the effect of a time-varying treatment and on the G-computation algorithm (Robins, 1986), may be applied even when the sampling fraction of non-diseased subjects is not known. We illustrate and compare conventional and new methods with data from a hypothetical study of breast cancer mortality.
List of speakers who are nonmembers: None