|
Activity Number:
|
136
|
|
Type:
|
Invited
|
|
Date/Time:
|
Monday, August 3, 2009 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Biopharmaceutical Section
|
| Abstract - #303117 |
|
Title:
|
Causal Modeling When the Treatment Is a Latent Class
|
|
Author(s):
|
Joseph L. Schafer*+ and Joseph Kang
|
|
Companies:
|
Penn State University and Northwestern University
|
|
Address:
|
The Methodology Center, State College, PA, 16801,
|
|
Keywords:
|
expected estimating equations, missing data
|
|
Abstract:
|
In the potential-outcomes approach to causal inference, causal effects are differences among outcomes that would be realized if different treatments were applied to the same individual. The treatment is usually assumed to be a binary variable measured without error. In many settings, however, the treatment is measured imperfectly by multiple questionnaire items, and failure to account for measurement error may bias the estimated effects. We present models in which the treatment is a latent class. Treatment propensities and potential outcomes are regressed on a rich set of confounders and prognostic variables. After estimating parameters by EM, we estimate average causal effects within each treatment group by solving expected estimating equations. We apply our model to estimate the effects of naturalistic weight-loss strategies on body-mass index among adolescent girls.
|
