|
Activity Number:
|
188
|
|
Type:
|
Topic Contributed
|
|
Date/Time:
|
Monday, August 7, 2006 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Section on Health Policy Statistics
|
| Abstract - #305417 |
|
Title:
|
Power of Tests for a Dichotomous Independent Variable Measured with Error
|
|
Author(s):
|
Daniel McCaffrey*+ and Marc Elliott
|
|
Companies:
|
RAND Corporation and RAND Corporation
|
|
Address:
|
201 N. Craig Street, Suite 202, Pittsburgh, PA, 15213,
|
|
Keywords:
|
maximum likelihood ; measurement error
|
|
Abstract:
|
We consider the power of three methods of testing differences in the means of groups defined by an unobserved dichotomous variable with known conditional probability p, given other predictors. The first method classifies observations into two groups by whether p exceeds a threshold. Method two tests the coefficient in a regression of the outcome on p. Method three estimates the difference in means by maximizing the outcomes' likelihood, given p. The efficiency of method one roughly scales with the square of one less the classification error. The efficiency of method two roughly scales with the R-square for predicting the unobserved dichotomous variable and is usually more powerful than method one. Method three is most powerful, but simulations show that for differences in means of 0.2 - 0.5 standard deviations, method two is about 90% as efficient as the MLE.
|
