American Statistical Association
Start with a response Y and predictors X, S and Z. In a general measurement error model, one of the variables is always missing, say X. Instead of observing X, we observe a guess at X, call it W. In a semiparametric measurement error model, one of the variables, say Z, is meant to affect the response via an unknown function, \theta(Z). Thus, for example, the partially linear measurement error model states that the mean of Y is \beta_1X+\beta_2S+\theta(Z), and instead of observing X we observe W. Alternatively, X is modeled semiparametrically, so the partially linear model takes the deconvolution form: \beta_1Z+\beta_2S+\theta(X).
There has been considerable work recently on semiparametric measurement error models, much of the deconvolution work motivated by the Nevada Test Site Thyroid Disease Study. I will review this work, and in passing mention how to actually compute standard errors in SIMEX for parametric and nonparametric parts.
|Date:||Thursday, November 29, 2007|
|Time:||4:00 - 5:00 P.M.|
Columbia Presbyterian Medical Center
Humphreys Auditorium - 14th Floor, Room 240
622 West 168th Street
(between Broadway and Fort Washington Avenue)
New York, New York