Set static beta weight in GENLIN (a Monte Carlo study)

[J. R. Carroll]

We're using GENLIN to estimate a logistic regression model. We want to test the fit of a model where we input the beta weight for a particular predictor rather than having it estimated. We ran a model where we set the value of the beta for that predictor equal to 1, by designating it as the "offset" variable. Is this achieving our goal? It seems to be, but we wanted to run it by the list.

Thanks,

----


J. R. Carroll

Researcher for Hurtz Labs

Instructor at California State University, Sacramento

Research Methods, Test Development, and Statistics

www.jrcresearch.net

Cell:  [phone]

Email: [hidden email]

          [hidden email]

          [hidden email]

[Ryan]

Carroll,

It's uncommon to incorporate an offset into a logistic regression equation. Typically, an offset is used for Poisson or Negative Binomial regression equations to account for varying degrees of exposure.  To do so for a Poisson or Negative Binomial regression equation, typically the natural log of the original variable is entered as an "offset" into the equation. This makes sense given the log-link function which is used for Poisson and Negative Binomial models. More can be said on this topic, but that is not your question.

Anyway, back to your original question...If you enter a variable in its original form as an "offset" in logistic regression via GENLIN, then you are setting the beta coefficient to 1.0 as follows:

logit(y) = b0 + b1*x1 + b2*x2 + ... + 1.0*offset

Why you would want to set the regression coefficient to 1.0 for a variable in a logit equation is beyond me. Before treating a variable as an offset, at the very least I would start by entering the variable as a covariate to see whether the estimate of the coefficient is in fact near 1.0.

Ryan

On Tue, Sep 6, 2011 at 1:23 AM, J. R. Carroll <[hidden email]> wrote:

We're using GENLIN to estimate a logistic regression model. We want to test the fit of a model where we input the beta weight for a particular predictor rather than having it estimated. We ran a model where we set the value of the beta for that predictor equal to 1, by designating it as the "offset" variable. Is this achieving our goal? It seems to be, but we wanted to run it by the list.

Thanks,

----


J. R. Carroll

Researcher for Hurtz Labs

Instructor at California State University, Sacramento

Research Methods, Test Development, and Statistics

www.jrcresearch.net

Cell:  <a href="tel:%28916%29%20628-4204" value="[phone]" target="_blank">[phone]

Email: [hidden email]

          [hidden email]

          [hidden email]

[Ling Ting]

I am out of office and will return on 9/7/2011.  I will have very limited access to email. If you need immediate assistance please contact 479-575-2905.  Thank you.