Bookmark and Share

Notice: On March 31, it was announced that Statalist is moving from an email list to a forum. The old list will shut down on April 23, and its replacement, is already up and running.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: simulation to derive power for GLM multiple regression

From   Austin Nichols <>
Subject   Re: st: simulation to derive power for GLM multiple regression
Date   Tue, 9 Mar 2010 11:36:18 -0500

Jake <> :
If you have the data on explanatory variables, you need only specify
true coefs and draw error terms.  This involves far fewer choice than
generating all the  explanatory variables, with various possible joint
distributions.  But you still have to make choices about what
combinations of parameter values to test, and what family of
distributions to draw errors from.  Or you can use your estimated
parameter values in the current dataset, and perhaps twice and half
each, and draw errors from the empirical distribution of errors in
your estimated model.  If you have a negative binomial, perhaps you'd
like to specify errors as being multiplicative, so y=f(X.b)e and the
estimated errors are y/f(X.bhat) which you can then sample from in
your simulation.

Is this a post-hoc power analysis?

On Tue, Mar 9, 2010 at 10:51 AM, Jake <> wrote:
> Hello,
> I am interested in learning details about how to conduct simulations to calculate power for a test of a single coefficient in a GLM (negative binomial) multiple regression model.  I am somewhat familiar with the relevant methods outlined by Feiveson (2002).  However, what I am curious about which Feiveson does not discuss is how best to simulate when there are k covariates, in addition to the coefficient of interest.
> I assume I would treat the covariates as fixed -- using the real (already collected) data in the simulation, in other words.  Then I imagine that I would simulate only the variable for the coefficient of interest.  I'm not certain of how to generate this random variable so that it would be (asymptotically) correlated with all of the other variables in specified ways.  I imagine I would use the cholesky method?  But many of the variables are not normally distributed -- some are dichotomous, etc.
> I would appreciate any help that you might provide.
> Thanks,
> Jacob Felson
> Assistant Professor
> Dept. of Sociology
> William Paterson University
> Reference
> Feiveson, A.H. 2002 "Power by simulation."  The Stata Journal 2: 107-124.

*   For searches and help try:

© Copyright 1996–2016 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   Site index