[R-sig-ME] Partial effects in mixed models

[Steven J. Pierce]

It is my understanding that each coefficient in a model with multiple predictors reflects the effect of that predictor conditional on the set of other predictors included in the model. Isn't that exactly what you're trying to obtain? 

If you want to explicitly model the effects of both predictors on the response and simultaneously model the correlation between those predictors, you could switch over to using a multilevel structural equation model. Mplus (a commercial software package) allows you to use Poisson response variables in such models. There may also be R packages that also allow such models, but I have not really looked to verify that.


Steven J. Pierce, Ph.D. 
Associate Director 
Center for Statistical Training & Consulting (CSTAT) 
Michigan State University 
E-mail: pierces1 at msu.edu 
Web: http://www.cstat.msu.edu 

-----Original Message-----
From: v_coudrain at voila.fr [mailto:v_coudrain at voila.fr] 
Sent: Friday, March 01, 2013 6:35 AM
To: Steven J. Pierce; r-sig-mixed-models at r-project.org
Subject: RE: [R-sig-ME] Partial effects in mixed models

Thank you. My concern was that the model with both variables within may not be optimal because both variables are correlated and I would like to know if the 
second variable has a "pure" effect on the response variable that is independent from the effect of the first variable. Since I have a generalized mixed model with 
poisson distribution, the statistics are based on Chi test and not F tests and I think that these tests are not sequential like in anova. Am I correct? 

Best


> Message du 01/03/13 à 03h23
> De : "Steven J. Pierce" 

> A : v_coudrain at voila.fr, r-sig-mixed-models at r-project.org
> Copie à : 
> Objet : RE: [R-sig-ME] Partial effects in mixed models
> 
> Why not just run a model with both predictors instead? See King (1986) for one perspective on why extracting the residuals to use as the dependent variable in 
another model is sub-optimal. That paper is about plain old OLS regression, but I suspect it still is applicable logic. 
> 
> King, G. (1986). How not to lie with statistics: Avoiding common mistakes in quantitative political science. American Journal of Political Science, 30(3), 666-687.
> 
> 
> Steven J. Pierce, Ph.D. 
> Associate Director 
> Center for Statistical Training & Consulting (CSTAT) 
> Michigan State University 
> E-mail: pierces1 at msu.edu 
> Web: http://www.cstat.msu.edu 
> 
> 
> -----Original Message-----
> From: v_coudrain at voila.fr [mailto:v_coudrain at voila.fr] 
> Sent: Thursday, February 28, 2013 11:25 AM
> To: r-sig-mixed-models at r-project.org
> Subject: [R-sig-ME] Partial effects in mixed models
> 
> Dear all,
> 
> I would like to test the effect of an explanatory variable after removing the effect of another one. I thought about calculating the model with the first explanatory 
> variable only, then take the model residuals and use the residuals as response variable to test the effect of the second explanatory variable. However, I do not 
> know if this is possible for a model containing random effects. Maybe it doesn't make sense anyway, but if it is possible, should I include the random effects in 
the 
> second model (residuals as response variable) or not, since variance explained by random effects should also have been accounted for in the first model?
> 
> Thank you for your help
> 
> Valérie
> ___________________________________________________________
> Alain Delon horrifié et dévasté de tristesse ! Mais pour quelle raison ? À lire sur Voila.fr http://people.voila.fr/people/actu-stars/tv-cine-musique/alain-delon-
horrifie-et-devaste-de-tristesse-repond-a-son-fils-people_9663.html
> 
> 
> 
> 

___________________________________________________________
News TV : Valérie Damidot annonce la fin de son émission D&Co ! À lire sur Voila.fr http://people.voila.fr/people/actu-stars/personnalites/valerie-damidot-annonce-la-fin-de-d-co-people_9679.html