s-news
[Top] [All Lists]

Logistic regression with random effects using glmmPQL

To: <s-news@lists.biostat.wustl.edu>
Subject: Logistic regression with random effects using glmmPQL
From: "Hunsicker, Lawrence" <lawrence-hunsicker@uiowa.edu>
Date: Mon, 3 Dec 2007 16:21:52 -0600
In-reply-to: <BF6EA021B2A77C46B86FF9112F4597217CE9CC@E2K3VS02.lsc.local>
Thread-index: Acd7gvwouOBaqbGlRRGrzmSl22+vGgAAbvDALpt61OA=
Thread-topic: [S] Logistic regression with random effects using glmmPQL

Good evening, all:

 

Back again.  I have now found, set up, and run glmmPQL.  I need some help in reading the output.  

 

I am trying to find whether there is meaningful residual center-to-center variability among dialysis centers in the odds that a patient will have a type of diagnostic test, corrected for the patients’ baseline characteristics.  The focus here is on whether there is real center-to-center variability, not on the baseline predictors.  I ran:

 

test.glmm <- glmmPQL(fixed = [two sided formula of fixed effects], random = ~1 | Center, family=binomial(link=logit), data="" dispersion = NULL,…)

 

This ran and converged.

 

The output included a lot of stuff that I understand, but also the following:

 

Random effects:

  Formula:  ~1 | Center

                        (Intercept)          (Residual)

StdDev:             0.4090162         0.9714077

 

This is the part that I don’t quite get.  

 

a.  The intercept, I assume, is the StdDev of the random effect.  How do I test whether the StdDev is significantly different from 0?  What is its scale?  (The naïve median of the chance of the test at each center is about 19%, with a range from 0 to 100%.)  

 

b:  What is the Residual?  Is this on the same scale as the Intercept?  I.e., can one say that 0.4090162/(0.4090162+0.9714077) of the variability is “explained” by including the random effect?  Does the fact that the Residual is very close to 1.0 mean that there is essentially no over dispersion once I include the random Center effect?

 

c.  Should I be testing the significance of the random effect by using a likelihood ratio test comparing with the same model above, but setting: 

            random = ~ 1 (without the “| Center”)?  This gives me a huge difference in log likelihood. 

 

d.  Is there somewhere that I can read up to understand this function?

 

Many thanks in advance for any help you can give me.

 

Larry Hunsicker

 

 

 


From: Jimenez-Leal William [mailto:william.jimenezleal@lsc.gov.uk]
Sent: Tuesday, April 10, 2007 10:32 AM
To: Hunsicker, Lawrence; s-news@lists.biostat.wustl.edu
Subject: RE: [S] Logistic regression with random effects

 

Did you try glmmPQL() ? It actually works by calling the lme.

 

William

 

 


From: s-news-owner@lists.biostat.wustl.edu [mailto:s-news-owner@lists.biostat.wustl.edu] On Behalf Of Hunsicker, Lawrence
Sent: 10 April 2007 16:15
To: s-news@lists.biostat.wustl.edu
Subject: [S] Logistic regression with random effects

 

Good morning, all:

I suppose that this must be the thousandth time someone has asked this question, and I apologize that I don’t know how to look up past questions and answers.  I am trying to study whether there is a significant difference in outcomes among centers providing a kind of medical service.  The outcome is binary.  There are a batch of individually varying covariates of importance, but the real focus is on whether, after correcting for these covariates, there is a meaningful variability among centers in outcome.  I am not interested in the specific centers, but rather in the overall distribution of underlying center effect.  It would seem that the appropriate statistical method is logistic regression with a random center effect.  I can do this with Egret, but I am wondering whether it is possible to do this in S-Plus.  In S-Plus we have lme for linear random effects, and we have glm for estimating logistic regression.  But is there something that combines these two?

As always, thanks in advance to anyone that can help me.

Larry Hunsicker

Confidentiality and Disclaimer: This email and its attachments are intended for the addressee
only and may be confidential or the subject of legal privilege.
If this email and its attachments have come to you in error you must take no action based on
them, nor must you copy them, distribute them or show them to anyone.
Please contact the sender to notify them of the error.
This email and any attached files have been scanned for the presence of computer viruses.
However, you are advised that you open any attachments at your own risk.
Please note that electronic mail may be monitored in accordance with the
Telecommunications (Lawful Business Practices)(Interception of Communications) Regulations
2000.
<Prev in Thread] Current Thread [Next in Thread>
  • Logistic regression with random effects using glmmPQL, Hunsicker, Lawrence <=