Statalist


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

RE: st: Multilevel modelling of survival data


From   Maarten buis <[email protected]>
To   [email protected]
Subject   RE: st: Multilevel modelling of survival data
Date   Mon, 16 Mar 2009 20:39:28 +0000 (GMT)


--- On Mon, 16/3/09, Justin B Echouffo Tcheugui wrote:
> I have smoking as one of my outcome in a cluster
> randomised trial. The unit of randomisation are clinics.
> I want to explore the difference between groups in term
> of smoking at follow-up 
>
> 1- I first used logistic
> regression using the following commands to derive the 
> odds ratio -gllamm- and -xtmelogit-
> 
> Is there any advantage of using one or the other command? 

-xtmelogit- is quicker, and since you are only estimating
a random intercept model you could also use -xtlogit-, which
is even quicker.
 
> 2- Later on, I was advised to analyse smoking as a
> continuous variable in a linear fashion expressing my
> result as the adjusted difference in proportion of smokers
> between the groups. 
>
> Is there reason why one should prefer -xtmixed- instead of
> xtmelogit for a binary variable? 

I am no fan of using linear models for categorical data, and
that is an understatement. The arguments that I have heard 
in their favor come roughly in two flovours:

1) "I (or my readers) don't know how to interpret odds ratios"
The answer is: it is not that hard, the odds is how many 
successes for every failure, and is a measure of the 
likelihood of success. The odds ratio is how many times 
larger this odds of success is for group 2 relative to group 1.

2) "The results of non-linear models like -logit- are a biased
biased estimate of the causal effect, even when analyzing data
obtained in a randomized experiment"
The answer is: true, but the same is also true for the linear 
model when applied to a categorical dependent variable. 

So, I just would not bother with -xtmixed-, and stick with
-xtlogit-.
 
> Do I need absolutely need to correct the confidence
> intervals given by the linear model? 

Probably, the residual error is not going to be homoskedastic.

> If yes, is the following command the right one to use?

I have no idea, primarily because I would not use the linear
model in this case anyhow.

-- Maarten

-----------------------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany

http://home.fsw.vu.nl/m.buis/
-----------------------------------------




      

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/



© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index