Skip to content

Ancova and lme use

2 messages · a.menicacci@fr.fournierpharma.com, Spencer Graves

a.menicacci@fr.fournierpharma.com
# 
Dear R-users,

We expect to develop statistic procedures and environnement for the
computational analysis of our experimental datas. To provide a proof of
concept, we plan to implement a test for a given experiment.

Its design split data into 10 groups (including a control one) with 2
mesures for each (ref at t0 and response at t1). We aim to compare each
group response with control response (group 1) using a multiple comparison
procedure (Dunnett test).

Before achieving this, we have to normalize our data : response values
cannot be compared if base line isn't corrected. Covariance analysis seems
to represent the best way to do this. But how to perform this by using R ?

Actually, we have identify some R functions of interest regarding this
matter (lme(), lm() and glm()).

For example we plan to do as describe :
glm(response~baseline) and then simtest(response_corrected~group,
type="Dunnett", ttype="logical")
If a mixed model seems to better fit our experiment, we have some problems
on using the lme function : lme(response~baseline) returns an error
("Invalid formula for groups").

So :
Are fitted values represent the corrected response ?
Is it relevant to perform these tests in our design ?
And how to use lme in a glm like way ?

If someone could bring us your its precious knowledge to validate our
analytical protocol and to express its point of view on implementation
strategy ?

Best regards.


Alexandre MENICACCI
Bioinformatics - FOURNIER PHARMA
50, rue de Dijon - 21121 Daix - FRANCE
a.menicacci at fr.fournierpharma.com
t??l : 03.80.44.76.17
5 days later
Spencer Graves
# 
Mon cher M. MENICACCI:

	  It looks to me like you ultimately want to use "lmer" in 
library(lme4) [which also requires library(Matrix)].  For documentation, 
I suggest you start with Doug Bates (2005) "Fitting Linear Mixed Models 
in R", R News, vol. 5/1: 27-30 (available from "www.r-project.org" -> 
Newsletter).  After install.packages("lme4"), I suggest you read 
"Implementation.pdf" in "~R\library\lme4\doc".  I also suggest you get 
Pinheiro and Bates (2000) Mixed-Effects Models in S and S-Plus 
(Springer).  For me, this book was essential documentation for "lme", 
the previous implementation of "lmer".  Studying that book might help 
you understand "lmer".

	  Also, I encourage you to use the extensive random number generation 
capabilities in R (including the nlme and lme4 packages) to produce 
simulated data like you expect to collect and try to analyze the 
simulated data.  You should simulate both what you expect to see and the 
null hypothesis as well.  If you encounter difficulties doing that, 
please submit another question to this listserve.  Before submitting 
another post, I suggest you help yourself by reading the posting guide! 
"www.R-project.org/posting-guide.html".  Anecdotal evidence suggests 
that posts that are more consistent with this "posting guide" generally 
get more useful replies quicker.

	  bon chance.
	  spencer graves
a.menicacci at fr.fournierpharma.com wrote: