[R] Log transformed values and contrasts in LME
Francisco J. Zagmutt Vergara
fzagmutt at hotmail.com
Tue Oct 7 00:16:33 CEST 2003
Dear All
This is probably a very basic question for this list but I just wanted to
make sure that I am doing things right:
I have an LME model with 4 categorical variables and 2 continuous variables
(analysis of covariance model). I had to use a log transformation on the
data to achieve normality (log(x)-.1) and then I used contrast treatment to
compare differences between a baseline level and the other level of the same
categorical variable (as far as I understand R picks automatically the level
with the smaller marginal mean to make these comparisons). When I want to
interpret my coefficients in terms of the original (non-transformed data) is
not as simple as using:
>exp(beta)-.1
since (I believe) the hypotheses testing with the contrasts is
log(Beta1)-Log(beta2)=0
So I though that a way to go around this is to remove the intercept from the
original model to get a "cell-means" model which would basically give me
the average log transformed value of the outcome variable for each category:
log(beta1) +log(beta2)+...log(betan)
and then transform those values to the original data form and subtract the
means to obtain an estimate of the difference between the means that I
tested with the contrast:
>b1<-exp(beta) -.1
>b2<-exp(beta)-.1
>b1-b2
Is this conceptually right? I would not be making any hypothesis testing
with the new model, just getting an estimate of the actual difference
between the level means so I think that his could be a valid approach. I
would still report the test statistics and significance values for the
contrasts from the original model but just would include the estimation of
the means from the second model. Am I in the right path?
Many thanks for your help!!
Francisco
_________________________________________________________________
¿Estás buscando un auto nuevo? http://www.yupimsn.com/autos/
More information about the R-help
mailing list