Paige is indeed correct. However, it is also true that the correlation
between two variables can be affected by distributional form. Did you
look to see if the distribution of the covariate was the same within
each level of the categorical variable that the covariate interacted
with? I suggest this only to help you understand your results. I always
feel uncomfortable when results vary depending on a transformation and
so wish to understand why. In fact, the best plan might be to run the
model both ways (transformed and untransformed covariate), output the
residuals and examine them by the grouping variable.

Paul R. Swank, Ph.D. 
Professor, Developmental Pediatrics
Medical School
UT Health Science Center at Houston 



-----Original Message-----
From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED]
On Behalf Of Paige Miller
Sent: Tuesday, September 16, 2003 6:40 AM
To: [EMAIL PROTECTED]
Subject: Re: skewed covariate


emma wrote:
> Hi there,
> 
> Am new to posting - so let me know if haven't included enough info 
> etc.
> 
> I have conducted an ancova with two between subjects (group - 3
> levels) and education (2 levels) IV's and one covariate.  The DV has 
> been transformed using sq root.  When I run the analysis I have an 
> interaction between group and the covariate suggesting a lack of 
> homogeneity of regression slopes.  In trying to understand the 
> interaction (which was unexpected) - I ran some diagnostic tests on 
> the covariate which I found to be skewed. I decided to transform the 
> covariate (using sq root in line with the DV) - and this normalised 
> the distribution.  When re-running the ancova using both of the 
> transformed variables the interaction disappeared.  Is this valid?  Is

> it necessary to transform a covariate if:  1) it is non-normally 
> distributed and/or 2) to be consistent with the DV?  I should probably

> mention that although I have homogeneity of variance the sample sizes 
> are small and unequal (46/15/15).

The standard assumption in fitting Linear Models is that the DV errors 
are normally distributed, and NOT that the covariate itself is normally 
distributed. The covariate can have any distribution, as long as the DV 
  errors are normally distributed.

Transforming the IVs can indeed change the presense or absence of an 
interaction.

-- 
Paige Miller
Eastman Kodak Company
[EMAIL PROTECTED]
http://www.kodak.com

"It's nothing until I call it!" -- Bill Klem, NL Umpire
"When you get the choice to sit it out or dance, I hope you dance" -- 
Lee Ann Womack

.
. =================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to