On 21 Apr 2002 09:04:03 -0700, [EMAIL PROTECTED] (wuzzy) wrote:

 ... 
> Maybe when using the variable as a covariate it is good to
> dichotomize?
> 
> Say you wanted to adjust for energy (in METs) expended during
> work-related activity. Is it better to adjust for the continuous
> variable, energy, or is it better to adjust for quartiles of energy.
> (ie. Q1=office-type Q4=fireman)

How bad is your variable scored?  Does it need to be
transformed for it to have linear effects on outcome?

If you have enough degrees of freedom, you *can*  
do things with covariates, for *control*  purposes,  that 
you would not do with something you are testing.
For instance, you can control by "narrow age group"
by having a factor with each age, by year.  That is 
exactly equivalent to having dummy variables for
covariates, for each year.  But the linear of effect for
age is usually 90% or more of what is there, and it 
is both easier to look at analyses, and to explain them,
if you don't have 40 extra covariates.  Having a second
variable for Age might do most of the rest -- either
providing a quadratic fit, or a spline (essentially, use
two regression lines).

> 
> the quartiles(dichotomized) might cause a loss of information but it
> might "adjust" for the variable better by filtering out groups of
> people?  am I wrong?

IF there are real groups, your best control is to have 
covariates that match/ describe the real groups
(with large enough N that you are screwing up your
analysis by capitalizing on chance).  "Effect coding"
is what it has been called, when you put in a covariate
that *uses*  the observed mean as its score:  This 
implies that a correction will be needed to adjust the 
residual d.f., according to how many 'means' have 
been employed.  - I don't know that anyone does this
in analyses these days, since I ran into the technique
in an old book that was advising how to perform 
repeated measures ANOVA with a little missing data,  
especially when using small-capacity computers.

Hope this helps.
-- 
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
.
.
=================================================================
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