Categorical and continuous predictors in GLM
Adrienne
I am using a GLM to look at the effects of four predictor variables on
multiple DVs. Three of my predictors are categorical, and one is
continuous. I don't want to dichotomize the continuous variable
(hence the GLM), but for some reason I can't figure out how to enter
it as a continuous predictor using SPSS. The only place to enter
continuous IVs in the GLM menu seems to be under "covariates" - and if
I do that, wouldn't I just be looking at the effects of the
categorical variables above and beyond the effect of the continuous
variable? Is there any way I can look at interactions among all four
variables while maintaining the continuous variable as continuous?
Thank you!
-Adrienne
Bruce Weaver
> Hi there,
>
> I am using a GLM to look at the effects of four predictor variables on
> multiple DVs. Three of my predictors are categorical, and one is
> continuous. I don't want to dichotomize the continuous variable
Good! Doing so would discard information needlessly.
> (hence the GLM), but for some reason I can't figure out how to enter
> it as a continuous predictor using SPSS. The only place to enter
> continuous IVs in the GLM menu seems to be under "covariates" - and if
> I do that, wouldn't I just be looking at the effects of the
> categorical variables above and beyond the effect of the continuous
> variable? Is there any way I can look at interactions among all four
> variables while maintaining the continuous variable as continuous?
>
> Thank you!
> -Adrienne
Enter your continuous variable as a covariate. The output should
include an F-test for the covariate alone, but also F-tests for its
interactions with other variables.
ONE WORD OF CAUTION: You should center your continuous predictor
variable on its mean (or some other meaningful in-range value) first.
For details, see:
GERARD J.P. VAN BREUKELEN and KOENE R.A. VAN DIJK. Use of covariates
in randomized controlled trials. Journal of the International
Neuropsychological Society (2007), 13, 903–904.
If you cannot get access through your library, drop me a line off-
list, and I'll send you a copy.
HTH.
--
Bruce Weaver
bwe...@lakeheadu.ca
http://sites.google.com/a/lakeheadu.ca/bweaver/Home
"When all else fails, RTFM."
Adrienne
interpreting the results incorrectly. My understanding was that when
a variable was entered as a "covariate", the model first removed all
variance in the DV due to that covariate variable, then looked at the
effects of the "fixed factor"predictors on the remaining variance.
This was what I wanted to avoid... but it sounds like you're saying
that's not the case?
Thanks for your clarification!
Rich Ulrich
<adrienn...@gmail.com> wrote:
>Thank you Bruce, that is helpful - I think I must have been
>interpreting the results incorrectly. My understanding was that when
>a variable was entered as a "covariate", the model first removed all
>variance in the DV due to that covariate variable, then looked at the
>effects of the "fixed factor"predictors on the remaining variance.
The ordinary method of ANOVA or regression is to look at
each of the variables after removing the effects of the
others - categorical or continuous. It is usually possible
to specify an evaluation in a particular order... and you
are right to be concerned about whether something labeled
"covariate" is treated the same as the categorical factors.
Read the documentation to find out about special provisions
for controlling the order of effects, or which interactions
you want to examine. - Preferably, the interactions should
be examined *after* the earlier effects, or else you (or the
program) should pay special attention to the coding - such
as, "centering" the covariate, or the choice of reference
category in ANOVA.
>This was what I wanted to avoid... but it sounds like you're saying
>that's not the case?
>
>Thanks for your clarification!
Bruce knows that procedure; I don't.
--
Rich Ulrich
Bruce Weaver
> On Wed, 20 Jan 2010 11:10:49 -0800 (PST), Adrienne
>
> >Thank you Bruce, that is helpful - I think I must have been
> >interpreting the results incorrectly. My understanding was that when
> >a variable was entered as a "covariate", the model first removed all
> >variance in the DV due to that covariate variable, then looked at the
> >effects of the "fixed factor"predictors on the remaining variance.
>
> The ordinary method of ANOVA or regression is to look at
> each of the variables after removing the effects of the
> others - categorical or continuous.
Which is why I'm a bit confused by Adrienne's desire to include
covariates, but at the same time to not have their effects partialled
out. But perhaps I've misunderstood. (It wouldn't be the first
time.)
> It is usually possible
> to specify an evaluation in a particular order... and you
> are right to be concerned about whether something labeled
> "covariate" is treated the same as the categorical factors.
>
> Read the documentation to find out about special provisions
> for controlling the order of effects, or which interactions
> you want to examine. - Preferably, the interactions should
> be examined *after* the earlier effects, or else you (or the
> program) should pay special attention to the coding - such
> as, "centering" the covariate, or the choice of reference
> category in ANOVA.
>
> >This was what I wanted to avoid... but it sounds like you're saying
> >that's not the case?
>
> >Thanks for your clarification!
>
> Bruce knows that procedure; I don't.
>
> --
> Rich Ulrich
The default for GLM (and most other things in SPSS) is Type III sums
of squares, for which the order in which you enter the variables has
no effect. If you change it to Type I SS, then order would matter.
The Help files contain a page with info on the 4 types of SS.
Adrienne
> Which is why I'm a bit confused by Adrienne's desire to include
> covariates, but at the same time to not have their effects partialled
> out. But perhaps I've misunderstood. (It wouldn't be the first
> time.)
>
Ideally, what I'd like to do is look at the main and interaction
effects of all the predictors as if they were all of the same type -
that is, the independent contribution of each predictor (categorical
or continuous) to explaining the variance in the DVs. My concern was
that by listing the continuous variable as a covariate, its
contribution to explaining the variance would be treated differently
(partialled out), and the analysis would only look at the
contributions of the categorical predictors above and beyond that
effect...
> The ordinary method of ANOVA or regression is to look at
> each of the variables after removing the effects of the
> others - categorical or continuous.
If this is the case, as Rich suggests, then maybe I am just worrying
too much! Either way, I appreciate your help!