ANCOVA and Categorical Covariates
Eric
I am trying to figure out the best way to proceed with an analysis and
appeal to the group for assistance. I have a continuous DV, a
categorical IV and 2 categorical covariates. The advice I have
received is that it would be best to conduct an ANCOVA and dummy code
the covariates using k-1 dummy variables. Yet, it also seems that the
same question could be answered using regression.
Can anyone offer any counsel on the best way to proceed? I'm unclear on
how dummy coding the covariates in ANCOVA helps me meet the assumption
of at least one continuous covariate.
Also, if I pursue the regression route is dummy coding still necessary?
Thank you in advance for your assistance.
Eric
Bruce Weaver
In the dialogs for various procedures (e.g., GLM), SPSS uses
"Factor" to refer to categorical predictor variables, and
"Covariate" to refer to interval/ratio scale predictor
variables.
Having said that, if you are using GLM-->Univariate to
perform your analysis, why not just include your two
categorical covariates as two more "factors" (to use the
SPSS terminology) in the design? There is no need to create
dummy codes if you use GLM. But it /would/ be necessary if
you were to use the REGRESSION procedure instead.
Cheers,
Bruce
--
Bruce Weaver
wea...@mcmaster.ca
www.angelfire.com/wv/bwhomedir/
Steve
Bruce's approach is appropriate for your situation. Let's suppose that you
used dummy variables and defined them as COVARIATES as opposed to FACTORS in
the SPSS GLM module. SPSS would partial out the impact of these covariates
from the error term. BUT, correct me if I'm wrong, it will not covary out
the INTERACTION between the covariates and the FACTOR and between the
covariates themselves. If the GLM procedure is used, the default will be to
create a fully-factorial design in which the variance due to the
main-effects of the covariate variables are accounted for as well as the
variance accounted for by all of the interaction terms. This is not
necessarily a bad thing, but it will cost you in terms of degrees of
freedom. This may or may not be a problem, depending on whether or not the
interaction terms account for large portions of variance and the sample size
of the study. A compromise would be to use the GLM procedure, define the
covariates as factors, and use the "Model" feature to build a custom model
that does not include the interaction terms, thus replicating the ANCOVA
results.
I think that the choice of approach should be driven by theory as well as
practical concerns.
Cheers,
Steve
Eric
Bruce/Steve,
Thank you both very much. I just completed the analysis using your
guidance and now have meaningful results to interpret. The custom
model was the way to go as none of the interactions were significant.
Regards,
Eric