Justin I think you need to reconsider this model. You don’t just have a little issue with power, it’s a huge issue. If you consider the 6 groups, basically
you are saying you want to get an overall effect and specific set of effects for 36 predictors between 6 groups. Running this through G*Power, I calculate that you would realistically need around 2-300 people. You might argue that you are going to test these
all separate, as seems to be indicated by your request for a t-test, but you really should be then correcting for this as multiple tests (in other words the same rule still applies for power).
I would consider reducing group sizes. I would also take a look at your 36 variables, and look at them carefully to see if they can be reduced or combined.
I don’t know what these 36 variables are, but you may find that they are indicative of some latent constructs when combined (they may have even been designed that way to begin with).
Now, as for how to do what you want, people have given you this already. You can’t do it with a t-test, you would switch to an ANOVA. When you “control” for
something, you are computing the variance in the Y variable accounted for by the “control” variable. In reality, you are removing the variance from the beta 1 coefficient for the beta 2 coefficient, which you are calling a control variable. The reality of
this is that the beta 2 coefficient is also now having the variance of beta 1 removed from it, i.e. both are controls for each other.
You can accomplish this in a few ways, but you won’t actually use the ANOVA command under means comparison, you want Univariate general linear model. You can
include the race variable as a fixed effect factor or as a covariate, and in the end, it will give you precisely the same results. Either place is fine, they are mathematically equivalent. For the model, you want this to be main effects only. If you include
the interaction, then you end up looking at the specific results within African American vs Caucasian, and not just controlling for the variance explained of being an African American. That’s a different research question altogether. So my recommendation
would be as follows.
UNIANOVA Yvar BY Smokecat WITH RaceX
/METHOD=SSTYPE(3)
/INTERCEPT=INCLUDE
/PRINT=PARAMETER
/CRITERIA=ALPHA(.05)
/DESIGN=Smokecat RaceX.
The code above includes syntax to give you the parameter estimates. Makes things easier to interpret, you get the beta’s then. No interactions, as mentioned.
Your only remaining problem is that Yvar is really Yvar 1 through 36. Many would argue you need to take your alpha criteria for each, divide by the number of Y’s, and input that. That would come to .001. We could get into a long discussion of when this
is appropriate and when it isn’t. If this is part of a true experimental study, and you want the results to be indicative of experimental trials, then you need to do this for anyone to take the results seriously. If its totally exploratory, then there is
an argument that no correction is needed. I would argue, however, that the results need to be presented in this way, that a follow-up study is necessary with sufficient sample size to allow for this.
Matthew J Poes
Research Data Specialist
Center for Prevention Research and Development
University of Illinois
510 Devonshire Dr.
Champaign, IL 61820
Phone: 217-265-4576
email:
[hidden email]
From: SPSSX(r) Discussion [mailto:[hidden email]]
On Behalf Of Justin Blehar
Sent: Saturday, September 29, 2012 5:18 PM
To: [hidden email]
Subject: Controlling for Race with SPSS 20
Hello All,
Not sure how much detail is needed so I'll give you a quick overall but I'm trying to control for race and am unsure how to best go about this. I know that I can run a partial correlation and control for race using the menu but is this really controlling for
race? If not is there a better way? How would I do this for a t-test?
This is a cross sectional design looking at cognition and smoking in a psychiatric population. There are six groups I'm looking at; Never Smokers, Former Smokers, Nonsmokers (includes both never smokers and former smokers), Heavy Smokers, Light Smokers, and
Smokers (includes heavy and light smokers). I have 36 scale variables that I want to compare between each of these groups. When I break out the groups by race (just looking at box plots and mean comparisons) there are clearly some large race effects (e.g.
parental education, level of functioning, IQ, etc...). I'd like to be able to correct for this in each analysis. I'm running both correlations and t-tests (maybe this isn't the best process?).
If I run a partial correlation and control for race is this really controlling for race?
When running the t-tests how do I control for race?
Any help would be greatly appreciated.
V/R
Justin
| Free forum by Nabble | Edit this page |