Working around a significant Levene's test in an ANCOVA?

#1
Hi there,

I'm working with some data in which I'm looking for mean differences across 15 subject categories of data on a variable (Clicks) that also covaries with another (Time). I'm trying to control for Time to neutralize its effect on the mean differences in which I'm interested.

I've hit a bit of a road block, however, in that when I run the ANCOVA, my Levene's Test is significant (see attached).

Everything I've read suggests this is a major violation of the ANCOVA's assumptions, but given that I have a large sample size (n=500), can I proceed anyway?

If not, is there another workaround? Eliminating outliers or combining categories (where appropriate)?

ANCOVAs are pretty much the outer limit of my statistical knowledge at this point.

Any guidance would be greatly appreciated.

Thanks,

Jordan
 

Karabiner

TS Contributor
#2
I've hit a bit of a road block, however, in that when I run the ANCOVA, my Levene's Test is significant (see attached).
Have you calculated the ratio between the largest versus the smallest variance?
Everything I've read suggests this is a major violation of the ANCOVA's assumptions,
Is it? With equal sample sizes, at least the oneway ANOVA is not much
affected by unequal variances; maybe the same applies to ANCOVA.
If not, is there another workaround? Eliminating outliers or combining categories (where appropriate)?
Eliminating "outliers" in order to achieve equal variance sounds too much
like forging of data. Probably someone else has a better solution than me,
but I'd suggest that you adjust clicks for time (i.e. regress clicks on time; but
you should first check whether there's truly a linear relationship between
counts an time [I doubt that]), then use the adjusted variable (i.e. the
resisuals from the regression) as dependent variable in your oneway
ANOVA. If variances are unequal, you can use the Brown-Forsythe-
correction or the Welch-correction.

With kind regards

K.
 
#3
Thanks so much for your help. So of this fixes may be beyond my statistical know-how at this point, but I'm going to try to muddle through them and see what I can come up with.
 
#4
Thanks so much for your help. I really appreciate you taking the time.

Some of these fixes may be beyond my statistical know-how at this point, but I'm going to try to muddle through them and see what I can come up with.

A colleague has also suggested that I could use a chi-square to determine whether the clicks/subject category depart significantly from the assumed equal distribution. I haven't yet figured out how to make SPSS give me the chi-square for clicks/category, however.

Thanks again.