Re: st: My ANOVA and regression results don't agree

[David Hoaglin <dchoaglin@gmail.com>]

Hi, Jess.

If the range of the outcome variable is only from 1 to 5, and the
values are very skewed, the frequencies of some values (the larger
ones?) are likely to be rather small. You may want to consider
treating the outcome as an ordered categorical variable, instead of
analyzing the data as "continuous."  (Did the subjects respond by
choosing a number from 1 to 5, or did they choose categories described
by verbal labels, to which you have assigned the numbers?)

Treating the outcome as ordered categories would lead you to ordinal
logistic regression.   With that approach, it might be appropriate to
combine adjacent categories if the counts in some categories are too
small.  In the extreme version of combining, you could wind up with a
dichotomous outcome and a logistic regression model.

It's helpful that (so far) the interaction term is not significant.
Interactions tend to make interpretation more difficult.

David Hoaglin

On Wed, Jan 8, 2014 at 4:00 PM, Pepper, Jessica <jkadis@live.unc.edu> wrote:
> Hi David,
> The outcome variable has a range of 1 to 5 and is very skewed (mean 2.04, SD 1.18, skewness 1.02). I do have balance of assignment of conditions (i.e. the type of ad) within sex.
> The interaction term (adcontent#sex) is not significant in either the ANOVA or the regression. And when I drop the interaction from the model, the regression and ANOVA match again (i.e. both models show significant main effects for both variables). Given that situation, dropping the interaction from the model may be the way to go.
> Jess

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/