Mixed Model:Post-hoc comparisons/contrasts for an interactio

Moderators: statman, Analyst Techy, andris, Fierce, GerineL, Smash

markusd
Posts: 1
Joined: Wed Feb 08, 2012 1:08 pm

Mixed Model:Post-hoc comparisons/contrasts for an interactio

Postby markusd » Wed Feb 08, 2012 3:16 pm

Hi All,

I have the following problem and I hope that someone here can enlighten me.

I am running the following analyses with the indicated post-hoc comparisons:

GLM agg1 agg2 agg3 BY sa
/WSFACTOR=Zeit 3 Polynomial
/METHOD=SSTYPE(3)
/EMMEANS=TABLES(OVERALL)
/EMMEANS=TABLES(sa)
/EMMEANS=TABLES(zeit)
/EMMEANS=TABLES(sa*zeit) compare (zeit) adj (bonferoni)
/CRITERIA=ALPHA(.05)
/WSDESIGN=Zeit
/DESIGN=sa.

The factor "Zeit" has three levels and with the compare (zeit) command above I am getting all comparsions (1.vs 2, 2 vs. 3. etc.). However, I need to test specifically one level (for example agg1) against another one (for example agg 2).
And I dod not know how to tell SPSS to do that.
Importantly, it is that I need to compare agg1 with agg2 on each level of sa seperately.
Additionally, I see the "significance" column in the "Pairwise comparison" table. This might be a hint, but I do not see a) the exat test statistic and b) degrees of freedom or the like, which I would need to properly report the results.
The multivariate table below, again, indicates the comparsion for all types of combinations of "Zeit".

I hope that this was clear enough, to grasp what the problem is. I am convince that this is not a statistical problem but more one of SPSS or - better - me and SPPS.

Thank you very much in advance!
Markus.
dnichols
Posts: 6
Joined: Tue Dec 20, 2011 5:36 pm

Re: Mixed Model:Post-hoc comparisons/contrasts for an intera

Postby dnichols » Tue Apr 10, 2012 8:58 pm

With the EMMEANS specification you've given, the pairwise comparisons for zeit are done separately within each level of ar (if you simply had zeit with COMPARE rather than zeit*ar and COMPARE(zeit) you would be comparing zeit averaging over the levels of ar). In this case tests would be based on t with N-k degrees of freedom, or equivalently F with 1 and N-k df (which is the square of the t), where k is the number of levels of ar. The p values and confidence intervals are Bonferroni adjusted since you specified that. The family used is the set of pairwise comparisons within a given level of ar, not the entire set over all levels of ar.

You can get GLM to provide you with the actual test statistics (without Bonferroni adjustments to p values or CIs) using LMATRIX and MMATRIX subcommands. For example, if there are two levels of ar, add:

/LMATRIX 'Compare zeit at first level of ar' ALL 1 1 0
/LMATRIX 'Compare zeit at second level of ar' ALL 1 0 1
/MMATRIX 'zeit 1 vs zeit 2' ALL 1 -1 0;
'zeit 1 vs. zeit 3' ALL 1 0 -1;
'zeit 2 vs. zeit 3' ALL 0 1 -1

You should be able to match the values in the Contrast Results table to the mean differences from the Pairwise Comparisons table you get from EMMEANS. The table with Univariate Test Results for each set of contrasts will give F statistics and df for the tests. Again, these are not Bonferroni corrected.

Who is online

Users browsing this forum: No registered users and 1 guest

cron