Complex vs. two-level
Message/Author
 Emily Rosenzweig posted on Thursday, September 10, 2015 - 8:24 am
I have a simple study, and I have figured out how to model it with type = complex (where I can account for the fact that I have multiple observations per person and thus the errors will be correlated), but am wondering if a two-way specification is more appropriate.

I present 66 people with 11 scenarios (everyone sees the same scenarios), and ask them to rate those scenarios on X and Y. Basically I want to know the correlation between X and Y, where that reflects the average of the 66 within-subject correlations. In essence I'd like to run 66 regressions of Y on X, and then get an estimate of the average beta for X across those 66. I also expect that Y may vary by scenario.

Based on the user guide and things I've found in the discussion forum, here is the syntax I have been using in trying to run this two-way (where scenario = s1 to s10)

WITHIN = X s1 s2 s3 s4 s5 s6 s7 s8 s9 s10;
CLUSTER = id;
DEFINE: CENTER X (GRANDMEAN);
ANALYSIS:TYPE = TWOLEVEL;
MODEL:
%WITHIN%
X ON Y s1 s2 s3 s4 s5 s6 s7 s8 s9 s10;
%BETWEEN%
Y;

Can someone tell me: does the within-level estimate I get for X reflect what I want it to (based on what I've mentioned above, where I want to know the average within-subjects beta for X)?
 Linda K. Muthen posted on Thursday, September 10, 2015 - 1:56 pm
If you simply have several variables for each person, you do not need to use COMPLEX or TWOELVEL. Multivariate modeling accounts for the correlations. You need COMPLEX or TWOLEVEL only if you have clustered data, for example, students nested in schools or employees nested in companies where you are interested in student and school performance or employee and company performance.
 Emily Rosenzweig posted on Thursday, September 10, 2015 - 5:24 pm
Perhaps I wasn't clear enough -- my IV and DV are measured 11 times per subject, so simply running a multivariate regression on long-form data would not account for the fact that each set of 11 rows should have correlated errors. But again, what I'd implicitly like to do is run one 11-observation regression for each of the 66 participants, and then get an estimate of the average beta across those 66. Can you help me understand if that is what my two-way syntax is giving me in the within subjects output?
 Linda K. Muthen posted on Friday, September 11, 2015 - 11:42 am