Re: st:iccvar

["Robert Fornango" <robert.fornango@f1analytics.com>]

Dr. Enzmann,

Thank you for pointing out my *obvious* mistake. Many apologies to everyone
on the list.

As Dirk notes in his posting, -quickicc- does *NOT* compute the intraclass
correlation coefficient (ICC) for random slopes. Here is the description
from http://ideas.repec.org/c/boc/bocode/s457268.html :

"quickicc calcuates the intraclass correlation (ICC) after fitting a
two-level xtmixed model where the intercept is the only random effect. In
addition to calculating the ICC, this program also calculates the standard
error of the maximum likelihood large sample ICC"

Clearly I misread the description when I previously suggested that
-quickicc- might be used to get the ICC for a random intercept AND slope
model.

Of course, calculating ICCs is relatively easy to do even without such
post-estimation commands. I would refer interested readers to:
Rabe-Hesketh, Sophia, and Anders Skrondal. 2012. Multilevel and Longitudinal
Modeling Using Stata. 2nd Edition. College Station: Stata Press.


Megan,

While the means of each toxin might appear to be correlated over time, it
doesn't necessarily follow that they would be correlated at the individual
level (i.e. ecological fallacy).

-correlate- simply produces a correlation matrix for the variables entered.
In this case, including 'timepoint' will not provide any adjustment or
control for autocorrelation. So the fact that the correlation remained with
its inclusion is not surprising.

Since I don't know more about the details of your research, I will not
comment on whether you should or should not explicitly account for potential
autocorrelation. 

If you need to account for an autoregressive, or lagged-response, structure
you can use either -xtreg- or -xtmixed-. In this model you can include a
one-period (or more) lag of the dependent variable as a predictor. However,
this will results in losing one time point on the within patient estimates.
Also, it can result in bias due to the endogeneity between the lagged
outcome and the random intercept (in that case -xtabond- may be
appropriate).

In contrast, if you need to model an autoregressive residual, you can do so
with -xtregar- and the 're' option for random effects. Again, I will refer
you to the Rabe-Hesketh and Skrondal text provided above. I would also
recommend finding one of the many good texts on panel data analysis.

I hope this at least gets you pointed in the right direction, and that my
previous post on -quickicc- did not cause any confusion.

Rob

Robert J Fornango, PhD
Chief Executive Officer
F1 Analytics LLC
Phone: 602.578.7167
www.f1analytics.com




*
*   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/