Re: st: can GLLAMM handle missing data?

[David Hoaglin <dchoaglin@gmail.com>]

Hi, Francesca.

I have no simple answer.  What you should do depends on the extent,
pattern, and nature of the missingness.  All I know about your data so
far (I think) is that you have repeated measures for respondents at 7
time points.  From the number of level-2 units (566) and the number of
level-1 units (2284), it appears that you have a substantial number of
missing observations (566 x 7 - 2284 = 1678).

I suggest that you start by examining the patterns of missing
observations and considering the possible reasons for the missingness.
 It will be important to understand those reasons in as much detail as
possible, to determine whether the observations that are missing are
"missing completely at random," "missing at random," or "missing not
at random."  (More than one of these categories may apply.)  The
information on the nature of the missingness will affect how you
interpret results from -gllamm-, and it will also guide choices that
you will need to make if you use multiple imputation.  Don't
underestimate the amount of work that you may need to do before you
proceed with any analysis.  Considerable judgment is involved, and the
process cannot be automated.

David Hoaglin

On Wed, Mar 26, 2014 at 9:15 AM, Pesola, Francesca
<francesca.pesola@kcl.ac.uk> wrote:
> Hi David,
>
> Thank you very much for your helpful reply.
>
> Do you think as good practice I should:
>
> 1. Run the analysis with the missing data  and rely on the GLLAMM flexibility;  and
>
> 2. Impute the data using MI as sensitivity analysis
>
> Thanks,
> Francesca
>
*
*   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/