Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: GEE and planned missingness


From   Steve Samuels <[email protected]>
To   [email protected]
Subject   Re: st: GEE and planned missingness
Date   Tue, 25 Jun 2013 19:35:53 -0400

Correction: You can do an analysis of x alone (not x1 and x2) in waves 1-3 and of u alone (not u1 and u2)  in waves 2-4 

Steve



I'm going to reduce your problem to two variables not measured at every
wave plus a third measured every time.

Suppose x is measured at waves 1-3; u is measured at waves 3-4; z is
measured at all waves

. gen x1 = x*(wave == 1)
. gen x2 = x*(wave == 2 | wave==3)
. gen u1 = u*(wave ==4)
. gen u2 = u*(wave ==2 | wave ==3)

These are defined for all waves

The following model will fit everything

. xtgee mjuser x1 x2 u1 u2 i.wave z

To test for a differential effect of x between wave 1 and waves
2 and 3

. lincom _b[x1] - _b[x2]

You might expect such an interaction, because collinearity of u and x
will change the meaning of the coefficient _b[x2].  To check this, fit
a model with x1 and x2 only, ignoring wave 4; and u1 and u2, ignoring
wave 1. In fact, I'd recommend that you do this first and test for
differential wave effects on x and u.

Another approach would be to multiply impute x and u in the wave
where each is missing.

A completely different question: is this survey data? If so, I
suggest that you -svyset- and use -svy: glm- with one of the
approaches above; you won't need id information, since
tests, SEs, and CIs will be based on the survey design factors.

Steve
[email protected]

.




To fit all the data at once, you create interactions of X1 with group
and X2 with group, letting group 2 (waves 2 and 3) be the baseline:

Then: On Jun 25, 2013, at 12:10 PM, Kris Anderson wrote:

Hi, all.

I am running GEE with panel data with four waves of data. I have time
varying predictors that were assessed at different time points (i.e.,
predictors at wave 1-3 [NMU] and a different set at waves 2-4 [CC EXPAND
SE]). I'd like to include them in the analysis but my examination of the
documentation (and past posts) suggests that cases with missingness at a
given time point will be dropped. Given that data was missing by design
for all cases at these waves, this is problematic. Is there a way around
this?

Here's my code for reference.

xtgee mjuser CC EXPAN NMU SE sexR i.wave, f(binomial) corr(uns)
vce(robust) eform estat wcorr


I'd appreciate suggestions of alternatives if this is a nonstarter.

Thanks, Kris


* * 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/

On Jun 25, 2013, at 12:10 PM, Kris Anderson wrote:

Hi, all.

I am running GEE with panel data with four waves of data. I have time varying predictors that were assessed at different time points (i.e., predictors at wave 1-3 [NMU] and a different set at waves 2-4 [CC EXPAND SE]). I'd like to include them in the analysis but my examination of the documentation (and past posts) suggests that cases with missingness at a given time point will be dropped. Given that data was missing by design for all cases at these waves, this is problematic. Is there a way around this?

Here's my code for reference.

xtgee mjuser CC EXPAN NMU SE sexR i.wave, f(binomial) corr(uns) vce(robust) eform
estat wcorr


I'd appreciate suggestions of alternatives if this is a nonstarter.

Thanks,
Kris


*
*   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/


*
*   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/


© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index