st: Explaining strange xtgee results

[Jason Bond <jbond@arg.org>]

Hello.  My dataset is attached.

list x y id

shows the 3 variables I'm using:

scatter x y

and

line y x if id == 1 || line y x if id == 2 || line y x if id == 3 || 
line y x if id == 4 || line y x if id == 5 || line y x if id == 6 
||  line y x if id == 7

show the linked lines.  So I have 7 pairs of observations, each of 
which is positively correlated.  However, the overall trend of the 
between the pairs of points is negative, similar to Figure 1.1b in 
Diggle et al.'s book.

Running the simple population averaged model:

xtgee y x, i(id)

gives a negative estimate of the slope for x.  If I separate out the 
(baseline) cross sectional effect from the longitudinal (time 2 - 
time 1) effect, and run the same model

xtgee y x1rep x2diff, i(id)

where x1rep is just x repeated and x2rep is defined as x2 - x1, I get 
a negative relationship for the cross-sectional effect and a positive 
relationship for the longitudinal effect (which I expected).  I am 
sort of surprised that I am getting a negative estimate for the slope 
of x in the gee model above though.

If I run a mixed model using:

xtmixed y x || id:

I get a positive estimate for the slope estimate of x.  Is there 
something that I'm simply not seeing here?  Thanks much,

  Jason

Attachment: Example-gee-data.dta
Description: Binary data