Title:
this is an example of modeling with data
not missing at random (NMAR) using a
pattern-mixture model
Generating missing data indicators m
(analyzed as an arbitrary model)
Note that the m's are renamed as z's
in the real-data run of ex11.4
Note that the Monte Carlo setup is the same as
for mcex11.2 and mcex11.3
MonteCarlo:
names = y0 y1-y5 m1-m5;
nobs = 1000;
nreps = 1;
missing = y1-y5;
generate = m1-m5(1);
categorical = m1-m5;
save = ex11.4.dat;
Model population:
i s | y0@0 y1@1 y2@2 y3@3 y4@4 y5@5;
i*.25 s*.1;
y0*.4 y1*.4 y2*.35 y3*.3 y4*.4 y5*.5;
[m1$1*2];
[m2$1*1];
[m3$1*0];
[m4$1*0];
[m5$1*-1];
! Missing at t as a function of y_t-1, y_t, and previous m's
m1 on y0*-.1 y1*1;
m2 on y1*-.1 y2*1 m1*.5;
m3 on y2*-.1 y3*1 m1-m2*.5;
m4 on y3*-.1 y4*1 m1-m3*.5;
m5 on y4*-.1 y5*1 m1-m4*.5;
[i*2 s*-1];
Model missing:
! missing on y if m=1 (high logit)
[y1-y5@-15];
y1 on m1@30;
y2 on m2@30;
y3 on m3@30;
y4 on m4@30;
y5 on m5@30;
Analysis:
algo = int;
integ = montecarlo;
Model:
! analyzed as Diggle-Kenward selection
! which is not the model that generated
! the data, and the m's are not scored
! as survival dropout indicators so DK
! analysis is not achieved
i s | y0@0 y1@1 y2@2 y3@3 y4@4 y5@5;
i*.25 s*.1;
y0*.4 y1*.4 y2*.35 y3*.3 y4*.4 y5*.5;
[m1$1*2];
[m2$1*1];
[m3$1*0];
[m4$1*0];
[m5$1*-1];
m1 on y0*-1 (11)
y1*1 (12);
m2 on y1*-1 (11)
y2*1 (12);
m3 on y2*-1 (11)
y3*1 (12);
m4 on y3*-1 (11)
y4*1 (12);
m5 on y4*-1 (11)
y5*1 (12);
[i*2 s*-1];
Output:
tech9;