I tried to specify a partial proportional odds regression in STATA using the
gologit2 runs extraordinarily slow in my dataset (108k observations of 9 vars). For example,
ologit PfSt age if bund == 1
runs in less than 1 sec, while
gologit2 PfSt age if bund == 1, pl
which is essentially the same, as the
pl option enforces parallel lines in all vars, i.e., the
ologit specification, takes about a minute on my computer.
Calculating the fully specified model is possible with
mlogit (~ 1 min), but not with
gologit2, even if I specify for which variables parallel lines should be assumed.
The regression command is
ologit PfSt age age_sq gender i.bundesland tt tt2 [fw=size], cluster(bundesland)
PfSt ranges from 0 to 7 and
tt capture time trends (
tt = year − 1996 and
tt2 = -1/tt +1, a “phasing out” trend)
Now, considering that relaxing the parallel lines assumption for some variables would be identical to adding interaction terms to my model specification, I could change my regression command to
ologit PfSt age* gender i.bundesland#i.PfSt tt*, cluster(bundesland) would yield me different coefficients per level of
PfSt for the one variable for which I would relax the parallel lines assumption. Is that correct?
However, the estimation process fails to converge, as now (1200th iteration) negative log pseudoelikelihood ratios rise (i.e., drop in absolute values, now <<0.0001), but they are
(non concave). I wonder why is this the case?, as both
mlogit did produce results after a few iterations.