lagged variable causes issues with the logitfe command

[Luis Mijares Castaneda]

Hello, I'm reposting the issue I have in the hope that more people will comment. I am having a problem with the -logitfe- command,when I add the lagged outcome variable L.dental_Wave to the model below, I receive the following error "Computing analytical correction initial values not feasible r(1400);" . I am interested in using the lagged outcome variable Dental_Wave, it is a binary variable that lists if a respondent had a dental visit in the year. Any advice would be appreciated.

. local contrls2 L.dental_Wave c.tract_ppl_density_wave c.tract_poverty_wave c.a
> ge_Wave  ///
> i.smoked_marijuana_alt_Wave c.smoked_cigarettes_Wave i.alcohol_Wave c.depressi
> ve_symptoms_Wave ///
> i.education_Wave

. 
. 
. logitfe dental_Wave `contrls2', analytical ibias(yes) tbias(yes) 

Computing uncorrected fixed effects estimator

note: multiple positive outcomes within groups encountered
note: multiple positive outcomes within time periods encountered
note: 6141 groups (11355 obs) dropped because of all positive or
      all zero outcomes
Iteration 0:  f(p) =  -3585.264  
Iteration 1:  f(p) = -2168.1412  
Iteration 2:  f(p) = -2035.5913  
Iteration 3:  f(p) = -2013.8838  
Iteration 4:  f(p) = -2009.6491  
Iteration 5:  f(p) = -2008.7317  
Iteration 6:  f(p) = -2008.5208  
Iteration 7:  f(p) = -2008.4705  
Iteration 8:  f(p) = -2008.4602  
Iteration 9:  f(p) = -2008.4584  
Iteration 10: f(p) = -2008.4582  
Iteration 11: f(p) = -2008.4582  
Iteration 12: f(p) = -2008.4582  

Computing analytical correction
initial values not feasible
r(1400);

end of do-file

r(1400);

[Mitch Downey]

Hi. My gut reaction is that one of three things is true:
1) Once logitfe has dropped the 6141 groups with all-positive or all-negative outcomes, then L.dental_Wave perfectly predicts dental_Wave.
2) Once logitfe has dropped the 6141 groups with all-positive or all-negative outcomes, then L.dental_Wave doesn't vary within groups.
3) Something went wrong with the tsset and L.dental_Wave is always missing (for instance, because you didn't use the delta option to set the gap between waves?).

Keep if mind that if you only have 3 time periods, including a lagged value drops the first time period, so then you only have two time periods and you really have to worry whether you have any usable variation left between those two time periods.

Also, it seems the problem is related to initial values. There's probably an option within logitfe to change those, and you could play with that. But, of course, you should worry that being able to estimate a model only by messing around with initial values might suggest that you've only found a local optimum and not a global one.

[Felix Bittmann]

I think Mitch has some very good points. As an alternative, did you try

xtreg (...), fe

[Luis Mijares Castaneda]

Felix Bittmann I used the -felogit- command since it was recommended to handle problems with incidental paramters and the fact that my outcome variable is binary

[Felix Bittmann]

Agreed, the felogit command is potentially the most sophisticated. However, if it does not run at all, xtreg (linear probability model) or xtlogit could be feasible alternatives.