Convergence not achieved with xtlogit, fe, but it does converge with xtlogit, re

[Lynn Morgan]

I am using xtlogit on the following data:

input str2 st int date byte club float(combo l ca com) byte vmse
"AK" 22412 0 .05 0 0 0 0
"AK" 22419 0 .05 0 0 0 0
"AK" 22426 0 .05 0 0 0 0
"AK" 22433 0 .06 0 0 0 0
"AK" 22440 0 .06 0 0 0 0
"AK" 22447 0 .06 0 0 0 0
"AK" 22454 0 .06 0 0 0 0
"AK" 22461 0 .06 0 0 0 0
"AK" 22468 0 .09 0 0 0 0
"AK" 22475 0 .09 0 0 0 0
"AK" 22482 0 .09 0 0 0 0
"AK" 22489 0 .09 0 0 0 0
"AK" 22496 0 .19 0 0 0 0
"AK" 22503 0 .1923 0 0 0 0
"AK" 22510 0 .1923 0 0 0 0
"AK" 22517 0 .29 0 0 0 0
"AK" 22524 1 .4111 0 0 0 0
"AK" 22531 1 .4111 1 0 0 0
"AK" 22538 1 1 1 0 0 0
"AK" 22545 1 1 1 0 0 0
"AK" 22552 1 1 1 0 0 0
"AK" 22559 1 1 1 0 0 0
"AK" 22566 1 1 1 0 0 0
"AK" 22573 1 1 1 0 0 0
"AK" 22580 1 1 1 0 0 0
"AK" 22587 1 1 1 0 0 0
"AK" 22594 1 1 0 0 0 0
"AK" 22601 1 1 0 0 0 0
"AK" 22608 1 1 0 0 0 0

The data is at the state and week level. My dependent variable is club, which is binary. Combo is a federal level variable between 0 and 1 that varies over time, but does not vary by state. L, ca, com, and vmse vary over time and state, and also have a range between 0 and 1.

When I run xtlogit, re I have no problem. When I run xtlogit, fe with all the variables, I get the message that convergence is not achieved. When I remove the combo regressor and re-run the fixed effects model then convergence is achieved, but the combo regressor is key to my regression, so I do not want to exclude it. I have checked collinearity, and all my VIF's are below 3.

There are 11 states that are dropped due to no variation in club in those states, but that leaves 40 states (DC included) that do have variation.

My questions are:
1. why does xtlogit, re converge while xtlogit, fe does not?
2. why does xtlogit, fe converge if combo is excluded but not if it is included?

I appreciate any feedback.

[Lynn Morgan]

I should have included my code in the above post:

xtset state date
xtlogit club combo
xtlogit club combo, fe

again, I get convergence with the random effects but not the fixed effects.

combo is a time-fixed effect variable, so it does not vary by state. Is that a problem for convergence?

[Andrew Musau]

These are very different estimators. You can try unconditional logit and plug in the estimated coefficient as a starting value for the conditional estimator.

logit club combo i.state
mat b= e(b)
xtset state date
xtlogit club combo, fe from(b, skip)

[Joro Kolev]

If you have many weeks per state, you can as well stay with what Andrew showed

 
 logit club combo i.state

this si consistent if you have many weeks per state.

[Lynn Morgan]

Many thanks for these suggestions, and apologies for the delayed reply.

I tried the code:

logit club combo i.state

and got the following:

Note: 615 failures and 973 successes completely determined.
convergence not achieved
r(430);

[Carlo Lazzaro]

Lynn:
can't you add other predictors?

[Lynn Morgan]

I have used other predictors, but when I plot the data, the combo variable moves very closely with club, so I would like to include it. There is only variation within states in the binary club variable for the first five weeks of data (and there are 40 weeks total). Could this be a problem? There are 13 states that are dropped because they have no variation in the club variable, but that leaves 38 (DC included).

Are there any general or theoretical reasons xtlogit, fe may not converge, but the random effects would?

[Carlo Lazzaro]

Lynn:
if -combo- and -club- are highly correlated and have lomoted within-panel variation, this may explain your problem with the (conditional) -fe- estimator.

[Lynn Morgan]

Okay, thank you!

[Carlo Lazzaro]

Lynn:
-lomoted- reads -limited- (I'm progressively losing the letters on my keyboard ).