Complex hierarchical model of panel data with fixed effect

[Ariyo Irhamna]

Dear all,

I have weakly balanced panel data with observations of around 45000 of 5 years: the outcome variable is Y at the plot level (categorical), and my interest independent variables are at the village (continuous) X1 and crop levels (categorical) X2.

It was suggested that I use an order logit regression of panel data with a fixed effect with three dummy dependent variables, but I am unfamiliar with that.

I would appreciate your suggestions on the most appropriate specification and code.

Many thanks!

[Carlo Lazzaro]

Ariyo:
the -fe- estimators is not available from -xtologit- for sound statistical reasons.
You might want to consider an -ologit- with -i.panelid-, -i.timevar-, and -vce(cluster panelid)- standard errors.
However, due to the massive number of your panels (45000/5=9000) the model will take tons of time for the MLE to converge (if feasible).
Another option, conditional on grouping in two levels the >2 levels of your depebdent variable, is to run a linear probabilistic model via -xtreg,fe- (with the usual caveats that comes with considering a 2-level categorical variable as it were continuous).
On a different note, I doubt that your -village- variable is continuous.

[Ariyo Irhamna]

Hi Carlo,

I would love your insights on why the -fe- isn’t available in the ordered logit model!
If I were to explore a probit panel model instead with a fixed effect at the district level, could you share the specifications and the corresponding code?
Additionally, I want to confirm that the independent variable in the village is continuous, and it’s suggested that fixed effects be applied at the district level.

Your expertise is greatly appreciated, and I’m eager to learn more! Thank you!


Best,
Ariyo

[Carlo Lazzaro]

Ariyo:
1) about -fe- and -xtologit-, see feologit: A new command for fitting fixed-effects ordered logit models;
2) and 3) -xtoprobit- has no -fe- either. You may want to consider_

probit y <independent variables>  <controls> i.district, vce(cluster clusterid)

.