Fixed effects regression with or without year dummies?

[Gabriela Kalibatseva]

Hi all,

I am working with balanced panel data covering the period of 32 consequent years. I am analyzing the effect of marital status on the life satisfaction. I am treating the variables as cardinal since reg and ologit provide more or less the same results, also Hausman points towards FE rather than RE.

However, I am not sure if I need to include year effects (i.year) in my regression. Currently I am running the conditional fixed effect model (xtlogit) with year dummies and it takes significant time and seems to never provide results. I didn't have this problem before adding the year dummies and I wonder now if that might be the trigger or the big massive of data.

In general answering the question of life satisfaction would require or not controlling for the year effects ?

Thank you!

Kind regards,
Gabriela

[Carlo Lazzaro]

Gabriela:
I'm not clear with your using -regress- to analyze panel data with a categorical regressand (as it would seem from your -xtlogit- mention).
Besides, as per your T dimension (32 years), you have a T>N panel dataset: I am not sure that -xtlogit- is the way to go, nor I'm aware of Stata official or community-contributed commands to deal with such panel dataset when you have a binomial dependent variable.

[Gabriela Kalibatseva]

Hi Carlo,

thank you for the quick reply ! I am also doing pooled OLS and ologit and comparing the partial correlation finding that the difference is small therefore I am using Pooled OLS since the interpretation of the coefficients is easier. Since it is panel data I do fixed effect and login fixed effect regression (for this one assuming life satisfaction =1 is over 7 on the scale of 0-10 and 0 otherwise). Let me know if you have a suggestion for a better regression model. I have 800K observations.

Regards,
Gabriela

[Carlo Lazzaro]

Gabriela:
if your dependent variable is ordered and you have panel data, you should go -xtologit-.
Beware that a simpler interepretation of a wrong regresion model is not helpful at all.

[Gabriela Kalibatseva]

Hi Carlo,

thank you for the heads up. I did try running with -xtologit- however it still takes forever to run and I am not sure I will get any output at all. Doe that mean that I have too many observations or my model specifications are not correct?

Regards,
Gabriela

[Carlo Lazzaro]

Gabriela:
complex model like -xtologit- take time to give back an output.
If you are concerned about the correctness of model specification, you can start with a more parsimonious model, add predictors one at time and see when problems begin.

[Gabriela Kalibatseva]

Hi Carlo,

just did that and turns out that the 'trouble-maker' are the year dummies (i.year). I will unfortunately get the same lack of output when I try regression with 'year' variable only. I would assume this Isi driven then by the size of the sample rather than misspecification of the model. Would you agree?

Kind regards,
Gabriela

[Carlo Lazzaro]

Gabriela:
I do not know how many panels your dataset is composed of; hence, I cannot say whether your diagnosis is on/off target here.
That said, it may also be that -years- express quasi-extreme collinearity with some other predictors.

[priya sawaliya]

Respected sir,
This is my first post on statalist forum. I am using xtologit command as my data is panel and dependent variable has three categories. After running this ordered logit regression i want to check the proportional odds assumption. i have tried a lot but could not get the correct command. please help me....