Panel data models

[Paul Leeds]

Hi everyone

I am doing a research which involve a panel data analysis with dependant variable that take the value of 1-100. This value is a result of a readability analysis called "Flesch reading ease" each score represent a reading description " ease/difficult/very difficult, etc"

This table helps to understand the scoring method (the Raw score are actually the dependent variable values and could take any value from 1 to 100):



The main independent variables is service fees (continuous and in logs)
Control variables are: Firm age (logs) + return (ratio %) + industry (sic codes) + std dev of profit + market capitalisation (logs)

I am now to decide on the appropriate panel data model on stata and have some questions:

1- Is my dependent variable categorical (since it can take only 1-100 values?) or continuous?
2- I have done the hausman test to decide which model between fixed and random effect model and turn out to be random. Am I doing this right?
3- I have winsorized the independent variables (because of the outliers) using stata, do I have to choose the robust regression as a result of this?

Any posts would be appreciated.

Many Thanks

Paul

[Jorge Eduardo Perez Perez]

1- Is my dependent variable categorical (since it can take only 1-100 values?) or continuous?

Although to be strictly precise you would treat it as categorical, there isn't much gain to this for such a large number of categories so it would be better to treat it as continuous. This is standard practice.

2- I have done the hausman test to decide which model between fixed and random effect model and turn out to be random. Am I doing this right?

Impossible to know if you don't provide more details about what you did.

3- I have winsorized the independent variables (because of the outliers) using stata, do I have to choose the robust regression as a result of this?

Quite on the contrary: if you believe that winsorization took care of the outliers, then you don't need robust regression. A good check would be to run the model with a robust regression method on the unwinsorized variables, and compare the results to those of a standard regression analysis on the winsorized variables.

[Paul Leeds]

Hello Jorge

Thanks for your reply.

What I did with the hausman test is conducting the panel data regression twice, once with fixed and the other with random effect and stored them in the memory.
Then, I run the hausman test ans see whether the prob is <= .05 and found out that it is greater than .05 and hence going for random effect model.
Does that sound alright?

Many thanks

Paul

[Jorge Eduardo Perez Perez]

That sounds about right. The Hausman test tells you that there aren't any systematic differences between the fixed and the random effects models. Consider using the sigmamore option as an additional test or a robust Hausman test.

[Paul Leeds]

Thank you very much Jorge

Your replies were of great help

[Umar Farooq]

Although to be strictly precise you would treat it as categorical, there isn't much gain to this for such a large number of categories so it would be better to treat it as continuous. This is standard practice.

I have panel data with categorical dummy variable as dependent variable.

Can I still use Fixed and Random effect model.
Moreover, I also want to use some data mining techniques such as stepwise regression. Are these kind of test are available in fixed and random effect models in STATA?
Your responses will be high appriciated

REGARDS

[Nick Cox]

Umar Farooq: You have asked this question already, so you need to wait for a reply. Asking the same question twice (or more) won't make it any clearer or make it more likely that you will be answered,

Meanwhile, a search in Stata [NB, not STATA] for stepwise will tell you what there is and why it is widely thought a bad idea.

search stepwise