Introducing dummies in the panel data model is not uncommon, but what gave me more concern is that you have decided about fixed effect model. As per I understand, you need, first, theoretical justification of employed model, and later the statistical results.

If you have continuous variable as DVs and binary variable on other hand and you're running FE, your independent variables may drop. It happened with me too. I solved the issue by addressing the issue of endogenous variables in my model, and further analyzed with system GMM. I've been also advised to try other form of dynamic panel. Perhaps, you may try some of them to resolve this issue. 

Thanks Iftekhar Ahmed.

I have tried dynamic panel option. But, I have weak instruments problem. And I get non stable coefficients when I use technics of dynamic panel (Arellano-Bond estimator (1991), Arellano and Bover (1995), Blundell and Bond (1998) ).

Kofi. You have not given us enough information to enable one give you a a very balanced answer. For example, what is your T and N? What are your variables and which ones are endogenous and which ones are exogenous? Without this information, I can only afford a proximate response as follows.

(1) You can not have a dummy when you are using a FE panel data model because the dummies will be differenced away during estimation.

(2) When dealing with panel data you have a choice between pooled OLS, FE, RE and dynamic panel model. To choose between OLS, on one hand and FE/RE on the other hand, you will have to conduct tests for individual and time effects. These include Breusch-Pagan LM test, King and WU, Honda standardized LM tests. To choose between FE and RE, we usually use the Hausman test.

(3) For you to use dynamic panel model, you will need to have many observations because the efficiency of GMM is based on generating instruments (many) which consume many degrees of freedom. If you have a few observations, your models will not yield efficient estimates. With few observations, you may wish to try RE or GLS or pooled OLS (with robust standard errors).

Let me know whether this helps you.

HI Eliud Moyi

My data contains 756 individuals on 29 periods (1985-2013). These statistical units live in five types of area. Areas are classified into five groups (g1, g2, g3, g3, g4 and g5).

I estimate these impacts of urban type on kilometers travelled (km) .

The final goal is to compare parameters that are associated to urban type and observe if before and after 2000 these coefficients are statistically different.

The basic model is:

Km(i,t)= mu(i)+beta (i)*Urban (i)+ gamma* X(i,t)+epsilon (i,t)

Technically, I want to know if the parameter beta differs before and after 2000

Koffi.

You have enough observations to run both pooled OLS, RE and system GMM. I wish to advice as follows;

1. Begin with pooled OLS (which is usually the baseline) and compare your findings with RE. Avoid FE because your main variable of interest is a dummy. Your pooled OLS and RE results shouldn't be very different.

2. If you can justify GMM, then you can go ahead and try system GMM using the xtabond2 stata command. This command will require you to select varying lags and has many other options. The theory behind system GMM is a bit dense for a non-econometrician but if you prefer to use it, there are many references that can help you. Usualy, we justify system GMM because it is an IV estimator.

Let me know how it goes.

Thanks Eliud Moyi

Please can I use cross sectional invariant dummy variable. I have T=20 and N=6 and I am using PMG. The dummy is specifically to assess how a change in my data set affects my dependent variables