Is it required for panel data to use dummy variables? I am doing research considering seven countries and I have panel data. My question is: do I need to include dummy variables every time I use panel data in regression, or is enough to do it as a fixed effect in Stata, without dummy variables?
 A: Assuming you have the aggregate data for seven countries and 10 years, running a OLS model with dummy variables is equivalent to fixed effects. 
If you include dummy variables for countries (there will be six, one omitted to avoid the dummy variable trap) or dummy variables for years (if there are 10 years, then there will be nine dummies, again to avoid the dummy variable trap) then that will be one-way fixed effects. 
However, if you include dummy variables for both countries as well as years, it will be two way fixed effects. You can find details here if you are using Stata.
A: David Fernando implies not only the panel data by his question, but also the specific method of estimating his model, namely the fixed effect estimator. David, however, did not mention what kind of model he works with. 
You have to be quite specific here. Fixed effect panel data models are conditional models. They provide inference by conditioning on the average (or total) response within the panel. This usually does not limit the interpretation (although sometimes conditional models get tricky; conditional heteroskedasticity ARCH models behave very differently from the corresponding marginal models that would use a fat tailed distribution like Student $t$ for regression errors). However, you need to have this distinction in mind when working with panel data. So, the conditioning means that


*

*In linear models, conditioning on the panel mean with a Gaussian model means subtracting the panel mean. And, by an unfortunate coincidence of the matrix algebra of linear regression, this is equivalent to introducing dummy variables. Sometimes you can hear of this implementation of the fixed effect estimator as DVLS -- dummy variable least squares.

*There are two more panel data models that can afford fixed effect estimation: the binary model with the logit link, and the Poisson model with the log link. They have a sufficient statistic (total response) that can be used for conditioning. xtlogit, fe and xtpoisson, fe do just that, but they internalize these computations, and since these are iterative maximum likelihood estimators, there are no matrix algebra tricks applicable to them.

*Adding dummy variables to your other estimators (e.g., Heckman) may be a good idea, but you have to realize that you cannot refer to the resulting estimators as "fixed effects". While they would still capture some sort of an individual effect, they may not produce the benefits of the fixed effects estimators such as a better control over endogeneity biases.


Generally, I would recommend looking into the black Wooldridge to get it right.
