I'm working with a database of longitudinal data with an unbalanced panel, data with monthly frequency. I'm using the "plm" package of R to analyze and run the models fixed and random.
NOTE: Data presented serial autocorrelation and heteroscedasticity, and I am correcting with the covariance of Arellano (1987) coming in plm package (R software).
1nd - It is redundant to run FE / RE and GMM as robust proof?
2nd - What is the impact of using the GMM if the results come out different?
3nd - What other suggestions would be to perform a robust test?
The analysis is very important.
You should use the Hausman test to choose between FE and RE models.
You should use the GMM method when you have a dynamic model.
I suggest you to use STATA which is very well for such models.
You should use Wooldridge test for autocorrelation and Breush Pagan test for heteroscedasticity.
Professor @Belloumi, thanks,
Note: In R I'm also using the hausman tests, Breusch Pagan and autocorrelation.
How can I determine if I should use a dynamic model?
What are the benefits of GMM compared to the FE / RE?
Dear Melquiades
You should use GLMM if the distributional assumption of your data suits for it. If the data seems to be normal distributed you should you LMM but if the data shows a differetn distribution, you should use GLMM with the appropriate distributional familly. Regarding the use aof FE or RE model, I found this presentation that may help.
http://www.princeton.edu/~otorres/Panel101.pdf
best
Elmer
Dear Melquiades
1) Linear models, with fixed or random effects, are a case in GLMM models. GLMM allows to analyze data with a different distribution than normal distribution.
2) GLMM also allows you to use random or fixed effect.
3) I do not agree to use a test to compare a fixed an a random effect model. One effect is random if the levels of the effect are a random sampling of a population of effects. If you select the levels of the factor then you have a fixed effect model. However I know that in practice a lot of people interchange the effect of some factor in order to obtain results or based in tests.
4) In the context of GLMM you can model the autocorrelation and the heterogeneity of variances.
Answer to your questions:
1) If you use different models you will have different results and I there is a risk to select the model which answer what we know. Nowadays it is very easy to run models, but in my opinion we need to think more on the data and the problem and then select the model. The questions that you have to answer are: Are my effects random or fixed? What is the distribution of my data? What kind of autocorrelation model has sense in my data?
2) Because you have counts you have to use GLMM with the correct specification of the distribution (Poisson for counts and Binomial for percentages) and an appropriate link function.
3) What do you mean with robust test?
Thanks professor @Juan Burgueño,
I understand the argument you made in relation to not base solely on statistical tests. My data is non-parametric and the dependent variable follows a behavior as in the attached graph.
Robust test: it is assumed that the results should be the same or similar regardless of how you analyze them. A different way to verify the hypothesis, is a different way you validate the hypothesis to confirm the evidence form.
thanks for your help.
Dear Melquiades
Regarding the figure, I think that you can use a log-normal distribution.
What I know as a robust test is a test which gives similar results when the data do not follow the assumptions of the test. If you have one data set and you analyze it with different methods or models the results will not be the same because you are assuming different things, for example that one effect is fixed or random. Beside, if you use several methods for test an hypothesis, what is the probability to take a wrong decision? It is not 0.05. How many similar test (models) do you need to confirm an hypothesis? They are hard question.
ok professor, i understand.
thanks for your help. In relation of the distribution is the same of my opinion. I think more about how to justify the method selected.
I WANT TO WORK ON POOLED DATA. DATA IS RELATED TO MIGRATED PEOPLE AS SUGAR CANE CUTTER. I HAVE DATA FOR 8 YEARS. I WANT TO KNOW ABOUT DYNAMIC PANEL DATA ANALYSIS.
IS IT APPLICABLE TO ANALYSIS THE SAID DATA SET?
WHAT ARE THE PREREQUISITE FOR THE SAME
ANY BOOK/ READING MATERIAL WHERE I CAN GET THE BASIC KNOWLEDGE OF DYNAMIC PANEL DATA
PLEASE HELP ME OUT
I HAVE WORKING KNOWLEDGE OF ECONOMETRICS, STATA AND EVIEW AS WELL AS I CAN HANDEL SPSS (VERSION 20) ALSO
THANKS
@Anurag Asawa
The model is according to the data panel and what you want answer. The problem of you work is very important.
Other important factor is the structure of the data. If it are balanced or unbalanced data.
The book that i work is: "Econometric Analysis of Panel Data". Baltagi
The software that you cite, the best of panel data is Stata.
i work with the R project software (free).
I hope this has helped.