I have panel data consisting of 4 different countries and time variant explanatory variables. To analyse the impact of these time variant variables I am using Fixed Effect model. However, when I apply stationarity tests over dependent and independent variables, I observed that all of the variables are stationary except one of the independent variables. In this case, how can I apply Fixed Effect model? By taking differences of all dependent/independent variables or should I switch to another model rather than Fixed Effect?
After taking the first difference of the non-stationary variable, did you see the variable become stationary? If so, you may use the first difference of that variable in your model, especially when economic interpretation of the variable is not a focus.
Thanks for the reply. Yes, when I take the difference of nonstationary variable, it becomes stationary. But I am not sure if I can use for example dependent variable at level and similarly all stationary explanatory variables at level but only the nonstationary variable at first difference. From statistical point of view, does it make sense? Or should I just use first order differenced variables (for all of the variables) regardless of the stationarity? ( Note: I also tried that case , but R-squared was extremely low i.e 2% )
Have you tried to replace all variables with their first order difference? R-squared would be lower. But you can calculate a "synthetic" R-squared for the dependent variable at level even though your model is at the first difference.
Yes , I have also tried that and came up with a R-squared value of 2 %. How can I calculate synthetic R-squared for the dependent variable at level? Could you give me a reference ?
Aplicaciones Econométricas, también se puede utilizar las cadenas de MARKOV para calcular su probabilidad de permanencia.
Differencing is nothing more than looking at a change in one variable instead of its level. Therefore, you can use a model where you have variables at levels and variables that are differenced. Just be sure to make a note of the way variables are introduced into the model when you are interpreting the coefficients.
for non-stationary panel model,
please have a look in my question :
https://www.researchgate.net/post/Which_model_in_panel_data_I_can_use_for_non-stationary_data
Dear Nur Yesilay,
The analysis of panel data can be categorized into traditional panel data analysis and non-stationary panel data analysis. Well, if you have N>T, you should go for traditional panel data analysis (fixed, random, pooled OLS), however, if you have T>N, you should go for non-stationary panel data analysis (FMOLS, DOLS, PMG etc).
For traditional panel data analysis, if Hausman test result is significant, you should go for Fixed effect, however, an insignificant Hausman test result favors random effect. If you get insignificant Hausman test result, then test Breusch Pagan LM for choosing between random effect and pooled OLS. if BP LM test is significant, go for Random effect, however, if this test result is insignificant, go for pooled OLS.
On the other hand, in non-stationary panel data analysis, if all variables are stationary at level, you can use least squares estimation. if you get all variables I(1), with cointegration you can go for panel VECM, FMOLS and DOLS, while if you get all variables I(1) without cointegration you can go for panel VAR. However, an evidence of a mix of integration leads us to use Pooled mean group of panel ARDL.
Best wishes,
- Badges
- Science topic
- Similar topics
- Biological Science
- Histology
- Sectioning