i ran a panel data with STATA these days, and find out my r-squared of Fixed effects model (FEM) is very low ( the following result)
R-sq:
within = 0.0191
between = 0.3425
overall = 0.1576
Ques1.
how should i improve my R-sq in this case ?
i have already test the VIF, Pwcorr, F test are all uncorrelated
Ques2.
is overall R-sq my FEM R-sq?
Lin Cy , in general there are few reasons that the Rsq is low:
1. the covariates itself is not important in the sense that theoretically and practically, it does not effect the dependent variable.
2. the covariates are not linearly linked to the dependent variable. Some transformation might be required.
3. Multicollinearity in the data. I spot some of the pairs have correlation more than 0.7. I am not sure the threshold of the correlation value in order to say the multicollinearity exists. But, for me it is a bit high. May be, there is a link between those two variables, unknowingly.
4. Perhaps, FEM is not suitable, you may want to try REM and Pooled OLS.
5. Panel data analysis and modelling is quite complex. There are several preliminary tests (unit root test, etc.) that you may have to run before using FEM because the test may indicate other model is suitable.
Try read this:
https://nomanarshed.wordpress.com/2015/09/18/tentative-panel-data-model-selection-plan/
and other sources as well.
Razik Ridzuan Mohd Tajuddin , I agrred wit your submission on R-square value. For multicollinearity, 0.7 is still OK, 0.9 is the benchmark.
Lin Cy , you did not report the Hausman test to indicated of fixed effects is appropriate, this is what Razik Ridzuan Mohd Tajuddin is pointing at in no. 5. Also, the coefficients are large while not log variables in absolute values and leave the ones in %. More so, your dataset have some attributes of system GMM approach, N>T, your dependent variable may also be persistence, just check for endogeneity bites.
Good luck
Lin Cy , the result is good to be reported in your study (fixed effect).
Have a productive week.
Lin Cy the fixed effect model seems fine to me except for covariate DPI, the p-value and CI suggest that it is not significant.
Seems to be fine.... in some research it is hard to have more than 30%
You may try to improve it with re-running with changes but if no improvement then those are maybe the best results.
Also remember to take a look at the Adjusted R-square too when adding covariates.
- Badges
- Science topic
- Similar topics
- Education