Namaskar
I am working on panel data. As per my regression analysis the R-square value of the model was R-squared 0.369134 and Adjusted R-squared 0.302597. Like wise another findings showed R-squared 0.085355 and Adjusted R-squared 0.078845. Kindly suggest ways to solve the issue. Also guide which value (among weighted or unweighted) is to be considered for analysis in results of random effect regression?
Urvi -
R-square may not be very informative. I am not very experienced in panel data problems, but for regression in general, perhaps you could (1) estimate variances of prediction errors, (2) consider graphical residual analyses, and (3) compare model performances, where models are each supported by subject matter theory, by removing data to save for testing purposes, where you then predict for those data using the candidate models, and then compare results to the withheld actual data.
Best wishes - Jim
Dear Urvi,
I do agree with the opinion of Prof. James R Knaub that R square is not very informative in this case. In panel data analysis, rely more on individual significance and overall significance of the model instead of R square or adjusted R square. Generally, R square is low in cross sectional data as compared to time series data. In panel data due to heterogeneity of cross sections, it is not too high. If your data is more time dominant, R square can be higher as compared to the case when panel data is more cross section dominant. In general, more related included explanatory variables boost the value of R square. Yet, one has to focus more on objectives of the research to be fulfilled from individual significance and overall significance of the model making sure that there is no model specification bias and avoid spurious regressions. Another point to mention here is that a very high R square in the presence of very few significant t values indicates the presence of multicollinearity and spuriousness of the regression.
Best wishes,
Hearty thanks to Sir James and Sir Atif for important insights. By considering the significance of the variables individually I found my results to be significant. Also Sir, few of my results showed very high R Square. To make data normal, homoscedastic and make sure absence of serial correlation I employed Square and Squareroot transformation, Log transformation, Reflect and Sqrt and Reflect and Log transformation. And found that very few variables could be transformed as shown by ShapiroWilk and KS test. And many variables had 0.001, 0.002,0.006 etc results of KZ test whereas Shapirow Wilk results were 0.000.And I ran panel data regression on those variables only and got such results which have extreme R Square values but individually many of them are significant too.
If you are running panel data, I would suggest looking into employing cross-section and period effects (be it random or fixed).
R2 is not the best, nor is Adj. R2, but when looking at the explanatory power of the model these are chiefly used.
If you have low R2 and your coefficients are statistically significant, then you should look at the fitted vs. actual data and see if they are close or not. It may be that you have low R2 simply because a great explanatory power is put on the constant term and the explanatory variables used are not good in explaining changers in your dependent variable.
It does not matter if you have cross-section or time-series data; the interpretation of R2 stays the same.
Muhammad Atif Nawaz has explained it very well and I agree that R square does not hold primary importance in model fitness criteria.
Really a good problem. It seems to me you have to search and work further on it.
I think it is not a big problem that R2 is less than 60%, your variables explain more than 35% if you have a good argument and justification for your model then you should not worry about that.
I agree with Yazan. If you have theory to back up your results you need no border so much about the R2 level. However, you have to ensure that you did not violate the relevant regression assumptions in you analysis.
I am also facing the similar situation with low adjusted R square value for panel data random effect model (~30%). But with the same set of variables fixed effect model (LSDV) shows more than 90% value for adjusted R-square. My data set is long panel i.e. number of cross sections is very high. Because of the long panel and confirmed by Hausman test, I need to use random effect model.
From the different posts of the current blog I understand that low adjusted r-square for random effect should not be worried as long as all variables are significant and backed up by literature.
I am curious to know why adjusted R-square value differs so much between fixed and random effect. Also, if anyone can provide reference of book or research paper where authors mentioned about low adjusted r-square of panel data random effect then I will be highly obliged. Thanks in advance.
In the similar lines, my fixed effect model to predict industrial growth using growth volatility has very low R-sq but all variables are significant. The issue is that the model violates all the assumptions of normality, serial independence and homoskedasticity. Is it ok to still accept the model?
It is a older tread, but the topic is still relevant.
Regardless of working with time series or cross-section data, it is important to clarify what are the objectives of your model. Is it explanatory or predictive.
Also, as panel data checks the consistency of the relationships in the model over an extended period of time, it can affect the explanatory and predictive power of the model in different ways.
Here is a link to an interesting paper that explains the difference: Article To Explain or to Predict?
Hello. The discussion is quite interesting. I was facing the same issue of low R-square in Fixed effect regression(~0.06). I discussed with many of the financial experts on this topic as my data was financial panel data with large number of cross sections with small Time period( T=96, N=260). What I understood from the discussion is that in panel models with financial data, R-square is always low as financial data is not normally distributed in most cases and is highly volatile. Hence, R-square is of no significant in explaining the fitness of the model. I would appreciate if someone provide a detail discussion of the validity of this argument.
that's a normal, since the fact that the model re-iteration based on time series data
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression