0

I have a Twoways "within fixed effects panel regression model and detected multicollinearity, autocorrelation and heteroskedasticity. For heteroskedasticity I want to use heteroskedasticity-consistent standard errors by White (1980). This results in different standard errors. For multicollinearity, I'd drop a variable and for autocorrelation I'll try to reduce the sample period.

Anyway, anything I'll apply will change my regression results or even my model. And that's the point where I'm stuck. Therefore, please allow me to ask:

  1. What should I do first and why?
  2. Which results/model should I use as my starting point?

Should I fix the 1. problem > use the new model/results > test for 2. problem again > fix it > use the new model/results > test for 3. problem again > fix it?

mizzoh
  • 11
  • Changing the standard errors won't change the coefficients of your model. – Helix123 Nov 18 '20 at 19:12
  • I was already wondering about that coefficient, too. However, since I test for 2 independent variables, I also use two models. Unfortunately, I compared the results of White's (1980) approach with the wrong model. That's why I erroneously assumed different coefficients, although I did not read about them in theory. Sorry. – mizzoh Nov 18 '20 at 22:22
  • Methods for standard errors exists which cope with heteroskedasticity and autocorrelation, see Arellano (1987). – Helix123 Nov 19 '20 at 21:38
  • Thanks. This has indirectly answered my questions. I removed 2 variables from the model, so I had no more multicollinearity after the vif test. Then I used robust standard errors by Arellano (1987) to tackle autocorrelation and heteroskedasticity. The order in which I do things is now clear. You can actually write this comment to the answer section and I'll mark your answer as "solved". – mizzoh Nov 20 '20 at 19:49

2 Answers2

0

First off, remove those predictors whose VIF is > 10 (or analogously, those with a correlation coefficient with the rest of the regressors > .8). Then, I'd use the Newey-West (1987) method to deal simultaneously with autocorrelation and heteroskedasticity. The main goal of this method is to find a covariance matrix of the errors whose diagonal entries are all equal (homoskedasticity) and the off-diagonal are 0 (no autocorrelation). Hope it helps

Mario Aguilar
  • 1
-1

I have the same problem with heteroscedasticity. I think it's best for that that you use the White correction. It won't change you coefficients, but it will probably solve your heteroscedasticity.

LillyAnn
  • 1