I believe there is a mistake when explaining the effects of sigma square alpha and sigma square u at 11:47. If sigma square u (within variation) is 0, sigma square alpha must be relatively higher, such that there is more between unit variation. If there its more between variation, it means that unit-specific effects are strong, such as in the case of FE. Same goes for the other cases but vice versa.
Hi. I am in a class, so I cannot listen to the video. Do you mean that there is an error in my slide or in how I explain the slide? Unfortunrately I do not understand what you mean by your explanation.
Thank you for this very informative video, Mikko! Can you please explain why when we use mean-centering approach for a FE model we interprete beta as one unit of change in the IV within a cluster is associated with n units change in the DV? Looking on the equation, I see that we regress deviation from the mean of the DV on the deviation from the mean of the IV. So my interpretation would be: one unit increase in deviation from the mean of the IV within the cluster is associated in n units change in deviation from the mean of the DV. So why the first interpretation is actually true? How it can be expained?
To me, both interpretations seem the same. When IV increases by one, DV increases by beta. If you increase IV by one, you also increase its deviation from mean by one. It is important to understand that these are within estimates, which means that if IV increases by one for one observation in the cluster, it causes an increase of the DV for that same observation but the change in IV has no influence on the other observations in the cluster. Think about the differences between the within, between, and contextual effect.
I do not have a video on the general idea of GLS. I focus more on application than statistical theory. That being said, I think that a general explanation of GLS might be useful. I will put that on my list of things to do.
@@mronkko thank you very much! It would be really helpful maybe doing a parallell between gls and linear mixed models with random intercepts, I've started studying it (I'm using R's lmer but I've been told to consider gls) but I'm quite lost about it
@@larissacury7714 What you are asking cannot be done because mixed model is a model and GLS is an estimation technique. GLS can be used to estimate mixed models (or at least a subset of model parameters). When someone tells you to consider GLS over ML, you should ask why they think so. (Note lmer does ML estimation by default.)
@@mronkko I have selected RE model over FE model via Hausman test . But found Heteroskedasticity in my RE model after the Pangan LM test. Now should I robust my RE model or run FGLS to take care of the heteroskedasticity?
I am not sure was you mean specifically. You can estimate fixed effects by using dummies, so yes. If you mean whether you can estimate GLS-FE and FGLS at the same time, if your statistical software implements such estimator, then yes.
I believe there is a mistake when explaining the effects of sigma square alpha and sigma square u at 11:47. If sigma square u (within variation) is 0, sigma square alpha must be relatively higher, such that there is more between unit variation. If there its more between variation, it means that unit-specific effects are strong, such as in the case of FE. Same goes for the other cases but vice versa.
Hi. I am in a class, so I cannot listen to the video. Do you mean that there is an error in my slide or in how I explain the slide? Unfortunrately I do not understand what you mean by your explanation.
Thank you for this very informative video, Mikko!
Can you please explain why when we use mean-centering approach for a FE model we interprete beta as one unit of change in the IV within a cluster is associated with n units change in the DV? Looking on the equation, I see that we regress deviation from the mean of the DV on the deviation from the mean of the IV. So my interpretation would be: one unit increase in deviation from the mean of the IV within the cluster is associated in n units change in deviation from the mean of the DV. So why the first interpretation is actually true? How it can be expained?
To me, both interpretations seem the same. When IV increases by one, DV increases by beta. If you increase IV by one, you also increase its deviation from mean by one. It is important to understand that these are within estimates, which means that if IV increases by one for one observation in the cluster, it causes an increase of the DV for that same observation but the change in IV has no influence on the other observations in the cluster. Think about the differences between the within, between, and contextual effect.
Hi! IThank dou! 'm used to lm and lmer, but I'm trying to get a better understanding of GLS. Do you have a more introductionary video to GLS?
I do not have a video on the general idea of GLS. I focus more on application than statistical theory. That being said, I think that a general explanation of GLS might be useful. I will put that on my list of things to do.
@@mronkko thank you very much! It would be really helpful maybe doing a parallell between gls and linear mixed models with random intercepts, I've started studying it (I'm using R's lmer but I've been told to consider gls) but I'm quite lost about it
@@larissacury7714 What you are asking cannot be done because mixed model is a model and GLS is an estimation technique. GLS can be used to estimate mixed models (or at least a subset of model parameters). When someone tells you to consider GLS over ML, you should ask why they think so. (Note lmer does ML estimation by default.)
In correcting for heteroskecedicity in RE model.. which is the best estimator? Robustness of RE model or FGLS
@@mronkko I have selected RE model over FE model via Hausman test . But found Heteroskedasticity in my RE model after the Pangan LM test.
Now should I robust my RE model or run FGLS to take care of the heteroskedasticity?
@@mronkko thanks you soo much
Can I use FGLS with Fixed effects at the same time?
I am not sure was you mean specifically. You can estimate fixed effects by using dummies, so yes. If you mean whether you can estimate GLS-FE and FGLS at the same time, if your statistical software implements such estimator, then yes.
Would you tell me how to perform FE GLS in Stata?
xtreg $ylist $xlist, fe *(specify y and x lists first and then add fe (fixed effects) to your xtreg)