Hi, thank you for the great video, I was relieved to find a simple explanation about what CREs are. One further question though: what other assumptions must hold in CRE? The ones from Fixed Effects, like parallel trends assumption, too? Is there any publication going into detail about the CRE assumptions?
@@mronkko Thank you so much for your quick answer! Because you have mentioned 2). Are the assumptions the same for a CRE Probit model, as there is no linear increase in Y but only from 0 to 1?
@@mronkko thanks a lot for your suggestion. I think that some IV are correlated with DP. I run CRE approach because I am interrested in evaluation of the effect of time-invariant covariates on DV and the RE assumptions are not respected.
Great summary, thank you. Can one include time dummies in the correlated random effects model (CRE) whether in an unbalanced panel or balanced panel data?
@@TrustGangaidzo That depends on your research question, but I find it difficult to see when both effects, if they both exist in the data, would not be of interest to report. That being said, our review of the literature shows that researchers rarely report and interpret both effects jyx.jyu.fi/bitstream/handle/123456789/66704/2/Antonakisym.pdf
Thank You Mikko, this is a very helpful video! :) I am using correlated random effects in my research. I am facing one challenge. I get the same within effect coefficient for fixed effect and correlated random effect model when I do not include year dummies in my equation. However, when I include year dummies (and also the average value of the dummies) in the equations, I get different within effect coefficients between the fixed effects and the correlated random effects model. Would you be able to help me understand why this happens? Regards, Siddharth
I do not understand the question. What you explain does not seem possible to do. For example, if you have three dummies, their average is 1/3, which is a constant. If you add that constant to your model, you are violating the no perfect collinearity assumption (MLR.3, Wooldridge 2013) and the model cannot been be estimated.
Dear Mr. Rönkkö, Thank you for the video. I have clustered cross-sectional data, that is, a certain number of households for a certain number of countries.I want to investigate the effects of a variable x that is constant within a country (so it is the same for every household of the same country, but differs for each country) on a consumption variable (that differs for every household). I have reason to believe that my independent variable of interest x is correlated with unobserved country charactersitics, but since it is the same for every household within a country I cannot use FE. I could use RE to include this country-constant variable, however, this would violate the assumption of no correlation between the unobserved and my x. It seems I also cannot use the CRE approach as I cannot add the country-mean of x since x is country-constant. Did I get this right, or is there a way to still consistently estimate this model?
If you want to estimate the effects of x, which varies only on the between country level, the simplest way to do it is to use a between regression. That is, just take country means of the dependent variable and regress it on x using country as the unit of analysis. Anything other than that is just adds complexity without adding any value if it is just the effect of x that you are interested in.
@@mronkko Ah that makes sense! Also, I wonder if it is possible to use multiple levels of cluster means within the model to get varied fixed effects. For example if my model is a country pair migration panel with years with variables that vary over the country pair, country of origin and country of destination. Would taking cluster means of the way the various variables average over be a logical thing to do? for example for country pair variables you take the cluster mean at country pair level and country of origin level variables you take the cluster mean at the country of origin level.
@@arpansagar6453 I cannot answer the question from the top of my head, but would need to think about it. I recommend that you simulate a large sample and then just test if you get the correct estimates using the approach that you propose. If you google: mundlak multiple levels you will get lots of relevant hits. economics.princeton.edu/wp-content/uploads/2021/08/two_way_mundlak-Wooldridge.pdf
@@AfricanHope If the data are clustered, you can use CRE, but I do not really see how that would be the case if your unit of analysis is a household. How would you calculate the required cluster means?
@@mronkko I have no idea 😭 I attempted using clogit and save fixed effects. But that did not work either. Could you please help me? How would my data be that I can use CRE? I’ve read a lot and it’s just making me more confused. Thank you
@@AfricanHope The techniques that you are discussing are multilevel techniques. You need to start by thinking what are the levels (L1, L2) and then how these map to your research question. If you have cross-sectional data with no level structure, then you cannot use any of the techniques that you mention. Here is some material that I have produced about multilevel data th-cam.com/video/dx8lZnoY8hU/w-d-xo.html
Hi, thank you for the great video, I was relieved to find a simple explanation about what CREs are. One further question though: what other assumptions must hold in CRE? The ones from Fixed Effects, like parallel trends assumption, too? Is there any publication going into detail about the CRE assumptions?
@@mronkko Thank you so much for your quick answer! Because you have mentioned 2). Are the assumptions the same for a CRE Probit model, as there is no linear increase in Y but only from 0 to 1?
Thank a lot for your precious explanation. I have a balanced panel data N=12 T=5, and I adopted CRE approach...But how fix for endogeneity?
@@mronkko thanks a lot for your suggestion. I think that some IV are correlated with DP. I run CRE approach because I am interrested in evaluation of the effect of time-invariant covariates on DV and the RE assumptions are not respected.
Great summary, thank you. Can one include time dummies in the correlated random effects model (CRE) whether in an unbalanced panel or balanced panel data?
Yes, you can include time dummies and it is in fact a common practice to do so if you want to eliminate macro-level time trends from the data.
@@mronkko thank you Prof
@@mronkko one more question….is it worth it to interpret both X and Xmean….or only will do?
@@TrustGangaidzo That depends on your research question, but I find it difficult to see when both effects, if they both exist in the data, would not be of interest to report. That being said, our review of the literature shows that researchers rarely report and interpret both effects jyx.jyu.fi/bitstream/handle/123456789/66704/2/Antonakisym.pdf
@@TrustGangaidzo You are welcome.
Thank You Mikko, this is a very helpful video! :) I am using correlated random effects in my research. I am facing one challenge. I get the same within effect coefficient for fixed effect and correlated random effect model when I do not include year dummies in my equation. However, when I include year dummies (and also the average value of the dummies) in the equations, I get different within effect coefficients between the fixed effects and the correlated random effects model. Would you be able to help me understand why this happens? Regards, Siddharth
I do not understand the question. What you explain does not seem possible to do. For example, if you have three dummies, their average is 1/3, which is a constant. If you add that constant to your model, you are violating the no perfect collinearity assumption (MLR.3, Wooldridge 2013) and the model cannot been be estimated.
Dear Mr. Rönkkö, Thank you for the video. I have clustered cross-sectional data, that is, a certain number of households for a certain number of countries.I want to investigate the effects of a variable x that is constant within a country (so it is the same for every household of the same country, but differs for each country) on a consumption variable (that differs for every household). I have reason to believe that my independent variable of interest x is correlated with unobserved country charactersitics, but since it is the same for every household within a country I cannot use FE. I could use RE to include this country-constant variable, however, this would violate the assumption of no correlation between the unobserved and my x. It seems I also cannot use the CRE approach as I cannot add the country-mean of x since x is country-constant. Did I get this right, or is there a way to still consistently estimate this model?
If you want to estimate the effects of x, which varies only on the between country level, the simplest way to do it is to use a between regression. That is, just take country means of the dependent variable and regress it on x using country as the unit of analysis. Anything other than that is just adds complexity without adding any value if it is just the effect of x that you are interested in.
@@mronkko Thank you very much for your valuable response!
what is the command for the correlated random effects model
Any ideas
Do correlated random effects require balanced panels or work just fine with unbalanced panels as well?
Unbalanced panels are fine.
@@mronkko ah nice! Also, how do they work in the context for two-way fixed effects (time and panel)?
@@arpansagar6453 You cannot do CRE and FE at the same time. But you can use a CRE model for the firm and then add year dummies to your analysis.
@@mronkko Ah that makes sense! Also, I wonder if it is possible to use multiple levels of cluster means within the model to get varied fixed effects. For example if my model is a country pair migration panel with years with variables that vary over the country pair, country of origin and country of destination. Would taking cluster means of the way the various variables average over be a logical thing to do? for example for country pair variables you take the cluster mean at country pair level and country of origin level variables you take the cluster mean at the country of origin level.
@@arpansagar6453 I cannot answer the question from the top of my head, but would need to think about it. I recommend that you simulate a large sample and then just test if you get the correct estimates using the approach that you propose.
If you google:
mundlak multiple levels
you will get lots of relevant hits.
economics.princeton.edu/wp-content/uploads/2021/08/two_way_mundlak-Wooldridge.pdf
Can I use CRE on cross sectional data? I’d appreciate your response. Thank you!
You can use it to any data that is clustered (multilevel). A cross section of e.g. multiple teams with multiple members each could be used.
@@mronkko I have household survey cross sectional data
@@AfricanHope If the data are clustered, you can use CRE, but I do not really see how that would be the case if your unit of analysis is a household. How would you calculate the required cluster means?
@@mronkko I have no idea 😭 I attempted using clogit and save fixed effects. But that did not work either. Could you please help me? How would my data be that I can use CRE? I’ve read a lot and it’s just making me more confused. Thank you
@@AfricanHope The techniques that you are discussing are multilevel techniques. You need to start by thinking what are the levels (L1, L2) and then how these map to your research question. If you have cross-sectional data with no level structure, then you cannot use any of the techniques that you mention. Here is some material that I have produced about multilevel data
th-cam.com/video/dx8lZnoY8hU/w-d-xo.html