Dear community,
OT: Originally I wanted to ask why my log-level -xtsur ..., onestep- regression in Stata12 on Windows (as well as on another computer) does not converge. Now, Stata was able to deliver results with the same specifications that did not work previously. Any general idea why? I.e. what could have influenced the non-convergence of a log-level or log-log onestep xtsur? (multistep/default xtsur did not converge even for level-level models). I might have to try different models and could use hints about what to avoid or what to change to get results.
My real question for this topic is still about -xtsur-: My dataset is a strongly balanced panel for 7 cities over 6 years with monthly data (7x72). The time dimension becomes smaller for statistics with lagged or forwarded variables, which I will have to include. I want to use the data to explain Rooms Sold (Nji) and Rooms Available (Bi) with ~12 variables.
Independent -hausman fe re, sigmamore/sigmaless- tests suggest to use FE for the first equation (Nji as dependent variable) and RE for the second equation (Bi as dependent variable. RE confirmed through -xttest0-). Furthermore, tests revealed the presence of
While the above tests suggest the use of -xtscc..., fe- and -xtscc-, respectively, this does not take into account a possible SUR problem with ADRi and OCCp=Nji/Bi.
A -sureg ..., corr- for the overall data (and seperately for 2 out of 7 cross-sections) suggests that it is a case of seemingly unrelated equations (for both log-level and level-level models).
However, since I am using panel data, -sureg- is not what I should use.
Is it hence possible to replicate the Breusch-Pagan test for -xtsur- or ascertain in another manner whether to use two -xtscc- or one -xtsur- model?
So far, the coefficients given by a level-level -xtsur..., onestep- as in the code above make the most sense when trying to interpret them.
If it is indeed a SUR situation, is -xtsur- still appropriate, given the presence of heteroskedasticity, autocorrelation and cross-sectional dependence and given that one equation is FE and the other RE?
Also, I already had a look at this thread, but I am hesitant to reshape my dataset due to its longitudinal nature (72 periods).
Thank you very much for any help you can give me and please don't hesitate to ask for further clarification or information!
Best,
Klaus
OT: Originally I wanted to ask why my log-level -xtsur ..., onestep- regression in Stata12 on Windows (as well as on another computer) does not converge. Now, Stata was able to deliver results with the same specifications that did not work previously. Any general idea why? I.e. what could have influenced the non-convergence of a log-level or log-log onestep xtsur? (multistep/default xtsur did not converge even for level-level models). I might have to try different models and could use hints about what to avoid or what to change to get results.
My real question for this topic is still about -xtsur-: My dataset is a strongly balanced panel for 7 cities over 6 years with monthly data (7x72). The time dimension becomes smaller for statistics with lagged or forwarded variables, which I will have to include. I want to use the data to explain Rooms Sold (Nji) and Rooms Available (Bi) with ~12 variables.
Independent -hausman fe re, sigmamore/sigmaless- tests suggest to use FE for the first equation (Nji as dependent variable) and RE for the second equation (Bi as dependent variable. RE confirmed through -xttest0-). Furthermore, tests revealed the presence of
- heteroskedasticity (-xttest3-, LRtest according to this guide),
- cross-sectional dependence (-xttest2-, -xtcsd, pesaran abs-),
- autocorrelation (-xtserial-).
Code:
xtwest lnNji YjM12 ADRi12 ADRc12 CPIji12 CPIjc12 H1N1i12, lags(0 6) bootstrap(400) xtwest lnBi ADRi12 OCCp12 LTp12 STp12, lags(0 6) bootstrap(400)
A -sureg ..., corr- for the overall data (and seperately for 2 out of 7 cross-sections) suggests that it is a case of seemingly unrelated equations (for both log-level and level-level models).
Code:
sureg (Nji12 Yj12 ADRi12 ADRc12 CPIji12 CPIjc12 H1N1i12 d_GFC d_OGHK d_OGL) (Bi12 ADRi12 OCCp12 STp12 LTp12) , corr . A/N: -skipped- . Correlation matrix of residuals: Nji12 Bi12 Nji12 1.0000 Bi12 0.5678 1.0000 Breusch-Pagan test of independence: chi2(1) = 162.496, Pr = 0.0000
Is it hence possible to replicate the Breusch-Pagan test for -xtsur- or ascertain in another manner whether to use two -xtscc- or one -xtsur- model?
So far, the coefficients given by a level-level -xtsur..., onestep- as in the code above make the most sense when trying to interpret them.
If it is indeed a SUR situation, is -xtsur- still appropriate, given the presence of heteroskedasticity, autocorrelation and cross-sectional dependence and given that one equation is FE and the other RE?
Also, I already had a look at this thread, but I am hesitant to reshape my dataset due to its longitudinal nature (72 periods).
Thank you very much for any help you can give me and please don't hesitate to ask for further clarification or information!
Best,
Klaus