dakila wrote:"ROT" means the rotated factors. In order to estimate the factors do the following steps:
1) estimate factors (C) using all variables except FFR (you did this step)
2) estimate slow moving factors (Fs) from the slow moving variables.
3) estimate the regression by OLS : C=b1*Fs + b2*FFR + e
4) estimate the rotated factors : C-b2*FFR
given your reply to adriana a couple of years back, I was wondering why exactly we would need to follow steps 1 to 4 separately? If I understood the example file correctly, factor estimation and rotation is taken care of in the background when executing your brilliant favar command. Would you agree?
Secondly, I would also like to double check that the observable factors (=policy variables, such as ffr) must not be included in the data set from which the principal components are to be extracted (compare your step 1) above).
Thirdly, in their working paper version, Bernanke et al. state that the number of principal components to be estimated from their macro data set is equal to K + M (the number of unobservable factors plus the number of policy variables included). Since I would like to estimate a FAVAR with three observale policy variables, I would first determine the appropriate number of unobservable factors to be estimated by a scree plot or numerical criterion, such as IC_2(K). Finally, I would add the number of policy variables (M = 3) to the result given by the scree plot or IC_2(K) such that I would estimate K + 3 factors rather than just K. Would you agree?
Fourthly, I would like to know how to set the ordering (in the Cholesky sense of the word) in the FAVAR by code. Suppose I seek to include three observable policy variables called Yt_1, Yt_2 and Yt_3. Economic theory stipulates that their ordering be (1) Yt_3, (2) Yt_2, (3) Yt_1. Would the command to derive impulse response functions for the Yt_2 variable with the specified Cholesky ordering be:
Code: Select all
favar(factor=6,horizon=48,rep=1000,ci=0.9, impulse=Yt_2) 13 xdata xslow xir tcode yx_name @ Yt_3 Yt_2 Yt_1
Fifthly, I would like to ascertain that the macro data used to estimate the unobservable factors is made stationary before the favar command. And that the vector of transformation codes reflects the transformations applied to the variables for which impulse responses are to be derived. In other words, the transformation codes following the favar command are there only to determine whether cumulative impulse responses should be calculated or not. Would you agree?
And lastly (thanks for making it so far!), I would like to know if it is possible to specify particular variables as exogenous (for robustness checks)? I know this question has been asked before but I don't believe it has been answered yet.
Thank you SO MUCH in advance for answering my questions.
All help will be much appreciated!