The file slides080205.pdf are slides used for my talk at
the MN Summer Workshop for Macroeconomic Theory at the
University of Minnesota.

For those present at Larry Christiano's talk following mine,
let me answer some questions that Larry did not.
(Note that larryslides.pdf contains his slides.)

  1. Does including government spending in the observer
     equation have a significant effect on the CKM baseline
     parameterization?

     Answer: No.  We put in government spending so that
     the data vector was the same when we added more variables
     and shocks to the VAR.  If we drop it, the baseline
     parameterization barely changes.

  2. Does it make sense to free up measurement error during
     the maximum likelihood estimation?

     Answer: Not if one is trying to give the SVAR approach
     it's best chance.  With error-ridden data, the SVAR
     researcher will have an even tougher time uncovering
     the true impulse responses.  

  3. Was Larry's MLE estimate on our sensitivity graph?

     Answer: Yes.

  4. Is Larry right to say that the standard errors will alert
     the researcher to problems with the SVAR procedure?

     Answer: No.  Take, for example, our QDSVAR(.99) results
     with quasi-differenced hours.  Technically, there should
     be no problem with over-differencing in this case and
     the SVAR procedure should uncover the truth.  The confidence
     bands in this case are such that the researcher would 
     confidently predict a fall in hours following a technology 
     shock.  That prediction is wrong and the standard errors
     won't help you.

  5. Is it logically inconsistent to choose parameterizations
     for the model that imply that technology shocks account 
     for a lot of the variance of output if the SVAR procedure
     predicts a small contribution?

     Answer: Yes.

  6. Larry said that short-run identifying assumptions "work 
     remarkably well."  Is that an answer to CKM's critique?

     Answer: No, because we view the main goal of the SVAR 
     literature as (possibly) identifying promising classes of 
     business cycle models using a simple time series procedure.   
     If the short-run identifying assumptions are not satisfied 
     by a broad class of models, then obviously it is hard to 
     say which models of a broad class are consistent with the 
     data and which are not.   If the SVAR approach is to be
     useful at all, the identifying assumptions must be 
     satisfied by a broad class of potential models. Period.
     Then Monte Carlo exercises must be done to assure that the
     procedure can in fact uncover the model impulse responses.


Ellen McGrattan
August 2005