home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: sci.math.stat
- Path: sparky!uunet!charon.amdahl.com!pacbell.com!sgiblab!sdd.hp.com!usc!elroy.jpl.nasa.gov!ucla-cs!ucla-mic!agsm!iwelch
- From: iwelch@agsm.ucla.edu (Ivo Welch)
- Subject: Questions on GS-Causality With Panel Data Set
- Message-ID: <1992Nov18.145329.4290@mic.ucla.edu>
- Nntp-Posting-Host: risc.agsm.ucla.edu
- Organization: UCLA, Anderson Graduate School Of Management
- Date: 18 Nov 92 14:53:28 PST
- Lines: 29
-
-
- I have a basic understanding of Granger-Simms causality (or, better, GS
- temporal ordering). I have seen bivariate time-series tests in two flavors:
-
- [1] a vector autoregression (Y_0= a Y_-1 + b X_-1 + n)
- [2] two time-series regressions and a subsequent correlation
- of the residuals (Y_0 = a Y_-1 + n_t; X_0 = b X_-1 + m_t;
- if n_t-1 correlates with m_t, then "Y sort of causes X".
-
- I presume that the first method is less restrictive than the first, but loses
- a little by mixing expected changes with unexpected changes. (I may be wrong,
- here, altogether.) Further, multivariate extensions are a bit less natural in
- the second procedure.
-
- Most studies seem to perform either of these analyses with long time series.
- Yet, I have a panel data set with only 5 years of data, but thousands of
- firms/people. My question is whether it makes much sense to run the above
- regressions cross-sectionally (and year-by-year or pooled), and claim GS-like
- causality. It does appear to be different; the relevant statement is now how
- does a cross-sectional outlier in X_{i,t-1} correlate with a cross-sectional
- outlier in Y_{i,t}.
-
- Are there comparable studies (especially if these studies use easy
- computational/notational methods; I am less interested in asymptotic theory
- than I am interested in actual use).
-
- Thanks.
-
- /ivo welch
-