On Wed, 17 Oct 2012, 佃鹏 于 wrote: The Granger test is not "for" autoregression, but it is conducted in the framework of a vector autoregression (VAR). The theory of Granger causality is well developed for the time series case. For panel data it is more complicated; for one thing you'd have to decide if the causality result is supposed to be in common ("homogeneous") across your panel units, or not. If you google "Granger causality panel data" you'll find discussions of this. In gretl, Granger causality tests are provided among the results when estimating a VAR, but the "var" command is not available for panel data, only time series. I think this is true of most econometric software. Gretl scripting has the flexibility to allow you to construct a related test for panel data, if you can figure out exactly what it is you want to test. Allin Cottrell