Re: [Gretl-users] Angrist-Newey or Chamberlain test

Am 06.07.2015 um 13:23 schrieb JOSE FRANCISCO PERLES RIBES: I'm not familiar with these tests, I just looked into a relevant Baltagi et al paper, where they say: "...the sum of T terms. Each term of this sum is simply the degrees of freedom times the R2 from a regression of within residuals for a particular period on all leads and lags of the independent variables." This sounds relatively simple, as in (1) run a panel-FE regression ("panel Y const X --fixed-effects"), (2) store the residuals ("series feresids = $uhat") (3) for each time period create the leads and lags of the RHS (4) regress the residuals on those, and record (and sum) the R2*dof. Here's a working implementation: <hansl> # example panel data (balanced?) open nc_crime.gdt list lX = wfed list lY = taxpc # calculate the "within residuals" (?) panel lY const lX --fixed-effects series feresids = $uhat # prelims genr time T = max(time) test = 0 loop tix = 1..T smpl time == tix --restrict --replace maxlead = T - tix maxlag = tix - 1 # non-negative if maxlead == 0 || maxlag == 0 # workaround for strange gretl requirement: list leadlagX = lX(xmin(maxlead,maxlag) to xmax(maxlead,maxlag)) else # both actual leads and lags list leadlagX = lX(+maxlead to -maxlag) endif ols feresids const leadlagX --quiet test += $rsq * $df endloop print test </hansl> But I have no idea whether this is really the right way to do it. (For example the constant term in the second-stage ols...?) If during your work with this you find anything to correct, it would be nice to let us know. hth, sven