SAS vs Stata Jackknife

[Bruce Bradbury]

I'm getting different results when using SAS (9.4) and Stata (v11) for jackknife regression. Anyone know why these are different?
Below are the key parts of my SAS and Stata code. If I run SAS surveyreg using the default Taylor series linearisation, and also run the corresponding Stata code on the same data, I get (near enough to) identical results. When I use the jackknife approach, the std errors are essentially the same as the Taylor series approach in Stata, but about 10% larger in SAS.

SAS CODE

  title "Regression: Taylor series linearisation";
  proc surveyreg data=temp1 total=lsac.stratum74with73   ;
     strata stratum74with73;
     cluster pcodes;
     weight defwt;
     model ready5_as = edl edm edh /noint;
  run;
  title "Regression: Jackknife method (10% larger SEs)";
  proc surveyreg data=temp1 varmethod=jk ;
     strata stratum74with73;
     cluster pcodes;
     weight defwt;
     model ready5_as = edl edm edh /noint;
  run;

STATA CODE

  svyset pcodes [pweight=defwt], strata(stratum74with73) fpc(totalstratums)
  * Taylor series linearisation
  svy: regress ready5_as edl edm edh, noconstant
  * Jackknife calculation
  svy jackknife: regress ready5_as edl edm edh, noconstant

[Bruce Bradbury]

I think it has something to do with the FPC. If I remove the FPC information from the code, all four approaches give essentially the same results (the larger std errors). But presumably the optimal estimates should take account of this? The SAS documentation implies that the jackknife estimate doesn't, and shouldn't, take account of the FPC, but it would seem the Stata code does in some way. Which is correct?

[Steve Samuels]

The fpc is certainly permitted with the jackknife, at least in some circumstances.. See, for example, Wolter, 2007, p. 174; Lumley, 2004, p. 8

References:

Lumley, T. (2004). Analysis of complex survey samples. Journal of Statistical Software, 9(1), 1-19.
http://www.jstatsoft.org/v09/i08/paper

Wolter, Kirk M. 2007. Introduction to variance estimation. New York: Springer.

[Bruce Bradbury]

Steve, Thanks for these helpful comments and references. I haven't read the Wolter reference yet, but have found some useful comments on the Stata support site at http://www.stata.com/support/faqs/st...stage-designs/ There, it is suggested that a FPC for the first stage PSU selection is not appropriate if there is sampling within the PSUs (which there certainly is in my case).

So, even though the survey providers do provide information to make an FPC, I think I should not be doing so - under either the linearisation or jackknife approaches.

(The design is a regional stratification, followed by a first stage selection of postal areas, followed by a clustered sampling of people within these postal areas. Only information on the first stage stratification and psu selection is provided.)