## Re: SV: SV: st: svy jackknife problems

Richard Williams wrote: > At 03:43 AM 3/15/2006, you wrote: > >Thanks Nick, this was also helful. I particularly like the advice to > >use -aweigth- to get the correct point estimates for the correlation > >coefficient with survey data. This should be public knowledge! > >Frankly speaking I never quite figured out why Stata needs to make > >these distinctions between different types of weight variables. It > >is very confusing, and I can't remember having this problem when > >using SAS or SPSS, so is it really necessary? > > That is because SPSS weighting is terrible! Basically, it is limited > to fweights/iweights. The SPSS documentation even says "If the > weighted number of cases exceeds the sample size, tests of > significance are inflated; if it is smaller, they are > deflated." Stata has a huge edge in this respect. Sure, Stata is > confusing, but it is confusing because you have to choose the correct > option from several choices, whereas with SPSS all you have to do is > pick the single (often incorrect) option that is offered!

I think SPSS now also offers a module for complex samples. With this module you can get correct standard errors with weighted data for just some procedures, however also at the cost of being more complicated (and some dollars).

BTW: With "normal" weights, SPSS seems to be inconsistent even for point-estimations. Consider the following little analysis:

. set obs 1000 . gen x = uniform()>.5 . gen y = 2 * x + invnorm(uniform()) . gen w = uniform()*100

. sum y [aw=w] if !x . local mean0 = r(mean) . sum y [aw=w] if x . local mean1 = r(mean)

. di as result `mean1'-`mean0'

As it should be, the result will be equal to the coefficient of the equivalent regression analysis:

. reg y x [aw=w]

Now do the same in SPSS. Calculate the weighted means of two groups and compare their difference with the analogous regression-coefficient. I cannot do it, because I don't have access to SPSS. But two years ago we have found, that in SPSS the mean difference was not equal to the regression-coefficients. Only the results of the regression analysis were correct (i.e. equal to the Stata results).

Uli

-- kohler@wz-berlin.de +49 (030) 25491-361 * * For searches and help try: * http://www.stata.com/support/faqs/res/findit.html * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/

Tag: