Actually, the differences run a little deeper than this. SPSS, like most
general purpose stat packages of its generation never did handle sampling
design issues properly when it came to the usual tests of significance.
When one could specify sampling weights, the estimates of parameters,
frequencies and so on were often correct, but the variance estimates were
not, because they ignored the (sampling) design effects. Times have
changed, but SPSS has not. Stata's implementation with its svy suite of
estimators is a vast improvement and is to be preferred over anything that
SPSS produces.
To answer Ann's question(s), SPSS does not calculate the Pearson chi-squared
(or any other test) correctly when using weights. The test statistic itself
is subject to rounding error, since SPSS uses the weighted cell counts
rounded to the nearest integer rather than the unrounded values. This might
seem the appropriate method given the fact that only the rounded cell counts
are displayed anyway, however, this is only part of the problem. This
statistic is not distributed as chi-squared, which, I believe, is the
rationale behind the F test reported by Stata. So, don't use any of the
tests from SPSS.
> -----Original Message-----
> From: [email protected]
> [mailto:[email protected]]On Behalf Of Richard
> Herrell
> Sent: Monday, March 31, 2003 7:16 AM
> To: [email protected]
> Subject: RE: st: stata and spss output
>
>
> On Mon, 31 Mar 2003, Nick Winter wrote:
>
> > > -----Original Message-----
> > > From: Ernest Berkhout [mailto:[email protected]]
> > > Sent: Monday, March 31, 2003 10:05 AM
> > > To: [email protected]
> > > Subject: Re: st: stata and spss output
> > >
> > >
> > > At 10:24 31-3-2003, you wrote:
> > > > From
> > > >memory, SPSS was doing things incorrectly; it was using the weighted
> > > >sample size, rather than actual number of obs. (Stata got things
> > > >right.)
> > >
> > > There have been reports of SPSS sometimes 'ignoring' very low
> > > weighted
> > > frequencies... (I actually saw it doing that)
> > >
> >
> > At least under some conditions, SPSS seems to round weights to the
> > nearest integer.
> >
> > My experience leads me to say that SPSS should be avoided at all costs,
> > especially when you've got survey weights.
> >
>
> Indeed. I was forced to use SPSS to teach fellows in an intro epi program
> once. After presenting the 2x2 table, odds ratio, and why the OR is used
> so much in epidemiology, I moved on to the logistic model to discuss how
> adjusted ORs are obtained. When I came to the ta-DA moment demonstrating
> that the OR from the logistic model is identical to that from a 2x2 when
> only the single dichotomous predictor is entered in the model, I came up
> with a different OR in SPSS. After much exploring and consulting with
> more mathematically sophisticated friends, we found that the SPSS
> algorithms simply stop short of the last iteration in some
> circumstances.
> Since ad/bc IS the MLE, they should be identical. As this is one of the
> few situations where the precise calculation is easy, I've become wary of
> SPSS when there is no easy check.
>
> *
> * 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/
>
*
* 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/
