Date: Fri, 23 Sep 2005 15:07:15 -0300
Reply-To: Hector Maletta <hmaletta@fibertel.com.ar>
Sender: "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From: Hector Maletta <hmaletta@fibertel.com.ar>
Subject: Re: Fractional weights - does SPSS round off?
In-Reply-To: <200509231754.j8NGNKrV022425@malibu.cc.uga.edu>
Content-Type: text/plain; charset="iso-8859-1"
SPSS does not round off the weights, but it rounds off the resulting
FREQUENCIES.
If the weighted number of cases is not an integer, it is rounded to the
nearest integer.
Instead, if the weighted total of some variable (e.g. total number of
children for all women in the sample) has decimals, then the total is
reported with the decimals. Sometimes these decimals make sense (total
income, for instance) sometimes they do not (as in total children). However,
both cases and totals, as well as averages and other measures, can be
DISPLAYED without the decimals, even if internally SPSS keeps those figures
with the full complement of decimal places.
As said before, the fractional weights are used without rounding. Each case
is multiplied by its weight before entering it into a computation. This may
cause distortion only for SMALL weights.
When the integer part of the weights is large relative to the decimals
(weights, for instance, like 150.45 or more), rounding the decimal part is
not very important. Instead, when weights are small numbers, where the
integer part is small relative to the decimals (weights like 0.80 or 1.45),
the rounding can cause considerable distorsion. For instance, one case
weighing 150.45 will be counted as 150 cases, and another weighting 80.45
will be counted as 80 cases, whereas two case weighting 1.45 and 0.80 would
be counted as 1 case each, thus defeating the purpose of giving one of them
more weight than the other.
Weights are likely to be small when they are simply PROPORTIONAL weights,
preserving the sample size but correcting for instance the proportions of
age groups or genders. When weights are also SCALE weights (expanding the
sample to population size) they are likely to be large figures, where the
decimal part is relatively irrelevant.
Hector
> -----Original Message-----
> From: SPSSX(r) Discussion [mailto:SPSSX-L@LISTSERV.UGA.EDU]
> On Behalf Of David Chapman
> Sent: Friday, September 23, 2005 2:55 PM
> To: SPSSX-L@LISTSERV.UGA.EDU
> Subject: Fractional weights - does SPSS round off?
>
> List members:
>
> I have a question about weights and how SPSS uses fractional
> weight values... I am working with a large database where the
> agency has weights for national level data. Weights are not
> integer values; they are fractional figures. For example, a
> weight may be be something like 595.08 or 2477.90 (rounded).
>
> SPSS Help says: "Fractional values are valid; they are used
> exactly where this is meaningful and most likely where cases
> are tabulated."
>
> Hmmmm... "...meaningful..." can mean a lot of things.
> Operationally, does SPSS actually use this properly in
> various forms of regresson (like logistics or ordinal) or
> would it round off? If the latter, under what conditions
> might SPSS do so?
>
> This issue came up because of something I'm proposing to do
> with relative weights; Hector Maletta has had numerous posts
> (over the years!) regarding relative weighting and I'm trying
> that with the data set in question. The downsampled weights
> become values as low as .03 and a high as 3.6
> (rounded) but I've been questioned about SPSS' rounding-off
> during statistical analysis, particularly the lower values
> (where the belief is that SPSS may drop out the low values
> for example).
>
> My belief is that the only problem may be a roundoff problem
> in calculating the weights (because of precision); however,
> there is a contention that the roundoff may be because of
> cases dropout when SPSS thinks they are 0 (i.e., could a
> value of 0.03 be seen as a "nothing"
> case?).
>
> Any insights would help...
>
> Many thanks!
>
> David Chapman
>
> __________ Información de NOD32 1.1230 (20050922) __________
>
> Este mensaje ha sido analizado con NOD32 Antivirus System
> http://www.nod32.com
>
>
| 
