SPSSX Discussion

Weights in SPSS

Classic

List

Threaded

5 messages Options

Lightfoot Russell

Weights in SPSS

I have heard a lot of debate about whether or not SPSS handles weights
"correctly" vs. SAS. I have only found one procedure in SPSS where I
question how SPSS handled the weights. Could someone on the list provide
some guidance as to SPSS' handling of weights? Are there certain
procedures where the handling of weights is problematic? Are there any
good online sources that discuss SPSS and weights? Etc.
Thanks,

Russ Lightfoot

====================To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

Hector Maletta

Re: Weights in SPSS

SPSS provides frequency weights through the WEIGHT command. These are not
adequate for significance calculations, except in case of simple random
samples. For complex samples there are other specialized modules. There are
also some special programs for rim weighting.
For an introduction to weighting in SPSS, not covering all these topics but
the basics, you may want to look at my paper on weighting at Ray Levesque's
website on SPSSTools.
Hector
-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Lightfoot Russell
Sent: 06 August 2008 10:43
To: [hidden email]
Subject: Weights in SPSS

I have heard a lot of debate about whether or not SPSS handles weights
"correctly" vs. SAS. I have only found one procedure in SPSS where I
question how SPSS handled the weights. Could someone on the list provide
some guidance as to SPSS' handling of weights? Are there certain
procedures where the handling of weights is problematic? Are there any
good online sources that discuss SPSS and weights? Etc.
Thanks,

Russ Lightfoot

To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

Graham Wright-2

Re: Weights in SPSS

In reply to this post by Lightfoot Russell

The main problem I've had with population weights in SPSS
is that they don't allow for any sorts of statistical tests because they
ignore the size of the sample when calculating standard errors. If you
surveyed 100 people and then had weights to make the sample look like
10,000 people, then SPSS's stats tests (chi square, regressions, t-tests
etc) will treat the sample as if it were actually 10,000 people, and
thus the SE's and such will be absurdly small and every relationship
will be significant. So any sorts of tests or confidence intervals using
the weights in SPSS's base module are generally very wrong when using
population weights. SPSS has a separate survey module that (I'm told)
deals with them correctly, but I don't have a suitcase full of diamonds
lying around, so I uses Stata for anything involving weighting schemes.

-Graham Wright

Lightfoot Russell wrote:

> I have heard a lot of debate about whether or not SPSS handles weights
> "correctly" vs. SAS. I have only found one procedure in SPSS where I
> question how SPSS handled the weights. Could someone on the list provide
> some guidance as to SPSS' handling of weights? Are there certain
> procedures where the handling of weights is problematic? Are there any
> good online sources that discuss SPSS and weights? Etc.
> Thanks,
>
> Russ Lightfoot
>
> ===================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>

Hector Maletta

Re: Weights in SPSS

Graham,
That problem is analyzed in my paper on weighting at Ray Levesque's site,
SPSSTools. For simple random sampling, the solution is estimating
significance with a non-expansionary set of weights. This is easily created
by dividing your original weights by the total expanded population and
multiplying by the size of your sample, i.e. multiplying your original
weights by n/N where n=sample size and N=expanded sample size or population
size. With complex random samples (stratified and/or clustered this
procedure would give somewhat biased results: stratification reduces the
margin of error, while clustering increases it, so it all depends on the
balance of the two biases in your case. The SPSS Complex Samples component
gives you the possibility of estimating significance in these cases.

Hector

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
Graham Wright
Sent: 06 August 2008 12:37
To: [hidden email]
Subject: Re: Weights in SPSS

The main problem I've had with population weights in SPSS
is that they don't allow for any sorts of statistical tests because they
ignore the size of the sample when calculating standard errors. If you
surveyed 100 people and then had weights to make the sample look like
10,000 people, then SPSS's stats tests (chi square, regressions, t-tests
etc) will treat the sample as if it were actually 10,000 people, and
thus the SE's and such will be absurdly small and every relationship
will be significant. So any sorts of tests or confidence intervals using
the weights in SPSS's base module are generally very wrong when using
population weights. SPSS has a separate survey module that (I'm told)
deals with them correctly, but I don't have a suitcase full of diamonds
lying around, so I uses Stata for anything involving weighting schemes.

-Graham Wright

Lightfoot Russell wrote:

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

wsu_wright

Re: Weights in SPSS

It's not as elaborate as Hector's but will do the trick of minimizing the false signficance, create a relative weight:

(be sure the weight is off when creating the relative weight)

relative weight=current weight/(mean of current weight)

You can automate this:

compute aggflag=1.
agg
/break=aggflag
/curwgt_mn=mean(curwgt).
compute relwgt=curwgt/curwgt_mn.
weight by relwgt.

This will retain the sample size of the original sample but have the distribution pattern of the generalied weighted sample.

DW

---- Hector Maletta <[hidden email]> wrote:

> Graham,
> That problem is analyzed in my paper on weighting at Ray Levesque's site,
> SPSSTools. For simple random sampling, the solution is estimating
> significance with a non-expansionary set of weights. This is easily created
> by dividing your original weights by the total expanded population and
> multiplying by the size of your sample, i.e. multiplying your original
> weights by n/N where n=sample size and N=expanded sample size or population
> size. With complex random samples (stratified and/or clustered this
> procedure would give somewhat biased results: stratification reduces the
> margin of error, while clustering increases it, so it all depends on the
> balance of the two biases in your case. The SPSS Complex Samples component
> gives you the possibility of estimating significance in these cases.
>
> Hector
>
> -----Original Message-----
> From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
> Graham Wright
> Sent: 06 August 2008 12:37
> To: [hidden email]
> Subject: Re: Weights in SPSS
>
> The main problem I've had with population weights in SPSS
> is that they don't allow for any sorts of statistical tests because they
> ignore the size of the sample when calculating standard errors. If you
> surveyed 100 people and then had weights to make the sample look like
> 10,000 people, then SPSS's stats tests (chi square, regressions, t-tests
> etc) will treat the sample as if it were actually 10,000 people, and
> thus the SE's and such will be absurdly small and every relationship
> will be significant. So any sorts of tests or confidence intervals using
> the weights in SPSS's base module are generally very wrong when using
> population weights. SPSS has a separate survey module that (I'm told)
> deals with them correctly, but I don't have a suitcase full of diamonds
> lying around, so I uses Stata for anything involving weighting schemes.
>
> -Graham Wright
>
> Lightfoot Russell wrote:
> > I have heard a lot of debate about whether or not SPSS handles weights
> > "correctly" vs. SAS. I have only found one procedure in SPSS where I
> > question how SPSS handled the weights. Could someone on the list provide
> > some guidance as to SPSS' handling of weights? Are there certain
> > procedures where the handling of weights is problematic? Are there any
> > good online sources that discuss SPSS and weights? Etc.
> > Thanks,
> >
> > Russ Lightfoot
> >
> > ===================
> > To manage your subscription to SPSSX-L, send a message to
> > [hidden email] (not to SPSSX-L), with no body text except the
> > command. To leave the list, send the command
> > SIGNOFF SPSSX-L
> > For a list of commands to manage subscriptions, send the command
> > INFO REFCARD
> >
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>
> =====================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD