Ordinary Least Square and use of Weight By

[Marsha and Mike SZYMCZUK]

I do not have access to the Complex Samples module.

I am working with a stratified sample from a national database with pre-defined weights (cannot change them).

Previous researchers have applied the Weight By command to the sample to adjust for sample bias.  This changes N of the sample from  about 1,000 to about 24,000.

When computing an OLS model ( s ) on the expanded sample, how can I adjust or test for incorrect error estimates?

Thanks.

Mike S.

====================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

[Brandon Paris]

Hi,

I may be wrong here, but here is how I understand it.

a) I think the more typically-derived weights (like the ones you describe that inflate or deflate your sample size) REQUIRES the complex samples module.  There isn't any way around this that I am aware of that produces the correct results.

b) The issue relates to degrees of freedom -- SPSS is assuming you have more DF than you do in reality.  If you know how, you can recompute the degrees of freedom by hand, which then feed into the analysis fairly straightforwardly.

c) The Weight By command is best utilized when the weights are centered around 1.0.  This allows for RELATIVE weighting to occur, but does not adjust the TOTAL sample count (so your statistical tests should be still appropriate).  You can achieve this by dividing your individual weights by the average weighting across all respondents.

Good luck.

Thanks,
Brandon Paris


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Marsha and Mike SZYMCZUK
Sent: Thursday, September 11, 2008 10:11 PM
To: [hidden email]
Subject: Ordinary Least Square and use of Weight By

I do not have access to the Complex Samples module.

I am working with a stratified sample from a national database with pre-defined weights (cannot change them).

Previous researchers have applied the Weight By command to the sample to adjust for sample bias.  This changes N of the sample from  about 1,000 to about 24,000.

When computing an OLS model ( s ) on the expanded sample, how can I adjust or test for incorrect error estimates?

Thanks.

Mike S.

=======
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

[Art Kendall]

Without the complex sampling module, SPSS uses the weights that you supply.
You might use expansion weights where each case "stands for" w cases in
the pop (24000 in OP)or adjusted weights where w "stands for" an
adjusted sample size (1000 in OP). SPSS does not assume that your
weights are expansion weights.  The person who provides the weights
decides that.
In either instance your point estimates ( means, r, R,  Bs, Betas) will
be ok. Except that for the estimated total with adjusted weights will be
the total for that portion of the pop that is in the sample. (multiply
by N/n to get the pop total).  In the OP, I assume that previous
researchers were interested in totals more than relations among variables.
If you use expansion weights, interval estimates (SDs, SEs )and
therefore tests of significance  will be based on too many df.

If all of your design factors are stratifications, the interval
estimates will be too *large*, and so your tests will show as
significant less often than they should
If all of your design factors are clusterings,      the interval
estimates will be too *small*, and so your tests will show as
significant more often than they should.

Art Kendall
Social Research Consultants

Brandon Paris wrote:

> Hi,
>
> I may be wrong here, but here is how I understand it.
>
> a) I think the more typically-derived weights (like the ones you describe that inflate or deflate your sample size) REQUIRES the complex samples module.  There isn't any way around this that I am aware of that produces the correct results.
>
> b) The issue relates to degrees of freedom -- SPSS is assuming you have more DF than you do in reality.  If you know how, you can recompute the degrees of freedom by hand, which then feed into the analysis fairly straightforwardly.
>
> c) The Weight By command is best utilized when the weights are centered around 1.0.  This allows for RELATIVE weighting to occur, but does not adjust the TOTAL sample count (so your statistical tests should be still appropriate).  You can achieve this by dividing your individual weights by the average weighting across all respondents.
>
> Good luck.
>
> Thanks,
> Brandon Paris
>
>
> -----Original Message-----
> From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Marsha and Mike SZYMCZUK
> Sent: Thursday, September 11, 2008 10:11 PM
> To: [hidden email]
> Subject: Ordinary Least Square and use of Weight By
>
> I do not have access to the Complex Samples module.
>
> I am working with a stratified sample from a national database with pre-defined weights (cannot change them).
>
> Previous researchers have applied the Weight By command to the sample to adjust for sample bias.  This changes N of the sample from  about 1,000 to about 24,000.
>
> When computing an OLS model ( s ) on the expanded sample, how can I adjust or test for incorrect error estimates?
>
> Thanks.
>
> Mike S.
>
> =======
> 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