Google Groups no longer supports new Usenet posts or subscriptions. Historical content remains viewable.
Dismiss
Adding Weights
37 views
Skip to first unread message
Joe
Jan 23, 2012, 3:19:14 PM1/23/12
to
The sample of a data set of a survey of customers I have needs to be
weighted to reflect the ages of known actual ages. I am clearly under
in the 18-34 cohort and way over in the 55+
Here are the actual age percentages and those of the sample we have
collected:
Actual Customer Ages:
18-34--21%
35-54--34%
55+--45%
Age of Customer Sample:
18-34--3%
35-54%--31%
55+--66%
I know to go to the weight cases tab and to enter age as the variable
I want to weight by, But I am not clear on what to do next.
Can someone please advise?
Thanks,
Joe Ferrara
Market Research Manager
North Jersey Media Group
weighted to reflect the ages of known actual ages. I am clearly under
in the 18-34 cohort and way over in the 55+
Here are the actual age percentages and those of the sample we have
collected:
Actual Customer Ages:
18-34--21%
35-54--34%
55+--45%
Age of Customer Sample:
18-34--3%
35-54%--31%
55+--66%
I know to go to the weight cases tab and to enter age as the variable
I want to weight by, But I am not clear on what to do next.
Can someone please advise?
Thanks,
Joe Ferrara
Market Research Manager
North Jersey Media Group
David
Jan 23, 2012, 4:37:49 PM1/23/12
to
"> I know to go to the weight cases tab and to enter age as the
variable
> I want to weight by, But I am not clear on what to do next."
--
variable
> I want to weight by, But I am not clear on what to do next."
A little knowledge is a very dangerous thing!!!
You **DO NOT** want to use your age variable as a WEIGHT!
You need to create a variable which will weight your 18-34 up and the
remaining groups down and weight by that variable.
Assuming you wish to maintain the same effective sample size as what
you have actually sampled then using Pct desired/Pct Observed works.
YMMV WRT standard errors and such.
----
RECODE AGE (18 THRU 34=1)(35 THRU 54=2)(55 THRU HI=3) INTO AGEGROUP.
FREQ AGEGROUP.
* Weights computed as 21/3, 34/31, 45/66 *.
RECODE AGEGROUP (1=7)(2=1.09677419354839)(3=.68181818181818) INTO
CASEWGT.
WEIGHT BY CASEWGT.
FREQ AGEGROUP.
AGEGROUP
Frequency Percent Valid Percent Cumulative Percent
Valid 1.00 21 21.0 21.0 21.0
2.00 34 34.0 34.0 55.0
3.00 45 45.0 45.0 100.0
Total 100 100.0 100.0
*Alternatively!!!*.
DO IF AGEGROUP EQ 1.
+ COMPUTE WEIGHT=21/3.
ELSE IF AGEGROUP=2.
+ COMPUTE WEIGHT=34/31.
ELSE.
+ COMPUTE WEIGHT=45/66.
END IF.
WEIGHT BY WEIGHT.
FREQ AGEGROUP.
Rich Ulrich
Jan 23, 2012, 9:08:49 PM1/23/12
to
On Mon, 23 Jan 2012 13:37:49 -0800 (PST), David
<david...@gmail.com> wrote:
[snip]
In particular - when you have weighted one group as 7 times the
actual N, it is pretty safe to say that you should not report any
test that reflects extreme means for that group.
If you not intending to report tests, then that's fine and you have
no problem. If you do want tests, then -- especially given the
large discrepancy for one group between sample and population --
I don't know how others have handled it. If I were telling the
story, I think I would do tests on both raw data and weighted
data, and then I would be careful to contrast and compare those
results, giving suitable warnings where the raw data do not
support conclusions that are implied by the weighted data.
--
Rich Ulrich
<david...@gmail.com> wrote:
[snip]
>Assuming you wish to maintain the same effective sample size as what
>you have actually sampled then using Pct desired/Pct Observed works.
>YMMV WRT standard errors and such.
[snip rest]
>you have actually sampled then using Pct desired/Pct Observed works.
>YMMV WRT standard errors and such.
In particular - when you have weighted one group as 7 times the
actual N, it is pretty safe to say that you should not report any
test that reflects extreme means for that group.
If you not intending to report tests, then that's fine and you have
no problem. If you do want tests, then -- especially given the
large discrepancy for one group between sample and population --
I don't know how others have handled it. If I were telling the
story, I think I would do tests on both raw data and weighted
data, and then I would be careful to contrast and compare those
results, giving suitable warnings where the raw data do not
support conclusions that are implied by the weighted data.
--
Rich Ulrich
Jon Peck
Jan 24, 2012, 9:17:47 AM1/24/12
to
What about other dimensions of representativeness? Is your dataset reasonably representative in other dimensions where you have control totals? Were some groups deliberately over/under sampled? If there was a sampling plan, consider using the procedures in the Complex Samples option to adjust for it.
From a testing viewpoint, you don't want to overstate the N - make sure that the weighted and unweighted total N's are the same.
You can adjust the sample proportions in up to 10 dimensions using the SPSSINC RAKE extension command available from the SPSS Community website (www.ibm.com/developerworks/spssdevcentral). The prerequisites can also be obtained there.
From a testing viewpoint, you don't want to overstate the N - make sure that the weighted and unweighted total N's are the same.
You can adjust the sample proportions in up to 10 dimensions using the SPSSINC RAKE extension command available from the SPSS Community website (www.ibm.com/developerworks/spssdevcentral). The prerequisites can also be obtained there.
0 new messages