Re: Inflated N's in Spearman's Rho
Posted by
Rick Oliver-3 on
Apr 04, 2014; 9:15pm
URL: http://spssx-discussion.165.s1.nabble.com/Inflated-N-s-in-Spearman-s-Rho-tp5725316p5725317.html
If you are using the WEIGHT command to
weight cases, the issue is fractional weights. From the documentation on
the WEIGHT command:
Weight values do not need to be integer, and some procedures,
such as FREQUENCIES, CROSSTABS,
and CTABLES, will use
fractional values on the WEIGHT variable.
However, most procedures treat the WEIGHT variable
as a replication weight and will simply round fractional weights to the
nearest integer. Some procedures ignore the WEIGHT
variable completely, and this limitation is noted
in the procedure-specific documentation. Procedures in the Complex Samples
add-on module can use inverse sampling weights specified on the CSPLAN
command.
If you are using NONPAR CORR to compute
Spearman's Rho, it is rounding the weights. The WEIGHT command is not the
ideal solution for fractional weights. I think the weight options in the
Complex Sample add-on module are more robust.
Rick Oliver
Senior Information Developer
IBM Business Analytics (SPSS)
E-mail: [hidden email]
From:
Adam Troy <[hidden email]>
To:
[hidden email],
Date:
04/04/2014 03:51 PM
Subject:
Inflated N's
in Spearman's Rho
Sent by:
"SPSSX(r)
Discussion" <[hidden email]>
Hi all,
Quick question (I did look for an answer and could not
find one). I have a dataset from a nationally representative sample with
a column for the weighting of each case which weights the sample to 1,000
total (original sample was 1,003). Weights range from .25 to 5.48 for each
case.
When I was running Spearman's correlations (Spearman's
rho) with weight cases on I noticed that the N's were being inflated above
1,000 in all cases. For example, in pairs where the pairwise N was 975,
the N for the Spearman's rho was 1041. This was not the case when I ran
these chi squares with weights applied, which should yield the exact same
p value for a comparison of two binary variables. The results also differed
from identical weighted Spearman correlation results in SAS.
Any ideas why SPSS was inflate these N's unnecessarily
for weighted Spearman correlations? I will be encountering a lot of weighted
samples and would like to be able to run all these analyses in SPSS. I'm
currently using SPSS 21.
Thanks again,
Adam