spearman's rank correlation + finite population correction
Locked
spearman's rank correlation + finite population correction
 
|
Dear SPSS community, What is an SPSS solution to doing the following? 1. Run spearman's rank correlations for ordinal variables 2. Adjust the sample for
Any pointers appreciated! Thank you so much! Shirin Shirin Nuesslein Research Assistant - Health, Environment and Indigenous Communities Research Group Administrative Coordinator - Nasivvik Centre for Inuit Health and Changing Environments Master's Student - Sustainability Studies Program at Trent University Trent University - Indigenous Studies Department 1600 West Bank Drive - Peterborough, ON, K9J 7B8 Phone: 705-748-1011 ext 7242 Email: [hidden email] |
Re: spearman's rank correlation + finite population correction
|
|
Do you really have a problem that justifies using the FPC, or are you
looking at what seems to be a magical way of getting a test that looks
significant? I think I have seen one problem /deserving/ FPC in 25 years of
answering questions, so I am curious if you can show me a second one.
IIRC - The Spearman rank coefficient is exactly the Pearson when r is
computed on the (properly) ranked scores: List all N cases, including
ties, with ranks 1 to N; re-assign each case with the average rank for all
its ties. RANK offers that.
Then you can apply weights and get a weighted r.
Using the rank-transform is not as desirable as using transforms that
create and preserve "equal intervals" - so I tend to be hostile to facile use
of rank-transforms.
The FPC is for testing and is very seldom appropriate in real life except
for deciding when a close election can be "called" by the TV networks.
I once saw an artificial example suggested in wildlife management where
the FPC could be useful in research. I forget what the "real" example was
that came up in one of the internet groups.
I don't remember any SPSS procedure that incorporates FPC, so I expect you
would have to compute it yourself from the F-test table for the R-squared,
for instance, if you compute the r by using multiple regression. It is probably
a good thing that SPSS does not provide it, because that must vastly limit the
number of people who otherwise would mis-apply it. For the same reason,
I will not spell out more precisely how to compute it -- to save myself from
blame for misuse. I worry that I may have already said too much.
--
Rich Ulrich
From: SPSSX(r) Discussion <[hidden email]> on behalf of Shirin Nuesslein <[hidden email]>
Sent: Wednesday, October 21, 2020 10:28 PM To: [hidden email] <[hidden email]> Subject: spearman's rank correlation + finite population correction Dear SPSS community,
What is an SPSS solution to doing the following?
1. Run spearman's rank correlations for ordinal variables
2. Adjust the sample for
Any pointers appreciated!
Thank you so much!
Shirin
Shirin Nuesslein
Research Assistant - Health, Environment and Indigenous Communities Research Group
Administrative Coordinator - Nasivvik Centre for Inuit Health and Changing Environments Master's Student - Sustainability Studies Program at Trent University Trent University - Indigenous Studies Department
1600 West Bank Drive - Peterborough, ON, K9J 7B8
Phone: 705-748-1011 ext 7242
Email: [hidden email]
|
Re: spearman's rank correlation + finite population correction
|
In reply to this post by Shirin Nuesslein
Do you have actual ranks, i.e., just about as many values as there are cases?
If not what is the response scale? What is the population you want to generalize to? How many cases are there in the population? How many cases in your data? Were the cases selected by a scientific sampling process? ----- Art Kendall Social Research Consultants -- Sent from: http://spssx-discussion.1045642.n5.nabble.com/ ===================== 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
Social Research Consultants |
Re: spearman's rank correlation + finite population correction
|
|
Did you receive the data as ranks or did you rank it?
----- Art Kendall Social Research Consultants -- Sent from: http://spssx-discussion.1045642.n5.nabble.com/ ===================== 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
Social Research Consultants |
Re: spearman's rank correlation + finite population correction
|
|
Thank you both for the responses! Lots of food for thought. Perhaps I can offer some context to my study: 1. My data is ordinal (not ranked), and according to my textbook it should be appropriate to use the Spearman's Rank Correlation for ordinal data. I am now questioning whether it's appropriate and whether a Biseral Correlation (one dichotomous, one interval variable) is better. I just learned about this test. Would you have any advice on which of the Biseral Correlation types would be a better fit for my data? Some of my variables of interest are: personal income (<$20,000, $21-40,000, $41-60,000, $61,000 plus) education status (no formal schooling, less than secondary school completed, secondary school completed, more than secondary school) – across food security levels (food secure, food insecure) 2. Regarding the Finite Population Correction, I am working with a small dataset (n=138). However, the sample is small because the total population of interest is also small (N=294). I am working in a small remote community context and I know the exact number of the group of people of interest. I thought the FPC would be appropriate in this context to adjust for error. I have a small sample yet technically I sampled 47% of the population, which is a decent size. From your experience with FPC, what other things do I need to consider to decide whether it is appropriate to apply it? Any pointers or resources appreciated! Warmly, Shirin On Thu, Oct 22, 2020 at 9:27 AM Art Kendall <[hidden email]> wrote: Did you receive the data as ranks or did you rank it? Shirin Nuesslein Research Assistant - Health, Environment and Indigenous Communities Research Group Administrative Coordinator - Nasivvik Centre for Inuit Health and Changing Environments Master's Student - Sustainability Studies Program at Trent University Trent University - Indigenous Studies Department 1600 West Bank Drive - Peterborough, ON, K9J 7B8 Phone: 705-748-1011 ext 7242 Email: [hidden email] |
Re: spearman's rank correlation + finite population correction
|
|
How did you select respondents?
Do you have a map to visualize the population of housing units? This might help you get a handle on the representativeness of your set of respondents. How did you operationalize the construct of food insecurity? If food insecurity was gathered in a less coarse measurement, I would use the less coarse measurement. I would start with a 3D scatterplot and in the viewer of the output file look at it from different angles. It would make the scatterplot more readable if you use distinct colors and marker shapes. Since your measures are coarse, use large markers to remind your self of the "around here" nature of your values. Before you run the graphic think about what patterns might occur. Ask your self if there appears to be any pattern. Is it really clear to you? Then use CATREG and test whether it makes any meaningful fit difference to use ordinal vs interval assumptions. Uncertainty includes: -- How the cases were selected. This is problematical when cases are not selected by scientific sampling methods. -- How the constructs are measured. Validity, reliability, and coarsness all effect this. If you do decide that the fit is inconsistent with sampling and measurement error, then you need to apply the SO WHAT test. Finite pop corrections change the standard error, hence the width of confidence intervals and levels of significance. If you see a good fit without the fpc, why bother with the fpc? The fpc might *narrow *your idea of how far off the fit line might be and if you want to go the effort go to If https://www.ibm.com/support/pages/confidence-intervals-correlations It would be nice for you to report back to the list difference you see in the confidence interval and statistical significance ----- Art Kendall Social Research Consultants -- Sent from: http://spssx-discussion.1045642.n5.nabble.com/ ===================== 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
Social Research Consultants |
Re: spearman's rank correlation + finite population correction
|
|
should have included this. The fpc is .7 ish
Use the suggested syntax twice. First, as it is. compute sez = 1/sqrt(n-3). Then add 2 lines so you have compute sez = 1/sqrt(n-3). compute fpc = Sqrt((294-138)/(293)). compute sez = sez*fpc, ----- Art Kendall Social Research Consultants -- Sent from: http://spssx-discussion.1045642.n5.nabble.com/ ===================== 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
Social Research Consultants |
Re: spearman's rank correlation + finite population correction
|
|
In reply to this post by Shirin Nuesslein
1a. Using Spearman can give you a correlation that is little
influenced by extreme outliers in a set of continuous scores.
Your grouped data has no extremes to correct; but it has /ties/;
ties undermine the precision of using the small-sample tables for p.
If you replace every score by its rank-transformation, there will be
at least a slight change in the spacing of the group scores. Does that
change create a clear improvement on the "intervals" between levels?
Backing up to assumptions: the groups are presumably created with
an eye to "equal intervals" -- which is the assumption for ordinary
ANOVA, etc. Here is why the 1950s fad for "nonparametric analyses"
properly faded: the rank-transformed spacing is seldom a clear improvement,
and ranking throws out the anchor labels that help us interpret the means.
1b. One version of the biserial is exactly equal to the Pearson r.
The other one gives an estimator that adjusts for extreme proportions,
and that is useful mainly when you want to use the r as a meaningful
"effect size" (which you don't). It does nothing to change the inherent
p-level of the table. A different p probably means a bad error estimation.
2. What is the purpose of your research? If you are creating
"administrative" options for this set of people ("What can the government
assume about them? How precisely can we describe them?"), FPC may
be relevant. On the other hand, if this is intended to be of interest to
a scientific community, concerning cause and effect, your implicit
"population" comprises the thousands/millions of people who may
live in similar circumstances, and whom you are generalizing to.
In other words, what is the conclusion that you may draw? (based,
of course, on the assumption of random sampling...).
--
Rich Ulrich
From: Shirin Nuesslein <[hidden email]>
Sent: Monday, October 26, 2020 11:12 PM To: Art Kendall <[hidden email]>; Rich Ulrich <[hidden email]> Cc: [hidden email] <[hidden email]> Subject: Re: spearman's rank correlation + finite population correction Thank you both for the responses! Lots of food for thought.
Perhaps I can offer some context to my study:
1. My data is ordinal (not ranked), and according to my textbook it should be appropriate to use the Spearman's Rank Correlation for ordinal data. I am now questioning whether it's appropriate and whether a Biseral Correlation (one dichotomous, one interval
variable) is better. I just learned about this test. Would you have any advice on which of the Biseral Correlation types would be a better fit for my data?
Some of my variables of interest are:
personal income (<$20,000, $21-40,000, $41-60,000, $61,000 plus)
education status (no formal schooling, less than secondary school completed, secondary school completed, more than secondary school)
– across food security levels (food secure, food insecure)
2. Regarding the Finite Population Correction, I am working with a small dataset (n=138). However, the sample is small because the total population of interest is also small (N=294). I am working in a small remote community context and I know the exact
number of the group of people of interest. I thought the FPC would be appropriate in this context to adjust for error. I have a small sample yet technically I sampled 47% of the population, which is a decent size.
From your experience with FPC, what other things do I need to consider to decide whether it is appropriate to apply it?
Any pointers or resources appreciated!
Warmly,
Shirin
On Thu, Oct 22, 2020 at 9:27 AM Art Kendall <[hidden email]> wrote:
Did you receive the data as ranks or did you rank it? Shirin Nuesslein
Research Assistant - Health, Environment and Indigenous Communities Research Group
Administrative Coordinator - Nasivvik Centre for Inuit Health and Changing Environments Master's Student - Sustainability Studies Program at Trent University Trent University - Indigenous Studies Department
1600 West Bank Drive - Peterborough, ON, K9J 7B8
Phone: 705-748-1011 ext 7242
Email: [hidden email]
|
«
Return to SPSSX Discussion
|
1 view|%1 views
| Free forum by Nabble | Edit this page |