Home > Statistics Every Writer Should Know > The Stats Board > Discusssion
i think my prof's wrong, so who's right???
Message posted by sue (via 137.132.3.5) on May 7, 2001 at 10:12 PM (ET)
i'm doing some research for my prof. it entails measurement of quality of life scores in patient populations.
test-retest was done and we did a correlation analyis between scores on the first occasion and those on the second occasion.
my prof says that i should use pearson cos it's commonly used. and a linear relationship is expected
i think spearmans is more appropriate cos the scale is ordinal.
who's right?
READERS RESPOND:
(In chronological order. Most recent at the bottom.)
Spearman vs. Pearson
Message posted by nancy diehl (via 129.176.151.121) on May 8, 2001 at 8:44 AM (ET)
I have to agree with you. Quality of Life data is typically ordinal data bounded by responses of values between say 1 to 7. This is not continuous data where a measure of the linear correlation between x and y would commonly use the Pearson correlation coefficient. You want a rank correlation to provide a measure of the degree of linearity between the ranking (QofL variables) variables. A rank correlation coefficient, or a coefficient of agreement for preference data would be either Spearman's or Kendall's.
Re: i think my prof's wrong, so who's right???
Message posted by Bill (via 192.231.71.108) on May 8, 2001 at 1:23 PM (ET)
I've always wondered why we use a Pearson (or Spearman) to measure "reliability." The lay definition of reliability would be something like: if the person scores x today he will tend to score x on retest if the test is reliable. We then proceed to correlate the pre and post measure with a Pearson to test for the "repeatability of scores." However, as I understand the Pearson, it indicates "pattern" agreement, not absolute agreement. Consider the follow pre/post measures on a 1-9 QOL scale for three patients: 3-9, 2-8, 1-7. A Pearson will indicate perfect agreement (r=1.0) but there is obvious disagreement from the pre to post testing. It seems that repeatability of scores should be measured as absolute agreement (with an Intra-class correlation coefficient) but that's not how we do it. Perhaps I don't understand what is meant by reliability.
Re: i think my prof's wrong, so who's right???
Message posted by sue (via 203.124.2.34) on May 9, 2001 at 9:40 AM (ET)
thanks for the comments.
to the second person who wsa wondering why spearman/pearson is used if it indicates pattern,
there is another way to do the reliability.that is to plot difference in test-retest by mean scores on the 2 settings.
you should get a horizontal line if it's correct.
Re: i think my prof's wrong, so who's right???
Message posted by Phil (via 216.175.114.127) on May 9, 2001 at 11:12 AM (ET)
Don't you have to have two ranked lists to use Spearmans? I doesn't sound to me like that is the type of data you have. If you have paired data (before and after scores), then Pearsons should work. How about a paired t-test to test the hypothesis that there was no difference in before and after scores?
Re: i think my prof's wrong, so who's right???
Message posted by Bill (via 192.231.71.108) on May 9, 2001 at 1:03 PM (ET)
Concerning Sue's suggestion of plotting the differences between pre and post tests (for each subject I assume), if the scores of the three subjects were 3-9, 3-, and 3-9 then the plot would be -6 for each subject and horizontal but you still have "substantial" disagreement between the pre and post test measure. If you use a t-test and the data were (for 4 subjects) 3-9, 9-3, 3-9, and 9-3 then the t-test would show no difference but there is still disagreement. That's why I assert that "reliability" in a test-retest situation is simply a question of agreement (assessed in this situation by an intra-class correlation coefficient). But the gods of psychometrics don't agree and simply use a Pearson correlation. In the course of life it's not a big deal although most who measure reliability in this manner probably don't know what they are actually measuring. Just food for thought.
Your $5 contribution helps cover part the $500 annual cost of keeping this site online.