0
$\begingroup$

I am computing the significance of an overlap between two subgroups, each in two related datasets 1 and 2. For instance:

Dataset1 total: 500
Dataset1 subgroup: 100

Dataset2 total: 300
Dataset2 subgroup: 50

Intersection between subgroups: 25
Union 1 and 2 (no duplicates): 600

I want to compute how significant is the overlap between subgroups against what would be gotten by chance. How would you do this in Python? I was looking at Fisher's exact test or hypergeometric test but have some problems putting my data into the analyses.

From what I understand, the contingency table would be:

                Dataset1    Dataset2
In_subgroup     100         50
Not_subgroup    400         250
Total           500         300

Here, note that the universe is comprised by 600 unique elements, and not 500+300 (as there are duplicates within dataset1 and dataset2). Given this, and based on another post, I would do this in R:

phyper(24, 100, 500, 50, lower.tail = FALSE)
[1] 9.15e-19

Translating this into Python I would use in scipy:

>>> scipy.stats.hypergeom.cdf(24, 600, 500, 300)
0.0

Can I assume that the difference between both results is numeric error?

$\endgroup$
2
  • 1
    $\begingroup$ What do you mean by duplicates? $\endgroup$
    – Michael M
    Apr 15 '20 at 19:23
  • $\begingroup$ I mean that I have two populations, each with a subgroup. At the population level (and possibly at the subgroup) there are some terms in common between populations. What I want to test is whether the subgroup vs subgroup overlap is statistically significant, given each of the populations' compositions. $\endgroup$
    – Sos
    Apr 23 '20 at 8:17
1
$\begingroup$

Your proposed table does not seem right: First, because it is not clear how it counts overlaps. Second, because its grand total 800 does not match the total number 600 of subjects.

For a correct table, you cannot expect the P-value of fisher.test (2-sided as shown), to match results from phyper, which would be for a one-sided test.

Suppose you have 600 subjects, of which 150 are in subgroups 450 are not. Then I guess you have 25 + 175 = 200 subjects involved in overlaps. Perhaps you want to know if 25 'subgroup' subjects involved in overlaps (1/5 of them) is surprisingly small--compared to the 175 subjects outside of subgroups involved on overlaps ($175/450\approx .39$ of them/)

If so, then you want the data table shown below. If not, then please revise your Question to show a data table with the correct grand total, which is directly relevant to what you want to test. And explain how it is relevant.

In or Out \     Unique   Overlap       TOTAL
--------------------------------------------
Subgroup           125        25         150
Remainder          275       175         450
--------------------------------------------
TOTAL              400       200         600

DTA = rbind(c(175, 25), c(275,175));  DTA

     [,1] [,2]
[1,]  175   25
[2,]  275  175

fisher.test(DTA, alt="g")  # one-sided test

        Fisher's Exact Test for Count Data

data:  DTA
p-value = 1.493e-12
alternative hypothesis: 
   true odds ratio is greater than 1
...

The 2-sided test has P-value 2.381e-12. A two-sided chi-squared test without the Yates correction gives the following result.

chisq.test(DTA, corr=F)

        Pearson's Chi-squared test

data:  DTA
X-squared = 45.264, df = 1, p-value = 1.722e-11

Output from Minitab 17 statistical software (for comparison):

        C1   C2  All

1      125   25  150
2      275  175  450
All    400  200  600

Cell Contents:      Count

Pearson Chi-Square = 25.000, DF = 1, P-Value = 0.000
Likelihood Ratio Chi-Square = 27.225, DF = 1, P-Value = 0.000

Fisher’s exact test: P-Value =  0.0000003
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.