I have a 2x2 table of single nucleotide substitutions in DNA. The data rows represent the nucleotides in one genome (A) and columns are the nucleotides in the same positions of a different genome (B). The numbers are the times such substitution occurred.

        A or T   C or G
A or T  3        12
C or G  12       3

I want to analyse whether these mutations are random. The problem is that since any "substitutions" to the same base (eg A->A) are not recorded you would expect to have twice as many substitutions from A/T to C/G than from A/T to A/T etc. The expected values for n=30:

        A or T   C or G
A or T  5        10
C or G  10       5

Tests like chi-squared or Fisher's exact test determine if there are nonrandom associations between rows and columns. But I know that there is an association anyway and want to test whether the data is significantly different from it.

What statistical tests would be appropriate?

When performing chi-squared for independence is it possible to use specific expected values instead of using row and column totals?

Would it be correct to transform the data by dividing the A/T->C/G and C/G->A/T cells by 2, to get the mean number of substitutions to the other base category? The n would then change.


1 Answer 1


In the formula of chi-square, for each observed frequency you:

-- 1) take the difference between the observed frequency and its expected one;

-- 2) square the above difference;

-- 3) divide the above squared difference by the expected frequency of point 1;

-- 4) sum all the results (of points 1-3 applied to every cell in your contingency table) to obtain the chi-square statistic.

Therefore, in the formula, you can set a different expected frequency for each cell in your contingency table and test the hypothesis you need.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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