I have N 2x2 observed contingency tables. Each one was sampled from the same distribution. The problem is the column labels are obfuscated so that I do not know the true label name of each column, and in some contingency table the first column refers to the same label that in a different contingency table is the second column.

If I had one contingency table, I would not care that the column labels are obfuscated in order to check independence between the variables, I can still calculate the chi-squared statistic and run the chi-squared test.

But since I have N contingency tables with the columns maybe permuted and no way to discern the column labels, I cannot build one big contingency table.

Would it make sense here to calculate the chi-squared statistic for each table separately then take the sum of all of these and calculate a p-value on that?

It seems to make sense but I want to make sure I am not missing something obvious


In fact that is a common technique in meta-analysis of significance values. The original approach by Fisher transforms the $p$-values to a $\chi^2_2$ because there is an easy formula for that but Lancaster suggested doing it for any number of degrees of freedom. So, yes, just add them up and you have a $\chi^2_k$ where you had $k$ tables.

  • $\begingroup$ ah great! Thanks. Do you have a source explaining this technique? $\endgroup$ Jun 30 at 6:07
  • 1
    $\begingroup$ I think it is in Lancaster, H O in Biometrika 1949 36 370-382 but it is quite an intensive paper and I do not have the time right now to read it in detail. $\endgroup$
    – mdewey
    Jun 30 at 12:27
  • $\begingroup$ Thanks! I'll try to find it. $\endgroup$ Jul 1 at 6:44

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.