Advertisement
Boards Golf Society are looking for new members for 2022...read about the society and their planned outings here!
How to add spoiler tags, edit posts, add images etc. How to - a user's guide to the new version of Boards

Fishers exact query?

  • #1
    Closed Accounts Posts: 3 ajf103


    Hi there..

    Hope this is apt forum for this query...

    I am currently analysing my data for a research thesis and have run into a little problem that I cannont find the answer to in the literature/online etc..

    I am using SPSS version 18 for my analysis, currently I am looking at some categorical variables for associations. I have been using Chi-squared however several of my tests have violated the assumptions and I have had to revert to Fishers Exact (FET). Unfortunately some of my contingency tables are greater than 2x2 and I cannot find if this is allowed/appropriate with FET. I have found some older sources that suggested you could not do this with SPSS but I think that may have related to older versions of SPSS.

    In any regard, SPSS will run the test and always gives me a P-value, however it also gives me a value for the Fishers Exact test which I do not get with 2x2 tables (and as I understand it I should not receive a test result just a p-value). I am wondering if this is significant/an indication that the test is not appropriate, does anyone know? also if it is acceptable to run the test with >2x2 should I report these Fishers exact test values?

    In addition my supervisor is very keen for me to report effect size values, from the literature I understand you can obtain and report phi (2x2 tables) and cramer's V (>2x2 tables) values for effect size for Chi-square tests, can anyone confirm if this is correct? and can I also report these values for Fishers?

    any advice appreciated, thanks & regards.


Comments



  • ajf103 wrote: »
    Hi there..

    Hope this is apt forum for this query...

    I am currently analysing my data for a research thesis and have run into a little problem that I cannont find the answer to in the literature/online etc..

    I am using SPSS version 18 for my analysis, currently I am looking at some categorical variables for associations. I have been using Chi-squared however several of my tests have violated the assumptions and I have had to revert to Fishers Exact (FET). Unfortunately some of my contingency tables are greater than 2x2 and I cannot find if this is allowed/appropriate with FET. I have found some older sources that suggested you could not do this with SPSS but I think that may have related to older versions of SPSS.

    FET handles larger tables ok but is recommended for 2x2 (I'm assuming you chose this because of sample issues?). The question of using / not using Fisher wouldn't have anything to do with the edition of SPSS you are using the rules are the same (i.e. compensation for low cell counts). As long as you dont have a huge number of cells I think you're ok. Just mention in your reporting that you've encountered contradictory interpretations, and if in doubt, shove all relevant output into an appendix.
    ajf103 wrote: »
    In any regard, SPSS will run the test and always gives me a P-value, however it also gives me a value for the Fishers Exact test which I do not get with 2x2 tables (and as I understand it I should not receive a test result just a p-value). I am wondering if this is significant/an indication that the test is not appropriate, does anyone know? also if it is acceptable to run the test with >2x2 should I report these Fishers exact test values?

    Not sure what you mean - it doesn't give you the value of the test statistic in the output?
    ajf103 wrote: »
    In addition my supervisor is very keen for me to report effect size values, from the literature I understand you can obtain and report phi (2x2 tables) and cramer's V (>2x2 tables) values for effect size for Chi-square tests, can anyone confirm if this is correct? and can I also report these values for Fishers?

    Your choice of effect size measure is determined by the level of measurement of your variables as well as the rxc so you are fine. Just be aware that there are different ranges by which certain effect size scores are interpreted as either strong or weak.

    Not sure on the use of Cramers' for FET, as the formula for Cramer's V uses the value of the chi-square statistic, and phi is generally used for 2x2 tables. If you are reporting 2x2 tables it would probably be handier to use the odds ratio.




  • Hi efla, thanks very much for your input, it is very helpfull.

    Yes, I am using FET as some of my cell counts are <5.

    Sorry if I was not clear on the fishers exact test value. What I meant by this was that when using FET with a 2x2 table I do not get a fishers test value in the output, just a p-value, which the literature tells me is normal. However when I use FET with >2x2 tables I get a test value for Fishers, in addition to the p-value, which going by the books is unexpected (although no books I can find discuss using FET with >2x2 tables). I am not sure if I should ignore this value or report it as I would a chi-sq.

    I tried to put two output tables in the post to illustrate but the format went doolaly, anyway hope that makes a bit more sense?

    And I will look up odds ratio....

    Thanks again..




  • ajf103 wrote: »
    Hi efla, thanks very much for your input, it is very helpfull.

    Yes, I am using FET as some of my cell counts are <5.

    Sorry if I was not clear on the fishers exact test value. What I meant by this was that when using FET with a 2x2 table I do not get a fishers test value in the output, just a p-value, which the literature tells me is normal. However when I use FET with >2x2 tables I get a test value for Fishers, in addition to the p-value, which going by the books is unexpected (although no books I can find discuss using FET with >2x2 tables). I am not sure if I should ignore this value or report it as I would a chi-sq.

    I tried to put two output tables in the post to illustrate but the format went doolaly, anyway hope that makes a bit more sense?

    And I will look up odds ratio....

    Thanks again..

    Ok, I see. Fisher's formula doesn't work like a typical test statistic - it calculates the probability of observing such results under the null hypothesis of equality. The formula works the probabilty so all you need to report is that (the usual p≤0.05*)




  • Great, will do so thanks for your help...


Advertisement