i wanted to find out the association between two variables and it reduced to 6X4 contingency table. grand total is 103. the spss output of chi square test displayed "82.1% cells have expected count less than 5". test shows as significant by the p value. is this reliable enough to reject the null hypothesis? or how to make it reliable? can i used monte carlo confidence intervals to check the pa value? please guide me. thanks
$\begingroup$
$\endgroup$
5
-
$\begingroup$ It means most of your cells have very small counts. Your test might not be reliable. Try to combine some groups. $\endgroup$– SmallChessCommented Jul 19, 2017 at 6:01
-
$\begingroup$ It's hard to tell from the information you give. The rule of thumb that specifies expected counts greater than 5 is generally quite a bit too strict. If the expected counts are fairly uniform you can go considerably lower. You may be okay. If you tell us the marginal counts it might be possible to make some assessment of the extent to which it might matter. [If you hadn't already seen the p-value, you might instead have considered a G test of the same hypothesis, it's slightly less sensitive to small expected counts; you can compute expected values from only the table margins.] $\endgroup$– Glen_bCommented Jul 19, 2017 at 6:26
-
1$\begingroup$ Essentially the same question has been asked and answered previously, see stats.stackexchange.com/questions/34703 . Just use Fisher's exact test, which is readily available in SPSS and which does not rely on asymptotics. $\endgroup$– Gordon SmythCommented Jul 19, 2017 at 8:12
-
1$\begingroup$ Fisher's exact test is always safe but is somewhat conservative. If you're worried about power, then try Frank Harrell's suggestions: stats.stackexchange.com/questions/14226 $\endgroup$– Gordon SmythCommented Jul 19, 2017 at 10:54
-
$\begingroup$ thanks for all reponses. Can I use montecarlo confidence intervals for the analysis because that output consists of the p value of exact $\endgroup$– V AbeygunawardenaCommented Jul 20, 2017 at 8:20
Add a comment
|