### Post hoc Fisher's Exact Test and Logistic Regression

Posted:

**Wed Mar 15, 2017 5:50 pm**I'm a medical student doing his research interneship. We're currently investigating if there's a link between motoric and cognitive outcome and MRI imaging done on neonates. Im wondering if the post-hoc analysis I did in SPSS was the right way to do it.

When inserting the data in SPSS, I computed lots of difference variables for the outcome and MRI's. Some as simple as dichotomous answers (for both outcome and MRI), and some very detailed (as much as 8 different groups for the different area's on MRI's; max 4 groups for outcome). For the 2x2 tables for the dichotomous variables, the result of the Fisher's Exact Test's were clear. But for the larger r x c tables (some as large as 8 x 4's), the p-value obtained only says something about the whole table, but doesn't tell you anything about about the relationship between the individual cells. So I did a post-hoc by crossing ''adjusted residuals'' in SPSS in crosstabs. This gives you the z-scores for all the cells. I squared these Z-scores to obtain the chi-square. I computed the p-value from these chi-sqaure scores with SPSS via transform - significance - p-value (df 1). To individual p-values form all these cells were compared with a p-value corresponding with the r x c table ((0,05 / (r x c)). Was this the right way to do this? I know you can also make 2x2 tables for all the r x c tables, but that would take ages...

I also did a binary logistic regressions for the dichotomous outcome variable. The thing that struck me as strange was that on 2 test I didn't get a significant p-value out of the post-hoc analysis, but I did get it on the logistic regression (all other variables lead to the same result on both tests). I'm wondering which test I should trust in this case? My guess would be the regression analysis as this presents (for me) a more cut and dry method to assess significance than the post-hoc analysis. Or should I present both findings in the article.

I hope it's all clear, we get the bare minimum about statistics during the length of our study. So my statistical knowledge is not advanced at all.

Thanks for your time and help!

J.

When inserting the data in SPSS, I computed lots of difference variables for the outcome and MRI's. Some as simple as dichotomous answers (for both outcome and MRI), and some very detailed (as much as 8 different groups for the different area's on MRI's; max 4 groups for outcome). For the 2x2 tables for the dichotomous variables, the result of the Fisher's Exact Test's were clear. But for the larger r x c tables (some as large as 8 x 4's), the p-value obtained only says something about the whole table, but doesn't tell you anything about about the relationship between the individual cells. So I did a post-hoc by crossing ''adjusted residuals'' in SPSS in crosstabs. This gives you the z-scores for all the cells. I squared these Z-scores to obtain the chi-square. I computed the p-value from these chi-sqaure scores with SPSS via transform - significance - p-value (df 1). To individual p-values form all these cells were compared with a p-value corresponding with the r x c table ((0,05 / (r x c)). Was this the right way to do this? I know you can also make 2x2 tables for all the r x c tables, but that would take ages...

I also did a binary logistic regressions for the dichotomous outcome variable. The thing that struck me as strange was that on 2 test I didn't get a significant p-value out of the post-hoc analysis, but I did get it on the logistic regression (all other variables lead to the same result on both tests). I'm wondering which test I should trust in this case? My guess would be the regression analysis as this presents (for me) a more cut and dry method to assess significance than the post-hoc analysis. Or should I present both findings in the article.

I hope it's all clear, we get the bare minimum about statistics during the length of our study. So my statistical knowledge is not advanced at all.

Thanks for your time and help!

J.