Discrepancy between Chi-square and sensitivity and specificity

[Naomiyiftach]

The performance of 19 impaired and 42 typically developing children was tested using 3 assessment parameters to see which parameter best discriminated between the groups (correctly detected impaired as impaired or typically developing as typically developing). In addition discrimination function analysis was conducted to see which combination of tools provided the best diagnostic accuracy. HOWEVER, the highest Chi-square did not correspond with the highest specificity and sensitivity.
As both chi-square and specificity and sensitivity to my knowledge reflect the proportion of participants who are correctly and incorrectly classified, I am struggling to understand this discrepancy.
ANY EXPLANATION would be highly appreciated!!!!!!!!!!!!!!!!!!!!

 

[hlsmith]

accuracy = combination of SEN and SPEC. We need to know more about your variables to help out. Prevalence influences the weights of SEN and SPEC in calculating accuracy.

 

[Naomiyiftach]

No problem- some more data.
5 Models were compared. Each model consisted of different combination of the 3 assessment tools. Following are the results of the discrimination function analysis

Model 1- tool A alone. chi-square 73.546. standardized coefficient 1. specificity 95.2%, sensitivity 100%, overall accuracy 96.7%

Model 2- tool A and tool B. chi-square 87.553. standardized coefficient tool A .922 tool B .541. specificity 92.9%, sensitivity 93.4%, overall accuracy 93.4%

Model 3- tool A and tool C. chi-square 75.780. standardized coefficient tool A .972 tool C .257. specificity 92.9%, sensitivity 89.5%, overall accuracy 91.8%

Model 4- tool A, tool B and tool C. chi-square 86.856. standardized coefficient tool A .925 tool B .560 tool C -.041. specificity 92.9%, sensitivity 64.7%, overall accuracy 93.4%

Model 5- tool B and tool C. chi-square 26.854. standardized coefficient tool B .981 tool C .039. specificity 88.1%, sensitivity 68.4%, overall accuracy 82%

Thanks!!

 

[hlsmith]

I have never run "discrimination function analysis", but it seem to parallel accuracy values, with exception of model five which may be a typo?


It would all come down to what the chi-sq test is actually testing?