Help with Nominal Logistic Fit: Confusion Matrix vs ROC Table Output?

Thierry_S

Hi JMP Community,

 

I have been studying the Nominal Logistic Fit to determine the value of a Baseline Biomarker to predict the outcome of a Clinical Treatment.

 

I thought that I understood the concept of the Confusion Matrix: it returns the numbers of True Positive, True Negative, False Positive, and False Negative for a given model for the Training data set and, if defined, the Validation set. However, when I compare the Confusion Matrix to the best outcome from the ROC Table (Maximum SENSITIVITY - (1 - SPECIFICITY) value), I struggle to reconcile the two.

 

For example I have a a model with a ROC AUC = 0.654 (rather weak association) where the Confusion Matrix returns:

		Predicted	Predicted
		YES	NO
ACTUAL	YES	1	82
ACTUAL	NO	1	278

 

--> which is really bad (actually worst than expected for the ROC AUC value).

For the same model, the ROC Table best combination of SENSITIVITY and SPECIFICITY is:

 

Prob	1-SPEC	SENS	SENS - (1-SPEC)	True Pos	True Neg	False Pos	False Neg
0.2797	0.2437	0.5060	0.2623	42	211	68	41

 

--> which is quite bad  but more in line with expected outcome of a model with a ROC AUC = 0.654

So, my questions are:

• What is the main difference between Confusion Matrix and the "best" row of the ROC Table?
  • Is it because the former use the highest Probability and the latter uses the best SENSITIVITY and SPECIFICITY combination?
• If I were to present these results, what would be the best option to present the Positive Predictive Value and the Negative Predictive Value?

Thank you for your help.

 

Sincerely,

 

TS

Mark_Bailey

The confusion matrix and ROC are different. You understand the confusion matrix as described. In this example, you have practically no sensitivity (1/83) but quite good specificity (278/279).

 

The ROC simultaneously evaluates both sensitivity and specificity so overall it looks a bit better than chance (AUC = 0.654).

 

The confusion matrix is for one cutoff and the ROC curve uses each observation as a cutoff, including the observation that produces the largest separation.