Inverse Prediction

[#1]

Dear JMP experts,

I appreciate any clarification in regard to "Inverse Prediction"!

In my databse I have variable represent a biological marker, and I have two level groups (patients A and healthy B). I would like to find a cutoff value for the  marker to say that this value predicts disease at specific probability (data attached).

I did logistic fit, then ROC curve, after then "Inverse Prediction" which gave me the "prediction probability" for every value of the marker (figure attached). My questions are:

1. Are the steps that I followed, correct, to find a cut off value for the marker to predict disease A.

2. What the term "Inverse Prediction" means? more specifically, what the word "inverse" stands for?

Thanks for any clarification

John

[#2]

We usually fit a model to data, generally Y = f(X). We can then use this model to predict Y for a given X. Inverse prediction is used when we want to predict the X value that corresponds to a given Y.

[#3]

Dear markbailey, I immensely appreciate your reponse! I would like to know the methods that JMP uses to measure "Inverse prediction" for a certain variable. Is there any reference that may help to clarify how it works?

Sincerely,

John

[#4]

John:

A great place to start learning 'how' JMP does anything is the documentation. Here's a landing page for 'inverse prediction' within the JMP documentation. Inverse prediction as a method is native to several JMP analysis platforms so the exact answer to 'How does JMP do inverse prediction' is dependent upon the analysis platform that was used.

https://www.jmp.com/support/help/14/inverse-prediction.shtml

[#5]

In addition to the link provided by Peter for the definition, you can find more information by selecting Help > Books > Fitting Linear Models. Go to the Appendix for the statistical details about the inverse prediction method.

[#6 · Accepted solution]

Start with the ROC Curve. Open the Details report. Use this table to decide the cutoff that you want that satisfies your goals for sensitivity and specificity. Peter's solution is a good one when you know the probability and want to determine the corresponding level of the biomarker.

[#7]

Peter_Bartell and  markbailey, I appreciate your responses. Thank you!