Biostat - how Cut-off values affect PPV/NPV, sensitivity/specificity

This forum made possible through the generous support of SDN members, donors, and sponsors. Thank you.

Stylus

New Member
7+ Year Member
Joined
Jan 16, 2014
Messages
2
Reaction score
0
Had an issue understanding this concept for quite some time and finally figured out. Sharing some details that helped me (attached as word doc as well since tables may not appear clearly)

How cut-0ff values affect PPV and NPV

The text in bold (italics) is from an explanation I found online and the rest is smth that I put together to understand the concept better

Consider the following hypothetical example: measurement of high endorphin levels in SpRs in Anaesthesia has been found to be associated with success in the final FRCA examination. A sample of SpRs is tested before the examination resulting in a range of endorphin values. The data are examined and an arbitrary cut-off point for endorphin levels is chosen above which most of the candidates passed with few failures. Despite choosing the cut-off value in such a way that the maximum possible number of SpRs is correctly classified, we may find that 10% of the cohort with endorphin levels above the cut-off level failed the exam (false positives) and 15% of the cohort with endorphin levels below the cut-off level passed the exam (false negatives).

Let say for e.g

N=100 doctors

Cut-off value for endorphin = 200

doc will pass if endorphin level above 200

doc will fail if endorphin level below 200


Exam passed Exam failed Total

Endorphin test + ve 54 (TP) 6 (FP) 10% 60 docs had endorphin ≥ 200


Endorphin test -ve 6 (FN) 15% 34 (TN) 40 docs had endorphin ≤ 200

100

If the cut-off point is raised, there are fewer false positives but more false negatives—the test is highly specific but not very sensitive. Similarly, if the cut-off point is low, there are fewer false negatives but more false positives—the test is highly sensitive but not very specific

The 6 (FP) people who actually failed the exam most likely had endorphin levels just slightly above 200. So if we raised the cut-off from 200 to 230 (for e.g) then probably all the 6 would correctly reflect that they failed (i.e. will become TN and that FP will disappear or decrease). But this would mean there could be more false negatives showing up.

Similarly the 6 (FN) people who actually passed the exam most likely had endorphin levels just slightly below 200. So if we lowered the cut-off from 200 to 170 (for e.g) then all the 6 would correctly reflect that they passed (i.e. will become TP and that FN will disappear or decrease). But this would mean there could be more false positives showing up.

Bottom Line:

If we raise cut-off

False positives decreases

False negatives increases

If FP ↓ then specificity increases (FP is in denominator in formula)

If FN ↑then sensitivity decreases (based on formula)

If FP ↓ then PPV increases (based on formula)

If FN ↑then NPV decreases (based on formula)


If we lower cut-off

False negatives decreases

False positives increases

If FN ↓ then sensitivity increases (based on formula)

If FP ↑then sensitivity decreases (based on formula)

If FN ↓ then NPV increases (based on formula)

If FP ↑then PPV decreases (based on formula)

Members don't see this ad.
 

Attachments

Last edited:
Huge wall of text for a pretty intuitive concept.

Take DM as an example and fasting [glc] as the cut-off. If you increase the cut-off from 126 to 150 you will catch fewer people, but be more confident about the diagnosis in those people you do catch (i.e. people with >150 [glc]). Thus, lower sensitivity, higher specificity, higher ppv, lower npv. Don't need any formulas to understand the relationships. Also, sensitivity and npv go together, so if one decreases, the other decreases. Same for spec/ppv.
 
Top