0

Hope all is well. I developed predictive models and I wonder if anyone has some gold standard for sensitivity and specificity from the literature so that I can refer my results to the ones in the literature to validate my study?

I really appreciate it.

Many Thanks

Ben

gung - Reinstate Monica
  • 132,789
  • 81
  • 357
  • 650
Benn
  • 97
  • 1
  • 3
  • 1
    The many reasons why sensitivity and specificity should be avoided are detailed [here](http://www.fharrell.com/post/class-damage). – Frank Harrell May 01 '18 at 11:26

1 Answers1

3

How long does a piece of string need to be?

It depends on what you need to do with it.

How high do you need your sensitivity and specificity to be?

It depends on what you need to do with your model. You need to do a cost benefit analysis based on your planned deployment environment to get a definitive answer.

It really is application specific. If there are application relevant standards to compare against then do so. If they have carried out a benefit risk analysis then you may be able to dovetail into that. If you perform better in both sensitivity and specificity then it is win win. More common is that one will be higher and one lower, then you really need to get into the meat of how the model will be used.

What are the costs associated with positive predictions?

What are the risks associated with actions undertaken (or inaction) in response positive predictions?

What are the costs associated with negative predictions?

What are the risks associated with actions undertaken (or inaction) in response to negative predictions?

What context will it be used?

Will it be completely randomly sampled from all possible targets?

Will there be some form of prefiltering that correlates with your predictive model? (see Is sensitivity or specificity a function of prevalence? for a relevant discussion)

Will there be follow up confirmatory testing or will affirmative action result directly from it?

Note on Gold Standard

The term 'gold standard' has a very specific meaning in classification/discrimination. It is what you used to define ground truth for the classification. The context of your use of the term suggests you mean an ideal target value for sensitivity and specificity.

ReneBt
  • 2,863
  • 1
  • 8
  • 24
  • Hello Rene . Thanks for the answer. The reason why I asked is because I tended to refer the sensitivity of the predictive models that I developed to some acceptable work in literature review . I don't know if you'have ever read an article talking about let's say "good or high enough" sensitivity or specificity for a predictive model in healthcare? – Benn May 01 '18 at 13:22
  • 1
    No, I haven't read such an article and I fear that if it exists it would be useless. Within healthcare, which is my application area, issues such as chronic vs acute, mortality rate, intervention side effect severity and disease prevalence vary so wildly that no useful rule of thumb is informative. Also, whether it is being used to screen the general population, diagnose people already deemed to be at risk or even people admitted to ward will change the characteristics substantially. – ReneBt May 01 '18 at 13:54