[R] Fw: Logistic regresion - Interpreting (SENS) and (SPEC)

[Frank E Harrell Jr]

Gabor Grothendieck wrote:
> On Mon, Oct 13, 2008 at 11:47 PM, Frank E Harrell Jr
> <f.harrell at vanderbilt.edu> wrote:
>> Gabor Grothendieck wrote:
>>> On Mon, Oct 13, 2008 at 11:21 PM, Frank E Harrell Jr
>>> <f.harrell at vanderbilt.edu> wrote:
>>>> cryan at binghamton.edu wrote:
>>>>> I recall a concept of Snout:  sensitivity that is high enough to
>>>>> essentially rule out the presence of disease.  And Spin:  specificity
>>>>> that
>>>>> is high enough to essentially rule in the presence of disease.
>>>>>
>>>>> So perhaps the below is backwards?  The higher the sensitivity, the
>>>>> greater the NPV?  And the higher the specificity, the
>>>> greater the PPV?
>>>> Why should we care when we can directly estimate Prob(disease | test
>>>> results
>>>> and risk factors)?
>>> Sensitivity and specificity are functions of the test only but ppv is
>>> also a function
>>> of the disease prevalence.   Just change the prevalence and the ppv
>>> changes
>>> whereas sensitivity and specificity are invariant.
>> Gabor,
>>
>> That's a very common belief but it turns out not to be true.  See references
>> from my earlier post.  Sensitivity and specificity are only invariant in you
>> don't analyze how they vary.
>>
>> Also, much research does not understand what prevalence really means. It
>> actually could be argued to not be a scientific quantity as its meaning
>> depends on unspecified mixtures of subjects.
> 
> Its the number of diseased patients in the population divided by the
> total population
> considered.

True, but researchers who attempt to adjust various estimators for 
prevalence in other populations tend to mix conditional and 
unconditional estimates.

When strong risk factors exist I find the concept of prevalence not very 
useful, just as I don't want to know the prevalence of pregnancy in the 
entire population.

> 
> Suppose we want to compare the PSA test for prostate cancer to some other
> new diagnostic.  We want a measure of the test itself, not of the population.
> We would like the numbers to be the same in Japan and North America even
> though the prevalence of prostate cancer varies widely between them.
> 
>>> If our aim is to assess a test one wants a measure that only measures the
>>> test
>>> itself.
>> There is no such measure.  The performance of a test depends on the type of
>> patient being tested as well as other things.
>>
> 
> There is no such thing as a normal distribution since if you get enough
> data you will find discrepancies but that does not mean that for all practical
> purposes that there is no normal distributions.

I'm not clear on the analogy.

> 
> Sensitivity and specificity are generally used to compare tests, not patients.

That's for sure.

There is one "pure" quantity although it doesn't measure absolute yield 
of the test: the adjusted odds ratio.

Cheers,
Frank

> 


-- 
Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt University