On Twitter I’ve seen lots of discussion on the sensitivity and specificity of tests for SARS-CoV2 (both diagnostic and antibody tests) but not a lot of discussion around the positive and negative predictive value of these tests (PPV and NPV). Previous discussions about sensitivity/specificity vs PPV on this site have argued that PPV and continuous risk measures should be prioritized for clinical decision making. With this in mind, why is there not more discussion about the PPV of these tests?
My initial thought is that we currently do not know prevalence (lots of testing studies underway), so it would be hard to calculate PPV. If so, is there a time when we should start thinking about the PPV of these tests and how PPV statistics factor into clinical and policy decisions?
1 Like
I’ll bet that the absence of the needed statistical quantities arises because the data do not come from studies but rather from haphazard case selection. To get the probabilities we need you need the right inception cohort.
2 Likes
@Stephen Senn posted this comment a few weeks ago, spawning an interesting conversation:
The paper he cites [1] offer some wonderful historical insight into sensitivity & specificity and associated debates. A further thread from Adam Rothman carries the discussion back further to Laennec!
- Guggenmoos-Holzmann I, van Houwelingen HC. The (in)validity of sensitivity and specificity. Stat Med. 2000;19(13):1783-1792. [link to PDF]
4 Likes
One issue is the one pointed out by f2harrell, namely the haphazard nature of source populations. The other is that we would normally regard sensitivity and specificity as being (more or less) characteristics of a test in itself, and PPV and NPV as the consequences of these particular values of sensitivity and specificity when the test is applied to a fresh population with a different prevalence to the training series.
2 Likes
The fundamental problem is that the predictive values of the tests have been tested in hospital settings with high representation of confirmed severe cases. However, these tests are currently being used in asymptomatic individuals, who are presumably at very low risk of being infected. The problem, I believe, is exacerbated by the fact that population estimates rely heavily on the PPVs of these tests, the parameter which is precisely unknown. Therefore, I believe that there is a lot of unknown data on the diagnosis of Covid-19 in oligosymptomatic patients.
1 Like