[2006alum]

OK, to figure out an acceptable specificity, you would have to start with the acceptable positive predictive value and work backward.  The positive predictive value is the proportion of people with a positive test who are infected, and it is not characteristic of the test; it depends on both test characteristics and the baseline prevalence of disease.

So, let's say you wanted to have a 95% positive predictive value (that is, if you had a positive test, you wanted to be 95% sure that you really had been exposed to Covid-19.  You can pick your own value, but let's take 95% to be our target.  Now, let's say that the prevalence of disease is 5%.  In order to hit your positive predictive value, you need very few false positives.  In fact, even if you detect every single positive case (100% sensitivity), because there are so few true positives, you would actually need the false positives to be 5%/19 (because you the positive tests in individuals with Covid-19 need to be 95% of the total number of positives, or 19-fold higher than your false positives), so your specificity has to be along the lines of 99.75%.  If your prevalence is a bit higher (more like 10%) then you will have more true positives and your false positives (and thus specificity) can be lower.

If you want, you can play around with the disease prevalence and test performance characteristics (sensitivity and specificity) and see how it affects the positive and negative predictive value (NPV is the probability that you don't have the disease if you tested negative).

https://docs.google.com/spreadsheets/d/1...sp=sharing

BC

Very belatedly, thanks for this! Looks like with tests with roughly 96% percent sensitivity and specificity, we would need to reach roughly 50% prevalence to get into the 90% confidence range for positive predictive value. Alternatively, if the tests could get to 98%, even 20% prevalence would get us to into the 90% confidence range. We obviously have to tolerate some amount of error, but I feel like anything less than 90% might be a gamble given a disease where both asymptomatic and presymptomatic carriers can shed the virus for over a week. (Though if we had rapid and widely available testing that would diminish those concerns, too...)

[2006alum]

OK, to figure out an acceptable specificity, you would have to start with the acceptable positive predictive value and work backward.  The positive predictive value is the proportion of people with a positive test who are infected, and it is not characteristic of the test; it depends on both test characteristics and the baseline prevalence of disease.

So, let's say you wanted to have a 95% positive predictive value (that is, if you had a positive test, you wanted to be 95% sure that you really had been exposed to Covid-19.  You can pick your own value, but let's take 95% to be our target.  Now, let's say that the prevalence of disease is 5%.  In order to hit your positive predictive value, you need very few false positives.  In fact, even if you detect every single positive case (100% sensitivity), because there are so few true positives, you would actually need the false positives to be 5%/19 (because you the positive tests in individuals with Covid-19 need to be 95% of the total number of positives, or 19-fold higher than your false positives), so your specificity has to be along the lines of 99.75%.  If your prevalence is a bit higher (more like 10%) then you will have more true positives and your false positives (and thus specificity) can be lower.

If you want, you can play around with the disease prevalence and test performance characteristics (sensitivity and specificity) and see how it affects the positive and negative predictive value (NPV is the probability that you don't have the disease if you tested negative).

https://docs.google.com/spreadsheets/d/1...sp=sharing

BC

Very belatedly, thanks for this! Looks like with tests with roughly 96% percent sensitivity and specificity, we would need to reach roughly 50% prevalence to get into the 90% confidence range for positive predictive value. Alternatively, if the tests could get to 98%, even 20% prevalence would get us to into the 90% confidence range. We obviously have to tolerate some amount of error, but I feel like anything less than 90% might be a gamble given a disease where both asymptomatic and presymptomatic carriers can shed the virus for over a week. (Though if we had rapid and widely available testing that would diminish those concerns, too...)