help on stats question: sens/spec/ppv

[bulldog]

---

Members don't see this ad.

I was looking thorugh brs behavioral sci epidemiology section, and there's a q where they have specificity of 70%, sensitivity of 90%, and ask you calculate estimate postive predictive value.

Answer is 25%. How do you get that? I got answer as 75%.

also, could someone point out the trends of sensivity, specifity, positive predictive value, and npv. i.e. when one variable goes up, what variable goes down, and what variables change?

 

[bulldog]

also, if you increased the cuttoff value (i.e. make the PSA cuttoff of 5 to become 7), how would it fx sensitivyt, specificity, npv, ppv?

 

[bludeviled]

I think the best way to answer those ?'s is to make up numbers you can easily fiddle around with
say you have a population of 1000, with a 10% prevalence or 100 with disease and thus 900 without disease.
so now they are telling you the sensitivity is 90% or in other words, out of the 100 people with disease 90% of them were picked up by the test, or 90 of them tested positive. (as for the leftover 10 who had the disease, yet tested negative, we call them false negatives)
If the specificity is 70%, that means the people w/o disease (900) were tested and 70% or 630 tested negative. (thus 900-630= 270 people who did not have the disease yet tested positive so they are the false positves)

so now the positive predictive value asks "what are the chances that a positive test result means I really have the disease?"
PPV = true positives/ (false + true positives) = 90 / (270+90) = 90/360 = 25%

As for trends, the prevalence does not change the sensitivity or specifcity of tests.
When prevalence increases, PPV increases and NPV decreases
When prevalence decreases, PPV decreases and NPV increases

As for cutoff questions, I'm not sure if there are any steadfast rules. I think its best to think about each situation logically vs. memorizing anything b/c they could ask you a simillar question about, say blood glucose ranges for diabetic cut-offs and they might either increases or decrease the value.
If you raised PSA cut-off from 5 to 7, you would miss those people with cancer and PSA's under 7. So you are being more selective and hence lose/decrease sensitivity. But you also decrease the rate of false positives so you have increased specificity. The PPV increases and NPV decreases.

 

[bulldog]

I think the best way to answer those ?'s is to make up numbers you can easily fiddle around with
say you have a population of 1000, with a 10% prevalence or 100 with disease and thus 900 without disease.
so now they are telling you the sensitivity is 90% or in other words, out of the 100 people with disease 90% of them were picked up by the test, or 90 of them tested positive. (as for the leftover 10 who had the disease, yet tested negative, we call them false negatives)
If the specificity is 70%, that means the people w/o disease (900) were tested and 70% or 630 tested negative. (thus 900-630= 270 people who did not have the disease yet tested positive so they are the false positves)

so now the positive predictive value asks "what are the chances that a positive test result means I really have the disease?"
PPV = true positives/ (false + true positives) = 90 / (270+90) = 90/360 = 25%

As for trends, the prevalence does not change the sensitivity or specifcity of tests.
When prevalence increases, PPV increases and NPV decreases
When prevalence decreases, PPV decreases and NPV increases

As for cutoff questions, I'm not sure if there are any steadfast rules. I think its best to think about each situation logically vs. memorizing anything b/c they could ask you a simillar question about, say blood glucose ranges for diabetic cut-offs and they might either increases or decrease the value.
If you raised PSA cut-off from 5 to 7, you would miss those people with cancer and PSA's under 7. So you are being more selective and hence lose/decrease sensitivity. But you also decrease the rate of false positives so you have increased specificity. The PPV increases and NPV decreases.

thanks!