Diagnoses, Predictive Values, and Whether You're Sick or Not: NPV and PPV
ฝัง
- เผยแพร่เมื่อ 31 พ.ค. 2015
- Housekeeping first:
1) You can support Healthcare Triage on Patreon: vid.io/xqXr Every little bit helps make the show better!
2) Check out our Facebook page: goo.gl/LnOq5z
We've covered test characteristics like sensitivity and specificity before, and we've even covered how you can use those values to make decisions about what results mean about your health. But sensitivity and specificity don't have a meaning that people usually understand.
What people want to know are other values, what we might call the positive predictive and negative predictive values. But there are problems with those metrics. They're the topic of this week's Healthcare Triage.
For those of you who want to read more, go here: theincidentaleconomist.com/wor...
John Green -- Executive Producer
Stan Muller -- Director, Producer
Aaron Carroll -- Writer
Mark Olsen -- Graphics
/ aaronecarroll
/ crashcoursestan
/ johngreen
/ olsenvideo
Was already planning on rewatching the sensitivity and specificity video today to help me make some decisions regarding cancer screening. I think Aaron is a mind reader.
This REALLY clears it up better than the se/sp video you put up before.
Thanks. The chart that shows the clear formulas between sensitivity, specificity and PPV and NPV is very helpful. I am trying to learn how to calculate confidence intervals and to do so I need to get the mean. Where do I find the mean of this? Are there 2 means? Do you have another video to help me out with that?
Dr. Carroll, you are amazing , and so are your videos, I love statistics episodes, please do more statistics episodes.
Great episode, thanks!
I love these videos, thank you
Fantastic episode!
Link to the sensitivity / specificity video, for us new subscribers?
More test statistics aren't always beneficial. This week on Healthcare Triage: th-cam.com/video/dHj7ygeqelw/w-d-xo.html
Hey, Aaron, could you do an episode on who gets the most radiation out of all people? I've seen a lot of research proving that it's smokers due to the polonium. It would be appreciated.
Inorganic Vegan Airline pilots?
Inorganic Vegan Veritasium has a great video on it, you should be able to find it easily.
Tyrrell
Nope. It's surprisingly smokers.
Lazhward
I saw it, but I was referencing the research I saw. Still a good video.
Inorganic Vegan Might be a fun topic I can cover sometime. The long non-answer is it gets complicated by the way to talk about radiation exposure and its related health effects. A whole body dose from an xray deposits energy different to different parts of the body than a smoker that focuses radiation into a specific body part, even if the total amount of energy deposited by both was the same. We have ways we compensate for this, IE, effective dose, but it gets complicated!
can you send the link of the previous episode of sensitivty and specifity episode
Thanks for doing this. I can't tell you how many women on pregnancy forums advocate skipping screening, such as the NT scan or the 1-hour diabetes screen, because "there are so many false positives."
Would it be reasonable to first check specificity and sensitivity to see if a test is reliable, and once you have confirmed that the test is reliable, check the npv/ppv to see how great the chance is that your diagnosis is right?!
And is there a way to condense sensitivity And specificity into one number that shows how accurate a test is e.g.
(S+P)/2
Or
S+P-1
or something like that?!
Hi, I am confused about the cutoff point for most of the true positives, the highest positive predictive value, the lowest negative predictive value and the highest specificity on graph. At what point, I get most of the true positives and at what point I get the highest specificity. I hope someone clarifies these. I would be much more appreciated.
Can u do a video about pregnancy risks and how likely they can really happen
at 1:35 you meant to say "relative to all people who do NOT have the disease, B + D). great videos, thanks.
You have a typo in one of your equations. Go to around 4:35 or so and look at Specificity. You list "50" when you mean "450."
sovereign578 He also says "all people who have the disease: B+D" .. those are the ones that don't have the disease
hoarf Exactly - this is at 1:34 to 1:35, by the way.
you wrote wrong on fakitus scenario #2, specificity = 450/(450*450) * 100 = 50%
and not 450/(450+50). (typing error I guess)
So whats your opinion on HIV Western Blot and related Rapid Tests. They claim both a very high Sensitivity and Specificity. Moreover prevalence is determined by using these test on a "Low" risk population. Basically the test themselves are used to determine how accurate the tests are. There is no independent test or diagnostics to determine a positive infection .
I can see how PPV and NPV are bad measures and sensitivity and specificity are better for determining if a test procedure is good or not. However, I feel the opposite is the case for patients actually getting tests done and understanding their results. PPV and NPV can give some insight to the patient on how likely they are actually ill or not, given their test result.
The "viewers like you" line reminds me of PBS.
I need help I know I have a parasite infection but my doctor after two years he said he never treated a infection like this then said I'm crazy please help I live in London KY I can tell you my symptoms plz help
I think it would be interesting to set up a calculator for this in excel and port it to google docs to take with me next time I go to the doctors office and they want to do something. Problem is I have almost no experience with excel.
PPV is essentially factoring in the people who are disease negative (B) by essentially saying "of those who test positive (A), some of those test + will also be will also be Disease Negative (B)". That is different to sensitivity, which looks purely at the diseased population and completely ignores (A) those who Tested + for disease and are (B) Negative Disease.
What about the cross-ratios?
Do A / (A + D) and C / (C + B) have any useful meanings?
PitaJ A / (A+D) is the disease positive percentage out of all people correctly diagnosed. C / (C+B) is the missed diagnosis (as opposed to the false alarm) percentage out of all people incorrectly diagnosed. I think the two metrics depend on both the quality of the test and the prevalence of the disease, so I don't think they're useful.
your math is wrong for specificity at 4:36
there is a typo @ 4:28
i'm unclear as to how the prevalence of a disease in a given population is such a detriment to the predictive values and yet prevalence (prior probability) is central to baye's theorem? could someone help me think more clearly on this? thnx
You should have use False Positive and False Negative. It is a happy to glad type change but it is easier to track.
I think I have fakitus, what do I do?
jdcguitar Did you get a positive test result? I heard that 90% of people who tested positive have the disease!
***** That's actually not bad advice for everyone! Apart from the gluten-free bit, of course. Only those sensitive to gluten in some way need avoid it. Also, make sure it's non-dairy milk so you're not drinking pus, blood, mucus etc. ;)
jdcguitar I think you should look into some homeopathic remedies.
In Fakitus scenario #2 the sensitivity is wrong on screen. It says 450/(450+50) when it should be (450+450). I just wanted to let you know so next time a small mistake like that doesn't make it through. Just trying to help but besides that it was a good video.
Healthcare Triage I think the point you are trying to make: "Given PPV & NPV what do we learn about the test" misses the question I want answered as a patient: "Given a test that passes sensitivity / specificity selection, I then get a positive result, am I actually sick?"
The example on general population screening for breast cancer proved the value of PPV and NPV. aka You got a negative result? you are fine. You got a positive result? you only have a 4% chance of actually being sick.
The value difference is to the patient in interpreting the results not the doctor in selecting a test.
Suggested New Topic:
Watch a few episodes of House and tell us where the writers have made a mistake, mislead us and what they got right.
At 1:30 you say specificity uses B + D and describe it as "the proportion of people who are healthy and test negative (D) relative to all people who have the disease (B + D)." Aren't all the people who have the disease A + C?
Multiplying the result by 100 isn't really necessary. For example, 450/(450+50) = ..9 and if you apply probable statistics it's highly likely. The multiple times 100 just shifts the decimal point two places to the right.
kd1s I think that one was a typo btw. It should be 450(450+450)
kd1s Yeah. You multiply by 100 to get a percent.
John Edwards That's kind of strange though, because in practice you normalize to 1, not 100.
Zuzu Superfly In practice?
John Edwards Yea. When you actually use it. Show me an instance when you normalize to 100 instead of 1.
I have an exam today I’m scared and anxious (:
I'm sure there are people who understand this
***** if a good test has a small false positive rate, but the condition is vanishingly rare, then most of the positive results will be the false positives occurring from the 99.999% of the population without the condition, rather than any very rare patient actually having the condition. The second issue is about the positive predicted value (what fraction of people with a positive test result actually have the condition): the person being tested needs to be of the same selection group as that used to derive the positive predictive value data, and most tests are developed to help confirm a diagnosis in a group of patients already (through history and examination) suspected of having the condition, rather than the population as a whole. The value of tests is to help progressively exclude patients who do not have the condition, leaving behind a smaller subset group with increasing chance of having the condition. Consider seemingly straight forward pregnancy tests, they have a given sensitivity/specificity rate, and its predictive values are derived from testing a very specific group of people: namely women of childbearing age who have not had hysterectomies. If one starts routinely undertaking pregnancy tests in female infants, elderly women or indeed the male half of the population, then the predictive value of a positive test is greatly less (we can surmise that the positive predicted value in 4year old boys is zero - all such positive results are false positives in this subgroup, and the PPV of perhaps 99% in fertile women is irrelevant). Even if one thinks the above example of pregnancy testing is too simplistic, let’s consider 2 scenarios to show increasing complexity and difficulty in knowing how to interpret the results. a) consider pregnancy tests run on women of child bearing age, the group we stated the test should apply to, or is it ? What of testing women the day after they have delivered - the hormone levels will yet to have returned to baseline levels and so the tests will be reported as positive, yet the woman clearly will not have ovulated and reconceived in one day. What of a test run 3 weeks later, or after 1 month ? b) what of a woman who has been breast feeding for 6months, has infrequent or absent periods, and then experiences abdominal pain. A pregnancy test may be undertaken to help decide if they have an ectopic pregnancy. This group of women is less likely than non-lactating women to be fertile, but less likely is not the same as a zero possibility. How much reliance should one then place on the pregnancy result - one needs to know the chance of such women eventually truly proving to be pregnant to know how useful pregnancy test results are. A positive result would in the circumstance be taken as very strong evidence of pregnancy and an ectopic needs to be considered. But a negative result does not always rule out an ectopic pregnancy, especially if the pain been occurring for a while and placental tissue hormone production has ceased.
Multiplying by 100 is wrong without explicitly stating that you're using percent. You can multiply by 100%, but that doesn't do anything.
bjb2 It wouldn't be wrong, it'd just be vague. But you're smart enough to know what he means, so it's entirely proper.
Not to to mention he does in fact give the results as a percentage; when he gives the ppv of the mammograms, he says 4.4 percent.
Zuzu Superfly The equations given are wrong. (50 / (50 + 450)) * 100 = (50/500) * 100 = 0.1 * 100 = 10.
10 is not equal to 10%. 10 is equal to 1000%. The '%' literally means 'divided by 100'. If you want to get a percentage, you multiply by 100%, not by 100.
rufusdark Percents are a dimensionless unit; there's nothing wrong with the equations. 0.1 is the pure number which is then multiplied by 100 and by the unit "percent".
rufusdark I'll apologize if you demonstrate that percents couldn't formally qualify as dimensionless units, or that operators and units are necessarily mutually exclusive.
Either the camera isn't straight or the studio is falling into the abyss...
... Fakitus? Not fakeitis or fakitis? I'm sorry to be concentrating on that, but it really caught my attention.
Am I dumb? I think I might be dumb.
so many errors...
You are talking too fast 👉🏻👈🏻