[jasonhansel]

IMHO, some of those criticizing the article for failing to understand statistics are missing the point.

The point is that people who get a "positive" result on these tests are often put through terrifying levels of anxiety when there is no actual problem; this anxiety is often exacerbated because they aren't informed of the false positive rate. This clearly has a harmful emotional effect on people, and explaining the false positives in Bayesian terms, or reframing it in terms of sensitivity and specificity, doesn't undo that damage.

That potential harm needs to be explained to patients, and it needs to be weighed carefully against the potential benefits of the test (as is done for PSA tests for prostate cancer, which also have a high false positive rate). Given that potential for harm, it's not unreasonable to ask that these tests be more tightly regulated.

To quote the OP:

> In interviews, 14 patients who got false positives said the experience was agonizing. They recalled frantically researching conditions they’d never heard of, followed by sleepless nights and days hiding their bulging bellies from friends. Eight said they never received any information about the possibility of a false positive, and five recalled that their doctor treated the test results as definitive.

(Edit: clarified)

[midjji]

If you get a positive for a horrid cancer with a 90 percent false positives you should be afraid. Its lunacy for tests to be regulated beyond requiring rough false positive false negative rates, and if anything smacks of "I dont understand statistics and therefore have to protect my children from understanding statistics." The article is most likely written by some anti abortion idiot.

[teruakohatu]

> if you get a positive for a horrid cancer with a 90 percent false positives you should be afraid.

No you should not panic, anymore than you should celebrate and buy a yatch if you think you have 10% chance of winning a lottery.

The doctor should phrase it as "the test indicates you have a 10% possibility of cancer. The majority of people who test positive do not have cancer. Further testing is required to confirm or rule it out."

Going with the expected value of 10% of a horrid thing is still bad, or 10% of $100m is still $10m, is not applicable to a single non repeated event.

[halpert]

The GP wasn’t saying go out and kill yourself because you are definitely condemned. They were highlighting that you still have an extremely elevated chance of having cancer. If you got into your car and found out you had a 10% chance of being in a fatal accident, would you not worry? Sure, there is a 90% probability you’ll be fine, but the probability for most other people is 99.999%. Pretending like there isn’t a potential issue is dangerous and willful ignorance.

[thaumasiotes]

> Its lunacy for tests to be regulated beyond requiring rough false positive false negative rates

There is a good reason tests are described in terms of sensitivity and specificity ("if the answer in reality is yes, how often will the test say no?") rather than in terms of false positive or false negative rates ("if the test says no, how often is the true answer yes?"). The sensitivity and specificity are facts about the test which can conceptually stay constant[1] as you apply the test to different people. False positive and false negative rates do not have that property; they are facts about the group you're performing tests on just as much as they're facts about the test.

[1] This is not to say that the sensitivity and specificity of a test do stay constant as you apply the test to different populations. Often they won't. But it is a theoretical possibility, and even that isn't true for false negative rates.

[midjji]

For a test manufacturer you are right, and they aren't responsible for accounting for general population frequencies, or conditional frequencies based on things like some sibling has it etc. But there is no need to regulate that. But the serivice in question isnt a test manufacturer, its diagnostics, and they are responsible for this. I think a company selling a diagnostics service should be required to prominently display reasonably accurate FP,TP,FN,TN rates for gen pop, as the next example will show.

This is is not a legal problem, its a customer experience one, and as such its very much what the free market does well. A clever company could say that the tests service we provide works in a two stage process where a preliminary test is used to identify which conditions need to be investigated further, the results of which are presented in a meeting with one of our doctors were the results are explained and additional tests are performed to resolve any ambiguity. This combines the benefits of sensitivity, and specificity in a resource efficient way, and does not provide the user with any scary results before a certain answer is available. The FP,FN,TP,TN probabilities for gen pop for such a company would be much better than those of companies trying to do so in a single pass. But for these figures to mean anything regulating that they are available and reasonably accurate is key to ensure an efficient market.

Another problem is of course that the tests themselves are shit, or the rates based on poor data. The requirement on reasonably accurately reported gen pop rates help with this however, and I see no way around legally forcing diagnostic services providers to prove this. And until someone does, my patent pending algorithm(){return false} tests for all rare genetic disorders with excellent accuracy^^

[OJFord]

I'm not sure GP was intending to make that distinction (rates globally for the test vs. the sample population being tested by.. what a given doctor/hospital?) - I haven't come across that before.

If the population is the same then your changed-order definitions are just inverses, and they're just different terms for the same thing.

[thaumasiotes]

> I'm not sure GP was intending to make that distinction (rates globally for the test vs. the sample population being tested by.. what a given doctor/hospital?) - I haven't come across that before.

> If the population is the same then your changed-order definitions are just inverses

No, you just haven't understood the concept.

Let's assume some condition has a prevalence of 20%, and a test for it will correctly identify presence of the condition 95% of the time, while correctly identifying absence of the condition 90% of the time. We can immediately answer the first question: when the answer is "yes", the test will say "no" 5% of the time.

You have proposed that when the test says "no", the answer is "yes" a share of the time that might be the inverse of 5%, or perhaps 5% itself. I have no idea what you meant -- and I suspect you didn't either -- but the correct rate of false negatives is not 5%, 95%, nor 2,000%.

In a model population of 10,000 people, we will see this:

                  |  condition present  |  absent  |
    test positive |               1900  |     800  |
         negative |                100  |    7200  |

From this table we can see that the false negative rate is 100/7300 or 1.4%. The false negative rate looks much better than the sensitivity and specificity figures because the condition is rare. The corollary to that is a horrific false positive rate of 800/2700 = 30%.

[OJFord]

I don't think I have, it sounds like you think I'm saying false negatives are the inverse of false positives? Not at all, that's obviously not true.

I was surprised at 'sensitivity and specificity' (jointly) being considered different from 'false negatives and false positives' (jointly).

The given reasoning was about population differences, which.. fair enough, I understand that makes a difference, I just wasn't aware that was a standard difference in definition (if it is) and suggested the up thread commenter wasn't (or wasn't meaning to use it) either.

> correctly identify presence of the condition 95% of the time, while correctly identifying absence of the condition 90% of the time. We can immediately answer the first question: when the answer is "yes", the test will say "no" 5% of the time. You have proposed that when the test says "no", the answer is "yes" a share of the time that might be the inverse of 5%, or perhaps 5% itself. I have no idea what you meant -- and I suspect you didn't either

10%. 'inverse', as I called it, of 90%, not 95%.

(That's why I think you think I think (..!) that false negatives/positives rates are derivable from one another. Sorry if not and I'm just still not getting it...)

I don't think I am misunderstanding though - Wikipedia calls them 'true pos/neg rate', and gives formulae for false pos/neg rates as 1-true: https://en.m.wikipedia.org/wiki/Sensitivity_and_specificity

[thaumasiotes]

> it sounds like you think I'm saying false negatives are the inverse of false positives? Not at all, that's obviously not true.

No, I think you're saying that the false negative rate as I defined it in my comment is the inverse of sensitivity. You've corrected me to say that you think the rate I defined is the inverse of specificity, which makes even less sense. And you did that despite the fact that I included a full calculation demonstrating that that isn't true.

Wikipedia's definition of the false negative rate differs from mine. Wikipedia indeed defines the false negative rate as (1 - sensitivity), though not, as you seem to believe, (1 - specificity). But you get no credit for this, because I explicitly defined what I meant by the false negative rate, and you echoed that definition in your response to my comment:

>>> your changed-order definitions

So: you think you haven't misunderstood what's happening. I ask you this: in my table above, I believe that prevalence is 20%, sensitivity is 95%, and specificity is 90%. Please verify that.

I have said that the conditional probability P(condition present | test negative) is 1.4%. You responded saying that that probability is actually 10%:

>> You have proposed that when the test says "no", the answer is "yes" a share of the time that might be the inverse of 5%, or perhaps 5% itself.

> 10%

Where are you getting that figure from? Show me in the table.

[derbOac]

I've worked clinically and taught this in professional training programs. It's a classical clinical decision making problem.

The solution is fairly straightforward I think: require the test companies to report the posterior probability of the illness given the test result. Probably with some kind of uncertainty interval.

[rscho]

The posterior depends on the prior. So not very feasible in practice.

[derbOac]

Generally speaking in clinical settings there are rough estimates of base rates of diseases available, either through biological theory (eg, mutation rates) or empirical studies or both. Not always to be sure, but often enough to mitigate against some of the problems discussed.