[mmmrtl]

That 40% number is accurate but misleading. Here's an old prediction of pretty much that result that explains the issue: https://liorpachter.wordpress.com/2013/11/30/23andme-genotyp...

Basically, almost all (~99.9%) of the results of tests like 23andme (SNP arrays) are correct, but because an interesting/scary variant/mutation is about as rare in the population as the test result being wrong for a given site, somewhere around half of those scary results are false positive. If 1000 people are tested for a mutation with this method, 10 will be false positive. The real frequency of mutations vary by site - some will be close to 50%, others at 10e-9, depending on how strongly that variant is selected against, so the most harmful ones are very rare. If you get a result from 23andme that your doctor thinks is worth retesting, it's probably because it's a harmful mutation, which is rare, and ergo there's a fair chance the 23andme result is wrong. So 23andme is almost always right, except when the result might affect your health in a very serious way. I should point out that for late-onset diseases like cancer, heart disease, alzheimer's, etc., there's little natural selection, so pathogenic variants are more common, making it more likely that the 23andme result is correct.

It's a simple bayes rule result. It would not be hard for 23andme to give you an estimate of how likely the result is to be a false positive, but maybe there's some regulatory barrier in the way?

[jstandard]

Thanks, this is fascinating.

How is error rate being defined in this case?

I took your explanation to mean: Cancer is common, so the genetic blueprint for Cancer risk factors should also be common. We have a higher number of samples to verify against which increases the detection accuracy.

For less common diseases, we have a smaller pool to verify against, so our blueprint is less accurate. Lower detection accuracy.

[rb23]

No, that's not what it is. Let me give an example:

Disease X is very bad and strikes around age 30, so it has been selected against and is very rare. Only 1/10,000 people have the gene that causes it. Disease Y is bad, but it strikes when you're 70 (Alzheimer's let's say), so it is not strongly selected against and is more common, with 1/100 people having it.

The technology 23andMe uses is wrong about 1/10,000 times, just randomly, by chance. This is true for every place in the genome it tests.

Of 10,000 people, only 2 will test positive for disease X while 101 will test positive for disease Y. The 2 people who test positive for disease X will be the 1 person who actually has it, and 1 false positive, so 50% of the people who tested positive are actually at risk. Of the 101 people who test positive for disease Y, 100 will be at risk and 1 will be a false positive, so 99% of people who test positive are actually at risk.

Incidentally, this is why testing populations at very low risk for disease is generally counterproductive. The false positive rate stays the same but as the true positive rate gets very low, one tends to cause more harm than good.

[jstandard]

Thanks, the overall explanation makes sense now.

> The technology 23andMe uses is wrong about 1/10,000 times

Is 23andMe's tech considered state-of-the-art? I'm assuming there are ways to reduce both the false positive and false negative rate and I'm curious how other labs/tech stack up against these rates.