this article is trivial nonsense. of course he's technically correct, but the article contains no useful information and boils down to just saying that people (including doctors) aren't looking at literally only 5 year survival rate charts.
like the colon cancer thing. he talks about how it would only be more effective to catch colon cancer early if you assume we have treatments for it that would work early. but we don't need to just assume blindly. we already know we do have those treatments!
Agreed. I think his blog title, "Probably Overthinking It", is appropriate named.
Essentially every assertion in the article is either an oversimification, cherry picking a random niche situation to highlight, or just flat out factually inaccurate.
Let's take this paragraph for example:
"Catching cancer early is beneficial only if (1) the cancers we catch would otherwise cause disease and death, and (2) we have treatments that prevent those outcomes, and (3) these benefits outweigh the costs of additional screening. This table does not show that any of those things is true."
To address these one by one:
1. Obviously cancer causes disease and death. The same graphic he references makes that abundantly clear. Sure, there might be some rare exceptions (elderly patients with slow growing colon cancer for example), but we're talking about the general population.
2. All cancers have treatment options available in some form (could be chemo, radiation, surgical resection, etc), so this assumption doesn't even make sense to include. Let's assume for a second though that treatments might not be available. Even if that were true, there ARE treatments that can help treat cancer symptoms, and but may not affect the tumor directly. Often these are specific to the specific type of cancer.
3. This assertion is dumb - is the author really trying to argue that providing symptomatic or other relief to a cancer patient isn't a sufficient benefit to warrant additional screening?
I could go on, but you get the point. Some people just like arguing for the sake of arguing I guess.
Of course they're misleading.
What did the doctor tell us when my mom was diagnosed? Don't do research, do not trust Dr. Google, Dr. Google lies. At best, Dr. Google is behind the times.
The specifics of your case will strongly affect what happens to you.
And even for cancers that are a guaranteed death sentence, survival has increased significantly in recent years.
I think this is a technical article about a narrow aspect of public health policy, not advice to individual patients.
One point in the article is that early detection would give you more years to live even if there were no treatment. Because "early" means "more years". This wasn't obvious to me right away.
But he is not saying don't get screened! He is not saying there are no cancer treatments! He's saying that the 5-year survival rate, considered alone, is a tricky measure that can fool our intuition. In my case he's right.
---------------------
Details.
Dumb toy model. Let Tumor X kill you exactly 8 years after it becomes detectable in screening. Assume screening is 100% accurate with no false positives. Assume X cancer kills you exactly 2 years after it causes symptoms. Imagine that there is no treatment for X cancer.
In this dumb model, everybody dies at exactly the same time after the tumor became detectable. The people who caught it in screening had more warning, but otherwise they didn't get a better outcome. Even though screening boosts the 5-year survival rate from 0% to 100%.
Never mind his like 7-state Markov model. OMG. Why.
This article, in the world as it exists right now, is wrong about colon cancer. Anyone reading this of a certain age: get that colonoscopy, and those polyps removed. Snip it in the bud. That's the great thing about a colonoscopies - all-in-one screening + treatment.
Large prospective cohorts (Nurses’ Health Study + Health Professionals Follow-Up Study) with long follow-up - screening colonoscopy was associated with a 68% lower risk of death from colorectal cancer overall (multivariable HR ≈ 0.32, 95% CI 0.24–0.45) and showed significant reduction for proximal colon mortality as well (HR ≈ 0.47, 95% CI 0.29–0.76).
>To be clear, all of these conclusions can be true, and in some cases we know they are true, at least in part. For some cancers, treatments have improved, and for some, additional screening would save lives. But to support these conclusions, we need other methods and metrics – notably randomized controlled trials that compare mortality.
the economist put out a piece a few months ago providing just that. Specifically it compares overall cancer mortality rates (and more interestingly, mortality rates adjusted for age) and shows that cancer deaths have been dropping.
The site, in very elaborate ways, is saying that Stage IV cancer (in the author's words 'tumor has spread to distant organs or lymph nodes') is worse than Stage I (localized) cancer.
I don't think there is any person who is aware of the idea of cancer mortality who would equate 'Stage IV' to lead to 'average' survival.
So maybe the article's only point (which is very obvious, and does not require Markov modeling) is that if you increase the number of people who live a long time in a sample, then the average of that sample will go up.
This feels like someone saw a fact on the internet and didn't try to read about it before writing a blog post.
This article is a criticism of reasoning, not health advice or suggestions for cancer screening. Maybe he should put a big warning at the top, rather than explain it throughout. A lot of people seem to be missing it.
We're so used to argument that criticizing logic is taken as criticizing the conclusion.
Near as I can tell, the only valid point the author makes is that since mortality rates increase as cancer progresses stages and since progression through stages takes time, a 5 year mortality rate is not a great metric and it would be better to also have 10 and 15 year mortality rates to determine the degree to which early detection + treatment actually increases life expectancy.
> Because it is based on past cases, it doesn’t apply to present cases if (1) the effectiveness of treatment has changed or – often more importantly – (2) diagnostic practices have changed.
This was my key takeaway. In a society organized around statistics, we're struggling through an era where those statistics expire faster everyday, and faster than new data can be generated. I can almost relate to the mindset that devalues "facts" because they're increasingly complicated, rapidly changing and come with too many caveats.
Weirder still that Taleb misses the base rate flaw in the logic of full-body MRI screening and cancer screening, an observation that is pretty up his alley and is kind of a well-known thing in this domain.
I'm surprised this doesn't really talk about the thing that was most obvious to me: assuming the 5 year survival rate is five years from diagnosis, that means that if a tumor is diagnosed earlier, even if the cancer kills you, your death is more likely to be outside the five year window.
So for example, if you have (hypothetically) an untreatable cancer that would take six years to kill you, if it is diagnosed right away, you would be counted as a survivor, but if you are diagnosed at year five, you'll only survive a year.
>The purpose of the model is to show that we can reproduce the survival rates we see in reality, even if there are no effective treatments.
That's a great argument in the abstract, but it ignores the fact that there are effective treatments for colon cancer. The fact that we can reproduce real survival rates in a counterfactual world where there are no effective treatments for colon cancer does not actually give us a model of the real world because the counterfactual explicitly contradicts known scientific facts.
What you have to do in order to make this argument is to show that there are Markov models where early detection does not work despite the fact that some cancers will cause death if untreated and not if treated. You cannot simply rely on models that have clearly impossible transition probabilities. You need possible models. Or you have to show that the absolutely massive amount of scientific literature and clinical experience about how to treat colon cancer is somehow flawed.
Some people are defending this because the blog post is attacking a specific argument, but I don't see how that can work. I am pretty sure that Nassim Taleb and most other people who are capable of putting together a coherent statistical argument (even a flawed one) understand that colon cancer can be treated sometimes.
Colon cancer is an interesting one, Hank Green [1] recently covered a new paper [2] that showed a massive reduction in colon cancer risk for folks that engaged in moderate, regular, exercise. The authors speculated that mechanical stress leading to increased shedding might play an important role.
Weirdly enough that's the same mechanism hypothesized to play a partial role in why breast feeding is also associated with a reduced cancer risk.
16 comments
[ 3.2 ms ] story [ 32.8 ms ] threadlike the colon cancer thing. he talks about how it would only be more effective to catch colon cancer early if you assume we have treatments for it that would work early. but we don't need to just assume blindly. we already know we do have those treatments!
Essentially every assertion in the article is either an oversimification, cherry picking a random niche situation to highlight, or just flat out factually inaccurate.
Let's take this paragraph for example:
"Catching cancer early is beneficial only if (1) the cancers we catch would otherwise cause disease and death, and (2) we have treatments that prevent those outcomes, and (3) these benefits outweigh the costs of additional screening. This table does not show that any of those things is true."
To address these one by one:
1. Obviously cancer causes disease and death. The same graphic he references makes that abundantly clear. Sure, there might be some rare exceptions (elderly patients with slow growing colon cancer for example), but we're talking about the general population.
2. All cancers have treatment options available in some form (could be chemo, radiation, surgical resection, etc), so this assumption doesn't even make sense to include. Let's assume for a second though that treatments might not be available. Even if that were true, there ARE treatments that can help treat cancer symptoms, and but may not affect the tumor directly. Often these are specific to the specific type of cancer.
3. This assertion is dumb - is the author really trying to argue that providing symptomatic or other relief to a cancer patient isn't a sufficient benefit to warrant additional screening?
I could go on, but you get the point. Some people just like arguing for the sake of arguing I guess.
The specifics of your case will strongly affect what happens to you. And even for cancers that are a guaranteed death sentence, survival has increased significantly in recent years.
One point in the article is that early detection would give you more years to live even if there were no treatment. Because "early" means "more years". This wasn't obvious to me right away.
But he is not saying don't get screened! He is not saying there are no cancer treatments! He's saying that the 5-year survival rate, considered alone, is a tricky measure that can fool our intuition. In my case he's right.
---------------------
Details.
Dumb toy model. Let Tumor X kill you exactly 8 years after it becomes detectable in screening. Assume screening is 100% accurate with no false positives. Assume X cancer kills you exactly 2 years after it causes symptoms. Imagine that there is no treatment for X cancer.
In this dumb model, everybody dies at exactly the same time after the tumor became detectable. The people who caught it in screening had more warning, but otherwise they didn't get a better outcome. Even though screening boosts the 5-year survival rate from 0% to 100%.
Never mind his like 7-state Markov model. OMG. Why.
Evidence: https://www.nejm.org/doi/full/10.1056/NEJMoa1301969
Large prospective cohorts (Nurses’ Health Study + Health Professionals Follow-Up Study) with long follow-up - screening colonoscopy was associated with a 68% lower risk of death from colorectal cancer overall (multivariable HR ≈ 0.32, 95% CI 0.24–0.45) and showed significant reduction for proximal colon mortality as well (HR ≈ 0.47, 95% CI 0.29–0.76).
the economist put out a piece a few months ago providing just that. Specifically it compares overall cancer mortality rates (and more interestingly, mortality rates adjusted for age) and shows that cancer deaths have been dropping.
https://www.economist.com/briefing/2025/07/17/the-world-is-m...
https://archive.is/TNjoi
I don't think there is any person who is aware of the idea of cancer mortality who would equate 'Stage IV' to lead to 'average' survival.
So maybe the article's only point (which is very obvious, and does not require Markov modeling) is that if you increase the number of people who live a long time in a sample, then the average of that sample will go up.
This feels like someone saw a fact on the internet and didn't try to read about it before writing a blog post.
We're so used to argument that criticizing logic is taken as criticizing the conclusion.
This was my key takeaway. In a society organized around statistics, we're struggling through an era where those statistics expire faster everyday, and faster than new data can be generated. I can almost relate to the mindset that devalues "facts" because they're increasingly complicated, rapidly changing and come with too many caveats.
So for example, if you have (hypothetically) an untreatable cancer that would take six years to kill you, if it is diagnosed right away, you would be counted as a survivor, but if you are diagnosed at year five, you'll only survive a year.
That's a great argument in the abstract, but it ignores the fact that there are effective treatments for colon cancer. The fact that we can reproduce real survival rates in a counterfactual world where there are no effective treatments for colon cancer does not actually give us a model of the real world because the counterfactual explicitly contradicts known scientific facts.
What you have to do in order to make this argument is to show that there are Markov models where early detection does not work despite the fact that some cancers will cause death if untreated and not if treated. You cannot simply rely on models that have clearly impossible transition probabilities. You need possible models. Or you have to show that the absolutely massive amount of scientific literature and clinical experience about how to treat colon cancer is somehow flawed.
Some people are defending this because the blog post is attacking a specific argument, but I don't see how that can work. I am pretty sure that Nassim Taleb and most other people who are capable of putting together a coherent statistical argument (even a flawed one) understand that colon cancer can be treated sometimes.
Weirdly enough that's the same mechanism hypothesized to play a partial role in why breast feeding is also associated with a reduced cancer risk.
Fascinating, weird, stuff.
1. https://www.youtube.com/watch?v=4RXSX93mvg8
2. https://www.nejm.org/doi/pdf/10.1056/NEJMoa2502760