> Stipulating that five years is the wrong period to use, what would the right period be?

The short answer: Whenever death rates plateau for the 99th (or whatever) percentile, relative to progression of the illness. If you wait XX years until the effect of moving around the diagnosis date is largely mitigated statistically, then the error begins to fade.

The long answer: This isn't quite the right question to ask at all, because it assumes away the possibility that the naive "years from diagnosis" is a flawed metric in the first place. The point of the comment you're responding to is that "n-year survival rate (from diagnosis)" as measured is a flawed statistic in general, because the baseline from which the counting starts can be different for people with the exact same outcomes. An illustrative _reductio ad absurdum_ thought experiment here is that of taking two people with the exact same cancer, detecting one earlier, and giving both exactly no treatment (or treatment on the exact same schedule). Despite having identical quality of care and outcome, the person whose cancer was detected earlier will show up as having a better 5-year survival rate, simply because we started counting earlier and her 5-year mark came earlier in the progression of the disease.

This can be mitigated to an extent by controlling for stage of cancer or whatever, and that's probably a good idea, but this is necessarily brittle and hard to scale and study.

You've proved too much. This same argument could be used to show that we can't decide e.g. which drugs or lifestyle choices make death due to cancer more or less likely, or indeed which behaviors make death due to any long-term illness more or less likely.

Congratulations I guess, but you've just put a bunch of medical researchers out of a job. They had to study a long time to do that job!

> This same argument could be used to show that we can't decide e.g. which drugs or lifestyle choices make death due to cancer more or less likely, or indeed which behaviors make death due to any long-term illness more or less likely.

I won't hazard a guess as to which argument you're imagining I made, but I'm not seeing how you jumped to this conclusion at all. As Dylan16807 points out below, "deciding which drugs or lifestyle choices make death due to longterm illness more likely" has approximately nothing to do with an apples to oranges comparison of time intervals starting at different points.

In fact, to the extent that attempts to figure this out do involve comparing intervals, they tend to be ages, which by definition are measured from birth. That's pretty much the highest standard for having a reasonable, stable beginning point when comparing time intervals across people's lives, in a way that "years after diagnosis" doesn't approach by a long shot.

> Congratulations I guess, but you've just put a bunch of medical researchers out of a job. They had to study a long time to do that job!

I'm assuming the juvenile tone is an attempt to cover up your lack of comprehension with bluster. "Congratulations I guess"

I don't see how you reach that conclusion. If your detection is the same, it's easy to see which treatments or lifestyles work better or worse. This seems to me to be a problem unique to measuring the value of different kinds of detection.

If the argument is that detecting cancer at year four isn't worse than detecting that cancer at year zero, then could we specify a year at which it would be worse to detect cancer? It's a deadly disease, so presumably if you wait too long the patient will have died already?

If I've misunderstood this complex and sophisticated argument and we're all ready to admit that detecting cancer later is worse, then the original point that USA is better than UK in this one tiny respect stands.

The argument is that while detecting cancer earlier is usually a good thing, you have to be careful how you measure the effects. Looking at survival X years from detection can give you false positives about what is best.

-

I'll lay out a particular scenario where this happens:

Currently we detect a cancer moderately far along, and aggressively treat it. The five year survival rate is 50%.

We figure out how to detect it two years earlier. We use the same aggressive treatments, and the five year survival rate is now 70%. Hooray! But looking closer, if we wait another two years to correct for the early detection, the survival rate is only 45%. Only some of those tumors would have continued growing. Of those, attacking early is only marginally helpful. In others, the tumor wouldn't have killed the patient, but the radio/chemotherapy killed a quarter of them.

In this case, we magically know we detected everything two years earlier, so we could look at the seven year survival rate instead of five year. But the real world is not so clean. It's very hard to figure out a timescale and normalize everyone to it.

-

Catching cancer earlier screws up your ability to measure survival. A naive analysis will see higher survival rates when smaller/earlier tumors are detected. A sophisticated analysis that corrects for this is actually hard to do.

Even when you do save lives by treating earlier, it's very hard to figure out how many lives are saved, and how much of a confounding factor your detection method is.