TLDR: Study on death avoidance, which interests a lot of people here, and commentary on what sort of informative priors we should have about health hypotheses.

From Steve Sailer, who is responding to Andrew Gelman, who got sent this study. An observational study showed that people who consumed nuts were less likely to die; Gelman points out that the study's statistics aren't obviously wrong. Sailer brings up an actual RCT of Lipitor from the 90s:

The most striking Lipitor study was one from Scandinavia that showed that among middle-aged men over a 5-year-period, the test group who took Lipitor had a 30% lower overall death rate than the control group. Unlike the nuts study, this was an actual experiment.

That seemed awfully convincing, but now it just seems too good to be true. A lot of those middle-aged deaths that didn't happen to the Lipitor takers didn't have much of anything to do with long-term blood chemistry, but were things like not driving your Saab into a fjord. How does Lipitor make you a safer driver? 

I sort of presumed at the time that if they had taken out the noisy random deaths, that would have made the Lipitor Effect even more noticeable. But, of course, that's naive. The good folks at Pfizer would have made sure that calculation was tried, so I'm guessing that it came out in the opposite direction of the one I had assumed. Guys who took Lipitor everyday for five years were also good about not driving into fjords and not playing golf during lighting storms and not getting shot by the rare jealous Nordic husband or whatever. Perhaps it was easier to stay in the control group than in the test group?

Here’s how I would approach claims of massive reductions in overall deaths from consuming some food or medicine:

Rank order the causes of death by how plausible it is that they are that they are linked to the food or medicine. For example:

1. Diabetes

2. Heart attacks

3. Strokes

4. Cancer

5. Genetic diseases

6. Car accidents

7. Drug overdoses

8. Homicides

9. Lightning strikes

If this nuts-save-your-life finding is valid, then the greater effects should be found in causes of death near the top of the list (e.g., diabetes). But if it turns out that eating nuts only slightly reduces your chances of death from diabetes but makes you vastly less likely to be struck by lighting, then we’ve probably gotten a selection effect in which nut eaters are more careful people in general and thus don’t play golf during thunderstorms, or whatever.

Table 3 of the paper breaks out the hazard ratios by cause of death. The most impressive effects (as measured by the right tail of the 95% CI for pooled men and women for any nut)¹ are Heart Disease, All Causes, Other Causes, Cancer, Respiratory Disease, Stroke, Infection, Diabetes, Neurodegenerative Disease, and Kidney Disease.

Steve's categories and the paper's categories don't overlap very well. But it looks to me like if you follow Steve's logic, it's reasonable to believe that nuts have a protective effect against heart disease, and then most of the other effects or non-effects have a common cause with nut consumption, like healthiness / conscientiousness / whatever, rather than being caused by nut consumption. Note the strong negative relationships between nut consumption and BMI or smoking, and the strong positive relationships between nut consumption and physical activity or intake of fruits, vegetables, or alcohol. The hazard ratios are calculated controlling for those variables, but it's still reasonable to see there being a hidden 'health-consciousness' node which noisily affects all of those nodes.

It's also interesting to look at the negative results- the hazard ratio for neurodegenerative disease and stroke was roughly 1, implying that nut-eaters and non-nut eaters had comparable risks, despite 'other causes' having a hazard ratio of 0.87. That weakly implies to me that either health consciousness has no impact on neurodegenerative disease and stroke, or that nuts are harmful for those two categories.

Since heart disease is a huge killer (24% of all deaths in the study group), this study seems like moderate evidence in favor of eating nuts, but it's likely that the total study's effect is overstated. (The study also suggests that tree nuts are probably superior to peanuts; I know various QS people have raised concerns that the kind of nut matters significantly.)

1. This is a heuristic for impressiveness, not the point estimate. It looks like nuts have the strongest effect for kidney disease, with a mean hazard ratio estimate of 0.69- but the upper bound of the 95% CI is 1.26, because only a handful of people died due to kidney disease. The heart disease hazard ratio estimate is 0.74 (0.68-0.81), which is much more believable, even though the point estimate is slightly higher. The point estimate for diabetes is 0.80 (0.54-1.18), which has a mean estimate that's only slightly worse, but diabetes again killed far fewer than heart disease. If you order them by point estimates, the paper is stronger evidence for nuts being useful for dietary reasons, and which method you prefer depends on your priors for how representative this sample is.