Comment by KempyKolibri | Hacker News Reader

Comment by KempyKolibri | Hacker News Reader

Parent Full thread

A null finding in a study does not mean it “doesn’t matter”. Let’s take an extreme example: we want to know if exercise reduces mortality, so we randomise one arm of a cohort of 20 people to an exercise regime, while the other arm does nothing. After four weeks, we compare the death rate in each group. There’s no significant difference.

Does this mean exercise “doesn’t matter” for mortality, or it’s in the “doesn’t affect category”? No, of course not - the finding was null because the timescale was too short and even then, the statistical power from a cohort of 20 is going to be very low.

Likewise, if you look at the study characteristics, most of the studies were too short to find a significant outcome on an insensitive endpoint like CVD mortality or ACM. FWIW, Dayton 1967 had a followup of 8 years and did find significant increases in mortality endpoints in the SFA group, but those results are pulled towards null in the meta summation by the other, shorter studies.

CVD events are more sensitive because they don’t require the participants to die of a heart attack in order to register as a data point.

So no, there’s no lying or cherry picking going on here. In order for that to be the case, you’d have to argue that angina and non-fatal heart attacks are not negative health outcomes. If they are, then it’s demonstrably the case that SFA consumption is associated with negative health outcomes, and Nina is telling porkies.

As for your accusation that I’m lying - what false claim did I make? Be specific.

Ok, lets nitpick further:

for cvd risk mortality if you exclude the largest single study that also had a long duration (WHI 2006) it made CVD risk mortality significant, so one of those long duration studies is dragging the results the opposite way.

We need to remember we are talking about relative risk so what does a 17% increase correspond to? 15 more incidents per 1000 participants with a ci of 24 to 2 incidents per 1000 participants among studies that were mostly moderate to high risk individuals. It's a rounding error on a low risk event that is probably an even lower risk event for most of the population. If you look at something that actually exists and is relevant you will see 100% plus changes to relative risk. All cause mortality relative risk is 2.29 in this study on the effect of smoking on women, for example: https://pmc.ncbi.nlm.nih.gov/articles/PMC6219821/

Figure six explores saturated fat cut offs and all but 2 of the events trend down for the last observation which is only at 13% energy from saturated fat, probably need to look at higher saturated fat levels to make sure the chart isn't an upside down U and we aren't stuck in the most deadly part of the curve arguing about whether we should cut more when we could increase consumption more to solve the issue.

I'm sure there's more but I am out of time.

> for cvd risk mortality if you exclude the largest single study that also had a long duration (WHI 2006) it made CVD risk mortality significant, so one of those long duration studies is dragging the results the opposite way.

Right, and the proposed pathway by which SFA increases CVD risk is via increases in ApoB/LDL-c. What LDL-c differences did they achieve in WHI 2006?

Additionally, the substitution is important. DGs recommend replacing SFA with PUFA, whereas in WHI the intervention group largely replaced SFA with CHO. Interestingly, the control group had higher levels of PUFA and MUFA, which may also explain the paltry change in serum cholesterol.

If you want to claim that 15 more incidents per 1000 of a disease that is one of the top killers in the western world is a rounding error, then go for it. I don’t think that’s a reasonable position, personally.

As for the speculation about higher levels being healthy - speculate away, but I don’t see why anyone should believe it’s the case when a) it runs contrary to the body of evidence on the subject b) there’s no actual evidence backing up that speculation.

But your example is not reflective of the study. Are you saying that the 17% reduction is for some reason significant but the other ones, all of which would inconveniently disagree with the result you want, are not, even though they are in the same study?

IOW, you're saying that among the study results, all that agree with your POV are valid, all that don't are invalid. That's quite some bias there.

The answer to your question is literally answered by my comment that you’re replying to. Frequentist statistics cannot be used to affirm the null. That is, you cannot say “cardiovascular deaths was not significantly associated, therefore SFA does not cause CVD mortality”.

So I’m not disagreeing with or omitting anything in the study. The study said no significant association with CVD mortality. Ok, no problem. That doesn’t mean SFA doesn’t cause CVD mortality.

However, the study does show that SFA is associated with CVD events. So there’s a significant finding. It’s not cherry picking, this is just how frequentist statistics works.

Reply on news.ycombinator.com