Linked in Zvi newsletter here

https://www.lesswrong.com/posts/wfz47Ez2r4rQZuYBY/medical-roundup-2

Would this answer include broader reverse causality related reasons, such as an omitted variable causing people to both have lower risk of diabetes and eat more ice cream?

Or a selection effect: patients who are at higher risk of diabetes are more likely to visit the doctor and are instructed to have less ice cream consumption, thus we see that lower risk of diabetes correlates with higher ice cream consumption?

@AmmonLam An omitted variable could be "demographics" depending on the variable. If not, could be "Not correlated" or "Other", depending.

The second effect you describe is an example of "being at low risk of diabetes causes people to eat more ice cream" where the causal mechanism via doctor instructions, so yes, "reverse correlation".

I have some concerns regarding the inference presented in the provided background information about the correlation between ice cream consumption and reduced diabetes risk. Perhaps there are some issues with this inference.

The main evidence presented in the article (https://www.theatlantic.com/magazine/archive/2023/05/ice-cream-bad-for-you-health-study/673487/) consists of two points.

Firstly, there is a study from 2018 that includes various dairy products such as skim/low-fat milk, whole milk, ice cream, yogurt, cottage/ricotta cheese, cream cheese, other cheese, and cream. The study found that the intake of total dairy products or individual dairy products was not associated with cardiovascular disease (CVD) risk, except for an inverse association between ice cream intake and CVD health outcomes (not diabetes).

Secondly, there is a study from 2005, which is Harvard's first observational study of dairy and type 2 diabetes, that suggests that higher low-fat dairy intake is associated with a lower risk of diabetes. It also mentions that trans fatty acids found in milk, cheese, yogurt, and butter (no mention of ice cream) can help reduce the risk of type 2 diabetes.

Based on these two articles, it can be concluded that, first, ice cream is different from other dairy products (thus, the results for other dairy products cannot be simply applied to ice cream); second, the 2005 study and its subsequent research do not specifically mention the impact of ice cream on the risk of type 2 diabetes (in fact, traditional ice cream is not included in the study's definition of low-fat dairy). Therefore, we cannot draw the conclusion that ice cream consumption is correlated with reduced diabetes risk.

Nutrition Science’s Most Preposterous Result: Could Ice Cream Possibly Be Good for You?

Studies show a mysterious health benefit to ice cream. Scientists don’t want to talk about it.

@cr I think you are wrong about the Choi 2005 paper, but I appreciate the prod to actually read it. You write:

Secondly, there is a study from 2005, which is Harvard's first observational study of dairy and type 2 diabetes, that suggests that higher low-fat dairy intake is associated with a lower risk of diabetes. It also mentions that trans fatty acids found in milk, cheese, yogurt, and butter (no mention of ice cream) can help reduce the risk of type 2 diabetes.

The Choi 2005 paper is here:

https://jamanetwork.com/journals/jamainternalmedicine/article-abstract/1152788

It does mention ice cream. It says:

When we examined the association with dairy products stratified by their fat contents, the significant inverse association was primarily limited to low-fat dairy consumption (Table 3). Most individual low-fat dairy products and ice cream showed a similar inverse trend but only skim milk reached statistical significance (multivariate RR per serving, 0.90; 95% CI, 0.83-0.97) (Table 4).

This is a subtle way of admitting that ice cream shows a similar inverse trend to low-fat dairy products. You have to look at Table 4 to see that the effects of ice cream (all types) is higher than skim milk or yoghurt. Also note that there is not a separate category for different types of ice cream.

The downplaying of the ice cream effect in the paper matches The Atlantic's reporting of the politics around this effect and the choices made by the researchers in what to emphasize.

@cr Regarding Korat 2018, it's a huge PhD, and I've not read all of it in detail.

The study does indeed note a link between ice cream and low CVD:

In chapter 2, we examined the association between dairy products and risk of CVD in a T2D patient population. We concluded that total dairy product intake was not associated with risk of CVD among participants with diabetes and neither were dairy products stratified by their fat content. We found a surprising inverse association between ice-cream intake and CVD risk that warrants further research. Replacing dairy products with red and processed read meat is associated with higher CVD risk, whereas replacing dairy product intake with nus is associated 139 with lower CVD risk.

I appreciate that this is a link between ice cream and lower CVD disk, not directly with reduced diabetes risk. Still, this is a study among people with Type 2 Diabetes, and T2D is a risk factor for CVD.

The PhD also has a more direct link in Chapter 1, which shows that dairy fat is a healthy replacement for animal fats and refined carbohydrates, but whole grains are a healthy replacement for dairy fat. This is a "Substitution Effects" result.

Conclusions: In conclusion, dairy fat intake was not associated with T2D risk. However, replacement of dairy fat with carbohydrates from whole grains was associated with lower T2D risk, whereas replacement of dairy fat with other animal fats or refined carbohydrates was associated with higher risk of T2D.

However, they don't break this down further, other than to say that the main contributors of dairy fat are "cheese, pizza, dairy desserts, milk, and butter".

Overall, I think glossing this as ice cream reducing diabetes risk, as The Atlantic does, is a reasonable summary for journalism, but I agree that the details are less clear.

I'm happy to add an answer for "It isn't" if you want to bet on it. I think given the other paper my odds on that are low, maybe 5%.

I may have linked the wrong paper for Choi 2005, sorry. I'll try to fix later, need to sleep now. Sleep also has positive health benefits.

Edit: fixed the link in description and in comment.

@MartinRandall I apologize for the delayed response. I have reviewed our conversation and consulted with my supervisor and colleagues during yesterday's research group meeting regarding the technical details mentioned in our discussion. Firstly, I would like to point out that my previous statement, "It also mentions that trans fatty acids found in milk, cheese, yogurt, and butter (no mention of ice cream) can help reduce the risk of type 2 diabetes," was incorrect. In fact, the research regarding trans fatty acids was a subsequent study published in 2010, not the Choi study from 2005.

@MartinRandall Regarding the article in The Atlantic that you mentioned, I need to point out that the article did not adequately utilize the evidence you mentioned. Even if the evidence you mentioned was fully utilized, it would not robustly lead to the conclusion you anticipated. Allow me to explain.

Unfortunately, neither of us has a particular understanding of P for trend, as it is commonly used in the field of biology and medicine but not widely utilized across various disciplines. P for trend is essentially a trend test. The value of P-trend can be extremely small, even reaching magnitudes of 10 to the power of -18. Currently, we generally consider P-trend < .0001 to be statistically significant. However, in Choi's 2005 report regarding ice cream, the P-trend mentioned is only .07, which means that the 95% confidence interval of the relative risk (RR) values for different levels of ice cream consumption mostly includes 1.00. Even in the groups where the 95% confidence interval of the RR values does not include 1.00, it is likely due to random fluctuations.

This is similar to filming someone on a roller coaster. If you randomly select video clips, you can always find some clips that make it seem like the person on the roller coaster is about to fly out. However, in reality, they are not actually going to fly out. My supervisor is a statistics Ph.D., an executive director of the Statistical Society, and a renowned academic journal reviewer. Although I do not understand the mathematical details of P for trend, I trust his perspective to be credible. Therefore, I am sharing his viewpoint with you, hoping it will be helpful.

@MartinRandall It is delightful to ponder the question you have posed. You consistently come up with intriguing inquiries, whether they are about society, politics, life. ❤ （P.S. I am curious about the green quoting feature that is available for comments. Where can it be found? It seems that the button to use it is not accessible on the PC version of Manifold.)

@cr Thanks. I'd like to add an answer for your supervisor's idea that this is a noise result that is not reflective of a real world correlation. How do you think I might phrase that?

@MartinRandall Your language proficiency and ability to summarize are clearly superior to mine. If I must provide you with advice, I would say, "Currently, no sufficient evidence to reject the null hypothesis." This kind of answer may sound somewhat boring and is unlikely to be favored by many. I, along with others, surely prefer the alternative hypotheses you proposed.

In statistical testing, we typically set the results we do not want or wish to reject as the "null hypothesis" (or establish the null hypothesis based on other more complex criteria), while we set the results we expect as the "alternative hypothesis." Here, the result we do not want to obtain (null hypothesis) would naturally be 'ice cream consumption is not related to diabetes risk.' The current evidence may be insufficient to reject the null hypothesis.

@cr Thanks! Going to add:

Not correlated - if ice cream consumption is not correlated with reduced diabetes risk, for example if studies that appear to show otherwise are due to chance.

I think I want to add an entry for "Happiness" or "Reduced Stress" or "Emotional Wellbeing" or "Hormonal Balance". The mechanism of action here is suggested by https://pulsetcm.sg/can-happiness-help-to-manage-diabetes/ and https://medium.com/@GlycoLeap/does-happiness-affect-diabetes-2b2c0e8f82ae , which don't explicitly mention ice cream, but to speculate:

• Eating ice cream reduces stress

• Reduced stress reduces stress hormones: adrenaline and cortisol

• Reduced stress hormones reduce glucose spikes

• Reduced glucose spikes reduce diabetes risk

This doesn't seem well studied, perhaps because the reverse causation effects are going to kill it. Having diabetes probably makes people less happy. You'd need some careful longitudinal study to show that being unhappy in your 20s makes you more diabetic in your 40s. That's expensive.

I think this is probably nonsense, but I added it and it's sitting at 16%. Aside from anything else, the idea probably dies to substitution effects. If someone is not eating ice cream for their endorphins they're probably eating chocolate, or watching a movie, or doing something else. Plus the unhappiness <--> diabetes link is random speculation. So several reasons this won't pan out.

"It's always income effects" is another nutritional confounder that might be worth tracking. But a quick search suggests the income effects on ice cream consumption aren't huge, and it's not at all clear to me they're different than for other desserts. (< 5 minutes of research.)

The calculated income elasticities were reported to be 0.047, −0.009, 0.070, and 0.149, respectively, suggesting that income increases did not generate substantialchanges in consumption, although they were generally positive.

https://www.sciencedirect.com/science/article/pii/S0022030209713402

@EvanDaniel How about "Demographics" as an answer, which would include income, gender, race, etc?

@EvanDaniel I added "Demographics", let's go.

It's hard to eat a lot of ice cream quickly, and it's mostly water, so sugar intake rate is limited.

@a2bb I could add "Slow Eating"?

@a2bb I added it. It's sitting at 20% which I think is too high.

I think the straightforward answer should be on the list: ice cream is just basically good for you, ice cream in moderation is healthy, and most people who eat some aren't eating it to excess. (Presumably the reason is that milk fats are good, not sugar, but things like "ice cream makes people happy" and "magic pixie dust in the ice cream" also fit this.)

If true, then substitution effects are probably also happening. People who eat ice cream are presumably eating instead of something else. The distinguishing prediction this theory makes is that people who eat half a cup of ice cream per day have better outcomes than otherwise similar peers eating no ice cream and also not eating the equivalent amount of desserts. So substituting ice cream for desserts is good, but so is substituting ice cream for the average diet item or even average non-dessert item (at least on the margin at half a cup a day levels).

If true we might also see this in things like people who drink a lot of cream with their coffee, but getting up to half a cup a day that way seems like it might be rare.

@EvanDaniel Maybe I could split this off as "Healthy milk fats" from "Other"?

@MartinRandall Works for me!

With the caveat that, conditional on "healthy milk fats" being true (and the result remaining robust) I expect the overall result to be something like:

• Milk fats good

• Sugar bad

• The milk fats : sugar ratio in ice cream is net good

• Substitution for other less healthy desserts is also occurring

So I think if ice cream ends up looking better than a typical dessert food but less good than a typical food this resolves as substitution effects, if better than a typical food and a typical dessert food it resolves as "healthy milk fats". Does that sound about right?

Do you also want "stops replicating" on the list? Or confounding problems other than reverse causation, that amount to variants of "observational studies continue to replicate but the result isn't real"?

@EvanDaniel Gah. I hadn't even considered that the entire thing might stop replicating.

@EvanDaniel Yes, that sounds right. I suppose it's not exactly "better than a typical food" because there is some notion of a balanced diet.

Maybe I should make the new answer "Nutritional value" so it's not specifically about milk fats vs fats in general vs milk proteins vs whatever else it turns out to be.

@MartinRandall Works for me. I assume if this is what's going on then things probably resolve as a mix anyway (including some substitution effects), so I'm betting on a fuzzy outcome regardless. The question of "typical food" is basically trying to get at what substitution effects "feel like" substitution effects. "Oh, people who eat ice cream are improving on eating candy" is one thing, "improving on eating a salad and sandwich" is another. I'm not even sure if we know whether "ice cream is good for you, all non-ice cream food consumption held constant" or "ice cream is good for you, total calories held constant" yet.

@EvanDaniel Added "Nutritional value" for you, and added some words distinguishing the called out options.