I love fries; you love fries; but should we only eat 6 per serving? Less is more is good advice but why 6 & not 5, or 7?

Unfortunately, there’s prob no way to know if 5 or 6 is better!

Why? Here’s a #tweetorial on estimating causal effects for nutrition. Grab a & get comfy!

Imagine you want to reduce your intake of French fries with the specific goal of reducing your chance of a heart attack.

You need to make 2 decisions: how often should I eat any fries; and how many fries should I eat in a serving?

To help you live your best life (ie eating max safe # of fries), researchers need to ask a pair of causal questions:

•what is the best frequency of French fry consumption to prevent heart attacks?
•what is the best serving size of French fries to prevent heart attacks?

If we had unlimited resources, we could easily design an #RCT to answer both these questions. So let’s start by thinking through the #targettrial.

First question, do we want to know the best fry strategy for life, or among adults?

Since we’re trying to help you decide if you should *reduce* your French fry consumption, let’s assume we want to know about the optimal fry eating strategy for adults.

So, the first step of our #targettrial (inclusion criteria) is sorted: we’ll enroll a bunch of 40 year olds.

Should we exclude anyone? If we want to know about 1st heart attack, we should exclude those who have already had one, but if we want to know about any heart attack then everyone can join.

What about those who have never eaten fries? That depends on our causal question...

Do we want to know the best fry eating strategy for to adopt on your 40th birthday, or do we want to know the best strategy for French fry eaters to adopt?

The first question is easier to answer and let’s us include the never eaters.

What are our target interventions?

We want joint interventions on the frequency & amount of fries:
•Eat fries only Z times per month & only X fries at once sitting.

We can specify ranges, say X = 1 to 100 & Z = 1 to 30.

But 100*30 = 3000 trial arms. That’s a lot of people!

Next, our #targettrial needs a time frame: how long should the trial last?

Our participants are 40 at baseline, but the average age at 1st heart attack for men is ~66 & for women, ~70. So maybe we should follow for ~30 years?

Good thing we have those unlimited resources!

Finally, we need a causal estimand. In an #RCT we usually like to estimate the intention-to-treat (eat?) effect, but that can under- *or* over-estimate the true effect of fries when the compator is active and some participants drop out or don’t adhere to their assigned strategy.

But, we’re asking people to eat X amount of fries Z times per month for 30 years! No way everyone is going to adhere, and prob some people will ghost us (ie drop out). Plus, we def don’t have a French fry placebo.

So, the intention-to-eat effect isn’t going to be very useful.

The other option is a per-protocol effect: heart attack risk if everyone ate the amount & frequency of fries we told them to and stayed in our study.

That’s more complicated, & we need people to record their actual fry eating behavior for 30 years, but in theory it can be done!

Two important caveats on how to estimate the per-protocol effect in our trial:

we absolutely require info on predictors or fry eating & heart attack over time (& same for ghosting)

we must use g-methods if past fry-eating predicts future fry-eating (& ditto ghosting).

Okay, now we have a basic #targettrial for finding the fry-eating strategies for 40 year olds to adopt that minimizes 1st heart attack risk.

Who wants to fund my 3000-arm trial w/ 30 years of follow-up & daily fry consumption diaries, plus regular measurement of confounders?

No takers??



Okay, don’t panic, we may still be able to answer our question: let’s do an observational study! We can use our #targettrial to design it, and we probably don’t even need to change much!

Let’s recap the trial: enroll 40-yr olds w/ no prior heart attack; randomize; record fry eating, health, behavior, & heart attacks for up to 30 years; adjust for post-baseline confounding with g-methods; & learn best fry eating plan based on per-protocol effect.

What’s different in our #obsdata study? Just that we don’t randomize! This means two things.

First, we need to worry about baseline randomization! But that’s hardly a big deal—we were already dealing with 30 years of counfounders, so what’s one more time point?

But the 2nd issue could be tricker.

We wanted 3000 trial arms to cover all the possible combos of fry eating frequency and amount, and that was too many for our trial to be reasonable.

But our #obsdata will *also* only work if there are people who eat every freq/amount combo.

That is, we need data on all 3000 strategies to estimate the heart attack risk under all of those strategies.

Plus we need something called positivity: we need everyone in our study to have had a chance of following every fry-eating strategy.

Returning to 6 fries: The 1st bar to deciding if 6 is the *right* number is frequency. 6 at once, an avg of 6 per day, something else?

Let’s assume:
•eat fries whenever, but only ever X at once.

That’s a bit weird, but cuts our # of interventions down to 100 from 3000

But we still need our data to have 2 things.

At least some people must always eat X=1 to 100 fries every time they eat fries for 30 years.

And everyone has a chance of always eating X=1 to 100 fries every time they eat fries for 30 years (positivity)

To decide between 5 or 6 or 7 fries, our biggest problem is probably going to be that there almost certainly aren’t many 40+ year olds out there who *only ever eat exactly* 5, 6, or 7 fries in one sitting!

And even if we’re willing to incur @f2harrell’s wrath & change our question to a dichotomous x = <50 fries versus x =50-100 fries, how many of you count your fries ever, let alone *every time*?

(PS: even if we switch to container size, we aren’t saved b/c size can vary widely!)

We started off so well! A #targettrial that could tell us exactly what we wantted to know! But it required lots and lots and lots of data and extremely high participant engagement, and we just can’t afford that.

We made one teeny change (baseline confounding) for #obsdata but it fell apart b/c we want to know about things no one does, & things people do, we can’t measure well!

But for *realistic* strategies we *can* measure, a #targettrial helps us design an obs study that *can* work.

The hardest part of #causalinference for nutritional epi is defining realistic interventions & measuring the components to find people who actually follow them.

Compared to that, using g-methods to control for time-varying exposure-confounding feedback is a piece of cake!

But, on the other hand, maybe 6 fries was never the real target anyway https://twitter.com/kareem_carr/status/1071932108082503683?s=21