Either (1) Conduct a lot of very expensive randomized controlled trials on every possible lifestyle intervention and wait a couple of decades for the results,
RCTS are less expensive than you think and correlational approaches more. The alternative is that instead, we run lots and lots of very expensive enormous national surveys and countless analyses of the form ‘blueberry consumption associates with better health (again)’ which still wind up being wrong something like 2/3rds the time, which wrongness itself wastes even more money by sending researchers down dead-ends (looking for the exact flavenoid which improves health) and distorting the general populations’ expenditures & quality of life & trust in science. Correlational trials are only a bargain if you’re trying to maximize citation count and confusion.
On a more positive note: #4 is unacceptable because human life is so valuable. Each year of life is worth scores of thousands of dollars, and good knowledge about lifestyle interventions like resistance exercise or aspirin can be applied to the entire American population of 300 million people indefinitely. So it’s worth paying for lots of trials from any kind of cost-benefit perspective.
And of course there’s all sorts of ways to optimize these trials to reduce the already-trivial cost of running them: factorial trials (why study just one intervention at a time?), trials designed ahead of time to fit into meta-analyses so they can borrow strength, informative priors on parameters like effect size (eg any RR <0.90 is implausible for these kinds of interventions), sequential trials to re-allocate across arms (like Thompson sampling) or just to halt early once enough information has accrued for a decision-theoretic judgement that an intervention has proven useless or useful (and can be rolled out to the population), and use of exotic covariates (the genetics of placebo response looks very interesting for increasing power)...
I agree with almost all of this. I do however think that it would be very hard to convince a sufficient number of people to let their lifestyle choices be assigned by chance, and even harder to convince them to adhere to the assigned randomization arm over several decades.
Note that if you use a factorial design, you are limiting yourself to study only joint interventions. For example, if you conduct an experiment where you first randomly assign alcohol, and then randomly assign smoking, you will be able to figure out the joint effect of these interventions and the interaction between them, but you will not be able to estimate the overall effect of using alcohol, because part of that effect may be mediated by an increased chance of taking up smoking. This can make it difficult to interpret the trials, particularly if we use high dimensional factorial designs.
I am also skeptical of re-allocation across arms, but I’ll have to read up on Thompson sampling.
RCTS are less expensive than you think and correlational approaches more. The alternative is that instead, we run lots and lots of very expensive enormous national surveys and countless analyses of the form ‘blueberry consumption associates with better health (again)’ which still wind up being wrong something like 2/3rds the time, which wrongness itself wastes even more money by sending researchers down dead-ends (looking for the exact flavenoid which improves health) and distorting the general populations’ expenditures & quality of life & trust in science. Correlational trials are only a bargain if you’re trying to maximize citation count and confusion.
On a more positive note: #4 is unacceptable because human life is so valuable. Each year of life is worth scores of thousands of dollars, and good knowledge about lifestyle interventions like resistance exercise or aspirin can be applied to the entire American population of 300 million people indefinitely. So it’s worth paying for lots of trials from any kind of cost-benefit perspective.
And of course there’s all sorts of ways to optimize these trials to reduce the already-trivial cost of running them: factorial trials (why study just one intervention at a time?), trials designed ahead of time to fit into meta-analyses so they can borrow strength, informative priors on parameters like effect size (eg any RR <0.90 is implausible for these kinds of interventions), sequential trials to re-allocate across arms (like Thompson sampling) or just to halt early once enough information has accrued for a decision-theoretic judgement that an intervention has proven useless or useful (and can be rolled out to the population), and use of exotic covariates (the genetics of placebo response looks very interesting for increasing power)...
I agree with almost all of this. I do however think that it would be very hard to convince a sufficient number of people to let their lifestyle choices be assigned by chance, and even harder to convince them to adhere to the assigned randomization arm over several decades.
Note that if you use a factorial design, you are limiting yourself to study only joint interventions. For example, if you conduct an experiment where you first randomly assign alcohol, and then randomly assign smoking, you will be able to figure out the joint effect of these interventions and the interaction between them, but you will not be able to estimate the overall effect of using alcohol, because part of that effect may be mediated by an increased chance of taking up smoking. This can make it difficult to interpret the trials, particularly if we use high dimensional factorial designs.
I am also skeptical of re-allocation across arms, but I’ll have to read up on Thompson sampling.