Cheerios vs the FDA

Via Tyler Cowen, there is this new article.

So, because health research is being cited, Cheerios are a drug? So what is the rationale for regulation of this product? Patient safety? I fear that denying patients food because clinical research has not been done would seem to be an unwise idea.

I do think that poor regulation of products can lead to harm but there does seem to be a difference of kind here. Nor do I think that merely citing research should be an issue: there should be more careful research into diet and not less.

What should your final model be?

In the decision to present the best possible model based on the weight of evidence, should you present the results of Bayesian Model Averaging or of a single selected model based on variables that had strong posterior probabilities in the BMA analysis?

I think you have to present the posterior probabilities. They actually contain MORE information than the selected model. More importantly, the p-values in a selected model don't seem to give a good sense of model uncertainty. This feature is something that I dislike about all automated model selection approaches in that they don't represent competing models that were close to the final model in likelihood but just happened not to be selected.

But it looks like almost no substantive papers use BMA. Why is this?

Measuring changes

Is there anything trickier than measuring changes? Probably, yes, but this has got to be competitive for the difficulty in translating clear concepts into analysis. We can all describe, using ordinary language, the idea that a person may become more physically active with time.

But how do we operationalize this using the tools available in the standard cohort study. Do we look at change? What about change conditional on baseline? But the second quantity is going to have issues with measurement error.

What about categorizing things? Sadly, this creates the annoying problem of regression to the mean and ignores the bigger question of how do we define a meaningful threshold. Let alone the issue of reducing the information content of a continuous variable by breaking it into two categories. This reduction might be defensible if these categories align with clear clinical guidelines and so make results easier to operationalize. But many variables don't have this lucky property.

The more I look at this, the trickier the whole thing becomes.

Mission Creep

Professor in Training has a post on the amazing amount of time that it takes to certify a lab and/or a researcher. I find it very interesting how there has been a slow increase in the amount of overhead to do research. Each step may be both reasonable and justified. But the total effect can be quite counter-productive.

And, I must admit, that I am not altogether clear that it is a better world when the key skill for a junior professor is negotiating bureaucracy and/or finding ways around regulation. I know why making these rules apply to everyone is useful (as people avoid these things if they have any chance at all -- or at least I do). But maybe we should learn to balance total impact on research with protection of lab-workers and research subjects?

Sick Days part 2

A nice follow-up to my post yesterday about the public health issues of poor sick leave policy. This area is a place where there is a bad tendency to spout what good practice is and then put incentives in place that make following such practices highly problematic at best.