For the new kids in back: If you hate statistics, you'll love my free lectures. Putting science before statistics, from basics of inference & causal modeling to multilevel models & dynamic state space models. It's all free, made with love and sympathy.

RT @rlmcelreath: Synchronized behavior without intelligent central control is common in nature. And it's something we actually understand f….

Join millions who have switched to Grok.

I have an old blog post series that discusses these approaches and shows the Bayesian approach in part 3:

There are multiple ways to estimate that ratio, and you should be careful about how you do it! 2-Stage-Least-Squares is popular, but it has poor variance, because estimating a ratio is one of the harder things to do robustly. I prefer a Bayesian approach, unsurprisingly.

Can't just regress Y ~ X, because of U. Let's write down what we know, then solve for b. The covariances are:. cov(Z,X) = a var(Z).cov(Z,Y) = ab var(Z). Solve for b:. b = cov(Z,Y) / [ a var(Z) ]. Ssolve for a:. a = cov(Z,X) / var(Z). Substitute:. b = cov(Z,Y) / cov(Z,X).

In my dept today we discussed deriving estimators from causal models. This can be opaque, but as a simple e.g. instrumental variable Z for estimating X on Y. Z –a–> X –b–> Y, and X <– U –> Y, where U is unmeasured, a and b are path coefs. We want to know b. How? >>.

Do you find the way statistics relates to scientific questions confusing and arbitrary? Do you like drawing owls? Then my free lectures are here for you.

RT @lpachter: This is a good article and I agree with its argument for comprehensible research.

I updated this post with links at bottom that go into more depth on mechanistic and philosophical reasons for troubles in science reform.

link

I haven't read this book, but the authors (Grant & Di Tanna) know their stuff, and they provide code for every script/engine. Meta- and mega-analysis one of those places where Bayes is natural and often easier than non-Bayes.

RT @MicrobiomDigest: Good morning from London, where we start Day 2 of the @royalsociety meeting "The Future of Scientific Publishing". Yes….

For the morning EU crew.

How can we reform science? I have some ideas. But I am not sure you’ll like them, because they don’t promise much.

everyone stop fighting and touch the fluffy belly (beware)

Every Mac knows what Rob Mac is going through

Here's a paper with a skeleton of the idea, but there is really lots of research to do on the structure of workflow networks, now to make them robust, how to development diagnostics and calculi for steps within them.

When colleagues ask me what I'm working on now, I mutter about "scientific workflow". Good scientific projects are entangled networks of assumptions and models and measurements in which the logic of each connection is transparent and can be checked. Working on 2 books and an ERC.

This is what anthropology is about

Central European parlor weasel, fully extended

Ppl asking me about Bayes and causal inference and I have nothing new to say but repeat myself below.