RT @FanLiDuke: I was often asked by practitioners about power calculations for causal inference with observational data, a hard problem wi…

@mzloteanu Good post. Sander Greenland’s work deserves a wider audience outside of epidemiology.

@infornomics @ajordannafa This looks fun. Looking forward to exploring. Thanks for the post.

@analisereal This is 👍 🤍

@matt_blackwell On this topic, a great interview of Jamie Robins by Stefan Wager, here:

RT @_MiguelHernan: 1/ That "immortal time" is so frequent in survival analyses for #causalinference is fascinating. Because "immortal time…

Using three waves of national panel data, we estimate well-defined causal effects of religious service attendance on charitable giving among New Zealanders and also include measures of help received:

@ryancbriggs Excellent wish! And even with rich panel data, a causal question of marriage’s effect on happiness remains to be defined, see section 4 here:

@SolomonKurz p.s. you’d know how to do this, was just thinking it out loud.

RT @mtdemangeeos: Google spent $80M studying effective teams. Their shocking discovery? Perks & ping pong tables don't matter. Here's wh…

Two Key Points: Point 1. Treatment Assignment vs. Treatment Received: Randomisation assigns the treatment but does not ensure compliance. We often need causal inference methods suited for observational data to assess the causal effect of treatment (per-protocol effect). Point 2. Avoid Conditioning on Post-Treatment Variables: Doing so can bias your results. Causal DAGs help clarify these issues. I’ve made this mistake myself -- I can do better; perhaps, so can you!

Recommended Reading: •Montgomery et al. on post-treatment conditioning: 🔗 •Hernán & Robins on estimating the intent-to-treat effect: 🔗

Many questions remain: * How might measurement error or selection biases threaten our experimental causal inferences? * What do we require to evaluate effect modification? * Which experimental designs should we implement? * Which statistical estimators are appropriate, when, and why? These questions are the basis of productive research in causal data science. I hope some readers are inspired to investigate these topics and contribute their own work.

Causal Directed Acyclic Graphs (DAGs) are like chainsaws: used well, they slice through heavy tasks; used poorly, and you might lose a limb. Here's the first of four tutorials in @Journal_EHS for human scientists new to causal inference: Tips: - Understand that causal inference involves a larger workflow that starts by stating a clear causal question & target population. - Understand that larger workflow. E.g. consider you must also address positivity & causal consistency assumptions. - Your DAG is a tool to clarify unmeasured confounding, not a paintbrush for describing all reality. - Group common causes, let the spatial arrangement reflect chronological order. - Like time's arrow, ensure your DAG doesn't loop back. - When in doubt, present multiple DAGs. - There's lots of other material I reference in the article. I got my start from @_MiguelHernan's course, which is *free* here: In later posts, I'll describe my other tutorials and why I'm so passionate about causal inference. The full special issue with lots of great work is here. Thanks to @EffersonCharles @tavitonst for driving it: #CausalInference #DAGs

@JkayFlake I should add I never get politics in my feed, only academic stuff, and nearly all of it mutually supporting. I really like that.