Alexander Etz

@stephenjwild "Overthrowing the Tyranny of Null Hypotheses Hidden in Causal Diagrams" by Sander Greenland

What are your favorite books/resources on causal inference? I'm looking for off-the-beaten-path resources, the hidden gems, the provocative hot takes.

@peder_isager It is a very different marginal landscape. In the first, p(C), A, and B are marginally dependent. In the second they are marginally independent; It’s as if they have all been experimentally manipulated. Your question becomes, essentially, “why do colliders block paths”

@AlxEtz That makes sense to me, but then why does the same not happen when proxy p(C) is a cause of C and C is a collider? In this case, I don't open up the collider path between A and B by conditioning on p(C), even if it is very highly correlated with C.

TIL (I think) that conditioning on the parent/ancestor of a confounder partially conditions on the confounder as well. AFAIK the only case where conditioning on an ancestor does the same thing as conditioning on a descendant.

@AdanBecerraPhD @SolomonKurz @stephenjwild For example, Figure 2c shows a confounder pointing to the interaction effect only, but not to G or E. So in that case how would one determine rules for when they can identify the effect of E on D?

DAG folks--thoughts?

@SolomonKurz @AdanBecerraPhD @stephenjwild DAGs are non-parametric. The just represent which variables enter into functions determining other variables. The nature of those functions (e.g. interactions, linearity, etc) are not represented.

@AdanBecerraPhD @stephenjwild Path models are often designed to highlight various beta coefficients in a regression model, hence the concern about an interaction in the original post. It's my understanding DAGs are not designed to distinguish concepts like interactions.

The HDI ROPE decision rule is logically incoherent but we can fix it osf.io/5p2qt/

@cameron_pfiffer @shoshievass @yizhoujin This is similar escholarship.org/content/qt3fk4…

I got a Bayes question for y'all. We (@shoshievass and @yizhoujin) have this Bayesian model that is (a) complicated and (b) run on a TON of data. We're considering just cutting the data up into tiny chunks and then doing a meta analysis to combine estimates.

"We induced different per-condition moods in our participants. Specifically, those in the 'sad' condition watched a harrowing video about a tenured professor losing a lucrative book contract, while those in the 'happy' condition watched a comedy montage of dogs eating homework".

@BA_Schuetze @dingding_peng

@sTeamTraen It is similar in form to a chi-square stats.stackexchange.com/questions/3362…

Is there a mapping from a Kruskal-Wallis H statistic to, say, an F or chi-square statistic?

@sTeamTraen I’d imagine minor differences could be due to using slightly different assumptions in converting things, eg equal vs unequal variances etc

@AlxEtz Thanks. effectsize::cohens_d gives me different results from psych::cohen.d. Both CIs seems to be asymmetrical, but to a different degree, and the asymmetry is very small for two rnorm(50) samples with d~0.8.

In a normal independent-samples t test, is there any obvious reason why the confidence interval for the effect size (Cohen's d) should not be symmetrical about the point estimate?

@sTeamTraen Can't speak to how they would have calculated it. But see e.g. the blurb in the effectsize package below. Easy enough to do by hand, in any case.

@AlxEtz I presume that in the paper I'm reading it has been calculated by hand from a formula that the authors found online (the paper is from 2010 and SPSS didn't do Cohen's d for t tests then, AFAIK).

@FredOswald @sTeamTraen See twitter.com/AlxEtz/status/… If the method used treats the test statistic like a pivot, then this will happen.

@sTeamTraen hm - checked out the psych package and d.ci is asymm (CI toward zero), except when d=0 then went into the d.ci code but couldn't figure out why (tried to make the tol smaller in case converging early, but was a read-only file...ack)

@AshleyJ_Thomas Not sure if there is a specific option. But in general you'd want to be careful testing everything possible. Everything may not be equally interpretable.

Every statistician will come to a point in their career where they feel like they have no choice but to get a tattoo with "what is your research question?" on their forearm

@Notawful Potentially interesting, but too bad there are no figures showing the raw data or somewhere the raw data is shared (that I could see)

@CostelloWilliam I did a lot, but many were just requests to review the stats. I'd recommend limiting yourself to ~3 per year.

Peers. How many peer reviews did you do during grad school?

Teaching evaluations are a funny thing. "Alexander was not a professor, he just regurgitated info on slides." I don't even teach from slides in that course🤷‍♂️