Scholar

Richard McElreath

Videos

Richard McElreath is an American professor of anthropology and a director of the Max Planck Institute for Evolutionary Anthropology in… more

H-index: 55
Psychology 22%
Sociology 21%
more

by Richard McElreath

rmcelreath.bsky.social
I guess because software allows it, ppl keep trying to estimate diversification rates on phylogenies. This is not in principle possible, because infinite combinations of diversification and extinction rates can explain almost any tree. Short, clear recent-ish paper: www.nature.com/articles/s41...
4 16

by Richard McElreath

by Richard McElreath

by Richard McElreath

rmcelreath.bsky.social
This is a nice paper. It faces down a common problem: in order to explain to why a common approach cannot work even in theory, we first need to teach a framework in which regression is not magic that tells us which variables on right cause the variable on left. It's exhausting.

Anyway great paper!
pwgtennant.bsky.social
Peter Tennant @pwgtennant.bsky.social · 1d
academic.oup.com/ije/article/...
Analyses of ‘change scores’ do not estimate causal effects in observational data
AbstractBackground. In longitudinal data, it is common to create ‘change scores’ by subtracting measurements taken at baseline from those taken at follow-u
academic.oup.com
1 9 54

by Richard McElreath

by Richard McElreath

Reposted by: Richard McElreath, Jill A. Jacobson

pwgtennant.bsky.social
academic.oup.com/ije/article/...
Analyses of ‘change scores’ do not estimate causal effects in observational data
AbstractBackground. In longitudinal data, it is common to create ‘change scores’ by subtracting measurements taken at baseline from those taken at follow-u
academic.oup.com
1 2 16

by Richard McElreath

Reposted by: Richard McElreath, Gordon Hodson, Martin Tomko , and 2 more

dingdingpeng.the100.ci
Happy to announce that I'll give a talk on how we can make rigorous causal inference more mainstream 📈

You can sign up for the Zoom link here: tinyurl.com/CIIG-JuliaRo...
Causal inference interest group, supported by the Centre for Longitudinal Studies Seminar series 20th October 2025, 3pm BST (UTC+1) "Making rigorous causal inference more mainstream" Julia Rohrer, Leipzig University Sign up to attend at tinyurl.com/CIIG-JuliaRohrer
3 51 140

Reposted by: Richard McElreath, Paul Goldsmith-Pinkham, Barrett

avehtari.bsky.social
I'm looking for a doctoral student with Bayesian background to work on Bayesian workflow and cross-validation (see my publication list users.aalto.fi/~ave/publica... for my recent work) at Aalto University.

Apply through the ELLIS PhD program (dl October 31) ellis.eu/news/ellis-p...
ELLIS PhD Program: Call for Applications 2025
The ELLIS mission is to create a diverse European network that promotes research excellence and advances breakthroughs in AI, as well as a pan-European PhD program to educate the next generation of AI...
ellis.eu
30 38

by Richard McElreath

rmcelreath.bsky.social
Someone sent this to me and I want to hug, um, share it with all of you (src @theunderfold.bsky.social )
comic by @theunderfold.bsky.social
1 9 110

by Richard McElreath

rmcelreath.bsky.social
Thinking about and discussing this more with colleagues, I'd really like a continuous time solution, like the Gillespie algorithm but for perfect conterfactuals as defined in thread below. I can't find that this has been done, but somehow believe some rogue chemist has already worked it out
rmcelreath.bsky.social
Richard McElreath 🐈‍⬛ @rmcelreath.bsky.social · 6d
Are we doing simulations wrong? This paper convinced me we are. doi.org/10.1098/rstb... Usually we run 2 sets of "worlds" w and w-out intervention. Gives large uncertainties that include negative (harm) effects of interventions that are actually always positive (beneficial)!
Figure 5. Time series showing cumulative number of cases averted at each time caused by the intervention calculated using our method (single-world) and a standard method. Shaded regions denote 90% confidence intervals. Note that there is more variation in the middle of the epidemic, so it may seem as though the number of cases averted is large during those times. (Online version in colour.)
18

by Richard McElreath

rmcelreath.bsky.social
The most romantic sentence in the German language: Möchtest du mit mir eine Kartoffel essen gehen? (Today and tommorrow in Leipzig at Augustusplatz) kartoffelverband-sachsen.de/saechsisches...
4 2 34

by Richard McElreath

rmcelreath.bsky.social
YES THE OLD MAGIC
2 34

Reposted by: Richard McElreath, Ingo Rohlfing, Andreas Haupt

dingdingpeng.the100.ci
Just finished reading this *excellent* article by Gabriel et al. which discusses which effects can be identified in randomized controlled trials. With DAGs!>

link.springer.com/article/10.1...
Elucidating some common biases in randomized controlled trials using directed acyclic graphs Although the ideal randomized clinical trial is the gold standard for causal inference, real randomized trials often suffer from imperfections that may hamper causal effect estimation. Stating the estimand of interest can help reduce confusion about what is being estimated, but it is often difficult to determine what is and is not identifiable given a trial’s specific imperfections. We demonstrate how directed acyclic graphs can be used to elucidate the consequences of common imperfections, such as noncompliance, unblinding, and drop-out, for the identification of the intention-to-treat effect, the total treatment effect and the physiological treatment effect. We assert that the physiological treatment effect is not identifiable outside a trial with perfect compliance and no dropout, where blinding is perfectly maintained Table 1 showing the Identifiability of target estimands depending on whether there is blinding, full compliance, and no drop-out An example DAG from the paper. Fig. 4: A blinded trial with noncompliance. U are unobserved confounders, Z is treatment assignment, C is compliance, X is the realized treatment, S is the subject's physical and mental health status, Xself and Xcln are the treatment that the participant and the clinician believed the participant received, Y is the outcome.
4 23 120

by Richard McElreath

rmcelreath.bsky.social
The algorithm is more intensive than usual method. So some algorithm research is needed here. And getting ppl to do the right thing will need friendly tool support. The authors of paper provide their code, which is a great start to templating this to generic ABMs or ODEs. github.com/HopkinsIDD/c...
1 9

by Richard McElreath

rmcelreath.bsky.social
We want instead "perfect" counterfactuals that change only the intervention (and consequences of it) and compare across same worlds in all other events. This can be done by computing a potential event graph first, and then forking it and applying intervention logic.
Figure 3. ‘Single-world’ simulation process (continued). In order to measure the impact of the intervention on the epidemic, we use the PEG (figure 2) to construct two epidemics: one uncontrolled (left), one with intervention (right). We start by setting the actual state (denoted by colouring the whole circle) of each individual at time t0 to their initial condition, and then changing their state according to the intervention, in this case setting p2 as vaccinated (dark green). We then prune events between t0 and t1: removing inconsistent events and allowing the intervention to stochastically remove events (b,c). We then use those events to determine the actual state of each individual at t1, allow the intervention to alter that state, then prune inconsistent events again (d,e). We repeat the process for times t2 ( f,g), t3 (h,i) and so on. This final graph can be used to extract our outcome of interest. Any graphs made from the same PEG represent the results of different interventions in a ‘single world’.
1 8

by Richard McElreath — Reposted by: David Rosnick

rmcelreath.bsky.social
Are we doing simulations wrong? This paper convinced me we are. doi.org/10.1098/rstb... Usually we run 2 sets of "worlds" w and w-out intervention. Gives large uncertainties that include negative (harm) effects of interventions that are actually always positive (beneficial)!
Figure 5. Time series showing cumulative number of cases averted at each time caused by the intervention calculated using our method (single-world) and a standard method. Shaded regions denote 90% confidence intervals. Note that there is more variation in the middle of the epidemic, so it may seem as though the number of cases averted is large during those times. (Online version in colour.)
4 12 89

by Richard McElreath

Reposted by: Richard McElreath, Frédéric Delsuc

richardsever.bsky.social
Another record month for bioRxiv - and further evidence the pandemic spike+dip was just that and growth continues. Thanks to all involved and that includes 🫵
1 43 140

by Richard McElreath

rmcelreath.bsky.social
So another example of this perhaps doi.org/10.1098/rsos...
The natural selection of bad science | Royal Society Open Science
Poor research design and data analysis encourage false-positive findings. Such poor methods persist despite perennial calls for improvement, suggesting that they result from something more than just m...
doi.org
1 2

by Richard McElreath

rmcelreath.bsky.social
I suspect researchers want to find p<0.05 and collinearity makes that harder, so they invented procedures for searching model space so they could get more p<0.05.

There is no statistical philosophy in which those procedures are good for prediction or inference. Good for publishing though
2

by Richard McElreath

rmcelreath.bsky.social
In summary
2 1 13

by Richard McElreath

rmcelreath.bsky.social
I am told that his Orthogonal series is his best work, so I might read those books next. But happy to receive opinions.
6 3

by Richard McElreath

rmcelreath.bsky.social
I recently read Egan's Incandescence. This is a book that people hate or love. I liked it! It doesn't condescend to the reader, but expects you to pay attention and figure out the major plot links on your own. Also endless pages of aliens talking about geometry.

Not going to be a movie anytime soon
cover of Greg Egan's Incandescence
1 19

by Richard McElreath

rmcelreath.bsky.social
My first attempt to dispel concern with this kind of naive correlation filtering: It can arise from different causes and vanish in a properly specified model. Example: Data generated by this DAG

X –> Z –> Y

can have highly correlated X and Z. But in model Y ~ X + Z it won't show any "collinearity"
1 1 9

by Richard McElreath

rmcelreath.bsky.social
A serious failure case is that many biologists seem to have been taught that they should inspect bivariate correlations between predictors and not include pairs with high correlations. That is not even "collinearity" as statisticians define it (where is about joint info in additive model).
2 3 9

by Richard McElreath — Reposted by: Richard Shaw, Jill A. Jacobson

rmcelreath.bsky.social
Was asked about collinearity again, so here's Vahove's 2019 post on why it isn't a problem that needs a solution. Design the model(s) to answer a formal question and free your mind janhove.github.io/posts/2019-0...
tl;dr Collinearity is a form of lack of information that is appropriately reflected in the output of your statistical model. When collinearity is associated with interpretational difficulties, these difficulties aren’t caused by the collinearity itself. Rather, they reveal that the model was poorly specified (in that it answers a question different to the one of interest), that the analyst overly focuses on significance rather than estimates and the uncertainty about them or that the analyst took a mental shortcut in interpreting the model that could’ve also led them astray in the absence of collinearity. If you do decide to “deal with” collinearity, make sure you can still answer the question of interest.
3 33 110

by Richard McElreath

by Richard McElreath

References

Fields & subjects

Updated 1m