Rarely comment, but really appreciated this post and a digest of the work of Wysocki et al. (2022), which offers a well-articulated critique of statistical control and its limitations when divorced from causal justification. It echoes the foundational arguments made by Judea Pearl—particularly in The Book of Why—against the overreliance on control groups and purely statistical associations for causal inference. While control groups can reveal associations, they often fail to address hidden confounders and cannot, on their own, establish causality.
Pearl’s “ladder of causation” that encompasses associations, interventions, and counterfactual reasoning, emphasizes that causal understanding requires structured assumptions and formal tools. Causal diagrams (DAGs) and the back-door criterion provide a framework for identifying which variables must be adjusted for to estimate causal effects in observational data. Wysocki et al. effectively highlight this point [citing Pearl in 4 instances], cautioning against common missteps, such as adjusting for colliders or mediators, which can introduce bias or obscure causal pathways. Their discussion of proxy variables is similarly nuanced, recognizing both their potential and their pitfalls.
While the share post and discussion draws heavily from Wysocki et al., it’s essential to recognize the broader context: the sophistication of modern causal inference should enhance, not diminish, public trust in science. The complexity we confront in modeling causation is not a weakness—it’s a necessary response to the multifactorial, dynamic nature of biological systems. Scientific claims like “X causes Y” are appealing in their simplicity, but the real work lies in rigorously navigating a web of interdependent variables, mediators, confounders, and evolving biological processes.
Pearl, J., & Mackenzie, D. (2018). The book of why: The new science of cause and effect. Basic Books.
All very valid points. My only query is to the regression equation on the introductory picture
y = beta0 + beta1 + beta2*x2 + ... + epsilon
Obviously having beta0 AND beta1 standing for beta0 would produce a not-well behaving multilinear regression matrix. I mean the stuff where the first row would read l 1 1 x21 x31 .... xk1 l . Was it on purpose? Or was it just forgetfulness of writing out the complete beta1*x1 term?
DAGs show technically possible and practically improbable issues. The base cases above start with 50% causal relations from the confounders, colliders, and mediators. Even with such improbably strong correlations, to work through two indirect connections reduces the impact from 50% to 50% * 50% = 25%. In more likely cases where the off-line interferers are 20%, that drops down to a 4% influence. Showing significance is hard. First, assume there's probably something there before theorizing about bystander interference.
Love the post but am most impressed and inspired you did this in 2 hours — care to share some of the process around that? did you have images and notes prepared beforehand?
Ironically, I've been meaning to write a post on writing fast, including the little piece of software I made for compliance, and I haven't gotten it done yet.
This is the second LLM-based account I've seen with a pretty substantial readership that goes around posting these meaningless commentaries from GPT or Claude. I guess it's an effective advertising tactic?
This is giving me epistemic vertigo
Thank you for writing this.
Rarely comment, but really appreciated this post and a digest of the work of Wysocki et al. (2022), which offers a well-articulated critique of statistical control and its limitations when divorced from causal justification. It echoes the foundational arguments made by Judea Pearl—particularly in The Book of Why—against the overreliance on control groups and purely statistical associations for causal inference. While control groups can reveal associations, they often fail to address hidden confounders and cannot, on their own, establish causality.
Pearl’s “ladder of causation” that encompasses associations, interventions, and counterfactual reasoning, emphasizes that causal understanding requires structured assumptions and formal tools. Causal diagrams (DAGs) and the back-door criterion provide a framework for identifying which variables must be adjusted for to estimate causal effects in observational data. Wysocki et al. effectively highlight this point [citing Pearl in 4 instances], cautioning against common missteps, such as adjusting for colliders or mediators, which can introduce bias or obscure causal pathways. Their discussion of proxy variables is similarly nuanced, recognizing both their potential and their pitfalls.
While the share post and discussion draws heavily from Wysocki et al., it’s essential to recognize the broader context: the sophistication of modern causal inference should enhance, not diminish, public trust in science. The complexity we confront in modeling causation is not a weakness—it’s a necessary response to the multifactorial, dynamic nature of biological systems. Scientific claims like “X causes Y” are appealing in their simplicity, but the real work lies in rigorously navigating a web of interdependent variables, mediators, confounders, and evolving biological processes.
Pearl, J., & Mackenzie, D. (2018). The book of why: The new science of cause and effect. Basic Books.
All very valid points. My only query is to the regression equation on the introductory picture
y = beta0 + beta1 + beta2*x2 + ... + epsilon
Obviously having beta0 AND beta1 standing for beta0 would produce a not-well behaving multilinear regression matrix. I mean the stuff where the first row would read l 1 1 x21 x31 .... xk1 l . Was it on purpose? Or was it just forgetfulness of writing out the complete beta1*x1 term?
DAGs show technically possible and practically improbable issues. The base cases above start with 50% causal relations from the confounders, colliders, and mediators. Even with such improbably strong correlations, to work through two indirect connections reduces the impact from 50% to 50% * 50% = 25%. In more likely cases where the off-line interferers are 20%, that drops down to a 4% influence. Showing significance is hard. First, assume there's probably something there before theorizing about bystander interference.
I covered this in the specific case for health studies: https://newsletter.unaging.com/p/why-epidemiology-rules-ee9
By the way, that's why - unlike pure math - statistics is an experimental science.
Love the post but am most impressed and inspired you did this in 2 hours — care to share some of the process around that? did you have images and notes prepared beforehand?
Ironically, I've been meaning to write a post on writing fast, including the little piece of software I made for compliance, and I haven't gotten it done yet.
And yes, I've discussed most of this beforehand.
would very much enjoy the post
Cremieux hey could I ask you a question about something
Depends on what. Shoot?
love this
This is the second LLM-based account I've seen with a pretty substantial readership that goes around posting these meaningless commentaries from GPT or Claude. I guess it's an effective advertising tactic?