TB: Causal Inference

Overview

Causal inference is the discipline that formalizes the difference between association (what co-occurs) and causation (what would happen if we intervened). Its central claim, articulated by Judea Pearl's do-calculus and the Neyman-Rubin potential outcomes framework, is that prediction and intervention are mathematically distinct problems — and confusing them produces expensive errors in science and policy. For quantitative scientists, including physicists, causal inference is the most underrated methodological toolkit available: HEP already performs implicit causal reasoning via simulation and control regions, but has not yet imported 30 years of formal identification theory from econometrics.

Core Tension

ML optimizes prediction error — the wrong objective when the question is "what happens if I intervene?" You can predict without understanding mechanism, but you cannot identify causal effects without either a formal identification strategy or a randomized experiment. The hardest version of the tension in HEP: physicists already reason causally via simulation (the detector model is a causal graph), but haven't adopted the vocabulary or toolkit to identify effects in the presence of arbitrary confounders without simulation.

Key Insights Across Sources

The Neyman counterfactual frame

  • On Causation: the fundamental problem of causal inference — for any unit, only one potential outcome is ever observed (the one that occurred). The counterfactual is never directly visible; it must be identified from assumptions about exchangeability or structural models.
  • Association = prediction (what co-occurs); Causation = intervention (what would happen if we acted). In CI, some variables must be included regardless of predictive power — unlike ML which optimizes purely for prediction error.

Pearl's do-calculus and the causal revolution

  • The do-operator forces a variable to take a value (intervention) rather than conditioning on its observed value (observation). This is the formal distinction between P(Y|X=x) and P(Y|do(X=x)).
  • Pearl's framework — DAGs, the backdoor criterion, do-calculus — is a complete formal toolkit for identification. It is the most important conceptual advance in statistics since Bayes, yet remains largely unabsorbed outside econometrics.
  • "A wedding between two non-intersecting languages: data and cause-and-effect." (Pearl)

The HEP connection — causal inference without the vocabulary

  • On Causation + Help Claude Help Us (causalinf): a physicist measuring ZZ→4b as a standard candle is doing causal inference — using a known process to identify a parameter in the presence of confounders (backgrounds). HEP has domain-specific names for everything Pearl would call DAG surgery and the backdoor criterion.
  • Simulation-based inference (SBI, arXiv:2010.06439) is the natural bridge: the causal structure of the detector simulation is doing implicit causal work. Formalizing this connection would import a 30-year identification literature that HEP hasn't touched.
  • Learning CI formally gives HEP physicists generalized identification strategies beyond the simulation paradigm — especially valuable as data-driven background modeling replaces MC-based approaches.

Domain expertise × CI = rocket fuel

  • Help Claude Help Us (causalinf): CI tools in the hands of domain experts vastly outperform CI tools alone. The dangerous failure mode is CI without domain knowledge — or domain knowledge that hasn't been formalized into a causal graph.

Contributing Sources

  • On Causation — association vs intervention; Neyman's counterfactual frame; the fundamental identification problem
  • Help Claude Help Us (causalinf) — Scott Cunningham's Substack; domain expertise as multiplier; The Mixtape as the entry point

Resources

  • Cunningham, Causal Inference: The Mixtape (free online) — the canonical econometrics toolkit: DiD, IV, RD, matching, synthetic control
  • Pearl, The Book of Why — conceptual framework for the causal revolution; accessible entry point
  • arXiv:2010.06439 — simulation-based inference in particle physics; the natural bridge from HEP simulation to causal identification

Related Concepts

  • The Map is Not the Territory — CI formalizes the gap between a predictive model (the map) and the actual causal structure (the territory); they can agree on observables and diverge on interventions
  • Second-Order Thinking — CI asks what happens downstream of an intervention, not just what correlates with it; second-order effects are exactly what prediction misses
  • First Principles Thinking — building a causal DAG requires identifying the underlying generative mechanism, not just fitting observational patterns

Cross-Topic Connections

  • AI and Experimental Particle Physics — SBI and ML-accelerated pipelines raise the stakes for causal identification; autonomous analysis agents that confuse association with causation propagate bias at scale
  • Physics and the Nature of Reality — the epistemic question of what "cause" means in a deterministic or probabilistic physical world; CI is an operational framework, not a metaphysical claim
  • Philosophy and Epistemology — CI makes counterfactual assumptions explicit; Hume's problem of induction is the ancestor of the identification problem

Backlinks