The difference between correlation and cause, formalised.
10 topics · 20 curated works
No prior grounding assumed.
The C-Word: Scientific Euphemisms Do Not Improve Causal Inference From Observational Data
Miguel A. Hernán · 2018
Argues that hedging a causal question behind associational language does not make the analysis more rigorous, it just hides the causal assumptions…
Assumes you know the vocabulary.
Statistics and Causal Inference
Paul W. Holland · 1986
Argues causal questions should be posed as comparisons of potential outcomes under different manipulations of the same unit, and that 'attribute…
+5 more at this level
Primary sources and full treatments.
Regression-Discontinuity Analysis: An Alternative to the Ex Post Facto Experiment
Donald L. Thistlethwaite & Donald T. Campbell · 1960
Shows that a treatment assigned purely by whether a score crosses a cutoff lets the jump in outcomes right at the cutoff be read as a treatment…
+12 more at this level
12 of 20 works
Miguel A. Hernán
Argues that hedging a causal question behind associational language does not make the analysis more rigorous, it just hides the causal assumptions being made from the scrutiny they need.
link checked 17 Sept 2026Paul W. Holland
Argues causal questions should be posed as comparisons of potential outcomes under different manipulations of the same unit, and that 'attribute causes' like race or sex fail this test in ways that make many causal claims about them incoherent.
link checked 17 Sept 2026Guido W. Imbens & Thomas Lemieux
Works through the practical choices in an RD estimate — bandwidth, functional form, manipulation of the running variable near the cutoff — that determine whether the design's credibility survives implementation.
link checked 17 Sept 2026Peter C. Austin
Walks through matching, stratification, weighting and covariate adjustment on the propensity score side by side on the same worked example, so a reader can see what each buys and costs relative to the others.
link checked 17 Sept 2026Tyler J. VanderWeele
Argues the common practice of adjusting for anything associated with the outcome is wrong, and gives a smaller graphical rule for which variables to include that avoids inducing new bias by controlling for the wrong ones.
link checked 17 Sept 2026Miguel A. Hernán & James M. Robins
Rebuilds the whole causal-inference toolkit — potential outcomes, identifiability conditions, g-methods for time-varying treatment — around a single running question: what randomised trial would this observational analysis have to emulate to be believed.
link checked 17 Sept 2026Esther Duflo and Sara Fisher Ellison (MIT OpenCourseWare)
Teaches the statistics social science actually relies on, with identification and what a coefficient can honestly be said to mean treated as the central problem rather than an afterthought.
link checked 17 Sept 2026Donald L. Thistlethwaite & Donald T. Campbell
Shows that a treatment assigned purely by whether a score crosses a cutoff lets the jump in outcomes right at the cutoff be read as a treatment effect, founding regression discontinuity as a design rather than a regression trick.
link checked 17 Sept 2026William G. Cochran
Shows that sorting units into five subclasses on a confounder removes most of the bias adjusting for it in full would remove, giving a simple rule of thumb still used to justify coarse stratification.
link checked 17 Sept 2026Donald B. Rubin
Defines a causal effect as the difference between the outcome a unit would show under treatment and the outcome it would show under control, and shows randomisation's real job is making that unobservable comparison estimable.
link checked 17 Sept 2026Paul R. Rosenbaum & Donald B. Rubin
Proves that matching or stratifying on the single scalar probability of receiving treatment given covariates balances those covariates just as adjusting for all of them individually would.
link checked 17 Sept 2026Judea Pearl
Shows that whether adjusting for a variable removes or introduces bias can be read directly off a causal graph, replacing case-by-case reasoning about confounding with a graphical criterion.
link checked 17 Sept 2026This subject genuinely sits in more than one domain. These fields approach the same ground with different methods.