Heterogeneous treatment effects with a causal forest (GRF recipe)
The full GRF HTE playbook: cross-fit nuisances → causal forest → calibration → AIPW ATE → BLP → RATE → policy.
Difference-in-differences with multiple periods (did)
Staggered-adoption DiD done right: group-time ATT(g,t) → event-study / group / calendar aggregations, with honest pre-trends.
Double machine learning for the 401(k) effect (DoubleML)
Effect of 401(k) eligibility on net assets via PLR / IRM / IIVM with cross-fit ML nuisances — four learners, one honest comparison.
Model, identify, estimate, refute — the DoWhy four-step recipe (DoWhy)
Make your assumptions explicit: draw a causal graph, identify the estimand by the backdoor criterion, estimate it, then actively try to refute it with placebo and confounding tests.
Synthetic control, the tidy way — weights, gaps and placebo inference (tidysynth)
Build a synthetic version of the treated unit from a convex blend of donors, read the treated-minus-synthetic gap, and test it against placebos run on every donor.
Mendelian randomization: genes as instruments for a causal effect (TwoSampleMR)
Use genetic variants as instruments to estimate the causal effect of an exposure on an outcome from GWAS summary data — with IVW plus pleiotropy-robust MR-Egger and weighted-median checks.
An introduction to GRF (getting started)
A minimal first-contact recipe: regression forest, quantile forest, and a causal forest on the same data.
Assessing heterogeneity with RATE (AUTOC & Qini)
Causal forest → train/eval split → RATE with both AUTOC and Qini → TOC plot.
Heterogeneous effects with causal-forest double ML (EconML)
Double machine learning with a forest final stage: partial out nuisance with flexible learners, then read the conditional effect τ(x) — with valid confidence intervals.
Honest sensitivity bounds for parallel-trends violations (HonestDiD)
Stop betting everything on a pre-trends test. Allow the post-treatment trend to deviate within a transparent class, and report the confidence set — and the breakdown value where the effect would vanish.
Matching for causal inference (MatchIt)
Preprocess by matching so groups are comparable, check balance, then estimate the effect on the matched sample — design before analysis.
Bayesian regression discontinuity with credible intervals (CausalPy)
Fit a model on each side of the cutoff, put a posterior on the jump, and report a credible interval for the discontinuity — plus an honest look at how it moves with the bandwidth.
Event-study DiD with Sun & Abraham (fixest)
Fast fixed-effects event study that survives staggered timing — sunab() vs naive TWFE, plotted against the truth.
Draw the DAG, find the adjustment set (ggdag & dagitty)
Before any estimation: encode your assumptions as a causal graph, enumerate the backdoor paths from treatment to outcome, and let the graph hand you the minimal set of covariates to adjust for.
Two-stage difference-in-differences (did2s)
Gardner's 2-stage estimator for staggered DiD: residualize on the untreated, then estimate the event study — fast and timing-robust.
Uplift modelling with S-, T-, X- and R-learners (CausalML)
Estimate who responds, not just the average: fit a family of meta-learners for the CATE, pick the best by validation error, then rank and target with an uplift curve.
Causal forest with time-to-event data (survival)
Censoring check → causal survival forest → RMST-scale AIPW ATE → calibration → report.
Confounder-adjusted survival curves for a treatment (adjustedCurves)
Compare survival between treatment groups after removing confounding — via IPTW, the g-formula or AIPW — instead of a raw Kaplan-Meier that quietly bakes in selection.
Goodman-Bacon decomposition: what your TWFE estimate is averaging (bacondecomp)
A two-way fixed-effects DiD is a weighted average of all possible 2×2 comparisons — including 'forbidden' ones that use already-treated units as controls. This shows you the weights.
Sensitivity analysis for unobserved confounding (sensemakr)
Don't just assume no unobserved confounding — quantify it: robustness value + contour plots benchmarked against your real covariates.
Evaluating a causal forest fit
Did the forest actually capture treatment-effect heterogeneity? Calibration → variable importance → BLP → omnibus tests.
Design & diagnose a randomized experiment (DeclareDesign)
Specify a study as model–inquiry–data–answer, simulate it, and read its diagnosands — bias, power, coverage — before you run it.
Group & conditional effects with DoubleML (GATE / CATE)
Slice the average effect: Group Average Treatment Effects and a CATE surface from a debiased IRM, with simultaneous confidence bands.
Quantile treatment effects of 401(k) eligibility (DoubleML)
Beyond the average: how 401(k) eligibility shifts net financial assets across the whole wealth distribution, estimated orthogonally.
Sharp regression discontinuity with robust bias correction (rdrobust)
Identify the effect at a cutoff: a local-polynomial RD with an MSE-optimal bandwidth and robust, bias-corrected confidence intervals.
Cross-fold validation of heterogeneity
K-fold cross-fitted CATEs → RATE on out-of-fold priorities → honest verdict on heterogeneity strength.
Policy learning via optimal decision trees
Causal forest → doubly-robust scores → policytree → evaluate policy value → plot the tree.
Covariate balance for matching & weighting (cobalt)
Before you trust an observational estimate, prove balance: SMDs, overlap, and a Love plot before vs after adjustment.
Smooth signals with a local linear forest
When the conditional mean is smooth: regression forest baseline → ll_regression_forest → tuning → diagnostics.
Synthetic difference-in-differences (synthdid)
Reweight both control units and pre-periods to build a synthetic control, then apply a DiD correction — robust where plain SC or TWFE struggle.
Learn an interpretable treatment policy (DoubleML policy tree)
Turn debiased CATEs into a rule: fit a shallow, readable decision tree that maximises the doubly-robust policy value.
Estimating ATEs on a new target population
Train a causal forest on the source sample → reweight AIPW to a target population → report transported ATE.
Causal mediation: natural direct & indirect effects (CMAverse)
Split a total effect into what flows through a mediator (indirect) and what doesn't (direct) — with a sensitivity analysis for mediator–outcome confounding.
Dose–response with average potential outcomes (DoubleML APO)
For a multi-valued or continuous treatment: estimate E[Y(d)] at each dose and the contrasts between them, all cross-fitted.
Instrumental variables & 2SLS for an endogenous treatment (ivreg)
When treatment is endogenous, an instrument identifies the complier (LATE) effect via two-stage least squares — after you check the instrument is strong.
Qini curves: automatic cost-benefit analysis
From CATEs to a budgeted treatment policy: causal forest → DR scores → cost matrix → maq Qini curve → pick the budget.
Staggered DiD done three ways (did · did2s · fixest)
My side-by-side for staggered adoption: Callaway–Sant'Anna vs Gardner's two-stage vs Sun–Abraham — do the event studies agree?
Distributional effects: potential quantiles & CVaR (DoubleML)
When the tails matter: estimate potential quantiles and the conditional value-at-risk of a treatment with Neyman-orthogonal scores.
An observational ATE you can defend (balance → estimate → sensitivity)
My checklist for an observational effect: match, prove balance with cobalt, estimate on the matched sample, then quantify hidden-confounding risk with sensemakr.