Sources checked 4 September 2026
References and further reading
Textbooks and original papers for the topics covered in the course.
Bruce E. Hansen
Open this when a projection argument or asymptotic approximation needs its full mathematical setting. Hansen pairs this volume with Probability and Statistics for Economists. The author also provides separate data, analysis programs, and source-data processing archives.
Work through the assumptions attached to the result you use. A formula surviving a change of notation does not establish that it survives a change of sampling process.
Scott Cunningham
An accessible companion for potential outcomes, DAGs, experiments, RD, IV, panels, and synthetic control. Use it for a second explanation of the designs in chapters 5 through 9.
The official online edition currently carries a work-in-progress notice. Check its corrections and the original methodological papers when implementing an estimator. Causal identification is one substantial part of econometrics; forecasting and statistical modeling ask additional questions.
Brantly Callaway and Pedro H. C. Sant’Anna
Read after the two-group, two-period calculation. Cohort-by-time effects make the target and comparison group visible when treatment begins at different dates.
Identification still needs a valid untreated comparison, suitable parallel trends, overlap, and restrictions on anticipation. Aggregation weights determine which effect a reported average describes. Distinguish simultaneous inference from separate pointwise intervals.
Ashesh Rambachan and Jonathan Roth
A route from asking whether pre-trends look flat to stating how much departure from parallel trends a conclusion can withstand.
The restrictions connecting pre-treatment and post-treatment deviations require economic justification. The procedure accounts for statistical and identification uncertainty. Our design experiment illustrates a simple population-level bound; it does not implement HonestDiD.
Victor Chernozhukov and coauthors
The original framework behind the orthogonal-score and cross-fitting discussion in chapter 10. Read the score construction alongside the assumptions supporting its limiting distribution.
Orthogonality reduces local sensitivity to nuisance estimation. Cross-fitting controls a different overfitting problem. The underlying identifying conditions, nuisance convergence rates, moments, and nondegeneracy requirements remain essential. Prediction quality alone does not establish these conditions.
Isaiah Andrews, James H. Stock, and Liyang Sun
Read when an IV estimate appears precise despite a weak first stage. The review treats identification-robust inference and the complications introduced by heteroskedastic, serially correlated, and clustered observations.
Instrument strength and instrument validity are separate questions. A first-stage F threshold does not test exclusion and does not supply a universal inference guarantee. Match the robust procedure to the design and error structure.
Anastasios N. Angelopoulos and Stephen Bates
A practical introduction to prediction sets and their coverage guarantees. Pair it with chapter 11 and ask exactly which randomness the advertised coverage averages over.
Ordinary split conformal relies on exchangeability of calibration and test observations. Marginal coverage does not imply coverage at every covariate value. Time dependence and distribution shift require additional treatment. A predictive interval is distinct from a confidence interval for a causal effect.
Issa J. Dahabreh and coauthors
Read when a result from enrolled participants is being used to justify a wider rollout. The paper combines trial outcomes with baseline information about the eligible population.
Beyond trial randomization, the argument needs conditional mean exchangeability across participation and participation positivity over the target population. The specified intervention and trial-engagement assumptions also matter. Random assignment within a trial does not establish its representativeness.
Susan Athey and Stefan Wager
A bridge from estimating effects to learning a feasible treatment rule. It makes room for constraints such as budget, simplicity, and the information available when a decision is made.
The welfare objective, policy class, identifying strategy, and nuisance-estimation conditions are inputs to the argument. Evaluate a learned policy using held-out information. A ranking of predicted outcomes is insufficient evidence for a treatment-allocation rule.
Achim Ahrens, Victor Chernozhukov, Christian Hansen, Damian Kozbur, Mark Schaffer, and Thomas Wiemann
A recent explanatory bridge between nuisance functions and a target parameter. Read this alongside the original DML paper when the notation obscures the estimation problem.
This is a current introduction to an established framework. Keep the nuisance definitions consistent across residual-on-residual and structural-score formulations. Changing a nuisance parameterization changes the equation you must derive.
Michal Kolesár, José Luis Montiel Olea, and Jonathan Roth
An advanced reading on sharp bounds for policies that change treatment assignment. The paper shows why the policy question can differ materially from the question answered by an IV coefficient.
Its framework avoids IV monotonicity while retaining other restrictions. Conclusions depend on the specified policy and invariance assumptions. The author’s status page lists revisions requested at Econometrica; this is a working paper, and its scope should be read accordingly.
These existing sites remain separate companions. Their scope, authorship, and development stage vary. A link here does not imply that every older exercise or software example has been revalidated for this edition.
The final two readings cover DML estimation and bounds for counterfactual policy evaluation using instruments.
Books and papers are linked through author or institutional sources. Downloaded textbook files are not distributed with this course. See the edition note for the boundaries of the material.
Return to the chapters →