Microeconometrics

Microeconometrics — Designing Credible Counterfactuals

An identification-first course in causal inference and microeconometric methods

Microeconometrics — Designing Credible Counterfactuals

Microeconometrics is not a catalogue of estimators. It is a disciplined evidence chain:

\text{question}\rightarrow\text{estimand}\rightarrow\text{assignment process} \rightarrow\text{identification}\rightarrow\text{estimation} \rightarrow\text{inference}\rightarrow\text{decision boundary}.

The estimator comes after the counterfactual argument. A precisely estimated coefficient can still answer the wrong question.

Two running cases

Northbridge Pathways Scholarship

Northbridge offers a scholarship to some school-leavers. Eligibility, offer, take-up, university entry, completion and later earnings are different variables.

Research object	Example
treatment	scholarship offer, receipt or university enrolment
outcome	entry within one year, completion within five years, or earnings
assignment	lottery, score cutoff, phased district rollout or self-selection
possible estimand	ITT of an offer, LATE of enrolment, ATT for recipients

Municipalities adopt a clean-air policy at different dates. Pollution, traffic, retail activity and health visits are observed over time. This case develops panel, difference-in-differences and synthetic-control reasoning.

Both cases are synthetic. Real studies are named and linked explicitly.

Learning outcomes

By the end, you should be able to:

define treatment, outcome, population, horizon and estimand before selecting a model;
represent selection and timing with potential outcomes and causal diagrams;
distinguish a linear projection from a causal effect and choose defensible controls;
explain IV, fixed effects, DID, RDD, matching and synthetic control through their counterfactual source;
diagnose weak instruments, non-parallel trends, manipulation, poor overlap and failed pre-treatment fit;
match standard errors and uncertainty to sampling and treatment assignment;
interpret binary, multinomial, count, censored and selected outcomes without confusing coefficients with effects;
use machine learning for nuisance estimation and heterogeneity without claiming that prediction creates identification;
report sensitivity, external validity, data provenance and a reproducible evidence chain.

The main route suits advanced undergraduates and taught postgraduates. Graduate extensions add formal identification, weak-identification-robust inference, staggered-treatment estimands, partial identification and policy learning.

Preparation

You need probability, expectations, confidence intervals, matrix-free OLS intuition and basic Python or R reading. Prior causal inference is not required.

Ten-minute readiness check

Why can the same person’s outcome under treatment and no treatment not both be observed?
Does adding more controls always reduce omitted-variable bias?
Is a strong first-stage relationship enough to validate an instrument?
Does an insignificant pre-trend prove parallel trends?
If an RDD estimate is credible at a cutoff, does it identify an effect for everyone?

Answers: the missing potential outcome is counterfactual; no—mediators and colliders can introduce bias; no—the exclusion and assignment arguments remain; no—pre-tests may have low power and the post-treatment counterfactual is still unobserved; no—the canonical estimand is local to the cutoff.

Course route

Module	Central question	Main deliverable
0. Causal questions and designs	What effect, for whom, under which assignment process?	identification memo
1. Regression and inference	What does OLS estimate, and when can it support a causal claim?	regression and uncertainty audit
2. Instruments and panel data	Which variation isolates treatment, and what remains unidentified?	IV or within-unit design brief
3. Policy evaluation designs	How do time, thresholds and comparison units construct $Y(0)$ ?	DID/RDD/weighting design
4. Choice and limited outcomes	How does the outcome’s support change modelling and interpretation?	marginal-effect report
5. Modern causal analysis	How should flexible prediction, sensitivity and policy learning enter?	cross-fitted or sensitivity analysis
6. Research workflow	Can another researcher audit the complete evidence chain?	replication-ready research record
Capstone	Which design best evaluates the Pathways expansion?	policy memo plus technical appendix
Appendix	Which formula, diagnostic or source is needed?	revision and evidence map

The identification memo

Complete this before running a regression:

Field	Pathways example
decision	expand, redesign or stop scholarship offers
unit	one eligible school-leaver
treatment	offer received by 31 August 2024
outcome	degree completion within five academic years
estimand	ITT among applicants in the 2024 eligibility window
assignment	lottery within score band 68–72
counterfactual	outcomes of comparable applicants assigned no offer
interference	peer and university-capacity spillovers
missingness	completion not mature; migration may hide outcomes
inference	assignment at applicant level, schools may induce dependence
scope	applicants in participating districts and score band

Changing “offer” to “enrolment” changes the estimand and usually the identification argument.

Four claims that must stay separate

Claim	Example	Evidence needed
descriptive	recipients complete at a higher rate	transparent sample and denominator
predictive	prior scores predict completion	time-valid out-of-sample performance
causal	an offer raises completion by 6 points	credible treatment assignment and counterfactual
policy	expand offers to another score band	causal effects, costs, capacity, distribution and transportability

A causal estimate need not answer a policy question if the affected population, treatment version or capacity changes.

A 13-week teaching route

Week	Preparation	Workshop output
1	potential outcomes and estimands	one-page identification memo
2	experiments and causal diagrams	assignment and control audit
3	OLS, FWL and selection	projection versus causal interpretation
4	uncertainty and clustering	inference plan tied to design
5	IV and LATE	first-stage, exclusion and weak-IV audit
6	panel and fixed effects	within-unit comparison and threat map
7	2×2 and staggered DID	group-time estimand and event-study plan
8	RDD	cutoff graph, bandwidth and falsification plan
9	matching and weighting	overlap and balance report
10	synthetic control	donor-pool and placebo audit
11	discrete, count and selected outcomes	marginal-effect interpretation
12	DML, heterogeneity and sensitivity	cross-fit or robustness-value exercise
13	capstone defence	recommendation, limitations and replication record

Suggested workload is 90 minutes of preparation, two hours of workshop activity and 90 minutes of follow-up per week. Every executable cell has a static worked result; students may complete equivalent calculations in a spreadsheet or on paper. Tables carry the key values, and colour is never the only evidence channel.

Assessment alignment

This is a teaching template, not an institutional grading rule:

Task	Suggested share	Evidence assessed
identification memo	20%	estimand, assignment, assumptions and scope
method replication	25%	data construction, estimator, inference and diagnostics
design comparison	25%	counterfactual quality, sensitivity and communication
integrated capstone	30%	complete evidence chain and policy boundary

Postgraduate work should distinguish identification from estimation formally, justify the asymptotic or randomisation reference, and report a sensitivity or partial-identification analysis.

Current evidence window

The reading map was checked on 1 August 2026. Three updates shape this edition:

Modern DID work separates group-time effects, aggregation and sensitivity rather than interpreting a single staggered two-way-fixed-effects coefficient automatically; see Callaway and Sant’Anna (2021), Borusyak, Jaravel and Spiess (2024) and Rambachan and Roth (2023).
IV practice now includes quantitative sensitivity to exclusion and ignorability violations, not only first-stage strength; see Cinelli and Hazlett (2025).
The AEA Data and Code Availability Policy, revised in February 2026, requires provenance and computational detail sufficient for replication, including when source data are restricted.

The reading and software map distinguishes foundational results, current methods, applied illustrations and live software documentation.

Study discipline

For every result, write six lines:

estimand: which counterfactual contrast;
assignment: where treatment variation came from;
assumption: what makes the comparison causal;
estimator: how the contrast was calculated;
uncertainty: what varies in repeated assignment or sampling;
scope: for whom, where, when and under which treatment version.

Start with causal questions and estimands.

Edit this pageorReport an issue

Causal Diagrams and Controls

Distinguish confounders, mediators, colliders and proxies before adjustment

1. Regression and Inference

Interpret linear projections, control choices and uncertainty without manufacturing causality