This page gives a week-by-week overview of RMDA II (Spring 2026). Each entry includes what students are learning, the core conceptual thread for that stretch of the course, and the key skill they should be able to demonstrate by the end of the week.
Week 1 — Jan 12 & 14 | Introduction + Measurement
We open the semester by asking a deceptively simple question: how do we know if a policy worked? The first lecture sets up the fundamental problem of causal inference — the gap between what we observe and what we want to know. We motivate the course by reading policy claims critically and asking what evidence would actually be convincing. The second lecture turns to measurement: before we can estimate anything, we need to be precise about what we are measuring, how, and whether the operational definition maps onto the concept we care about.
Skill: Critically read an empirical claim in a policy brief or news article — identify what is being measured, what the implied counterfactual is, and what assumptions would be needed to make the causal claim credible.
Week 2 — Jan 21 | Measurement & Magnitude
(MLK Day Jan 19 — no class)
With measurement foundations in place, we move to the question of size: even if a finding is real, does it matter? Students learn to translate raw regression coefficients into policy-relevant quantities — percent changes, standard deviation shifts, cost-per-unit comparisons, and benchmark comparisons. The goal is to move beyond statistical significance and toward economic or practical significance. A coefficient that is precisely estimated but tiny may be irrelevant to policymakers; a noisy estimate with a large effect size may be worth acting on.
Skill: Express an estimated effect in at least two alternative units (e.g., raw units → SD units → cost-effectiveness ratio) and make an argument about whether the magnitude is policy-relevant.
Week 3 — Jan 28 | Mean Comparison & Potential Outcomes
(Jan 26 class cancelled)
This week introduces the formal language of causal inference. We define potential outcomes — the outcomes a unit would experience under treatment and under control — and use this framework to see exactly why naive comparisons between treated and untreated units are often misleading. Selection bias emerges naturally: the people who receive a treatment are systematically different from those who do not, in ways that also affect the outcome. The lecture connects this framework to the simplest causal estimator: the difference in means between two groups, and when that difference is (and is not) interpretable as a causal effect.
Skill: Write the observed difference in means using conditional expectation notation, decompose it into ATT + bias, and explain concretely what the bias term represents in a given policy context.
Week 4 — Feb 2 & 4 | Mean Comparison with Covariates & OLS
We now add covariates to the mean comparison. The key insight is that OLS with a binary treatment variable is just a formalization of comparing means while holding other things constant. Students learn to interpret regression output: what the slope coefficient means, how to read standard errors and t-statistics, and how to predict outcomes for specific types of individuals. We introduce the omitted variable bias (OVB) formula and use it to reason about what happens to a coefficient when an important variable is left out of the regression — and which direction the bias runs.
Skill: Given two nested regressions (short and long), use the OVB formula to sign and explain the bias in the short regression — both algebraically and with an intuitive story about the omitted variable.
Week 5 — Feb 9 & 11 | OLS Mechanics + Regression Mechanics 1
This week opens the black box of OLS. Where does β̂ come from? We derive the OLS estimator geometrically and algebraically, and build intuition for what it means to "project" an outcome onto a set of regressors. The second lecture begins the regression mechanics unit: non-linear specifications. Specifically, we cover quadratic and higher-order polynomial terms, which allow the marginal effect of X to vary with the level of X. Students work through the derivative-based formula for marginal effects and calculate turning points.
Skill: Given a regression with a quadratic term, compute the marginal effect of X at a specified value, find the turning point algebraically, and interpret both in a substantive sentence.
Week 6 — Feb 16 & 18 | Regression Mechanics 2 & 3 (Interactions + Logs)
We extend the regression mechanics toolkit to interaction terms and logarithmic transformations — two of the most commonly used and most commonly misinterpreted tools in applied work. Interaction terms allow the effect of one variable to depend on another (continuous × binary, binary × binary, continuous × continuous). Log transformations convert the scale of coefficients from dollars to percent changes, and allow us to model elasticities. By the end of the week, students can read any standard applied regression table and correctly interpret every specification.
Skill: Interpret a coefficient in each of the four log specifications (level-level, log-level, level-log, log-log), and correctly compute the marginal effect of X for each group in a regression with an interaction term.
Week 7 — Feb 23 & Feb 25 | Review + Midterm 1
The first half of the course concludes with a review session and the first midterm exam. The review session synthesizes the regression toolkit built over Weeks 1–6: measurement, magnitude, potential outcomes, OLS, OVB, quadratics, interactions, and logs. Students should be able to move fluidly between concepts — connecting OVB to selection bias, connecting log specifications to economic interpretation, and using the OVB formula to reason about what controls to include and why.
Skill: (Synthesis) Given a policy evaluation scenario, choose an appropriate regression specification, anticipate the direction of OVB from omitted variables, and interpret all coefficients correctly — including non-linear and interacted terms.
☀️ Week 8 — Spring Break (Mar 2 & 4)
No class. Rest up — the second half of the course introduces the quasi-experimental toolkit.
Week 9 — Mar 9 & 11 | Instrumental Variables (Part 1)
The second half of the course opens with the most powerful — and most demanding — tool in the causal inference toolkit: instrumental variables (IV). OLS breaks down when the treatment variable is endogenous (correlated with the error term). IV solves this by finding an external source of variation — an instrument — that shifts treatment assignment without directly affecting the outcome. We motivate IV with the college proximity example, derive the Wald estimator, and show how IV is implemented in Stata via two-stage least squares (2SLS). The first-stage F-statistic is introduced as the key diagnostic for instrument strength.
Skill: Given a proposed instrument, evaluate its plausibility against the two core conditions (relevance and exclusion restriction), run 2SLS in Stata, and interpret the first-stage F-statistic.
Week 10 — Mar 16 & 18 | Instrumental Variables (Part 2)
We deepen the IV framework. The central result of this week is the Local Average Treatment Effect (LATE): IV does not recover the ATE or the ATT — it recovers the causal effect for compliers, the subgroup whose treatment status is actually shifted by the instrument. We cover the four compliance types (compliers, always-takers, never-takers, defiers) and the monotonicity assumption that makes LATE interpretable. We also work through the 16 and Pregnant application and discuss what it means that an IV estimate is internally valid but potentially limited in external validity.
Skill: Given an IV setup, describe who the compliers are in plain language, explain why the IV estimate represents a LATE rather than an ATE, and assess whether the complier subpopulation is policy-relevant.
Week 11 — Mar 23 & 25 | Regression Discontinuity
Regression discontinuity (RD) exploits the fact that many policies assign treatment based on a threshold — a test score cutoff, an income limit, an age rule. Units just below the threshold are nearly identical to units just above it, creating a local randomization. We cover the logic and vocabulary of RD (running variable, cutoff, bandwidth, local linear regression), distinguish sharp from fuzzy RD, and introduce the graphical tools used to present and validate RD estimates. Assumptions are emphasized: continuity of potential outcomes, no manipulation of the running variable, and no other discontinuity at the threshold.
Skill: Given an RD setting, produce and interpret the RD plot, run the local linear regression in Stata, and conduct the key validity tests (density test, covariate balance at the cutoff, placebo cutoffs).
Week 12 — Mar 30 & Apr 1 | RD (continued) + Panel Data & Fixed Effects (intro)
We close out regression discontinuity and pivot to panel data and fixed effects. Panel data — observing the same units (people, firms, states) across multiple time periods — opens a powerful identification strategy: comparing each unit to itself over time. The key idea is that if the main confounders are time-invariant (ability, geography, culture), they can be eliminated entirely by absorbing unit fixed effects. We build intuition through two examples: the effect of sex education on teen pregnancy, and police on crime. We show how pooled OLS can give the wrong sign — and how fixed effects corrects it.
Skill: Explain in one sentence why unit fixed effects eliminate time-invariant confounders, and identify which variables in a given dataset will be dropped or attenuated when fixed effects are included.
Week 13 — Apr 6 | Panel Data & Fixed Effects (continued)
(Apr 8 — 48-hour project, no class)
We formalize the fixed effects estimator and its algebraic relationship to the within-estimator (de-meaned OLS). We extend to two-way fixed effects — controlling for both unit fixed effects and time fixed effects simultaneously — which eliminates both time-invariant unit characteristics and common time shocks. Students implement FE in Stata using xtreg, fe and learn to interpret the variation that identifies each coefficient. The critical diagnostic question: does your key variable of interest actually vary within units over time? If not, fixed effects cannot identify its effect.
Skill: Runxtreg, fein Stata, usextsumto assess the within-unit variation of key variables, and explain which covariates will survive fixed effects estimation and why. Conceptual questions about what is a FE? What does it capture? What does it not capture?
Week 14 — Apr 13 & 15 | Difference-in-Differences (Part 1)
Difference-in-differences (DiD) combines the logic of panel data with the logic of a natural experiment. The setup requires a treatment group affected by a policy, a control group not affected, and observations before and after the policy change. The core estimator is the "difference of differences": how much did the treatment group change, relative to how much the control group changed? The parallel trends assumption — that both groups would have trended the same way in the absence of treatment — is the crucial identifying assumption. We introduce it, visualize it, and think carefully about when it is and is not plausible.
Skill: Construct the 2×2 DiD table (group × time) by hand from summary statistics, compute the DiD estimate, write the equivalent regression, and explain what parallel trends requires in the context of a specific policy application.
Week 15 — Apr 20 & 22 | Difference-in-Differences (continued)
We extend DiD to more realistic and more demanding settings. Event study plots are introduced as the standard visual tool for assessing parallel pre-trends and estimating dynamic treatment effects. We discuss staggered DiD — the case where different units adopt the policy at different times — and the complications this creates for the standard 2×2 estimator. Modern DiD methods (Callaway-Sant'Anna, Bacon decomposition) are introduced conceptually. Throughout, we emphasize that DiD is one of the most widely used empirical strategies in applied policy work precisely because clean natural experiments abound when policies roll out unevenly.
Skill: Produce and interpret an event study plot in Stata, explain what a violation of parallel pre-trends would look like in the plot, and describe in plain language why staggered timing complicates the standard DiD estimator.
Week 16 — Apr 27 | Review + Final Exam Period
The final week brings together the full causal inference toolkit built across the semester: mean comparison and OLS, instrumental variables, regression discontinuity, panel fixed effects, and difference-in-differences. The review session focuses on the logic connecting these methods — when each is appropriate, what identifying assumption it relies on, and what the estimated quantity (ATE, ATT, LATE, local estimate at a threshold) actually means. The course ends where it began: can we make a credible causal claim about a policy?
Skill: (Synthesis) Given a policy evaluation scenario with a described data structure and variation, select the most appropriate causal inference method, state the key identifying assumption, and describe what the estimate would and would not tell a policymaker.