Projekte › Kausalinferenz
Turning policy questions into transparent empirical designs: difference-in-differences, regression discontinuity, robustness checks, and clear uncertainty communication.
Research-based case studyRole relevance
Demonstrates careful statistical reasoning for policy questions, including identification strategy, assumptions, robustness checks, uncertainty communication, and limitations of observational data.
Policy questions — whether a cash transfer programme increases political support, whether a lockdown damages wellbeing — cannot be answered by simply comparing outcomes between groups that chose differently. Treated and untreated groups typically differ in many ways that also affect the outcome. Quasi-experimental methods provide principled ways to isolate causal effects from this confounding, by exploiting naturally occurring sources of variation that approximate random assignment.
This project page summarises two complete analyses conducted during my MSc at the London School of Economics, each addressing a different policy question with a different empirical design. The analyses are documented in detail on separate pages; this page explains the methods, assumptions, workflow, and what the work demonstrates as a data science competency.
Simple correlations between a policy and an outcome are typically not causal. The groups that receive treatment differ from those that do not — in income, geography, prior behaviour, or dozens of other characteristics that also influence the outcome. This is the fundamental problem of causal inference: in observational data, we never observe the same unit with and without treatment simultaneously.
Quasi-experimental designs address this by finding situations where treatment assignment is as-good-as-random for a well-defined subset of the data. A sharp eligibility cutoff produces near-random variation near the threshold. A policy change that affects one region but not an otherwise-similar one can reveal causal effects if pre-treatment trends were parallel. The skill lies in recognising these structures in the data, formalising the assumptions they require, and testing those assumptions before interpreting results.
Compares the change in outcomes for a treated group before and after a policy event, relative to a control group that was not exposed. The key assumption is that the two groups would have followed parallel trends in the absence of treatment.
Exploits a sharp eligibility threshold in a running variable. Units just below and just above the cutoff are nearly identical in all respects except treatment assignment — yielding a Local Average Treatment Effect (LATE) at the cutoff.
Results are only credible if they hold up to alternative specifications. Every analysis includes systematic robustness checks designed to challenge the main finding rather than confirm it.
Average treatment effects can mask important heterogeneity. In the WELLBY dissertation, the distributional impact was the central question: does the harm fall on those already worst-off? A Prioritarian Social Welfare Function was used to formally evaluate this.
The figures below illustrate the two designs with synthetic simulated data. They show the structural logic of each approach — not results from any specific analysis.
Fig. 1 — Difference-in-Differences. Pre-treatment parallel trends (left of t₀) validate the identifying assumption; the gap between the treated group and the counterfactual (dashed) after treatment is the DiD estimate. Illustrative simulated data.
Fig. 2 — Regression Discontinuity Design. Observations just left and right of the cutoff are nearly identical; the discontinuity in the fitted line at the threshold identifies the Local Average Treatment Effect (LATE). Illustrative simulated data.
Policy question: Does receiving a government cash transfer causally increase support for the incumbent government? And does this effect persist after the programme ends?
Context: Uruguay's PANES programme (2005–2007) distributed cash and food-card transfers to households below a predicted income score threshold. Manacorda, Miguel & Vigorito (2011) identified the causal effect using the sharp eligibility cutoff as an instrument. This analysis replicates and extends their methodology with a full robustness battery.
Identification: Sharp RDD at the income score cutoff. Households just below and just above the threshold are comparable on all pre-treatment characteristics, except treatment assignment. This was confirmed by a McCrary density test (p = 0.41, no manipulation) and placebo regressions on seven baseline covariates (six of seven non-significant).
Core finding: ~10% causal increase in political support — robust across OLS, donut-hole, bandwidth sensitivity, and local polynomial specificationsPolicy question: What was the causal impact of Victoria's second COVID-19 lockdown (July–October 2020) on life satisfaction, and did the harm fall disproportionately on those already worst-off?
Context: Victoria imposed one of the most stringent lockdowns outside China, while other Australian states remained open. The HILDA panel (215,733 observations, 16,513 individuals, 20 waves) allows comparison of Victorians and non-Victorians before and after the lockdown — the structure a DiD design requires.
Identification: Difference-in-Differences with Propensity Score Matching pre-processing (to improve covariate balance between treated Victorians and control states). The parallel-trends assumption was tested using pre-treatment waves. Results were fed into an Atkinson Social Welfare Function (γ = 1) to formally evaluate the distributional consequences.
Core finding: −0.030 points life satisfaction (95% CI −0.049 to −0.012) — harm concentrated among worst-off (LS < 4: additional −0.334 points)The typical pipeline for a quasi-experimental analysis, from raw data to written findings:
Harmonise raw survey data; construct treatment indicator, running variable, outcome variable, and control variables. Validate sample composition against the research question.
Plot the distribution of the running variable (RDD) or the raw group trends over time (DiD). Visual inspection often reveals whether the design is viable before any regression is run.
Formally test the identifying assumptions: McCrary density test and covariate placebo regressions (RDD); parallel-trends pre-test and covariate balance check (DiD/PSM). Document what is found, not just what supports the design.
Estimate the preferred specification: OLS with robust standard errors (RDD), or two-way fixed-effects regression (DiD). Report the point estimate, standard error, and confidence interval alongside sample size.
Systematically vary the specification: bandwidth, functional form, exclusion zones (donut), control variables. A credible result holds across all meaningful alternatives. Any instability is reported transparently.
Produce RD plots, event-study plots, bandwidth sensitivity charts, and coefficient plots with confidence intervals. Figures are generated programmatically (Stata, R, or Python) — no manual editing. Write findings for both technical and non-technical readers.
Every quasi-experimental design rests on identifying assumptions that are testable in part but not fully provable. Honest analysis requires stating them explicitly:
RDD assumptions
DiD assumptions
Reproducibility note
The two illustrative design figures on this page are fully reproducible from projects/causal-policy-analysis/scripts/generate_illustrative_figures.py using synthetic data. The underlying Stata analysis scripts exist for both completed analyses; public replication files will be added where the licensing of the source datasets permits. Where data cannot be shared, this page documents the analytical design, assumptions, workflow, and outputs in sufficient detail to evaluate the methodology.