TGM AL MAMUN
Empirical Research Guide · Causal Inference

Which causal design fits your data?

A causal estimate is only as good as its source of variation. Test a real dataset, upload your own, or enter what you know about your design, and follow the path to the method your paper needs, with the diagnostics a referee will ask for.

7 design guides7 real example datasetsPython · statsmodels

Choose a dataset

Seven real datasets from well-known studies. Each one leads to a different design.

Design guides

Routes to a finished paper

Each guide takes one real dataset through estimation, diagnostic tests and writing, with Python code, real output and sample text for the paper.

How the tool decides. The design follows from the source of variation, which you declare through the column roles; the data then decide among the designs in that family. With a running variable and a cutoff, the tool estimates a local linear regression discontinuity with the Imbens–Kalyanaraman bandwidth and checks manipulation (McCrary density test), covariate continuity, placebo cutoffs and bandwidth sensitivity; if the treatment is not fully determined by the cutoff, the design is fuzzy. With an instrument, it compares OLS and 2SLS and reports the first-stage F statistic (below 10 signals a weak instrument), the weak-instrument-robust Anderson–Rubin test and an endogeneity test. With a treatment that starts at a point in time, the number of treated units decides: with 3 or fewer, a synthetic control if at least 5 untreated units are available and an interrupted time series otherwise; with more, an event study tests for parallel pre-trends (leads −5 to −2, cluster bootstrap), sending designs that fail to generalized synthetic control, and the others to DiD (one adoption date) or a heterogeneity-robust staggered DiD (several dates). These rules organise the evidence; they do not establish identification, which rests on the argument for the source of variation.