Your first analysis — 10 minutes¶

We’ll load one of the 41 built-in datasets, estimate an average treatment effect via inverse-probability weighting, and get bootstrap confidence intervals. This runs offline, with no LLM, no cloud.

Prereqs¶

ESML installed (esml --version works)
2 minutes

1. Launch the REPL¶

esml repl

Or directly in Python:

import esml
from esml.data import load_dataset
from esml.fn import ate

2. Load data¶

df = load_dataset("ccs_2022_2023")   # Canadian Cannabis Survey
df.shape
# (18000, ...)

3. Estimate an ATE¶

result = ate(
    df,
    treatment="cannabis_past_year_use",
    outcome="mental_health_good",
    covariates=["age_group", "province", "sex", "education"],
    method="ipw",
    bootstrap=True,
    n_boot=500,
    seed=42,
)
print(result.summary())

Output:

Average Treatment Effect (IPW)
──────────────────────────────────────────
   estimate:    -0.0432
   SE:           0.0187
   95% CI:      (-0.0799, -0.0065)
   n:           17842
   n effective: 12451
   method:      Inverse probability weighting, bootstrap CI (B=500)

4. Sanity-check with a different estimator¶

from esml.fn import aipw

result_dr = aipw(
    df,
    treatment="cannabis_past_year_use",
    outcome="mental_health_good",
    covariates=["age_group", "province", "sex", "education"],
)
print(result_dr.summary())

AIPW is doubly-robust — consistent if either the propensity model or the outcome model is correctly specified. If the two estimates disagree meaningfully, that’s a signal something’s off in one of the models.

5. Visualize¶

result.plot_forest()        # forest plot of treatment effect
result.plot_weight_diag()   # weight diagnostics for IPW

Or, from the terminal:

esml run-module propensity-scores \
    --dataset ccs_2022_2023 \
    --treatment cannabis_past_year_use \
    --outcome mental_health_good

Next steps¶

Dataset-agnostic: every ESML function takes column names as keyword arguments. Swap in your own CSV via load_dataset("path/to/mine.csv").
Causal inference deep dive: see the esml API docs.
Ask Perseus: esml ? "what's the difference between IPW and AIPW?" (requires Ollama installed, or falls back to free-tier FreeAPI).