Atlas/Heterogeneous Treatment Effects

Heterogeneous Treatment Effects: Elections → Shutdowns

For each country, the expected lift in 7-day shutdown probability when an election occurs in the next 7 days. Causal Forest DML (Athey & Wager 2019, Chernozhukov et al. 2018). 90% honest forest confidence intervals from EconML.

Avg ATE (train)

9.6pp

on training pop.

Avg ATE (holdout)

9.6pp

temporal hold-out

CI Coverage

99.5%

vs bootstrap ATE

Promote

PROMOTED

floor: coverage ≥ 75%, nonzero-CI ≥ 50%

Honest caveats

Only 5 countries have ≥ 3 elections in the training window. Other estimates are extrapolations via the country one-hot plus risk-tier and unrest features — wide CI, flagged noisy=true.
Identification assumes unconfoundedness conditional on W = (risk_tier, month, day-of-week, prior shutdowns, lagged block rate, weekend). Unobserved confounders (e.g. coups not in our event calendar) still bias the estimate.
ATE is the average over the country's day distribution. A specific election may differ by stakes, incumbent risk, opposition coordination — not captured here.

Per-country ATE (sorted by magnitude)

Bars scale by ATE; brackets show the 90% honest forest CI. Negative ATE = elections reduce shutdown probability.

Per-country detail (click country for raw JSON)

Country	ATE	90% CI	Train events	Status
IRIran	44.8pp	[-54.0pp, 143.6pp]	0	noisy / extrapolated
VNVietnam	32.2pp	[14.1pp, 50.2pp]	8	signal
KPNorth Korea	17.7pp	[-2.8pp, 38.2pp]	0	noisy / extrapolated
CNChina	17.7pp	[-2.8pp, 38.2pp]	0	noisy / extrapolated
BYBelarus	16.7pp	[9.2pp, 24.3pp]	8	signal
SASaudi Arabia	16.7pp	[-3.0pp, 36.3pp]	0	noisy / extrapolated
CUCuba	15.8pp	[3.1pp, 28.6pp]	0	noisy / extrapolated
SYSyria	15.2pp	[2.7pp, 27.7pp]	0	noisy / extrapolated
TMTurkmenistan	13.4pp	[1.7pp, 25.1pp]	0	noisy / extrapolated
PKPakistan	7.1pp	[-7.7pp, 21.8pp]	0	noisy / extrapolated
RURussia	3.3pp	[-22.3pp, 29.0pp]	0	noisy / extrapolated
TRTurkey	3.1pp	[-13.9pp, 20.1pp]	0	noisy / extrapolated
MYMalaysia	2.6pp	[-4.2pp, 9.4pp]	0	noisy / extrapolated
MMMyanmar	1.9pp	[-6.9pp, 10.7pp]	0	noisy / extrapolated
BDBangladesh	1.5pp	[-10.7pp, 13.7pp]	0	noisy / extrapolated
THThailand	0.6pp	[-3.5pp, 4.7pp]	0	noisy / extrapolated
KZKazakhstan	0.5pp	[-0.3pp, 1.2pp]	8	signal
IDIndonesia	0.4pp	[-1.8pp, 2.6pp]	0	noisy / extrapolated
EGEgypt	0.3pp	[-0.2pp, 0.8pp]	8	signal
UZUzbekistan	0.2pp	[-3.3pp, 3.7pp]	0	noisy / extrapolated
VEVenezuela	-9.6pp	[-15.2pp, -3.9pp]	8	signal

Methodology

We use Double Machine Learning (DML) with a Causal Forest second stage (EconML's CausalForestDML). The DML framework cross-fits two nuisance models — an outcome regression E[Y | W] and a propensity model P(T=1 | W) — and then a causal forest learns the conditional average treatment effect tau(X) = E[Y(1) - Y(0) | X] from residualised data. This makes the estimate doubly robust: it is consistent if either the outcome model or the propensity model is correctly specified.

Treatment

T = 1 if the country has an election in the next 7 days (election_in_7days flag from the forecast feature panel). 40 treated rows in training, across 7 distinct countries (KZ, VN, BD, EG, BY, MM, VE).

Outcome

Y = target_7day — binary indicator: did at least one labelled shutdown event occur for this country in the 7-day window starting at this country-day? Same label as the production XGBoost forecast.

Covariates W

risk_tier (1–5 baseline censorship intensity)
month, day_of_week, is_weekend (seasonality)
prior_shutdowns_30d (rolling 30-day target sum, shifted)
block_rate_lag7 (7-day-old block-rate baseline)

Effect modifiers X

Country one-hot (21 dummies)
risk_tier
gdelt_unrest_30d

Identification assumptions

SUTVA: country i's outcome does not depend on whether country j has an election. Plausible at 7-day horizon.
Unconfoundedness given W: after controlling for risk tier, seasonality, and lagged severity, election timing is as good as random. Violated if governments call snap elections strategically when censorship infrastructure is ready.
Positivity: every covariate stratum has a non-zero chance of treatment. Approximated by the DML cross-fit; some country/risk-tier cells are sparse.

Evaluation

We hold out the most recent 120 days of country-days. The held-out ATE is computed two ways:

Model ATE: est.ate_inference(X_holdout).conf_int_mean(alpha=0.10).
Empirical ATE: bootstrap-resample the held-out treated and control outcome means 200 times.

CI coverage = fraction of bootstrap empirical ATEs that fall inside the model's 90% CI. Election treatment: 99.5%.

Promote criteria

CI coverage ≥ 75% on held-out
≥ 50% of countries with ≥ 3 events have a CI that excludes zero (signal-to-noise floor)

Election treatment passes both. Protest treatment fails (empirical ATE near zero, CI excludes empirical mean) and is not promoted.

/sentinel/attribute — single-event synthetic DiD attribution
/atlas/forecast — calibrated 7-day risk predictions
/atlas/elections — upcoming election calendar with risk overlay

References

Athey, S. & Wager, S. (2019). "Estimation and Inference of Heterogeneous Treatment Effects using Random Forests." JASA.
Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., & Robins, J. (2018). "Double/debiased machine learning for treatment and structural parameters." Econometrics Journal.
Microsoft EconML library: econml.azurewebsites.net

API: GET /v1/sentinel/hte?country={cc}&treatment=election · leaderboard /v1/sentinel/hte/leaderboard · info /v1/sentinel/hte/info

Data refreshes every 30 minutes. Model retrained on demand from forecast_training_data.parquet.