Who responds to treatment? (FindIt)

1

Input · what goes in

Experimental data with a treatment (or factorial/conjoint factors), an outcome, and pre-treatment covariates.

Show data format & exampleHide example

re78	treat	age	educ	black	married
9.9	1	37	11	1	1
3.6	0	22	9	1	0
14.2	1	45	16	0	1
8.8	0	29	10	1	0

2

Pipeline · the recipe

↑ Click any step in the diagram to read its logic, code, assumptions & discussion.

1

Data prep

Load the data

Data preparation — shapes the raw inputs into what the estimator expects.

What happens here

Use the Lalonde experiment with the outcome, treatment and covariates whose interactions to search.

Reads from the input data Feeds into #2

Key code

# Install:  install.packages("FindIt")
library(FindIt)
data(LaLonde)
set.seed(1)

Reference / docs ↗

Discussion on this step (0)

No comments on this step yet — be the first.

2

Estimation

Search for heterogeneous effects

The core estimate — where the causal quantity itself is computed.

What happens here

FindIt fits the regularized model with treatment-by-covariate interactions, selecting the penalty by cross-validation.

Formula

\hat\tau(x)=\hat\beta^\top x,\quad \hat\beta=\arg\min_\beta \|Y-X\beta\|^2+\lambda\|\beta\|_1

Reads from #1 Feeds into the final output

Key code

out <- FindIt(model.treat = re78 ~ treat,
              model.main = ~ age + educ + black + hisp + married,
              model.int  = ~ age + educ + black + hisp + married,
              data = LaLonde, type = "continuous", treat.type = "single")
summary(out)

Reference / docs ↗

Discussion on this step (0)

No comments on this step yet — be the first.

3

Output · what you get

Fig 1Distribution of FindIt's estimated individual treatment effects — a spread around the average reveals who responds.

Result figure rendered by StatsOtter from the package's documented example — unofficial community showcase; all credit to the original authors.

Result · the numbers

\hat\tau(x)=\hat\beta^\top x,\quad \hat\beta=\arg\min_\beta \|Y-X\beta\|^2+\lambda\|\beta\|_1

⚠️ Unofficial community showcase of FindIt (docs). Not affiliated with the authors — all credit to Naoki Egami, Marc Ratkovic & Kosuke Imai; this summarizes public documentation.

What it does. FindIt implements the Imai–Ratkovic method for heterogeneous treatment effects: it searches over covariates and their interactions with treatment to discover who is helped or hurt, while controlling false discoveries through regularization.

How it works. It fits a Support Vector Machine / LASSO model with treatment–covariate interactions, choosing the penalty by cross-validation so only well-supported interactions survive. predict() returns each unit's estimated treatment effect; the method also handles factorial and conjoint designs (causal interaction among factors).

Assumptions. Randomized (or unconfounded) treatment; the LASSO penalty trades a little bias for far lower variance and guards against overfitting the interactions.

What you get — A per-unit treatment-effect estimate, the selected interaction terms, and the cross-validated penalty.

Example output

FindIt: Finding Heterogeneous Treatment Effects

Treatment effect estimates by subgroup (ATE range):
  Min. 1st Qu. Median  Mean 3rd Qu.  Max.
 -1842    312    948  1021   1693  3204

Number of selected interaction terms: 4
Optimal lambda (LASSO): 0.083

Links: package · paper