Instrumental Variable Forest
An instrumental variable forest is a generalized random forest method that uses an instrument to estimate how a causal treatment effect varies with individual characteristics when the treatment is endogenous.
도구변수 포레스트는 처치가 내생적일 때 도구변수를 이용해 인과적 처치효과가 개인 특성에 따라 어떻게 달라지는지 추정하는 일반화 랜덤포레스트 방법이다.
How do I estimate heterogeneous treatment effects when the treatment is confounded but I have an instrument?
처치가 교란되었지만 도구변수가 있을 때 이질적 처치효과를 어떻게 추정하나요?
An instrumental variable forest grows honest causal trees that split on covariates to maximize heterogeneity in instrument-based treatment effect estimates, yielding conditional local average treatment effects. It fits large observational datasets with a valid instrument where you suspect the effect differs across subgroups.
When to use Instrumental Variable Forest
- An endogenous treatment and a valid instrument in a large sample
- Exploring how the causal effect varies across covariates
- Identifying subgroups with larger or smaller treatment effects
When to use something else
- A single average effect is enough: use two-stage least squares
- Treatment is unconfounded given observed covariates: use a causal forest without an instrument
Assumptions and requirements
- Standard instrumental variable assumptions (relevance, exclusion, independence, monotonicity)
- Large sample size for stable splits
- Honest sample splitting for valid confidence intervals
- Instrument strength sufficient within subgroups
What the analysis reports
- Conditional local average treatment effect for each observation
- Average effect with confidence interval
- Variable importance for effect heterogeneity
- Effects by subgroup or covariate
How to report it
The instrumental variable forest estimated an average effect of [value] (95% CI [[lower], [upper]]), with effects varying most strongly by [covariate].
Related methods
Run Instrumental Variable Forest on Quorant
- Data you need
- a table (CSV or Excel) with an outcome variable, a treatment indicator and instrument
- Price
- 400 credits for one analysis up to 10,000 rows (complexity level 3); larger or wider data costs more. Optional methodologist review: 1800 credits.
- Checking
- Python results are cross-checked against R reference outputs in every deployment
- Before you pay
- A free pre-analysis diagnosis checks your data and gives an exact quote; nothing is deducted until you approve it.