LATE implementation details

[mili-yini]

trafficstars

Does the implementation of LATE（instrumental variables (IV) regression） in grf have a detailed algorithm explanation, or are there any documents that can be referred to

[erikcs]

If you want an abstract algorithm description, instrumental_forest is Algorithm 1 in https://arxiv.org/abs/1610.01271 with the estimating equation $\psi$ given by the moment functions described in section 7.1 of the same paper. The grf package also takes sample weights into account.

If you want something less abstract, then the description of causal_forest given here can be applied to IV forests too: IV forest first centers the outcome Y, treatment W, and instrument Z (ala Robinson for causal forest), then uses grf to find forest weights for a treatment effect $\tau(x)$ calculated as the forest-weighted two-stage least squares regression of centered Y on centered W and Z (i.e, where causal forest runs residualized regressions, instrumental forest runs two-stage residualized regressions).

[mili-yini]

If you want an abstract algorithm description, instrumental_forest is Algorithm 1 in https://arxiv.org/abs/1610.01271 with the estimating equation ψ given by the moment functions described in section 7.1 of the same paper. The grf package also takes sample weights into account.

If you want something less abstract, then the description of causal_forest given here can be applied to IV forests too: IV forest first centers the outcome Y, treatment W, and instrument Z (ala Robinson for causal forest), then uses grf to find forest weights for a treatment effect τ(x) calculated as the forest-weighted two-stage least squares regression of centered Y on centered W and Z (i.e, where causal forest runs residualized regressions, instrumental forest runs two-stage residualized regressions).

Thank you very much,and how to understand in instrumental variable regression,