Advanced

Counterfactual Prediction Methods for Causal Inference in Observational Studies with Continuous Treatments

Persson, Joel LU (2019) STAN40 20191
Department of Statistics
Abstract
We develop methods for estimation, inference and optimization of causal effects from observational data with continuous treatments. We present a counterfactual prediction method based on the potential outcomes framework that estimates the expected value of a potential outcome given a treatment level and confounders. We show that the method identifies the average generalized treatment effect (AGTE) and the average dose-response function (ADRF) and propose estimators of these functional causal estimands. Our estimators work under high-dimensional confounding and when the treatment takes many distinct values. Under multiple treatments, the method identifies the effect of a single treatment and the joint effect of multiple treatments.... (More)
We develop methods for estimation, inference and optimization of causal effects from observational data with continuous treatments. We present a counterfactual prediction method based on the potential outcomes framework that estimates the expected value of a potential outcome given a treatment level and confounders. We show that the method identifies the average generalized treatment effect (AGTE) and the average dose-response function (ADRF) and propose estimators of these functional causal estimands. Our estimators work under high-dimensional confounding and when the treatment takes many distinct values. Under multiple treatments, the method identifies the effect of a single treatment and the joint effect of multiple treatments. Treatment effects can further be estimated from unobserved treatment levels. We provide non-parametric and computationally efficient parametric estimation procedures of uncertainty intervals of the ADRF and AGTE and develop algorithms for implementation of the estimators. Finally, we show that the counterfactual prediction method can be used to estimate the treatment level that maximizes the expected individual and population outcome. (Less)
Please use this url to cite or link to this publication:
author
Persson, Joel LU
supervisor
organization
course
STAN40 20191
year
type
H1 - Master's Degree (One Year)
subject
keywords
Causal Inference, Observational Study, Treatment Effects, Continuous Treatments, Dose-Response Function, Optimization
language
English
id
8977715
date added to LUP
2019-06-14 14:10:54
date last changed
2019-06-14 14:10:54
@misc{8977715,
  abstract     = {We develop methods for estimation, inference and optimization of causal effects from observational data with continuous treatments. We present a counterfactual prediction method based on the potential outcomes framework that estimates the expected value of a potential outcome given a treatment level and confounders. We show that the method identifies the average generalized treatment effect (AGTE) and the average dose-response function (ADRF) and propose estimators of these functional causal estimands. Our estimators work under high-dimensional confounding and when the treatment takes many distinct values. Under multiple treatments, the method identifies the effect of a single treatment and the joint effect of multiple treatments. Treatment effects can further be estimated from unobserved treatment levels. We provide non-parametric and computationally efficient parametric estimation procedures of uncertainty intervals of the ADRF and AGTE and develop algorithms for implementation of the estimators. Finally, we show that the counterfactual prediction method can be used to estimate the treatment level that maximizes the expected individual and population outcome.},
  author       = {Persson, Joel},
  keyword      = {Causal Inference,Observational Study,Treatment Effects,Continuous Treatments,Dose-Response Function,Optimization},
  language     = {eng},
  note         = {Student Paper},
  title        = {Counterfactual Prediction Methods for Causal Inference in Observational Studies with Continuous Treatments},
  year         = {2019},
}