Regularization and confounding in linear regression for treatment effect estimation

Paul Hahn, Carlos M. Carvalho, David Puelz, Jingyu He

Research output: Contribution to journalArticle

4 Citations (Scopus)

Abstract

This paper investigates the use of regularization priors in the context of treatment effect estimation using observational data where the number of control variables is large relative to the number of observations. First, the phenomenon of "regularization-induced confounding" is introduced, which refers to the tendency of regularization priors to adversely bias treatment effect estimates by over-shrinking control variable regression coefficients. Then, a simultaneous regression model is presented which permits regularization priors to be specified in a way that avoids this unintentional "re-confounding". The new model is illustrated on synthetic and empirical data.

Original languageEnglish (US)
Pages (from-to)163-182
Number of pages20
JournalBayesian Analysis
Volume13
Issue number1
DOIs
StatePublished - Jan 1 2018
Externally publishedYes

Fingerprint

Confounding
Treatment Effects
Linear regression
Regularization
Shrinking
Regression Coefficient
Variable Coefficients
Regression Model
Estimate
Model

Keywords

  • Causal inference
  • Observational data
  • Shrinkage estimation

ASJC Scopus subject areas

  • Statistics and Probability
  • Applied Mathematics

Cite this

Regularization and confounding in linear regression for treatment effect estimation. / Hahn, Paul; Carvalho, Carlos M.; Puelz, David; He, Jingyu.

In: Bayesian Analysis, Vol. 13, No. 1, 01.01.2018, p. 163-182.

Research output: Contribution to journalArticle

Hahn, Paul ; Carvalho, Carlos M. ; Puelz, David ; He, Jingyu. / Regularization and confounding in linear regression for treatment effect estimation. In: Bayesian Analysis. 2018 ; Vol. 13, No. 1. pp. 163-182.
@article{00a0f01010774e449c83bbb4fcaef63c,
title = "Regularization and confounding in linear regression for treatment effect estimation",
abstract = "This paper investigates the use of regularization priors in the context of treatment effect estimation using observational data where the number of control variables is large relative to the number of observations. First, the phenomenon of {"}regularization-induced confounding{"} is introduced, which refers to the tendency of regularization priors to adversely bias treatment effect estimates by over-shrinking control variable regression coefficients. Then, a simultaneous regression model is presented which permits regularization priors to be specified in a way that avoids this unintentional {"}re-confounding{"}. The new model is illustrated on synthetic and empirical data.",
keywords = "Causal inference, Observational data, Shrinkage estimation",
author = "Paul Hahn and Carvalho, {Carlos M.} and David Puelz and Jingyu He",
year = "2018",
month = "1",
day = "1",
doi = "10.1214/16-BA1044",
language = "English (US)",
volume = "13",
pages = "163--182",
journal = "Bayesian Analysis",
issn = "1936-0975",
publisher = "Carnegie Mellon University",
number = "1",

}

TY - JOUR

T1 - Regularization and confounding in linear regression for treatment effect estimation

AU - Hahn, Paul

AU - Carvalho, Carlos M.

AU - Puelz, David

AU - He, Jingyu

PY - 2018/1/1

Y1 - 2018/1/1

N2 - This paper investigates the use of regularization priors in the context of treatment effect estimation using observational data where the number of control variables is large relative to the number of observations. First, the phenomenon of "regularization-induced confounding" is introduced, which refers to the tendency of regularization priors to adversely bias treatment effect estimates by over-shrinking control variable regression coefficients. Then, a simultaneous regression model is presented which permits regularization priors to be specified in a way that avoids this unintentional "re-confounding". The new model is illustrated on synthetic and empirical data.

AB - This paper investigates the use of regularization priors in the context of treatment effect estimation using observational data where the number of control variables is large relative to the number of observations. First, the phenomenon of "regularization-induced confounding" is introduced, which refers to the tendency of regularization priors to adversely bias treatment effect estimates by over-shrinking control variable regression coefficients. Then, a simultaneous regression model is presented which permits regularization priors to be specified in a way that avoids this unintentional "re-confounding". The new model is illustrated on synthetic and empirical data.

KW - Causal inference

KW - Observational data

KW - Shrinkage estimation

UR - http://www.scopus.com/inward/record.url?scp=85039841453&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85039841453&partnerID=8YFLogxK

U2 - 10.1214/16-BA1044

DO - 10.1214/16-BA1044

M3 - Article

AN - SCOPUS:85039841453

VL - 13

SP - 163

EP - 182

JO - Bayesian Analysis

JF - Bayesian Analysis

SN - 1936-0975

IS - 1

ER -