Relative Bias Under Imperfect Identification in Observational Causal Inference
Extracted source text · arXiv 2507.23743
% Options for packages loaded elsewhere
\PassOptionsToPackage{unicode}{hyperref}
\PassOptionsToPackage{hyphens}{url}
\documentclass[11pt, ]{article}
\usepackage{mathtools}
\usepackage{amsmath}
\usepackage{amsthm}
\usepackage{amssymb}
\usepackage[italicdiff]{physics}
% SPACING AND FONTS %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\usepackage{iftex}
% CAREFUL: the order of font includes here is very important!
\ifPDFTeX
\usepackage[OT1,T1]{fontenc}
\usepackage[utf8]{inputenc}
\usepackage{textcomp} % provide euro and other symbols
\usepackage[p,osf,swashQ]{cochineal}
\usepackage[cochineal,vvarbb]{newtxmath}
\usepackage[scale=0.95]{biolinum}
\usepackage[scale=0.95,varl]{inconsolata}
\else % if luatex or xetex
\usepackage[scale=0.95,varl]{inconsolata}
\usepackage{newpxtext}
\usepackage{mathpazo}
\usepackage[scale=0.95]{biolinum}
\fi
\ifLuaTeX
\usepackage{selnolig} % disable illegal ligatures
\fi
\IfFileExists{microtype.sty}{% use microtype if available
\usepackage[]{microtype}
\UseMicrotypeSet[protrusion]{basicmath} % disable protrusion for tt fonts
}{}
\setlength{\emergencystretch}{3em} % prevent overfull lines
%\widowpenalty=10000
%\clubpenalty=10000
\flushbottom
\allowdisplaybreaks
\sloppy
% CORE PACKAGES %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\usepackage[dvipsnames,svgnames,x11names]{xcolor}
\usepackage[lmargin=1in,rmargin=1in,tmargin=1in,bmargin=1in]{geometry}
\usepackage[format=plain,
labelfont={bf,sf,small,singlespacing},
textfont={sf,small,singlespacing},
justification=justified,
margin=0.25in]{caption}
% SECTIONS AND HEADINGS %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\setcounter{secnumdepth}{4}
\usepackage{sectsty}
\usepackage[compact]{titlesec}
% short title
\makeatletter
\newcommand\@shorttitle{}
\newcommand\shorttitle[1]{\renewcommand\@shorttitle{#1}}
\usepackage{fancyhdr}
\fancyhf{}
\pagestyle{fancy}
\renewcommand{\headrulewidth}{0pt}
\fancyheadoffset{0pt}
\lhead{\scshape \@shorttitle}
\rhead{\scshape July 31, 2025}
\cfoot{\thepage}
\makeatother
% abstract styling
\renewenvironment{abstract}{
\centerline
{\large\sffamily\bfseries Abstract}\vspace{-1em}
\begin{quote}\small
}{
\end{quote}
}
% PANDOC INCLUDES %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\providecommand{\tightlist}{%
\setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}}\usepackage{longtable,booktabs,array}
\usepackage{calc} % for calculating minipage widths
% Correct order of tables after \paragraph or \subparagraph
\usepackage{etoolbox}
\makeatletter
\patchcmd\longtable{\par}{\if@noskipsec\mbox{}\fi\par}{}{}
\makeatother
% Allow footnotes in longtable head/foot
\IfFileExists{footnotehyper.sty}{\usepackage{footnotehyper}}{\usepackage{footnote}}
\makesavenoteenv{longtable}
\usepackage{graphicx}
\makeatletter
\newsavebox\pandoc@box
\newcommand*\pandocbounded[1]{% scales image to fit in text height/width
\sbox\pandoc@box{#1}%
\Gscale@div\@tempa{\textheight}{\dimexpr\ht\pandoc@box+\dp\pandoc@box\relax}%
\Gscale@div\@tempb{\linewidth}{\wd\pandoc@box}%
\ifdim\@tempb\p@<\@tempa\p@\let\@tempa\@tempb\fi% select the smaller of both
\ifdim\@tempa\p@<\p@\scalebox{\@tempa}{\usebox\pandoc@box}%
\else\usebox{\pandoc@box}%
\fi%
}
% Set default figure placement to htbp
\def\fps@figure{htbp}
\makeatother
% END PANDOC %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% USER INCLUDES %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\usepackage{mathtools}
\usepackage{amsmath}
\usepackage{amsthm}
\usepackage{amssymb}
\usepackage[italicdiff]{physics}
\usepackage[p,osf,swashQ]{cochineal}
\usepackage[cochineal,vvarbb]{newtxmath}
\usepackage{float}
\usepackage{placeins}
\usepackage{setspace}
\newcommand{\papersp}{1.0} % 1.75 = 28lpp w/ a4, 12pt
% \usepackage{lineno}
% \linenumbers
\usepackage{tikz}
\usetikzlibrary{arrows.meta}
\usepackage[dvipsnames]{xcolor}
\definecolor{sooblue}{HTML}{10577F}
\definecolor{proxred}{HTML}{9A5510}
\definecolor{ivpink}{HTML}{C040BA}
\definecolor{gray}{HTML}{777777}
\newcommand{\soo}{\mathrm{soo}}
\newcommand{\prox}{\mathrm{prox}}
\newcommand{\iv}{\mathrm{iv}}
\newcommand{\bias}{\mathrm{Bias}}
\newcommand{\mse}{{\normalfont\textsc{mse}}}
\newcommand{\GRV}{\mathrm{TRV}}
\newcommand{\TRV}{\mathrm{TRV}}
\newcommand{\hlsoo}[1]{\textcolor{sooblue}{#1}}
\newcommand{\hlprox}[1]{\textcolor{proxred}{#1}}
\newcommand{\hliv}[1]{\textcolor{ivpink}{#1}}
\newcommand{\ct}[1]{\textcolor{NavyBlue}{#1}}
\newcommand{\rt}[1]{\textcolor{red}{#1}}
\newcommand{\biasSOO}{\mathrm{Bias}(\tau_\soo; \hlsoo{R_{Y \sim U \mid Z, W_Y, W_Z}}, \hlsoo{R_{Z \sim U \mid W_Y, W_Z})}}
\renewcommand*{\arraystretch}{1.5}
\newcommand{\sooRsq}{\{\hlsoo{R_{Y \sim U \mid Z, W_Y, W_Z}}, \hlsoo{R_{Z \sim U \mid W_Y, W_Z}}\}}
%\renewcommand{\hat}{\widehat}
% theorem environments
% \theoremstyle{plain}
% \newtheorem{prop}{Proposition}[section]
% \newtheorem{claim}[prop]{Claim}
% \newtheorem{corollary}{Corollary}[prop]
% \newtheorem{theorem}[prop]{Theorem}
% \newtheorem{lemma}[prop]{Lemma}
% add \numberthis to the end of a line in an `align*` environment to get a number
\newcommand\numberthis{\addtocounter{equation}{1}\tag{\theequation}}
% Blackboard bold sets and the empty set
\newcommand{\C}{\ensuremath{\mathbb{C}}}
\newcommand{\R}{\ensuremath{\mathbb{R}}}
\newcommand{\Q}{\ensuremath{\mathbb{Q}}}
\newcommand{\Z}{\ensuremath{\mathbb{Z}}}
\newcommand{\N}{\ensuremath{\mathbb{N}}}
\renewcommand{\empty}{\varnothing}
% quick matrices and common fractions
\newcommand{\m}[1]{\begin{pmatrix}#1\end{pmatrix}} % a matrix or vector
\newcommand{\sm}[1]{\begin{psmallmatrix}#1\end{psmallmatrix}}
\newcommand{\half}{\frac{1}{2}}
\newcommand{\thalf}{\tfrac{1}{2}}
\newcommand{\third}{\frac{1}{3}}
\newcommand{\quarter}{\frac{1}{4}}
\newcommand{\eps}{\varepsilon} % A cursivelike epsilon
% common operators
\DeclareMathOperator{\id}{id}
\DeclareMathOperator{\sgn}{sgn}
\DeclareMathOperator*{\argmax}{arg\,max}
\DeclareMathOperator*{\argmin}{arg\,min}
% probability operators, abbrevations, and notions of convergence
\DeclareMathOperator{\E}{\mathbb{E}}
\let\Pr\undefined
\DeclareMathOperator{\Pr}{\mathbb{P}}
\newcommand{\indep}{\mathbin{\perp\!\!\!\!\!\:\perp}} % use \perp for orthogonality only
\newcommand{\gvn}{\;\middle|\;} % use \gvn rather than \mid inside a tall expectation
\newcommand{\F}{\ensuremath{\mathcal{F}}} % generic sigma algebra
\newcommand{\B}{\ensuremath{\mathcal{B}}}
\newcommand{\ind}{\mathbf{1}}
\newcommand{\iid}{\mathrel{\stackrel{iid}{\sim}}} % iid on top of ~
\newcommand{\rv}{random variable}
\newcommand{\rvs}{random variables}
\newcommand{\iidt}{independent and identically distributed}
\newcommand{\cvp}{\xrightarrow{\:p\,}}
\newcommand{\cvw}{\xrightarrow{\,w\,}}
\newcommand{\cvd}{\xrightarrow{\,d\,}}
\newcommand{\cvas}{\xrightarrow{\;\!a.s.\:\!}}
\newcommand{\cvlp}[1]{\xrightarrow{\;\!L^{#1}\,\!}} % e.g. \cvlp{2} for L2 conv.
% additional operators and common distributions
\DeclareMathOperator{\Var}{\mathbb{V}}
\DeclareMathOperator{\V}{\mathbb{V}}
\DeclareMathOperator{\cov}{cov}
\DeclareMathOperator{\cor}{corr}
\let\var\undefined
\DeclareMathOperator{\var}{var}
\DeclareMathOperator{\sd}{sd}
\DeclareMathOperator{\Unif}{Unif}
\DeclareMathOperator{\Expo}{Expo}
\DeclareMathOperator{\Cauchy}{Cauchy}
\DeclareMathOperator{\logit}{logit}
\newcommand{\Inv}{\mathrm{Inv-}}
\newcommand{\Pois}{\mathrm{Pois}\qty}
\newcommand{\Beta}{\mathrm{Beta}\qty}
\newcommand{\Dirichlet}{\mathrm{Dirichlet}\qty}
\newcommand{\Gam}{\mathrm{Gamma}\qty}
\newcommand{\Wei}{\mathrm{Wei}\qty}
\newcommand{\Hyper}{\mathrm{HGeom}\qty}
\newcommand{\Binom}{\mathrm{Bin}\qty}
\newcommand{\NBinom}{\mathrm{NBin}\qty}
\newcommand{\Multinom}{\mathrm{Multinom}\qty}
\newcommand{\Bern}{\mathrm{Bern}\qty}
\newcommand{\Bernoulli}{\mathrm{Bernoulli}\qty}
\newcommand{\Norm}{\mathcal{N}\qty}
\newcommand{\MVNorm}[1][]{\mathcal{N}_{#1}\qty} % e.g. \MVNorm{p} for a p-dim distr.
\newcommand{\Student}{\mathrm{Student-}\mathit{t}\qty}
\newcommand{\mhc}[1]{\textbf{\textcolor{NavyBlue}{(Melody: #1)}}}
% delimiters
\DeclarePairedDelimiter\br{\langle}{\rangle}
\DeclarePairedDelimiter\ceil{\lceil}{\rceil}
\DeclarePairedDelimiter\floor{\lfloor}{\rfloor}
\DeclarePairedDelimiter\round{\lceil}{\rfloor}
\DeclarePairedDelimiter\set{\{}{\}}
% NAMED ASSUMPTIONS %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\makeatletter
\newcommand{\assumplabel}[2]{%
\protected@write \@auxout {}{\string\newlabel{#1}{{#2}{\thepage}{#2}{#1}{}}}%
\hypertarget{#1}{#2}%
}
\makeatother
\newenvironment{assump}[2][]{%
\par\medskip\noindent%
\def\tempa{}\def\tempb{#1}%
\lowercase{\def\lblkey{a-#2}}%
\textbf{Assumption \assumplabel{\lblkey}{#2}}%
\ifx\tempa\tempb\else~(#1)\fi%
\textbf{.}\itshape
}{\medskip\normalfont}
\makeatletter
\@ifpackageloaded{caption}{}{\usepackage{caption}}
\AtBeginDocument{%
\ifdefined\contentsname
\renewcommand*\contentsname{Table of contents}
\else
\newcommand\contentsname{Table of contents}
\fi
\ifdefined\listfigurename
\renewcommand*\listfigurename{List of Figures}
\else
\newcommand\listfigurename{List of Figures}
\fi
\ifdefined\listtablename
\renewcommand*\listtablename{List of Tables}
\else
\newcommand\listtablename{List of Tables}
\fi
\ifdefined\figurename
\renewcommand*\figurename{Figure}
\else
\newcommand\figurename{Figure}
\fi
\ifdefined\tablename
\renewcommand*\tablename{Table}
\else
\newcommand\tablename{Table}
\fi
}
\@ifpackageloaded{float}{}{\usepackage{float}}
\floatstyle{ruled}
\@ifundefined{c@chapter}{\newfloat{codelisting}{h}{lop}}{\newfloat{codelisting}{h}{lop}[chapter]}
\floatname{codelisting}{Listing}
\newcommand*\listoflistings{\listof{codelisting}{List of Listings}}
\usepackage{amsthm}
\theoremstyle{plain}
\newtheorem{proposition}{Proposition}[section]
\theoremstyle{plain}
\newtheorem{lemma}{Lemma}[section]
\theoremstyle{plain}
\newtheorem{theorem}{Theorem}[section]
\theoremstyle{plain}
\newtheorem{corollary}{Corollary}[section]
\theoremstyle{remark}
\AtBeginDocument{\renewcommand*{\proofname}{Proof}}
\newtheorem*{remark}{Remark}
\newtheorem*{solution}{Solution}
\newtheorem{refremark}{Remark}[section]
\newtheorem{refsolution}{Solution}[section]
\makeatother
\makeatletter
\makeatother
\makeatletter
\@ifpackageloaded{caption}{}{\usepackage{caption}}
\@ifpackageloaded{subcaption}{}{\usepackage{subcaption}}
\makeatother
% END USER INCLUDES %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% BIBLIOGRAPHY %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\usepackage[]{natbib}
\bibliographystyle{apalike}
% Give it this name so that it works with ::: #refs
\newenvironment{CSLReferences}[2]{
\bibliography{references.bib}
\clearpage
}{}
% LINKS %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\usepackage{hyperref}
\usepackage{url}
\hypersetup{
pdftitle={Relative Bias Under Imperfect Identification in Observational Causal Inference},
pdfauthor={Melody Huang; Cory McCartan},
pdfkeywords={sensitivity analysis, proximal inference, instrumental
variables},
colorlinks=true,
linkcolor={black},
filecolor={Maroon},
citecolor={VioletRed4},
urlcolor={DodgerBlue4},
pdfcreator={LaTeX via pandoc}}
% TITLE, AUTHOR, DATE %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\title{\sffamily\bfseries\huge\parfillskip=0pt
\rightskip=0pt plus .5\textwidth
\leftskip=0pt plus .5\textwidth
\emergencystretch=.3\textwidth Relative Bias Under Imperfect
Identification in Observational Causal Inference}
\shorttitle{Relative Bias Under Imperfect Identification in Observational Causal Inference}
\author{\textbf{Melody Huang}\footnote{
To whom correspondence should be addressed.
Email: \texttt{\href{mailto:melody.huang@yale.edu}{melody.huang@yale.edu}}.
The authors thank Chad Hazlett, Kosuke Imai, Sam Pimentel, Eric Tchetgen
Tchetgen, Carlos Cinelli, participants at PolMeth XLII, and the Oxford
Causal Inference Reading Group for helpful comments and suggestions.}
\\Department of Political Science%
%
\vspace{2pt}
\\Department of Statistics \& Data Science%
\\Yale University%
\vspace{2pt}
\and \textbf{Cory McCartan}
\\Department of Statistics%
\\Pennsylvania State University%
\vspace{2pt}
}
\date{July 31, 2025}
\begin{document}
\allsectionsfont{\sffamily}
\maketitle
\begin{abstract}
To conduct causal inference in observational settings, researchers must
rely on certain identifying assumptions. In practice, these assumptions
are unlikely to hold exactly. This paper considers the bias of
selection-on-observables, instrumental variables, and proximal inference
estimates under violations of their identifying assumptions. We develop
bias expressions for IV and proximal inference that show how violations
of their respective assumptions are amplified by any unmeasured
confounding in the outcome variable. We propose a set of sensitivity
tools that quantify the sensitivity of different identification
strategies, and an augmented bias contour plot visualizes the
relationship between these strategies. We argue that the act of choosing
an identification strategy implicitly expresses a belief about the
degree of violations that must be present in alternative identification
strategies. Even when researchers intend to conduct an IV or proximal
analysis, a sensitivity analysis comparing different identification
strategies can help to better understand the implications of each set of
assumptions. Throughout, we compare the different approaches on a
re-analysis of the impact of state surveillance on the incidence of
protest in Communist Poland.
\end{abstract}
\textbf{\textit{Keywords}}\quad sensitivity
analysis~\textbullet~proximal inference~\textbullet~instrumental
variables
% USER BODY %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\clearpage
\section{Introduction}\label{sec-intro}
Causal inference in observational studies requires identification
assumptions which account for potential confounding that occurs from
non-random assignment. The most familiar set identification assumptions
is that of no unobserved confounding, also known as ignorable treatment
assignment, or \emph{selection on observables} (SOO): given observed
covariates, potential outcomes are independent of treatment. In many
observational settings, however, SOO is considered implausible. The
presence of unobserved confounders can bias treatment effects which are
estimated under an SOO assumption.
Alternative identification strategies have been proposed as an
alternative to the selection-on-observables assumptions. Researchers may
try to find an \emph{instrumental variable} (IV) which is associated
with treatment but satisfies an exclusion restriction with the outcome.
Under certain additional assumptions, an IV allows for estimating the
effect of interest. More recently, \emph{proximal} inference has been
developed as an alternative identification approach
\citep{miao2015identification, miao2018identifying, cui2024semiparametric}.
Instead of assuming researchers measure all relevant confounders,
proximal inference assumes researchers have access to two informative
proxies: a treatment proxy and an outcome proxy. If these proxies
satisfy certain conditional independence assumptions, not unlike the
exclusion restriction in IV, then a treatment effect can be estimated,
even though the proxies themselves are confounded.
The existence of three distinct identification approaches leads to
natural questions about how the corresponding estimators compare in
practice. Much work has been conducted to compare the relative
efficiency of the different approaches \citep[see,
e.g.,][]{andrews2019weak}. In particular, weak instruments, and by
extension, weak proxies, can inflate the variance of IV and proximal
estimators, such that SOO estimators will have lower mean squared error
even in settings when the IV and proximal assumptions hold, and the SOO
assumptions do not.
Fewer studies have examined how the bias of different identification
strategies compares. The nature of identification assumptions is that
they are untestable; in real-world observational data, it is unlikely
that any set of identifying conditions, SOO, IV, or proximal, will hold
exactly. It is important, then, to understand the bias that may result
from plausible violations of each set of identifying assumptions, and
how this bias compares between identification strategies. Such a
comparison is useful not only in picking an identification strategy, but
also in understanding the assumptions made as part of a primary causal
analysis under a single strategy.
Recent literature has introduced different sensitivity frameworks to
evaluate the sensitivity of estimators to potential violations in the
underlying identifying assumptions for SOO
\citep[e.g.,][]{rosenbaum1987sensitivity, frank2000impact, imbens2003sensitivity, tan2006distributional, carnegie2016assessing, ding2019decomposing, zhao2019sensitivity, cinelli2020making, hong2021did, dorn2023sharp, huang2025variance},
for IV
\citep{kang2021ivmodel, freidling2022optimization, cinelli2025iv}, and
for proximal \citep{cobzaru2024bias}, to name a few examples. However,
the different frameworks and approaches make it challenging to compare
bias \emph{across} identification strategies.
This paper introduces a framework for researchers to reason about the
robustness of different identification approaches under the realistic
setting that none of the identification assumptions hold exactly. Our
framework enables researchers to carefully consider the trade-offs made
between using a traditional SOO, IV, and proximal assumptions, and
understand the conditions under which each identification strategy will
provide the best estimates of the underlying effect.
First, in Section~\ref{sec-bias}, we derive expressions for the bias of
both IV and proximal estimators in terms of the violations of their
underlying assumptions. We parametrize the bias expressions using a
unified set of partial \(R^2\) parameters, and show that the bias that
arises from violations in one set of identification assumptions can
impact the bias in alternative identification approaches. Specifically,
the bias that arises in IV will additionally depend on violations in the
SOO assumptions, while the bias for a proximal estimator will depend on
\emph{both} violations in the IV assumptions and the SOO assumptions.
Second, in Sections \ref{sec-sens-comp} and \ref{sec-unify}, we propose
a set of sensitivity tools in the form of numerical and visual summary
measures. To compare the sensitivities of the various identification
approaches, we introduce a generalization of standard robustness values
called the \emph{total robustness value}, which measures the minimum
amount of total violation of a set of identification assumptions that
result in a certain amount of bias. Then, to connect the identifying
assumptions themselves, we show that the bias of both IV and proximal
inference can be written as a function of just the
selection-on-observables parameters and the data. This unification
allows us to relate the three identification strategies on a single bias
contour plot.
Our sensitivity framework can aid practitioners in several ways. Often,
applied researchers identify a causal effect of interest, consider
plausible identification strategies, and then collect data to enable
inference under their chosen strategy. In contrast to existing
sensitivity analyses, which work with a single identification strategy,
our proposed measures are constructed to directly account for the nested
dependency in violations across the different identification
assumptions. This allows researchers to directly compare different
identification strategies and choose one most appropriate for their
data.
The proposed framework is still useful, however, for researchers who
have already settled on an identification strategy for their primary
analysis. As both the bias expressions in Section~\ref{sec-bias} and the
unification in Section~\ref{sec-unify} show, the SOO, IV, and proximal
assumptions are deeply interrelated. Even when the SOO assumptions are
not plausible, the value of the SOO \emph{estimate} places certain
restrictions on plausible values for the different sensitivity
parameters. For example, the observed data may imply that \emph{either}
the IV assumptions hold \emph{or} a certain minimum amount of treatment
confounding must exist. If an analyst is confident about the direction
of confounding bias, that may also immediately rule out the plausibility
of certain other identifying assumptions. Thus our framework provides
for better interrogation and integration of identifying assumptions,
allowing researchers to consider their assumptions from a different
angle and move beyond fully \emph{a priori} reasoning about their
plausibility.
\subsection{Running example: Surveillance and Protest in Communist
Poland}\label{sec-intro-ex}
To illustrate our discussion and demonstrate the proposed sensitivity
analyses in a concrete setting, we re-analyze data from
\citet{hk2022poland}, who studied the effects of state surveillance on
resistance to the Polish regime, as measured by group and individual
protests. The authors digitized records from the Służba Bezpieczeństwa
(SB; Department of Security) from 1949--1989, covering the entire
existence of the Polish People's Republic. These records include the
number of secret agents assigned to each municipality each year. The
authors focus on the region of Upper Silesia, where data was
particularly comprehensive. They also collected data on protests
organized by the Solidarność (Solidarity) movement, which was the center
of resistance to the governing regime in the 1980s, as well as indirect
data on individual refusals to work on Saturdays. The research question
is whether additional surveillance of a municipality caused either group
or individual protest to increase or decrease.
Hager and Krakowski's primary analysis is based on an instrumental
variable: the number of local Catholic priests who were secretly
cooperating with the SB. The police intentionally recruited priests, who
in turn recruited other local informants. However, priests were randomly
assigned to parishes by the Catholic Church, making the number of
\emph{corrupt} priests plausibly exogenous. The authors include a
detailed study and discussion of the validity of the instrument, which
we also discuss below. As a robustness check, they also conducted a
cross-sectional analysis under a SOO assumption. Both analyses found
that state surveillance increased group protests but decreased
individual protests. Here, we study the sensitivity of both the IV and
SOO findings, and also conduct a sensitivity analysis for proximal
inference using the same data.
\section{Setup and Background}\label{setup-and-background}
This section provides an overview of three common identification
strategies for the ATE in cross-sectional observational settings:
selection-on-observables, instrumental variables, and proximal
inference. We express these strategies in terms of a common statistical
model, which we introduce next, and we discuss the plausibility of the
different identifying assumptions in the context of our running example.
We study the following partially linear structural model for an outcome
\(Y\), binary treatment \(Z\), vector of covariates \(X\), unobserved
confounder \(U\), and two additional variables \(W_Y\) and \(W_Z\).
\begin{equation}\phantomsection\label{eq-struct}{
\begin{aligned}
Y &= \beta_u U + \beta_{w_y} W_Y + \beta_{w_z} W_Z + \tau Z + f_y(X) + \eps_y \\
Z &= \gamma_u U + \gamma_{w_y} W_Y + \gamma_{w_z} W_Z + f_z(X) + \eps_z \\
W_Y &= \alpha_u U + \alpha_{w_z} W_Z + f_{w_y}(X) + \eps_{w_y} \\
W_Z &=\varphi_u U + f_{w_z}(X) + \eps_{w_z}
\end{aligned}
}\end{equation} Above, \(\eps_y, \eps_z, \eps_{w_z}, \eps_{w_y}\), and
\(\eps_u\) are independent noise terms, and \(f_y\), \(f_z\),
\(f_{w_y}\), and \(f_{w_z}\) are generic functions of the covariates
\(X\). This can be thought of as a special case of the fully
non-parametric setting studied in \citet{miao2018identifying}. For
expositional clarity, the rest of the paper works with the reduced model
that (nonparametrically) partials out \(X\); this also has the effect of
centering the remaining variables at zero. In a slight abuse of
notation, we will use the same notation for the variables and
coefficients in the reduced model.
The structural assumptions encoded in Eq.~\ref{eq-struct} mean that the
potential outcomes for \(Y\) under control and treatment, which we write
as \(Y(0)\) and \(Y(1)\), can be obtained by substituting \(Z=0\) and
\(Z=1\) into the equation for \(Y\).\footnote{The stable unit treatment
value assumption (SUTVA) is also implied by this setup.} As a result,
the average treatment effect (ATE) is \(\E[Y(1) - Y(0)] = \tau\).
Figure~\ref{fig-dag} illustrates the setup and the assumptions discussed
below.
\begin{figure}[t]
\begin{minipage}{0.33\linewidth}
\centering{
\tikzset{> = {Latex[length=6pt]},
every path/.append style = {arrows = ->, thick},
every node/.append style = {inner sep=0, font = \footnotesize}}
\tikz[scale=0.7]{
\node[draw, circle, minimum size=18pt] (u) at (0, 0.25) {$U$};
\node[draw, circle, minimum size=18pt] (z) at (-2, -2) {$Z$};
\node[draw, circle, minimum size=18pt] (y) at (2, -2) {$Y$};
\node[draw, circle, minimum size=18pt] (wz) at (-3, 1) {$W_Z$};
\node[draw, circle, minimum size=18pt] (wy) at (3, 1) {$W_Y$};
\path[dashed] (u) edge (z);
\path[dashed] (u) edge (y);
\path (u) edge (wz);
\path (u) edge (wy);
\path (z) edge (y);
\path (wz) edge (z);
\path (wz) edge (wy);
\path (wz) edge (y);
\path (wy) edge (z);
\path (wy) edge (y);
}
}
\subcaption{\label{fig-dag-soo}Selection-on-observables}
\end{minipage}%
%
\begin{minipage}{0.33\linewidth}
\centering{
\tikzset{> = {Latex[length=6pt]},
every path/.append style = {arrows = ->, thick},
every node/.append style = {inner sep=0, font = \footnotesize}}
\tikz[scale=0.7]{
\node[draw, circle, minimum size=18pt] (u) at (0, 0.25) {$U$};
\node[draw, circle, minimum size=18pt] (z) at (-2, -2) {$Z$};
\node[draw, circle, minimum size=18pt] (y) at (2, -2) {$Y$};
\node[draw, circle, minimum size=18pt] (wz) at (-3, 1) {$W_Z$};
\node[draw, circle, minimum size=18pt] (wy) at (3, 1) {$W_Y$};
\path (u) edge (z);
\path (u) edge (y);
\path (u) edge (wy);
\path (z) edge (y);
\path (wz) edge (z);
\path (wz) edge (wy);
\path (wy) edge (z);
\path (wy) edge (y);
}
}
\subcaption{\label{fig-dag-iv}Instrumental variables}
\end{minipage}%
%
\begin{minipage}{0.33\linewidth}
\centering{
\tikzset{> = {Latex[length=6pt]},
every path/.append style = {arrows = ->, thick},
every node/.append style = {inner sep=0, font = \footnotesize}}
\tikz[scale=0.7]{
\node[draw, circle, minimum size=18pt] (u) at (0, 0.25) {$U$};
\node[draw, circle, minimum size=18pt] (z) at (-2, -2) {$Z$};
\node[draw, circle, minimum size=18pt] (y) at (2, -2) {$Y$};
\node[draw, circle, minimum size=18pt] (wz) at (-3, 1) {$W_Z$};
\node[draw, circle, minimum size=18pt] (wy) at (3, 1) {$W_Y$};
\path (u) edge (z);
\path (u) edge (y);
\path (u) edge (wz);
\path (u) edge (wy);
\path (z) edge (y);
\path (wz) edge (z);
\path (wy) edge (y);
}
}
\subcaption{\label{fig-dag-prox}Proximal inference}
\end{minipage}%
\caption{\label{fig-dag}The causal structure studied in this paper,
after partialing out any covariates \(X\). Selection-on-observables
assumes that at least one of the dashed arrows is absent.}
\end{figure}%
Throughout the paper, we take a population view of the linear models.
Thus, the `hat' above different terms (i.e., \(\hat Z\) and
\(\hat W_Y\)) is used to denote that these are predicted values (at the
population-level), not that they correspond to sample estimates.
\subsection{Selection on observables}\label{selection-on-observables}
The ATE \(\tau\) can be estimated from Eq.~\ref{eq-struct} only if \(U\)
is observed, which it is not. The simplest way to estimate \(\tau\) in
the presence of an unobserved \(U\) is to assume that the observed
covariates \(X\) contain all relevant confounders, an assumption known
as selection-on-observables (SOO). In our model, SOO means that at least
one of the following two assumptions holds.
\begin{assump}[Unconfounded treatment]{SOO-Z}
$Z\indep U\mid X, W_Z, W_Y$.
\end{assump}
\begin{assump}[Unconfounded outcome]{SOO-Y}
$Y\indep U\mid X, Z, W_Z, W_Y$.
\end{assump}
Assumption \ref{a-soo-z} corresponds to \(\gamma_u=0\) in
Eq.~\ref{eq-struct}, and Assumption \ref{a-soo-y} corresponds to
\(\beta_u=0\). Either of these assumptions is sufficient for \(\tau\) to
be identified, and estimated via a regression on the observed variables,
with \(\tau_\soo\) the coefficient on \(Z\) in this regression. The
following standard result, whose proof is omitted, records this fact.
\begin{proposition}[SOO
validity]\protect\hypertarget{prp-soo-valid}{}\label{prp-soo-valid}
If either \ref{a-soo-z} or \ref{a-soo-y} holds, \(\tau_\soo=\tau\).
\end{proposition}
In practice, the selection-on-observables assumptions cannot be tested
and are often untenable. Researchers are often constrained by the
pre-treatment covariates that are available in observational contexts.
In the presence of latent, mismeasured, or unknown confounders, relying
on a SOO identification strategy will generally result in biased
estimates.
Our running example \citep{hk2022poland} involves several
municipality-level covariates, including subregion indicators, counts of
different types of facilities, demographic covariates, and historical
information about coal deposits and Russian occupation. A full
description of these covariates, and some discussion of their
plausibility in satisfying SOO, may be found in
Appendix~\ref{sec-app-covar}. In short, while several possible
confounders are included, there are several reasonable confounding
pathways that remain, as well as potential issues with measurement error
and dynamic effects. For these reasons, the authors relied on an IV
strategy for their primary analysis.
\subsection{Instrumental Variables}\label{instrumental-variables}
Instrumental variables have long been used in contexts where researchers
believe that confounding remains after controlling for observed
covariates. Informally, this identification approach relies on on
finding a variable \(W_Z\), the \emph{instrument}, that can explain
variation in the treatment \(Z\), but can only affect the outcome \(Y\)
through its relationship with \(Z\). The variation in \(W_Z\) can then
be used to identify the treatment effect \(\tau\), even in the presence
of an unobserved confounder \(U\). While IV no longer requires
Assumptions \ref{a-soo-z} or \ref{a-soo-y}, identification under our
model requires that \(W_Z\) satisfy the following two assumptions.
\begin{assump}[Exogenous instrument]{IV-EXOG}
$W_Z\indep U\mid X, W_Y$.
\end{assump}
\begin{assump}[Exclusion Restriction]{IV-EXCL}
$Y\indep W_Z\mid X, U, Z, W_Y$.
\end{assump}
An instrument \(W_Z\) is considered a \emph{valid instrument} if both
exclusion restriction, as well as exogeneity, holds. In the context of
Eq.~\ref{eq-struct}, Assumption \ref{a-iv-exog} corresponds to
\(\varphi_u=0\) and Assumption \ref{a-iv-excl} corresponds to
\(\beta_{w_z}=0\). The instrument must also be \emph{relevant}, meaning
that \(\gamma_{w_z}\neq 0\), so it has some effect on \(Z\). Given a
valid instrument, the most common way to estimate the treatment effect
is via two-stage least squares, which will recover \(\tau\). In general
settings beyond the model in Eq.~\ref{eq-struct}, additional assumptions
about treatment effect homogeneity or monotonicity are required to
identify a causal effect.
\begin{proposition}[IV
validity]\protect\hypertarget{prp-iv-valid}{}\label{prp-iv-valid}
Let \(\tau_\iv\) be estimated via two-stage least squares:
\begin{enumerate}
\def\labelenumi{\arabic{enumi}.}
\tightlist
\item
Regress \(Z\) on \(W_Z\) and \(W_Y\) and generate predicted values
\(\hat Z\).
\item
Regress \(Y\) on \(\hat Z\), \(W_Y\), and \(W_Z\). The least-squares
coefficient on \(\hat Z\) is denoted \(\tau_\iv\).
\end{enumerate}
\noindent Then, under Assumptions \ref{a-iv-exog} and \ref{a-iv-excl},
\(\tau_\iv=\tau\).\footnote{The derivation of this result relies on a
homogeneous treatment effect \(\tau\), which we have assumed from the
partial linearity of the structural model. Alternative identification
approaches, which we do not study, replace homogeneity with other
structural assumptions such as monotonicity.}
\end{proposition}
As Section~\ref{sec-intro-ex} describes, in our running example from
\citet{hk2022poland}, the authors used the number of local Catholic
priests who were secretly cooperating with the SB as an instrumental
variable. They show the number of corrupted priests is strongly
correlated with the number of SB officers, and justify both IV
identification assumptions using both substantive knowledge and
auxiliary data. However, like all identification approaches, there still
remain existing mechanisms that could violate either assumption. For
example, if the Catholic Church reserved assignments to important
cities, such as those with a history of protest or surveillance, for
priests that they judged to be more resistant to recruitment, then the
exogeneity assumption would no longer hold. Similarly, the authors note
that the possibility of indoctrination of parishioners by corrupted
priests would result in a violation of exclusion restriction. We provide
more discussion in Section~\ref{sec-app-covar}. All in all, like many
other real-world settings, the complex social dynamics in
\citet{hk2022poland} make it unlikely that both IV assumptions hold
exactly, and warrant investigation of the sensitivity of the estimates
to violations of either assumption.
\subsection{Proximal Inference}\label{proximal-inference}
While IV allows researchers to relax the selection-on-observables
assumptions in exchange for two instrumental variables assumptions,
finding a valid instrument can be challenging. In particular, an
instrument \(W_Z\) must explain variation in \(Z\), but be exogenous.
Recently, \citet{miao2015identification} proposed an alternative
identification strategy, \emph{proximal inference}, which relaxes the IV
assumption of exogeneity.
Proximal inference relies on finding two `proxy variables': (1) an
outcome proxy \(W_Y\) that is unrelated to the treatment but is related
to the outcome, and (2) a treatment proxy \(W_Z\) that is unrelated to
the outcome, but is related to the treatment. The treatment proxy is
similar to an instrumental variable, but does not need to be exogenous.
As such, proximal inference can be viewed as an alternative to
instrumental variables, in settings when researchers are not confident
they have an exogenous instrument.
More formally, the proximal inference approach of
\citet{miao2015identification} replaces SOO or IV assumptions with two
assumptions about the proxy variables.
\begin{assump}[Outcome proxy unrelated to treatment]{PROX-Y}
$W_Y\indep Z, W_Z\mid X, U$.
\end{assump}
\begin{assump}[Treatment proxy unrelated to outcome]{PROX-Z}
$W_Z\indep Y\mid X, U, Z, W_Y$.
\end{assump}
When both Assumptions \ref{a-prox-y} and \ref{a-prox-z} hold, we say
that \(W_Y\) and \(W_Z\) are valid proxy variables. In the context of
Eq.~\ref{eq-struct}, Assumption \ref{a-prox-y} corresponds to
\(\gamma_{w_y}=0\) and \(\alpha_{w_z}=0\), and Assumption \ref{a-prox-z}
means \(\beta_{w_z}=0\). Notice that Assumption \ref{a-prox-z} is
exactly the same as Assumption \ref{a-iv-excl}.
Like instrumental variables, a two-stage least squares procedure can be
used to estimate the treatment effect. However, instead of regressing
the treatment indicator \(Z\) with the instrument \(W_Z\), the first
stage relies on regressing the outcome proxy \(W_Y\) with \(Z\) and
\(W_Z\). Because \(W_Y\) has no direct causal connection with \(Z\) and
\(W_Z\) under Assumption \ref{a-prox-y}, any correlation between the two
sets of variables can only be due to the unobserved confounder \(U\),
and thus the first stage amounts to estimating a proxy for \(U\). That
estimated \(\hat W_Y\) is then included in the second-stage, along with
\(Z\). When both \(W_Y\) and \(W_Z\) are valid proxies, the estimated
treatment effect under proximal inference (denoted \(\tau_{\prox}\))
will recover the average treatment effect \(\tau\).
\begin{proposition}[Proximal
validity]\protect\hypertarget{prp-prox-valid}{}\label{prp-prox-valid}
Define \(\tau_{\prox}\) as the solution to the following two-stage least
squares procedure:
\begin{enumerate}
\def\labelenumi{\arabic{enumi}.}
\tightlist
\item
Regress \(W_Y\) on \(Z\) and \(W_Z\) and generate predicted values
\(\hat W_Y\).
\item
Regress \(Y\) on \(Z\) and \(\hat W_Y\). The estimated coefficient on
\(Z\) is denoted \(\tau_\prox\).
\end{enumerate}
\noindent Then, under Assumptions \ref{a-prox-y} and \ref{a-prox-z},
\(\tau_\prox = \tau\).
\end{proposition}
The proof, whose details appear in Appendix~\ref{sec-app-proofs}, relies
on the Frisch-Waugh-Lovell theorem: \begin{align*}
\tau_{\prox} &= \frac{\cov(Y^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})} \\
&= \frac{\cov\qty((\beta_u U + \beta_{w_y} W_Y + \beta_{w_z} W_Z + \tau Z + \eps_y)^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})} \\
&= \tau +
\beta_u \frac{\cov(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})} +
\beta_{w_y} \frac{\cov(W_Y^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})} +
\beta_{w_z} \frac{\cov(W_Z^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})}.
\end{align*} The construction of \(\hat W_Y\) ensures
\(\cov(W_Y^{\bot \hat W_Y}, Z^{\bot \hat W_Y})=0\) (see
Lemma~\ref{lem-wy-y}), which also makes sense if \(\hat W_Y\) is viewed
as a proxy for \(U\). Assumption \ref{a-prox-z} guarantees
\(\beta_{w_z}=0\), and it turns out that Assumption \ref{a-prox-y}
guarantees that \(\cov(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y})=0\), so the
two-stage approach is valid.
\citet{hk2022poland} did not originally run a causal analysis under
proximal assumptions. Yet, as described above, both SOO and IV
identification strategies face plausible challenges to their validity,
which proximal identification might address. The treatment proxy \(W_Z\)
is the count of corrupted priests, corresponding to the instrument under
IV. For the outcome proxy \(W_Y\), we argue that the indicator for
former Russian occupation is a reasonable candidate. Hager and Krakowski
explain that Russia implemented ``russification'' policies in these
areas ``which were intended to destroy Polish communities and
identities,'' in turn making protest against the Polish regime less
likely. A mechanism for affecting the amount of surveillance is less
obvious; the authors suggest in passing that Russia's experience in the
occupied territories may have made administering spy networks easier,
but we find this mechanism less plausible. Indeed, the occupation
variable has a partial correlation with the group protest outcome about
three times larger than the correlation with the treatment, after
controlling for other covariates, and a ten-times larger correlation
with the individual protest outcome than treatment.
Proximal identification is no panacea, however. Valid inference still
hinges on the exclusion restriction holding, and the outcome proxy must
satisfy two additional exclusion restrictions. Specifically, Russian
occupation cannot directly affect the number of corrupted priests or the
amount of surveillance. While plausible for some of the reasons Hager
and Krakowski outline in their discussion of the exogeneity of the
instrument, it is not unreasonable to assume that the same weakening of
Polish communities in occupied areas might affect which priests are
assigned and their success in recruiting informants. Thus, as with the
other two identification strategies, it remains important to examine the
sensitivity of estimates to violations of the identifying assumptions.
\subsection{Summary and illustration on running
example}\label{summary-and-illustration-on-running-example}
Table \ref{tbl-id} summarizes the three identification approaches, their
assumptions, and the regressions used in estimation. It also displays
the estimated effect of surveillance on both group and individual
protests in our running example, under each estimation approach. The
scale of the effects is difficult to interpret due to several layers of
rescaling and pre-processing in the underlying data. Here, we have
rescaled both the outcome and treatment variables to have unit variance,
to ease comparison and discussion of the estimates. Thus a value of
\(0.3\) below means a one-s.d. increase in surveillance is associated
with an \(0.3\)-s.d. increase in
protest.\footnote{The same estimate corresponds to a value of $0.068$ in the original paper.}
Like \citet{hk2022poland}, we find opposite signs for the effect on
group and individual protest. Notice that the estimates are all
statistically significant at traditional levels, except the proximal
estimate for the group outcome. One immediately apparent aspect of the
estimates in Table \ref{tbl-id} is that the three approaches give nearly
identical estimates for the effect on group protests (though different
standard errors), but disagree on the effect of individual protest.
\setstretch{1.0}
\begin{table}
\caption{\label{tbl-id}Summary of identification approaches, with estimates from the running example. Sample size is 216 for the group outcome and 206 for the individual outcome due to missing outcome data.}
\centering{
\centering
\begin{tabular}{>{\raggedright\arraybackslash}p{2.7cm}p{4cm}p{4cm}p{4cm}} \toprule
\textbf{} & \textbf{SOO} (Figure \ref{fig-dag-soo}) & \textbf{IV} (Figure \ref{fig-dag-iv}) & \textbf{Proximal} (Figure \ref{fig-dag-prox}) \\
\midrule
Identification assumptions
& Either \ref{a-soo-z} \textit{or} \ref{a-soo-y}
& \ref{a-iv-exog} and \ref{a-iv-excl}
& \ref{a-prox-z} and \ref{a-prox-y} \\
Implied coefficients
& $\beta_u \gamma_u = 0$
& $\beta_{w_z}=\varphi_u=0$
& $\gamma_{w_y}=\alpha_{w_z}=\beta_{w_z}=0$ \\
Estimation
& $Y \sim Z + W_Z + W_Y$
& \begin{tabular}[t]{@{}l@{}}
TSLS: \\
1. $Z \sim W_Z + W_Y$ \\
2. $Y \sim \hat{Z} + W_Y + W_Z$
\end{tabular}
& \begin{tabular}[t]{@{}l@{}}
TSLS: \\
1. $W_Y \sim Z + W_Z$ \\
2. $Y \sim Z + \hat{W_Y}$
\end{tabular} \\
Estimate (s.e.) for group protest
& $\phantom{-}0.347\ \ (0.058)$
& $\phantom{-}0.347\ \ (0.128)$
& $\phantom{-}0.348\ \ (0.373)$ \\
Estimate (s.e.) for individual protest
& $-0.296\ \ (0.051)$
& $-0.805\ \ (0.108)$
& $-0.627\ \ (0.068)$ \\ \bottomrule
\end{tabular}
}
\end{table}%
\setstretch{\papersp}
\section{Bias decompositions}\label{sec-bias}
This section presents decompositions of the bias that arises from
violations in any of the underlying identification assumptions. We find
that identification strategies are more sensitive to violations of their
assumptions when there is more unobserved confounding. In other words,
violations in Assumption \ref{a-soo-y} will \emph{amplify} any
violations in the alternative identification assumptions.
The results here are derived for the population estimates. Because our
model (Eq.~\ref{eq-struct}) is linear, however, they hold equally when
all correlations and coefficients are interpreted as their finite-sample
versions.
\subsection{Review: Bias of selection-on-observables
estimator}\label{sec-bias-soo}
For selection-on-observables to be a valid identification approach,
there cannot be an omitted confounder \(U\) that affects both the
outcome \(Y\) and the treatment assignment process \(Z\). In the
presence of an omitted \(U\), the bias for \(\tau_\soo\) will be a
function of \(U\)'s relationship with \(Y\) and \(Z\), respectively.
This can be expressed in terms of the coefficients in
Eq.~\ref{eq-struct}, or, in an equivalent formulation derived by
\citet{cinelli2020making}, in terms of partial correlations:
\begin{equation}\phantomsection\label{eq-bias-soo}{
\bias(\tau_\soo) = \beta_u \cdot \gamma_u \cdot \frac{\var(U^{\bot W_Z, W_Y})}{\var(Z^{\bot W_Z, W_Y})}
= \frac{\hlsoo{R_{Y \sim U \mid W_Z, W_Y, Z}} \cdot \hlsoo{R_{Z \sim U \mid W_Z, W_Y}}}{
\sqrt{1-\hlsoo{R^2_{Z \sim U \mid W_Z, W_Y}}}} \cdot
\frac{\sd(Y^{\bot W_Y, W_Z})}{\sd(Z^{\bot W_Y, W_Z})}
}\end{equation} where e.g., \(R_{Y\sim U\mid W_Z,W_Y}\) is notation for
\(\cor(Y^{\perp W_Z,W_Y}, U^{\perp W_Z,W_Y})\).\footnote{The original
bias decomposition proposed by \citet{cinelli2020making} considers the
absolute bias, and uses the squared correlations (i.e.,
\(R^2_{Y \sim U \mid W_Z, W_Y}\)). Because we care about the
directionality of the bias and not just the magnitude, we use the
partial correlation representation of the sensitivity parameters in
lieu of the \(R^2\) representation.} See Lemma~\ref{lem-soo} for the
full derivation.
Consider a potential confounder of \emph{state capacity} in the running
example. Then, \(\hlsoo{R_{Y \sim U \mid W_Z,W_Y,Z}}\) represents the
relationship between state capacity and the outcome (either group or
individual protest) that is not already captured by the observed proxies
of state capacity (i.e., number of schools). Similarly,
\(\hlsoo{R_{Z \sim U \mid W_Z,W_Y}}\) represents the residual imbalance
of state capacity across the treatment and control groups, after
controlling for the observed covariates. If researchers believe that
proxies of state capacity sufficiently capture the variation in state
capacity, then we expect \(\hlsoo{R_{Y \sim U \mid W_Z,W_Y,Z}}\) and
\(\hlsoo{R_{Z \sim U \mid W_Z,W_Y}}\) to be relatively small in
magnitude, and little bias to result for not controlling for the latent
variable of state capacity. In contrast, researchers could be worried
that the measurement error in using the number of schools as a proxy of
state capacity could be related to protests, in which case, we might
expect there to be more bias from omitting a variable like state
capacity.
Importantly, both representations of bias formalize the well-studied
result that if omitted confounder \(U\) is related to \emph{both} the
outcome process \emph{and} the treatment assignment process, there will
be bias \citep{imbens2003sensitivity}. Even if \(U\) is imbalanced
between the treatment and control groups, if it does not explain
variation in the outcome, then omitting it will not lead to bias, since
\(\beta_u = 0\) (and \(R_{Y \sim U \mid W_Z, W_Y} = 0\)). Similarly, if
\(U\) explains variation in the outcome but does not explain variation
in the treatment assignment process, then \(\gamma_u = 0\) (and
\(R_{Z \sim U \mid W_Z, W_Y} = 0\)), while it may be advantageous to
control for \(U\) for precision, doing so is not strictly necessary to
eliminate bias.
\subsection{Bias of instrumental variables}\label{sec-bias-iv}
Theorem~\ref{thm-iv} presents a decomposition of the bias of the
instrumental variables estimate. The bias decomposition highlights that
there are dependencies between the bias of the \(\tau_\iv\) and the
selection-on-observables assumptions. As such, even though \(\tau_\iv\)
relies on different identification assumptions as \(\tau_\soo\), the
bias in \(\tau_\iv\) can be amplified by violations in the
selection-on-observables assumptions. Appendix~\ref{sec-app-exog}
further discusses the special case when researchers are willing to
assume exogeneity of the instrument holds exactly, and wish to examine
sensitivity to the exclusion restriction.
\begin{theorem}[Bias of IV
estimate]\protect\hypertarget{thm-iv}{}\label{thm-iv}
The bias of \(\tau_\iv\) is given by
\begin{equation}\phantomsection\label{eq-iv-bias-r2}{
\begin{aligned}
\bias(\tau_{\iv}) &= \frac{1}{R_{Z\sim W_Z\mid W_Y}} \cdot
\frac{\sd(W_Z^{\bot W_Y})}{\sd(Z^{\bot W_Y})} \left\{\beta_{w_z} + \beta_u \cdot \varphi_u \right\} \\
&= \underbrace{\frac{1}{R_{Z\sim W_Z\mid W_Y}} \cdot \frac{\sd(Y^{\bot Z,W_Y})}{\sd(Z^{\bot W_Y})}}_{\text{(a) Instrument Strength}} \times \underbrace{\left\{~\hliv{R_{Y \sim W_Z \mid Z,W_Y,U}} \cdot \sqrt{\frac{1-R^2_{Y \sim U \mid Z, W_Y}}{1-R^2_{W_Z \sim U \mid Z,W_Y}} \cdot \frac{1}{1-R^2_{Z \sim W_Z \mid W_Y}}} \right.}_{\text{(b) Violations in Exclusion Restriction}} \\
&\qquad+\underbrace{\left. \hlsoo{R_{Y \sim U \mid Z, W_Z, W_Y}} \hliv{R_{W_Z \sim U \mid W_Y}} \cdot \sqrt{\frac{1}{1-\hlsoo{R^2_{Z \sim U \mid W_Z, W_Y}}} \cdot \frac{1}{1-\hliv{R^2_{W_Z \sim U \mid W_Y}}} \cdot R^2_{Y \sim W_Z \mid Z,W_Y}}~\right\}}_{\text{(c) Violations in Exogneity}},
\end{aligned}
}\end{equation} where \(\beta_{w_z}, \beta_u, \varphi_u\) correspond to
the parameters in Eq.~\ref{eq-struct}.
\end{theorem}
The bias in Eq.~\ref{eq-iv-bias-r2} depends on three key terms: (a) a
scaling factor depends on the strength of the instrument, (b) bias from
violations in the exclusion restriction, and (c) the bias from
violations in instrument exogeneity. We discuss each of these terms in
more detail below.
\subparagraph{Violations in Exclusion
Restriction.}\label{violations-in-exclusion-restriction.}
The bias that arises from violations in the exclusion restriction,
Eq.~\ref{eq-iv-bias-r2}(b), is directly proportional to the
\(\beta_{w_z}\) coefficient from Eq.~\ref{eq-struct}. The main driver is
\(\hliv{R_{Y \sim W_Z \mid W_Y, Z, U}}\), which directly maps to the
violation of Assumption \ref{a-iv-excl}, measuring how related the
instrument \(W_Z\) is to the outcome \(Y\), after conditioning on
\(\{W_Y, Z, U\}\). Here, \(R_{Y \sim W_Z \mid W_Y, Z, U}\) is scaled by
\((1-R^2_{Z \sim W_Z \mid W_Y})^{-1}\), which represents how strong or
weak the underlying instrument is. Paradoxically, when
\(R^2_{Z \sim W_Z \mid W_Y}\) is high, this implies that the instrument
is stronger, which \emph{increases} the sensitivity to potential
violations of exclusion restriction. This term interacts with the
scaling term, Eq.~\ref{eq-iv-bias-r2}(a). While ostensibly
counterintuitive, consider the extreme setting in which
\(R^2_{Z \sim W_Z \mid W_Y} \approx 1\). In other words, the instrument
\(W_Z\) almost exactly covaries with the treatment \(Z\). Then the total
effect on \(Y\) is a combination of the treatment effect \(\tau\) and
the effect of the exclusion restriction. The IV estimator will assign
all of the total effect to \(\tau\), causing more bias.
\subparagraph{Violations in
Exogeneity.}\label{violations-in-exogeneity.}
The bias from violations of exogeneity, Eq.~\ref{eq-iv-bias-r2}(c), is
driven by the \(\hliv{R_{W_Z \sim U \mid W_Y}}\) term, directly maps to
the violation of Assumption \ref{a-iv-exog}, measuring the residual
variation in \(W_Z\) that can be explained by the omitted confounder
\(U\), after controlling for observed covariates. The bias is amplified
by \(\sooRsq\), parameters that correspond to violations in the
underlying SOO assumptions. We can informally think of the SOO
parameters as controlling the overall `difficulty' of the identification
problem. When \(U\) is a relatively weak confounder such that
\(R_{Y \sim U \mid Z, W_Z, W_Y}\) and \(R_{Z \sim U \mid W_Z, W_Y}\) are
close to zero, then there will be less exposure to bias for \emph{both}
\(\tau_\soo\) and \(\tau_\iv\). In this setting, \(\tau_\iv\) will be
more robust to violations of exogeneity. Conversely, more treatment or
outcome confounding directly increases the effect of violations of
exogeneity.
\subparagraph{Strength of the
Instrument.}\label{strength-of-the-instrument.}
The unobserved quantities are scaled by an observable scaling factor,
which depends primarily on the strength of the underlying instrumental
variable as measured by \(R_{Z\sim W_Z\mid W_Y}\). In settings when
there is a weak instrument, \(R_{Z\sim W_Z\mid W_Y}\) will be low, and
even in the presence of very weak violations in the underlying
assumptions, the overall bias incurred by an IV estimator will be
relatively large. This is on top of any inferential challenges, such as
high variance and non-normality, that arise from weak instruments even
when the IV assumptions hold exactly. \vspace{2mm}
Recently, \citet{cinelli2025iv} developed a sensitivity analysis for
instrumental variables in a linear model based on an inversion of an
Anderson--Rubin test \citep{anderson1949estimation}. Their approach
allows for a sensitivity analysis that incorporates sampling uncertainty
and applies to upper and lower confidence bounds, even in the presence
of weak instruments. This allows researchers to understand the minimum
amount of confounding that would change an IV estimate to no longer be
statistically significant. The resulting sensitivity analysis is
parameterized in terms of two partial correlations, one of which
involves an (unobservable) potential outcome. Here, in contrast, we
focus solely on the effect of assumption violations on the bias of the
IV estimate. This leads us to directly consider the 2SLS estimate and
parametrize the bias in terms of coefficients and partial correlations
involving the observed variables and \(U\) only, without potential
outcomes. Theorem~\ref{thm-iv} is a helpful complement to the
sensitivity analysis of \citet{cinelli2025iv} in understanding the
behavior of IV estimates.
\subsection{Bias of proximal estimate}\label{sec-bias-prox}
Theorem~\ref{thm-prox-bias} presents a decomposition of the bias of the
proximal estimate. Like the results in Section~\ref{sec-bias-iv}, the
bias of \(\tau_\prox\) will implicitly depend on not only the degree to
which the proximal assumptions are violated, but also the
selection-on-observables and IV assumptions.
\begin{theorem}[Bias of the proximal
estimator]\protect\hypertarget{thm-prox-bias}{}\label{thm-prox-bias}
The bias of \(\tau_\prox\) is: \begin{align}
\bias&(\hat \tau_\prox)= S_{W_Z, Z} \cdot \left\{ \beta_{w_z} - \beta_u \frac{\alpha_{w_z}}{\alpha_u}\right\} \frac{\sd(W_Z^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})} \nonumber \\
=& S_{W_Z, Z} \cdot \underbrace{\sqrt{\frac{1-R^2_{\hat W_Y \sim W_Z}}{1-R^2_{\hat W_Y \sim Z}} \cdot \frac{1}{1-R^2_{W_Z \sim U \mid W_Y, Z}}} \cdot \frac{\sd(Y)}{\sd(Z)}}_{\text{(a) Scaling Factor}} \nonumber \\
&~~\times \underbrace{\left\{ \hlprox{R_{Y \sim W_Z \mid W_Y, Z, U}} \cdot \sqrt{\left(1-R^2_{Y \sim U \mid W_Y, Z} \right) \frac{1-R^2_{Y \sim W_Y, Z}}{1-R^2_{W_Z \sim W_Y, Z}}} \right.}_{\text{(b) Violation in treatment proxy}} \label{eq-prox-bias}\\
&~~~~~~~~~~ - \hlprox{R_{W_Y \sim W_Z \mid U}} \cdot \hlsoo{R_{Y \sim U \mid W_Z, W_Y, Z}}\nonumber \\
&~~~~~~~~~~\underbrace{~~~~~~~\left. \times\sqrt{\frac{1-R^2_{Y \sim W_Y, W_Z, Z}}{(1-R^2_{Z \sim U \mid W_Y})\cdot (1-R^2_{W_Y \sim U})} \cdot \frac{1-R^2_{W_Y \sim W_Z}}{\hlprox{R^2_{W_Y \sim W_Z \mid U}} (1-R^2_{W_Y \sim U}) + R^2_{W_Y \sim U} - R^2_{W_Y \sim W_Z}}} ~\right\}}_{\text{(c) Violation in outcome proxy}}, \nonumber
\end{align} where
\(S_{W_Z, Z}\coloneq \cor(W_Z^{\perp \hat W_Y}, Z^{\perp \hat W_Y})\) is
either \(1\) or \(-1\).
\end{theorem}
The theorem consists of three drivers of bias when the proximal
assumptions are violated: (a) a scaling factor, (b) bias from an invalid
treatment proxy, and (c) bias from an invalid outcome proxy.
\subparagraph{Scaling Factor.}\label{scaling-factor.}
The scaling factor depends on how much variation in the constructed
\(\hat W_Y\) can be attributed to \(W_Z\) and \(Z\), how much inherent
variation there is in \(Y\) and \(Z\), and the degree to which the
treatment proxy is endogenous. The scaling factor is made up of mostly
observable components (i.e.,
\(R^2_{\hat W_Y \sim W_Z}, R^2_{\hat W_Y \sim Z}, \sd(Y)\) and
\(\sd(Z)\)), with the exception of a \(R^2_{W_Z \sim U \mid W_Y, Z}\)
term. We can theoretically bound this term as a function of the IV
sensitivity parameter \(\hliv{R^2_{W_Z \sim U \mid W_Y}}\), which
corresponds to violations of exogeneity in \(W_Z\):
\begin{equation}\phantomsection\label{eq-r2-bound}{
1 \leq \frac{1}{1-R^2_{W_Z \sim U \mid W_Y, Z}}
\leq
\frac{1}{1-\hliv{R^2_{W_Z \sim U \mid W_Y}}}.
}\end{equation} While the bound in Eq.~\ref{eq-r2-bound} still relies on
an unobserved term, it highlights a theoretical dependency between
proximal estimators and IV estimators. While proximal inference allows
researchers to relax assumptions of exogeneity in \(W_Z\), the
sensitivity of bias in the proximal estimator will be a function of the
degree of endogeneity of the treatment proxy \(W_Z\). If there is little
residual confounding in \(W_Z\), then the scaling factor will be close
to the lower bound of the range in Eq.~\ref{eq-r2-bound}. In contrast,
if there is a large degree of residual confounding, then the scaling
factor will be larger, implying that there is a greater degree of
sensitivity to potential violations in the underlying proximal
assumptions.
\subparagraph{Violations in the treatment proxy
assumption.}\label{violations-in-the-treatment-proxy-assumption.}
Invalid proxies contribute bias measured by Eq. \ref{eq-prox-bias}(b),
due mainly to the \(\hlprox{R_{Y \sim W_Z \mid W_Y, Z, U}}\) parameter.
This parameter is identical to the parameter used in the partial
correlation decomposition of the IV estimator, and represents the degree
to which the exclusion restriction has been violated. While the terms
that scale this parameter differ slightly from the IV
setting,\footnote{The difference is a result of algebraic manipulation used to factor out similar terms between the error that arises from violations in the treatment proxy and error that arises from violations in the outcome proxy, which are now absorbed in the scaling factor component.}
the interpretation of the partial correlation parameter is identical
across the two settings. This clarifies that while proximal estimators
allow researchers to relax the exogeneity assumption associated with an
instrumental variable, violations of the exclusion restriction impact
both approaches similarly.
\subparagraph{Violations in the outcome proxy
assumption.}\label{violations-in-the-outcome-proxy-assumption.}
The bias due to invalid outcome proxies, Eq. \ref{eq-prox-bias}(c),
depends on the parameter \(\hlprox{R_{W_Y \sim W_Z \mid U}}\), which
directly captures violations of Assumption \ref{a-prox-y} with regards
to a direct effect between \(W_Y\) and \(W_Z\); the strength of the
outcome proxy, represented by \(R_{W_Y \sim U}\); and the
selection-on-observables parameter \(R_{Y \sim U \mid W_Z, W_Y, Z}\).
Analogously to IV, violations from an invalid outcome proxy are
amplified by the SOO parameter \(R_{Y \sim U \mid W_Z, W_Y, Z}\): when
\(U\) strongly affects the outcome, there is a greater sensitivity to
bias from an invalid outcome proxy.
\section{Comparing the sensitivity of identification
approaches}\label{sec-sens-comp}
The bias results in the previous section, Theorem~\ref{thm-iv} and
Theorem~\ref{thm-prox-bias}, highlight how violations of the different
identifying assumptions interact and lead to bias in estimates of
\(\tau\). Reasoning about the relative bias from each identification
strategy is therefore challenging, as this requires researchers to
account the nested dependency across the identification strategies.
Across the three bias expressions, there are six total unobservable
coefficients and far more unobservable partial \(R^2\) parameters,
making it difficult to use the bias decompositions directly to compare
the sensitivities across identification strategies.
In the following section, we propose a suite of sensitivity tools that
allow researchers to compare and quantify the robustness of the
different identification strategies. We introduce a numerical summary
measure called the \emph{total robustness value} that quantifies the
minimum total violation of the identifying assumptions needed to result
in a substantively meaningful change in the result. The proposed measure
generalizes the robustness value introduced in
\citet{cinelli2020making}, and allows researchers to compare the
sensitivities across the different identification approaches by
re-expressing the bias using a population covariance matrix
representation, which collapses the bias expressions into a set of four
partial correlation terms, two of which are the SOO parameters. Using
the TRV, we also introduce a measure called the \emph{robustness
allocation}. The robustness allocation quantifies the relative
contribution of violations allowed in each assumption for a given
identification strategy to induce a specified amount of bias. Finally,
we propose a benchmarking approach, which allows researchers to reason
about the plausibility of potential violations in their underlying
identification assumptions.
\subsection{Re-parameterizing the bias using the population covariance
matrix}\label{re-parameterizing-the-bias-using-the-population-covariance-matrix}
Any approach to comparing identification strategies must require
parametrizing the estimates, the bias, and the identifying assumptions
on a common space. To start, because each estimator is simply the
coefficient from a linear regression, we can write each estimator in
terms of the population covariance matrix \(\Sigma \in \R^{7 \times 7}\)
of \((Y, Z, W_Z, W_Y, \hat Z, \hat W_Y, U)\): \[
\tau_\soo = \Sigma_{zw_zw_y,zw_zw_y}^{-1}\Sigma_{zw_zw_y,y}, \quad
\tau_\iv = \Sigma_{\hat zw_y,\hat zw_y}^{-1}\Sigma_{\hat zw_y,y}, \qand
\tau_\prox = \Sigma_{z\hat w_y,z\hat w_y}^{-1}\Sigma_{z\hat w_y,y},
\] where subscripts here indicate the submatrix of \(\Sigma\) that
corresponds to the variables in the subscript (i.e.,
\(\Sigma_{zw_zw_y,y}\) is the 2-by-1 submatrix containing the covariance
of \((Z, W_Z, W_Y)\) with \(Y\)). Similarly, the true effect \(\tau\)
can be expressed as
\(\tau = \Sigma_{zw_zw_yu,zw_zw_yu}^{-1}\Sigma_{zw_zw_yu,y}\).
While the different estimators do not depend on \(U\), \(\tau\) depends
on \(U\). Because the observed data fixes most of the matrix \(\Sigma\),
only the submatrix \(\Sigma_{u,yzw_zw_y}\) is unidentified.\footnote{Because
\(\hat Z\) and \(\hat W_y\) are deterministic functions of observed
variables only, the submatrix \(\Sigma_{u,\hat z\hat w_y}\) will
depend deterministically on the other observed entries, leaving only
the submatrix \(\Sigma_{u,yzw_zw_y}\) as being unidentified.} Thus,
another view of sensitivity analysis is to vary the entries in the
submatrix \(\Sigma_{u,yzw_zw_y}\) to evaluate how the resulting
estimates of \(\tau\) change.
The four entries in \(\Sigma_{u,yzw_zw_y}\) are not unconstrained---they
must be such that \(\Sigma\) is positive semidefinite. Without loss of
generality, we can set \(\Sigma_{uu}=1\), since the scale of the
confounder does not affect any observables. We then parametrize
\(\Sigma_{u,yzw_zw_y}\) by a set of partial correlations of \(U\) and
each observable variable: \[
\rho \coloneq (R_{U\sim W_Z}, R_{U\sim W_Y\mid W_Z}, \hlsoo{R_{U\sim Z\mid W_Y,W_Z}}, \hlsoo{R_{U\sim Y\mid Z,W_Z,W_Y}}).
\] Notice that the latter two partial correlations are exactly the SOO
sensitivity parameters. Critically, this choice of partial correlations
admits a bijective mapping to the Cholesky decomposition of \(\Sigma\)
and thus to \(\Sigma\) itself \citep{lewandowski2009generating}. As a
result, unlike the entries in \(\Sigma_{u,yzw_zw_y}\), these parameters
are variation independent: any values of \(\rho\in[-1, 1]^4\)
corresponds to a valid covariance matrix \(\Sigma(\rho)\).
The mapping \(\rho\mapsto\Sigma\) fixes the true effect \(\tau\). It
also determines the value of the various partial correlations that
measure how much each identifying assumption is violated. In other
words, we can re-express each of the partial correlation values that
correspond to the different violations in identifying assumptions as
functions of \(\rho\). For example, the violation of the exclusion
restriction corresponds to the value of
\(\hliv{R_{Y\sim W_Z\mid Z,W_Y,U}}\), and can be written as \[
\hliv{R_{Y\sim W_Z\mid Z,W_Y,U}}(\rho) = \frac{\Sigma(\rho)^{-1}_{y,w_z}}{
\sqrt{\Sigma(\rho)^{-1}_{y,y}\Sigma(\rho)^{-1}_{w_z,w_z}}}.
\] Thus even though the exclusion restriction violation
\(R_{Y\sim W_Z\mid Z,W_Y,U}\) does not appear anywhere in \(\rho\),
unlike the SOO sensitivity parameters, it is still related nonlinearly
to the SOO parameters and the other entries in \(\rho\), due to the
common parametrization.
\subsection{Total Robustness Value}\label{total-robustness-value}
The total robustness value depends on an analyst-chosen bias threshold
\(b\), representing the smallest change in a causal estimate so that is
considered meaningful.\footnote{So that, for example, a true value of
\(\tau=\tau_\iv-b\) is considered substantively different to the
causal estimate \(\tau_\iv\) actually obtained. A common choice could
be that the bias is equal to the estimate itself, and thus large
enough to reverse the sign of the causal effect. Alternatively, \(b\)
could be chosen to be equal to some multiple of the estimate's
standard error. For a Normal sampling distribution, a bias of one
standard error reduces the coverage of nominally 95\% confidence
intervals to around 80\%; a two-s.e. bias reduces coverage to around
50\%. This is similar to the notion of a `killer confounder', first
presented in \citet{huang2024sensitivity} and
\citet{hartman2024sensitivity}.} To begin, for a fixed \(\rho\), we
define the vector of parameters associated with each identification
strategy as follows: \[
\begin{aligned}
A_\soo(\rho) &= \qty( \hlsoo{R_{Z\sim U\mid W_Z,W_Y}(\rho)},~\hlsoo{R_{Y\sim U\mid Z,W_Z,W_Y}(\rho)} ) \\
A_\iv(\rho) &= \qty( \hliv{R_{W_Z\sim U\mid W_Y}(\rho)},~\hliv{R_{Y\sim W_Z\mid Z,W_Y,U}(\rho)} ) \qand \\
A_\prox(\rho) &= \qty( \hlprox{R_{Y\sim W_Z\mid Z,W_Y,U}(\rho)},~\hlprox{R_{W_Y\sim Z\mid W_Z,U}(\rho)},~ \hlprox{R_{W_Y\sim W_Z\mid U}(\rho)} ).
\end{aligned}
\] Because the SOO sensitivity parameters are part of \(\rho\) itself,
\(A_\soo\) is simply comprised of the last two entries in \(\rho\);
however, \(A_\iv\) and \(A_\prox\) are nonlinear functions of \(\rho\).
While the bias of both the IV and proximal estimators additionally
depend on other partial correlations, \(A_\iv(\rho)\) and
\(A_\prox(\rho)\) represent the partial correlations that directly map
to the underlying identification assumptions.
For each identification strategy, we define the \emph{total assumption
violation} as the squared norm \(||A(\rho)||_2^2\) of the corresponding
parameter vector \(A(\rho)\). We can then formally define the
\emph{total robustness value} as the minimum total violation of the
identifying assumptions that would result in absolute bias of \(b\):
\begin{equation}\phantomsection\label{eq-trv}{
\begin{aligned}
\TRV_\soo(b) = \inf \big\lvert\big\lvert A_\soo(\{\rho\,&:\, \bias(\tau_\soo(\rho))=b,
{R^2_{Z\sim U\mid W_Z,W_Y}(\rho)=R^2_{Y\sim U\mid Z,W_Z,W_Y}(\rho)} \}) \big\rvert\big\rvert^2_2, \\
\TRV_\iv(b)\, = \,\inf \big\lvert\big\lvert A_\iv(\{\rho\,&:\, \bias(\tau_\iv(\rho))=b,\,\,
{R^2_{Z\sim U\mid W_Z,W_Y}(\rho)=R^2_{Y\sim U\mid Z,W_Z,W_Y}(\rho)} \})\big\rvert\big\rvert^2_2, \\
\TRV_\prox(b) = \inf \big\lvert\big\lvert A_\prox(\{\rho\,&:\, \bias(\tau_\prox(\rho))=b,
{R^2_{Z\sim U\mid W_Z,W_Y}(\rho)=R^2_{Y\sim U\mid Z,W_Z,W_Y}(\rho)} \})\big\rvert\big\rvert^2_2.
\end{aligned}
}\end{equation} A \emph{smaller} total robustness value means that a
smaller violation of the identifying assumptions will be sufficient to
result in a threshold bias of \(b\). In contrast, a larger total
robustness value means that the total violation in the identifying
assumptions must be larger for there to obtain the same threshold bias.
In other words, there can be larger violations in the underlying
identifying assumptions before we expect there to be a substantively
meaningful change in the underlying result.
The TRV computes the minimum total violation in the identifying
assumptions assuming there is an equivalent amount of treatment
confounding and outcome confounding.\footnote{Notably, this is the same
condition imposed by \citet{cinelli2020making} in defining robustness
values. Here, it serves a critical purpose of avoiding pathological
\(\rho\), which would cause the total robustness values for IV and
proximal inference to shrink to zero. Specifically, without the
restriction of equal treatment and outcome confounding, the minimizing
value of \(\rho\) in Eq.~\ref{eq-trv} would occur near the point
\(R^2_{Z\sim U\mid W_Z,W_Y}(\rho)=1\) and
\(R^2_{Y\sim U\mid Z,W_Z,W_Y}(\rho)=0\)---extreme treatment
confounding and near-zero outcome confounding. In this region of
\(\rho\)-space, minuscule changes in any part of \(\rho\) lead to
extreme changes in the value of \(\tau\). This can be seen visually in
a traditional sensitivity contour plot where the contours of bias are
extremely dense around the point \((1, 0)\). Restricting \(\rho\) to
\(R^2_{Z\sim U\mid W_Z,W_Y}=R^2_{Y\sim U\mid Z,W_Z,W_Y}\) avoids this
pathology and aids interpretability by reducing the effective
dimension by one. Fortunately, we have found the relative positions of
the total robustness values---and even the values themselves---to be
nearly unchanged if \(R^2_{Z\sim U\mid W_Z,W_Y}\) and
\(R^2_{Y\sim U\mid Z,W_Z,W_Y}\) are constrained to have a ratio other
than one, as long as pathological values of \(\rho\) are avoided.} By
construction, \(\TRV_\soo(b)\) is exactly equal to 2 times the
robustness value defined by Cinelli and Hazlett, and can be viewed as a
generalization of the standard robustness value for alternative
identification strategies.
Because of the common parametrization, it is possible to compare the TRV
of different identification approaches. In an extreme case, If the TRV
for one identification strategy is very small, while it relatively large
for the others, then the identification approach may be much less robust
to violations of its assumptions, and researchers should more carefully
justify those assumptions. However, each TRV measures violations of
different assumptions, and so when TRV values are closer, direct
comparisons may be more difficult. We develop a benchmarking procedure
for the TRV based on observed covariates in
Section~\ref{sec-benchmarking} to aid in these comparisons.
To estimate the TRV in practice, researchers can use numerical
optimization with sample estimates of the covariance matrix. Our
parametrization \(\Sigma(\rho)\) is differentiable, and so researchers
can use reverse-mode automatic differentiation to calculate exact
gradients for the minimization problem in Eq.~\ref{eq-trv}. The
constraints are imposed by strongly penalizing any deviations.
Researchers can quantify estimation uncertainty by bootstrapping, as we
demonstrate below.
The TRV provides a helpful, single number summary of how much the
underlying assumptions can be violated to result in a substantively
meaningful change in our research conclusion, one that accounts for the
nested dependency structure across the different identification
assumptions. However, like the robustness value, it cannot be used to
determine which identification strategy has less (or more) bias
necessarily. However, a small TRV \emph{does} signify that even a small
violation in any underlying assumption will result in a substantial
amount of bias, and warrants potential caution in evaluating the
credibility of an estimated causal effect.
\subsection{Robustness Allocation}\label{robustness-allocation}
To further understand the sensitivity of each identification strategy,
we consider the partial \(R^2\) values that together yield the
worst-case scenario measured by the TRV. In general, there exists a
\emph{set} of possible partial \(R^2\) values that could result in the
amount of total assumption violation quantified by the TRV. We restrict
our attention to the value closest to the origin, \(\rho^*\), since this
corresponds to a minimal violation of the identifying assumptions,
\begin{equation}\phantomsection\label{eq-rho-star}{
\begin{aligned}
\rho^*_\iv := \argmin_\rho ||\rho||_2, \text{ s.t. } A_{\iv}(\rho) = \TRV_\iv(b) \\
\rho^*_\prox := \argmin_\rho ||\rho||_2, \text{ s.t. } A_{\prox}(\rho) = \TRV_\prox(b)
\end{aligned}
}\end{equation}
We can then calculate the values of each partial correlation at
\(\rho^*\). For example, in the IV setting, researchers can report
\(R^2_{W_Z\sim U \mid W_Y}(\rho^*)\) and
\(R^2_{Y\sim W_Z \mid Z, W_Y, U}(\rho^*)\), the minimum violations of
the exogeneity assumption and exclusion restriction, respectively, that
yield a bias of \(b\). If the former is much larger than the latter,
that means that for this worst-case scenario, the estimate is more
sensitive to violations of the exogeneity assumption than the exclusion
restriction.
To aid in these comparisons, we propose a normalized measure, the
\emph{robustness allocation} (RA), which evaluates the relative
contribution of violations in each assumption on the total sensitivity
of an underlying result. More formally, the RA of a given assumption is
the corresponding partial \(R^2\) evaluated at \(\rho^*\), normalized by
the total robustness value (i.e., \(A(\rho^*)/\TRV(b)\)). For example,
for instrumental variables, the robustness allocation values are defined
as: \[
\text{RA}_\text{iv-exog} = \frac{\hliv{R^2_{W_Z\sim U\mid W_Y}}(\rho^*_\iv)}{\TRV_\iv(b)}, \ \ \ \text{RA}_\text{iv-excl} := \frac{\hliv{R^2_{Y\sim W_Z\mid Z,W_Y,U}}(\rho^*_\iv)}{\TRV_\iv(b)}
\] The RA for the three proximal assumptions are defined similarly.
Since the robustness allocations sum to 1, if a given assumption's RA is
close to 1, then that assumption drives most of the worst-case
sensitivity of the causal estimate.
Robustness allocations should be interpreted with some caution, since
they measure the \emph{local} allocation of sensitivity to the
identifying assumptions. Other values of \(\rho\), and thus combinations
of assumption violations, could lead to an equivalent amount of bias
\(\rho\). The TRV and RA merely examine the minimal such violations that
lead to the chosen amount of bias.
\subsection{Benchmarking}\label{sec-benchmarking}
While the TRV and RA allow researchers to quantify and compare
sensitivities, they each measure different assumptions, and so comparing
them is not always straightforward. One approach to aid in these
comparisons is \emph{benchmarking}, which allows researchers to use the
observed covariates as hypothetical unobserved confounders, to calibrate
plausible values for the partial correlations parametrizing assumption
violations. This strategy cannot be directly applied to the partial
correlations in Section~\ref{sec-bias}, because the unobserved
confounder \(U\) is often part of the conditioning set.\footnote{ When
\(X\) and \(U\) are both in the conditioning set, treating a single
covariate \(X_j\) as the confounder \(U\) does not change the value of
the partial correlation at all.} The shared population covariance
matrix parameterization here, however, allows us to benchmark the
sensitivity parameters and the TRV for all the identification
strategies.
For the \(j\)-th covariate (where \(j \in \{1, ..., |X|\}\)), we define
\[
\hat \rho(j) = \qty( \hat R_{X^{(j)} \sim W_Z \mid X^{-(j)}}, \hat R_{X^{(j)} \sim W_Y \mid W_Z, X^{-(j)}}, \hat R_{X^{(j)} \sim Z \mid W_Y, W_Z, X^{-(j)}}, \hat R_{X^{(j)} \sim Y \mid Z, W_Z, W_Y, X^{-(j)}} ),
\] where \(X^{-(j)}\) is the set of covariates except the \(j\)-th. In
this way, \(\hat \rho(j)\) represents a set of partial correlations that
would correspond to an omitted confounder \(U\) with equivalent
confounding strength as an observed covariate \(X^{(j)}\). Then, because
each partial correlation term in the \(A_\iv\) and \(A_\prox\) is a
function of the elements of \(\rho\), we can directly estimate the
benchmarked parameters, as well as the corresponding total assumption
violation that would occur from omitting a confounder \(U\) with
equivalent confounding strength to \(X^{(j)}\) (denoted by
\(||A(\hat \rho(j))||^2_2\)). The benchmarked total assumption violation
can be compared to the TRV to evaluate how much stronger (or weaker) a
hypothetical \(U\) would have to be to result in a total assumption
violation that would saturate the minimum violation strength represented
by the TRV.
Benchmarking is most useful in settings when researchers have strong
substantive priors about the types of covariates that are prognostic of
different aspects of the data generating process. In general, we caution
that while benchmarking allows researchers to quantify the relative
strength of a confounder in terms of observed covariates, these types of
statements cannot be used to \emph{rule out} the existence of
confounding.
\subsection{Illustration on running example}\label{sec-trv-demo}
To see the proposed sensitivity summary measures in action, we return to
our running example from \citet{hk2022poland}. We calculate total
robustness values and the respective robustness allocation values for
the individual protest outcome under each identification strategy. For
the bias threshold \(b\), we use twice the standard error of the
respective estimates. This choice of \(b\) favors proximal and
especially IV, since their standard errors are larger. Coincidentally,
these values of \(b\) each correspond to roughly a one-quarter decrease
in the magnitude of each estimated causal effect. Table~\ref{tbl-trv}
shows these bias thresholds and their corresponding total robustness
values for each identification strategy. To account for uncertainty, we
use a fractional-weighted bootstrap and report the median absolute
deviation of the estimates across bootstrap iterations. The
fractional-weighted bootstrap, a kind of Bayesian bootstrap, avoids
collinearity issues when projecting out indicator variables that are
only present in a few observations \citep{xu2020applications}, and the
use of MAD is appropriate given the heavy skew of some of the bootstrap
distributions, which results from the quantities being bounded below by
0.
For both outcomes, selection-on-observables has the largest total
robustness value across the identification strategies; for the
individual protest outcome, \(\TRV_\soo=0.262\). This corresponds to a
robustness value of \(0.131\) under the definition of
\citet{cinelli2020making}. In other words, if the unobserved confounder
explains \(13.1\%\) of the variation in both treatment and outcome, for
a total violation of \(\norm{A_\soo}_2^2=0.262\), that would lead to
bias equal to \(b=0.102\). In comparison, the total robustness values
for IV and proximal inference are much smaller. For IV,
\(\TRV_\iv=0.0104\) means that a total violation of the exclusion
restriction plus exogeneity as small as \(0.0104\) on the \(R^2\) scale
is sufficient to create bias as large as \(b=0.216\). For proximal,
\(\TRV_\prox=1.43\times 10^{-5}\) leaves extremely little room for even
small violations of the proximal assumptions. This does not necessarily
mean that there is a greater amount of bias in \(\tau_\prox\). However,
it \emph{does} mean that \emph{any} deviation from the proximal
assumptions will result in a substantial amount of bias, and researchers
should be cautious when drawing causal conclusions from \(\tau_\prox\).
We can also decompose the calculated TRVs into robustness allocations
for each assumption, as shown in the right column of
Table~\ref{tbl-trv}. As is required by the constraints in
Eq.~\ref{eq-trv}, each SOO assumption contributes equally to the TRV.
For IV, the partial \(R^2\) corresponding to Assumption \ref{a-iv-exog}
is around three times larger than that for Assumption \ref{a-iv-excl},
suggesting that the IV estimate may be more sensitive to exclusion
restriction violations than instrument endogeneity. This might prompt a
researcher to more carefully consider mechanisms which might result in a
dependency between the number of corrupt Catholic priests and the
likelihood of protest.
\setstretch{1.0}
\begin{longtable}[]{@{}lrcl@{}}
\caption[]{Total robustness values and robustness allocations for the
reanalysis of \citet{hk2022poland}, individual protest outcome. The
median absolute deviations are shown in
parentheses.}\label{tbl-trv}\tabularnewline
\toprule\noalign{}
Strategy & \(b\) & \(\TRV(b)\) & Allocation to identifying
assumptions \\
\midrule\noalign{}
\endfirsthead
\toprule\noalign{}
Strategy & \(b\) & \(\TRV(b)\) & Allocation to identifying
assumptions \\
\midrule\noalign{}
\endhead
\bottomrule\noalign{}
\endlastfoot
\textbf{SOO} & \(0.102\) & \(0.262~~\textcolor{gray}{(0.0783)}\) &
\(\hlsoo{\text{RA}_\text{soo-z}}: 50.0\%~~\textcolor{gray}{(0\%)},\quad
\hlsoo{\text{RA}_\text{soo-y}}: 50.0\%~~\textcolor{gray}{(0\%)}\) \\
\textbf{IV} & \(0.216\) & \(0.0104~~\textcolor{gray}{(0.00783)}\) &
\(\hliv{\text{RA}_\text{iv-exog}}: 73.0\%~~\textcolor{gray}{(18.0\%)},\quad
\hliv{\text{RA}_\text{iv-excl}}: 27.0\%~~\textcolor{gray}{(18.0\%)}\) \\
\textbf{Proximal} & \(0.136\) &
\(1.43\times 10^{-5}~~\textcolor{gray}{(6.59\times 10^{-4})}\) &
\(\hlprox{\text{RA}_\text{prox-z}}: 0.0\%~~\textcolor{gray}{(1.2\%)},\quad
\hlprox{\text{RA}_\text{prox-y-$z$}}: 0.1\%~~\textcolor{gray}{(16.3\%)},\) \\
& & &
\(\hlprox{\text{RA}_\text{prox-y-$w_z$}}: 99.8\%~~\textcolor{gray}{(20.5\%)}\) \\
\end{longtable}
\setstretch{\papersp}
To help reason about whether it is plausible that the magnitude of
violations captured by the TRVs could occur, we perform the benchmarking
discussed above, estimate the total assumption violation that would
occur for an omitted \(U\) that is as strong as an observed covariate
\(X^{(j)}\). Table~\ref{tbl-bench} shows that across the different
covariates, the total assumption violations for IV and proximal are
consistently larger than the benchmarked total assumption violations for
selection-on-observables. More specifically, SOO's TRV is around 5 times
larger than the largest benchmarked total assumption violation, whereas
the TRVs for IV and proximal are are around ten to a thousand times
\emph{smaller} than the benchmarked total assumption violations.
Appendix~\ref{sec-app-sens} contains benchmarked values for each
identifying assumption, and a parallel analysis of the group protest
outcome, where we observe similar patterns. Notably, even though all
three estimators have nearly identical point estimates, both IV and
proximal have very small TRVs and so may be susceptible to bias in the
presence of even small violations of their underlying assumptions.
\setstretch{1.0}
\begin{longtable}[]{@{}lrrrr@{}}
\caption{\label{tbl-bench}Benchmarking values for individual protest
outcome. Shown are the change in \(\tau\) from leaving out each
covariate, and the benchmarked total assumption violations. Median
absolute deviations for each quantity are shown in parentheses.}
\tabularnewline
\toprule\noalign{}
Covariate & Change in \(\tau\) & \(\norm{A_\soo(\hat \rho(j))}_2^2\) &
\(\norm{A_\iv(\hat \rho(j))}_2^2\) &
\(\norm{A_\prox(\hat \rho(j))}_2^2\) \\
\midrule\noalign{}
\endhead
\bottomrule\noalign{}
\endlastfoot
Wealth & \(9.0\times 10^{-4}~~\textcolor{gray}{(0.005)}\) &
\(0.0020~~\textcolor{gray}{(0.004)}\) &
\(0.095~~\textcolor{gray}{(0.07)}\) &
\(0.090~~\textcolor{gray}{(0.07)}\) \\
Schools & \(-8.8\times 10^{-4}~~\textcolor{gray}{(0.001)}\) &
\(0.0022~~\textcolor{gray}{(0.004)}\) &
\(0.105~~\textcolor{gray}{(0.07)}\) &
\(0.093~~\textcolor{gray}{(0.07)}\) \\
Industrialization & \(-0.00123~~\textcolor{gray}{(0.002)}\) &
\(0.0025~~\textcolor{gray}{(0.003)}\) &
\(0.093~~\textcolor{gray}{(0.07)}\) &
\(0.091~~\textcolor{gray}{(0.08)}\) \\
Log population & \(-0.02884~~\textcolor{gray}{(0.046)}\) &
\(0.0465~~\textcolor{gray}{(0.022)}\) &
\(0.097~~\textcolor{gray}{(0.08)}\) &
\(0.097~~\textcolor{gray}{(0.08)}\) \\
Diversity & \(-0.00707~~\textcolor{gray}{(0.008)}\) &
\(0.0191~~\textcolor{gray}{(0.013)}\) &
\(0.093~~\textcolor{gray}{(0.07)}\) &
\(0.092~~\textcolor{gray}{(0.08)}\) \\
Katowickie region & \(-0.00672~~\textcolor{gray}{(0.025)}\) &
\(0.0396~~\textcolor{gray}{(0.019)}\) &
\(0.118~~\textcolor{gray}{(0.08)}\) &
\(0.106~~\textcolor{gray}{(0.08)}\) \\
Opolskie region & \(0.00421~~\textcolor{gray}{(0.004)}\) &
\(0.0111~~\textcolor{gray}{(0.008)}\) &
\(0.091~~\textcolor{gray}{(0.07)}\) &
\(0.091~~\textcolor{gray}{(0.07)}\) \\
\end{longtable}
\setstretch{\papersp}
\section{Implications of choosing an identification
strategy}\label{sec-unify}
The common parametrization from Section~\ref{sec-sens-comp} has deeper
implications for comparing estimators. In particular, when researchers
choose an identification strategy, they are implicitly expressing a
belief about the degree of violations that must be present in
alternative identification strategies. In this section, we show that all
three estimator's biases can be re-expressed in terms of the standard
SOO sensitivity parameters
\(\{\hlsoo{R_{Y \sim U \mid Z, W_Z, W_Y}}, \hlsoo{R_{Z \sim U \mid W_Z, W_Y}}\}\).
This means, for example, that if a researcher believes that the IV
identification assumptions hold, then a certain magnitude of unobserved
confounding must also be present. In some cases, this degree of
confounding may be implausible, or in conflict with a benchmarking
analysis under SOO. Researchers might then conclude that the IV
assumptions cannot hold exactly. The bias re-expression also allows us
to visualize the relative bias of each approach on a single bias contour
plot.
\subsection{Re-writing bias with respect to selection-on-observables
parameters}\label{re-writing-bias-with-respect-to-selection-on-observables-parameters}
Section~\ref{sec-sens-comp} parametrized the entire data-generating
process in terms of four partial correlations, and used this
parametrization to define and evaluate a set of numerical sensitivity
summary measures. A parametrization in terms of four values is a
significant reduction from the ten coefficients in Eq.~\ref{eq-struct}
and the many different partial correlations in Section~\ref{sec-bias}.
However, it turns out that for the purposes of evaluating bias in
\(\tau_\soo\), \(\tau_\iv\), and \(\tau_\prox\), we can reduce the
number of parameters even further, to just two.
Specifically, we can re-write the bias of \(\tau_\iv\) and
\(\tau_\prox\) with the selection-on-observables parameters
\(\hlsoo{R_{Y\sim U\mid Z,W_Z,W_Y}}\) and
\(\hlsoo{R_{Z\sim U\mid W_Z,W_Y}}\). The key observation is that for a
particular \(\sooRsq\), the bias of the selection-on-observables
estimate will be fixed, which implicitly also fixes the value of
\(\tau\), the true treatment effect. As such, we can re-express the bias
of \(\tau_\iv\) and \(\tau_\prox\) as
\begin{equation}\phantomsection\label{eq-bias-unif}{
\begin{aligned}
\bias(\tau_\iv) &= \tau_\iv - \underbrace{ \qty{ \tau_\soo -
\bias(\tau_\soo; \hlsoo{R_{Y\sim U\mid Z,W_Z,W_Y}}, \hlsoo{R_{Z\sim U\mid W_Z,W_Y}}) }}_{\coloneq \tau} \\
\bias(\tau_\prox) &= \tau_\prox - \qty{ \tau_\soo -
\bias(\tau_\soo; \hlsoo{R_{Y\sim U\mid Z,W_Z,W_Y}}, \hlsoo{R_{Z\sim U\mid W_Z,W_Y}}) }
\end{aligned}
}\end{equation} Both of these expressions depend only on the observed
data and the value \(\biasSOO\), which is fixed by the parameters
\(\sooRsq\). Note that the partial correlations corresponding to other
assumptions, like the exclusion restriction
\(R_{Y\sim W_Z\mid Z,W_Y,U}\) \emph{do} depend on other parts of
\(\rho\); it is only the bias of each estimate that can be reduced to
exactly the SOO parameters.
This unification is advantageous for several reasons. First,
Eq.~\ref{eq-bias-unif} gives us a route to translate assumptions made
for one identification strategy into the language of treatment and
outcome confounding. Suppose a researcher believes that their IV
strategy is highly plausible, perhaps in part due to the IV-specific
sensitivity analyses developed in this paper, and furthermore they are
able to estimate the likely direction of IV bias. As a result, they
might believe that most likely \(0<\bias(\tau_\iv)<0.1\).
Eq.~\ref{eq-bias-unif} then connects this belief to an equivalent belief
about likely values of \(\sooRsq\). The researcher's judgement of the
plausibility of these values can further inform their understanding of
the IV bias, or even their choice of identification strategy.
Second, Eq.~\ref{eq-bias-unif} can be directly solved for the set of SOO
sensitivity parameters where each identification strategy has lower bias
than the others. Because Eq.~\ref{eq-bias-unif} is linear in all bias
terms, each estimator will dominate the others on a particular interval
of values of \(\tau\), corresponding in turn to a particular interval of
values of \(\biasSOO\). These intervals are exactly those obtained by
cutting the real line at the midpoints between the values of
\(\tau_\soo\), \(\tau_\iv\), and \(\tau_\prox\). For example, if
\(\tau_\soo=0.1\), \(\tau_\iv=0.4\), and \(\tau_\prox=0.2\), then SOO
dominates for the \(\sooRsq\) corresponding to any true \(\tau<0.15\),
proximal inference dominates sensitivity parameters yielding
\(0.15<\tau<0.3\), and IV dominates for parameters giving \(\tau>0.3\).
We can map these intervals to the traditional sensitivity plot,
discussed in the next section, as contour lines that bound regions where
each estimator dominates. These regions can be visualized together,
providing insight into the assumptions and sensitivity of each
identification approach.
Readers will note that the manipulations in Eq.~\ref{eq-bias-unif} are
in no way reliant on using the SOO sensitivity parameters specifically:
the same re-expression could be applied in terms of the partial
correlations for IV or proximal bias. However, as is clear from Theorems
\ref{thm-iv} and \ref{thm-prox-bias}, those bias expressions are
substantially more complicated, and rely on unobservable partial
correlations with no direct connection to the IV and proximal
assumptions. Moreover, the subtraction that is present in those bias
expressions means that bias is not monotonic in either sensitivity
parameter the way that it is for SOO, which further complicates
interpretation and visualization. Finally, treatment and outcome
confounding are in many ways the core challenges of causal inference,
and researchers are familiar with considering confounding mechanisms in
these terms. Thus we focus specifically on the SOO parametrization, as
we anticipate it will be easiest to reason about for applied
researchers.
\subsection{Augmented bias contour
plots}\label{augmented-bias-contour-plots}
The unified bias framework also allows us to extend the standard SOO
sensitivity plot and allow researchers to visually compare the relative
biases across the identification strategies. Traditionally, bias contour
plots have been employed to evaluate SOO bias by varying the two
sensitivity parameters \(\sooRsq\). A contour line corresponding to
\(\tau=0\) is often plotted to help identify which values of sensitivity
parameters would change the sign of the estimate. To help provide a
plausible range of the sensitivity parameters, the existing covariates
can be used as a benchmark. Each is left out in turn, and its partial
correlation with the outcome and treatment is plotted as if that
covariate were the unobserved confounder.
\begin{figure}
\centering{
\includegraphics[width=6.25in,height=\textheight,keepaspectratio]{figures/pol_sabotage_bias_contour.pdf}
}
\caption{\label{fig-contour}Sensitivity plot for the running example.
The contour lines are on a logarithmic scale and show the amount of SOO
confounding bias. Contours are colored according to the regions where
each estimator has lowest bias. Gray crosses indicate benchmark
sensitivity parameter values for each of the covariates.}
\end{figure}%
The augmented sensitivity plot we propose builds on this existing
visualization in three ways. First, because the sign of the bias, and
thus the dominating estimator, depends on the sign of the sensitivity
parameters, we generate the sensitivity plot using the signed parameters
rather than their square. Second, as described in the preceding section,
we can separate the contour lines according to the regions where each
identification strategy dominates, indicating this partition by lines
and colors. Third, we can plot a contour line that corresponds to zero
bias in either the proximal or IV estimate. If the assumptions of either
of those strategies hold exactly, then the SOO sensitivity parameters
\emph{must} lie along the corresponding contour.
While our primary focus throughout the paper has been on the bias of
different identification strategies, researchers may instead be
interested in the relative mean-squared error of the different
approaches. Notably, the same re-expression in Eq.~\ref{eq-bias-unif}
can be applied for mean-squared error, and a similar augmented
sensitivity plot can be produced. In some cases this may show that the
zero-bias contour for proximal or IV is actually contained in the
lowest-MSE region for another estimator. In other words, that \emph{even
if} the IV or proximal assumptions hold exactly, another estimator may
be preferred due to lower mean-squared error. This phenomenon has been
observed for IV estimates when the instrument is weak \citep[see,
e.g.,][]{bound1995problems, hahn2005estimation}. See
Appendix~\ref{sec-app-mse} for more discussion.
\subsection{Illustration on running
example}\label{illustration-on-running-example}
We now return to our running example, where we generate the proposed
augmented sensitivity plot, including the benchmarked values of
\(\sooRsq\) based on the observed covariates (see
Figure~\ref{fig-contour}) Even without applying the total robustness
values introduced above, the additional annotations are immediately
helpful to practitioners. If an analyst had originally designed their
observational study around IV, for instance, then their beliefs about
the values of the SOO sensitivity parameters must lie close to the curve
labeled ``No IV Bias'' in Figure~\ref{fig-contour}. In other words,
\emph{belief} about an identification strategy implies a relatively
narrow set of beliefs on the values of the sensitivity parameters.
Practitioners can compare that set of parameter values to the benchmark
values of the covariates, which are plotted as gray crosses in
Figure~\ref{fig-contour}. In this case, the benchmark values lie
squarely in the region labeled ``SOO least bias.'' The largest
benchmarked \(R_{Z\sim U\mid W_Z,W_Y}\) is \(0.0378\) and corresponds to
(log) municipality population. In contrast, the \emph{smallest} possible
value of \(R_{Z\sim U\mid W_Z,W_Y}\) consistent with a valid proximal
strategy is \(0.413\)--around \(11\) times larger. For IV, the factor
grows to \(15\). Both of those factors assume worst-case outcome
confounding, with \(R_{Y\sim U\mid Z,W_Z,W_Y}=1\), which is also several
times larger than the confounding associated with any of the observed
covariates.
Thus if a researcher is to proceed with a proximal or IV strategy, they
must believe that the unobserved confounder is significantly stronger
than any of the observed covariates' benchmark values, and that the
treatment and outcome confounding are related in a way as to fall close
to one of the no-bias contour lines in Figure~\ref{fig-contour}. One
immediate qualitative implication is that treatment and outcome
confounding must have the same sign for a proximal or IV approach to
produce less bias. If they do not, then the sensitivity parameters lie
in quadrants II or IV of Figure~\ref{fig-contour}, where the SOO
estimate always dominates.
Figure~\ref{fig-contour} additionally provides insight on the total
robustness values in Section~\ref{sec-sens-comp}. Near the origin, where
the SOO assumptions approximately hold, the contour lines are sparsely
spaced (note that the plotted contours in Figure~\ref{fig-contour} are
on a logarithmic scale). Thus, even moderate violations of the SOO
assumptions lead to small changes in the value of \(\tau\), and thus
small amounts of bias. In contrast, the ``No Proximal Bias'' and ``No IV
Bias'' lines in Figure~\ref{fig-contour} are much farther from the
origin, where the contour lines are denser. Smaller changes in the
sensitivity parameters in these regions lead to correspondingly larger
changes in bias, which in turn lowers the total robustness values.
From Eq.~\ref{eq-bias-unif}, the distance of the ``No Proximal Bias''
and ``No IV Bias'' lines from the origin is directly related to the
difference between the point estimates for these identification
strategies and the SOO estimate. The more the estimates differ, the more
confounding is necessarily implied under each set of identification
assumptions. As both Theorem~\ref{thm-iv} and
Theorem~\ref{thm-prox-bias} show, the bias of the IV and proximal
estimates grows with the amount of outcome confounding. Thus a larger
observed difference between the SOO estimate and IV or proximal
estimates implies that the outcome confounding is larger, which in turn
will generally translate to lower robustness.
\section{Conclusion}\label{sec-conclude}
This paper proposes a unified sensitivity framework for researchers to
consider the relative bias of different identification strategies. While
the unified model imposes some constraints compared to a fully
nonparametric setup, the relative simplicity also clarifies the key
drivers of bias under IV and proximal identification, and highlights
connections between each estimator and their biases. As
Section~\ref{sec-sens-comp} and Section~\ref{sec-unify} show, the
dimensionality of unidentified parts of the causal model is limited, and
this creates (usually nonlinear) relationships between the various
causal assumptions. Thus, by comparing benchmarking values and point
estimates under each strategy, more can be said about each of the
identifying assumptions. The bias expressions in Section~\ref{sec-bias}
and the proposed total robustness values also help researchers
understand the sensitivity of each approach to violations of these
assumptions.
There are several interesting avenues of future work. One natural
direction would be to explore the extent to which these findings hold in
a fully nonparametric setting. As \citet{chernozhukov2024long} show for
SOO, key intuition and structure from the linear setting can carry over
with little modification to a nonparametric model. The two-stage
estimation of IV and proximal, however, complicates such an analysis and
would warrant careful investigation.
Second, recent work in observational causal inference has introduced an
idea known as design sensitivity for observational studies, which allows
researchers to consider sensitivity to potential violations in the
underlying identification assumptions \emph{a priori}
\citep[e.g.,][]{rosenbaum2004design, rosenbaum2010design, huang2025design}.
Design sensitivity quantifies the tradeoffs in robustness that arise
from different `design' choices (i.e., choosing an estimand, treatment
definitions, study populations). However, much of this work focuses
explicitly on the SOO setting. Future work could build on the unified
sensitivity framework presented in this paper to consider design
sensitivity across different identification strategies.
\hypertarget{refs}{}
\begin{CSLReferences}{0}{0}\end{CSLReferences}
\clearpage
\appendix
\appendix
\renewcommand\thefigure{\thesection\arabic{figure}}
\setcounter{figure}{0}
\section{Additional Discussion}\label{sec-app-discussion}
\subsection{Covariates and identifying assumption plausibility in the
running example}\label{sec-app-covar}
\citet{hk2022poland} use the following covariates in their analysis:
\begin{itemize}
\tightlist
\item
indicators for the three subregions of Upper Silesia;
\item
a wealth index calculated as an average of the number of shops,
restaurants, and cinemas in each municipality;
\item
the number of schools in each municipality, which the authors argue is
a measure of state capacity;
\item
an industrialization index built as an average of the presence of
other mineral deposits and size of coal deposits, which the authors
note is also correlated with poor working conditions and more
grievances against the government;
\item
an ethnic diversity index measured by the number of migrants into
Upper Silesia after World War II;
\item
the population of each municipality;\footnote{Unlike the original
authors, we took the logarithm of this variable and also used the
square root of population to normalize the counts of different
buildings and establishments. In general, we find the original
analysis highly sensitive to treatment of municipality populations.
We do not further these concerns in the present analysis, however,
since they are not directly related to the causal assumptions under
study.} and
\item
an indicator for which municipalities were formerly under Russian
occupation.
\end{itemize}
While these covariates capture several plausible causes of both
surveillance and antiregime sentiment, they are far from perfect. Many
of the covariates were justified by Hager and Krakowski as proxies for
broader underlying causes, such as ``state capacity;'' if the error in
the observed proxies (e.g., the number of schools) is at all correlated
with treatment and outcome, then the SOO assumptions would be violated.
Moreover, it is easy to imagine that municipalities that experienced
above-expected antiregime activity in some years---that is, more than
average given their covariates---might then be subjected to increased
surveillance. Since the variables here are aggregated over multiple
years, some of these dynamic effects could also lead to violations of
the SOO assumptions. We stress that the original authors conducted
extensive investigations into all of their identifying assumptions, and
collected additional data to test certain assumptions, and we lack the
space here to fully summarize their justifications. However, the threats
to causal inference under SOO were plausible enough that the authors
used an instrumental variable strategy for their primary analysis.
For the instrumental variable analysis, the authors point to several
facts to justify the exogeneity assumption. First, new priests were sent
to parishes upon the retirement or death of previous priests, which are
plausibly exogenous. Second, the Catholic Church excommunicated priests
found to have been corrupted, making it less likely that a priest's
corruptibility increased their chances of assignment. To justify the
exclusion restriction, Hager and Krakowski note the inherent secrecy of
the corrupted priests---the citizens protesting would not have had
knowledge of which priests were corrupted, so a backlash effect is less
likely. They also point to the roughly three-decade lag between the
collection of their corruption data and the protests. Finally, they
conduct several regression analyses on auxiliary data to argue against
the presence of several disallowed causal pathways.
However, it is not difficult to imagine plausible mechanisms that would
violate either assumption. Suppose the Catholic Church reserved
assignments to important cities, such as those with a history of protest
or surveillance, for priests that they judged to be more resistant to
recruitment. Alternatively, suppose the SB stepped up priest recruiting
efforts in cities they deemed more likely to protest or resist. Even
municipality population could be a confounder, since the count of
corrupted priests was not normalized by population, and population
enters the estimation equations linearly only.
For the exclusion restriction, the very lag between the collection of
the corruption data and the protests also allowed time for priests'
corruption to be exposed, which could affect protest in either direction
(instilling fear or provoking backlash). Hager and Krakowski also raise
the possibility of indoctrination of parishioners by the corrupted
priests.
\subsection{IV bias under exogenous instruments}\label{sec-app-exog}
In some settings, researchers are able to exploit specific features of a
given study to ensure the instrument is exogenous by design (i.e.,
lotteries, random assignment to judges). When there is \emph{no}
violation in exogeneity of the instrument, we can simplify the bias
expression for \(\tau_{\iv}\) to only depend on the instrument strength
and violations in the exclusion restriction.
\begin{corollary}[Relative bias with exogenous
instruments]\protect\hypertarget{cor-relative-bias-iv}{}\label{cor-relative-bias-iv}
In settings where the instrument is exogenous, the bias of an IV
estimator simplifies to: \begin{align*}
\underbrace{\frac{1}{R_{Z \sim W_Z \mid W_Y}} \cdot \frac{\sd(Y^{\bot W_Y, Z})}{\sd(Z^{\bot W_Y})}}_{\text{(a) Instrument Strength}} \times \underbrace{ R_{Y \sim W_Z \mid W_Y, Z, U} \cdot \sqrt{\left( 1-R^2_{Y \sim U \mid W_Y, Z}\right) \cdot \frac{1}{1-R^2_{Z \sim W_Z \mid W_Y}}}}_{\text{(b) Violations in Exclusion Restriction}}.
\end{align*} The relative bias between an IV estimator and an SOO
estimator will then depend on the degree to which the exclusion
restriction is violated. In general, when the following holds:
\[R^2_{Y \sim W_Z \mid W_Y, Z, U} \leq \frac{R^2_{Y \sim U \mid W_Y, W_Z, Z}}{1- \gamma \cdot R^2_{Y \sim U \mid W_Y,W_Z, Z}} \cdot \frac{R^2_{Z \sim U \mid W_Y, W_Z}}{1-R^2_{Z \sim U \mid W_Y, W_Z}} \cdot (1-R^2_{Y \sim W_Z \mid W_Y, Z}) \cdot R^2_{Z \sim W_Z \mid W_Y},\]
then \(\tau_{\iv}\) will be less biased than \(\tau_\soo\), where
\(\gamma\) represents how different \(R^2_{Y \sim U \mid W_Y, W_Z, Z}\)
is to \(R^2_{Y \sim U \mid W_Y, Z}\).
\end{corollary}
For a fixed \(\gamma\) value, we can vary the selection-on-observables
parameters \(\sooRsq\) to find the maximum
\(R^2_{Y \sim W_Z \mid W_Y, Z, U}\) value that would still result in a
less biased estimator for \(\tau_\iv\) over \(\tau_\soo\). There exists
a set of \(\sooRsq\) that correspond to an upper bound on
\(R^2_{Y \sim W_Z \mid W_Y, Z, U}\) greater than 1. In that setting,
\(\tau_{\iv}\) will be less biased than \(\tau_\soo\). The bound on
\(R^2_{Y \sim W_Z \mid W_Y, Z, U}\) will increase as
\(R^2_{Y \sim U \mid W_Y, W_Z, Z}\) and \(R^2_{Z \sim U \mid W_Y, W_Z}\)
increase. Intuitively, this is because as \(\sooRsq\) increases in
magnitude, this implies that there is a greater amount of bias in
\(\tau_\soo\) from a stronger omitted confounder. As such, even under
relatively large violations of the exclusion restriction, \(\tau_{\iv}\)
may result in less biased estimates.
Notably, Corollary~\ref{cor-relative-bias-iv} provides a simple way to
compare the relative bias between \(\tau_\soo\) and \(\tau_\iv\).
However, this requires a relatively strong assumption that the
instrument used is exogenous. In settings when the instrumental variable
is not exogenous by design, researchers must simultaneously reason about
both violations in the exclusion restriction and exogeneity. This is
less straightforward, because as Theorem~\ref{thm-iv} highlights, the
bias from violations of exogeneity will interact with the
selection-on-observables bias.
We now turn to comparing IV with proximal in this special setting. We
now additionally assume that the outcome proxy is in fact valid, such
that there is no dependency between \(W_Y\) and the instrument \(W_Z\).
In this setting, the bias between the IV and proximal estimator depends
only on the exclusion restriction of the instrument \(W_Z\).
\begin{corollary}[]\protect\hypertarget{cor-relative-bias-prox}{}\label{cor-relative-bias-prox}
Assume the IV is exogenous and the outcome proxy is valid. Then, when
the following holds:
\begin{equation}\phantomsection\label{eq-rel-bias-prox}{
S_{W_Z, Z} \cdot \frac{\sd(W_Z^{\bot \hat W_Y})}{\sd(W_Z^{\bot W_Y})} \cdot \frac{\sd(Z^{\bot W_Y})}{\sd(Z^{\bot \hat W_Y})} \cdot \cor(W_Z^{\bot W_Y}, Z^{\bot W_Y}) \leq 1,
}\end{equation} the proximal estimator will incur less bias from an
exclusion restriction violation.
\end{corollary}
Eq.~\ref{eq-rel-bias-prox} comprises completely observable components.
This means in practice, researchers can directly calculate the sample
analog of each component to evaluate whether the ratio is greater than
1. When it is greater than 1, this implies that an IV estimator will
incur less bias under violations of the exclusion restriction. In
contrast, when it is less than 1, then this implies that a proximal
estimator will incur less bias under violations of the exclusion
restriction.
\section{Additional sensitivity analysis and
benchmarking}\label{sec-app-sens}
\subsection{Detailed benchmarking values for individual protest
outcome}\label{detailed-benchmarking-values-for-individual-protest-outcome}
\footnotesize
\begin{longtable}[]{@{}lrrrrrr@{}}
\caption{\label{tbl-bench-ext}Additional benchmarking values for
individual protest outcome. Shown are the partial correlations measuring
each identifying assumption's violation and the benchmarked total
assumption violations. Median absolute deviations for each quantity are
shown in parentheses.}
\tabularnewline
\toprule\noalign{}
Covariate & \(\hlsoo{R_{Z\sim U\mid W_Z,W_Y}}\) &
\(\hlsoo{R_{Y\sim U\mid Z,W_Z,W_Y}}\) & \(\hliv{R_{W_Z\sim U\mid W_Y}}\)
& \(\hliv{R_{Y\sim W_Z\mid Z,W_Y,U}}\) &
\(\hlprox{R_{W_Y\sim Z\mid W_Z,U}}\) &
\(\hlprox{R_{W_Y\sim W_Z\mid U}}\) \\
\midrule\noalign{}
\endhead
\bottomrule\noalign{}
\endlastfoot
Wealth & \(0.0075~~\textcolor{gray}{(0.05)}\) &
\(0.044~~\textcolor{gray}{(0.04)}\) &
\(-0.0746~~\textcolor{gray}{(0.03)}\) &
\(-0.30~~\textcolor{gray}{(0.1)}\) &
\(-0.0172~~\textcolor{gray}{(0.05)}\) &
\(0.01109~~\textcolor{gray}{(0.07)}\) \\
Schools & \(-0.0185~~\textcolor{gray}{(0.02)}\) &
\(0.043~~\textcolor{gray}{(0.04)}\) &
\(0.1084~~\textcolor{gray}{(0.05)}\) &
\(-0.30~~\textcolor{gray}{(0.1)}\) &
\(-0.0157~~\textcolor{gray}{(0.05)}\) &
\(-5.5\times 10^{-4}~~\textcolor{gray}{(0.07)}\) \\
Industrialization & \(-0.0413~~\textcolor{gray}{(0.03)}\) &
\(0.028~~\textcolor{gray}{(0.05)}\) &
\(-0.0452~~\textcolor{gray}{(0.06)}\) &
\(-0.30~~\textcolor{gray}{(0.1)}\) &
\(-0.0160~~\textcolor{gray}{(0.05)}\) &
\(0.00513~~\textcolor{gray}{(0.07)}\) \\
Log population & \(0.0378~~\textcolor{gray}{(0.05)}\) &
\(-0.212~~\textcolor{gray}{(0.05)}\) &
\(0.0041~~\textcolor{gray}{(0.02)}\) &
\(-0.31~~\textcolor{gray}{(0.1)}\) &
\(-0.0180~~\textcolor{gray}{(0.05)}\) &
\(0.00439~~\textcolor{gray}{(0.07)}\) \\
Diversity & \(-0.1257~~\textcolor{gray}{(0.05)}\) &
\(0.057~~\textcolor{gray}{(0.05)}\) &
\(-0.0326~~\textcolor{gray}{(0.04)}\) &
\(-0.30~~\textcolor{gray}{(0.1)}\) &
\(-0.0049~~\textcolor{gray}{(0.05)}\) &
\(0.00754~~\textcolor{gray}{(0.07)}\) \\
Katowickie region & \(-0.0130~~\textcolor{gray}{(0.04)}\) &
\(0.199~~\textcolor{gray}{(0.05)}\) &
\(0.1135~~\textcolor{gray}{(0.03)}\) &
\(-0.32~~\textcolor{gray}{(0.1)}\) &
\(-0.0158~~\textcolor{gray}{(0.05)}\) &
\(-0.00164~~\textcolor{gray}{(0.07)}\) \\
Opolskie region & \(0.0303~~\textcolor{gray}{(0.03)}\) &
\(0.101~~\textcolor{gray}{(0.04)}\) &
\(-0.0205~~\textcolor{gray}{(0.02)}\) &
\(-0.30~~\textcolor{gray}{(0.1)}\) &
\(-0.0155~~\textcolor{gray}{(0.05)}\) &
\(0.00379~~\textcolor{gray}{(0.07)}\) \\
\end{longtable}
\normalsize
\subsection{Sensitivity analysis and benchmarking for group protest
outcome}\label{sensitivity-analysis-and-benchmarking-for-group-protest-outcome}
Table~\ref{tbl-trv-grp}, Table~\ref{tbl-bench-grp}, and
Table~\ref{tbl-bench-grp-ext} show the total robustness value and
benchmarking results for the group protest outcome, paralleling the
results presented in Section~\ref{sec-trv-demo} and the preceding
section for the individual protest outcome.
\setstretch{1.0}
\begin{longtable}[]{@{}lrrl@{}}
\caption[]{Total robustness values and robustness allocations for the
reanalysis of \citet{hk2022poland}, group protest outcome. Median
absolute deviations are shown in
parentheses.}\label{tbl-trv-grp}\tabularnewline
\toprule\noalign{}
Strategy & \(b\) & \(\TRV(b)\) & Allocation to identifying
assumptions \\
\midrule\noalign{}
\endfirsthead
\toprule\noalign{}
Strategy & \(b\) & \(\TRV(b)\) & Allocation to identifying
assumptions \\
\midrule\noalign{}
\endhead
\bottomrule\noalign{}
\endlastfoot
\textbf{SOO} & \(0.117\) & \(0.256~~\textcolor{gray}{(0.0609)}\) &
\(\hlsoo{\text{RA}_\text{soo-z}}: 50.0\%~~\textcolor{gray}{(0\%)},\quad
\hlsoo{\text{RA}_\text{soo-y}}: 50.0\%~~\textcolor{gray}{(0\%)}\) \\
\textbf{IV} & \(0.256\) & \(0.0166~~\textcolor{gray}{(0.0115)}\) &
\(\hliv{\text{RA}_\text{iv-exog}}: 24.8\%~~\textcolor{gray}{(9.9\%)},\quad
\hliv{\text{RA}_\text{iv-excl}}: 75.2\%~~\textcolor{gray}{(9.9\%)}\) \\
\textbf{Proximal} & \(0.747\) &
\(9.36\times 10^{-6}~~\textcolor{gray}{(0.0037)}\) &
\(\hlprox{\text{RA}_\text{prox-z}}: 0.0\%~~\textcolor{gray}{(0.0\%)},\quad
\hlprox{\text{RA}_\text{prox-y-$z$}}: 8.3\%~~\textcolor{gray}{(4.2\%)},\) \\
& & &
\(\hlprox{\text{RA}_\text{prox-y-$w_z$}}: 91.7\%~~\textcolor{gray}{(4.1\%)}\) \\
\end{longtable}
\begin{longtable}[]{@{}lrrrr@{}}
\caption{\label{tbl-bench-grp}Benchmarking values for group protest
outcome. Shown are the benchmarked SOO parameters and the benchmarked
total assumption violations and the benchmarked total assumption
violations. Median absolute deviations for each quantity are shown in
parentheses.}
\tabularnewline
\toprule\noalign{}
Covariate & Change in \(\tau\) & \(\norm{A_\soo(\hat \rho(j))}_2^2\) &
\(\norm{A_\iv(\hat \rho(j))}_2^2\) &
\(\norm{A_\prox(\hat \rho(j))}_2^2\) \\
\midrule\noalign{}
\endhead
\bottomrule\noalign{}
\endlastfoot
Wealth & \(9.0\times 10^{-4}~~\textcolor{gray}{(0.005)}\) &
\(0.0020~~\textcolor{gray}{(0.004)}\) &
\(0.095~~\textcolor{gray}{(0.07)}\) &
\(0.090~~\textcolor{gray}{(0.07)}\) \\
Schools & \(-8.8\times 10^{-4}~~\textcolor{gray}{(0.001)}\) &
\(0.0022~~\textcolor{gray}{(0.004)}\) &
\(0.105~~\textcolor{gray}{(0.07)}\) &
\(0.093~~\textcolor{gray}{(0.07)}\) \\
Industrialization & \(-0.00123~~\textcolor{gray}{(0.002)}\) &
\(0.0025~~\textcolor{gray}{(0.003)}\) &
\(0.093~~\textcolor{gray}{(0.07)}\) &
\(0.091~~\textcolor{gray}{(0.08)}\) \\
Log population & \(-0.02884~~\textcolor{gray}{(0.046)}\) &
\(0.0465~~\textcolor{gray}{(0.022)}\) &
\(0.097~~\textcolor{gray}{(0.08)}\) &
\(0.097~~\textcolor{gray}{(0.08)}\) \\
Diversity & \(-0.00707~~\textcolor{gray}{(0.008)}\) &
\(0.0191~~\textcolor{gray}{(0.013)}\) &
\(0.093~~\textcolor{gray}{(0.07)}\) &
\(0.092~~\textcolor{gray}{(0.08)}\) \\
Katowickie region & \(-0.00672~~\textcolor{gray}{(0.025)}\) &
\(0.0396~~\textcolor{gray}{(0.019)}\) &
\(0.118~~\textcolor{gray}{(0.08)}\) &
\(0.106~~\textcolor{gray}{(0.08)}\) \\
Opolskie region & \(0.00421~~\textcolor{gray}{(0.004)}\) &
\(0.0111~~\textcolor{gray}{(0.008)}\) &
\(0.091~~\textcolor{gray}{(0.07)}\) &
\(0.091~~\textcolor{gray}{(0.07)}\) \\
\end{longtable}
\footnotesize
\begin{longtable}[]{@{}lrrrrrr@{}}
\caption{\label{tbl-bench-grp-ext}Additional benchmarking values for
group protest outcome. Shown are the partial correlations measuring each
identifying assumption's violation and the benchmarked total assumption
violations. Median absolute deviations for each quantity are shown in
parentheses.}
\tabularnewline
\toprule\noalign{}
Covariate & \(\hlsoo{R_{Z\sim U\mid W_Z,W_Y}}\) &
\(\hlsoo{R_{Y\sim U\mid Z,W_Z,W_Y}}\) & \(\hliv{R_{W_Z\sim U\mid W_Y}}\)
& \(\hliv{R_{Y\sim W_Z\mid Z,W_Y,U}}\) &
\(\hlprox{R_{W_Y\sim Z\mid W_Z,U}}\) &
\(\hlprox{R_{W_Y\sim W_Z\mid U}}\) \\
\midrule\noalign{}
\endhead
\bottomrule\noalign{}
\endlastfoot
Wealth & \(0.0075~~\textcolor{gray}{(0.05)}\) &
\(0.044~~\textcolor{gray}{(0.04)}\) &
\(-0.0746~~\textcolor{gray}{(0.03)}\) &
\(-0.30~~\textcolor{gray}{(0.1)}\) &
\(-0.0172~~\textcolor{gray}{(0.05)}\) &
\(0.01109~~\textcolor{gray}{(0.07)}\) \\
Schools & \(-0.0185~~\textcolor{gray}{(0.02)}\) &
\(0.043~~\textcolor{gray}{(0.04)}\) &
\(0.1084~~\textcolor{gray}{(0.05)}\) &
\(-0.30~~\textcolor{gray}{(0.1)}\) &
\(-0.0157~~\textcolor{gray}{(0.05)}\) &
\(-5.5\times 10^{-4}~~\textcolor{gray}{(0.07)}\) \\
Industrialization & \(-0.0413~~\textcolor{gray}{(0.03)}\) &
\(0.028~~\textcolor{gray}{(0.05)}\) &
\(-0.0452~~\textcolor{gray}{(0.06)}\) &
\(-0.30~~\textcolor{gray}{(0.1)}\) &
\(-0.0160~~\textcolor{gray}{(0.05)}\) &
\(0.00513~~\textcolor{gray}{(0.07)}\) \\
Log population & \(0.0378~~\textcolor{gray}{(0.05)}\) &
\(-0.212~~\textcolor{gray}{(0.05)}\) &
\(0.0041~~\textcolor{gray}{(0.02)}\) &
\(-0.31~~\textcolor{gray}{(0.1)}\) &
\(-0.0180~~\textcolor{gray}{(0.05)}\) &
\(0.00439~~\textcolor{gray}{(0.07)}\) \\
Diversity & \(-0.1257~~\textcolor{gray}{(0.05)}\) &
\(0.057~~\textcolor{gray}{(0.05)}\) &
\(-0.0326~~\textcolor{gray}{(0.04)}\) &
\(-0.30~~\textcolor{gray}{(0.1)}\) &
\(-0.0049~~\textcolor{gray}{(0.05)}\) &
\(0.00754~~\textcolor{gray}{(0.07)}\) \\
Katowickie region & \(-0.0130~~\textcolor{gray}{(0.04)}\) &
\(0.199~~\textcolor{gray}{(0.05)}\) &
\(0.1135~~\textcolor{gray}{(0.03)}\) &
\(-0.32~~\textcolor{gray}{(0.1)}\) &
\(-0.0158~~\textcolor{gray}{(0.05)}\) &
\(-0.00164~~\textcolor{gray}{(0.07)}\) \\
Opolskie region & \(0.0303~~\textcolor{gray}{(0.03)}\) &
\(0.101~~\textcolor{gray}{(0.04)}\) &
\(-0.0205~~\textcolor{gray}{(0.02)}\) &
\(-0.30~~\textcolor{gray}{(0.1)}\) &
\(-0.0155~~\textcolor{gray}{(0.05)}\) &
\(0.00379~~\textcolor{gray}{(0.07)}\) \\
\end{longtable}
\normalsize
\setstretch{\papersp}
\section{Mean-squared error contour plot}\label{sec-app-mse}
Figure~\ref{fig-contour-mse} is analogous to Figure~\ref{fig-contour}:
the plotted contours are the SOO bias. The contours are colored and
divided however, according to which estimator has the lowest
mean-squared error. Because in this case the three estimators have
similar variance, the plot is very similar to Figure~\ref{fig-contour}.
\begin{figure}
\centering{
\includegraphics[width=6.25in,height=\textheight,keepaspectratio]{figures/pol_sabotage_mse_contour.pdf}
}
\caption{\label{fig-contour-mse}Sensitivity plot for the running
example. The contour lines are on a logarithmic scale and show the
amount of SOO confounding bias. Contours are colored according to the
regions where each estimator has the lowest MSE. Gray crosses indicate
benchmark sensitivity parameter values for each of the covariates.}
\end{figure}%
\FloatBarrier
\section{Proofs}\label{sec-app-proofs}
For clarity in what follows, we abuse notation and manipulate random
variables as if they were random vectors of sufficient length \(N\), and
then implicitly send \(N\to\infty\) when converting back to statements
in terms of variances, covariances, etc.
\begin{lemma}[]\protect\hypertarget{lem-wy-y}{}\label{lem-wy-y}
\(\cov(W_Y^{\bot \hat W_Y}, Z^{\bot \hat W_Y})=0\).
\end{lemma}
\begin{proof}
Let \(\tilde W_Z = \mqty[Z & W_Z]\), and \(P_{\tilde W_Z}\) and
\(P_{\hat W_Y}\) be the projection matrices onto \(\tilde W_Z\) and
\(\hat W_Y\), respectively. Then \begin{align*}
P_{\hat W_Y} W_Y &= P_{\tilde W_Z} W_Y (W_Y P_{\tilde W_Z}^\top P_{\tilde W_Z} W_Y)^{-1} W_Y^\top P_{\tilde W_Z}^\top W_Y \\
&= P_{\tilde W_Z} W_Y = \hat W_Y.
\end{align*} Then \begin{align*}
\cov(W_Y^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) &= \qty{(I- P_{\hat W_Y}) W_Y}^\top \qty{(I- P_{\hat W_Y}) Z} \\
&= W_Y^\top (I- P_{\hat W_Y})^\top Z \\
&= (W_Y - \hat W_Y)^\top Z \\
&= 0,
\end{align*} since least-squares residuals are orthogonal to the
predictors (of which \(Z\) is one).
\end{proof}
\begin{lemma}[]\protect\hypertarget{lem-soo}{}\label{lem-soo}
Let \(\tau_\soo\) be the estimated coefficient in front of \(Z\) with
the regression specification \(Y \sim Z, W_Y\). Then, the bias that
arises from omitting \(U\) is: \[
\bias(\tau_\soo)
= \beta_u \frac{\cov(U^{\bot W_Z,W_Y}, Z^{\bot W_Z,W_Y})}{\var(Z^{\bot W_Z,W_Y})}
= \frac{R_{Y\sim U\mid Z,W_Z, W_Y} \cdot R_{Z \sim U \mid W_Z,W_Y}}{
\sqrt{1-R^2_{Z \sim U \mid W_Z,W_Y}}} \cdot
\frac{\sd(Y^{\bot Z,W_Z, W_Y})}{\sd(Z^{\bot W_Z,W_Y})}
\]
\end{lemma}
\begin{proof}
This result follows directly from \citet{cinelli2020making}. We provide
the full derivation for reference: \begin{align*}
\bias(\hat\tau_\soo) &= \beta_u \frac{\cov(U^{\bot W_Y}, Z^{\bot W_Y})}{\var(Z^{\bot W_Y})}\\
&= \frac{\cov(U^{\bot W_Z,W_Y}, Z^{\bot W_Z,W_Y})}{\var(Z^{\bot W_Z,W_Y})}
\cdot \frac{\cov(Y^{\bot Z,W_Z, W_Y}, U^{\bot Z, W_Z,W_Y})}{\var(U^{\bot Z,W_Z, W_Y})} \\
&= \frac{\cor(U^{\bot W_Z,W_Y}, Z^{\bot W_Z,W_Y}) \cdot
\sd(U^{\bot W_Z,W_Y})}{\sd(Z^{\bot W_Z,W_Y})} \cdot
\frac{\cor(Y^{\bot Z, W_Z,W_Y}, U^{\bot Z, W_Z,W_Y}) \cdot
\sd(Y^{\bot Z,W_Z, W_Y})}{\sd(U^{\bot Z,W_Z, W_Y})}\\
&= R_{Z\sim U\mid W_Z,W_Y}R_{Y\sim U\mid Z, W_Z,W_Y} \cdot \frac{\sd(U^{\bot W_Z, W_Y})}{\sd(U^{\bot Z,W_Z, W_Y})} \cdot \frac{\sd(Y^{\bot Z,W_Z, W_Y})}{\sd(Z^{\bot W_Z, W_Y})}\\
&= \frac{R_{Y\sim U\mid Z,W_Z, W_Y} \cdot R_{Z \sim U \mid W_Z,W_Y}}{
\sqrt{1-R^2_{Z \sim U \mid W_Z,W_Y}}} \cdot
\frac{\sd(Y^{\bot Z,W_Z, W_Y})}{\sd(Z^{\bot W_Z,W_Y})}
\end{align*}
\end{proof}
\subsection{\texorpdfstring{Proposition~\ref{prp-prox-valid}}{Proposition~}}\label{proposition-prp-prox-valid}
\begin{proof}
As in the main text, use Frisch-Waugh-Lovell to write
\(\tau_{\prox}-\tau\) as \begin{align*}
\tau_{\prox} -\tau &=
\beta_u \frac{\cov(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})} +
\beta_{w_y} \frac{\cov(W_Y^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})} +
\beta_{w_z} \frac{\cov(W_Z^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})}.
\end{align*}
The second covariance is zero by Lemma~\ref{lem-wy-y}. Moreover,
\ref{a-prox-z} is equivalent to \(\beta_{w_z}=0\), so the third term is
zero.
We next show \(\cov(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y})=0\). First,
notice that \[
\cov(W_Z,\eps_{w_y})=\varphi_u\cov(U,\eps_{w_y})+\varphi_{w_y}\cov(W_Y,\eps_{w_y})=\varphi_{w_y}\var(\eps_{w_y}),
\] so under \ref{a-prox-y} we have \(W_Z\perp\eps_{w_y}\). Similarly, \[
\cov(Z,\eps_{w_y})=\gamma_u\cov(U,\eps_{w_y})+\gamma_{w_y}\cov(W_Y,\eps_{w_y})+\gamma_{w_z}\cov(W_Z,\eps_{w_y}),
=(\gamma_{w_y} + \gamma_{w_z}\varphi_{w_y})\var(\eps_{w_y}),
\] which is likewise zero under \ref{a-prox-y}, implying
\(Z\perp\eps_{w_y}\).
Then let \(\tilde W_Z = \mqty[Z & W_Z]\), so we can write \(\hat W_Y\)
as \[
\hat W_Y = \alpha_u P_{\tilde W_Z} U + P_{\tilde W_Z} \eps_w = \alpha_u P_{\tilde W_Z} U
\] where \(P_{\tilde W_Z}\) represents the projection matrix onto
\(\tilde W_Z\), and \(P_{\tilde W_Z} \eps_w = 0\) since \(\eps_w\) is
independent. Notice that \(P_{\tilde W_Z}Z=Z\), since \(Z\) is in the
column space of \(\tilde W_Z\) by construction. Thus we can write (with
the \(\alpha_u\) cancelling) \begin{align*}
P_{\hat W_Y} Z &= P_{\tilde W_Z} U \big(U^\top P_{\tilde W_Z}^\top P_{\tilde W_Z} U \big)^{-1} U^\top P_{\tilde W_Z}^\top Z \\
&= P_{\tilde W_Z} U \big(U^\top P_{\tilde W_Z} U \big)^{-1} U^\top Z.
\end{align*} Then: \begin{align*}
\cov(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) &= \qty{ (I- P_{\hat W_Y}) U}^\top \qty{(I- P_{\hat W_Y}) Z} \\
&= U^\top (I- P_{\hat W_Y}) Z \\
&= U^\top Z - U^\top P_{\hat W_Y} Z) \\
&= U^\top Z - U^\top P_{\tilde W_Z} U \big(U^\top P_{\tilde W_Z} U \big)^{-1} U^\top Z \\
&= 0. \qedhere
\end{align*}
\end{proof}
\subsection{\texorpdfstring{Theorem~\ref{thm-iv}}{Theorem~}}\label{theorem-thm-iv}
\begin{proof}
By Frisch-Waugh-Lovell and Lemma~\ref{lem-wy-y}, we can write the bias
in \(\tau_\iv\) as \begin{equation}\phantomsection\label{eq-bias-iv-1}{
\tau_\iv - \tau = \underbrace{\beta_{w_z} \cdot \frac{\cov(\hat Z^{\bot W_Y}, W_Z^{\bot W_Y})}{\var(\hat Z^{\bot W_Y})}}_{(1)} + \underbrace{\beta_u \frac{\cov(\hat Z^{\bot W_Y}, U^{\bot W_Y})}{\var(\hat Z^{\bot W_Y})}}_{(2)}.
}\end{equation}
Term 1 represents violations in the exclusion restriction. We can
re-write term 1 as follows, where \(\hat \gamma\) represent the
coefficient on \(W_Z\) in the regression of \(Z\) on \(W_Z\) and
\(W_Y\): \begin{align*}
\beta_{w_z} \cdot \frac{\cov(\hat Z^{\bot W_Y}, W_Z^{\bot W_Y})}{\var(\hat Z^{\bot W_Y})}
&= \beta_{w_z} \cdot \frac{\cov((\hat\gamma W_Z)^{\bot W_Y}, W_Z^{\bot W_Y})}{\var((\hat\gamma W_Z){\bot W_Y})} \\
&= \beta_{w_z} \cdot \frac{\hat \gamma \cdot \cov(W_Z^{\bot W_Y}, W_Z^{\bot W_Y})}{\hat \gamma^2 \var(W_Z^{\bot W_Y})}\\
&= \beta_{w_z} \cdot \frac{1}{\hat \gamma} \\
&= \beta_{w_z} \cdot \frac{\var(W_Z^{\bot W_Y})}{\cov(W_Z^{\bot W_Y}, Z^{\bot W_Y})} \\
&= \beta_{w_z} \cdot \frac{1}{\cor(W_Z^{\bot W_Y}, Z^{\bot W_Y})} \cdot \frac{\sd(W_Z^{\bot W_Y})}{\sd(Z^{\bot W_Y})}
\end{align*}
Term 2 represents violations in instrument exogeneity: \begin{align*}
\beta_u \frac{\cov(\hat Z^{\bot W_Y}, U^{\bot W_Y})}{\var(\hat Z^{\bot W_Y})}
&= \beta_u \cdot \frac{1}{\hat \gamma} \cdot \frac{\cov( W_Z^{\bot W_Y}, U^{\bot W_Y})}{\var( W_Z^{\bot W_Y})} \\
&= \beta_u \cdot \frac{\var(W_Z^{\bot W_Y})}{\cov(W_Z^{\bot W_Y}, Z^{\bot W_Y})}\cdot \frac{\cov( W_Z^{\bot W_Y}, U^{\bot W_Y})}{\var( W_Z^{\bot W_Y})} \\
&= \beta_u \cdot \frac{\cov(W_Z^{\bot W_Y}, U^{\bot W_Y})}{\cov(W_Z^{\bot W_Y}, Z^{\bot W_Y})} \\
&= \beta_u \cdot \frac{\cor(W_Z^{\bot W_Y}, U^{\bot W_Y}) \sd(U^{\bot W_Y})}{\cor(W_Z^{\bot W_Y}, Z^{\bot W_Y}) \sd(Z^{\bot W_Y})}
\end{align*} The correlation term (i.e.,
\(\cor(W_Z^{\bot W_Y}, U^{\bot W_Y})\)) will equal zero if IV exogeneity
holds.
\noindent Putting it together: \begin{align*}
\tau_\iv - \tau =&\beta_{w_z} \cdot \frac{1}{\cor(W_Z^{\bot W_Y}, Z^{\bot W_Y})} \cdot \frac{\sd(W_Z^{\bot W_Y})}{\sd(Z^{\bot W_Y})} + \beta_u \cdot \frac{\cor(W_Z^{\bot W_Y}, U^{\bot W_Y}) \sd(U^{\bot W_Y})}{\cor(W_Z^{\bot W_Y}, Z^{\bot W_Y}) \sd(Z^{\bot W_Y})}\\
=& \frac{1}{R_{W_Z \sim Z \mid W_Y}} \cdot \frac{\sd(W_Z^{\bot W_Y})}{\sd(Z^{\bot W_Y})}\left( \beta_{w_z} + \beta_u \cdot \varphi_u \right)
\end{align*}
We can then re-write \(\beta_{w_z}\) and \(\beta_u\) in terms of \(R^2\)
values. Throughout, we will make use of the following decomposition for
partial \(R^2\) values: \begin{align*}
R^2_{Y \sim A \mid B, C} = \frac{R^2_{Y \sim A, B, C} - R^2_{Y \sim B, C}}{1-R^2_{Y \sim B, C}} \implies R^2_{Y \sim A, B, C} = R^2_{Y \sim A \mid B, C}(1-R^2_{Y \sim B, C}) + R^2_{Y \sim B, C}
\end{align*}
\begin{align}
\beta_{w_z} &= \frac{\cov(W_Z^{\bot W_Y, U, Z}, Y^{\bot W_Y, U, Z})}{\var(W_Z^{\bot W_Y, U, Z})} \nonumber \\
&= R_{W_Z \sim Y \mid W_Y, U, Z} \cdot \frac{\sd(Y^{\bot W_Y, U, Z})}{\sd(W_Z^{\bot W_Y, U, Z})} \nonumber \\
&= R_{Y \sim W_Z \mid W_Y, Z, U} \sqrt{\frac{\var(Y)}{\var(W_Z)} \cdot \frac{(1-R^2_{Y \sim U \mid W_Y, Z}) (1-R^2_{Y \sim W_Y, Z})}{(1-R^2_{W_Z \sim U \mid W_Y, Z})(1-R^2_{W_Z \sim W_Y, Z})}}
\label{eqn:beta_wz_r2_1}\\
&= R_{Y \sim W_Z \mid W_Y, Z, U} \sqrt{\frac{\var(Y^{\bot W_Y, Z})}{\var(W_Z^{\bot W_Y})} \cdot \frac{(1-R^2_{Y \sim U \mid W_Y, Z})}{(1-R^2_{W_Z \sim U \mid W_Y, Z})(1-R^2_{Z \sim W_Z \mid W_Y})}}
\label{eqn:beta_wz_r2_2}
\end{align}
We can re-write \(\beta_u \cdot \varphi_u\) as: \begin{align*}
\beta_u &\cdot \varphi_u \\
=& R_{Y\sim U \mid W_Z, W_Y, Z} \cdot R_{W_Z \sim U \mid W_Y} \cdot \frac{\sd(Y^{\bot W_Z, W_Y, Z})}{\sd(U^{\bot W_Z, W_Y, Z})} \cdot \frac{\sd(U^{\bot W_Y})}{\sd(W_Z^{\bot W_Y})}\\
=& R_{Y\sim U \mid W_Z, W_Y, Z} \cdot R_{W_Z \sim U \mid W_Y} \cdot\frac{\sd(U^{\bot W_Y})}{\sd(U^{\bot W_Z, W_Y, Z})} \frac{\sd(Y^{\bot W_Z, W_Y, Z})}{\sd(W_Z^{\bot W_Y})} \\
=& R_{Y\sim U \mid W_Z, W_Y, Z} \cdot R_{W_Z \sim U \mid W_Y} \cdot \sqrt{\frac{1}{1-R^2_{U \sim W_Z, Z \mid W_Y}}} \frac{\sd(Y^{\bot W_Z, W_Y, Z})}{\sd(W_Z^{\bot W_Y})} \\
=& R_{Y\sim U \mid W_Z, W_Y, Z} \cdot R_{W_Z \sim U \mid W_Y} \\
& \times \sqrt{\frac{1}{(1-R^2_{U \sim Z \mid W_Z, W_Y})(1-R^2_{U \sim W_Z \mid W_Y})}} \frac{\sd(Y^{\bot W_Z, W_Y, Z})}{\sd(W_Z^{\bot W_Y})}
\end{align*}
Combining together: \begin{align*}
&\frac{1}{R_{W_Z \sim Z \mid W_Y}} \cdot \frac{\sd(W_Z^{\bot W_Y})}{\sd(Z^{\bot W_Y})} \left\{R_{Y \sim W_Z \mid W_Y, Z, U} \sqrt{\frac{\var(Y^{\bot W_Y, Z})}{\var(W_Z^{\bot W_Y})} \cdot \frac{(1-R^2_{Y \sim U \mid W_Y, Z})}{(1-R^2_{W_Z \sim U \mid W_Y, Z})(1-R^2_{Z \sim W_Z \mid W_Y})}}
\right. \\
&+\left. R_{Y\sim U \mid W_Z, W_Y, Z} \cdot R_{W_Z \sim U \mid W_Y}\sqrt{\frac{1}{1-R^2_{W_Z \sim U \mid W_Y}} \cdot \frac{1}{1-R^2_{Z \sim U \mid W_Y, W_Z}}} \cdot \frac{\sd(Y^{\bot W_Z, W_Y, Z})}{\sd(W_Z^{\bot W_Y})}\right\} \\
=&\frac{R_{Y \sim W_Z \mid W_Y, Z, U} }{R_{W_Z \sim Z \mid W_Y}} \cdot \frac{\sd(Y^{\bot W_Y, Z})}{\sd(Z^{\bot W_Y})} \cdot \left\{ \sqrt{\frac{1-R^2_{Y \sim U \mid W_Y, Z}}{1-R^2_{W_Z \sim U \mid W_Y, Z}} \cdot \frac{1}{1-R^2_{Z \sim W_Z \mid W_Y}}} \right. \\
&+\left. R_{Y\sim U \mid W_Z, W_Y, Z} \cdot R_{W_Z \sim U \mid W_Y}\sqrt{\frac{1}{1-R^2_{W_Z \sim U \mid W_Y}} \cdot \frac{1}{1-R^2_{Z \sim U \mid W_Y, W_Z}}} \cdot \frac{\sd(Y^{\bot W_Z, W_Y, Z})}{\sd(Y^{\bot W_Y, Z})}\right\} \\
=&\frac{R_{Y \sim W_Z \mid W_Y, Z, U} }{R_{W_Z \sim Z \mid W_Y}} \cdot \frac{\sd(Y^{\bot W_Y, Z})}{\sd(Z^{\bot W_Y})} \cdot \left\{ \sqrt{\frac{1-R^2_{Y \sim U \mid W_Y, Z}}{1-R^2_{W_Z \sim U \mid W_Y, Z}} \cdot \frac{1}{1-R^2_{Z \sim W_Z \mid W_Y}}} \right. \\
&+\left. R_{Y\sim U \mid W_Z, W_Y, Z} \cdot R_{W_Z \sim U \mid W_Y}\sqrt{\frac{1}{1-R^2_{W_Z \sim U \mid W_Y}} \cdot \frac{R^2_{Y \sim W_Z \mid W_Y, Z}}{1-R^2_{Z \sim U \mid W_Y, W_Z}}}\right\} \\
=&\frac{R_{Y \sim W_Z \mid W_Y, Z, U} }{R_{W_Z \sim Z \mid W_Y}} \cdot \frac{\sd(Y^{\bot W_Y, Z})}{\sd(Z^{\bot W_Y})} \cdot \left\{ \sqrt{\frac{1-R^2_{Y \sim U \mid W_Y, Z}}{1-R^2_{W_Z \sim U \mid W_Y, Z}} \cdot \frac{1}{1-R^2_{Z \sim W_Z \mid W_Y}}} \right. \\
&+\left. R_{Y\sim U \mid W_Z, W_Y, Z} \cdot R_{W_Z \sim U \mid W_Y}\sqrt{\frac{1}{1-R^2_{Z \sim U \mid W_Y, W_Z}} \cdot \frac{1}{1-R^2_{W_Z \sim U \mid W_Y}} \cdot R^2_{Y \sim W_Z \mid W_Y, Z}}\right\}
\end{align*}
\end{proof}
\subsection{\texorpdfstring{Corollary~\ref{cor-relative-bias-iv}}{Corollary~}}\label{corollary-cor-relative-bias-iv}
\begin{proof}
\begin{align*}
R^2_{Y \sim W_Z \mid W_Y, Z, U} &\leq \frac{R^2_{Y \sim U \mid W_Y, W_Z, Z}}{1- R^2_{Y \sim U \mid W_Y, Z}} \cdot \frac{R^2_{Z \sim U \mid W_Y, W_Z}}{1-R^2_{Z \sim U \mid W_Y, W_Z}} \cdot \left( 1- R^2_{Z \sim W_Z \mid W_Y} \right) \cdot \frac{\var(Z^{\bot W_Y})}{\var(Y^{\bot W_Y, Z})} \cdot R^2_{Z \sim W_Z \mid W_Y} \cdot \frac{\var(Y^{\bot W_Y, W_Z, Z})}{\var(Z^{\bot W_Y, W_Z})}\\
&= \frac{R^2_{Y \sim U \mid W_Y, W_Z, Z}}{1- R^2_{Y \sim U \mid W_Y, Z}} \cdot \frac{R^2_{Z \sim U \mid W_Y, W_Z}}{1-R^2_{Z \sim U \mid W_Y, W_Z}} \cdot \left( 1- R^2_{Z \sim W_Z \mid W_Y} \right) \cdot \frac{ \var(Y^{\bot W_Y, W_Z, Z})}{\var(Y^{\bot W_Y, Z})} \cdot R^2_{Z \sim W_Z \mid W_Y} \cdot\frac{\var(Z^{\bot W_Y})}{\var(Z^{\bot W_Y, W_Z})}\\
&= \frac{R^2_{Y \sim U \mid W_Y, W_Z, Z}}{1- R^2_{Y \sim U \mid W_Y, Z}} \cdot \frac{R^2_{Z \sim U \mid W_Y, W_Z}}{1-R^2_{Z \sim U \mid W_Y, W_Z}} \cdot \left( 1- R^2_{Z \sim W_Z \mid W_Y} \right) \cdot (1-R^2_{Y \sim W_Z \mid W_Y, Z}) \cdot \frac{R^2_{Z \sim W_Z \mid W_Y}}{1-R^2_{Z \sim W_Z \mid W_Y}} \\
&= \frac{R^2_{Y \sim U \mid W_Y, W_Z, Z}}{1- R^2_{Y \sim U \mid W_Y, Z}} \cdot \frac{R^2_{Z \sim U \mid W_Y, W_Z}}{1-R^2_{Z \sim U \mid W_Y, W_Z}} \cdot (1-R^2_{Y \sim W_Z \mid W_Y, Z}) \cdot R^2_{Z \sim W_Z \mid W_Y} \\
&= \frac{R^2_{Y \sim U \mid W_Y, W_Z, Z}}{1- \gamma \cdot R^2_{Y \sim U \mid W_Y +W_Z, Z}} \cdot \frac{R^2_{Z \sim U \mid W_Y, W_Z}}{1-R^2_{Z \sim U \mid W_Y, W_Z}} \cdot (1-R^2_{Y \sim W_Z \mid W_Y, Z}) \cdot R^2_{Z \sim W_Z \mid W_Y}
\end{align*}
\end{proof}
\subsection{\texorpdfstring{Theorem~\ref{thm-prox-bias}}{Theorem~}}\label{theorem-thm-prox-bias}
\begin{proof}
As in the main text, use Frisch-Waugh-Lovell to write
\(\tau_{\prox}-\tau\) as \begin{align*}
\tau_{\prox} -\tau &=
\underbrace{\beta_u \frac{\cov(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})}}_{(1)} +
\underbrace{\beta_{w_y} \frac{\cov(W_Y^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})}}_{(2)} +
\underbrace{\beta_{w_z} \frac{\cov(W_Z^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})}}_{(3)}.
\end{align*}
\noindent We re-write each term.
\begin{itemize}
\item
Term 1: \begin{align*}
\beta_u \frac{\cov(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})} &= \beta_u \cdot \cor(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) \cdot \frac{\sd(U^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})}
\end{align*}
\item
Term 2: \begin{align*}
\beta_{w_y}& \cdot \frac{\cov(W_Y^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})}\\
&= \beta_{w_y} \cdot \frac{\alpha_u \cov(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) + \alpha_{w_z} \cov(W_Z^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})} \\
&= \beta_{w_y} \alpha_u \cor(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) \cdot \frac{\sd(U^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})} + \beta_{w_y} \cdot \alpha_{w_z} \cor(W_Z^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) \cdot \frac{\sd(W_Z^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})}
\end{align*} Since this term equals zero (Lemma~\ref{lem-wy-y}), we
can re-write
\(\cor(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) \cdot \sd(U^{\bot \hat W_Y})/\sd(Z^{\bot \hat W_Y})\)
as:
\end{itemize}
\begin{align*}
\cor(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) \cdot \frac{\sd(U^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})} &= -\frac{\alpha_{w_z}}{\alpha_u}\cor(W_Z^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) \cdot \frac{\sd(W_Z^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})}
\end{align*}
\begin{itemize}
\tightlist
\item
Term 3: \begin{align*}
\beta_{w_z} &\cdot \frac{\cov(W_Z^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})}\\
&= \beta_{w_z} \cdot \cor(W_Z^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) \frac{\sd(W_Z^{\hat W_Y})}{\sd(Z^{\bot \hat W_Y})}
%\\
%&= \beta_{w_z} \cdot \varphi_u \cdot \frac{\cov(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y})}{\var(Z^{\bot \hat W_Y})} \\
%&= \beta_{w_z} \cdot \varphi_u \cdot \cor(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) \cdot \frac{\sd(U^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})}
\end{align*}
\end{itemize}
Putting it together, and defining
\(S_{W_Z,Z}\coloneq\cor(W_Z^{\bot \hat W_Y}, Z^{\bot \hat W_Y})\), we
have: \begin{align*}
& \left(\beta_{w_z} - \alpha_{w_z} \frac{\beta_u }{\alpha_u} \right) \cor(W_Z^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) \cdot \frac{\sd(W_Z^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})} \\
&= S_{W_Z, Z}\cdot \left(\beta_{w_z} - \alpha_{w_z} \frac{\beta_u }{\alpha_u} \right) \frac{\sd(W_Z^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})}
% \\
%=& \beta_u \left(1 + \frac{\beta_{w_y}}{\beta_u} \cdot \alpha_u + \frac{\beta_{w_z}}{\beta_u} \cdot \varphi_u \right) \cor(U^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) \cdot \frac{\sd(U^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})}\\
%& + \beta_u \cdot \frac{\beta_{w_y}}{\beta_u} \cdot \alpha_{w_z} \cdot \cor(W_Z^{\bot \hat W_Y}, Z^{\bot \hat W_Y}) \cdot \frac{\sd(W_Z^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})}\\
\end{align*} Notice that because \(\hat W_Y\) is a linear combination of
\(W_Z\) and \(Z\), we must have
\(|\cor(W_Z^{\bot \hat W_Y}, Z^{\bot \hat W_Y})|=1\), so
\(S_{W_Z,Z}\in\{-1, 1\}\).
Turning to the partial \(R^2\) decomposition, we have
\begin{align*}
\frac{\alpha_{w_z}}{\alpha_u} &= \frac{\cov(W_Y^{\bot U}, W_Z^{\bot U})}{\cov(U^{\bot W_Z}, W_Y^{\bot W_Z})} \cdot \frac{\var(U^{\bot W_Z})}{\var(W_Z^{\bot U})} \\
&= \frac{\cor(W_Y^{\bot U}, W_Z^{\bot U})}{\cor(U^{\bot W_Z}, W_Y^{\bot W_Z})} \cdot \frac{\sd(W_Y^{\bot U})}{\sd(W_Y^{\bot U})} \frac{\sd(U^{\bot W_Z})}{\sd(W_Z^{\bot U})} \\
&= \frac{\cor(W_Y^{\bot U}, W_Z^{\bot U})}{\cor(U^{\bot W_Z}, W_Y^{\bot W_Z})}\cdot \frac{\sd(U^{\bot W_Z})}{\sd(W_Z^{\bot U})} \\
&= \sqrt{\frac{1-R^2_{W_Y \sim W_Z}}{R^2_{W_Y \sim W_Z \mid U} (1-R^2_{W_Y \sim U}) + R^2_{W_Y \sim U} - R^2_{W_Y \sim W_Z}} \cdot R^2_{W_Y \sim W_Z \mid U} }\frac{\sd(U)}{\sd(W_Z)}.
\end{align*} Let
\(\eta = \frac{1-R^2_{W_Y \sim W_Z}}{R^2_{W_Y \sim W_Z \mid U} (1-R^2_{W_Y \sim U}) + R^2_{W_Y \sim U} - R^2_{W_Y \sim W_Z}} \cdot R^2_{W_Y \sim W_Z \mid U}\)
for simplicity of notation.
Then, note that we can re-write \(\beta_u\) as: \begin{align}
\beta_u &= \frac{\cov(U^{\bot W_Z, W_Y, Z}, Y^{\bot W_Z, W_Y, Z})}{\var(U^{\bot W_Z, W_Y, Z})} \nonumber \\
&= \cor(U^{\bot W_Z, W_Y, Z}, Y^{\bot W_Z, W_Y, Z}) \cdot \frac{\sd(Y^{\bot W_Z, W_Y, Z})}{\sd(U^{\bot W_Z, W_Y, Z})} \nonumber \\
&= R_{Y \sim U \mid W_Z, W_Y, Z} \sqrt{\frac{\var(Y)}{\var(U)} \cdot \frac{1-R^2_{Y \sim W_Y, W_Z, Z}}{(1-R^2_{U \sim W_Z \mid W_Y, Z})(1-R^2_{U \sim W_Y, Z})}} \nonumber \\
&= R_{Y \sim U \mid W_Z, W_Y, Z} \sqrt{\frac{\var(Y)}{\var(U)} \cdot \frac{1-R^2_{Y \sim W_Y, W_Z, Z}}{(1-R^2_{W_Z \sim U \mid W_Y, Z})(1-R^2_{U \sim W_Y, Z})}}
\label{eqn:beta_u_r2}
\end{align}
Putting these results together with the expansion of \(\beta_{w_z}\)
from the proof of Theorem~\ref{thm-iv}: \begin{align*}
&\beta_{w_z} - \beta_u \cdot \frac{\alpha_{w_z}}{\alpha_u}\\
&= R_{Y \sim W_Z \mid W_Y, Z, U} \sqrt{\frac{\var(Y)}{\var(W_Z)} \cdot \frac{(1-R^2_{Y \sim U \mid W_Y, Z}) (1-R^2_{Y \sim W_Y, Z})}{(1-R^2_{W_Z \sim U \mid W_Y, Z})(1-R^2_{W_Z \sim W_Y, Z})}} \\
& - R_{Y \sim U \mid W_Z, W_Y, Z} \sqrt{\frac{\var(Y)}{\var(U)} \cdot \frac{1-R^2_{Y \sim W_Y, W_Z, Z}}{(1-R^2_{W_Z \sim U \mid W_Y, Z})(1-R^2_{U \sim W_Y, Z})}} \cdot \sqrt{\eta \cdot \frac{\var(U)}{\var(W_Z)}} \\
&= R_{Y \sim W_Z \mid W_Y, Z, U} \sqrt{\frac{\var(Y)}{\var(W_Z)} \cdot \frac{(1-R^2_{Y \sim U \mid W_Y, Z}) (1-R^2_{Y \sim W_Y, Z})}{(1-R^2_{W_Z \sim U \mid W_Y, Z})(1-R^2_{W_Z \sim W_Y, Z})}} \\
&- R_{Y \sim U \mid W_Z, W_Y, Z} \sqrt{\frac{\var(Y)}{\var(W_Z)} \cdot \frac{1-R^2_{Y \sim W_Y, W_Z, Z}}{(1-R^2_{W_Z \sim U \mid W_Y, Z})(1-R^2_{U \sim W_Y, Z})} \cdot \eta} \\
&= \frac{\sd(Y)}{\sd(W_Z)} \cdot \left\{ R_{Y \sim W_Z \mid W_Y, Z, U} \sqrt{\frac{(1-R^2_{Y \sim U \mid W_Y, Z}) (1-R^2_{Y \sim W_Y, Z})}{(1-R^2_{W_Z \sim U \mid W_Y, Z})(1-R^2_{W_Z \sim W_Y, Z})}} \right. \\
& \left. - R_{Y \sim U \mid W_Z, W_Y, Z} \sqrt{\frac{1-R^2_{Y \sim W_Y, W_Z, Z}}{(1-R^2_{W_Z \sim U \mid W_Y, Z})(1-R^2_{U \sim W_Y, Z})} \cdot \eta} \right\} \\
&= \frac{\sd(Y)}{\sd(W_Z)} \cdot \frac{1}{\sqrt{1-R^2_{W_Z \sim U \mid W_Y, Z}}} \left\{ R_{Y \sim W_Z \mid W_Y, Z, U} \sqrt{\frac{(1-R^2_{Y \sim U \mid W_Y, Z}) (1-R^2_{Y \sim W_Y, Z})}{1-R^2_{W_Z \sim W_Y, Z}}} \right. \\
& \left. - R_{Y \sim U \mid W_Z, W_Y, Z} \sqrt{\frac{1-R^2_{Y \sim W_Y, W_Z, Z}}{1-R^2_{U \sim W_Y, Z}} \cdot \eta} \right\} \\
&= \frac{\sd(Y)}{\sd(W_Z)} \cdot R_{Y \sim U \mid W_Z, W_Y, Z} \sqrt{\frac{1}{1-R^2_{W_Z \sim U \mid W_Y, Z}}} \left\{ R_{Y \sim W_Z \mid W_Y, Z, U} \sqrt{\frac{1-R^2_{Y \sim U \mid W_Y, Z}}{R^2_{Y \sim U \mid W_Z, W_Y, Z}} \frac{1-R^2_{Y \sim W_Y, Z}}{1-R^2_{W_Z \sim W_Y, Z}}} \right. \\
& \left. - \sqrt{\frac{1-R^2_{Y \sim W_Y, W_Z, Z}}{1-R^2_{U \sim W_Y, Z}} \cdot \eta} \right\}
\intertext{Then, we can write $1-R^2_{U \sim W_Y, Z} = (1-R^2_{U \sim Z \mid W_Y}) (1-R^2_{U \sim W_Y}) = (1-R^2_{Z \sim U \mid W_Y})(1-R^2_{W_Y \sim U})$:}
&= \frac{\sd(Y)}{\sd(W_Z)} \cdot R_{Y \sim U \mid W_Z, W_Y, Z} \sqrt{\frac{1}{1-R^2_{W_Z \sim U \mid W_Y, Z}}} \left\{ R_{Y \sim W_Z \mid W_Y, Z, U} \sqrt{\frac{1-R^2_{Y \sim U \mid W_Y, Z}}{R^2_{Y \sim U \mid W_Z, W_Y, Z}} \frac{1-R^2_{Y \sim W_Y, Z}}{1-R^2_{W_Z \sim W_Y, Z}}} \right. \\
& \left. - \sqrt{\frac{1-R^2_{Y \sim W_Y, W_Z, Z}}{(1-R^2_{Z \sim U \mid W_Y})(1-R^2_{W_Y \sim U})} \cdot \frac{1-R^2_{W_Y \sim W_Z}}{R^2_{W_Y \sim W_Z \mid U} (1-R^2_{W_Y \sim U}) + R^2_{W_Y \sim U} - R^2_{W_Y \sim W_Z}} \cdot R^2_{W_Y \sim W_Z \mid U}} \right\}
\end{align*}
Putting it together: \begin{align*}
&S_{W_Z, Z}\cdot \left(\beta_{w_z} - \alpha_{w_z} \frac{\beta_u }{\alpha_u} \right) \frac{\sd(W_Z^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})} \\
&= S_{W_Z, Z} \cdot \frac{\sd(Y)}{\sd(W_Z)} \cdot R_{Y \sim U \mid W_Z, W_Y, Z} \sqrt{\frac{1}{1-R^2_{W_Z \sim U \mid W_Y, Z}}} \left\{ R_{Y \sim W_Z \mid W_Y, Z, U} \sqrt{\frac{1-R^2_{Y \sim U \mid W_Y, Z}}{R^2_{Y \sim U \mid W_Z, W_Y, Z}} \frac{1-R^2_{Y \sim W_Y, Z}}{1-R^2_{W_Z \sim W_Y, Z}}} \right. \\
& \left. - \sqrt{\frac{1-R^2_{Y \sim W_Y, W_Z, Z}}{(1-R^2_{Z \sim U \mid W_Y})(1-R^2_{W_Y \sim U})} \cdot \frac{1-R^2_{W_Y \sim W_Z}}{R^2_{W_Y \sim W_Z \mid U} (1-R^2_{W_Y \sim U}) + R^2_{W_Y \sim U} - R^2_{W_Y \sim W_Z}} \cdot R^2_{W_Y \sim W_Z \mid U}} \right\} \cdot \frac{\sd(W_Z^{\bot \hat W_Y})}{\sd(Z^{\bot \hat W_Y})}\\
&= S_{W_Z, Z} \cdot \frac{\sd(Y)}{\sd(Z)} \cdot \sqrt{\frac{1-R^2_{\hat W_Y \sim W_Z}}{1-R^2_{\hat W_Y \sim Z}}} \cdot R_{Y \sim U \mid W_Z, W_Y, Z} \sqrt{\frac{1}{1-R^2_{W_Z \sim U \mid W_Y, Z}}} \left\{ R_{Y \sim W_Z \mid W_Y, Z, U} \sqrt{\frac{1-R^2_{Y \sim U \mid W_Y, Z}}{R^2_{Y \sim U \mid W_Z, W_Y, Z}} \frac{1-R^2_{Y \sim W_Y, Z}}{1-R^2_{W_Z \sim W_Y, Z}}} \right. \\
& \left. - \sqrt{\frac{1-R^2_{Y \sim W_Y, W_Z, Z}}{(1-R^2_{Z \sim U \mid W_Y})(1-R^2_{W_Y \sim U})} \cdot \frac{1-R^2_{W_Y \sim W_Z}}{R^2_{W_Y \sim W_Z \mid U} (1-R^2_{W_Y \sim U}) + R^2_{W_Y \sim U} - R^2_{W_Y \sim W_Z}} \cdot R^2_{W_Y \sim W_Z \mid U}} \right\}
\end{align*}
\end{proof}
% END BODY %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\end{document}
\begin{thebibliography}{}
\bibitem[Anderson and Rubin, 1949]{anderson1949estimation}
Anderson, T.~W. and Rubin, H. (1949).
\newblock Estimation of the parameters of a single equation in a complete system of stochastic equations.
\newblock {\em The Annals of mathematical statistics}, 20(1):46--63.
\bibitem[Andrews et~al., 2019]{andrews2019weak}
Andrews, I., Stock, J.~H., and Sun, L. (2019).
\newblock Weak instruments in instrumental variables regression: Theory and practice.
\newblock {\em Annual Review of Economics}, 11(1):727--753.
\bibitem[Bound et~al., 1995]{bound1995problems}
Bound, J., Jaeger, D.~A., and Baker, R.~M. (1995).
\newblock Problems with instrumental variables estimation when the correlation between the instruments and the endogenous explanatory variable is weak.
\newblock {\em Journal of the American statistical association}, 90(430):443--450.
\bibitem[Carnegie et~al., 2016]{carnegie2016assessing}
Carnegie, N.~B., Harada, M., and Hill, J.~L. (2016).
\newblock Assessing sensitivity to unmeasured confounding using a simulated potential confounder.
\newblock {\em Journal of Research on Educational Effectiveness}, 9(3):395--420.
\bibitem[Chernozhukov et~al., 2024]{chernozhukov2024long}
Chernozhukov, V., Cinelli, C., Newey, W., Sharma, A., and Syrgkanis, V. (2024).
\newblock Long story short: Omitted variable bias in causal machine learning.
\newblock Technical report, National Bureau of Economic Research.
\bibitem[Cinelli and Hazlett, 2020]{cinelli2020making}
Cinelli, C. and Hazlett, C. (2020).
\newblock Making sense of sensitivity: Extending omitted variable bias.
\newblock {\em Journal of the Royal Statistical Society Series B: Statistical Methodology}, 82(1):39--67.
\bibitem[Cinelli and Hazlett, 2025]{cinelli2025iv}
Cinelli, C. and Hazlett, C. (2025).
\newblock An omitted variable bias framework for sensitivity analysis of instrumental variables.
\newblock {\em Biometrika}, page asaf004.
\bibitem[Cobzaru et~al., 2024]{cobzaru2024bias}
Cobzaru, R., Welsch, R., Finkelstein, S., Ng, K., and Shahn, Z. (2024).
\newblock Bias formulas for violations of proximal identification assumptions in a linear structural equation model.
\newblock {\em Journal of Causal Inference}, 12(1):20230039.
\bibitem[Cui et~al., 2024]{cui2024semiparametric}
Cui, Y., Pu, H., Shi, X., Miao, W., and Tchetgen~Tchetgen, E. (2024).
\newblock Semiparametric proximal causal inference.
\newblock {\em Journal of the American Statistical Association}, 119(546):1348--1359.
\bibitem[Ding et~al., 2019]{ding2019decomposing}
Ding, P., Feller, A., and Miratrix, L. (2019).
\newblock Decomposing treatment effect variation.
\newblock {\em Journal of the American Statistical Association}, 114(525):304--317.
\bibitem[Dorn and Guo, 2023]{dorn2023sharp}
Dorn, J. and Guo, K. (2023).
\newblock Sharp sensitivity analysis for inverse propensity weighting via quantile balancing.
\newblock {\em Journal of the American Statistical Association}, 118(544):2645--2657.
\bibitem[Frank, 2000]{frank2000impact}
Frank, K.~A. (2000).
\newblock Impact of a confounding variable on a regression coefficient.
\newblock {\em Sociological Methods \& Research}, 29(2):147--194.
\bibitem[Freidling and Zhao, 2022]{freidling2022optimization}
Freidling, T. and Zhao, Q. (2022).
\newblock Optimization-based sensitivity analysis for unmeasured confounding using partial correlations.
\newblock {\em arXiv preprint arXiv:2301.00040}.
\bibitem[Hager and Krakowski, 2022]{hk2022poland}
Hager, A. and Krakowski, K. (2022).
\newblock Does state repression spark protests? evidence from secret police surveillance in communist poland.
\newblock {\em American Political Science Review}, 116(2):564--579.
\bibitem[Hahn and Hausman, 2005]{hahn2005estimation}
Hahn, J. and Hausman, J. (2005).
\newblock Estimation with valid and invalid instruments.
\newblock {\em Annales d'Economie et de Statistique}, pages 25--57.
\bibitem[Hartman and Huang, 2024]{hartman2024sensitivity}
Hartman, E. and Huang, M. (2024).
\newblock Sensitivity analysis for survey weights.
\newblock {\em Political Analysis}, 32(1):1--16.
\bibitem[Hong et~al., 2021]{hong2021did}
Hong, G., Yang, F., and Qin, X. (2021).
\newblock Did you conduct a sensitivity analysis? a new weighting-based approach for evaluations of the average treatment effect for the treated.
\newblock {\em Journal of the Royal Statistical Society: Series A (Statistics in Society)}, 184(1):227--254.
\bibitem[Huang and Pimentel, 2025]{huang2025variance}
Huang, M. and Pimentel, S.~D. (2025).
\newblock Variance-based sensitivity analysis for weighting estimators results in more informative bounds.
\newblock {\em Biometrika}, 112(1):asae040.
\bibitem[Huang et~al., 2025]{huang2025design}
Huang, M., Soriano, D., and Pimentel, S.~D. (2025).
\newblock Design sensitivity and its implications for weighted observational studies.
\newblock {\em Journal of the Royal Statistical Society Series A: Statistics in Society}, page qnaf067.
\bibitem[Huang, 2024]{huang2024sensitivity}
Huang, M.~Y. (2024).
\newblock Sensitivity analysis for the generalization of experimental results.
\newblock {\em Journal of the Royal Statistical Society Series A: Statistics in Society}, 187(4):900--918.
\bibitem[Imbens, 2003]{imbens2003sensitivity}
Imbens, G.~W. (2003).
\newblock Sensitivity to exogeneity assumptions in program evaluation.
\newblock {\em American Economic Review}, 93(2):126--132.
\bibitem[Kang et~al., 2021]{kang2021ivmodel}
Kang, H., Jiang, Y., Zhao, Q., and Small, D.~S. (2021).
\newblock Ivmodel: an r package for inference and sensitivity analysis of instrumental variables models with one endogenous variable.
\newblock {\em Observational Studies}, 7(2):1--24.
\bibitem[Lewandowski et~al., 2009]{lewandowski2009generating}
Lewandowski, D., Kurowicka, D., and Joe, H. (2009).
\newblock Generating random correlation matrices based on vines and extended onion method.
\newblock {\em Journal of multivariate analysis}, 100(9):1989--2001.
\bibitem[Miao et~al., 2018]{miao2018identifying}
Miao, W., Geng, Z., and Tchetgen~Tchetgen, E.~J. (2018).
\newblock Identifying causal effects with proxy variables of an unmeasured confounder.
\newblock {\em Biometrika}, 105(4):987--993.
\bibitem[Miao et~al., 2015]{miao2015identification}
Miao, W., Liu, L., Tchetgen, E.~T., and Geng, Z. (2015).
\newblock Identification, doubly robust estimation, and semiparametric efficiency theory of nonignorable missing data with a shadow variable.
\newblock {\em arXiv preprint arXiv:1509.02556}.
\bibitem[Rosenbaum, 1987]{rosenbaum1987sensitivity}
Rosenbaum, P.~R. (1987).
\newblock Sensitivity analysis for certain permutation inferences in matched observational studies.
\newblock {\em Biometrika}, 74(1):13--26.
\bibitem[Rosenbaum, 2004]{rosenbaum2004design}
Rosenbaum, P.~R. (2004).
\newblock Design sensitivity in observational studies.
\newblock {\em Biometrika}, 91(1):153--164.
\bibitem[Rosenbaum, 2010]{rosenbaum2010design}
Rosenbaum, P.~R. (2010).
\newblock Design sensitivity and efficiency in observational studies.
\newblock {\em Journal of the American Statistical Association}, 105(490):692--702.
\bibitem[Tan, 2006]{tan2006distributional}
Tan, Z. (2006).
\newblock A distributional approach for causal inference using propensity scores.
\newblock {\em Journal of the American Statistical Association}, 101(476):1619--1637.
\bibitem[Xu et~al., 2020]{xu2020applications}
Xu, L., Gotwalt, C., Hong, Y., King, C.~B., and Meeker, W.~Q. (2020).
\newblock Applications of the fractional-random-weight bootstrap.
\newblock {\em The American Statistician}, 74(4):345--358.
\bibitem[Zhao et~al., 2019]{zhao2019sensitivity}
Zhao, Q., Small, D.~S., and Bhattacharya, B.~B. (2019).
\newblock Sensitivity analysis for inverse probability weighting estimators via the percentile bootstrap.
\newblock {\em Journal of the Royal Statistical Society Series B: Statistical Methodology}, 81(4):735--761.
\end{thebibliography}