Causal inference for social network data
Extracted source text · arXiv 1705.08527
%% LyX 2.1.3 created this file. For more info, see http://www.lyx.org/.
%% Do not edit unless you really know what you are doing.
\documentclass[english]{article}
\usepackage[T1]{fontenc}
\usepackage[latin9]{inputenc}
\usepackage{geometry}
\geometry{verbose,tmargin=3cm,bmargin=3cm,lmargin=3cm,rmargin=3cm}
\usepackage{array}
\usepackage{multirow}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{graphicx}
\usepackage{setspace}
\usepackage{esint}
\usepackage[authoryear]{natbib}
\onehalfspacing
\makeatletter
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% LyX specific LaTeX commands.
%% Because html converters don't know tabularnewline
\providecommand{\tabularnewline}{\\}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% User specified LaTeX commands.
\usepackage{amsthm}
\usepackage{thmtools}\usepackage{latexsym}
\newtheorem{definition}{}\newtheorem{result}{}\newtheorem{theorem}{}\newtheorem{corollary}{}\newtheorem{lemma}{}\newtheorem{proposition}{}
\declaretheorem[name=Example,qed={$\Box$}]{example}
\declaretheorem[name=Remark,qed={$\Box$}]{remark}
\newcommand{\nl}{\newline}
\newcommand{\pl}{\parallel}
\newcommand{\openr}{\hbox{${\rm I\kern-.2em R}$}}
\newcommand{\openn}{\hbox{${\rm I\kern-.2em N}$}}
\newcommand{\argmax}[1]{\underset{#1}{\operatorname{argmax}}\;}
\newcommand{\argmin}[1]{\underset{#1}{\operatorname{argmin}}\;}
\newcommand{\norm}[1]{\left\lVert#1\right\rVert}
\DeclareMathOperator{\E}{\mathbb{E}}
\usepackage{babel}
\makeatother
\usepackage{babel}
\begin{document}
\title{Causal inference for social network data}
\author{Elizabeth L. Ogburn\thanks{Department of Biostatistics, Johns Hopkins Bloomberg School of Public
Health, Baltimore, MD, USA}, Oleg Sofrygin\thanks{Department of Biostatistics, University of California, Berkeley, CA,
USA}, Iván Díaz\thanks{Division of Biostatistics and Epidemiology, Weill Cornell Medicine,
New York, NY, USA}, and Mark J. van der Laan$^{\dagger}$ }
\date{~}
\maketitle
\begin{abstract}
We extend recent work by van der Laan (2014) on causal inference for
causally connected units to more general social network settings.
Our asymptotic results allow for dependence of each observation on
a growing number of other units as sample size increases. We are not
aware of any previous methods for inference about network members
in observational settings that allow the number of ties per node to
increase as the network grows. While previous methods have generally
implicitly focused on one of two possible sources of dependence among
social network observations, we allow for both dependence due to contagion,
or transmission of information across network ties, and for dependence
due to latent similarities among nodes sharing ties. We describe estimation
and inference for causal effects that are specifically of interest
in social network settings.
\end{abstract}
\section{Introduction and Background\label{sec:Introduction}}
Social networks have long been of interest to sociologists \citep{moreno1937sociometry}
and in the past decade they have gained prominence across a broad
range of disciplines, from economics to epidemiology \citep{knoke2008social}.
A social network is, roughly, a set of individuals who share some
kind of social relationships with one another -- e.g. friends, neighbors,
family members, or coworkers. If one individual's outcome can be affected
by his or her social contacts' outcomes, then we say that the outcome
exhibits \textit{induction} or \textit{contagion}, and the causal
effects are called \textit{peer effects}. If an individual's outcome
may be affected by his or her contacts' treatments of exposures, then
those treatments or exposures are said to exhibit \textit{interference}.
Many aspects of social networks are of interest to researchers, from
the clustering of individuals into communities to the probability
distributions that describe the generation of new relationships between
individuals in the network. There is increasing interest in identifying
and estimating causal effects in the contexts of social networks,
that is causal effects that one individual's behavior, treatment assignment,
beliefs, or health outcome could have on his or her social contacts'
behaviors, exposures, beliefs, or health statuses. There have been
a number of high profile articles that use standard methods like generalized
linear models (GLM) and generalized estimating equations (GEE) to
attempt to infer causal peer effects from network data (e.g. \citealp{christakis2007spread,christakis2008collective,christakis2010social}),
and this work has inspired several research programs that study peer
effects using the same statistical methods \citep{ali2010social,cacioppo2009alone,madan2010social,rosenquist2010spread,wasserman2013comment}.
However, these methods have come under considerable criticism from
the statistical community \citep{cohen2008obesity,lyons2011spread,shalizi2011homophily}.
These statistical models are not equipped to deal with dependence
across individuals and are rarely appropriate for estimating effects
using network data \citep{ogburn2014vaccines}. In some settings it
may be possible to use them to test for the presence of network dependence,
but some properties of such tests are unknown \citep{vanderweele2012statistics,shalizi2012statistics,ogburn2014vaccines}.
Spatial autoregressive (SAR) models have been applied to the study
of peer effects and induction in network settings (e.g. \citealp{goetzke2008network,lee2004asymptotic,lin2005peer,o2008analysis}).
Because the endogenous and exogenous variables are measured at the
same time, they parameterize an equilibrium state rather than causal
relationships. Few data generating processes give rise to true equilibrium
states \citep{besag1977spatial,lauritzen2002chain,thomas2013comment};
therefore SAR models may often be misspecified or uninformative about
causal relationships.
Very recently, researchers interested in causal inference for interconnected
subjects have begun to develop methods designed specifically for the
network setting. Many methods for interference--the effect of one
individual's treatment or exposure on others' outcomes--are relevant
to the analysis of network data \citep{aronow2012estimating,athey2016exact,Bowers2013interference,eckles2014design,graham2010measuring,halloran1995causal,halloran2011causal,hong2006evaluating,hong2008causal,hudgens2008toward,rosenbaum2007interference,rubin1990application,sobel2006randomized,tchetgen2010causal,vanderweele2010direct,vanderweele2011bounding,vanderweele2011effect}.
However, the inferential methods developed in this context generally
require observing multiple independent groups of units, which corresponds
to observing multiple independent networks, or else they require treatment
to be randomized. Ideally, we would like to be able to perform inference
even when all observations are sampled from a single social network
and in observational settings in addition to randomized experiments.
Methodology has not kept apace with interest in causal inference using
data from individuals connected in a social network, and researchers
have resorted to using standard statistical methods to analyze this
new type of data, or to collecting multiple independent groups of
individuals and using the groups, rather than the individuals, as
the unit of statistical inference. The former strategy is not statistically
appropriate and the the latter strategy, while statistically valid,
can be an inefficient use of data. Furthermore, it limits the settings
and effects that one can study, because in some settings it may be
too expensive or labor intensive to collect data from many independent
groups or independent groups may not exist (e.g. a global infectious
disease epidemic).
In this paper we extend recent work by \citet{vanderlaan2012} on
causal inference for causally connected units to more general social
network settings. \citet{vanderlaan2012} introduced methods for causal
inference from a single collection of interconnected units when each
unit is known to be independent of all but a small number of other
units. Asymptotic results rely on the number of dependent units being
fixed as the total number of units goes to infinity. We propose new
methods that allow similar causal and statistical inference without
requiring the number of dependent units to be fixed as sample size
increases. We are not aware of any previous methods for inference
about network members in observational settings that allow the number
of ties per node to increase as the network grows. As we discuss in
Section \ref{sub:Asymptotics}, below, this is a crucial feature of
most realistic models for social network generation. While previous
methods have generally implicitly focused on one of two possible sources
of dependence among social network observations, we allow for both
dependence due to contagion or transmission of information across
network ties, and dependence due to latent similarities among nodes
sharing ties. We describe estimation and inference for causal effects
that are specifically of interest in social network settings.
In Section (\ref{sec:setting}) we give some background on causal
inference for social network data, discussing briefly the relationship
between causal structural equation models and network edges, the types
of statistical dependence likely to be found in social network data,
and asymptotic growth. In Section (\ref{sub:SEM}) we present our
target of inference and the identifying assumptions that we will use
in the methods that follow. We present the efficient influence function
for our target parameter under the conditional independence assumptions
from \citet{vanderlaan2012}. When these independence assumptions
are relaxed, this will still be an influence function for our target
parameter but it may not be efficient. In Section (\ref{sub:Estimation})
we describe estimation procedures--using the (efficient) influence
function from (\ref{sub:SEM})--which will be efficient under the
stronger independence assumptions but still consistent and asymptotically
normal under the weaker independence assumptions. In Section (\ref{sub:Asymptotic-normality})
we prove our main result, which is the asymptotic normality of our
estimator under an asymptotic regime in which the number of ties per
node grows unboundedly with $n$. In Section (\ref{sec:Extensions})
we discuss estimation of rather esoteric causal effects that are specifically
of interest in social network settings. Section (\ref{sec:Simulations})
includes simulation results, and Section (\ref{sec:Conclusion}) concludes.
\section{Background and setting\label{sec:setting}}
\subsection{Networks and structural equation models}
A network is a collection of units and information about the presence
or absence of pairwise ties between them. The presence of a tie between
two units indicates that the units share some kind of a relationship;
what types of relationships are encoded by network ties depends on
the context. For example, in a social network we might define a tie
to include familial relatedness, friendship, or shared place of work.
In the study of sexually transmitted diseases, such as HIV, the sexual
contact network is of great interest; here ties represent sexual partnership
within a given time frame. Some types of relationships are mutual,
for example familial relatedness and shared place of work. Others,
like friendship, can go in only one direction: Tom may consider Sue
to be his friend, while Sue does not consider Tom to be her friend.
Edges can be binary (present or absent), multiplex (categorical with
different levels for different types of relationships), or weighted
(with continuous or ordinal weights) according to the strength of
the relationship.
Networks are most often studied through the lens of graph theory.
Units are represented by \textit{nodes} and ties are represented by
\textit{edges}. Two nodes connected by an edge are called neighbors.
A node whose characteristics we wish to explain or model is called
an \textit{ego}; the ego's neighbors are its \textit{alters}. If ties
must be mutual then the graph is \textit{undirected}; if a tie can
go in only one direction then the graph is \textit{directed}. For
simplicity we will assume all networks are undirected in what follows,
but our methods are equally applicable to directed networks. In an
undirected network, the \textit{degree} of a node is the number of
edges it has or, equivalently, the number of alters. The graph theoretic
approach to networks is immensely powerful and useful. It has been
used to study all manner of network structures, network generating
models, and diffusion processes on networks (see \citealp{newman2009networks}
and references therein for examples). However, this approach largely
reduces networks to topological features, and it is not well suited
for statistical and causal inquiries into mechanisms of diffusion
and influence across network nodes.
Underlying inquiries into causal effects across network nodes is a
representation of the network as a structural equation model. Consider
a network of $n$ subjects, indexed by $i$, with binary undirected
ties $A_{ij}\equiv I\left\{ \mbox{subjects }i\mbox{ and }j\mbox{ share a tie}\right\} $.
The matrix $\mathbf{A}$ with entries $A_{ij}$ is the adjacency matrix
for the network. Associated with each subject is a vector of random
variables, $O_{i}$, including an outcome $Y_{i}$, covariates $C_{i}$,
and an exposure or treatment variable $X_{i}$, all possibly indexed
by time $t$. In numerous applications across the social, political,
and health sciences, researchers are interested in ascertaining the
presence of and estimating causal interactions across alter-ego pairs.
Is there interference, i.e. does the treatment of subject $i$ have
a causal effect on the outcome of subject $j$ when $i$ and $j$
share a network tie? Is there peer influence, i.e. does the outcome
of subject $i$ at time $t$ have a causal effect on a future outcome
of subject $j$ when $i$ and $j$ are adjacent in the network? These
inquiries can be formalized with the help of a causal structural equation
model, informed by the network.
A structural equation model is a system of equations of the form $y_{i}=f_{i}\left[pa_{i}(Y),\varepsilon_{i}\right]$,
where $pa_{i}(Y)$, the set of parents of $Y$, is a collection of
variables that are causes of $Y$ for subject $i$, and $\varepsilon_{i}$
is an error term that may include omitted causes of $Y$. In general
$C_{i}$ and $X_{i}$ will be included in $pa_{i}(Y)$. See \citet{pearl2000causality}
for further discussion of SEMs. When causal inference is performed
on network data, the network ties inform which variables are to be
included in $pa_{i}(Y)$. For example, if interference might be present,
then the collection of treatment variables for $i$'s alters, $\left\{ X_{j}:A_{ij}=1\right\} $,
must be included in the set $pa_{i}(Y)$. If contagion might be present
then $\left\{ Y_{j,t-k}:A_{ij}=1\right\} $ must be included in the
set $pa_{i}(Y_{t})$, where $t$ indexes time and $k$ is an outcome-specific
lag time such that no causal effect can be transmitted from one person
to another in less than $k$ time steps (see \citealp{ogburn2013}
for discussion of lag times).
It is important that the network be completely and accurately specified;
missing ties are akin to missing components of a multidimensional
treatment vector because they result in important elements of exposure
of interest being left out of the SEM. Whenever an inquiry into causal
effects is informed by a social network, measurement error in the
network is tantamount to measurement error in the exposure of interest.
The network for which data is collected must be calibrated to the
causal question of interest. If we are interested in peer effects
on academic achievement among elementary school children and think
that being in the same classroom is the relationship that determines
whether or not two children affect one another's outcomes, then being
in the same classroom is the relationship that determines whether
or not a network tie exists, and a network that captures interaction
during playground sports is not informative or useful. In other words,
a tie between nodes $i$ and $j$ represents the possibility of a
causal effect of an element of $O_{i}$ on an element of $O_{j}$
at a later time, and vice versa. These issues have not been made explicit
in much of the existing literature on causal inference for network
data; equating a network with the underlying SEM can help to make
them precise.
\subsection{Networks and dependence}
Perhaps the greatest challenge and barrier to causal and statistical
inference using observations from a single, interconnected social
network is dependence among observations. The literature on statistics
for dependent data is vast and multifaceted, but very little has been
written on the dependence that arrises when observations are sampled
from a single network. Most of the literature on dependent random
variables assumes that the domain from which observations are sampled
(e.g. time or geographic space) has an underlying Euclidean topology.
The principles behind asymptotic results in the Euclidean dependence
literature are simple and intuitive. They rely on a combination of
stationarity assumptions, i.e. assumptions that certain features of
the data generating process do not depend on an observation's location
in the sample domain, and assumptions that bound the nature and the
amount of dependence in the data. Most frequently these are mixing
assumptions, which describe the decay of the correlation between observations
as a function of the distance between them. Sometimes the stronger
assumption of m-dependence is made, according to which two observations
are independent if they are sampled from locations that are $m$ or
more units apart. Intuitively, in order to extract an increasing amount
of information from a growing sample of dependent observations, old
observations must be predictive of new observations, which is ensured
by stationarity assumptions, and the amount of independence in the
sample must grow faster than the amount of dependence, which is ensured
by mixing conditions or m-dependence.
This literature is not immediately applicable to the network setting.
Roughly, this is due to the difference between Euclidean and network
topology. While it is possible to embed a network in $\mathbb{R}^{d}$
in such a way that preserves distances, to do so is to allow $d$
to increase as $n$ increases. Euclidean dependence results generally
require $d$ to be fixed, implying that, as new observations are sampled
at the boundary of a Euclidean domain, the average and maximum pairwise
distance between observations increases. Networks, on the other hand,
often do not have a clear boundary to which we can add observations
in such a way that ensures growth in the sample domain. In a large
sample with Euclidean dependence, most observations will be distant
from most other observations. This is not necessarily the case in
networks. The maximum distance between two nodes can be small even
in very large networks, and even if the maximum distance between two
nodes is large, there may be many nodes that are close to one another.
Even under m-dependence with small $m$, networks exist in which most
observations remain less than $m$ units apart as $n\rightarrow\infty$.
Therefore, mixing conditions and m-dependence do not necessarily result
in more independence than dependence in a large sample from a network.
Research indicates that social networks generally have the small-world
property (sometimes referred to as the ``six degrees of separation''
property), meaning that the average distance between two nodes is
small \citep{watts1998collective}. Therefore distances in real-world
networks may grow slowly with sample size. Of course some types of
networks, e.g. lattices, embed in $\mathbb{R}^{d}$ as $n$ grows,
but these are generally trivial cases that are not useful for naturally
occurring networks like social networks.
Dependence in networks is of two varieties, each with its own implications
for inference. In the literature on spatial and temporal dependence,
dependence is often implicitly assumed to be the result of latent
traits that are more similar for observations that are close in Euclidean
distance than for distant observations. This type of dependence is
likely to be present in many network contexts as well. In networks,
edges present opportunities to transmit traits or information, and
this direct transmission is an important additional source of dependence
that depends on the underlying network structure.
Latent variable dependence will be present in data sampled from a
network whenever observations from nodes that are close to one another
are more likely to share unmeasured traits than are observations from
distant nodes. Homophily, or the tendency of people who share similar
traits to form network ties, is a paradigmatic example of latent variable
dependence. If the outcome under study in a social network has a genetic
component, then we would expect latent variable dependence due the
fact that family members, who share latent genetic traits, are more
likely to be close in social distance than people who are unrelated.
If the outcome were affected by geography or physical environment,
latent variable dependence could arise because people who live close
to one another are more likely to be friends than those who are geographically
distant. Of course, these traits can create dependence whether they
are latent or observed. But if they are observed then conditioning
on them renders observations independent; therefore the methodological
challenges are greater when they are latent. Just like in the spatial
and temporal dependence context, there is often little reason to think
that we could identify, let alone measure, all of these sources of
dependence. In order to make any progress towards valid inference
in the presence of latent trait dependence, some structure must be
assumed, namely that the range of influence of the latent traits is
primarily local in the network and that any long-range effects are
negligible. In a structural equation model, latent trait dependence
would be captured by dependence among the error terms across subjects.
Whenever one subject's treatments, outcomes, or covariates affect
other subjects' treatments, outcomes, or covariates, observations
across subjects are dependent. However, this kind of dependence, which
arrises from causal effects between subjects, has structure lacking
in latent trait dependence. Figure 1 depicts contagion in a network
of three individuals. This diagram is the directed acyclic graph representation
\citep{pearl1995causal,ogburn2013} of the following structural equation
model: At each time $t$, $Y_{i}^{t}$ is affected by $i$'s own past
outcomes and those of $i$'s social contacts. Individual 2 shares
ties with 1 and 3 but individuals 1 and 3 are not connected. This
structure implies conditional independences: because any transmission
from individual 1 to 3 must pass through 2 we have that $Y_{1}^{t-2}\perp Y_{3}^{t}\mid Y_{2}^{t-1}$;
because information cannot be transmitted instantaneously we have
that $Y_{1}^{t-2}\perp Y_{2}^{t-2}$ . If observations are observed
at closely spaced time intervals then these conditional independences
can be harnessed for inference. There is no reason to think that any
such conditional independences would hold with latent variable dependence.
\begin{figure}
\protect\protect\caption{Dependence due to direct transmission}
\centering{}\includegraphics[scale=0.5]{contagionDAG}
\end{figure}
In this paper, we accommodate both dependence due to direct transmission
and dependence due to latent traits. We assume that both kinds of
dependence are limited to \textit{dependence neighborhoods} determined
by the underlying social network: each subject, or node, $i$ can
directly transmit to information, outcomes, or exposures to the nodes
with which $i$ shares a network tie, and each node $i$ can share
latent traits with the nodes with which $i$ shares a network tie
or a mutual connection. That a subject can only transmit to his or
her immediate social contacts may be a reasonable assumption (indeed,
we can define network ties in such a way as to make this true by definition),
but it is likely unrealistic to assume that latent variable dependence
only affects nodes at a distance of one or two ties, as we assume
throughout. This represents a first step towards valid statistical
and causal inference under more realistic assumptions than have been
required by previous work, but future work is needed to address more
realistic--i.e. longer range--forms of latent variable dependence.
\subsection{A note on asymptotic growth\label{sub:Asymptotics}}
There are many complex issues surrounding asymptotic growth of networks
(e.g. \citealp{diaconis2007graph,shalizi2011consistency}), and a
large literature on graph limits \citep{lovasz2012large}. These issues
are largely beyond the scope of this paper, but we believe that our
methods are consistent with the large and realistic class of network-generating
models.In particular, observed social networks and models proposed
for generating social networks tend to have heavy-tailed degree distributions,
with most nodes having low degree but a non-trivial proportion of
nodes having high degree, with the maximum degree dependent on the
size of the network. Some researchers speculate that the heavy right
tails of social network degree distributions tend to approximately
follow power laws: $Pr(degree=k)\sim k^{-\alpha}$ for $2<\alpha<3$
\citep{barabasi1999emergence,lovasz2012large,newman2003social}, in
which case $Pr(degree>K)=O(K^{1-\alpha})$ for any fixed $K$. Even
if degree distributions depart from power law distributions \citep{clauset2009power}
they are frequently incompatible with the assumption of bounded degrees,
which has been used in previous methods for inference about observations
sampled from a single social network. Our new methods are not able
to accommodate the most highly connected nodes from a power law degree
sequence, but they can nevertheless be used to perform inference about
the other nodes in a network that has a power law degree distribution
(see Section \ref{sec:hubs}).
Our theoretical results require an asymptotic regime in which the
number of nodes in the network, $n$, goes to infinity. Formalizing
asymptotic growth of network-generating models is beyond the scope
of this paper; we take for granted sequence of networks with increasing
$n$ such that key features of the network--whatever features inform
the causal effects of interest--are preserved. The structural equation
model that specifies the distributions of covariates, treatment, and
outcome does not change as $n$ increases.
\section{Methods\label{sec:methods}}
In this section we describe estimation of and inference about the
causal effect of a treatment or exposure, $X$, which can depend upon
covariates but is otherwise independent of outcomes and network topology.
This case includes randomized and non-randomized exposures that may
or may not be subject to interference, specifically direct interference
\citep{ogburn2013}, where the effect of one individual's exposure
on another's outcome is unmediated by the first individual's outcome.
(In Section \ref{sec:Extensions} we discuss causal effects that correspond
to interference due to contagion, where the effect of one individual's
exposure on another's outcome \textit{is} mediated by the first individual's
outcome, and allocational interference, where treatment changes network
ties.) The approach we describe below is different from traditional
approaches to interference in that they are justified when partial
interference does not hold. As far as we are aware, this is the first
approach to interference that references an asymptotic regime in which
the number of ties for a given individual may grow with sample size.
The estimating procedure that we describe in this section was first
proposed by \citet{vanderlaan2012}, but we generalize the results
to a broader class of causal effects and to more general and pervasive
forms of dependence among observations. The conditions under which
the resulting estimators are consistent and asymptotically Normally
distributed are different and weaker here than those in \citet{vanderlaan2012}.
For the remainder of Section \ref{sec:methods}, we describe CAN estimators
of causal effects under two different sets of assumptions. One set
of assumptions allows dependence due to direct transmission but not
latent variable dependence, as in \citet{vanderlaan2012}; under this
set of assumptions our estimators inherit the efficiency properties
from \citet{vanderlaan2012}. The other set of assumptions allows
dependence due to direct transmission and latent variable dependence;
under this set of assumptions our estimators are CAN but may not be
efficient. Our main result is the proof of asymptotic normality under
an asymptotic regime in which the number of ties for a given individual
may grow with sample size in Section \ref{sub:Asymptotic-normality}.
In Section \ref{sec:inference} we describe statistical inference
for the estimators introduced in Section \ref{sec:methods}. We consider
two different classes of estimators: estimators that marginalize over
baseline covariates and estimators that condition on baseline covariates.
In some cases, variance estimation is facilitated by conditioning
on covariates. Under the assumptions encoded in the structural equation
model in Section \ref{sub:SEM}, the conditional estimator is in fact
consistent for the marginal estimand. However, conditional estimators
have smaller variance and inference about the conditional estimand
cannot be interpreted as inference about the marginal estimand. All
of our estimands and estimators condition on the observed network
as given by the adjacency matrix $\mathbf{A}$. Table 1 summarizes
the relationships among the two sets of assumptions (with and without
latent variable dependence) and the two classes of estimators (marginal
over $\mathbf{C}$ and conditional on $\mathbf{C}$) according to
their properties and according to the limitations of our proposed
methods.
\begin{table}
\protect\caption{Properties of marginal estimands and of estimands conditional on $\mathbf{C}$}
\begin{tabular}{|c||c|c|}
\hline
\multirow{2}{*}{\textbf{Properties that we have demonstrated for the two classes of
estimands}} & \multicolumn{2}{c|}{\textbf{Estimand class}}\tabularnewline
\cline{2-3}
& Marginal & Conditional\tabularnewline
\hline
\hline
nonparametrically identified with or without latent variable (LV)
dependence & yes & yes\tabularnewline
\hline
estimator is CAN with or without LV dependence & yes & yes\tabularnewline
\hline
efficient estimator is available with LV dependence & no & no\tabularnewline
\hline
efficient estimator is available without LV dependence & yes & yes\tabularnewline
\hline
consistent and computationally tractable variance estimation with
LV dependence & no & yes\tabularnewline
\hline
consistent and computationally tractable variance estimation without
LV dependence & yes & yes\tabularnewline
\hline
\end{tabular}
\end{table}
In \ref{sec:Extensions} we describe more elaborate interventions
and causal questions that can be addressed by our methods and that
may be of interest specifically in the context of social network data.
However, we focus throughout on single treatment. Longitudinal interventions
are also possible but we leave the details for future work. In addition,
we state our results under the assumption that all variables take
values on discrete sets. Analogous results are valid for other types
of random variables: it is straightforward to extend our notation
and results to continuous covariates and outcomes, but continuous
treatments are more complicated (see \citealp{vanderlaan2012}).
\subsection{Structural equation model\label{sub:SEM}}
Recall that $C_{i}$ denotes a vector of covariates, $X_{i}$ is a
treatment variable, and $Y_{i}$ is the outcome for subjects $i\in\{1,\ldots,n\}$.
In addition, let $K_{i}=\sum_{j=1}^{n}A_{ij}$, that is, $K_{i}$
is degree of node $i$ or the number of individuals sharing a tie
with individual $i$. The degree of subject $i$ and the degrees of
$i$'s alters may be included in the covariate vector $C_{i}$. We
define $\mathbf{Y}=(Y_{1},...,Y_{n})$ and $\mathbf{C}$ and $\mathbf{X}$
analogously. We use a structural equation model to define the causal
effects of interest, as in Section \ref{sec:setting}, but note that
analogous definitions may be achieved within the potential outcome
framework \citep{vanderlaan2012}.
We assume that the data are generated by sequentially evaluating the
following set of equations:
\begin{align}
C_{i} & =f_{C}\left[\varepsilon_{C_{i}}\right] & i=1,\ldots,n\nonumber \\
X_{i} & =f_{X}\left[\left\{ C_{j}:A_{ij}=1\right\} ,\varepsilon_{X_{i}}\right] & i=1,\ldots,n\nonumber \\
Y_{i} & =f_{Y}\left[\left\{ X_{j}:A_{ij}=1\right\} ,\left\{ C_{j}:A_{ij}=1\right\} ,\varepsilon_{Y_{i}}\right] & i=1,\ldots,n,\label{eq:original SEM}
\end{align}
where $f_{C}$, $f_{X}$, and $f_{Y}$ are unknown and unspecified
functions and $\varepsilon_{i}=(\varepsilon_{C_{i}},\varepsilon_{X_{i}},\varepsilon_{Y_{i}})$
is a vector of exogenous, unobserved errors for individual $i$. This
set of equations encodes the assumptions of direct interference in
an observational or randomized setting: the treatment $X$ may or
may not depend on $C$, depending on the specification of $f_{X}$,
and the outcome $Y$ depends on $X$ and $C$ but not on previous
outcomes or on $\mathbf{A}$, except possibly through $C$. Time ordering
is a fundamental component of a structural causal model. For example,
we assume that $C$ is first drawn for all units, so that, in addition
to $C_{i}$, the other components of the vector $C$ may affect the
value of $X_{i}$.
In addition, nonparametric identification of causal effects requires
the following assumptions on the error terms from the SEM:
\begin{align}
& (\varepsilon_{X_{1}},...,\varepsilon_{X_{n}})\perp(\varepsilon_{Y_{1}},...,\varepsilon_{Y_{n}})\mid\mathbf{C},\tag{A1}\label{eq:independent errors}\\
& \varepsilon_{X_{1}},...,\varepsilon_{X_{n}}\mid\mathbf{C}\mbox{ are identically distributed and \ensuremath{\varepsilon_{Y_{1}},...,\varepsilon_{Y_{n}}\mid\mathbf{C},\mathbf{X}} are identically distributed,}\tag{A2a}\label{eq:(X,Y) iid}\\
& \varepsilon_{X_{i}}\perp\varepsilon_{X_{j}}\mid\mathbf{C}\mbox{ and }\varepsilon_{Y_{i}}\perp\varepsilon_{Y_{j}}\mid\mathbf{C},\mathbf{X}\mbox{ for }i,j\mbox{ s.t. \ensuremath{A_{ij}=0} and \ensuremath{\exists!k} with }A_{ik}=A_{kj}=1\tag{A2b}\label{eq:(X,Y)ind}\\
& \varepsilon_{C_{i}},i=1,...,n\mbox{, are identically distributed, and}\tag{A3a}\label{eq:C identically distributed}\\
& \varepsilon_{C_{i}}\perp\varepsilon_{C_{j}}\ensuremath{\mbox{ for }i,j\mbox{ s.t. \ensuremath{A_{ij}=0} and \ensuremath{\ensuremath{\exists!k}} with }A_{ik}=A_{kj}=1.}\tag{A3b}\label{eq:independent C errors}
\end{align}
Assumption (\ref{eq:independent errors}) is a \textit{no unmeasured
confounding} assumption; it ensures that $\mathbf{C}$ suffices to
control for confounding of the effect of $\mathbf{X}$ on $\mathbf{Y}$.
It implies that any latent variable dependence affects $\mathbf{X}$
and $\mathbf{Y}$ separately; in general a latent variable that affected
$\mathbf{X}$ and $\mathbf{Y}$ jointly would constitute a violation
of this assumption. Assumptions (\ref{eq:(X,Y)ind}) and (\ref{eq:independent C errors})
ensure that any unmeasured sources of dependence--i.e. latent trait
dependence--only affect pairs of observations up to a distance of
two network ties--that is, friends or friends-of-friends. Assumption
(A3) can be omitted if attention is restricted to causal effects conditional
on $\mathbf{C}$.
Although our main result, given in Theorem 1 below, holds for inference
in the SEM defined by assumptions (\ref{eq:independent errors})--(\ref{eq:independent C errors}),
some asymptotic properties are guaranteed only when stronger versions
of assumptions (\ref{eq:(X,Y)ind}) and (\ref{eq:independent C errors})
hold. We therefore introduce alternative assumptions
\begin{align}
& \varepsilon_{X_{1}},...,\varepsilon_{X_{n}}\mid\mathbf{C}\mbox{ are i.i.d. and \ensuremath{\varepsilon_{Y_{1}},...,\varepsilon_{Y_{n}}\mid\mathbf{C},\mathbf{X}} are i.i.d.},\mbox{ and}\tag{A4}\label{eq:iid x and y}\\
& \varepsilon_{C_{i}},i=1,...,n\mbox{, are i.i.d.}\tag{A5}\label{eq:iid c}
\end{align}
Note that, although the variance-covariance structure of the SEM given
in (\ref{eq:original SEM}) is affected by the dependence allowed
in (A2b) and (A3b), the mean structure is unaltered by the choice
of assumptions (A2) and (A3) or (A4) and (A5). Therefore, any estimator
that is unbiased under (A4) and (A5) will remain unbiased when these
are relaxed to (A2) and (A3). In Section \ref{sub:identification}
we discuss nonparametric identification of causal parameters, which
is agnostic to the choice of the weaker or stronger independence assumptions.
In Section \ref{sub:Estimation} we derive estimators under assumptions
(\ref{eq:independent errors}), (\ref{eq:iid x and y}), and (\ref{eq:iid c})--that
is, in the presence of dependence due to direct transmission but not
latent variable dependence. We use the stronger assumptions because
the resulting model is amenable to familiar tools for deriving semiparametric
estimators. In Section (\ref{sub:Asymptotic-normality}) we prove
that the estimators derived under assumptions (\ref{eq:independent errors}),
(\ref{eq:iid x and y}), and (\ref{eq:iid c}) are CAN under the weaker
set of assumptions (\ref{eq:independent errors})--(\ref{eq:independent C errors}).
In Section \ref{sec:inference} we discuss inference under each of
the two sets of assumptions.
\subsection{Definition and nonparametric identification of causal effects\label{sub:identification}}
In principle it is possible to perform statistical inference in the
model defined by assumptions (\ref{eq:independent errors})-(\ref{eq:independent C errors})
or by assumptions (\ref{eq:independent errors}), (\ref{eq:iid x and y}),
and (\ref{eq:iid c}). However, in practice, with finite data and
finite computing resources, we may need to make dimension-reducing
assumptions on the forms of $f_{X}$ and $f_{Y}$. This is done by
considering summary functions $s_{X}$ and $s_{Y}$ and random variables
$W_{i}=s_{X,i}\left(\left\{ C_{j}:A_{ij}=1\right\} \right)$ and $V_{i}=s_{Y,i}\left(\left\{ C_{j}:A_{ij}=1\right\} ,\left\{ X_{j}:A_{ij}=1\right\} \right)$
such that the model may be written as
\begin{align*}
C_{i} & =f_{C}\left[\varepsilon_{C_{i}}\right] & i=1,\ldots,n\\
X_{i} & =f_{X}\left[W_{i},\varepsilon_{X_{i}}\right] & i=1,\ldots,n\\
Y_{i} & =f_{Y}\left[V_{i},\varepsilon_{Y_{i}}\right] & i=1,\ldots,n.
\end{align*}
For example, $s_{X,i}\left(\left\{ C_{j}:A_{ij}=1\right\} \right)=\left(C_{i},\sum_{j:A_{ij}=1}C_{j}\right)$
implies that the treatment status of unit $i$ only depends on $i$'s
own covariate value and on the sum of the covariate values of the
units sharing a tie with $i$. Analogously, $s_{Y,i}\left(\left\{ C_{j}:A_{ij}=1\right\} ,\left\{ X_{j}:A_{ij}=1\right\} \right)=\left(C_{i},\sum_{j:A_{ij}=1}C_{j},X_{i},\sum_{j:A_{ij}=1}X_{j}\right)$
is an example of a summary function for $f_{Y}$. For convenience
we use the notation $s_{X,i}(\mathbf{C})$ and $s_{Y,i}(\mathbf{C},\mathbf{X})$
below; however, this notation should not undermine the important fact
that $W_{i}$ can only depend on the subset and $\left\{ C_{j}:A_{ij}=1\right\} $
and $V_{i}$ can only depend on the subsets $\left\{ C_{j}:A_{ij}=1\right\} $
and $\left\{ X_{j}:A_{ij}=1\right\} $ of $\mathbf{C}$ and $\mathbf{X}$,
as these are the only components of $\mathbf{C}$ and $\mathbf{X}$
that are parents of $\mathbf{X}$ and $\mathbf{Y}$, respectively,
in the causal model defined by the network-as-structural-causal-model.
For notational convenience, in what follows we augment the observed
data random vector with $V_{i}$ and $W_{i}$, recognizing that these
are deterministic functionals of $C_{i}$ and $X_{i}$, defined by
$s_{Y,i}$ and $s_{X,i}$, and are therefore technically redundant:
$O_{i}=(C_{i},W_{i},X_{i},V_{i},Y_{i})$.
Potential or counterfactual outcomes can be defined in terms of hypothetical
interventions to the underlying structural causal model. For example,
a hypothetical intervention that deterministically sets $X_{i}$ to
a user-given value $x_{i}^{*}$ for $i=1,...,n$ is given by
\begin{align*}
C_{i} & =f_{C}\left[\varepsilon_{C_{i}}\right] & i=1,\ldots,n\\
X_{i} & =x_{i}^{*} & i=1,\ldots,n\\
Y_{i}(\mathbf{x}^{*}) & =f_{Y}\left[V_{i}(\mathbf{x}^{*}),\varepsilon_{Y_{i}}\right] & i=1,\ldots,n,
\end{align*}
where $\mathbf{x}^{*}=(x_{1}^{*},\ldots,x_{n}^{*})$. Here $Y_{i}(\mathbf{x}^{*})$
denotes the potential outcome of individual $i$ in a hypothetical
world in which $P(\mathbf{X}=\mathbf{x}^{*})=1$. Analogously, $V_{i}(\mathbf{x}^{*})=s_{Y,i}(\mathbf{C},\mathbf{x}^{*})$
is a counterfactual random variable in a hypothetical world in which
$P(\mathbf{X}=\mathbf{x}^{*})=1$\textbf{. }Note that, although $V_{i}(\mathbf{x}^{*})$
is counterfactual, its value is determined by the observed realization
of $\mathbf{C}$ and by the user-specified value $\mathbf{x}^{*}$,
and it is therefore known. In order to streamline notation as we describe
increasingly complex interventions, we denote the counterfactual variables
$V_{i}(\mathbf{x}^{*})$ and $Y_{i}(\mathbf{x}^{*})$ by $V_{i}^{*}$
and $Y_{i}^{*}$, respectively. The causal parameter of interest throughout
is the expected average potential outcome in this same hypothetical
world, i.e. $E\left[\bar{Y}^{*}\right],$ where $\bar{Y}^{*}=\frac{1}{n}\sum_{i=1}^{n}Y_{i}^{*}$.
We are now ready define notation that we will use throughout the remainder
of the paper for functionals of the distribution of $\mathbf{O}$.
Let $p_{C}(\mathbf{c})=P\left(\mathbf{C}=\mathbf{c}\right)$, $g(\mathbf{x}|\mathbf{w})=P\left(\mathbf{X}=\mathbf{x}\mid\mathbf{W}=\mathbf{w}\right)$,
$g_{i}(x|w)=P\left(X_{i}=x|W_{i}=w\right)$, $p_{Y}(\mathbf{y}|\mathbf{v})=P\left(\mathbf{Y=y}|\mathbf{V=v}\right)$,
and $p_{Y,i}(y|v)=P\left(Y_{i}=y|V_{i}=v\right)$. Define the two
marginal distributions $h_{i}(v)=P\left(V_{i}=v\right)$ and $h_{i,x^{*}}(v)=P\left(V_{i}^{*}=v\right)$,
noting that both $h_{i}$ and $h_{i,x^{*}}$ are determined by $g$
and $p_{C}$ and are therefore observed data quantities. Finally,
$m(v)=\sum_{y}y\,p_{Y}(y|v)$ is the conditional expectation of $Y$
given $V=v$.
In addition to assumptions (\ref{eq:independent errors})-(\ref{eq:independent C errors})
or (A1), (A4), and (A5), identification of $E\left[\bar{Y}^{*}\right]$
requires the positivity assumption that
\begin{align}
P(V=v|\mathbf{C}=\mathbf{c})>0 & \mbox{\,\,\ for all \ensuremath{v} in the range of \ensuremath{V_{i}^{*}} and for all \,}\mathbf{c}\mbox{ such that }P(\mathbf{C}=\mathbf{c})>0.\tag{A6}\label{eq:positivity}
\end{align}
This assumption states that, within levels of $\mathbf{C}$, the values
of $V$ determined by the hypothetical intervention $\mathbf{x}^{*}$
have positive probability under the observed-data-generating distribution.
Now the causal parameter $E\left[\bar{Y}^{*}\right]$ is identified
by
\begin{equation}
\psi=\frac{1}{n}\sum_{i=1}^{n}E\left[m(V_{i}^{*})\right]=\frac{1}{n}\sum_{i=1}^{n}\sum_{v}m(v)h_{i,x^{*}}(v).\label{eq:identification form 1}
\end{equation}
This identification result is equivalent to
\begin{equation}
\psi=\frac{1}{n}\sum_{i=1}^{n}\sum_{\mathbf{c}}m(s_{Y,i}(\mathbf{c},\mathbf{x}^{*}))p_{C}(\mathbf{c}).\label{eq:identification form 2}
\end{equation}
From \eqref{eq:identification form 2}, it is clear that the conditional
causal parameter $E\left[\bar{Y}^{*}\mid\mathbf{C}=\mbox{\textbf{c}}\right]$
is identified by $\frac{1}{n}\sum_{i=1}^{n}m(s_{Y,i}(\mathbf{c},\mathbf{x}^{*}))$.
\subsection{Estimation\label{sub:Estimation}}
Estimation of and inference about $E\left[\bar{Y}^{*}\right]$ requires
a statistical model $\mathcal{M}$ for the distribution of the observed
data $P(\mathbf{O})$. That is, $\mathcal{M}$ is a collection of
distributions over $\mathbf{O}$ of which one element is the true
data-generating distribution. Our target of inference is a pathwise
differentiable mapping $\Psi:\mathcal{M}\rightarrow\mathbb{R}$ such
that $\psi$ is $\Psi(P)$, the mapping evaluated at the true data-generating
distribution. Under assumptions (A1), (A4), and (A5) the probability
distribution of the observed data may be factorized as
\begin{equation}
P\left(\mathbf{O}=\mathbf{o}\right)=P\left(\mathbf{C}=\mathbf{c}\right)g(\mathbf{x}|\mathbf{w})p_{Y}(\mathbf{y}|\mathbf{v}),\label{eq:factorization}
\end{equation}
suggesting that $\mathcal{M}$ requires three components: a model
for $p_{C}$, a model for $g$, and a model for $P\left(Y|V\right)$.
Furthermore, the identification results in (\ref{eq:identification form 1})
and (\ref{eq:identification form 2}) indicate that $\psi$ depends
on $P(Y|V)$ only through $m$. The empirical distribution $\hat{p}_{C}$
can be used throughout to nonparametrically estimate $p_{C}$, but,
when $\mathbf{C}$ is high-dimensional, $g$ and $m$ cannot be non-parametrically
estimated at rates of convergence that are fast enough to satisfy
the regularity conditions of Theorem 1 (see Appendix). Therefore,
in order to define the parameter mapping we require a statistical
model $\mathcal{M}=\mathcal{M}_{g}\times\mathcal{M}_{m}$, where $\mathcal{M}_{g}$
is a collection of conditional distributions for $X$ given $W$ such
that the true conditional distribution is a member, and $\mathcal{M}_{m}$
is a collection of expectations of $Y$ relative to conditional distributions
of $Y$ given $V$ such that the true conditional expectation of $Y$
given $V$ is a member. Estimation of $\psi$ is based on the efficient
influence function for the parameter mapping $\Psi:\mathcal{M}\rightarrow\mathbb{R}$.
Under assumptions (A1), (A4), and (A5), the efficient influence function,
$D$, evaluated at a fixed value $\mathbf{o}$ of $\mathbf{O}$, is
given by
\begin{equation}
D(\mathbf{o})=\sum_{j=1}^{n}\frac{1}{n}\sum_{i=1}^{n}E\left[m\left(V_{i}^{*}\right)\mid C_{j}=c_{j}\right]-\psi+\frac{1}{n}\sum_{i=1}^{n}\frac{\bar{h}_{x^{*}}(v_{i})}{\bar{h}(v_{i})}\left\{ y_{i}-m\left(v_{i}\right)\right\} ,\label{eq:EIF}
\end{equation}
where $\bar{h}(v_{i})=\frac{1}{n}\sum_{j=1}^{n}h_{j}(v_{i})$, $\bar{h}_{x^{*}}(v_{i})=\frac{1}{n}\sum_{j=1}^{n}h_{j,x^{*}}(v_{i})$,
$v_{i}=s_{Y,i}(\mathbf{c},\mathbf{x})$, and $V_{i}^{*}=s_{Y,i}(\mathbf{C},\mathbf{x}^{*})$.
The influence function has expected value equal to $0$ at the true
$\psi$; this fact can be used to generate unbiased estimating equations
for $\psi$. Estimating equations based on the efficient influence
function are doubly robust: the right hand side of Equation (\ref{eq:EIF})
has expected value equal to $0$ if $m(\cdot)$ is replaced with an
arbitrary functional of $V$ or if $g(\cdot)$ is replaced with an
arbitrary functional of $W$, as long as one of the two remains correctly
specified. (Recall that $g(\cdot)$, along with $p_{C}$, determines
$\bar{h}_{x^{*}}(v_{i})$ and $\bar{h}(v_{i})$.) This implies that
an estimating equation based on Equation (\ref{eq:EIF}) will be unbiased
for $\psi$ if either model $\mathcal{M}_{m}$ for $m(\cdot)$ or
model $\mathcal{M}_{g}$ for $g(\cdot)$ is correctly specified, i.e.
contains the truth, even if one is not. This influence function is
efficient in that, when $m(\cdot)$ is correctly specified, it has
the smallest variance among all influence functions in model $\mathcal{M}_{g}$.
The efficient influence function in a model that does not make any
distributional assumptions about $\mathbf{C}$, that is under assumptions
(A1) and (A4) only, is given in equation (\ref{eq:EIF-1}) below\textbf{.
\begin{equation}
D'(\mathbf{o})=\dfrac{1}{n}\sum_{i=1}^{n}\left(E\left[m\left(V_{i}^{*}\right)\mid\mathbf{C}=\mathbf{c}\right]-\psi+\frac{\bar{h}_{x^{*}}(v_{i})}{\bar{h}(v_{i})}\left\{ y_{i}-m\left(v_{i}\right)\right\} \right).\label{eq:EIF-1}
\end{equation}
}We use this influence function in the estimating procedure below.
This is also the influence function used to derive estimators conditional
on $\mathbf{C}$, in which case the first two terms cancel out; we
will denote the conditional influence function wit $D_{c}(\mathbf{o})$.
In social network settings $\mathbf{C}$ is likely to be high-dimensional
-- recall that $C_{i}$ includes all relevant covariates for individual
$i$ and for $i$'s social contacts, in addition to a list of network
ties. Even though we assume that $m(\cdot)$ and $g(\cdot)$ depend
on $\mathbf{C}$ only through the dimension-reducing functions $s_{Y}(\cdot)$
and $s_{X}(\cdot)$, if $m(\cdot)$ and $g(\cdot)$ are estimated
nonparametrically care should be taken to ensure rates of convergence
satisfying the regularity conditions of Theorem 1. When $m(\cdot)$
and $g(\cdot)$ are estimated with working models, the doubly robust
property is advantageous, affording two independent opportunities
for valid inference about $\psi$. If models for both $m(\cdot)$
and $g(\cdot)$ are correctly specified, then the solution to estimating
equations based on the efficient influence function are asymptotically
efficient in the semiparametric model defined by the identifying assumptions
and by the correctly specified model for $g(\cdot)$. See \citet{van2003unified,tsiatis2007semiparametric}
for a review of semiparametric inference in i.i.d. settings.
A traditional estimating equation approach to inference about $\psi$
has two disadvantages in finite samples. First, it does not result
in substitution estimator, leading to the possibility that finite-sample
estimates will fall outside of the parameter space when the parameter
space for $\psi$ is bounded. Second, it is sensitive to extreme values
of the term $\bar{h}_{i,x^{*}}(v)/\bar{h}(v_{i})$ in the righthand
side of Equation (\ref{eq:EIF}). For these reasons, we propose a
targeted maximum loss-based estimator (TMLE) of $\psi$; TMLEs are
substitution estimators and are not as sensitive to the near violations
of the positivity assumption that can occur in finite samples and
result in extreme values of $\bar{h}_{x^{*}}(v_{i})/\bar{h}(v_{i})$.
Targeted maximum likelihood estimation is a general template for estimation
of smooth parameters in semi- and nonparametric models. The estimation
algorithm is constructed to solve the efficient influence function
estimating equation, thereby yielding, under regularity conditions,
asymptotically linear estimators with the same semiparametric efficiency
property as the estimating equation approach described above. In our
setting, a TMLE is constructed using three elements: (i) a valid loss
function $L$ for the outcome regression model $m$, (ii) initial
working estimators $\hat{m}$ of $m$ and and $\hat{g}$ of $g$,
and (iii) a parametric submodel $m_{\epsilon}$ of $\mathcal{M}$,
the score of which corresponds to a particular component of the score
based on the efficient influence function $D(\mathbf{o})$ and such
that $m_{0}=m(\cdot)$. The TMLE is then defined by an iterative procedure
that, at each step, estimates $\epsilon$ by minimizing the empirical
risk of the loss function $L$ at $m_{\epsilon}$. An updated estimate
is then computed as $\hat{m}_{\hat{\epsilon}}$, and the process is
repeated until convergence. The TMLE is the estimator obtained in
the final step of the iteration. The result of the previous iterative
procedure is that, at the final step, the efficient influence function
estimating equation is solved. For more details about targeted maximum
likelihood estimation, see \citet{van2011targeted}. Below we describe
a TMLE for $\psi$ based on \eqref{eq:EIF-1} that requires only one
iteration for convergence. We use influence function $D'(\mathbf{o})$
to derive the TMLE, instead of $D(\mathbf{o})$, because it is computationally
more tractable and because the choice of influence function does not
matter for the conditional parameter that we are interested in when
latent variable dependence is present.
In order to use targeted maximum likelihood estimation to estimate
$\psi$, the parameter identifying $E\left[\bar{Y}^{*}\right]$, initial
estimators $\hat{m}$ and $\hat{g}$ of $m$ and $g$ may be found
through standard maximum likelihood or loss-based estimation methods.
For example, if $Y_{i}$ is binary, or if it is continuous and bounded
and rescaled to fall between $0$ and $1$, then the negative log
likelihood function
\[
\sum_{i=1}^{n}\log\left\{ m(V_{i})^{Y_{i}}(1-m(V_{i}))^{1-Y_{i}}\right\}
\]
is a valid loss function for $m$. Assuming a parametric model $m_{\theta}(\cdot)$
for $m(\cdot)$, a maximum likelihood estimator is given by $m_{\hat{\theta}}$,
where
\[
\hat{\theta}=\arg\max_{\theta}\sum_{i=1}^{n}\log\left\{ m_{\theta}(V_{i})^{Y_{i}}(1-m_{\theta}(V_{i}))^{1-Y_{i}}\right\}
\]
may be estimated by running standard binary regression software of
the outcomes $Y_{i}$ on $V_{i}$. An initial estimator $\hat{g}$
may be found analogously by optimizing the log likelihood function
$\sum_{i=1}^{n}\log g(X_{i}|W_{i})$. Alternatively, these loss functions
may be used to implement other loss-based estimation techniques such
as super learning \citep{van2007super}, which selects the weighted
average of a collection (``library'') of estimators that minimizes
the cross-validated loss-based risk; the library may include any machine
learning algorithms and logistic regression estimators. The empirical
distribution $\hat{p}_{C}$ is used to estimate $p_{C}$.
Estimation of $\bar{h}$ and $\bar{h}_{x}$ can be carried out by
substituting $\hat{g}$ and $\hat{p}_{C}$ for $g$ and $p_{C}$ in
the expressions
\begin{align*}
\bar{h}(v)= & \frac{1}{n}\sum_{j}\sum_{\mathbf{c},\mathbf{x}}I(s_{Y,j}(\mathbf{c},\mathbf{x})=v)g(\mathbf{x}|\mathbf{w})p_{C}(\mathbf{c})\mbox{ and}\\
\bar{h}_{x^{*}}(v)= & \frac{1}{n}\sum_{j}\sum_{\mathbf{c}}I(s_{Y,j}(\mathbf{c},\mathbf{x}^{*})=v)p_{C}(\mathbf{c}),
\end{align*}
where $I(E)$ is the indicator function of the event $E$. We denote
by $\hat{\bar{h}}$ and $\hat{\bar{h}}_{x^{*}}$ the corresponding
estimates of $\bar{h}$ and $\bar{h}_{x^{*}}$.
An alternative and computationally more feasible approach is to estimate
$\bar{h}(v)$ by directly optimizing the log likelihood function $\sum_{i=1}^{n}\log\bar{h}(V_{i}|W_{i})$,
as if the pooled sample $(V_{i},W_{i})$ were i.i.d. It can be shown
that this results in a valid loss function for $\bar{h}$, even for
dependent observations $(V_{i},W_{i})$, for $i=1,\ldots,n$. Similarly,
one can construct a direct estimator of $\bar{h}_{x^{*}}$, by first
creating a sample $(V_{i}^{*},W_{i})$ and then directly optimizing
the log likelihood function $\sum_{i=1}^{n}\log\bar{h}_{x^{*}}(V_{i}^{*}|W_{i})$,
as if the pooled sample $(V_{i}^{*},W_{i})$ were i.i.d.
Now the targeted minimum loss based estimator of $\psi$ is computed
as follows:
\begin{enumerate}
\item Define the auxiliary weights $H_{i}$ as the ratio of estimated densities
of $V^{*}$ and $V$ evaluated at the observed value $V_{i}$. Compute
the auxiliary weights as
\[
H_{i}=\frac{\hat{\bar{h}}_{x^{*}}(V_{i})}{\hat{\bar{h}}(V_{i})}.
\]
\item Compute initial predicted outcome values $\hat{Y}_{i}=\hat{m}(V_{i})$
and predicted potential outcome values $\hat{Y}_{i}^{*}=\hat{m}(V_{i}^{*})$
evaluated at the counterfactual value $V_{i}^{*}=s_{Y,i}(\mathbf{C},\mathbf{x}^{*})$.
\item Construct a TMLE model update $\hat{m}_{\hat{\epsilon}}$ of $\hat{m}$
by running a weighted intercept-only logistic regression model with
weights $H_{i}$ defined in step (1), $Y_{i}$ as the outcome and
including $\hat{Y}_{i}$ as an offset. That is, define $\hat{\epsilon}$
as the estimate of the intercept parameter $\epsilon$ from the following
\emph{weighted} logistic regression model
\[
\mbox{logit}\hat{m}_{\epsilon}(v)=\mbox{logit}\hat{m}(v)+\epsilon,
\]
where $\mbox{logit}(x)=\log\left(\frac{x}{1-x}\right)$.
\item Compute updated predicted potential outcomes $\tilde{Y_{i}}^{*}$
as the fitted values of the regression from step 3, evaluated at $v^{*}$
rather than $v$ (that is, at $\hat{Y}_{i}^{*}$ instead of $\hat{Y}_{i}$):
\[
\tilde{Y_{i}}^{*}=\mbox{expit}\{\mbox{logit}\hat{Y}_{i}^{*}+\hat{\epsilon}\},
\]
where $\mbox{expit}(x)=\frac{1}{1+e^{-x}}$, i.e., the inverse of
the $\mbox{logit}$ function.
\item Compute the TMLE $\hat{\psi}$ as
\[
\hat{\psi}=\frac{1}{n}\sum_{i=1}^{n}\tilde{Y_{i}}^{*}.
\]
\end{enumerate}
The TMLE inherits the double robustness property of the estimating
equation estimator we described above: it will be consistent for $\psi$
if either the working model $\hat{g}$ for $g$ or the working model
$\hat{m}$ for $m$ is correctly specified. This resulting estimator
remains CAN for $\psi$ under assumptions (A2) and (A3) instead of
(A4) and (A5), and the same procedure can be used to estimate the
parameter conditional on $\mathbf{C}$.
\subsection{Asymptotic normality\label{sub:Asymptotic-normality}}
\citet{vanderlaan2012} proved the asymptotic normality of the estimators
described above under assumptions that the degree $K_{i}$ is uniformly
bounded above as $n\rightarrow\infty$ by a constant $K$. In order
to accommodate more realistic models of asymptotic growth in the network
context, we consider a new asymptotic regime in which $K_{i}$ may
grow as $n\rightarrow\infty$.
\textbf{Theorem 1:\label{Theorem 1} }\textit{Let $K_{max,n}=max_{i}\{K_{i}\}$
for a fixed network with $n$ nodes. Suppose that $K_{max,n}^{2}/n\rightarrow0$
as $n\rightarrow\infty$. If at least one of the two models for $m(\cdot)$
and $g(\cdot)$ is correctly specified and under independence assumptions
(\ref{eq:independent errors}) through (\ref{eq:independent C errors}),
positivity assumption (\ref{eq:positivity}), and regularity conditions
(see Appendix),
\begin{align*}
& \sqrt{C_{n}}\left(\hat{\psi}-\psi\right)\overset{d}{\longrightarrow}N(0,\sigma^{2}),
\end{align*}
$n/K_{max,n}^{2}\leq C_{n}\leq n$. The asymptotic variance of $\hat{\psi}$,
$\sigma^{2}$, is given by the variance of the influence curve of
the estimator.}
The proof of Theorem 1 is in the Appendix. In Section \ref{sec:hubs},
below, we discuss settings in which the conditions for this theorem
fail to hold, and ways to recover valid inference for conditional
estimands in some of these settings. Broadly, the proof has two parts:
first, to show that the second order terms in the expansion of $\hat{\psi}-\psi$
are stochastically less than $1/\sqrt{C_{n}}$, and second, to show
that the first order terms converge to a normal distribution when
scaled by a factor of order $\sqrt{C_{n}}$. The proof that the second
order terms are stochastically less than $1/\sqrt{C_{n}}$ is an extension
of the empirical process theory of \citet{van1996weak} and follows
the same format as the proof in \citet{vanderlaan2012}.\textbf{ }For
the proof that the first order terms converge to a normal distribution,
we rely on Stein's method of central limit theorem proof \citep{stein1972bound}.
Stein's method allows us to derive a bound on the distance between
our first order term (properly scaled) and a standard normal distribution;
this bound depends on the degree distribution $K_{1},...,K_{n}$.
We show that this bound converges to $0$ as $n\rightarrow\infty$
under regularity conditions and our running assumption that $K_{max,n}^{2}=o(n)$.
When all nodes have the same number of ties, i.e. $K_{i}=K_{max,n}$
for all $i$, then the rate of convergence will be given \textit{$\sqrt{C_{n}}=\sqrt{n/K_{max,n}^{2}}$}.
When $K_{max,n}$ is bounded above as \textit{$n\rightarrow\infty$},
then the rate of convergence will be $\sqrt{n}$. This is the setting
covered in \citet{vanderlaan2012}. When $K_{max,n}\rightarrow\infty$
but some nodes have fewer than $K_{max,n}$ ties, the exact rate of
convergence is between $\sqrt{n/K_{max,n}^{2}}$ and $\sqrt{n}$ (inclusive)
but is difficult or impossible to determine analytically, as it may
depend intricately on the topology of the network. The inferential
procedures that we describe below do not require knowledge of the
rate of convergence.
\subsection{Inference}
\label{sec:inference}
A 95\% confidence interval for $\psi$ is given by $\hat{\psi}_{n}\pm1.96\sigma/\sqrt{C_{n}}$.
In practice neither $\sigma$ nor $C_{n}$ are likely to be known,
but available variance estimation methods estimate the variance of
$\hat{\psi}_{n}$ directly, incorporating the rate of convergence
without requiring it to be known a priori.
In i.i.d settings, the variance of doubly robust estimators can be
bootstrapped or, under the assumption that both nuisance models are
correctly specified, estimated with the plug-in estimator of the square
of the estimator's influence function. When the nuisance parameters
are estimated with parametric models, standard sandwich variance estimators
are appropriate whether one or both of the nuisance models are correctly
specified. In the present, non-i.i.d. setting, variance estimation
is considerably more challenging. In principle, the variance of $\hat{\psi}$
can be estimated using the empirical average of the square of the
influence function, substituting $\hat{\psi}$ for $\psi$ and the
fitted values from the working models $\hat{g}$ and $\hat{m}$ for
$g$ and $m$. Although this variance may be anticonservative if one,
but not both, of the working models $\hat{g}$ and $\hat{m}$ is correctly
specified, using flexible or non-parametric specifications for these
models increases opportunities to estimate both consistently. However,
unlike in i.i.d. settings, the expectation of the square of the empirical
version of the influence function given in Equation \eqref{eq:EIF}
does not reduce to the sum of squared influence terms for each observation.
Instead, it includes double sums for all pairs of observations that
are not marginally independent of one another. These terms capture
covariances between dependent observations; these extra covariance
terms reflect a larger variance and a slower rate of convergence due
to dependence across observations.
When dependence is due to direct transmission, that is, under assumptions
(A1), (A4), and (A5), two alternative variance estimation procedures
are available. One option is to\textbf{ }estimate the variance of
the influence function $D'(\mathbf{o})$ given by Equation \eqref{eq:EIF-1}.
Our TMLE is based on $D'(\mathbf{o})$, but because this is the efficient
IF in a model that makes fewer assumptions than (A1), (A4), and (A5),
it has larger variance than $D(\mathbf{o})$ and provides a valid
(asymptotically conservative) variance estimate even when estimation
is based on $D(\mathbf{o})$. For consistent and computationally feasible
estimators for the variance of $D'(\mathbf{o})$ see \citet{sofryginTechreport}.
An alternative approach to estimate the variance of $\hat{\psi}$
under assumptions (A1), (A4), and (A5) is to employ the following
version of a parametric bootstrap. This approach might offer improvements
in finite-sample performance over the previously described approach.
Briefly, our proposed procedure proceeds by iteratively sampling $n$
observations from the existing fit of the likelihood, fitting the
univariate least favorable submodel parameter $\epsilon$, and computing
the bootstrapped TMLE. After enough iterations one can obtain the
Monte Carlo variance estimator of $\hat{\psi}$ by evaluating the
empirical variance of the bootstrap TMLEs. In more detail, for each
of $M$ bootstrap iterations, indexed as $b=1,\ldots,M$, first $n$
covariates $\mathbf{C}^{b}=(C_{1}^{b},\ldots,C_{n}^{b})$ are sampled
with replacement (assuming $\mathbf{C}$ are i.i.d.), then the existing
model fit $\hat{g}$ is applied to sampling of $n$ exposures $\mathbf{X}^{b}=(X_{1}^{b},\ldots,X_{n}^{b})$,
followed by a sample of $n$ outcomes $\mathbf{Y}^{b}=(Y_{1}^{b},\ldots,Y_{n}^{b})$
based on the existing outcome model fit $\hat{m}$. Note that we are
also assuming that the corresponding bootstrap random summaries $W_{i}^{b}$
and $V_{i}^{b}$, for $i=1,\ldots,n$, were constructed by applying
the summary functions $s_{X}$ and $s_{Y}$ to $\mathbf{C}^{b}$ and
$(\mathbf{C}^{b},\mathbf{X}^{b})$, respectively. This bootstrap sample
is then applied to obtain the predicted values from the existing auxiliary
covariate fit $(\hat{\bar{h}}_{x^{*}}/\hat{\bar{h}})(V_{i}^{b})$,
for $i=1,\ldots,n$, followed by a bootstrap-based fitting of $\epsilon$,
and finally, evaluation of bootstrap TMLE. Note that the TMLE model
update is the only model fitting step needed at each iteration of
the bootstrap, which significantly lowers the computational burden
of this procedure. The variance estimate is then obtained by taking
the empirical variance of bootstrap TMLE samples $\hat{\psi}^{b}$.
In the present setting, due to dependence across observations, one
must be judicious with applications of the bootstrap. For example,
the parametric bootstrap procedure described above requires conditional
independence of $X_{i}$ given $W_{i}$ and $Y_{i}$ given $V_{i}$,
along with the consistent modeling of the corresponding factors of
the likelihood. It may seem natural to sample $V_{i}$ directly from
its corresponding auxiliary model fit, but this is likely to result
in an anti-conservative variance estimates, since the conditional
independence of $V_{i}$ is unlikely to hold by virtue of its construction
as a summary measure of the network.
When latent variable dependence is present, that is under assumptions
(A1) through (A3), consistent and computationally feasible variance
estimation procedures are not currently available for either $D'(\mathbf{o})$
or $D(\mathbf{o})$, because existing methods require bootstrapping
some of the observed data. Without latent variable dependence we can
take advantage of marginal and conditional independences to employ
i.i.d. or parametric bootstrap methods, but latent variable dependence
requires new methods for dependent data bootstrap. For this reason,
we instead estimate the conditional parameter with influence function
$D_{C}(\mathbf{o}).$ A simple plug-in estimator is available for
the variance of this influence function (see the Appendix and \citealp{vanderlaan2012}).
\section{Extensions\label{sec:Extensions}}
In this section we extend the estimation procedure to two causal effects
of great interest in the context of social networks: to social contagion,
or peer effects; to stochastic interventions; and to interventions
on the network topology itself, i.e. interventions on $\mathbf{A}=\left[A_{ij}:i,j\in\{1,\ldots,n\}\right]$
where, as above, $A_{ij}\equiv I\left\{ \mbox{subjects }i\mbox{ and }j\mbox{ share a tie}\right\} $.
First we introduce dynamic and stochastic interventions.
\subsection{Dynamic and stochastic interventions}
Interventions that assign treatment as a user-specified function $d_{X}(\cdot)$
of $\mathbf{C}$ correspond to substituting $d_{X,i}(\mathbf{C})$
for $x_{i}^{*}$ in the intervention model, definitions, and estimating
procedure above. This is an example of a ``dynamic'' intervention:
treatment is deterministically specified conditional on covariates
but is but allowed to depend (``dynamically'') on covariates.
We can also identify the effects of interventions that replace $f_{X}$
with a new, user-specified function. This is an example of a stochastic
intervention: the intervention changes the distribution of $X$ but
does not eliminate the stochasticity introduced by $\varepsilon_{X}$;
it is represented by an intervention SEM that replaces the equation
for $X_{i}$ with $X_{i}^{*}=r_{X}\left[W_{i}^{*},\varepsilon_{X_{i}}\right]$
for a user specified function $r_{X}$. For discussions of stochastic
interventions and their identifying assumptions in i.i.d. settings,
see \citet{munoz2012population,haneuse2013estimation,young2014identification}.
In the social network setting, stochastic interventions that change
the dependence of $X_{i}$ on \textbf{$\mathbf{C}$ }and of and $Y_{i}$
on $\mathbf{C}$ and $\mathbf{X}$ are of particular interest. For
example, consider data generated by an SEM in which $f_{X}$ depends
on $\mathbf{C}$ only through $W_{i}=\frac{1}{\left|A_{i}\right|}\sum_{j:A_{ij}=1}C_{j}$,
i.e. the mean of $C$ among the set of alters of $i$. We might be
interested in the mean counterfactual outcome under a stochastic intervention
that forces $f_{X}$ to depend instead on $W_{i}^{*}=max_{j:A_{ij}=1}\left\{ C_{j}\right\} $,
i.e. the maximum value $C$ among the alters of $i$. This particular
stochastic intervention modifies $f_{X}$ only through $W$; it is
represented by an intervention SEM that replaces the equation for
$X_{i}$ with $X_{i}^{*}=f_{X}\left[W_{i}^{*},\varepsilon_{X_{i}}\right]$.
This is not a deterministic intervention, rather it assigns $X$ according
to a stochastic distribution: for each $x$ in the support of $X$,
$X_{i}$ is set by the intervention to $x$ with probability $P\left[X=x|W=max_{j:A_{ij}=1}\left\{ C_{j}\right\} \right]$.
We formally define the class of stochastic interventions that alter
the dependence of $X_{i}$ on \textbf{$\mathbf{C}$ }and of and $Y_{i}$
on $\left(\mathbf{C},\mathbf{X}\right)$, discuss identifying assumptions
and estimation procedures, and then describe some such interventions
of particular interest. Let $s_{X,i}^{*}(\cdot)$ and $s_{Y,i}^{*}(\cdot,\cdot)$
be user-specified functionals. They are denoted by an asterisk because
they index hypothetical interventions rather than realized data-generating
mechanisms. Let $W_{i}^{*}=s_{X,i}^{*}(\mathbf{C})$ and $V_{i}^{*}=s_{Y,i}^{*}(\mathbf{C},\mathbf{X}^{*})$.
We are concerned with the class of stochastic interventions given
by
\begin{align}
C_{i} & =f_{C}\left[\varepsilon_{C_{i}}\right] & i=1,\ldots,n\nonumber \\
X_{i}^{*} & =f_{X}\left[W_{i}^{*},\varepsilon_{X_{i}}\right] & i=1,\ldots,n\nonumber \\
Y_{i}^{*} & =f_{Y}\left[V_{i}^{*},\varepsilon_{Y_{i}}\right] & i=1,\ldots,n.\label{eq: intervene on sX, sY}
\end{align}
This can be interpreted as an intervention where, for each $x^{*}$
in the support of $X$ and for $i=1,...,n$, $X_{i}$ is set to $x^{*}$
with probability $P\left[X=x^{*}|W=s_{X,i}^{*}(\mathbf{C})\right]$
and $V_{i}$ is set to $s_{Y,i}^{*}(\mathbf{C},\mathbf{x}^{*})$ deterministically
for each possible realization \textbf{$\mathbf{x}^{*}$}. Because
$Y$ depends on $\mathbf{X}$ only through $V$, this is equivalent
to an intervention that sets $V_{i}$ to $v$ with probability $P\left[\mathbf{X}\in\left\{ \mathbf{x}^{*}:s_{Y,i}^{*}(\mathbf{C},\mathbf{x}^{*})=v\right\} \mid\mathbf{W}=\mathbf{s}_{X}^{*}(\mathbf{C})\right]$,
where $\mathbf{s}_{X}^{*}(\mathbf{C})=\left(s_{X,1}^{*}(\mathbf{C}),...,s_{X,n}^{*}(\mathbf{C})\right)$.
This intervention is identified under the same assumptions as the
deterministic interventions described above, with the exception of
a positivity assumption that is a slight modification of (\ref{eq:positivity}).
Define $\mathcal{X}^{*}=\left\{ \mathbf{x}^{*}:P\left[\mathbf{X}=\mathbf{x}^{*}|\mathbf{W}=\mathbf{s}_{X}^{*}(\mathbf{C})\right]>0\right\} $
to be the set of treatment vectors $\mathbf{x}^{*}$ that have positive
probability under the stochastic intervention defined by (\ref{eq: intervene on sX, sY}).
We assume that
\begin{align}
min_{v\in\mathcal{V}^{*}}P(V=v|\mathbf{C}=\mathbf{c})>0 & \mbox{\,\,\ for \ensuremath{\mathcal{V}^{*}}=\ensuremath{\left\{ s_{Y,i}^{*}(\mathbf{C},\mathbf{x}^{*}):\mathbf{x}^{*}\in\mathcal{X}^{*}\right\} }and for all\,}\mathbf{c}\mbox{ in the support of }\mathbf{C}.\label{eq:positivity stochastic}
\end{align}
Note that, in order for this positivity assumption to hold, the supports
of $s_{X}^{*}(\cdot)$ and $s_{Y}^{*}(\cdot,\cdot)$ must be of the
same dimensions as the supports of $s_{X}(\cdot)$ and $s_{Y}(\cdot,\cdot)$,
respectively.
The causal parameter of interest is the expected average potential
outcome under this hypothetical intervention, $E\left[\bar{Y}^{*}\right]$.
Define $h_{i}^{*}(v)=P\left[V_{i}^{*}=v\right]=P\left[s_{Y,i}^{*}(\mathbf{C},\mathbf{X}^{*})=v\right]$.
Then $E\left[\bar{Y}^{*}\right]$ is identified by
\begin{align*}
\psi & =\frac{1}{n}\sum_{i=1}^{n}\sum_{\mathbf{c},\mathbf{x}}E\left[Y_{i}|s_{Y,i}^{*}(\mathbf{c},\mathbf{x})\right]P\left[\mathbf{X}=\mathbf{x}|\mathbf{W}=\mathbf{s}_{X}^{*}(\mathbf{c})\right]p_{C}(\mathbf{c})\\
& =\frac{1}{n}\sum_{i=1}^{n}E\left[m(V_{i}^{*})\right]=\frac{1}{n}\sum_{i=1}^{n}\sum_{v}m(v)h_{i}^{*}(v).
\end{align*}
An influence function for $\psi$, evaluated at a fixed value of the
observed data, $\mathbf{o}$, is given by
\[
D(\mathbf{o})=\sum_{j=1}^{n}\frac{1}{n}\sum_{i=1}^{n}E\left[m\left(V_{i}^{*}\right)\mid C_{j}=c_{j}\right]-\psi+\frac{1}{n}\sum_{i=1}^{n}\frac{\bar{h}^{*}(v_{i})}{\bar{h}(v_{i})}\left\{ y_{i}-m\left(v_{i}\right)\right\} ,
\]
where $\bar{h}^{*}(v_{i})=\frac{1}{n}\sum_{j=1}^{n}h_{_{j}}^{*}(v_{i})$.
(This is the efficient influence function under assumptions (\ref{eq:iid x and y})
and (\ref{eq:iid c}).) Estimation of $\bar{h}^{*}$ is carried out
by substituting $\hat{g}$ and $\hat{p}_{C}$ for $g$ and $p_{C}$
in the expression
\begin{align*}
\bar{h}^{*}(v)= & \frac{1}{n}\sum_{j}\sum_{\mathbf{c},\mathbf{x}}I\left(s_{Y,i}^{*}(\mathbf{c},\mathbf{x})=v\right)g\left(\mathbf{x}|\mathbf{s}{}_{X}^{*}\mathbf{(c)}\right)p_{C}(\mathbf{c}).
\end{align*}
Since $\hat{p}_{C}$ is an empirical distribution that puts mass one
on the observed value $\mathbf{c}$, the estimator $\hat{\bar{h}}^{*}$
reduces to
\begin{align*}
\hat{\bar{h}}^{*}(v)= & \frac{1}{n}\sum_{j=1}^{n}\sum_{\mathbf{x}}I\left(s_{Y,i}^{*}(\mathbf{x},\mathbf{C})=v\right)\hat{g}(\mathbf{x}|\mathbf{s}_{X}^{*}\mathbf{(C)}).
\end{align*}
We denote by $\hat{\bar{h}}$ and $\hat{\bar{h}}^{*}$ the corresponding
estimates of $\bar{h}$ and $\bar{h}^{*}$.
Now the targeted minimum loss based estimator of $\psi$ is computed
according to the steps outlined in Section \ref{sec:methods}, but
with $V^{*}$ and $Y^{*}$ defined as immediately above.
A special case of this class of stochastic interventions intervenes
only on $s_{X}$, like the example discussed above in which the intervention
forces $f_{X}$ to depend on $W_{i}^{*}=max_{j:A_{ij}=1}\left\{ C_{j}\right\} $
but does not alter the functional form of $s_{Y}$. $E[\bar{Y}^{*}]$
under this type of intervention is identified by
\begin{align*}
\psi & =\frac{1}{n}\sum_{i=1}^{n}\sum_{\mathbf{c},\mathbf{x}}E\left[Y_{i}|\mathbf{C}=\mathbf{c},\mathbf{X}=\mathbf{x}\right]P\left[\mathbf{X}=\mathbf{x}|\mathbf{W}=s_{X}^{*}(\mathbf{c})\right]p_{C}(\mathbf{c})\\
& =\frac{1}{n}\sum_{i=1}^{n}E\left[m(V_{i}^{*})\right]=\frac{1}{n}\sum_{i=1}^{n}\sum_{v}m(v)h_{i}^{*}(v).
\end{align*}
With $V_{i}^{*}$ defined as $s_{Y,i}(\mathbf{C},\mathbf{X}^{*})$,
estimation of this class of intervention proceeds as immediately above.
The fact that $\mathbf{X}^{*}$ is random does not affect the estimation
algorithm.
\subsection{Peer effects}
Define $Y_{i}^{0}$ to be the outcome variable measured at a time
previous to the primary outcome measurement $Y_{i}$. Peer effects
are the class of causal effects of $Y_{j}^{0}$ on $Y_{i}$ for $A_{ij}=1$:
the effects of individuals' outcomes on the subsequent outcome of
their alters. We can operationalize peer effects as the effects of
dynamic interventions where the counterfactual exposure for subject
$i$ is given by a user-specified function $d_{X}(\cdot)$ of $\{Y_{j}^{0}:A_{ij}=1\}$.
In order to maintain the identifying assumptions A2b and A3b, the
time elapsed between $Y^{0}$ and $Y$ must permit transmission only
between nodes and their immediate alters. Otherwise, if between $Y^{0}$
and $Y$ the outcome could have spread contagiously more broadly,
there will be more dependence present than our methods can account
for, and also possible confounding of the effect of $Y_{i}^{0}$ on
$Y_{j}$ for $A_{ij}=1$ due to mutual friends.
\subsection{Interventions on the network topology}
An intervention on the network, i.e. an intervention that adds, removes,
or relocates ties in the network, is a special case of a joint intervention
on $s_{X}(\cdot)$ and $s_{Y}(\cdot)$. To see this, note that the
network topology, codified by the adjacency matrix $\mathbf{A}$,
enters the data-generating structural equation model (\ref{eq:original SEM})
only through $s_{X}(\cdot)$ and $s_{Y}(\cdot)$; therefore we can
represent any modification to $\mathbf{A}$ via the corresponding
modification to $s_{X}(\cdot)$ and $s_{Y}(\cdot)$. Consider an intervention
that replaces the observed adjacency matrix $\mathbf{A}$ with a user-specified
adjacency matrix $\mathbf{A}^{*}$. This is an example of the stochastic
interventions described on page \pageref{eq:positivity stochastic},
with $s_{X,i}^{*}(\mathbf{C})$ replaced by $s_{X,i}^{\mathbf{A^{*}}}(\mathbf{C})\equiv s_{X,i}\left(\left\{ C_{j}:A_{ij}^{*}=1\right\} \right)$
and $s_{Y,i}^{*}(\mathbf{C},\mathbf{X^{*}})$ by $s_{Y,i}^{\mathbf{A^{*}}}(\mathbf{C},\mathbf{X}^{*})\equiv s_{Y,i}\left(\left\{ X_{j}^{*}:A_{ij}^{*}=1\right\} ,\left\{ C_{j}:A_{ij}^{*}=1\right\} \right)$.
The intervention SEM differs from the data-generating SEM only in
that $X_{i}$ depends on the covariate values for the individuals
with whom $i$ shares ties in the intervention adjacency matrix $\mathbf{A}^{*}$
and $Y_{i}$ depends on the counterfactual treatments and observed
covariate values for those same individuals.
Interventions on summary features of the adjacency matrix can also
be viewed as stochastic interventions. Instead of replacing $\mathbf{A}$
with $\mathbf{A}^{*}$, an intervention on features of the network
topology replaces $\mathbf{A}$ with the members of a class $\mathcal{A}^{*}$
of $n\times n$ adjacency matrices that share the intervention features,
stochastically according to some probability distribution $g_{\mathbf{A}^{*}}$
over $\mathcal{A}^{*}$. For example, we might be interested in interventions
that constrain the degree distribution of the network, e.g. fixing
the maximum degree to be smaller than some number $D$ or ensuring
that no node has degree equal to $0$. We might specify $g_{\mathbf{A}^{*}}(A)=\frac{1}{\left|\mathcal{A}^{*}\right|}I\left\{ A\in\mathcal{A}^{*}\right\} $,
giving equal weight to each realization in the class $\mathcal{A}^{*}$.
Effectively, this kind of intervention sets $V_{i}$ to $v$ with
probability $P\left[\mathbf{X}\in\left\{ \mathbf{x}^{*}:s_{Y,i}^{\mathbf{T}^{*}}(\mathbf{C},\mathbf{x}^{*})=v\right\} \mid\mathbf{W}=\mathbf{s}_{X}^{\mathbf{A}*}(\mathbf{C})\mbox{ for some }\mathbf{A}^{*}\in\mathcal{A}^{*}\right]$,
where $\mathbf{s}_{X}^{\mathbf{A}^{*}}(\mathbf{C})=\left(s_{X,1}^{\mathbf{A}^{*}}(\mathbf{C}),...,s_{X,n}^{\mathbf{A}^{*}}(\mathbf{C})\right)$.
As with the stochastic interventions discussed in the previous section,
positivity is a crucial assumption for identifying interventions on
$\mathbf{A}$. The support of $V^{*}$ must be the same as the support
of $V$. If replacing $\mathbf{A}$ with $\mathbf{A}^{*}$ (either
deterministically or as a random selection from the class $\mathcal{A}^{*}$)
assigns to unit $i$ a value of $V$ that not observed in the real
data for a unit in the same $\mathbf{C}$ stratum as $i$, then the
effect of the intervention that that replaces $\mathbf{A}$ with $\mathbf{A}^{*}$
is not identified for unit $i$.
We can expand the arguments of $s_{X}(\cdot)$ and $s_{Y}(\cdot)$
to include network topology explicitly, in addition to $\mathbf{C}$,
or we can include features of network topology as individual-level
covariates. For example, we could include in $C_{i}$ the total number
of ties for subject $i$ has, a measure of subject $i$'s centrality
in the network, or a local clustering coefficient for the neighborhood
comprised of subject $i$'s contacts. Whether or not we can define,
identify, and estimate interventions involving these features of network
topology hinges crucially on the positivity assumption.
Both the degree of subject $i$ and local clustering around subject
$i$ are local features of network topology: they depend on $\mathbf{T}$
only through subject $i$ and subject $i$'s immediate contacts. A
local clustering coefficient for node $i$ can be defined as the proportion
of potential triangles that include $i$ as one vertex and that are
completed, or the number of pairs of neighbors of $i$ who are connected
divided by the total number of pairs of neighbors of $i$ \citep{newman2009networks}.
This measure of triangle completion captures the extent to which ``the
friend of my friend is also my friend'': triangle completion is high
whenever two subjects who share a mutual contact are more likely to
themselves share a tie than are two subjects chosen at random from
the network. Positivity could hold if, within each level of $\mathbf{C}$,
subjects were observed to have a wide range of degrees and of triangle
completion among their contacts.
In contrast with degree and local clustering, network centrality is
a node-specific attribute that nevertheless depends on the entire
network topology. It captures the intuitive notion that some nodes
are central and some nodes are fringe in any given network. It can
be measured in many different ways, based, for example, on the number
of network paths that intersect node $i$, on the probability that
a random walk on the network will intersect node $i$, or on the mean
distance between node $i$ and the other nodes in network (see Chapter
7 of \citealp{newman2009networks} for a comprehensive discussion
of these and other centrality measures). Centrality is given by a
univariate measure for each node in a network, but each node's measure
depends crucially on the entire graph. In reality it is not generally
possible to intervene on centrality without altering the entire adjacency
matrix $\mathbf{A}$, which is likely to result in changes to $s_{X}(\cdot)$
and $s_{Y}(\cdot)$. Unless it is hypothetically possible to intervene
on centrality without changing any other input into the SEM, in this
case $s_{X}(\cdot)$ and $s_{Y}(\cdot)$, interventions on centrality
cannot be identified.
\subsection{Too many friends, too much influence\label{sec:hubs}}
The conditions in the theorem on page \pageref{Theorem 1} will be
violated for any asymptotic regime in which the degree of any one
or more nodes grows at the same or similar rate to the sample size
$n$. This is problematic because social networks frequently have
a small number of ``hubs''--that is, nodes with very high degree
\citep{newman2009networks}, and the occurrence of hubs is a feature
of many of the network-generating models that have been proposed for
social networks. When a small number of individuals wield influence
over a significant portion of the rest of the population, two problems
arise for statistical inference about the people connected in the
underlying network. First, the number of hubs may grow slowly with
$n$ or it could be fixed and not grow at at all. If the hubs are
systematically different from the rest of the population (e.g. they
have different covariates values or the exposure affects them differentially),
then a fixed or slowly growing number of hubs would not allow for
consistent inference about this distinct subpopulation. Second, and
more importantly, the sweeping influence of hubs creates dependence
among all of the influenced nodes that undermines inference. Our methods
rely on the independence of $Y_{i}$ and $Y_{j}$ whenever nodes $i$
and $j$ do not share a tie or a mutual alter. When hubs are present,
a significant proportion of nodes will share a connection to one of
these hubs, undermining our methods.
We can recover valid inference using our methods if we condition on
the hubs, treating them as features of the background network environment
rather than as observations. This results in different causal effects
or statistical estimands, as all of our inference is conditional on
the identity and characteristics of the hubs. Imagine a social network
comprised of the residents of a city in which a cultural or political
leader is connected to almost all of the other nodes. It may be impossible
to disentangle the influence of this leader, which affects every other
node, from other processes simultaneously occurring among the other
residents of the city. It will certainly be impossible to statistically
learn about the hub, as the sample size for the hub subgroup is 1.
But it may make sense to consider the hub as a feature of the city
rather than a member of the network. We could then learn about other
processes occurring among the other residents of the city, conditional
on the behavior and characteristics of the leader. For example, we
could evaluate the effect of a public health initiative encouraging
residents to talk to their friends about the importance of exercise,
but we could not evaluate a similar program targeting the leader's
communication about exercise. It is easy to see that, in the latter
case, positivity would be violated as there would be no control group
of subjects unexposed to the leader's communications. Equally important
is the fact that conditioning on the leader is necessary to engender
the independence among units required for our theoretical results
to hold.
Practically speaking, for real and finite datasets, this implies that
the methods we have proposed are inappropriate for networks in which
the degree is large, compared to $n$, for one or more nodes. If many
nodes are connected to a significant fraction of other nodes, this
problem is intractable. However, if only a small number of nodes are
highly connected we can condition on them to recover approximately
valid inference using our methods for conditional estimands. There
is a theoretical tradeoff between the rate of convergence of our estimators
and the order of $K$ relative to $n$ that, in finite samples, becomes
a practical tradeoff between generality and variance. Increasing the
number of nodes classified as hubs (and conditioned on in subsequent
analysis) will increase the rate of convergence by decreasing the
order of $K$ for the remaining, non-hub nodes -- assuming that the
number of hubs remains small compared to $n$ so that the sample size
does not decrease significantly when we exclude hubs from the analysis.
On the other hand, classifying more nodes as hubs results in analyses
that are increasingly specific: conditioning on a single hub may preserve
generalizability to other networks (similar cities with similar leaders),
but conditioning on many hubs is likely to limit the generalizability
of the resulting inference.
\section{Simulations\label{sec:Simulations}}
We conducted a simulation study that evaluated the finite sample and
asymptotic behavior of the TMLE procedure described in Section \ref{sub:Estimation}.
We generated social networks according to the preferential attachment
model \citep{barabasi1999emergence}, where the node degree (number
of friends) distribution followed a power law with $\alpha=0.5$,
and according to the small world network model \citep{watts1998collective}
with a rewiring probability of $0.1$. We generated data with two
different types of dependence: first with dependence due to direct
transmission only, and second with both latent variable dependence
and dependence due to direct transmission.
Our simulations mimicked a hypothetical study designed to increase
the level of physical activity in a population comprised of members
of a social network. For each community member indexed by $i=1,\ldots,n$,
the study collected data on $i$'s baseline covariates, denoted $C_{i}$,
which included the indicator of being physically active, denoted $PA_{i}$
and the network of friends on each subject, $F_{i}$. The exposure
or treatment, $X_{i}$, was assigned randomly to 25\% of the community.
For example, one can imagine a study where treated individuals received
various economic incentives to attend a local gym. The outcome $Y_{i}$
was a binary indicator of maintaining gym membership for a pre-determined
follow-up period. We assumed that it was of interest to examine and
estimate the average of the mean counterfactual outcomes $E\left[\bar{Y}^{*}\right]$
under various hypothetical interventions $g^{*}$ on such a community.
First, we considered a stochastic intervention $g_{1}^{*}$ which
assigned each individual to treatment with a constant probability
of $0.35$; this differs from the observed allocation of treatment
to $25\%$ of the community members. We also considered a scenario
in which the aforementioned economic incentive was resource constrained
and could only be allocated to up to 10\% of community members and
estimated the effects of various targeted approaches to allocating
the exposure. For example, we considered an intervention $g_{2}^{*}$
that targeted only the top 10\% most connected members of the community,
as such a targeted intervention would be expected to have a higher
impact on the overall average probability of maintaining gym membership
among the community, when compared to purely random assignment of
exposure to 10\% of the community. Another hypothetical intervention
$g_{3}^{*}$ assigned an additional physically active friend to individuals
with fewer than 10 friends. Notably, $g_{3}^{*}$ can be thought of
as intervening on the structure of the social network itself. Finally,
we estimated the combined effect of simultaneously implementing intervention
$g_{2}^{*}$ and the network-based intervention $g_{3}^{*}$ on the
same community; this is $g_{4}^{*}$ . For simplicity, this simulation
study only reports the expected outcome under each of these interventions;
causal effects defined as contrasts of these interventions can be
easily estimated based on the same methods.
All simulation and estimation was carried out in $\mathbf{R}$ language
\citep{r} with packages \emph{$\mathtt{simcausal}$} \citep{R-simcausal}
and \emph{$\mathtt{tmlenet}$} \citep{R-tmlenet}. The full $\mathbf{R}$
code for this simulation study is available in a separate github repository\footnote{github.com/osofr/socialnets.sim.study}.
The simulations were repeated for community sizes of $n=500$, $n=1,000$
and $n=10,000$. The estimation was repeated by sampling $1,000$
such datasets, conditional on the same network (sampled only once
for each sample size). For the simulations with dependence due to
direct transmission,\textbf{ }the baseline covariates were independently
and identically distributed. The probability of success for each $Y_{i}$
was a logit-linear function of $i$'s exposure $X_{i}$ (indicator
of receiving the economic incentive), the baseline covariates $C_{i}$
and the three summary measures of $i$'s friends exposures and baseline
covariates. In particular, we also assumed that the probability of
maintaining gym membership increased on a logit-linear scale as a
function of the following network summaries: the total number of $i$'s
friends who were exposed $(\sum_{j:A_{ij}=1}X_{j})$, the total number
of $i$'s friends who were physically active at baseline $(\sum_{j:A_{ij}=1}PA{}_{j})$
and the product of the two summaries $(\sum_{j:A_{ij}=1}X{}_{j}\times\sum_{j:A_{ij}=1}PA{}_{j})$.
The economic incentive to attend local gym had a small direct effect
on each individual who was not physically active at baseline and no
direct effect on those who were already physically active. However,
physically active individuals were more likely to maintain gym membership
over the follow-up period if they had at least one physically active
friend at baseline. We repeated these simulations with the addition
of latent variable dependence, which we introduced by generating unobserved
latent variables for each node which affected the node's own outcome
as well as the outcomes of its friends.
We estimated the expected outcome under interventions $g_{1}^{*}$
through $g_{4}^{*}$ using the aforementioned TMLE approach and evaluated
its finite sample bias. For the simulations under dependence due to
direct transmission, we estimated the marginal parameter $E\left[\bar{Y}^{*}\right]$
and compared three different estimators of the asymptotic variance
and the coverage of the corresponding confidence intervals. First,
we looked at the naive plug-in i.i.d. estimator (``\emph{IID Var}'')
for the variance of the influence curve which treated observations
as if they were i.i.d. Second, we used the plug-in variance estimator
based on the efficient influence curve which adjusted for the correlated
observations (``\emph{dependent IC Var}''), as previously described
in \citet{sofryginTechreport}. Finally, we used the parametric bootstrap
variance estimator (``\emph{bootstrap Var}''), where each bootstrap
iteration sampled $n$\textbf{ }observations from the previous fit
of the likelihood (the estimated exposure model $\hat{g}$ and the
mean outcome model $\hat{m}$) and used this bootstrap sample to re-fit
the least favorable parametric submodel update $\epsilon$. The simulation
results showing the mean length and coverage of these three CI types
are shown in Figure \ref{fig:CIres.EY.prefattach} (preferential attachment
network) and Figure \ref{fig:CIres.EY.smwld} (small world network).
For the simulations with latent variable dependence, we estimated
the conditional parameter $E\left[\bar{Y}^{*}\right]$ and we compared
two plug in variance estimators based on the conditional influence
function $D_{C}$: one that assumes conditionally i.i.d outcomes (conditional
on $\mathbf{X}$ and $\mathbf{C}$), which would be true if all dependence
were due to direct transmission but is violated in the presence of
latent variable dependence (``\emph{IID Var}''), and one that does
not make this assumption (``\emph{dependent IC Var}'').
We examined the empirical distribution of the transformed TMLEs, comparing
their histogram estimates against the predicted normal limiting distribution,
with the results shown in Figure \ref{fig:hist.TMLE.EY}, where the
histogram plots are displayed by sample size (horizontal axis) and
the intervention type (vertical axis). The estimates were first centered
at the corresponding true parameter values and then re-scaled by their
corresponding true standard deviation (SD). We note that our results
indicate that the estimators converge to their normal theoretical
limiting distribution, even in networks with power law node degree
distribution, such as the preferential attachment network model, as
well as in the densely connected networks obtained under the small
world network model. The results shown in Figure \ref{fig:hist.TMLE.EY}
were generated from simulations with dependence due to direct transmission;
simulations with latent variable dependence (not shown) evinced similar
approximate normality.
One of the lessons of our simulation study is that by leveraging the
structure of the network it might be possible to achieve a larger
overall intervention effect on a population level \citep{harling2016leveraging}.
For example, the results in the left panels of Figures \ref{fig:CIres.EY.prefattach}
and \ref{fig:CIres.EY.smwld} show that by targeting the exposure
assignment to highly connected and physically active individuals,
intervention $g_{2}^{*}$ increases the mean probability of sustaining
gym membership compared to the similar level of un-targeted coverage
of the exposure. We also demonstrated the feasibility of estimating
effects of interventions on the observed network structure itself,
such as intervention $g_{3}^{*}$, which can be also combined with
economic incentives, as it was mimicked by our hypothetical intervention
$g_{2}^{*}+g_{3}^{*}$. These combined interventions could be particularly
useful in resource constrained environments, since they may result
in larger community level effects at the lower coverage of the exposure
assignment.
We evaluated the performance of the different approaches to variance
estimation described above, as measured by the coverage of the 95\%
CIs. Our results from simulations with dependence due to direct transmission
show that conducting inference while ignoring the nature of the dependence
in such datasets generally results in anticonservative variance estimates
and under-coverage of CIs, which can be as low as 50\% even for very
large sample sizes (``\emph{IID Var}'' in the right panels of Figures
\ref{fig:CIres.EY.prefattach} and \ref{fig:CIres.EY.smwld}). The
CIs based on the dependent variance estimates (``\emph{dependent
IC Var}'' in the right panel of the same figures) obtain nearly nominal
coverage of 95\% for large enough sample sizes, but can suffer in
smaller sample sizes due to lack of asymptotic normality and near-positivity
violations. Notably, the CIs based on the parametric bootstrap variance
estimates provide the most robust coverage for smaller sample sizes,
while attaining the nominal 95\% coverage in large sample sizes for
nearly all of the simulation scenarios (``\emph{bootstrap Var}''
in the right panel of the same figures). The apparent robustness of
the parametric bootstrap method for inference in small sample sizes,
even as low as $n=500$, was one of the surprising finding of this
simulation study. Future work will explore the assumptions under which
this parametric bootstrap works and its sensitivity towards violations
of those assumptions. Similarly, in the simulations with latent variable
dependence the variance estimates that assume conditionally i.i.d.
outcomes, i.e. that dependence may be due to direct transmission but
not to latent variables, are anti-conservative.
\begin{figure}
\begin{centering}
\includegraphics[scale=0.8]{knitR-CIres_EY_prefattach_bw-1}
\par\end{centering}
\protect\protect\caption{Mean 95\% CI length (left panel) and coverage (right panel) for the
TMLE in preferential attachment network with dependence due to direct
transmission, by sample size, intervention and CI type. Results are
shown for the estimates of the average expected outcome under four
hypothetical interventions ($g_{1}^{*}$, $g_{2}^{*}$, $g_{3}^{*}$
and $g_{2}^{*}+g_{3}^{*}$).\label{fig:CIres.EY.prefattach}}
\end{figure}
\begin{figure}
\begin{centering}
\includegraphics[scale=0.8]{knitR-CIres_EY_smwld_bw-1}
\par\end{centering}
\protect\protect\caption{Mean 95\% CI length (left panel) and coverage (right panel) for the
TMLE in small world network\textbf{ }with dependence due to direct
transmission, by sample size, intervention and CI type. Results are
shown for the estimates of the average expected outcome under four
hypothetical interventions ($g_{1}^{*}$, $g_{2}^{*}$, $g_{3}^{*}$
and $g_{2}^{*}+g_{3}^{*}$). \label{fig:CIres.EY.smwld}}
\end{figure}
\begin{figure}
\centering{}\includegraphics[scale=0.25]{knitR-hist_TMLE_EY_prefattach-1}\includegraphics[scale=0.25]{knitR-hist_TMLE_EY_smwld-1}\protect\protect\caption{Comparing re-scaled empirical TMLE distributions (black) to their
theoretical normal limit (red) with varying sample size (x-axis) and
intervention type (y-axis). TMLEs were centered at the truth and then
re-scaled by true SD. Results shown for the preferential attachment
network (left) and the small world network (right).\label{fig:hist.TMLE.EY}}
\end{figure}
\begin{figure}
\begin{centering}
\includegraphics[scale=0.8]{COND_DEP_Y_knitR-CIres_EY_prefattach_bw-1}
\par\end{centering}
\protect\protect\caption{Mean 95\% CI length (left panel) and coverage (right panel) for the
TMLE in preferential attachment network with latent variable dependence,
by sample size, intervention and CI type. Results are shown for the
estimates of the average expected outcome under four hypothetical
interventions ($g_{1}^{*}$, $g_{2}^{*}$, $g_{3}^{*}$ and $g_{2}^{*}+g_{3}^{*}$).\label{fig:CIres.EY.prefattach.COND.DEP.Y}}
\end{figure}
\begin{figure}
\begin{centering}
\includegraphics[scale=0.8]{COND_DEP_Y_knitR-CIres_EY_smwld_bw-1}
\par\end{centering}
\protect\protect\caption{Mean 95\% CI length (left panel) and coverage (right panel) for the
TMLE in small world network with latent variable dependence, by sample
size, intervention and CI type. Results are shown for the estimates
of the average expected outcome under four hypothetical interventions
($g_{1}^{*}$, $g_{2}^{*}$, $g_{3}^{*}$ and $g_{2}^{*}+g_{3}^{*}$).
\label{fig:CIres.EY.smwld.COND.DEP.Y}}
\end{figure}
\section{Conclusion\label{sec:Conclusion}}
We proposed new methods that allow for causal and statistical inference
using observations sampled from members of a single interconnected
social network when the observations evince dependence due to network
ties. In contrast to existing methods, our methods do not require
randomization of an exogenous treatment and they have proven performance
under asymptotic regimes in which the number of network ties grows
(slowly) with sample size. In future work we plan to address a key
limitation of the present proposal, namely the assumption that the
network is observed fully and without error. We also plan to develop
data-adaptive methods for estimating the summary measures $s_{X}$
and $s_{Y}$, as it may be unreasonable to expect these to be known
\textit{a priori}. Finally, we plan to develop estimating algorithms
for longitudinal settings; the influence function and asymptotic results
for these settings are straightforward extensions of the results presented
here, but estimation can be challenging.
\section*{Acknowledgements}
Elizabeth Ogburn was supported by ONR grant N000141512343. Oleg Sofrygin
and Mark van der Laan were supported by NIH grant R01 AI074345-07.
\part*{Appendix: Proof of Theorem 1}
\section*{Regularity conditions}
For a real-valued function $\mathbf{c}\mapsto f(\mathbf{c})$, let
the $L^{2}(P)$-norm of $f(\mathbf{c})$ be denoted by $\norm{f}=E[f(\mathbf{C})^{2}]^{1/2}$.
Define \textbf{$\mathcal{M}_{m}$ }and \textbf{$\mathcal{M}_{\tilde{h}}$
}as the classes of possible functions that can be used for estimating
the two nuisance parameters $m$ and $\tilde{h}\equiv\bar{h}_{x^{*}}/\bar{h}$,
respectively. Note that a model for $g$ plus the empirical distribution
of covariates $\mathbf{C}$ determines $\tilde{h}$. Equivalent assumptions
could be stated in terms of $g$ instead of $\tilde{h}$, but we focus
on $\tilde{h}$ because that is the functional of $g$ and $\mathbf{C}$
that we model in our estimating procedure. Assume that the TMLE update
$\hat{m}_{\hat{\epsilon}}\in{\cal M}_{m}$ with probability 1 and
assume that $\hat{\bar{h}}_{x^{*}}/\hat{\bar{h}}\in{\cal M}_{\tilde{h}}$
with probability 1. Finally, define the following dissimilarity measure
on the cartesian product of $\mathcal{F}\equiv{\cal M}_{m}\times{\cal M}_{\tilde{h}}$:
\[
d\left(\left(h,m\right),\left(\tilde{h},\tilde{m}\right)\right)=\max\left(\sup_{v\in{\cal V}}\mid h-\tilde{h}\mid(v),\sup_{v\in{\cal V}}\mid m-\tilde{m}\mid(v)\right).
\]
The following are the regularity conditions required for Theorem 1,
i.e. for asymptotic normality of the TMLE $\hat{\psi}^{*}$.
\begin{description}
\item [{Uniform consistency:}] Assume that
\[
d\left(\left(\hat{\bar{h}}_{x^{*}}/\hat{\bar{h}},\hat{m}_{\hat{\epsilon}}\right),\left(\bar{h}_{x^{*}}/\bar{h},m\right)\right)\rightarrow0
\]
in probability as $n\rightarrow\infty$. Note that this assumption
is only needed for proving the asymptotic equicontinuity of our process;
it is not needed for proofs of relevant convergence rates for the
second order terms.
\item [{Bounded entropy integral:}] Assume that there exists some $\eta>0$,
so that $\int_{0}^{\eta}\sqrt{\log\left(N(\epsilon,{\cal F},d)\right)}d\epsilon<\infty$,
where $N(\epsilon,{\cal F},d)$ is the number of balls of size $\epsilon$
w.r.t. metric $d$ needed to cover ${\cal F}$.
\item [{Universal bound:}] Assume $\sup_{f\in{\cal F},\mathbf{O}}\mid f\mid(\mathbf{O})<\infty$,
where the supremum of $\mathbf{O}$ is over a set that contains $\mathbf{O}$
with probability one. This assumption will typically be a consequence
of the choosing a specific function class ${\cal F}$ that satisfies
the above entropy condition.
\item [{Positivity:}] Assume
\[
\sup_{v\in{\cal V}}\frac{\bar{h}_{x^{*}}(v)}{\bar{h}(v)}<\infty.
\]
\item [{Consistency and rates for estimators of nuisance parameters:}] Assume
that $\norm{\hat{m}-m}\norm{\hat{\bar{h}}-\bar{h}}=o_{P}\left(\left(C_{n}\right)^{-1/2}\right)$.
Note that this rate is achievable if, for example, estimation of $\bar{h}$
relies on some pre-specified parametric model, or if both $\bar{h}$
and $m$ are estimated at rate $C_{n}^{-1/4}$.
\item [{Rate of the second order term:}] Assume that
\[
R_{n1}\equiv-\int_{v}\left\{ \left(\frac{\hat{\bar{h}}_{x^{*}}}{\hat{\bar{h}}}-\frac{\bar{h}_{x^{*}}}{\bar{h}}\right)(\hat{m}_{\hat{\epsilon}}-m)(v)\bar{h}(v)d\mu(v)\right\} =o_{P}\left(1/\sqrt{C_{n}}\right).
\]
Note that this condition is provided here purely for the sake of completeness,
since it will satisfied based on the previously assumed rates of convergence
for $\norm{\hat{m}-m}\norm{\hat{\bar{h}}-\bar{h}}$. This follows
from the fact that the parametric TMLE update step $\hat{m}_{\hat{\epsilon}}$
of $\hat{m}$ will have a negligible effect on the rate of convergence
of the initial estimator $\hat{m}$, that is, $\hat{m}_{\hat{\epsilon}}$
will converge at ``nearly'' the same rate as $\hat{m}$.
\item [{Limited connectivity and limited dependence of $\mathbf{Y}$,$\mathbf{X}$ and $\mathbf{C}$:}] Let
$K_{max,n}=max_{i}\{K_{i}\}$ for a fixed network with $n$ nodes.
Assume that $K_{max,n}^{2}/n$ converges to $0$ in probability as
$n\rightarrow\infty$.
\item [{Bounded fourth moments:}] $E[f_{i}(\mathbf{O})^{4}]<\infty$,
where $f_{i}(\mathbf{O})$ is the contribution of the $i$th observation
to the estimator and is defined below.
\end{description}
A key condition is \textit{consistency and rates for estimators of
nuisance parameters}. This condition will be satisfied, for example,
if both nuisance models are parametric and if one of the two is correctly
specified, or if both models converge to the truth at rate $C_{n}^{1/4}$.
It can in fact be weakened, but for a more general discussion and
the corresponding technical conditions we refer to the Appendix of
\citet{vanderlaan2012}. With the exception of the rates of convergence,
the more general conditions for asymptotic normality of the TMLE presented
in that paper apply to our setting as well.
\section{Overview of the proof of Theorem 1 }
We want to show that $\sqrt{C_{n}}(\hat{\psi}-\psi)$ converges in
law to a Normal limit as $n$ goes to infinity for some rate $\sqrt{C_{n}}$
such that $\sqrt{n/\left(K_{max}(n)\right)^{2}}\leq\sqrt{C}_{n}\leq\sqrt{n}$,
where the rate $\sqrt{C_{n}}$ is the order of the variance of the
sum of the first-order linear approximation of $(\hat{\psi}-\psi)$.
Broadly, the proof has two parts: First, we require that the second
order terms in the expansion of $\hat{\psi}-\psi$ are stochastically
less than $1/\sqrt{C_{n}}$, that is that
\begin{eqnarray*}
\hat{\psi_{n}}-\psi & = & \frac{1}{n}\sum_{i=1}^{n}\left\{ f_{i}(\mathbf{O})-E[f_{i}(\mathbf{O})]\right\} +o_{p}\left(1/\sqrt{C_{n}}\right),\\
\end{eqnarray*}
where $f_{i}(\mathbf{O})$ is the contribution of the $i$th observation
to the estimator. Specifically, for our influence function
\begin{eqnarray*}
D(\mathbf{o}) & = & \sum_{j=1}^{n}\frac{1}{n}\sum_{i=1}^{n}E\left[m\left(V_{i}^{*}\right)\mid C_{j}=c_{j}\right]-\psi+\frac{1}{n}\sum_{i=1}^{n}\frac{\bar{q}_{x^{*}}(v_{i})}{\bar{q}(v_{i})}\left\{ y_{i}-m\left(v_{i}\right)\right\} ,
\end{eqnarray*}
the contribution of the $i$th observation is
\begin{eqnarray*}
f_{i}(\mathbf{o}) & = & \sum_{i=j}^{n}E\left[m\left(V_{i}^{*}\right)\mid C_{j}=c_{j}\right]+\frac{\bar{q}_{x^{*}}(v_{i})}{\bar{q}(v_{i})}\left\{ y_{i}-m\left(v_{i}\right)\right\} .
\end{eqnarray*}
Then proving asymptotic normality of the TMLE amounts to the asymptotic
analysis of the sum $\sum_{i=1}^{n}\left\{ f_{i}(\mathbf{O})-E[f_{i}(\mathbf{O})]\right\} $,
and the second part of the proof establishes that the first order
terms converge to a normal distribution when scaled by $\sqrt{C_{n}}$,
that is that $\sqrt{C_{n}}\sum_{i=1}^{n}\left\{ f_{i}(\mathbf{O})-E[f_{i}(\mathbf{O})]\right\} \rightarrow_{d}N(0,\sigma^{2})$
for some finite $\sigma^{2}$.
The proof that the second order terms are stochastically less than
$1/\sqrt{C_{n}}$ is an extension of the empirical process theory
of \citet{van1996weak} and follows the same format as the proof in
\citet{vanderlaan2012}. Indeed, the proof offered by \citet{vanderlaan2012}
holds immediately after replacing the rate or scaling factor $\sqrt{n}$
with $\sqrt{C_{n}}$ throughout. Only one step in the \citet{vanderlaan2012}
proof relies on the network topology, which is the major difference
between the setting in that paper, where the number of network connections
is fixed and bounded as $n$ goes to infinity, and the present setting:
the proof requires bounding the Orlicz norms of several empirical
processes corresponding to components of the influence function for
$\psi$, and a key step is bounding the expectation of $E\left[\left|X_{n}(f)\right|^{p}\right]$
, where $X_{n}(f)$ is the stochastic process that describes the difference
between the empirical (indexed by $n$) and the true distribution
functions of a component of the influence function for $\psi$. This
step relies on a combinatorial argument about nature of overlapping
friend groups in the underlying network, and the argument for the
case of growing $K_{i}$ is subsumed by the argument for fixed $K$
in \citet{vanderlaan2012}.
The proof that the first order terms converge to a normal distribution
requires a central limit theorem for dependent data with growing and
possibly irregularly sized dependency neighborhoods, where a dependency
neighborhood for unit $i$ is a collection of observations on which
the observations for unit $i$ may be dependent. We prove such a CLT
in Lemmas 1 and 2. In next section we use the CLT for growing and
irregular dependency neighborhoods, along with an orthogonal decomposition
of the first order terms, to prove the remainder of Theorem 1.
\section*{Central limit theorem for first order terms}
Proving asymptotic normality of the TMLE amounts to the asymptotic
analysis of the sum $\sum_{i=1}^{n}\left\{ f_{i}(\mathbf{O})-E[f_{i}(\mathbf{O})]\right\} $.
As a start, decompose $\sum_{i=1}^{n}\left\{ f_{i}(\mathbf{O})-E[f_{i}(\mathbf{O})]\right\} $
into a sum of three orthogonal components:
\begin{eqnarray*}
f_{\mathbf{Y},i}(\mathbf{Y},\mathbf{X},\mathbf{C}) & = & f_{i}(\mathbf{O})-E\left[f_{i}(\mathbf{O})\mid\mathbf{X},\mathbf{C}\right],\\
f_{\mathbf{X},i}(\mathbf{X},\mathbf{C}) & = & E[f_{i}(\mathbf{O})\mid\mathbf{X},\mathbf{C}]-E[f_{i}(\mathbf{O})\mid\mathbf{C}],\mbox{ and}\\
f_{\mathbf{C},i}(\mathbf{C}) & = & E[f_{i}(\mathbf{O})\mid\mathbf{C}]-E[f_{i}(\mathbf{O})].
\end{eqnarray*}
Note that
\[
f_{i}(\mathbf{O})-E[f_{i}(\mathbf{O})]=f_{\mathbf{Y},i}(\mathbf{Y},\mathbf{X},\mathbf{C})+f_{\mathbf{X},i}(\mathbf{X},\mathbf{C})+f_{\mathbf{C},i}(\mathbf{C})
\]
and with slight abuse of notation we will also write $f_{\mathbf{Y},i}(\mathbf{O})$,
$f_{\mathbf{X},i}(\mathbf{O})$ and $f_{\mathbf{C},i}(\mathbf{O})$.
Let $f_{\mathbf{Y}}(\mathbf{O})=\sum_{i=1}^{n}f_{\mathbf{Y},i}(\mathbf{O})$,
$f_{\mathbf{X}}(\mathbf{O})=\sum_{i=1}^{n}f_{\mathbf{X},i}(\mathbf{O})$
and $f_{\mathbf{C}}(\mathbf{O})=\sum_{i=1}^{n}f_{\mathbf{C},i}(\mathbf{O})$.
For $i=1,\ldots,n$, let
\begin{eqnarray*}
Z{}_{Y,i} & = & \frac{f_{\mathbf{Y},i}(\mathbf{Y},\mathbf{X},\mathbf{C})}{\sqrt{Var(\sum_{i=1}^{n}f_{\mathbf{Y},i}(\mathbf{Y},\mathbf{X},\mathbf{C}))}}\\
Z{}_{X,i} & = & \frac{f_{\mathbf{X},i}(\mathbf{X},\mathbf{C})}{\sqrt{Var(\sum_{i=1}^{n}f_{\mathbf{X},i}(\mathbf{X},\mathbf{C}))}}\\
Z{}_{C,i} & = & \frac{f_{\mathbf{C},i}(\mathbf{C})}{\sqrt{Var(\sum_{i=1}^{n}f_{\mathbf{C},i}(\mathbf{C}))}}.
\end{eqnarray*}
and
\begin{eqnarray*}
Z'_{Y,i} & = & \frac{f_{\mathbf{Y},i}(\mathbf{Y},\mathbf{X},\mathbf{C})\left|(\mathbf{X},\mathbf{C})\right.}{\sqrt{Var(\sum_{i=1}^{n}f_{\mathbf{Y},i}(\mathbf{Y},\mathbf{X},\mathbf{C})\left|(\mathbf{X},\mathbf{C})\right.)}}\\
Z'_{X,i} & = & \frac{f_{\mathbf{X},i}(\mathbf{X},\mathbf{C})\left|\mathbf{C}\right.}{\sqrt{Var(\sum_{i=1}^{n}f_{\mathbf{X},i}(\mathbf{X},\mathbf{C})\left|\mathbf{C}\right.)}}\\
\end{eqnarray*}
We use the prime to denote conditional random variables: $Z'_{Y,i}$
conditions $f_{\mathbf{Y},i}(\mathbf{O})$ on $(\mathbf{X},\mathbf{C})$
and rescales it by the standard error of $f_{\mathbf{Y}}(\mathbf{O})|\left(\mathbf{X},\mathbf{C}\right)$.
Similarly, $Z'_{X,i}$ conditions $f_{\mathbf{X},i}(\mathbf{O})$
on $\mathbf{C}$ and rescales it by the standard error of $f_{\mathbf{X}}(\mathbf{O})|\mathbf{C}$.
Let
\begin{align*}
& \sigma_{nY}^{2}(\mathbf{x},\mathbf{\mathbf{c}})=Var\left(\sum_{i=1}^{n}f_{\mathbf{Y},i}(\mathbf{Y},\mathbf{x},\mathbf{c})\left|(\mathbf{X}=\mathbf{x},\mathbf{C}=\mathbf{\mathbf{c}})\right.\right)\\
& \sigma_{nY}^{2}=E_{P_{\mathbf{X},\mathbf{C}}}\left[\sigma_{nY}^{2}(\mathbf{X},\mathbf{\mathbf{C}})\right],
\end{align*}
\begin{align*}
& \sigma_{nX}^{2}(\mathbf{\mathbf{c}})=Var\left(\sum_{i=1}^{n}f_{\mathbf{X},i}(\mathbf{X},\mathbf{c})\left|\mathbf{C}=\mathbf{c}\right.\right)\\
& \sigma_{nX}^{2}=E_{P_{\mathbf{C}}}\left[\sigma_{nX}^{2}(\mathbf{\mathbf{C}})\right],
\end{align*}
and
\[
\sigma_{nC}^{2}=Var\left(\sum_{i=1}^{n}f_{\mathbf{C},i}(\mathbf{C})\right).
\]
Note that by the law of total variance $\sigma_{nX}^{2}=Var(\sum_{i=1}^{n}f_{\mathbf{X},i}(\mathbf{X},\mathbf{C}))$
and $\sigma_{nY}^{2}=Var\left(\sum_{i=1}^{n}f_{\mathbf{Y},i}(\mathbf{Y},\mathbf{X},\mathbf{C})\right)$.
Let $Z_{nY}'$ denote $\sum_{i=1}^{n}Z_{Y,i}'$, $Z_{nX}'$ denote
$\sum_{i=1}^{n}Z_{X,i}'$, $Z_{nY}$ denote $\sum_{i=1}^{n}Z_{Y,i}$,
$Z_{nX}$ denote $\sum_{i=1}^{n}Z_{X,i}$, and $Z_{nC}$ denote $\sum_{i=1}^{n}Z_{C,i}$.
We will establish convergence in distribution of each of the three
terms separately. Because $Z_{nY}^{'}$ and $Z_{nX}^{'}$ converge
to distributions that do not depend on their conditioning events,
conditional convergence in distribution implies convergence of $Z_{nY}$
and $Z_{nX}$ to the same limiting distributions. Since $f_{Y}(\mathbf{O}),$$f_{X}(\mathbf{O})$,
and $f_{C}(\mathbf{O})$ are orthogonal by construction, the variance
of the limiting distribution of their sum is the sum of their marginal
variances. If the three processes converge at the same rate the limiting
variance will be the sum of the variances of the three processes.
However, the three terms may converge at different rates, in which
case the limiting distribution of $\hat{\psi}-\psi$ will be given
by the limiting distribution of the term(s) with the slowest rate
of convergence.
In order to show that $Z_{nX}'$, $Z_{nY}'$, and $Z_{nC}$ all converge
in distribution to a $N(0,1)$ random variable, we can use three separate
applications of the central limit theorem given in Lemma 1, which
is based on Stein's method.
Stein's method \citep{stein1972bound} quantifies the error in approximating
a sample average with a normal distribution. (For an introduction
to Stein's method see \citealp{ross2011fundamentals}.) Stein's method
has been used to prove CLTs for dependent data with dependence structure
given by \textit{dependency neighborhoods} \citep{chen2004normal}:
the dependency neighborhood for observation $i$ is a set of indices
$D_{i}$ such that observation $i$ is independent of observation
$j$, for any $j\notin D_{i}$. Conditionally on $\mathbf{C}$, $f_{X,i}$
and $f_{X,j}$ are independent for any nodes $i$ and $j$ such that
$A_{ij}=0$ and there is no $k$ with $A_{ik}=A_{jk}=1$, that is
for any nodes that do not share a tie or have any mutual network contacts.
The same is true for $f_{Y,i}$ and $f_{Y,j}$ conditional on $\mathbf{X}$
and $\mathbf{C}$ and for $f_{C,i}$ and $f_{C,j}$. Thus the three
collections of random variables $Z_{X,1}',...,Z_{X,n}'$, $Z'_{Y,1},...,Z'_{Y,n}$,
and $Z_{C,1},...,Z_{C,n}$ each has a dependency neighborhood structure
with $D_{i}=i\cup\{j:A_{ij}=1\}\cup\{k:A_{jk}=1\mbox{ for }j:A_{ij}=1\}$,
that is the ``friends'' and ``friends of friends'' of node $i$.
Define the indicators $R(i,j)$ for any $(i,j)\in\{1,\ldots,n\}^{2}$
to be an indicator of dependence between $Z_{X,i}$ and $Z_{X,j}$,
$R(i,j)=1$ iff $j\in D_{i}$ or, equivalently, if $i\in D_{j}$.
For any $i\in\{1,\ldots,n\}$ the set $\{Z'_{X,j}:\left(R(i,j)=1,j\in\{1,\ldots,n\}\right)\}$
forms the dependency neighborhood of $Z'_{X,i}$ and the collection
$\{Z'_{X,j}:\left(R(i,j)=0,j\in\{1,\ldots,n\}\right)\}$ is independent
of $Z'_{X,i}$. The same logic applies to defining the dependency
neighborhoods for $Z'_{Y,1},...,Z'_{Y,n}$ conditional on $\mathbf{X}$
and $\mathbf{C}$, and for $Z_{C,1},...,Z_{C,n}$ based on (unconditional)
independence of each $f_{C,i}(\mathbf{O})$ and $f_{C,j}(\mathbf{O})$,
as determined by the network structure and the distributional assumptions
made for the baseline covariates $\mathbf{C}$.
Applied to $Z'_{nX}$ , Stein's method provides the following upper
bound
\begin{eqnarray*}
d(Z'_{nX},Z) & \leq & \sum_{i=1}^{n}\sum_{j,k\in D_{i}}E\left|Z'_{X,i}Z'_{X,j}Z'_{X,k}\right|\\
& & +\sqrt{\frac{2}{\pi}}\sqrt{Var\left(\sum_{i=1}^{n}\sum_{j\in D_{i}}Z'_{X,i}Z'_{X,j}\right)},
\end{eqnarray*}
where $Z\sim N(0,1)$ and $d(\cdot,\cdot)$ is the Wasserstein distance
metric \citep{vallender1974calculation}.
In order to show that $Z'_{nX}$ converges in distribution to $Z$,
we must show that the righthand side of the inequality converges to
zero as $n$ goes to infinity. We will first show that this convergence
holds when $K_{i}=\left|F_{i}\right|=K_{max}(n)$ for all $i$, that
is when all nodes have the same number of ties. We will then show
that removing any tie from the network preserves an upper bound on
the righthand side of the inequality. This completes our proof that
for any network such that $K_{i}\leq K_{max}(n)$ for all $i$ and
$\frac{K_{max}^{2}(n)}{n}$ converges to zero as $n$ goes to infinity,
$Z'_{nX}$ converges in distribution to a standard normal distribution.
The same argument applied to $Z{}_{nC}$ proves that it has a Normal
limiting distributions as well.\\
\begin{lemma}{[}Applying Stein's Method to the dependent sum{]}\label{lem:Steins}\label{stein's lemma}
Consider a network of nodes given by adjacency matrix $A$. Let $U_{1},...,U_{n}$
be bounded mean-zero random variables with finite fourth moments and
with dependency neighborhoods $D_{i}=i\cup\{j:A_{ij}=1\}\cup\{k:A_{jk}=1\mbox{ for }j:A_{ij}=1\}$,
and let $K_{i}$ be the degree of node $i$. If $K_{i}=K_{max}(n)$
for all $i$ and $K_{max}(n)^{2}/n\rightarrow0$, then $\frac{\sum U_{i}}{\sqrt{var(\sum U_{i})}}\overset{D}{\rightarrow}N(0,1)$.
\end{lemma}
\begin{proof}{[}Proof of Lemma~\ref{lem:Steins}{]}\textbf{ }Let
$U'_{i}=\frac{U_{i}}{\sqrt{var(\sum U_{i})}}$. Application of Stein's
method often involves defining the so-called ``Stein coupling''
$(W,W',G)$ \citep{fang2011thesis,fang2015rates}. Consider the following
sum of dependent variables $W=\sum_{i=1}^{n}U'_{i}$. Define a discrete
random variable $I$ distributed uniformly over $\{1,\ldots,n\}$
and define another random variable $W'=(W-\sum_{j=1}^{n}R(I,j)U'_{j})$.
Finally, define $G=-nU'_{I}$ and note that $(W,W',G)$ forms a Stein
coupling \textbackslash{}citep\{fang2011thesis,fang2015rates\}. We
also let $D=(W'-W)=-\sum_{j=1}^{N}R(I,j)U'{}_{j}$. This Stein coupling
allows us then to derive the upper bound
\begin{align}
d(W,Z) & \leq\sum_{i=1}^{n}\sum_{j,k\in D_{i}}E\left|U'{}_{i}U'_{j}U'_{k}\right|+\sqrt{\frac{2}{\pi}}\sqrt{Var\left(\sum_{i=1}^{n}\sum_{j\in D_{i}}U'_{i}U'_{j}\right)},\label{eq:stein}
\end{align}
as shown in \citet{ross2011fundamentals}. We will now show that,
for any network structure,
\begin{eqnarray}
& & \sum_{i=1}^{n}\sum_{j,k\in D_{i}}E\left|U'{}_{i}U'{}_{j}U'_{k}\right|+\sqrt{\frac{2}{\pi}}\sqrt{Var\left(\sum_{i=1}^{n}\sum_{j\in D_{i}}U'_{i}U'_{j}\right)}\nonumber \\
& = & O\left(\frac{\sum_{i,j,k}R(i,j)R(i,k)}{\left[\sum_{i,j}R(i,j)\right]^{3/2}}\right).\label{eq:to show}
\end{eqnarray}
The righthand side of the above equation is equal to $\sqrt{\frac{\left(K_{max}(n)\right)^{2}}{n}}$
under the assumption of $K_{max}(n)$ ties for each node $i=\{1,\ldots,n\}$.
By assumption, we also have that $\frac{K_{max}(n)}{\sqrt{n}}$ converges
to zero as $n$ goes to infinity, and therefore if we can show equation
\eqref{eq:to show} we have proved that $\frac{\sum U_{i}}{\sqrt{var(\sum U_{i})}}\overset{D}{\rightarrow}N(0,1)$.
Consider the term
\[
\sum_{i=1}^{n}\sum_{j,k\in D_{i}}E\left|U'{}_{i}U'{}_{j}U'_{k}\right|=\dfrac{1}{var(\sum U_{i})^{3/2}}\sum_{i=1}^{n}E\left\{ \left|U_{i}\left(\sum_{j\in D_{i}}U_{k}\right)^{2}\right|\right\} .
\]
By the assumption of bounded 4th moments, $var(\sum U_{i})^{3/2}=O\left(\left[\sum_{i,j}R(i,j)\right]^{3/2}\right)$,
that is, $var(\sum U_{i})$ stabilizes to a constant when scaled by
$\sum_{i,j}R(i,j)$. Using the fact that each $|U_{i}|$ is bounded
we get
\begin{eqnarray*}
& & \sum_{i=1}^{N}E\left\{ \left|U_{i}\left(\sum_{j\in D_{i}}U_{j}\right)^{2}\right|\right\} \\
& \leq & M\sum_{i=1}^{n}\left\{ \sum_{j,k}R(i,j)R(i,k)\right\} \\
& = & M\sum_{i,j,k}R(i,j)R(i,k),
\end{eqnarray*}
for some positive constant $M<\infty$. Combining the above expressions,
we get
\[
\sum_{i=1}^{n}\sum_{j,k\in D_{i}}E\left|U'{}_{i}U'{}_{j}U'_{k}\right|=O\left(\frac{\sum_{i,j,k}R(i,j)R(i,k)}{\left[\sum_{i,j}R(i,j)\right]^{3/2}}\right).
\]
Now consider the second term:
\begin{align*}
\sqrt{Var\left(\sum_{i=1}^{n}\sum_{j\in D_{i}}U'_{i}U'_{j}\right)} & =\frac{\sqrt{Var\left(\sum_{i=1}^{n}\sum_{j\in D_{i}}U{}_{i}U{}_{j}\right)}}{var(\sum U_{i})^{2}}.
\end{align*}
There are $\sum_{i,j}R(i,j)$ terms in $\sum_{i=1}^{n}\sum_{j\in D_{i}}U{}_{i}U{}_{j}$,
and the number of terms $U_{k}U_{l}$ with which $U_{i}U_{j}$ has
non-zero covariance is $\left|D_{i}\cup D_{j}\right|\leq\sum_{k}R(i,k)+\sum_{k}R(i,k)$,
so $Var\left(\sum_{i=1}^{n}\sum_{j\in D_{i}}U{}_{i}U{}_{j}\right)\leq M\sum_{i,j}R(i,j)\sum_{k}R(i,k)$
for some finite $M$. Therefore $Var\left(\sum_{i=1}^{n}\sum_{j\in D_{i}}U{}_{i}U{}_{j}\right)=O\left(\sum_{i,j,k}R(i,j)R(i,k)\right)$.
$Var(\sum U_{i})^{2}=O\left(\left[\sum_{i,j}R(i,j)\right]^{2}\right)$,
so the second term is of smaller order than the first term. Therefore
we have only to consider the first term and we have completed the
proof. \end{proof}
\begin{lemma}{[}Bound goes to zero when $K_{i}\leq K_{max}(n)$ for
all $i${]}\label{lem:lowerbound} Convergence to zero of the righthand
side of Equation (\ref{eq:stein}) is preserved under the removal
of ties and holds as long as $K_{i}\leq K_{max}(n)$ for all $i$
and $\frac{K_{max}^{2}(n)}{n}$ converges to zero as $n$ goes to
infinity. \end{lemma}
\begin{proof}{[}Proof of Lemma~\ref{lem:lowerbound}{]} Consider
a sequence of networks with $n$ going to infinity such that the righthand
side of Equation (\ref{eq:stein}) converges to $0$, i.e.
\begin{eqnarray*}
\sum_{i=1}^{n}\sum_{j,k\in D_{i}}E\left|U'_{i}U'_{j}U'_{k}\right|+\sqrt{\frac{2}{\pi}}\sqrt{Var\left(\sum_{i=1}^{n}\sum_{j\in D_{i}}U'_{i}U'_{j}\right)} & \rightarrow & 0.
\end{eqnarray*}
Because the second term is of the same or smaller order than the first,
we only have to consider the first term. For this sequence of networks,
define $A_{n}=\sum_{i=1}^{n}\sum_{j,k\in D_{i}}E\left|U'_{i}U'_{j}U'_{k}\right|$
. Removing a single tie from the underlying network has the effect
of rendering independent some pairs that were previously dependent;
We now consider the effect of rendering a single dependent pair independent
but otherwise leaving the distributions of the random variables the
same. Suppose the pair rendered independent is $(l,m)$. Define a
new sequence of networks with $n$ going to infinity to be identical
to the previous sequence but with pair $(l,m)$ independent, and let
$A'_{n}$ be the first term in the righthand side of Equation (\ref{eq:stein})
for this new sequence. Then
\begin{align*}
A'_{n} & =A_{n}-2\sum_{k\in D_{l}\cup D_{m}}E\left|U'_{l}U'_{m}U'_{k}\right|
\end{align*}
which is bounded above by $A_{n}$.\end{proof}
This completes the proof that $Z_{nX}'$, $Z'_{nY}$, and $Z{}_{nC}$
have Normal limiting distributions.
\begin{lemma}{[}Conditional CLT implies marginal CLT{]}\label{lem:marginalCLT}$Z_{nX}'$
converges to Normal distribution after marginalizing over \textbf{$\mathbf{C}$
}(but conditioning on the network as captured by the adjacency matrix
\textbf{$\mathbf{A}$}) and $Z_{nY}'$ converges to Normal distribution
after marginalizing over $(\mathbf{X},\mathbf{C})$. That is, $Z_{nX}$
and $Z_{nY}$ both converge to Normal distributions.\end{lemma}
\begin{proof}{[}Proof of Lemma~\ref{lem:marginalCLT}{]} For illustration
consider $Z_{nX}'=\sum_{i=1}^{n}Z_{2,i}'$, where
\[
Z_{X,i}'=\left(f_{\mathbf{X},i}(\mathbf{X},\mathbf{C})\left|\mathbf{C}\right.\right)/\sqrt{\sigma_{nX}^{2}(\mathbf{\mathbf{C}})}
\]
and note that the proof of the convergence of $Z_{nY}$ is nearly
identical. The conditional CLT results from Lemma \ref{lem:Steins}
show that
\[
P\left[Z_{nX}'\leq x\left|\mathbf{C}=\mathbf{c}\right.\right]=P\left[\left(\sum_{i=1}^{N}\dfrac{f_{\mathbf{X},i}(\mathbf{X},\mathbf{c})}{\sqrt{\sigma_{nX}^{2}(\mathbf{c})}}\leq x\right)\left|\mathbf{C}=\mathbf{c}\right.\right]
\]
converges to $\Phi(x)$ for each $x$ and almost every $\mathbf{c}$,
where $\Phi$ is the cumulative distribution function of the standard
Normal random variable and $\mathbf{C}$ is a given sequence $(C_{i}:i=1,\ldots,n)$.
Let $P_{\mathbf{C}}$ denote the distribution of $\mathbf{C}$. Then
\begin{align*}
P(Z_{nX}\leq x) & \equiv P\left[\left(\sum_{i=1}^{N}\dfrac{f_{\mathbf{X},i}(\mathbf{X},\mathbf{C})}{\sqrt{\sigma_{nX}^{2}}}\leq x\right)\right]\\
& =\int_{\mathbf{c}}P(Z_{nX}'\leq x|\mathbf{C}=\mathbf{c})dP_{\mathbf{C}}(\mathbf{c}).
\end{align*}
For a given $x$, the dominated convergence theorem is now applied
with $f_{n}(\mathbf{c})=P(Z_{nX}'\leq x|\mathbf{C}=\mathbf{c})$ and
the limit given by $f(\mathbf{c})=\Phi(x)=m$, where $m$ is some
constant that doesn't depend on $\mathbf{c}$. From the previous conditional
CLT result it follows that $f_{n}(\mathbf{c})$ converges to $f(\mathbf{c})$
pointwise for each $\mathbf{c}$. The next step is to find an integrable
function $g$, such that $f_{n}<g$ and $\int g(\mathbf{c})dP_{\mathbf{C}}(\mathbf{c})<\infty$.
The proof is then completed by choosing $g=1$. \end{proof}
We have now shown that $Z_{nY}$, $Z_{nX}$, and $Z_{nC}$ are asymptotically
Normally distributed. We now show that the sum of the three processes
converges in distribution to a Normal random variable. Consider three
cases: (1) the three processes have the same rate of marginal convergence
in distribution, (2) one of the three processes converges faster than
the other two, and (3) two of the processes converge faster than the
third. In all three cases the rate of convergence for the sum will
be the slowest of the three marginal rates. In case (3), the limiting
distribution of the sum is determined entirely by the one process
that converges with a slower rate than the other two: the other two
processes will converge to constants (specifically to their expected
values of 0) when standardized by the slower rate; Slutsky's theorem
concludes the proof. We focus on case (1) below; case (2) follows
immediately by applying the proof below to the two processes that
converge at the same slower rate and applying Slutsky's to the third,
faster converging process.
For convenience, in order to show that the sum of the three dependent
processes also converges to Normal, define
\[
C_{n}^{*}:=\sigma_{nY}^{2}+\sigma_{nX}^{2}+\sigma_{nC}^{2}.
\]
Note that $C_{n}^{*}$ is related to $C_{n}$ as follows: $C_{n}=O(n^{2}/C_{n}^{*})$.
\begin{lemma}{[}CLT for the sum of the three orthogonal processes{]}\label{lem:sumCLT}If
all three processes have the same marginal rate of convergence, then
\[
\dfrac{1}{\sqrt{C_{n}^{*}}}\left(f_{\mathbf{Y}}(\mathbf{Y},\mathbf{\mathbf{X}},\mathbf{C})+f_{\mathbf{\mathbf{X}}}(\mathbf{\mathbf{X},}\mathbf{C})+f_{\mathbf{C}}(\mathbf{C})\right)\rightarrow N(0,1).
\]
\end{lemma}
\begin{proof}{[}Proof of Lemma~\ref{lem:sumCLT}{]} Without the
loss of generality, we prove that $Z_{nX}+Z_{nC}\rightarrow N(0,2)$
and note that the general result for $(Z_{nY}+Z_{nX}+Z_{nC})$ follows
by applying a similar set of arguments.
Consider the following random vector $(Z_{nX},Z_{nC})$ taking values
in $\openr^{2}$. Let $F_{n}(x_{1},x_{2})\equiv P(Z_{nX}\leq x_{1},Z_{nC}\leq x_{2})$,
where $(x_{1},x_{2})\in\openr^{2}$. Let $\Phi^{2}(x_{1},x_{2})\equiv P(Z_{X}\leq x_{1})P(Z_{C}\leq x_{2})$,
for $Z_{X}\sim N(0,1)$ and $Z_{C}\sim N(0,1)$, that is, $\Phi^{2}(x_{1},x_{2})$
defines the CDF of the bivariate standard normal distribution, for
$(x_{1},x_{2})\in\openr^{2}$. The goal is to show that $F_{n}(x_{1},x_{2})\rightarrow\Phi^{2}(x_{1},x_{2})$,
for any $(x_{1},x_{2})\in\openr^{2}$. The convergence in distribution
for $Z_{nX}+Z_{nC}$ will follow by applying the Cramer and Wold Theorem
(1936).
Note that
\begin{align*}
& P(Z_{nX}\leq x_{1},Z_{nC}\leq x_{2})\\
= & P(Z_{nX}\leq x_{1}\left|Z_{nC}\leq x_{2}\right.)P(Z_{nC}\leq x_{2}).
\end{align*}
First, from the previous application of Stein's method, we have that
\[
P(Z_{nC}\leq x_{2})\rightarrow\Phi(x_{2}),
\]
where $\Phi(x_{2})\equiv P(Z_{C}\leq x_{2})$, $Z_{C}\sim N(0,1)$
and $x_{2}\in\openr^{2}$. Also note that
\begin{eqnarray*}
& & P(Z_{nX}\leq x_{1}\left|Z_{nC}\leq x_{2}\right.)\\
& = & \sum_{\mathbf{c}\in\mathcal{C}}P(Z_{nX}\leq x_{1}\left|\mathbf{C}=\mathbf{c}\right.)P(\mathbf{C}=\mathbf{c}\left|Z_{nC}\leq x_{2}\right.),
\end{eqnarray*}
where $\mathcal{C}$ denotes the support of $\mathbf{C}$, $Z_{nX}=\frac{1}{\sqrt{C_{n}^{*}}}f_{\mathbf{X}}(\mathbf{\mathbf{X},}\mathbf{C})$,
$Z_{nC}=\frac{1}{\sqrt{C_{n}^{*}}}f_{\mathbf{C}}(\mathbf{C})$ and
\[
P(\mathbf{C}=\mathbf{c}\left|Z_{nC}\leq x_{2}\right.)=\dfrac{P(\mathbf{C}=\mathbf{c})I(\left(1/\sqrt{C_{n}^{*}}\right)f_{\mathbf{C}}(\mathbf{c})\leq x_{2})}{P(\left(1/\sqrt{C_{n}^{*}}\right)f_{\mathbf{C}}(\mathbf{c})\leq x_{2})}.
\]
By another application of Stein's method, it was shown that
\[
P(Z_{nX}\leq x_{1}\left|\mathbf{C}=\mathbf{c}\right.)\rightarrow\Phi(x_{2}),
\]
for any realization of $\mathbf{c}\in\mathcal{C}$. That is, we've
shown that the limiting distribution of $Z_{nX}$ conditional on $\mathbf{C=\mathbf{c}}$,
does not itself depend on the conditioning event $\mathbf{C}=\mathbf{c}$.
Applying Lemma 3, we finally conclude that $F_{n}(x_{1},x_{2})\rightarrow\Phi^{2}(x_{1},x_{2})$,
for any $(x_{1},x_{2})\in\openr^{2}$ and the result follows.\end{proof}
\section*{Variance estimation}
The estimate of the variance of the TMLE $\hat{\psi}$ can be obtained
from the sum, scaled by $1/n^{2}$, of the three plug-in estimators
of
\begin{eqnarray*}
\sigma_{nY}^{2} & = & \sum_{i,j}E(f_{\mathbf{Y},i}(\mathbf{O})f_{\mathbf{Y},j}(\mathbf{O}))\\
\sigma_{nX}^{2} & = & \sum_{i,j}E(f_{\mathbf{X},i}(\mathbf{O})f_{\mathbf{X},j}(\mathbf{O}))\\
\sigma_{nC}^{2} & = & \sum_{i,j}E(f_{\mathbf{C},i}(\mathbf{O})f_{\mathbf{C},j}(\mathbf{O})).
\end{eqnarray*}
Alternatively, one can estimate the variance from a single plug-in
estimator
\[
\frac{1}{n^{2}}\sum_{i,j}E(f_{i}(\mathbf{O})f_{j}(\mathbf{O})).
\]
Note that contribution to these variances of any pair $i,j$ not in
each others dependency neighborhoods will be $0$. Therefore, it is
acceptable to sum only over pairs $i,j$ sharing a tie or a mutual
contact in the underlying network. Finally, note that we do not need
to know the true rate of convergence $\sqrt{C_{n}}$ to obtain a valid
estimate of the C.I. for $\psi$; this rate is captured by the number
of non-zero terms in the variance sums.
\bibliographystyle{jasa}
\bibliography{references}
\end{document}
\begin{thebibliography}{71}
\newcommand{\enquote}[1]{``#1''}
\expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi
\expandafter\ifx\csname url\endcsname\relax
\def\url#1{{\tt #1}}\fi
\expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi
\bibitem[{Ali and Dwyer(2010)}]{ali2010social}
Ali, M.~M. and Dwyer, D.~S.
\newblock \enquote{Social network effects in alcohol consumption among
adolescents.}
\newblock {\em Addictive behaviors\/}, 35(4):337--342 (2010).
\bibitem[{Aronow and Samii(2013)}]{aronow2012estimating}
Aronow, P.~M. and Samii, C.
\newblock \enquote{Estimating average causal effects under general
interference.}
\newblock Technical report, Yale University (2013).
\bibitem[{Athey et~al.(2016)Athey, Eckles, and Imbens}]{athey2016exact}
Athey, S., Eckles, D., and Imbens, G.~W.
\newblock \enquote{Exact P-values for Network Interference*.}
\newblock {\em Journal of the American Statistical Association\/},
(just-accepted) (2016).
\bibitem[{Barab{\'a}si and Albert(1999)}]{barabasi1999emergence}
Barab{\'a}si, A.-L. and Albert, R.
\newblock \enquote{Emergence of scaling in random networks.}
\newblock {\em science\/}, 286(5439):509--512 (1999).
\bibitem[{Besag(1974)}]{besag1977spatial}
Besag, J.
\newblock \enquote{On spatial-temporal models and Markov fields.}
\newblock In {\em Transactions of the Seventh Prague Conference on Information
Theory, Statistical Decision Functions, and Random Processes\/}, 47--55.
Springer (1974).
\bibitem[{Bowers et~al.(2013)Bowers, M, and C}]{Bowers2013interference}
Bowers, J., M, F.~M., and C, P.
\newblock \enquote{Reasoning about interference between units: A general
framework.}
\newblock {\em Political Analysis\/}, 21:97--124 (2013).
\bibitem[{Cacioppo et~al.(2009)Cacioppo, Fowler, and
Christakis}]{cacioppo2009alone}
Cacioppo, J.~T., Fowler, J.~H., and Christakis, N.~A.
\newblock \enquote{Alone in the crowd: the structure and spread of loneliness
in a large social network.}
\newblock {\em Journal of personality and social psychology\/}, 97(6):977
(2009).
\bibitem[{Chen and Shao(2004)}]{chen2004normal}
Chen, L.~H. and Shao, Q.-M.
\newblock \enquote{Normal approximation under local dependence.}
\newblock {\em The Annals of Probability\/}, 32(3):1985--2028 (2004).
\bibitem[{Christakis and Fowler(2007)}]{christakis2007spread}
Christakis, N. and Fowler, J.
\newblock \enquote{The spread of obesity in a large social network over 32
years.}
\newblock {\em New England Journal of Medicine\/}, 357(4):370--379 (2007).
\bibitem[{Christakis and Fowler(2008)}]{christakis2008collective}
---.
\newblock \enquote{The collective dynamics of smoking in a large social
network.}
\newblock {\em New England journal of medicine\/}, 358(21):2249--2258 (2008).
\bibitem[{Christakis and Fowler(2010)}]{christakis2010social}
---.
\newblock \enquote{Social network sensors for early detection of contagious
outbreaks.}
\newblock {\em PloS one\/}, 5(9):e12948 (2010).
\bibitem[{Clauset et~al.(2009)Clauset, Shalizi, and Newman}]{clauset2009power}
Clauset, A., Shalizi, C.~R., and Newman, M.~E.
\newblock \enquote{Power-law distributions in empirical data.}
\newblock {\em SIAM review\/}, 51(4):661--703 (2009).
\bibitem[{Cohen-Cole and Fletcher(2008)}]{cohen2008obesity}
Cohen-Cole, E. and Fletcher, J.
\newblock \enquote{Is obesity contagious? Social networks vs. environmental
factors in the obesity epidemic.}
\newblock {\em Journal of Health Economics\/}, 27(5):1382--1387 (2008).
\bibitem[{Diaconis and Janson(2007)}]{diaconis2007graph}
Diaconis, P. and Janson, S.
\newblock \enquote{Graph limits and exchangeable random graphs.}
\newblock {\em arXiv preprint arXiv:0712.2749\/} (2007).
\bibitem[{Eckles et~al.(2014)Eckles, Karrer, and Ugander}]{eckles2014design}
Eckles, D., Karrer, B., and Ugander, J.
\newblock \enquote{Design and analysis of experiments in networks: Reducing
bias from interference.}
\newblock {\em arXiv preprint arXiv:1404.7530\/} (2014).
\bibitem[{Fang(2011)}]{fang2011thesis}
Fang, X.
\newblock \enquote{Multivariate, combinatorial and discretized normal
approximations by Stein's method.}
\newblock Ph.D. thesis (2011).
\bibitem[{Fang et~al.(2015)Fang, R{\"o}llin et~al.}]{fang2015rates}
Fang, X., R{\"o}llin, A., et~al.
\newblock \enquote{Rates of convergence for multivariate normal approximation
with applications to dense graphs and doubly indexed permutation statistics.}
\newblock {\em Bernoulli\/}, 21(4):2157--2189 (2015).
\bibitem[{Goetzke(2008)}]{goetzke2008network}
Goetzke, F.
\newblock \enquote{Network effects in public transit use: evidence from a
spatially autoregressive mode choice model for New York.}
\newblock {\em Urban Studies\/}, 45(2):407--417 (2008).
\bibitem[{Graham et~al.(2010)Graham, Imbens, and Ridder}]{graham2010measuring}
Graham, B., Imbens, G., and Ridder, G.
\newblock \enquote{Measuring the Effects of Segregation in the Presence of
Social Spillovers: A Nonparametric Approach.}
\newblock Technical report, National Bureau of Economic Research (2010).
\bibitem[{Halloran and Hudgens(2011)}]{halloran2011causal}
Halloran, M. and Hudgens, M.
\newblock \enquote{Causal Inference for Vaccine Effects on Infectiousness.}
\newblock {\em The University of North Carolina at Chapel Hill Department of
Biostatistics Technical Report Series\/}, 20 (2011).
\bibitem[{Halloran and Struchiner(1995)}]{halloran1995causal}
Halloran, M. and Struchiner, C.
\newblock \enquote{Causal inference in infectious diseases.}
\newblock {\em Epidemiology\/}, 142--151 (1995).
\bibitem[{Haneuse and Rotnitzky(2013)}]{haneuse2013estimation}
Haneuse, S. and Rotnitzky, A.
\newblock \enquote{Estimation of the effect of interventions that modify the
received treatment.}
\newblock {\em Statistics in medicine\/}, 32(30):5260--5277 (2013).
\bibitem[{Harling et~al.(2016)Harling, Wang, Onnela, and
DeGruttola}]{harling2016leveraging}
Harling, G., Wang, R., Onnela, J.-P., and DeGruttola, V.
\newblock \enquote{Leveraging Contact Network Structure in the Design of
Cluster Randomized Trials.}
\newblock {\em Harvard University Biostatistics Working Paper Series\/},
(Working Paper 199) (2016).
\bibitem[{Hong and Raudenbush(2006)}]{hong2006evaluating}
Hong, G. and Raudenbush, S.
\newblock \enquote{Evaluating Kindergarten Retention Policy.}
\newblock {\em Journal of the American Statistical Association\/},
101(475):901--910 (2006).
\bibitem[{Hong and Raudenbush(2008)}]{hong2008causal}
---.
\newblock \enquote{Causal inference for time-varying instructional treatments.}
\newblock {\em Journal of Educational and Behavioral Statistics\/},
33(3):333--362 (2008).
\bibitem[{Hudgens and Halloran(2008)}]{hudgens2008toward}
Hudgens, M. and Halloran, M.
\newblock \enquote{Toward causal inference with interference.}
\newblock {\em Journal of the American Statistical Association\/},
103(482):832--842 (2008).
\bibitem[{Knoke and Yang(2008)}]{knoke2008social}
Knoke, D. and Yang, S.
\newblock {\em Social network analysis\/}, volume 154.
\newblock Sage (2008).
\bibitem[{Lauritzen and Richardson(2002)}]{lauritzen2002chain}
Lauritzen, S.~L. and Richardson, T.~S.
\newblock \enquote{Chain graph models and their causal interpretations.}
\newblock {\em Journal of the Royal Statistical Society: Series B\/},
64(3):321--348 (2002).
\bibitem[{Lee(2004)}]{lee2004asymptotic}
Lee, L.-F.
\newblock \enquote{Asymptotic Distributions of Quasi-Maximum Likelihood
Estimators for Spatial Autoregressive Models.}
\newblock {\em Econometrica\/}, 72(6):1899--1925 (2004).
\bibitem[{Lin(2005)}]{lin2005peer}
Lin, X.
\newblock \enquote{Peer effects and student academic achievement: an
application of spatial autoregressive model with group unobservables.}
\newblock {\em Unpublished manuscript, Ohio State University\/} (2005).
\bibitem[{Lov{\'a}sz(2012)}]{lovasz2012large}
Lov{\'a}sz, L.
\newblock {\em Large networks and graph limits\/}, volume~60.
\newblock American Mathematical Soc. (2012).
\bibitem[{Lyons(2011)}]{lyons2011spread}
Lyons, R.
\newblock \enquote{The spread of evidence-poor medicine via flawed
social-network analysis.}
\newblock {\em Statistics, Politics, and Policy\/}, 2(1) (2011).
\bibitem[{Madan et~al.(2010)Madan, Moturu, Lazer, and
Pentland}]{madan2010social}
Madan, A., Moturu, S.~T., Lazer, D., and Pentland, A.~S.
\newblock \enquote{Social sensing: obesity, unhealthy eating and exercise in
face-to-face networks.}
\newblock In {\em Wireless Health 2010\/}, 104--110. ACM (2010).
\bibitem[{Moreno(1937)}]{moreno1937sociometry}
Moreno, J.~L.
\newblock \enquote{Sociometry in relation to other social sciences.}
\newblock {\em Sociometry\/}, 1(1/2):206--219 (1937).
\bibitem[{Mu{\~n}oz and van~der Laan(2012)}]{munoz2012population}
Mu{\~n}oz, I.~D. and van~der Laan, M.
\newblock \enquote{Population intervention causal effects based on stochastic
interventions.}
\newblock {\em Biometrics\/}, 68(2):541--549 (2012).
\bibitem[{Newman(2009)}]{newman2009networks}
Newman, M.
\newblock {\em Networks: an introduction\/}.
\newblock Oxford: Oxford University Press (2009).
\bibitem[{Newman and Park(2003)}]{newman2003social}
Newman, M.~E. and Park, J.
\newblock \enquote{Why social networks are different from other types of
networks.}
\newblock {\em Physical Review E\/}, 68(3):036122 (2003).
\bibitem[{Ogburn and VanderWeele(2013)}]{ogburn2013}
Ogburn, E. and VanderWeele, T.~J.
\newblock \enquote{Causal diagrams for interference.}
\newblock Technical report, Harvard University (2013).
\bibitem[{Ogburn and VanderWeele(2014)}]{ogburn2014vaccines}
Ogburn, E.~L. and VanderWeele, T.~J.
\newblock \enquote{Vaccines, Contagion, and Social Networks.}
\newblock {\em arXiv preprint arXiv:1403.1241\/} (2014).
\bibitem[{O'Malley and Marsden(2008)}]{o2008analysis}
O'Malley, J.~A. and Marsden, P.~V.
\newblock \enquote{The analysis of social networks.}
\newblock {\em Health services and outcomes research methodology\/},
8(4):222--269 (2008).
\bibitem[{Pearl(1995)}]{pearl1995causal}
Pearl, J.
\newblock \enquote{Causal diagrams for empirical research.}
\newblock {\em Biometrika\/}, 82(4):669--688 (1995).
\bibitem[{Pearl(2000)}]{pearl2000causality}
---.
\newblock {\em Causality: models, reasoning and inference\/}.
\newblock Cambridge Univ Press (2000).
\bibitem[{{{R} Core Team}(2015)}]{r}
{{R} Core Team}.
\newblock {\em {R}: A Language and Environment for Statistical Computing\/}.
\newblock {R} Foundation for Statistical Computing, Vienna, Austria (2015).
\bibitem[{Rosenbaum(2007)}]{rosenbaum2007interference}
Rosenbaum, P.
\newblock \enquote{Interference between units in randomized experiments.}
\newblock {\em Journal of the American Statistical Association\/},
102(477):191--200 (2007).
\bibitem[{Rosenquist et~al.(2010)Rosenquist, Murabito, Fowler, and
Christakis}]{rosenquist2010spread}
Rosenquist, J.~N., Murabito, J., Fowler, J.~H., and Christakis, N.~A.
\newblock \enquote{The spread of alcohol consumption behavior in a large social
network.}
\newblock {\em Annals of Internal Medicine\/}, 152(7):426--433 (2010).
\bibitem[{Ross(2011)}]{ross2011fundamentals}
Ross, N.~F.
\newblock \enquote{Fundamentals of Stein's method.}
\newblock {\em Probability Surveys\/}, 8:210--293 (2011).
\bibitem[{Rubin(1990)}]{rubin1990application}
Rubin, D.
\newblock \enquote{Comment: Neyman (1923) and causal inference in experiments
and observational studies.}
\newblock {\em Statistical Science\/}, 5(4):472--480 (1990).
\bibitem[{Shalizi and Thomas(2011)}]{shalizi2011homophily}
Shalizi, C. and Thomas, A.
\newblock \enquote{Homophily and contagion are generically confounded in
observational social network studies.}
\newblock {\em Sociological Methods \& Research\/}, 40(2):211--239 (2011).
\bibitem[{Shalizi(2012)}]{shalizi2012statistics}
Shalizi, C.~R.
\newblock \enquote{Comment on ``Why and When 'Flawed' Social Network Analyses
Still Yield Valid Tests of no Contagion''.}
\newblock {\em Statistics, Politics, and Policy\/}, 5(1) (2012).
\bibitem[{Shalizi and Rinaldo(2013)}]{shalizi2011consistency}
Shalizi, C.~R. and Rinaldo, A.
\newblock \enquote{Consistency under Sampling of Exponential Random Graph
Models.}
\newblock {\em Annals of Statistics\/}, 41(2):508--535 (2013).
\bibitem[{Sobel(2006)}]{sobel2006randomized}
Sobel, M.
\newblock \enquote{What Do Randomized Studies of Housing Mobility Demonstrate?}
\newblock {\em Journal of the American Statistical Association\/},
101(476):1398--1407 (2006).
\bibitem[{Sofrygin and van~der Laan(2015)}]{sofryginTechreport}
Sofrygin, O. and van~der Laan, M.~J.
\newblock \enquote{{Semi-Parametric Estimation and Inference for the Mean
Outcome of the Single Time-Point Intervention in a Causally Connected
Population}.}
\newblock {\em U.C. Berkeley Division of Biostatistics Working Paper Series\/},
(Working Paper 344) (2015).
\bibitem[{Sofrygin and {van der Laan}(2015)}]{R-tmlenet}
Sofrygin, O. and {van der Laan}, M.~J.
\newblock {\em {tmlenet}: Targeted Maximum Likelihood Estimation for Network
Data\/} (2015).
\newblock R package version 0.1.0.
\bibitem[{Sofrygin et~al.(2015)Sofrygin, van~der Laan, and
Neugebauer}]{R-simcausal}
Sofrygin, O., van~der Laan, M.~J., and Neugebauer, R.
\newblock {\em {simcausal}: Simulating Longitudinal Data with Causal Inference
Applications\/} (2015).
\newblock R package version 0.5.0.
\newline\urlprefix\url{http://CRAN.R-project.org/package=simcausal}
\bibitem[{Stein(1972)}]{stein1972bound}
Stein, C.
\newblock \enquote{A bound for the error in the normal approximation to the
distribution of a sum of dependent random variables.}
\newblock In {\em Proc. Sixth Berkeley Symp. Math. Stat. Prob.\/}, 583--602
(1972).
\bibitem[{Tchetgen~Tchetgen and VanderWeele(2012)}]{tchetgen2010causal}
Tchetgen~Tchetgen, E.~J. and VanderWeele, T.
\newblock \enquote{On causal inference in the presence of interference.}
\newblock {\em Statistical Methods in Medical Research\/}, 21(1):55--75 (2012).
\bibitem[{Thomas(2013)}]{thomas2013comment}
Thomas, A.~C.
\newblock \enquote{The social contagion hypothesis: Comment on `Social
contagion theory: Examining dynamic social networks and human behavior'.}
\newblock {\em Statistical in Medicine\/}, 32(4):581--590 (2013).
\bibitem[{Tsiatis(2007)}]{tsiatis2007semiparametric}
Tsiatis, A.
\newblock {\em Semiparametric theory and missing data\/}.
\newblock Springer Science \& Business Media (2007).
\bibitem[{Vallender(1974)}]{vallender1974calculation}
Vallender, S.
\newblock \enquote{Calculation of the Wasserstein distance between probability
distributions on the line.}
\newblock {\em Theory of Probability \& Its Applications\/}, 18(4):784--786
(1974).
\bibitem[{van~der Laan(2014)}]{vanderlaan2012}
van~der Laan, M.~J.
\newblock \enquote{Causal Inference for a Population of Causally Connected
Units.}
\newblock {\em Journal of Causal Inference\/}, 0(0):2193--3677 (2014).
\bibitem[{Van~der Laan and Robins(2003)}]{van2003unified}
Van~der Laan, M.~J. and Robins, J.~M.
\newblock {\em Unified methods for censored longitudinal data and causality\/}.
\newblock Springer Science \& Business Media (2003).
\bibitem[{Van~der Laan and Rose(2011)}]{van2011targeted}
Van~der Laan, M.~J. and Rose, S.
\newblock {\em Targeted learning: causal inference for observational and
experimental data\/}.
\newblock Springer Science \& Business Media (2011).
\bibitem[{van~der Laan~Mark et~al.(2007)van~der Laan~Mark, Polley~Eric
et~al.}]{van2007super}
van~der Laan~Mark, J., Polley~Eric, C., et~al.
\newblock \enquote{Super learner.}
\newblock {\em Statistical Applications in Genetics and Molecular Biology\/},
6(1):1--23 (2007).
\bibitem[{Van Der~Vaart and Wellner(1996)}]{van1996weak}
Van Der~Vaart, A.~W. and Wellner, J.~A.
\newblock \enquote{Weak Convergence.}
\newblock In {\em Weak Convergence and Empirical Processes\/}, 16--28. Springer
(1996).
\bibitem[{VanderWeele(2010)}]{vanderweele2010direct}
VanderWeele, T.
\newblock \enquote{Direct and indirect effects for neighborhood-based clustered
and longitudinal data.}
\newblock {\em Sociological Methods \& Research\/}, 38(4):515--544 (2010).
\bibitem[{VanderWeele and
Tchetgen~Tchetgen(2011{\natexlab{a}})}]{vanderweele2011bounding}
VanderWeele, T. and Tchetgen~Tchetgen, E.
\newblock \enquote{Bounding the Infectiousness Effect in Vaccine Trials.}
\newblock {\em Epidemiology\/}, 22(5):686 (2011{\natexlab{a}}).
\bibitem[{VanderWeele and
Tchetgen~Tchetgen(2011{\natexlab{b}})}]{vanderweele2011effect}
---.
\newblock \enquote{Effect partitioning under interference in two-stage
randomized vaccine trials.}
\newblock {\em Statistics \& probability letters\/}, 81(7):861--869
(2011{\natexlab{b}}).
\bibitem[{VanderWeele et~al.(2012)VanderWeele, Ogburn, and
Tchetgen}]{vanderweele2012statistics}
VanderWeele, T.~J., Ogburn, E.~L., and Tchetgen, E. J.~T.
\newblock \enquote{Why and when 'flawed' social network analyses still yield
valid tests of no contagion.}
\newblock {\em Statistics, Politics, and Policy\/}, 3(1) (2012).
\bibitem[{Wasserman(2013)}]{wasserman2013comment}
Wasserman, S.
\newblock \enquote{Comment on ``Social contagion theory: Examining dynamic
social networks and human behavior'' by Nicholas Christakis and James
Fowler.}
\newblock {\em Statistics in Medicine\/}, 32(4):578--580 (2013).
\bibitem[{Watts and Strogatz(1998)}]{watts1998collective}
Watts, D.~J. and Strogatz, S.~H.
\newblock \enquote{Collective dynamics of 'small-world' networks.}
\newblock {\em Nature\/}, 393(6684):440--442 (1998).
\bibitem[{Young et~al.(2014)Young, Hern{\'a}n, and
Robins}]{young2014identification}
Young, J.~G., Hern{\'a}n, M.~A., and Robins, J.~M.
\newblock \enquote{Identification, Estimation and Approximation of Risk under
Interventions that Depend on the Natural Value of Treatment Using
Observational Data.}
\newblock {\em Epidemiologic Methods\/}, 3(1):1--19 (2014).
\end{thebibliography}