EC10CH18_Abadie ARI 28 June 2018 11:30
along which it is expanding, discussing recent developments and the areas where research may be
particularly fruitful in the near future.
Other recent surveys on the estimation of causal treatment effects and the econometrics of
program evaluation from different perspectives and disciplines include those by Abadie (2005a),
Angrist & Pischke (2008, 2014), Athey & Imbens (2017c), Blundell & Costa Dias (2009), DiNardo
& Lee (2011), Heckman & Vytlacil (2007), Hern
´
an & Robins (2018), Imbens & Rubin (2015),
Imbens & Wooldridge (2009), Lee (2016), Manski (2008), Pearl (2009), Rosenbaum (2002, 2010),
Van der Laan & Robins (2003), and VanderWeele (2015), among many others.
2. CAUSAL INFERENCE AND PROGRAM EVALUATION
Program evaluation is concerned with the estimation of the causal effects of policy interven-
tions. These policy interventions can be of very different natures depending on the context of the
investigation, and they are often generically referred to as treatments. Examples include condi-
tional transfer programs (Behrman et al. 2011), health care interventions (Finkelstein et al. 2012,
Newhouse 1996), and large-scale online A/B studies in which IP addresses visiting a particular
web page are randomly assigned to different designs or contents (see, e.g., Bakshy et al. 2014).
2.1. Causality and Potential Outcomes
We represent the value of the treatment by the random variable W. We aim to learn the effect of
changes in treatment status on some observed outcome variable, denoted by Y. Following Neyman
(1923), Rubin (1974), and many others, we use potential outcomes to define causal parameters:
Y
w
represents the potential value of the outcome when the value of the treatment variable, W,is
set to w. For each value of w in the support of W, the potential outcome Y
w
is a random variable
with a distribution over the population. The realized outcome, Y, is such that, if the value of the
treatment is equal to w for a unit in the population, then for that unit, Y = Y
w
, while other
potential outcomes Y
w
with w
= w remain counterfactual.
A strong assumption lurks implicit in the last statement. Namely, the realized outcome for
each particular unit depends only on the value of the treatment of that unit and not on the
treatment or on outcome values of other units. This assumption is often referred to as the stable
unit treatment value assumption (SUTVA) and rules out interference between units (Rubin 1980).
SUTVA is a strong assumption in many practical settings; for example, it may be violated in an
educational setting with peer effects. However, concerns about interference between units can
often be mitigated through careful study design (see, e.g., Imbens & Rubin 2015).
The concepts of potential and realized outcomes are deeply ingrained in economics. A demand
function, for example, represents the potential quantity demanded as a function of price. Quantity
demanded is realized for the market price and is counterfactual for other prices.
While, in practice, researchers may be interested in a multiplicity of treatments and outcomes,
we abstract from that in our notation, where Y and W are scalar random variables. In addition
to treatments and potential outcomes, the population is characterized by covariates X,a(k × 1)
vector of variables that are predetermined relative to the treatment. That is, while X and W may
not be independent (perhaps because X causes W, or perhaps because they share common causes),
the value of X cannot be changed by active manipulation of W. Often, X contains characteristics
of the units measured before W is known.
Although the notation allows the treatment to take on an arbitrary number of values, we
introduce additional concepts and notation within the context of a binary treatment, that is,
W ∈{0, 1}.Inthiscase,W = 1 often denotes exposure to an active intervention (e.g., participation
www.annualreviews.org
•
Econometric Methods for Program Evaluation 467
Annu. Rev. Econ. 2018.10:465-503. Downloaded from www.annualreviews.org
Access provided by Massachusetts Institute of Technology (MIT) on 03/19/19. For personal use only.