Structural Behavioral Economics*
UC Berkeley and NBER
What is the role of structural estimation in behavioral economics? I discuss advantages, and
limitations, of the work in Structural Behavioral Economics, cover common modeling choices,
and how to get started. Among the advantages, I argue that structural estimation builds on,
and expands, a classical behavioral tool, back-of-the-envelope calibrations, and that it benefits
from the presence of a few parsimonious behavioral models which can be taken to the data.
Estimation is also well suited for experimental work, common in behavioral economics, as it can
lead to improvements in the experimental design. Also, at a time where policy implications of
behavioral work are increasingly discussed, it is important to ground these policy implications
in (estimated) models. Structural work, however, has important limitations, which are relevant
to its behavioral applications. Estimation takes much longer and the extra degree of complexity
can make it difficult to know which of a series of assumptions is driving the results. For related
reasons, it is also easy to over-reach with the welfare implications. Taking this into account, I
provide a partial how-to guide to structural behavioral economics: (i) the choice of estimation
method; (ii) the modeling of heterogeneity; (iii) identification and sensitivity; and (iv) common
issues for the estimation of leading behavioral models. I illustrate this discussion with selected
coverage of existing work in the literature.
*In Preparation for the 1st Handbook of Behavioral Economics, edited by Douglas Bernheim, Stefano DellaVigna,
and David Laibson, Elsevier. I thank Hunt Allcott, Charles Bellemare, Daniel Benjamin, Douglas Bernheim, Colin
Camerer, Vincent Crawford, Thomas Dohmen, Philipp Eisenhauer, Lorenz Goette, Johannes Hermle, Lukas Kiessling,
Nicola Lacetera, David Laibson, John List, Edward O’Donoghue, Gautam Rao, Alex Rees-Jones, John Rust, Jesse
Shapiro, Charles Sprenger, Dmitry Taubinsky, Bertil Tungodden, Hans-Martin von Gaudecker, George Wu, and the
audience of presentations at the 2016 Behavioral Summer Camp, at the SITE 2016 conference, and at the University
of Bonn for their comments and suggestions. I thank Bryan Chu, Avner Shlain, and Alex Steiny for outstanding
Behavioral economics, with its lessons regarding non-standard preferences, beliefs, and decision-
making, has important applications to most fields of economics. This Handbook is a clear illustration
of this broad reach, with chapters on a variety of fields, including finance, public economics, and
The empirical applications in these fields utilize a range of data sources—observational studies,
survey collection, laboratory experiments, and field experiments, among others. The applications
also come with a variety of estimation methods, including simple treatment-control comparisons in
experiment, correlations, instrumental variables, but also structural estimation.
In this chapter I ask: Is there an important role for structural estimation1 in behavioral eco-
nomics, or for short Structural Behavioral Economics? Are there special lessons beyond the well-
known advantages, such as the ability to do welfare and policy evaluations, but also the well-known
pitfalls, such as the complexity of the analysis?
I argue that the answer is: Yes, and Yes. In Section 2 I discuss six advantages of structural
estimation, several of which have roots in key features of behavioral research. One of the most
important ones is that estimation builds on a long-standing tradition in behavioral economics of
back-of-the-envelope calibration of models. Indeed, one of the most influential results in behavioral
economics is Rabin’s calibration theorem for risk (Rabin, 2000). As I argue, estimation takes the
calibration one step further, including cases in which a simple back-of-the-envelope calibration is
Structural estimation also benefits from the fact that behavioral economics has a small number
of widely-used parsimonious models, such as beta-delta preferences (Laibson, 1997; O’Donoghue
and Rabin, 1999a) and reference dependence (Kahneman and Tversky, 1979). The presence of
commonly-used models makes it more useful to test for the stability of estimates across settings and
to examine the out-of-sample performance of models.
Relatedly, in behavioral economics there has always been a healthy exchange of ideas between
theorists and applied researchers. Partly because of the focus on calibrating models, behavioral
theorists have paid attention, arguably more than in some other fields, to empirical evidence. Con-
versely, empirical researchers, given the importance of testing the null hypothesis of the standard
model, have typically paid close attention to the development of behavioral theory, or at least ap-
plied theory. Structural estimation builds on, and reinforces, this closeness, as it forces empirical
researchers to take seriously models which they take to the data.
An additional feature of the behavioral field is the importance of experimental evidence, both
from the laboratory, where much of the initial behavioral evidence came from, and from the field.
In experiments, there are extra advantages to estimation: paying close attention to the models at
the design stage can lead to different designs that allow for a clearer test of models. In observational
studies, in contrast, the design is limited by the the data and the setting (though models can of
course motivate the search for the right observational design). This particular advantage of models,
1I define structural as the “estimation of a model on data that recovers parameter estimates (and confidence
intervals) for some key model parameters”. For definitions of structural estimation see Reiss and Wolak (2007),
Wolpin (2013), and Rust (2014).
interestingly, so far has played a larger role in laboratory experiments than in field experiments
(Card, DellaVigna, and Malmendier, 2011). There is an opportunity for more work of this type.
An additional motivation for structural analysis is clearly shared by all applications of struc-
tural estimation: welfare and policy analysis. The timing for that in behavioral economics is just
right. While behavioral economics has mostly shied away from policy implications until the more
recent decade, the recent emphasis on cautious paternalism (Camerer et al., 2003), nudges (Thaler
and Sunstein, 2008), and behavioral welfare economics (Bernheim and Rangel, 2009) substantially
increased the policy reach of behavioral. Yet, many policy applications of behavioral findings do not
have a fully worked out welfare or policy evaluation. Structural estimation has a role to play.
Having said this, should all of behavioral economics be structural? Absolutely not. To start with,
many studies do not lend themselves well to structural estimation, for example because they explore
a channel for which we do not have yet a well-understood model, e.g., framing effects, or the interest
is on a reduced-form finding. In addition, even in cases in which there is an obvious model-data
link, an alternative strategy is to derive comparative statics from the model (e.g., Andreoni and
Bernheim, 2009), including in some cases even an axiomatic characterization (e.g., Halevy, 2015),
to derive empirical testable predictions. This strategy allows for clear model testing, without the
extra assumptions and time involved in structural estimation.
For the studies where structural estimation makes sense, in Section 3 we outline common limita-
tions of structural estimation. These limitations are shared with applications of structural estimation
in other fields, but I emphasize examples, and specific issues, within behavioral economics.
First, and perhaps most obviously, structural estimation typically takes much more time, given
the number of necessary steps, starting from the reduced-form results, spelling out the full model,
the estimation strategy, and getting to reliable estimates. The estimation itself can be a very time-
consuming step, and indeed much of the training for work in the structural area revolves around
computational short-cuts and techniques to ensure that the results are robust. An implication is
that structural analysis, being more complex, also increases the chance that programming errors may
drive the results, or that the estimates may not be stable. These important time and complexity
costs must be weighed against the benefits above.
A possible saving grace from this cost is well-known in the literature: sufficient statistics (Chetty,
2009). In some cases, a parameter, or combination of parameters, can be estimated using a key
statistic, or a combination of statistics, that is sufficient for estimation (hence the name). Sufficient
statistics, thus, allow for structural estimation (of some parameters) using reduced-form findings,
without a full estimation. This is a desirable route when possible, and has been used for example
to obtain estimates of simple limited attention model