Lecture 4 - Interference, Spillovers and Dynamics

Paul Goldsmith-Pinkham
January 25, 2024

This lecture note will discuss what it means to relax the following
assumptions from the previous lecture:

1. Binary scalar treatment

2. Single time period (e.g. one treatment within the person)

3. SUTVA – Stable Unit Treatment Value Assignment

Multivalued treatments

So far, our discussion of treatment effects has focused on single bi-

nary treatments. This made life very easy, but we have a lot of other
more complex settings. We’ll consider a few different cases. First, a
multi-valued treatment. Then, we’ll consider a continuous treatment.
Finally, we’ll consider an unordered multi-valued treatment.

Discrete multi-valued treatment

Let’s start with a discrete, multi-valued treatment to start: Di ∈ D =
{0, 1, . . . , d}. This captures a simple setting like “what is the impact
of 0, 1, 2, or 3 children on labor force participation?” In this setting,
we can easily shift the scale up and down (“what is the impact of
5, 10, 15, or 20 minutes on a task”) but the order and spacing may
matter, depending on how we choose to parameterize and estimate
the treatment effect.
In this setting, how should we consider the treatment effects?
First, what should we consider the “control”? We could consider the
control to be the lowest value of the treatment, but that depends a bit
on the context. For example, if we are considering the impact of 0, 1,
2, or 3 children on labor force participation, we might consider the
control to be 0 children. However, if we are considering the impact of
5, 10, 15, or 20 minutes on a task, we might consider the control to be
whatever the status quo was.
Formally, we define the potential outcome for any d ∈ D as Yi (d),
and we consider the individual and average treatment effect differ-
ence between d and d′ as:

τi (d, d′ ) = Yi (d) − Yi (d′ )

E(τi (d, d′ )) = E(Yi (d) − Yi (d′ )).
 lecture 4 - interference, spillovers and dynamics 2

If strong ignorability holds, then this is also identified by simply

conditioning on each observed value:1 1
The overlap condition in strong ignor-
ability with multiple treatments is more
E(τi (d, d′ )) = E(Yi | Di = d) − E(Yi | Di = d′ ). complicated, but effectively entails that
for any X, there are observations for
every d in D .
This type of estimation is non-parametric in nature: we’ve as-
sumed no functional form between the treatment and the potential
outcome. A consequence of that, just like in the case with many
covariates X, is that it requires a lot more data to provide precise
estimates. If we wanted to consider the CATE:

E(τi (d, d′ )|X = x ) = E(Yi | Di = d, X = x ) − E(Yi | Di = d′ , X = x ),

then we’ll need to condition on treatment catgories within each cell,

which can be very data hungry, and less precise.
Often, instead of estimating the effect for every point separately,
we will postulate a model for the potential outcomes:

Yi (d) = Yi (0) + τi d.

Notice that in this case, this implies that for all d, d′ pairs

τi (d, d′ ) = τi ,

which is the slope parameter. Hence, estimation can be made more

precise by pooling all of these estimands together into a single esti-
mand E(τi ).2 2
Other, more flexible parametric forms
for Yi (d) could be chosen as well.
The insights will carry through so
Example 1 long as the functional form is finite-
Consider the following simulated data, where the true effect is linear dimensional.
and simulated such that E(τi (d, d′ ) = d′ − d and strong ignorability
holds.
Each dot in Figure 1 is the estimated mean at the point, and we
find a positive treatment effect. Imposing the model helps a lot com-
pared to non-parametric form. To see this, consider the treatment
effect comparing d to d − 1 in Figure 2
The direction of the effect is much more ambiguous. This is a com-
mon tradeoff in estimation: imposing a model can help with precision,
but can also lead to bias if the model is misspecified.

How should we consider these functional forms once we include

controls? To see what I mean, consider the same context, but we now
assume strong ignorability conditional on X. Then, we would need
to estimate the slope τi for each value of X. How would that map
over to a linear regression model? The simplest version would be
one where the heterogeneity, τi , is uncorrelated with X. Then, one
 lecture 4 - interference, spillovers and dynamics 3

Outcome
10

5 b=1.178

−5
1 2 3 4 5
Treatment D
Figure 1: Linear effects estimated in
simulated data with a true linear model
for Example 1

Estimated effects
5.0

2.5

0.0

−2.5

2 3 4 5
Treatment D vs D−1
Figure 2: Non-parametric effects esti-
mated in simulated data with a true
linear model for Example 1
 lecture 4 - interference, spillovers and dynamics 4

way to estimate the ATE is to assume that E( Di |X) = Xi β (e.g. the

propensity score is linear in X), we could estimate E(τi ) by simply
running the following regression:

Yi = α + Di τ + Xi β + ϵi (1)

and using τ as our estimate of E(τi ). But, if τi is not uncorrelated

with X, then to estimate the ATE we would need to estimate the slope
for each value of X, and pool separately. See Comment 1 for more
details.

Comment 1
It is worth thinking about why Equation (1) will correctly estimate
the ATE in this setting. To do this, let E∗ ( Di |Xi ) denote the best
linear predictor of Di conditional on Xi . Now, note that by Frisch-
Waugh-Lovell,
Ỹi = α̃ + D̃i τ + ui , (2)
where Ỹi = Yi − E∗ (Yi |Xi ) and D̃i = Di − E∗ ( Di |Xi ). Then,

E( D̃i Yi )
τ=
E( D̃i2 )
E( D̃i Yi (0)) E( D̃i Di τ )
= +
E( D̃i2 ) E( D̃i2 )

The first term is zero because the residual D̃i is mean independent of
X by the linearity of E( Di |X). Therefore,

E( D̃i Yi (0)) = E( E( D̃i Yi (0)|X)) = E( E( D̃i |X) E(Yi (0)|X)) = 0.

Now note the second term by similar arugments:

E( D̃i Di τ ) E(Var ( Di |X) E(τ |X))

τ= 2
= .
E( D̃i ) E(Var ( Di |X))

But, since we assumed E(τ |X) = E(τ ), this is just the ATE. If there
is correlation of τ with X, we will get a different estimand.
 lecture 4 - interference, spillovers and dynamics 5

Discussion Questions 1
Under what assumptions about E(Yi (0)|Xi ), instead of E( Di |Xi ),
could we estimate the ATE using Equation (1)?

1. Hint: think about a linear model for E(Yi (0)) that is linear in Xi .

2. Second hint: the estimand could also be written as

E( D̃i Ỹi )
τ=
E( D̃i2 )

Continuous valued treatment

In many cases, the jump from discrete ordered treatments to contin-
uous valued treatments is not large. Often, it just has to do with how
many repeated observations we have of the same treatment; if each
treatment value is unique, we’re more likely to treat it as continu-
ous. None of what we discussed above changes, except that direct
non-parametric estimation becomes infeasible.
Instead, to do non-parametric estimation we’ll need to make other
assumptions and use other methods, like kernel regression or local
linear regression. We’ll discuss these in future classes, but the key
point is that we will want to make some amount of smoothness as-
sumptions the effect of the treatment on the potential outcome.
Instead of non-parametric estimation, it is also reaosnable to as-
sume a functional form, as above, and proceed from there. Then
everything is exactly the same.

Unordered multi-valued treatment

Finally, we might have a setting where the treatment is unordered.
For example, we might consider the impact of different CEOs on
firms’ performance. In this case, we can’t assume any particular
ordering of the treatment, and it’s not clear how to presume a func-
tional form. Instead, there is a set of K treatments in D , and we can
consider a set of different contrasts between them: E(τi (d, d′ )) for all
d, d′ ∈ D .
A straightforward special case would be to consider a factorial
design: a randomized treatment where two treatments are cross-
randomized, such that an individual can receive either no treatment,
treatment 1, treatment 2, or both. Then, our potential outcomes look
like Table 1.
Given this, we have a number of potential estimands to con-
sider. For example, we could consider the average treatment effect
of treatment 1, but we would need to make a decision on what to
 lecture 4 - interference, spillovers and dynamics 6

do about the individuals who received both treatments. If the treat-

ments interact in some way, then the average treatment effect of treat-
ment 1 is not well-defined. Instead, we might consider the average
treatment effect of treatment 1 for those who received treatment 2
(E(Yi (1, 1) − Yi (0, 1))), and the average treatment effect of treatment 1
for those who did not receive treatment 2 (E(Yi (1, 0) − Yi (0, 0))).

Yi (Di ) D1i = 0 D1i = 1 Table 1: Potential outcomes for a

factorial design
D2i = 0 Yi (0, 0) Yi (1, 0)
D2i = 1 Yi (0, 1) Yi (1, 1)

Of course, when data is sparse, we might want to do more with

this, and just pool all of the data together: E(Yi (1, Di2 ) − Yi (0, Di2 )).
This is a reasonable estimand if we believe that the treatment effects
are constant across the different levels of the other treatment, but if
there are interactions, the external validity of this estimate will be
suspect.3 See Banerjee et al. [2021] for a very interesting discussion 3
Think about why this is the case, if it’s
on how to think about these types of estimands in the context of not clear.

factorial designs when there are many treatments.

Factorial designs are the simplest case to consider, because the
treatments are cross-randomized for many binary treatments. Often,
we have just an ordered set of treatments. In this case, the same
logic applies, but we have to be more careful about how we define
the estimands. For example, if we consider the impact of different
CEOs on firm performance, how do we define a “control”? It is often
not obvious, and we need to be careful about how we define the
estimands. We may instead focus on the conditional means for each
treatment, and then consider the full distribution of effects. This is
the type of consideration in work thinking about place-based effects,
for example, such as Chetty and Hendren [2018].4 4
It is interesting to think about these
Estimating these effects seem like they should be straightforward unordered treatments can sometimes
be projected into continuous scalar
extensions of the binary treatment case. However, the partial linear measures. For example, when consid-
regression model fails us in this case. The variation in the propensity ering the impact of a CEO on a firm,
you might measure a CEO’s experience,
score across strata (controls) combined with heterogeneity in the and project the overall CEO’s effect
treatment effects across strata will lead to contamination bias in the onto experience to capture a contin-
linear regression model. See Goldsmith-Pinkham et al. [2022] for uous measure of the CEO’s effect. If
you are additionally willing to assume
more details, which we will revisit in the linear regression lectures. that the effect of experience is the only
Another issue that can arise is when the treatments are not cross- channel controlling the CEO’s effect,
then a more efficient procedure would
randomized, but instead are correlated on one another. A simple ex- use experience as the treatment effect,
ample of this is sequenced treatments: e.g. treatment 2 is only given instead of the CEO. But that may not
after treatment 1, and only a subset of individuals receive treatment be a reasonable assumption. This issue
arises when considering the effect of
1. See Figure 3 for an example of this. In this case, it is not possible judges as in Arnold et al. [2022].
to identify the effect of D2 separately from D1 : E(Yi (0, 1) − Yi (0, 0))
is not identified because E(Yi (0, 1)) is never observed. This rarely
 lecture 4 - interference, spillovers and dynamics 7

D1
happens in many cross-sectional settings, but is quite common in
dynamic settings (our next topic).

D2 Y
Treatment dynamics Figure 3: Correlated treatments

We will briefly discuss the impact of treatments over time to set

the stage for our study of panel data later in the class. Consider
a setting where we now observe T time periods for a unit: Yi =
(Yi1 , Yi2 , . . . , YiT ). Now, for each time period, there is a treatment Dit .
It would be convenient to simply consider Yit ( Dit ) as the potential
outcome for an individual i in period t, but that would be very re-
strictive: it would assume that only the period t treatment affects
the outcome in period t. A more general form would define a vector
Di = ( Di1 , Di2 , . . . , DiT ), and define the potential outcome in period
t as Yit (Di ). In this case, however, we are perhaps too general: this
allows for treatments in the future to affect current outcomes, which
may be too strong.5 5
This is often referred to as “anticipa-
There are a large number of ways to simplify these potential out- tory effects” in the literature.

comes. One simple way would be to restrict treatments to only affect

outcomes in the future, and not the past. This is often referred to
as the “no anticipation” assumption. The second is to assume that
treatments will only turn on once: this allows the researcher to only
consider the adoption date as the relevant period.6 6
This is common in the staggered
As you can see, things become much more complex as soon as difference-in-difference setting.

you allow for dynamic effects. In order to make progress, it will

often be necessary to make restrictions on teh dynamics to make the
estimands identified. We will discuss these in more detail when we
discuss difference-in-differences.

The SUTVA hits the fan

In the discussion so far, the “interference” between treatments just

comes from having multiple treatments to worry about, or from
spillover across time. However, there are many other ways that treat-
ments can interfere with one another. For example, what if treat- D1 Y2
ments spill across units? What if the treatment of one unit affects the
potential outcomes of another unit?
Recall the key assumption of Stable Unit Treatment Value As- D2 Y1
sumption (SUTVA): the potential outcomes of a unit do not vary with Figure 4: Interference between units
the treatment of other units. When could this be violated?

So many places

Why does failure of SUTVA create an issue? Recall our discussion

regarding marginal estimands when there were multiple treatments:
 lecture 4 - interference, spillovers and dynamics 8

even with random assignment, the estimates effect will be contam-

inated by others’ treatment status, thereby leading to estimates that
are not informative for the policy maker.
This type of problem is generally referred to as “interference.” It
is challenging for identification, estimation and inference. For now,
we’ll focus on identification. I will flag three versions of this problem:

1. Social interactons and peer effects

2. Spatial spillovers

3. Economic interactions – budget constraints, etc.

All these problems are versions of violation of SUTVA. With a clean,

well-identified experiment, it is still possible to identify interesting
estimands, but we may have to substantially modify our traditional
estimators or make strong assumptions to make progress. One way
to view this fact is that our original setting — SUTVA, binary treat-
ment and a single time period — is a very special (and somewhat
unrealistic) case.

Social interactions and peer effects

A variety of terms in common use connote endogenous social effects, wherein the propensity of an individual to
behave in some way varies with the prevalence of that behaviour in some reference group containing the indi-
vidual. These effects may, depending on the context, be called "social norms", "peer influences", "neighbourhood
effects", "conformity", "imitation", "contagion", "epidemics", "bandwagons", "herd behaviour", "social interac-
tions", or "interdependent preferences".
— Manski [1993]
Manski (1993) spawned a huge literature, much of which focused
on the linear-in-means model.7 . An inherent issue, in my view, is 7
There are theoretical models micro-
that many empirical papers jumped to this construction immedi- founding a linear-in-means outcome
model, which typically involve some
ately. They did not have a structural interpretation in mind, but were kind of quadratic cost to deviating
instead interested in testing for the statistical presence of spillovers from the group. See Shue [2013] for an
example
across individuals.
 lecture 4 - interference, spillovers and dynamics 9

Comment 2 (Historical context on peer effects)

Manski [1993] focused on a linear-in-means structural equation

Y = βE(Y | g) + γ1 E( X | g) +γ2 X
| {z } | {z }
endogeneous exogeneous

+ γ3 g +u.
|{z}
contextual

Peers were not well-defined in the model, but empirically, were usu-
ally groups like classrooms or clubs. What is important to note about
this model is that it is a structural model of the outcome, Y. The
reduced form is

Y = γ1 /(1 − β) E( X | g) + (γ2 /(1 − β)) X

+ (γ3 /(1 − β)) g + ũ,

which is estimable under special exogeneity assumptions on X, g

and E( X | g).
An innovation in this literature was to start using network data
to define the group structure. Bramoullé et al. [2009] was a key
paper in this literature, that reframed the Manski linear-in-means
model to

Y = βAY + γ1 AX + γ2 X + ϵi ,
Y = ( I − βA)−1 γ1 AX + ( I − βA)−1 γ2 X + ( I − βA)−1 ϵi

where A was an n × n matrix of individuals’ connections. This was

still a structural model, but allowed for richer data and more easily
identified the effect of peers.

Given that most researchers studying peer effects were not initially
motivated by a structural model, it seems more natural to initially
take a statistical approach to the problem. Namely, we would like to
identify the effect of spillovers across units. How can we approach
this problem using the tools we’ve developed so far?
Given n individuals, for person i, how much interference can we
allow? What types?
Yi ( D1 , D2 , . . . , Dn )

is far more extreme than

Yi ( Di , ADn ).

This question is analogous to our setting with treatment dynamics:

how much spillover should we allow? SUTVA is complicated by the
 lecture 4 - interference, spillovers and dynamics 10

fact that there is no natural “no anticipation” condition due to the

natural flow of time. As you might expect, there is no “one solution”
in this setting. Certain restrictions need to be made to identify some
estimands.
Manski [2013] is a very nice discussion of this in a very high-level
way. One key assumption he highlights is that “anonymity” of treat-
ment spillover is a very important assumption. This implies that if
I have peers who are treated, it does not matter which of those peers
are treated – the impact on me is identical. This is a very strong as-
sumption, but it is a necessary one to identify the effect of peers. If
each peer’s effect is allowed to be unique, then the effect of peers is
not identified, since there is no way to separate out the treatment’s
spillover effect from the effect of the peer itself.8 8
It might be doable in some networks,
A key question to keep in mind when considering spillovers: are but it would be very challenging to do
so, and require exogeneous network
you attempting to estimate the spillover effect, or are you attempt- connections.
ing to identify individual ATE in the presence of spillovers? These
are very different estimands, and require different assumptions to
identify. For the purposes of external validity, the latter is really only
relevant if the context you apply the treatment in would have limited
spillovers as well.
We now briefly discuss two papers in this space to give intuition.

Aronow and Samii [2017] is a lynchpin paper in this setting that pro-
vides a framework for thinking about estimation and identification
under general forms ofb interference. They use design-based infer-
ence, and consider the following generalized mapping.
Definition 1
For any generalized vector of interventions, Dn , there’s an experimental
design which assigns probabilities over Dn . There is then an exposure
mapping f (Dn , θi ) from these vectors to a treatment for an individual,
which includes traits of an individual, θi (e.g. their network location) and
the treatment vector, and maps it to an exposure outcome.

This exposure mapping does two things. First, it makes restric-

tions on types of interactions (e.g. who can affect you and what
type of effect it is).9 Second, it maps the experimental design to a 9
Concretely: consider a network of
peers affecting you. Is it the sum of
propensity score of the exposure treatment. This allows the use of
your connected individuals in your
Horvitz-Thompson estimators. network? Any exposure at all? Does it
So where are the bodies buried in this method? You have to have matter who in your network exposes
you?
a correctly defined exposure mapping, and you have to have a cor-
rectly defined experimental design. In the case of a randomized
experiemnt, the latter is straightforward, but the former is not, and
typically needs to be motivated by theory, or asssessed for robust-
ness. This is an active literature.
 lecture 4 - interference, spillovers and dynamics 11

Athey et al. [2018] studies null hypothesis tests in networks under

intereference. A key feature that this paper adds: testing specific
types of analysis by creating “artificial” experiments. This paper is
particularly powerful because it allows for testing in settings where
there is uncertainty about the exposure mapping, and gives a frame-
work for thinking about testing in a single network.

Comment 3
When thinking about experiments in networks (and other settings),
the structure of spillovers is very important. It is extraordinarily
helpful to identify settings where there are zero spillovers. Having
units (such as villages, roommate pairs, etc.) that are isolated from
one another is a very helpful way to identify the effect of the treat-
ments. If we permute the treaments across these groups, then we can
assess the spillover effects in a very clean way.
If, instead, we have only a single network, then we need to make
strong assumptions about the structure of the network to identify the
spillover effects. Namely, we need to have a well-defined exposure
mapping that asserts that some units are sufficiently independent
from others to serve as control units.

It is already very hard to do research on spillovers. Make sure to

not ignore the difficult identification challenges and assumptions that
you’ll need to make. If you need a model, that’s great! But often you
are just interested in starting from a statistical perspective, which
suggests you shoudl focus on a design-based approach as in Aronow Figure 5: My views on social interac-
and Samii [2017]. tions summed up

Spatial Spillovers
Much of the spatial literature has sat in the same literature as social
interactions. Distance on a network graph can be viewed as a sim-
ilar distance metric to geographic (or economic) distance.Similar A
matrix, and consequentially similar structural models are proposed.
The Aronow and Samii [2017] setting allows for this as well. From
an identification standpoint, there isnothing deeply different here rel-
ative to networks, except that distance is potentially more continuous
/ complex. When we revisit simulated instruments, we will discuss
some interesting implications raised by Borusyak and Hull [2020].

Economic interactions
Consider the following simple experiment – I give one half of people
in the economy checks for $2000 dollars. I then study the impact of
 lecture 4 - interference, spillovers and dynamics 12

these checks on their consumption. Why might the effects be differ-

ent than if I had run this experiment on a small share of individuals?
The economic spillovers coming through budget constraints are
hugely important, but also deeply challenging as well. They require,
often, modeling assumptions about spillovers. I will discuss two ex-
amples from the literature to give a flavor of the issues and solutions.

Chodorow-Reich [2019] studies the impact of fiscal stimulus on local

employment. The key identification strategy is to use cross-region in-
cidence of fiscal stimulus to identify multipliers on local employment.
The paper argues that cross-region evidence bounds the estimand of
interest, the impact of a national stimulus, from below.
Drawing on theoretical explorations, I argue that the typical empirical cross-sectional multiplier study provides a
rough lower bound for a particular, policy-relevant type of national multiplier, the closed economy, no-monetary-
policy-response, deficit-financed multiplier. The lower bound reflects the high openness of local regions, while
the “rough” accounts for the small effects of outside financing common in cross-sectional studies.

Sraer and Thesmar [2023] use cross-firm experiment to influence the

allocation of credit. Some firms got lots more credit! Some did not.
How to aggregate up this affect? E.g. the policy effect is estimated
by differencing the impact of the change on those who were more
directly exposed vs. not – however, this doesn’t tell us about the ag-
gregate impact on the economy. The paper argues, using economic
theory, that these issues can be safely ignored under certain assump-
tions.
Our paper bridges these two approaches. We offer a method to measure allocative efficiency in a (quasi-) experi-
mental settings. This method works as follows. An econometrician observes firm-level data in an economy where
a (quasi-) natural experiment has taken place. This experiment changes the set of frictions faced by treated firms
while leaving control firms unaffected. Under the appropriate identifying assumption, the econometrician can
estimate the causal effect of the experimenton firm-level outcomes, using classic difference-in-difference estima-
tors. Standard policy evaluations typically estimate treatment effects on firm size or employment. However, these
treatment effects alone cannot speak to allocative efficiency. To do so, we show that the econometrician needs
to estimate treatment effects on the distribution of log marginal products of capital (lMRPKs). These estimates
can then be injected in a simple aggregation formula to answer two simple questions: (i) how much did the ac-
tual policy change contribute to changes in aggregate efficiency (ex post evaluation)? (ii) how would aggregate
efficiency have changed if the policy had been extended to all firms in the economy (scale-up)?

References

David Arnold, Will Dobbie, and Peter Hull. Measuring racial dis-
crimination in bail decisions. American Economic Review, 112(9):
2992–3038, 2022.

Peter M Aronow and Cyrus Samii. Estimating average causal effects

under general interference, with application to a social network
experiment. Annals of Applied Statistics, 11(4):1912–1947, 2017.
 lecture 4 - interference, spillovers and dynamics 13

Susan Athey, Dean Eckles, and Guido W Imbens. Exact p-values for
network interference. Journal of the American Statistical Association,
113(521):230–240, 2018.

Abhijit Banerjee, Arun G Chandrasekhar, Suresh Dalpath, Esther Du-

flo, John Floretta, Matthew O Jackson, Harini Kannan, Francine N
Loza, Anirudh Sankar, Anna Schrimpf, et al. Selecting the most
effective nudge: Evidence from a large-scale experiment on immu-
nization. Technical report, National Bureau of Economic Research,
2021.

Kirill Borusyak and Peter Hull. Non-random exposure to exogenous

shocks: Theory and applications. Technical report, National Bureau
of Economic Research, 2020.

Yann Bramoullé, Habiba Djebbari, and Bernard Fortin. Identification

of peer effects through social networks. Journal of econometrics, 150
(1):41–55, 2009.

Raj Chetty and Nathaniel Hendren. The impacts of neighborhoods on

intergenerational mobility ii: County-level estimates. The Quarterly
Journal of Economics, 133(3):1163–1228, 2018.

Gabriel Chodorow-Reich. Geographic cross-sectional fiscal spending

multipliers: What have we learned? American Economic Journal:
Economic Policy, 11(2):1–34, 2019.

Paul Goldsmith-Pinkham, Peter Hull, and Michal Kolesár. Contami-

nation bias in linear regressions. Technical report, National Bureau
of Economic Research, 2022.

Charles F Manski. Identification of endogenous social effects: The

reflection problem. The review of economic studies, 60(3):531–542,
1993.

Charles F Manski. Identification of treatment response with social

interactions. The Econometrics Journal, 16(1):S1–S23, 2013.

Kelly Shue. Executive networks and firm policies: Evidence from the
random assignment of mba peers. The Review of Financial Studies, 26
(6):1401–1442, 2013.

David Sraer and David Thesmar. How to use natural experiments to

estimate misallocation. American Economic Review, 113(4):906–938,
2023.