ARTICLE IN PRESS

Journal of Econometrics 142 (2008) 615–635

www.elsevier.com/locate/jeconom

Regression discontinuity designs: A guide to practice

Guido W. Imbensa, Thomas Lemieuxb,
a
Department of Economics, Harvard University and NBER, M-24 Littauer Center, Cambridge, MA 02138, USA
b
Department of Economics, University of British Columbia and NBER, 997-1873 East Mall, Vancouver, BC, V6T 1Z1, Canada
Available online 21 May 2007

Abstract

In regression discontinuity (RD) designs for evaluating causal effects of interventions, assignment to a treatment is
determined at least partly by the value of an observed covariate lying on either side of a fixed threshold. These designs were
first introduced in the evaluation literature by Thistlewaite and Campbell [1960. Regression-discontinuity analysis: an
alternative to the ex-post Facto experiment. Journal of Educational Psychology 51, 309–317] With the exception of a few
unpublished theoretical papers, these methods did not attract much attention in the economics literature until recently.
Starting in the late 1990s, there has been a large number of studies in economics applying and extending RD methods. In
this paper we review some of the practical and theoretical issues in implementation of RD methods.
r 2007 Elsevier B.V. All rights reserved.

JEL classification: C14; C21

Keywords: Regression discontinuity; Treatment effects; Nonparametric estimation

1. Introduction

Since the late 1990s there has been a large number of studies in economics applying and extending
regression discontinuity (RD) methods, including Van Der Klaauw (2002), Black (1999), Angrist and Lavy
(1999), Lee (2007), Chay and Greenstone (2005), DiNardo and Lee (2004), Chay et al. (2005), and Card et al.
(2006). Key theoretical and conceptual contributions include the interpretation of estimates for fuzzy
regression discontinuity (FRD) designs allowing for general heterogeneity of treatment effects (Hahn et al.,
2001, HTV from hereon), adaptive estimation methods (Sun, 2005), specific methods for choosing bandwidths
(Ludwig and Miller, 2005), and various tests for discontinuities in means and distributions of non-affected
variables (Lee, 2007; McCrary, 2007).
In this paper, we review some of the practical issues in implementation of RD methods. There is relatively
little novel in this discussion. Our general goal is instead to address practical issues in implementing RD
designs and review some of the new theoretical developments.
After reviewing some basic concepts in Section 2, the paper focuses on five specific issues in the
implementation of RD designs. In Section 3 we stress graphical analyses as powerful methods for illustrating

Corresponding author. Tel.: +1 604 822 2092; fax: +1 604 822 5915.
E-mail addresses: imbens@harvard.edu (G.W. Imbens), tlemieux@interchange.ubc.ca (T. Lemieux).

0304-4076/$ - see front matter r 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.jeconom.2007.05.001
 ARTICLE IN PRESS
616 G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635

the design. In Section 4 we discuss estimation and suggest using local linear regression methods using only the
observations close to the discontinuity point. In Section 5 we propose choosing the bandwidth using cross-
validation. In Section 6 we provide a simple plug-in estimator for the asymptotic variance and a second
estimator that exploits the link with instrumental variable methods derived by HTV. In Section 7 we discuss a
number of specification tests and sensitivity analyses based on tests for (a) discontinuities in the average values
for covariates, (b) discontinuities in the conditional density of the forcing variable, as suggested by McCrary,
and (c) discontinuities in the average outcome at other values of the forcing variable.

2. Sharp and FRD designs

2.1. Basics

Our discussion will frame the RD design in the context of the modern literature on causal effects and
treatment effects, using the Rubin Causal Model (RCM) set up with potential outcomes (Rubin, 1974;
Holland, 1986; Imbens and Rubin, 2007), rather than the regression framework that was originally used in this
literature. For a general discussion of the RCM and its use in the economic literature, see the survey by Imbens
and Wooldridge (2007).
In the basic setting for the RCM (and for the RD design), researchers are interested in the causal effect of a
binary intervention or treatment. Units, which may be individuals, firms, countries, or other entities, are either
exposed or not exposed to a treatment. The effect of the treatment is potentially heterogenous across units. Let
Y i ð0Þ and Y i ð1Þ denote the pair of potential outcomes for unit i: Y i ð0Þ is the outcome without exposure to the
treatment and Y i ð1Þ is the outcome given exposure to the treatment. Interest is in some comparison of Y i ð0Þ
and Y i ð1Þ. Typically, including in this discussion, we focus on differences Y i ð1Þ  Y i ð0Þ. The fundamental
problem of causal inference is that we never observe the pair Y i ð0Þ and Y i ð1Þ together. We therefore typically
focus on average effects of the treatment, that is, averages of Y i ð1Þ  Y i ð0Þ over (sub)populations, rather than
on unit-level effects. For unit i we observe the outcome corresponding to the treatment received. Let W i 2
f0; 1g denote the treatment received, with W i ¼ 0 if unit i was not exposed to the treatment, and W i ¼ 1
otherwise. The outcome observed can then be written as
(
Y i ð0Þ if W i ¼ 0;
Y i ¼ ð1  W i Þ  Y i ð0Þ þ W i  Y i ð1Þ ¼
Y i ð1Þ if W i ¼ 1:

In addition to the assignment W i and the outcome Y i , we may observe a vector of covariates or pretreatment
variables denoted by ðX i ; Z i Þ, where X i is a scalar and Z i is an M-vector. A key characteristic of X i and Z i is
that they are known not to have been affected by the treatment. Both X i and Z i are covariates, with a special
role played by X i in the RD design. For each unit we observe the quadruple ðY i ; W i ; X i ; Zi Þ. We assume that
we observe this quadruple for a random sample from some well-defined population.
The basic idea behind the RD design is that assignment to the treatment is determined, either completely or
partly, by the value of a predictor (the covariate X i ) being on either side of a fixed threshold. This predictor
may itself be associated with the potential outcomes, but this association is assumed to be smooth, and so any
discontinuity of the conditional distribution (or of a feature of this conditional distribution such as the
conditional expectation) of the outcome as a function of this covariate at the cutoff value is interpreted as
evidence of a causal effect of the treatment.
The design often arises from administrative decisions, where the incentives for units to participate in a
program are partly limited for reasons of resource constraints, and clear transparent rules rather than
discretion by administrators are used for the allocation of these incentives. Examples of such settings abound.
For example, Hahn et al. (1999) study the effect of an anti-discrimination law that only applies to firms with at
least 15 employees. In another example, Matsudaira (2007) studies the effect of a remedial summer school
program that is mandatory for students who score less than some cutoff level on a test (see also Jacob and
Lefgren, 2004). Access to public goods such as libraries or museums is often eased by lower prices for
individuals depending on an age cutoff value (senior citizen discounts and discounts for children under some
age limit). Similarly, eligibility for medical services through medicare is restricted by age (Card et al., 2004).
 ARTICLE IN PRESS
G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635 617

1
0.8
0.6
0.4
0.2
0
0 1 2 3 4 5 6 7 8 9 10

Fig. 1. Assignment probabilities (SRD).

5
4
3
2
1
0
0 1 2 3 4 5 6 7 8 9 10

Fig. 2. Potential and observed outcome regression functions.

2.2. The sharp regression discontinuity design

It is useful to distinguish between two general settings, the sharp and the fuzzy regression discontinuity
(SRD and FRD from hereon) designs (e.g., Trochim, 1984, 2001; HTV). In the SRD design the assignment W i
is a deterministic function of one of the covariates, the forcing (or treatment-determining) variable X 1:
W i ¼ 1fX i Xcg.
All units with a covariate value of at least c are assigned to the treatment group (and participation is
mandatory for these individuals), and all units with a covariate value less than c are assigned to the control
group (members of this group are not eligible for the treatment). In the SRD design we look at the
discontinuity in the conditional expectation of the outcome given the covariate to uncover an average causal
effect of the treatment:
lim E½Y i jX i ¼ x  lim E½Y i jX i ¼ x,
x#c x"c

which is interpreted as the average causal effect of the treatment at the discontinuity point
tSRD ¼ E½Y i ð1Þ  Y i ð0ÞjX i ¼ c. (2.1)
Figs. 1 and 2 illustrate the identification strategy in the SRD setup. Based on artificial population values, we
present in Fig. 1 the conditional probability of receiving the treatment, PrðW ¼ 1jX ¼ xÞ against the covariate
x. At x ¼ 6 the probability jumps from 0 to 1. In Fig. 2, three conditional expectations are plotted. The two
continuous lines (partly dashed, partly solid) in the figure are the conditional expectations of the two potential
outcomes given the covariate, mw ðxÞ ¼ E½Y ðwÞjX ¼ x, for w ¼ 0; 1. These two conditional expectations are
continuous functions of the covariate. Note that we can only estimate m0 ðxÞ for xoc and m1 ðxÞ for xXc.
1
Here we take X i to be a scalar. More generally, the assignment can be a function of a vector of covariates. Formally, we can write this
as the treatment indicator being an indicator for the vector X i being an element of a subset of the covariate space, or

W i ¼ 1fX i 2 X1 g,
where X1  X, and X is the covariate space.
 ARTICLE IN PRESS
618 G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635

In addition we plot the conditional expectation of the observed outcome

E½Y jX ¼ x ¼ E½Y jW ¼ 0; X ¼ x  PrðW ¼ 0jX ¼ xÞ
þ E½Y jW ¼ 1; X ¼ x  PrðW ¼ 1jX ¼ xÞ
in Fig. 2, indicated by a solid line. Although the two conditional expectations of the potential outcomes mw ðxÞ
are continuous, the conditional expectation of the observed outcome jumps at x ¼ c ¼ 6.
Now let us discuss the interpretation of limx#c E½Y i jX i ¼ x  limx"c E½Y i jX i ¼ x as an average causal effect
in more detail. In the SRD design, the widely used unconfoundedness assumption (e.g., Rosenbaum and
Rubin, 1983; Imbens, 2004) underlying most matching-type estimators still holds:
Y i ð0Þ; Y i ð1Þ@W i jX i .
This assumption holds in a trivial manner, because conditional on the covariates there is no variation in the
treatment. However, this assumption cannot be exploited directly. The problem is that the second assumption
that is typically used for matching-type approaches, the overlap assumption which requires that for all values
of the covariates there are both treated and control units, or
0o PrðW i ¼ 1jX i ¼ xÞo1,
is fundamentally violated. In fact, for all values of x the probability of assignment is either 0 or 1, rather than
always between 0 and 1 as required by the overlap assumption. As a result, there are no values of x with
overlap.
This implies there is an unavoidable need for extrapolation. However, in large samples the amount of
extrapolation required to make inferences is arbitrarily small, as we only need to infer the conditional
expectation of Y ðwÞ given the covariates e away from where it can be estimated. To avoid non-trivial
extrapolation we focus on the average treatment effect at X ¼ c:
tSRD ¼ E½Y ð1Þ  Y ð0ÞjX ¼ c ¼ E½Y ð1ÞjX ¼ c  E½Y ð0ÞjX ¼ c. (2.2)
By design, there are no units with X i ¼ c for whom we observe Y i ð0Þ. We therefore will exploit the fact that we
observe units with covariate values arbitrarily close to c.2 In order to justify this averaging we make a
smoothness assumption. Typically this assumption is formulated in terms of conditional expectations.

Assumption 2.1. (Continuity of Conditional Regression Functions)

E½Y ð0ÞjX ¼ x and E½Y ð1ÞjX ¼ x,
are continuous in x.

More generally, one might want to assume that the conditional distribution function is smooth in the
covariate. Let F Y ðwÞjX ðyjxÞ ¼ PrðY ðwÞpyjX ¼ xÞ denote the conditional distribution function of Y ðwÞ given
X . Then the general version of the assumption is:

Assumption 2.2. (Continuity of Conditional Distribution Functions)

F Y ð0ÞjX ðyjxÞ and F Y ð1ÞjX ðyjxÞ,
are continuous in x for all y.

Both these assumptions are stronger than required, as we will only use continuity at x ¼ c, but it is rare that
it is reasonable to assume continuity for one value of the covariate, but not at other values of the covariate.
We therefore make the stronger assumption.
Under either assumption,
E½Y ð0ÞjX ¼ c ¼ lim E½Y ð0ÞjX ¼ x ¼ lim E½Y ð0ÞjW ¼ 0; X ¼ x ¼ lim E½Y jX ¼ x
x"c x"c x"c

2
Although in principle the first term in the difference in (2.2) would be straightforward to estimate if we actually observed individuals
with X i ¼ x, with continuous covariates we also need to estimate this term by averaging over units with covariate values close to c.
 ARTICLE IN PRESS
G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635 619

and similarly,
E½Y ð1ÞjX ¼ c ¼ lim E½Y jX ¼ x.
x#c

Thus, the average treatment effect at c, tSRD , satisfies

tSRD ¼ lim E½Y jX ¼ x  lim E½Y jX ¼ x.
x#c x"c

The estimand is the difference of two regression functions at a point. Hence, if we try to estimate this
object without parametric assumptions on the two regression functions, we do not obtain root-N consistent
estimators. Instead we get consistent estimators that converge to their limits at slower, nonparametric
rates.
As an example of an SRD design, consider the study of the effect of party affiliation of a congressman on
congressional voting outcomes by Lee (2007). See also Lee et al. (2004). The key idea is that electoral districts
where the share of the vote for a Democrat in a particular election was just under 50% are on average similar
in many relevant respects to districts where the share of the Democratic vote was just over 50%, but the small
difference in votes leads to an immediate and big difference in the party affiliation of the elected representative.
In this case, the party affiliation always jumps at 50%, making this an SRD design. Lee looks at the
incumbency effect. He is interested in the probability of Democrats winning the subsequent election,
comparing districts where the Democrats won the previous election with just over 50% of the popular vote
with districts where the Democrats lost the previous election with just under 50% of the vote.

2.3. The FRD design

In the FRD design, the probability of receiving the treatment needs not change from 0 to 1 at the threshold.
Instead, the design allows for a smaller jump in the probability of assignment to the treatment at the threshold:
lim PrðW i ¼ 1jX i ¼ xÞa lim PrðW i ¼ 1jX i ¼ xÞ,
x#c x"c

without requiring the jump to equal 1. Such a situation can arise if incentives to participate in a program
change discontinuously at a threshold, without these incentives being powerful enough to move all units from
nonparticipation to participation. In this design we interpret the ratio of the jump in the regression of the
outcome on the covariate to the jump in the regression of the treatment indicator on the covariate as an
average causal effect of the treatment. Formally, the estimand is
limx#c E½Y jX ¼ x  limx"c E½Y jX ¼ x
tFRD ¼ .
limx#c E½W jX ¼ x  limx"c E½W jX ¼ x
Let us first consider the interpretation of this ratio. HTV, in arguably the most important theoretical paper
in the recent RD literature, exploit the instrumental variables connection to interpret the FRD design when
the effect of the treatment varies by unit, as in Imbens and Angrist (1994).3 Let W i ðxÞ be potential treatment
status given cutoff point x, for x in some small neighborhood around c. W i ðxÞ is equal to 1 if unit i would take
or receive the treatment if the cutoff point was equal to x. This requires that the cutoff point is at least in
principle manipulable. For example, if X is age, one could imagine changing the age that makes an individual
eligible for the treatment from c to c þ . Then it is useful to assume monotonicity (see HTV).
Assumption 2.3. W i ðxÞ is non-increasing in x at x ¼ c.
Next, define compliance status. This concept is similar to the one used in instrumental variables settings
(e.g., Angrist et al., 1996). A complier is a unit such that
lim W i ðxÞ ¼ 0 and lim W i ðxÞ ¼ 1.
x#X i x"X i

3
The close connection between FRD and instrumental variables models led researchers in a number of cases to interpret RD designs as
instrumental variables settings. See, for example, Angrist and Krueger (1991) and Imbens and Van Der Klaauw (1995). The main
advantage of thinking of these designs as RD designs is that it suggests the specification analyses from Section 7.
 ARTICLE IN PRESS
620 G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635

Compliers are units that would get the treatment if the cutoff were at X i or below, but that would not get the
treatment if the cutoff were higher than X i . To be specific, consider an example where individuals with a test
score less than c are encouraged for a remedial teaching program (Matsudaira, 2007). Interest is in the effect of
the program on subsequent test scores. Compliers are individuals who would participate if encouraged (if the
test score is below the cutoff for encouragement), but not if not encouraged (if test score is above the cutoff for
encouragement). Nevertakers are units with
lim W i ðxÞ ¼ 0 and lim W i ðxÞ ¼ 0,
x#X i x"X i

and alwaystakers are units with

lim W i ðxÞ ¼ 1 and lim W i ðxÞ ¼ 1.
x#X i x"X i

Then,
limx#c E½Y jX ¼ x  limx"c E½Y jX ¼ x
tFRD ¼
limx#c E½W jX ¼ x  limx"c E½W jX ¼ x
¼ E½Y i ð1Þ  Y i ð0Þjunit i is a complier and X i ¼ c.

The estimand is an average effect of the treatment, but only averaged for units with X i ¼ c (by RD), and only
for compliers (people who are affected by the threshold).
In Fig. 3 we plot the conditional probability of receiving the treatment for an FRD design. As in the SRD
design, this probability still jumps at x ¼ 6, but now by an amount less than 1. Fig. 4 presents the expectation
of the potential outcomes given the covariate and the treatment, E½Y ðwÞjW ¼ w; X ¼ x, represented by the
dashed lines, as well as the conditional expectation of the observed outcome given the covariate (solid line):

E½Y jX ¼ x ¼ E½Y ð0ÞjW ¼ 0; X ¼ x  PrðW ¼ 0jX ¼ xÞ

þ E½Y ð1ÞjW ¼ 1; X ¼ x  PrðW ¼ 1jX ¼ xÞ.

Note that it is no longer necessarily the case here that E½Y ðwÞjW ¼ w; X ¼ x ¼ E½Y ðwÞjX ¼ x. Under some
assumptions (unconfoundedness) this will be true, but this is not necessary for inference regarding causal
effects in the FRD setting.

1
0.8
0.6
0.4
0.2
0
0 1 2 3 4 5 6 7 8 9 10

Fig. 3. Assignment probabilities (FRD).

5
4
3
2
1
0
0 1 2 3 4 5 6 7 8 9 10

Fig. 4. Potential and observed outcome regression (FRD).

 ARTICLE IN PRESS
G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635 621

As an example of an FRD design, consider the study of the effect of financial aid on college attendance by
Van Der Klaauw (2002). Van Der Klaauw looks at the effect of financial aid on acceptance on college
admissions. Here X i is a numerical score assigned to college applicants based on the objective part of the
application information (SAT scores, grades) used to streamline the process of assigning financial aid offers.
During the initial stages of the admission process, the applicants are divided into L groups based on
discretized values of these scores. Let
8
> 1 if 0pX i oc1 ;
>
>
>
< 2 if c1 pX i oc2 ;
Gi ¼ .
>
> ..
>
>
:
L if cL1 pX i ;
denote the financial aid group. For simplicity, let us focus on the case with L ¼ 2, and a single cutoff point c.
Having a score of just over c will put an applicant in a higher category and increase the chances of financial aid
discontinuously compared to having a score of just below c. The outcome of interest in the Van Der Klaauw
study is college attendance. In this case, the simple association between attendance and the financial aid offer
is ambiguous. On the one hand, an aid offer makes the college more attractive to the potential student. This is
the causal effect of interest. On the other hand, a student who gets a generous financial aid offer is likely to
have better outside opportunities in the form of financial aid offers from other colleges. College aid is
emphatically not a deterministic function of the financial aid categories, making this an FRD design. Other
components of the application that are not incorporated in the numerical score (such as the essay and
recommendation letters) undoubtedly play an important role. Nevertheless, there is a clear discontinuity in the
probability of receiving an offer of a larger financial aid package.

2.4. The FRD design and unconfoundedness

In the FRD setting, it is useful to contrast the RD approach with estimation of average causal effects under
unconfoundedness. The unconfoundedness assumption (e.g., Rosenbaum and Rubin, 1983; Imbens, 2004)
requires that
Y ð0Þ; Y ð1Þ@W jX .
If this assumption holds, then we can estimate the average effect of the treatment at X ¼ c as
E½Y ð1Þ  Y ð0ÞjX ¼ x ¼ E½Y jW ¼ 1; X ¼ c  E½Y jW ¼ 0; X ¼ c.
This approach does not exploit the jump in the probability of assignment at the discontinuity point. Instead it
assumes that differences between treated and control units with X i ¼ c are interpretable as average causal effects.
In contrast, the assumptions underlying an FRD analysis implies that comparing treated and control units
with X i ¼ c is likely to be the wrong approach. Treated units with X i ¼ c include compliers and alwaystakers,
and control units at X i ¼ c consist of nevertakers. Comparing these different types of units has no causal
interpretation under the FRD assumptions. Although, in principle, one cannot test the unconfoundedness
assumption, one aspect of the problem makes this assumption fairly implausible. Unconfoundedness is
fundamentally based on units being comparable if their covariates are similar. This is not an attractive
assumption in the current setting where the probability of receiving the treatment is discontinuous in the
covariate. Thus, units with similar values of the forcing variable (but on different sides of the threshold) must
be different in some important way related to the receipt of treatment. Unless there is a substantive argument
that this difference is immaterial for the comparison of the outcomes of interest, an analysis based on
unconfoundedness is not attractive.

2.5. External validity

One important aspect of both the SRD and FRD designs is that they, at best, provide estimates of the
average effect for a subpopulation, namely the subpopulation with covariate value equal to X i ¼ c. The FRD
 ARTICLE IN PRESS
622 G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635

design restricts the relevant subpopulation even further to that of compliers at this value of the covariate.
Without strong assumptions justifying extrapolation to other subpopulations (e.g., homogeneity of the
treatment effect), the designs never allow the researcher to estimate the overall average effect of the treatment.
In that sense the design has fundamentally only a limited degree of external validity, although the specific
average effect that is identified may well be of special interest, for example in cases where the policy question
concerns changing the location of the threshold. The advantage of RD designs compared to other non-
experimental analyses that may have more external validity, such as those based on unconfoundedness is that
RD designs may have a relatively high degree of internal validity (in settings where they are applicable).

3. Graphical analyses

3.1. Introduction

Graphical analyses should be an integral part of any RD analysis. The nature of RD designs suggests that
the effect of the treatment of interest can be measured by the value of the discontinuity in the expected value of
the outcome at a particular point. Inspecting the estimated version of this conditional expectation is a simple
yet powerful way to visualize the identification strategy. Moreover, to assess the credibility of the RD strategy,
it is useful to inspect two additional graphs for covariates and the density of the forcing variable. The
estimators we discuss later use more sophisticated methods for smoothing but these basic plots will convey
much of the intuition. For strikingly clear examples of such plots, see Lee et al. (2004), Lalive (2007), and Lee
(2007). Note that, in practice, the visual clarity of the plots is often improved by adding smoothed regression
lines based on polynomial regressions (or other flexible methods) estimated separately on the two sides of the
cutoff point.

3.2. Outcomes by forcing variable

The first plot is a histogram-type estimate of the average value of the outcome for different values of the
forcing variable, the estimated counterpart to the solid line in Figs. 2 and 4. For some binwidth h, and for
some number of bins K 0 and K 1 to the left and right of the cutoff value, respectively, construct bins ðbk ; bkþ1 ,
for k ¼ 1; . . . ; K ¼ K 0 þ K 1 , where
bk ¼ c  ðK 0  k þ 1Þ  h.
Then calculate the number of observations in each bin
X
N
Nk ¼ 1fbk oX i pbkþ1 g
i¼1

and the average outcome in the bin

1 X N
Yk ¼  Y i  1fbk oX i pbkþ1 g.
N k i¼1
The first plot of interest is that of the Y k , for k ¼ 1; . . . ; K against the mid point of the bins,
b~k ¼ ðbk þ bkþ1 Þ=2. The question is whether around the threshold c there is any evidence of a jump in the
conditional mean of the outcome. The formal statistical analyses discussed below are essentially just
sophisticated versions of this, and if the basic plot does not show any evidence of a discontinuity, there is
relatively little chance that the more sophisticated analyses will lead to robust and credible estimates with
statistically and substantially significant magnitudes. In addition to inspecting whether there is a jump at this
value of the covariate, one should inspect the graph to see whether there are any other jumps in the conditional
expectation of Y given X that are comparable to, or larger than, the discontinuity at the cutoff value. If so,
and if one cannot explain such jumps on substantive grounds, it would call into question the interpretation of
the jump at the threshold as the causal effect of the treatment. In order to optimize the visual clarity it is
important to calculate averages that are not smoothed over the cutoff point.
 ARTICLE IN PRESS
G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635 623

3.3. Covariates by forcing variable

The second set of plots compares average values of other covariates in the K bins. Specifically, let Z i be the
M-vector of additional covariates, with mth element Z im . Then calculate

1 X N
Z km ¼  Z im  1fbk oX i pbkþ1 g.
N k i¼1

The second plot of interest is that of the Z km , for k ¼ 1; . . . ; K against the mid point of the bins, b~k , for all
m ¼ 1; . . . ; M. In the case of FRD designs, it is also particularly useful to plot the mean values of the treatment
variable W i to make sure there is indeed a jump in the probability of treatment at the cutoff point (as in Fig. 3.
Plotting other covariates is also useful for detecting possible specification problems (see Section 7.1) in the
case of either SRD or FRD designs.

3.4. The Density of the forcing variable

In the third graph, one should plot the number of observations in each bin, N k , against the mid points b~k .
This plot can be used to inspect whether there is a discontinuity in the distribution of the forcing variable X at
the threshold. Such discontinuity would raise the question of whether the value of this covariate was
manipulated by the individual agent, invalidating the design. For example, suppose that the forcing variable is
a test score. If individuals know the threshold and have the option of retaking the test, individuals with test
scores just below the threshold may do so, and invalidate the design. Such a situation would lead to a
discontinuity of the conditional density of the test score at the threshold, and thus be detectable in the kind of
plots described here. See Section 7.2 for more discussion of tests based on this idea.

4. Estimation: local linear regression

4.1. Nonparametric regression at the boundary

The practical estimation of the treatment effect t in both the SRD and FRD designs is largely a standard
nonparametric regression problem (e.g., Pagan and Ullah, 1999; Härdle, 1990; Li and Racine, 2007).
However, there are two unusual features. In this case we are interested in the regression function at a single
point, and in addition that single point is a boundary point. As a result, standard nonparametric kernel
regression does not work very well. At boundary points, such estimators have a slower rate of convergence
than they do at interior points. Here we discuss a more attractive implementation suggested by HTV, among
others. First define the conditional means
ml ðxÞ ¼ lim E½Y ð0ÞjX ¼ z and mr ðxÞ ¼ lim E½Y ð1ÞjX ¼ z.
z"x z#x

The estimand in the SRD design is, in terms of these regression functions,
tSRD ¼ mr ðcÞ  ml ðcÞ.
A natural approach is to use standard R nonparametric regression methods for estimation of ml ðxÞ and mr ðxÞ.
Suppose we use a kernel KðuÞ, with KðuÞ du ¼ 1. Then the regression functions at x can be estimated as
P P
i:X i oc Y i  KððX i  xÞ=hÞ i:X Xc Y i  KððX i  xÞ=hÞ
m^ l ðxÞ ¼ P and m^ r ðxÞ ¼ P i ,
i:X i oc KððX i  xÞ=hÞ i:X i Xc KððX i  xÞ=hÞ

where h is the bandwidth.

The estimator for the object of interest is then
P P
i:X i Xc Y i  KððX i  xÞ=hÞ i:X oc Y i  KððX i  xÞ=hÞ
t^ SRD ¼ m^ r ðxÞ  m^ l ðxÞ ¼ P  Pi .
i:X i Xc KððX i  xÞ=hÞ i:X i oc KððX i  xÞ=hÞ
 ARTICLE IN PRESS
624 G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635

In order to see the nature of this estimator for the SRD case, it is useful to focus on a special case. Suppose we
use a rectangular kernel, e.g., KðuÞ ¼ 12 for 1ouo1, and 0 elsewhere. Then the estimator can be written as
PN PN
i¼1 Y i  1fcpX i pc þ hg i¼1 Y i  1fc  hpX i ocg
t^ SRD ¼ PN  P N
i¼1 1fcpX i pc þ hg i¼1 1fc  hpX i ocg
¼ Y hr  Y hl ,
the difference between the average outcomes for observations within a distance h of the cutoff point on the
right and left of the cutoff, respectively. N hr and N hl denote the number of observations with X i 2 ½c; c þ h
and X i 2 ½c  h; cÞ, respectively. This estimator can be interpreted as first discarding all observations with a
value of X i more than h away from the discontinuity point c, and then simply differencing the average
outcomes by treatment status in the remaining sample.
This simple nonparametric estimator is in general not very attractive, as pointed out by HTV and Porter
(2003). Let us look at the approximate bias of this estimator through the probability limit of the estimator for
fixed bandwidth. The probability limit of m^ r ðcÞ, using the rectangular kernel, is
R cþh
mðxÞf ðxÞ dx q h
plim½m^ r ðcÞ ¼ c R cþh ¼ mr ðcÞ þ lim mðxÞ  þ Oðh2 Þ.
f ðxÞ dx x#c qx 2
c

Combined with the corresponding calculation for the control group, we obtain the bias
 
h q q
plim½m^ r ðcÞ  m^ l ðcÞ  mr ðcÞ  ml ðcÞ ¼  lim mðxÞ þ lim mðxÞ þ Oðh2 Þ.
2 x#c qx x"c qx

Hence the bias is linear in the bandwidth h, whereas when we nonparametrically estimate a regression function
in the interior of the support we typically get a bias of order h2 .
Note that we typically do expect the regression function to have a non-zero derivative, even in cases where the
treatment has no effect. In many applications the eligibility criterion is based on a covariate that does have some
correlation with the outcome, so that, for example, those with poorest prospects in the absence of the program
are in the eligible group. Hence it is likely that the bias for the simple kernel estimator is relatively high.
One practical solution to the high order of the bias is to use a local linear regression (e.g., Fan and Gijbels,
1996). An alternative is to use series regression or sieve methods. Such methods could be implemented in the
current setting by adding higher-order terms to the regression function. For example, Lee et al. (2004) include
fourth-order polynomials in the covariate to the regression function. The formal properties of such methods
are equally attractive to those of kernel type methods. The main concern is that they are more sensitive to
outcome values for observations far away from the cutoff point. Kernel methods using kernels with compact
support rule out any sensitivity to such observations, and given the nature of RD designs this can be an
attractive feature. Certainly, it would be a concern if results depended in an important way on using
observations far away from the cutoff value. In addition, global methods put effort into estimating the
regression functions in areas (far away from the discontinuity point) that are of no interest in the current
setting.

4.2. Local linear regression

Here we discuss local linear regression. See for a general discussion Fan and Gijbels (1996). Instead of
locally fitting a constant function, we can fit linear regression functions to the observations within a distance h
on either side of the discontinuity point
X
min ðY i  al  bl  ðX i  cÞÞ2 ,
al :bl
i:choX i oc

and
X
min ðY i  ar  br  ðX i  cÞÞ2 .
ar :br
i:cpX i ocþh
 ARTICLE IN PRESS
G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635 625

The value of ml ðcÞ is then estimated as

md ^ l þ b^ l  ðc  cÞ ¼ a^ l ,
l ðcÞ ¼ a

and the value of mr ðcÞ is then estimated as

md ^ r þ b^ r  ðc  cÞ ¼ a^ r .
r ðcÞ ¼ a

Given these estimates, the average treatment effect is estimated as

t^ SRD ¼ a^ r  a^ l .
Alternatively one can estimate the average effect directly in a single regression, by solving
X
N
min 1fc  hpX i pc þ hg  ðY i  a  b  ðX i  cÞ  t  W i  g  ðX i  cÞ  W i Þ2 ,
a;b;t;g
i¼1

which will numerically yield the same estimate of tSRD .

An alternative is to impose the restriction that the slope coefficients are the same on both sides of the
discontinuity point, or limx#c ðq=qxÞmðxÞ ¼ limx"c ðq=qxÞmðxÞ. This can be imposed by requiring that bl ¼ br .
Although it may be reasonable to expect the slope coefficients for the covariate to be similar on both sides of
the discontinuity point, this procedure also has some disadvantages. Specifically, by imposing this restriction
one allows for observations on Y ð1Þ from the right of the discontinuity point to affect estimates of E½Y ð0ÞjX ¼
c and, similarly, for observations on Y ð0Þ from the left of discontinuity point to affect estimates of
E½Y ð1ÞjX ¼ c. In practice, one might wish to have the estimates of E½Y ð0ÞjX ¼ c based solely on observations
on Y ð0Þ, and not depend on observations on Y ð1Þ, and vice versa.
We can make the nonparametric regression more sophisticated by using weights that decrease smoothly as
the distance to the cutoff point increases, instead of the 0/1 weights based on the rectangular kernel. However,
even in this simple case the asymptotic bias can be shown to be of order h2 , and the more sophisticated kernels
rarely make much difference. Furthermore, if using different weights from a more sophisticated kernel does
make a difference, it likely suggests that the results are highly sensitive to the choice of bandwidth. So the only
case where more sophisticated kernels may make a difference is when the estimates are not very credible
anyway because of too much sensitivity to the choice of bandwidth. From a practical point of view, one may
just focus on the simple rectangular kernel, but verify the robustness of the results to different choices of
bandwidth.
For inference we can use standard least squares methods. Under appropriate conditions on the rate at which
the bandwidth goes to 0 as the sample size increases, the resulting estimates will be asymptotically normally
distributed, and the (robust) standard errors from least squares theory will be justified. Using the results from
HTV, the optimal bandwidth is h / N 1=5 . Under this sequence of bandwidths the asymptotic distribution of
the estimator t^ will have a non-zero bias. If one does some undersmoothing, by requiring that h / N d with
1=5odo2=5, then the asymptotic bias disappears and standard least square variance estimators will lead to
valid confidence intervals. See Section 6 for more details.

4.3. Covariates

Often there are additional covariates available in addition to the forcing covariate that is the basis of the
assignment mechanism. These covariates can be used to eliminate small sample biases present in the basic
specification, and improve the precision. In addition, they can be useful for evaluating the plausibility of the
identification strategy, as discussed in Section 7.1. Let the additional vector of covariates be denoted by Z i .
We make three observations on the role of these additional covariates.
The first and most important point is that the presence of these covariates rarely changes the identification
strategy. Typically, the conditional distribution of the covariates Z given X is continuous at x ¼ c. In fact, as
we discuss in Section 7, one may wish to test for discontinuities at that value of x in order to assess the
plausibility of the identification strategy. If such discontinuities in other covariates are found, the justification
of the identification strategy may be questionable. If the conditional distribution of Z given X is continuous at
 ARTICLE IN PRESS
626 G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635

x ¼ c, then including Z in the regression

X
N
min 1fc  hpX i pc þ hg  ðY i  a  b  ðX i  cÞ
a;b;t;d
i¼1

 t  W i  g  ðX i  cÞ  W i  d0 Z i Þ2 ,
will have little effect on the expected value of the estimator for t, since conditional on X being close to c, the
additional covariates Z are independent of W .
The second point is that even though the presence of Z in the regression does not affect any bias when X is
very close to c, in practice we often include observations with values of X not too close to c. In that case,
including additional covariates may eliminate some bias that is the result of the inclusion of these additional
observations.
Third, the presence of the covariates can improve precision if Z is correlated with the potential outcomes.
This is the standard argument, which also supports the inclusion of covariates in analyses of randomized
experiments. In practice the variance reduction will be relatively small unless the contribution to the R2 from
the additional regressors is substantial.

4.4. Estimation for the FRD design

In the FRD design, we need to estimate the ratio of two differences. The estimation issues we discussed
earlier in the case of the SRD arise now for both differences. In particular, there are substantial biases if
we do simple kernel regressions. Instead, it is again likely to be better to use local linear regression. We use a
uniform kernel, with the same bandwidth for estimation of the discontinuity in the outcome and treatment
regressions.
First, consider local linear regression for the outcome, on both sides of the discontinuity point. Let
X
ð^ayl ; b^ yl Þ ¼ arg min ðY i  ayl  byl  ðX i  cÞÞ2 , (4.3)
ayl ;byl
i:chpX i oc

X
ð^ayr ; b^ yr Þ ¼ arg min ðY i  ayr  byr  ðX i  cÞÞ2 . (4.4)
ayr ;byr
i:cpX i pcþh

The magnitude of the discontinuity in the outcome regression is then estimated as

t^ y ¼ a^ yr  a^ yl .
Second, consider the two local linear regression for the treatment indicator:
X
ð^awl ; b^ wl Þ ¼ arg min ðW i  awl  bwl  ðX i  cÞÞ2 , (4.5)
awl ;bwl
i:chpX i oc

X
ð^awr ; b^ wr Þ ¼ arg min ðY i  awr  bwr  ðX i  cÞÞ2 . (4.6)
awr ;bwr
i:cpX i pcþh

The magnitude of the discontinuity in the treatment regression is then estimated as

t^ w ¼ a^ wr  a^ wl .
Finally, we estimate the effect of interest as the ratio of the two discontinuities:
t^ y a^ yr  a^ yl
t^ FRD ¼ ¼ . (4.7)
t^ w a^ wr  a^ wl
Because of the specific implementation we use here, with a uniform kernel, and the same bandwidth for
estimation of the denominator and the numerator, we can characterize the estimator for t as a two-stage least
squares (TSLS) estimator. HTV were the first to note this equality, in the setting with standard kernel
 ARTICLE IN PRESS
G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635 627

regression and no additional covariates. It is a simple extension to show that the equality still holds when we
use local linear regression and include additional regressors. Define
0 1
0 1 ayl
1
Bb C
B 1fX ocg  ðX  cÞ C B yl C
Vi ¼ @ i i A and d ¼ B C. (4.8)
@ byr A
1fX i Xcg  ðX i  cÞ

Then we can write

Y i ¼ d0 V i þ t  W i þ e i . (4.9)
Estimating t based on the regression function (4.9) by TSLS methods, with the indicator 1fX i Xcg as the
excluded instrument and V i as the set of exogenous variables is numerically identical to t^ FRD as given in (4.7).

5. Bandwidth selection

An important issue in practice is the selection of the smoothing parameter, the binwidth h. In general there
are two approaches to choose bandwidths. A first approach consists of characterizing the optimal bandwidth
in terms of the unknown joint distribution of all variables. The relevant components of this distribution can
then be estimated, and plugged into the optimal bandwidth function. The second approach, on which we focus
here, is based on a cross-validation procedure. The specific methods discussed here are similar to those
developed by Ludwig and Miller (2005, 2007). In particular, their proposals, like ours, are aimed specifically at
estimating the regression function at the boundary. Initially we focus on the SRD case, and in Section 5.2 we
extend the recommendations to the FRD setting.
To set up the bandwidth choice problem we generalize the notation slightly. In the SRD setting we are
interested in
tSRD ¼ lim mðxÞ  lim mðxÞ.
x#c x"c

We estimate the two terms as

limd
mðxÞ ¼ a^ r ðcÞ
x#c

and

limd
mðxÞ ¼ a^ l ðcÞ,
x"c

where a^ l ðxÞ and b^ l ðxÞ solve

X
ð^al ðxÞ; b^ l ðxÞÞ ¼ arg min ðY j  a  b  ðX j  xÞÞ2 . (5.10)
a;b
jjxhoX j ox

and a^ r ðxÞ and b^ r ðxÞ solve

X
ð^ar ðxÞ; b^ r ðxÞÞ ¼ arg min ðY j  a  b  ðX j  xÞÞ2 . (5.11)
a;b
jjxoX j oxþh

Let us focus first on estimating limx#c mðxÞ. For estimation of this limit we are interested in the bandwidth h
that minimizes
" 2 #
Qr ðx; hÞ ¼ E lim mðzÞ  a^ r ðxÞ ,
z#x
 ARTICLE IN PRESS
628 G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635

at x ¼ c. In principle this could be different from the bandwidth that minimizes the corresponding criterion on
the left-hand side,
"  # 2
Ql ðx; hÞ ¼ E lim mðzÞ  a^ l ðxÞ ,
z"x

at x ¼ c. However, we will focus on a single bandwidth for both sides of the threshold, and therefore focus on
minimizing

1
Qðc; hÞ ¼  ðQl ðc; hÞ þ Qr ðc; hÞÞ
2 " 2 # " 2 #!
1
¼  E lim mðxÞ  a^ l ðcÞ þ E lim mðxÞ  a^ r ðcÞ .
2 x"c x#c

We now discuss two methods for choosing the bandwidth.

5.1. Bandwidth selection for the SRD design

For a given binwidth h, let the estimated regression function at x be

(
a^ l ðxÞ if xoc;
^
mðxÞ ¼
a^ r ðxÞ if xXc;

where a^ l ðxÞ, b^ l ðxÞ, a^ r ðxÞ, and b^ r ðxÞ solve (5.10) and (5.11). Note that in order to mimic the fact that we are
interested in estimation at the boundary, we only use the observations on one side of x in order to estimate the
regression function at x, rather than the observations on both sides of x, that is, observations with
x  hoX j ox þ h. In addition, the strict inequality in the definition implies that mðxÞ ^ evaluated at x ¼ X i does
not depend on Y i .
Now define the cross-validation criterion as
1X N
CVY ðhÞ ¼ ^ i ÞÞ2
ðY i  mðX (5.12)
N i¼1
with the corresponding cross-validation choice for the binwidth
hopt
CV ¼ arg min CVY ðhÞ.
h

The expected value of this cross-validation

R function is, ignoring the term that does not involve h, equal to
E½CVY ðhÞ ¼ C þ E½QðX ; hÞ ¼ C þ Qðx; hÞf X dx. Although the modification to estimate the regression
using one-sided kernels mimics more closely the estimand of interest, this is still not quite what we are
interested in. Ultimately, we are solely interested in estimating the regression
R function in the neighborhood of
a single point, the threshold c, and thus in minimizing Qðc; hÞ, rather than x Qðx; hÞf X ðxÞ dx. If there are quite
a few observations in the tails of the distribution, minimizing the criterion in (5.12) may lead to larger bins
than is optimal for estimating the regression function around x ¼ c, if c is in the center of the distribution. We
may therefore wish to minimize the cross-validation criterion after first discarding observations from the tails.
Let qX ;d;l be the d quantile of the empirical distribution of X for the subsample with X i oc, and let qX ;1d;r be
the 1  d quantile of the empirical distribution of X for the subsample with X i Xc. Then, we may wish to use
the criterion
1 X
CVdY ðhÞ ¼ ^ i ÞÞ2 .
ðY i  mðX (5.13)
N i:q pX pq
X ;d;l i X ;1d;r
 ARTICLE IN PRESS
G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635 629

The modified cross-validation choice for the bandwidth is

hd;opt d
CV ¼ arg min CVY ðhÞ. (5.14)
h

The modified cross-validation function has expectation, again ignoring terms that do not involve h,
proportional to E½QðX ; hÞjqX ;d;l oX oqX ;1d;r . Choosing a smaller value of d makes the expected value of the
criterion closer to what we are ultimately interested in, that is, Qðc; hÞ, but has the disadvantage of leading to a
noisier estimate of E½CVdY ðhÞ. In practice, one may wish to choose d ¼ 12, and discard 50% of the observations
on either side of the threshold, and afterwards assess the sensitivity of the bandwidth choice to the choice of d.
Ludwig and Miller (2005) implement this by using only data within 5% points of the threshold on either side.
Note that, in principle, we can use a different binwidth on either side of the cutoff value. However, it is likely
that the density of the forcing variable x is similar on both sides of the cutoff point. If, in addition, the
curvature is similar on both sides close to the cutoff point, then in large samples the optimal binwidth will be
similar on both sides. Hence, the benefits of having different binwidths on the two sides may not be sufficient
to balance the disadvantage of the additional noise in estimating the optimal value from a smaller sample.

5.2. Bandwidth selection for the FRD design

In the FRD design, there are four regression functions that need to be estimated: the expected outcome
given the forcing variable, both on the left and right of the cutoff point, and the expected value of the
treatment variable, again on the left and right of the cutoff point. In principle, we can use different binwidths
for each of the four nonparametric regressions.
In the section on the SRD design, we argued in favor of using identical bandwidths for the regressions on
both sides of the cutoff point. The argument is not so clear for the pairs of regression functions by outcome we
have here. In principle, we have two optimal bandwidths, one based on minimizing CVdY ðhÞ, and one based on
minimizing CVdW ðhÞ, defined correspondingly. It is likely that the conditional expectation of the treatment
variable is relatively flat compared to the conditional expectation of the outcome variable, suggesting one
should use a larger binwidth for estimating the former.4 Nevertheless, in practice it is appealing to use the
same binwidth for numerator and denominator. To avoid asymptotic biases, one may wish to use the smallest
bandwidth selected by the cross-validation criterion applied separately to the outcome and treatment
regression
 
opt d d
hCV ¼ min arg min CVY ðhÞ; arg min CVW ðhÞ ,
h h

where CVdY ðhÞ is as defined in (5.12), and CVdW ðhÞ is defined similarly. Again, a value of d ¼ 12 is likely to lead
to reasonable estimates in many settings.

6. Inference

We now discuss some asymptotic properties for the estimator for the FRD case given in (4.7) or its
alternative representation in (4.9).5 More general results are given in HTV. We continue to make some
simplifying assumptions. First, as in the previous sections, we use a uniform kernel. Second, we use the same
bandwidth for the estimator for the jump in the conditional expectation of the outcome and treatment. Third,
we undersmooth, so that the square of the bias vanishes faster than the variance, and we can ignore the bias in
the construction of confidence intervals. Fourth, we continue to use the local linear estimator.
4
In the extreme case of the SRD design where the conditional expectation of W given X is flat on both sides of the threshold, the optimal
bandwidth would be infinity. Therefore, in practice it is likely that the optimal bandwidth for estimating the jump in the conditional
expectation of the treatment would be larger than the bandwidth for estimating the conditional expectation of the outcome.
5
The results for the SRD design are a special case of those for the FRD design. In the SRD design, only the first term of the asymptotic
variance in equation (6.18) is left since V tw ¼ C ty ;tw ¼ 0, and the variance can also be estimated using the standard robust variance for
OLS instead of TSLS.
 ARTICLE IN PRESS
630 G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635

Under these assumptions we do two things. First, we give an explicit expression for the asymptotic variance.
Second, we present two estimators for the asymptotic variance. The first estimator follows explicitly the
analytic form for the asymptotic variance, and substitutes estimates for the unknown quantities. The second
estimator is the standard robust variance for the TSLS estimator, based on the sample obtained by discarding
observations when the forcing covariate is more than h away from the cutoff point. The asymptotic variance
and the corresponding estimators reported here are robust to heteroskedasticity.

6.1. The asymptotic variance

To characterize the asymptotic variance we need a couple of additional pieces of notation. Define the four
variances
s2Y l ¼ lim VarðY jX ¼ xÞ; s2Y r ¼ lim VarðY jX ¼ xÞ,
x"c x#c

s2W l ¼ lim VarðW jX ¼ xÞ; s2W r ¼ lim VarðW jX ¼ xÞ,

x"c x#c

and the two covariances

C YW l ¼ lim CovðY ; W jX ¼ xÞ; C YW r ¼ lim CovðY ; W jX ¼ xÞ.
x"c x#c

Note that, because of the binary nature of W , it follows that s2W l ¼ mW l  ð1  mW l Þ, where
mW l ¼ limx"c PrðW ¼ 1jX ¼ xÞ, and similarly for s2W r . To p
discuss
ffiffiffiffiffiffiffi the asymptotic variance of t^ , it is useful
to break it up in three pieces. The asymptotic variance of Nhð^ty  ty Þ is
4  
V ty ¼  s2Y r þ s2Y l . (6.15)
f X ðcÞ
pffiffiffiffiffiffiffi
The asymptotic variance of Nhð^tw  tw Þ is
4
V tw ¼  ðs2W r þ s2W l Þ. (6.16)
f X ðcÞ
pffiffiffiffiffiffiffi pffiffiffiffiffiffiffi
The asymptotic covariance of Nhð^ty  ty Þ and Nhð^tw  tw Þ is
4
C ty ;tw ¼  ðC YW r þ C YW l Þ. (6.17)
f X ðcÞ

Finally, the asymptotic distribution has the form

!
pffiffiffiffiffiffiffi d 1 t2y ty
Nh  ð^t  tÞ ! N 0; 2  V ty þ 4  V tw  2  3  C ty ;tw . (6.18)
tw tw tw

This asymptotic distribution is a special case of that in HTV (p. 208), using the rectangular kernel, and with
h / N d , for 1=5odo2=5 (so that the asymptotic bias can be ignored).

6.2. A plug-in estimator for the asymptotic variance

We now discuss two estimators for the asymptotic variance of t^ . First, we can estimate the asymptotic
variance of t^ by estimating each of the components, tw , ty , V tw , V ty , and C ty ;tw and substituting them into the
expression for the variance in (6.18). In order to do this we first estimate the residuals
e^i ¼ Y i  m^ y ðX i Þ ¼ Y i  1fX i ocg  a^ yl  1fX i Xcg  a^ yr ,

Z^ i ¼ W i  m^ w ðX i Þ ¼ W i  1fX i ocg  a^ wl  1fX i Xcg  a^ wr .

 ARTICLE IN PRESS
G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635 631

Then we estimate the variances and covariances consistently as

1 X 1 X
s^ 2Y l ¼ e^2i ; s^ 2Y r ¼ e^2 ,
N hl i:chpX oc N hr i:cpX pcþh i
i i

1 X 1 X
s^ 2W l ¼ Z^ 2i ; s^ 2W r ¼ Z^ 2 ,
N hl i:chpX N hr i:cpX pcþh i
i oc i

1 X 1 X
C^ YW l ¼ e^i  Z^ i ; C^ YW r ¼ e^i  Z^ i .
N hl i:chpX oc N hr i:cpX pcþh
i i

Finally, we estimate the density consistently as

N hl þ N hr
f^X ðxÞ ¼ .
2N h
Then we can plug-in the estimated components of V ty , V tW , and C tY ;tW from (6.15)–(6.17), and finally
substitute these into the variance expression in (6.18).

6.3. The TSLS variance estimator

The second estimator for the asymptotic variance of t^ exploits the interpretation of the t^ as a TSLS
estimator, given in (4.9). The variance estimator is equal to the robust variance for TSLS based on the
subsample of observations with c  hpX i pc þ h, using the indicator 1fX i Xcg as the excluded instrument, the
treatment W i as the endogenous regressor and the V i defined in (4.8) as the exogenous covariates.

7. Specification testing

There are generally two main conceptual concerns in the application of RD designs, sharp or fuzzy. A first
concern about RD designs is the possibility of other changes at the same cutoff value of the covariate. Such
changes may affect the outcome, and these effects may be attributed erroneously to the treatment of interest.
For example, at age 65 individuals become eligible for discounts at many cultural institutions. However, if one
finds that there is a discontinuity in the number of hours worked at age 65, this is unlikely to be the result of
these discounts. The more plausible explanation is that there are other institutional changes that affect
incentives to work at age 65. The effect of discounts on attendance at these cultural institutions, which may
well be present, may be difficult to detect due to the many other changes at age 65.
The second concern is that of manipulation of the forcing variable. Consider the Van Der Klaauw example
where the value of an aggregate admission score affected the likelihood of receiving financial aid. If a single
admissions officer scores the entire application packet of any one individual, and if this person is aware of the
importance of this cutoff point, they may be more or less likely to score an individual just below the cutoff
value. Alternatively, if applicants know the scoring rule, they may attempt to change particular parts of their
application in order to end up on the right side of the threshold, for example by retaking tests. If it is costly to
do so, the individuals retaking the test may be a selected sample, invalidating the basic RD design.
We also address the issue of sensitivity to the bandwidth choice, and more generally small sample concerns.
We end the section by discussing how, in the FRD setting, one can compare the RD estimates to those based
on unconfoundedness.

7.1. Tests involving covariates

One category of tests involves testing the null hypothesis of a zero average effect on pseudo outcomes
known not to be affected by the treatment. Such variables includes covariates that are, by definition, not
affected by the treatment. Such tests are familiar from settings with identification based on unconfoundedness
assumptions (e.g., Heckman and Hotz, 1989; Rosenbaum, 1987; Imbens, 2004). In the RD setting, they have
 ARTICLE IN PRESS
632 G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635

been applied by Lee et al. (2004) and others. In most cases, the reason for the discontinuity in the probability
of the treatment does not suggest a discontinuity in the average value of covariates. If we find such a
discontinuity, it typically casts doubt on the assumptions underlying the RD design. In principle, it may be
possible to make the assumptions underlying the RD design conditional on covariates, and so a discontinuity
in the conditional expectation of the covariates does not necessarily invalidate the approach. In practice,
however, it is difficult to rationalize such discontinuities with the rationale underlying the RD approach.

7.2. Tests of continuity of the density

The second test is conceptually somewhat different, and unique to the RD setting. McCrary (2007) suggests
testing the null hypothesis of continuity of the density of the covariate that underlies the assignment at the
discontinuity point, against the alternative of a jump in the density function at that point. Again, in principle,
one does not need continuity of the density of X at c, but a discontinuity is suggestive of violations of the no-
manipulation assumption. If in fact individuals partly manage to manipulate the value of X in order to be on
one side of the cutoff rather than the other, one might expect to see a discontinuity in this density at the cutoff
point. For example, if the variable underlying the assignment is age with a publicly known cutoff value c, and
if age is self-reported, one might see relatively few individuals with a reported age just below c, and relatively
many individuals with a reported age of just over c. Even if such discontinuities are not conclusive evidence of
violations of the RD assumptions, at the very least, inspecting this density would be useful to assess whether it
exhibits unusual features that may shed light on the plausibility of the design.

7.3. Testing for jumps at non-discontinuity points

A third set of tests involves estimating jumps at points where there should be no jumps. As in the treatment
effect literature (e.g., Imbens, 2004), the approach used here consists of testing for a zero effect in settings
where it is known that the effect should be 0.
Here we suggest a specific way of implementing this idea by testing for jumps at the median of the two
subsamples on either side of the cutoff value. More generally, one may wish to divide the sample up in
different ways, or do more tests. As before, let qX ;d;l and qX ;d;r be the d quantiles of the empirical distribution of
X in the subsample with X i oc and X i Xc, respectively. Now take the subsample with X i oc, and test for a
jump at the median of the forcing variable. Splitting this subsample at its median increases the power of the
test to find jumps. Also, by only using observations on the left of the cutoff value, we avoid estimating the
regression function at a point where it is known to have a discontinuity. To implement the test, use the same
method for selecting the binwidth as before, and estimate the jump in the regression function at qX ;1=2;l . Also,
estimate the standard errors of the jump and use this to test the hypothesis of a zero jump. Repeat this using
the subsample to the right of the cutoff point with X i Xc. Now estimate the jump in the regression function
and at qX ;1=2;r , and test whether it is equal to 0.

7.4. RD designs with misspecification

Lee and Card (2007) study the case where the forcing variable X is discrete. In practice this is of course
always the case. This implies that ultimately one relies for identification on functional form assumptions for
the regression function mðxÞ. Lee and Card consider a parametric specification for the regression function that
does not fully saturate the model, that is, it has fewer free parameters than there are support points. They then
interpret the deviation between the true conditional expectation E½Y jX ¼ x and the estimated regression
function as random specification error that introduces a group structure on the standard errors. Lee and Card
then show how to incorporate this group structure into the standard errors for the estimated treatment effect.
This approach will tend to widen the confidence intervals for the estimated treatment effect, sometimes
considerably, and leads to more conservative and typically more credible inferences. Within the local
linear regression framework discussed in the current paper, one can calculate the Lee–Card standard errors
(possibly based on slightly coarsened covariate data if X is close to continuous) and compare them to the
conventional ones.
 ARTICLE IN PRESS
G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635 633

7.5. Sensitivity to the choice of bandwidth

All these tests are based on estimating jumps in nonparametric regression or density functions. This brings
us to the third concern, the sensitivity to the bandwidth choice. Irrespective of the manner in which the
bandwidth is chosen, one should always investigate the sensitivity of the inferences to this choice, for example,
by including results for bandwidths twice (or four times) and half (or a quarter of) the size of the originally
chosen bandwidth. Obviously, such bandwidth choices affect both estimates and standard errors, but if the
results are critically dependent on a particular bandwidth choice, they are clearly less credible than if they are
robust to such variation in bandwidths. See Lee et al. (2004) and Lemieux and Milligan (2007) for examples of
papers where the sensitivity of the results to bandwidth choices is explored.

7.6. Comparisons to estimates based on unconfoundedness in the FRD design

When we have a FRD design, we can also consider estimates based on unconfoundedness (Battistin and
Rettore, 2007). In fact, we may be able to estimate the average effect of the treatment conditional on any value
of the covariate X under that assumption. Inspecting such estimates and especially their variation over the
range of the covariate can be useful. If we find that, for a range of values of X , our estimate of the average
effect of the treatment is relatively constant and similar to that based on the FRD approach, one would be
more confident in both sets of estimates.

8. Conclusion: a summary guide to practice

In this paper, we reviewed the literature on RD designs and discussed the implications for applied
researchers interested in implementing these methods. We end the paper by providing a summary guide of
steps to be followed when implementing RD designs. We start with the case of SRD, and then add a number
of details specific to the case of FRD.
Case 1: SRD designs

1. Graph the data (Section 3) by computing the average value of the outcome variable over a set of bins. The
binwidth has to be large enough to have a sufficient amount of precision so that the plots looks smooth on
either side of the cutoff value, but at the same time small enough to make the jump around the cutoff value
clear.
2. Estimate the treatment effect by running linear regressions on both sides of the cutoff point. Since we
propose to use a rectangular kernel, these are just standard regression estimated within a bin of width h on
both sides of the cutoff point. Note that:
 Standard errors can be computed using standard least square methods (robust standard errors).
 The optimal bandwidth can be chosen using cross-validation methods (Section 5).
3. The robustness of the results should be assessed by employing various specification tests.
 Looking at possible jumps in the value of other covariates at the cutoff point (Section 7.1).
 Testing for possible discontinuities in the conditional density of the forcing variable (Section 7.2).
 Looking whether the average outcome is discontinuous at other values of the forcing variable (Section 7.3).
 Using various values of the bandwidth (Section 7.5), with and without other covariates that may be
available.

Case 2: FRD designs

A number of issues arise in the case of FRD designs in addition to those mentioned above.

1. Graph the average outcomes over a set of bins as in the case of SRD, but also graph the probability of
treatment.
2. Estimate the treatment effect using TSLS, which is numerically equivalent to computing the ratio
in the estimate of the jump (at the cutoff point) in the outcome variable over the jump in the treatment
variable.
 ARTICLE IN PRESS
634 G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635

 Standard errors can be computed using the usual (robust) TSLS standard errors (Section 6.3), though a
plug-in approach can also be used instead (Section 6.2).
 The optimal bandwidth can again be chosen using a modified cross-validation procedure (Section 5).
3. The robustness of the results can be assessed using the various specification tests mentioned in the case of
SRD designs. In addition, FRD estimates of the treatment effect can be compared to standard estimates
based on unconfoundedness.

Acknowledgments

We are grateful for discussions with David Card and Wilbert Van Der Klaauw. Financial support for this
research was generously provided through NSF Grant SES 0452590 and the SSHRC of Canada.

References

Angrist, J.D., Krueger, A.B., 1991. Does compulsory school attendance affect schooling and earnings? Quarterly Journal of Economics
106, 979–1014.
Angrist, J.D., Lavy, V., 1999. Using Maimonides’ rule to estimate the effect of class size on scholastic achievement. Quarterly Journal of
Economics 114, 533–575.
Angrist, J.D., Imbens, G.W., Rubin, D.B., 1996. Identification of causal effects using instrumental variables. Journal of the American
Statistical Association 91, 444–472.
Battistin, E., Rettore, E., 2007. Ineligibles and eligible non-participants as a double comparison group in regression-discontinuity designs.
Journal of Econometrics, this issue.
Black, S., 1999. Do better schools matter? Parental valuation of elementary education. Quarterly Journal of Economics 114, 577–599.
Card, D., Dobkin, C., Maestas, N., 2004. The impact of nearly universal insurance coverage on health care utilization and health: evidence
from Medicare. NBER Working Paper No. 10365.
Card, D., Mas, A., Rothstein, J., 2006. Tipping and the dynamics of segregation in neighborhoods and schools. Unpublished Manuscript.
Department of Economics, Princeton University.
Chay, K., Greenstone, M., 2005. Does air quality matter? Evidence from the housing market. Journal of Political Economy 113, 376–424.
Chay, K., McEwan, P., Urquiola, M., 2005. The central role of noise in evaluating interventions that use test scores to rank schools.
American Economic Review 95, 1237–1258.
DiNardo, J., Lee, D.S., 2004. Economic impacts of new unionization on private sector employers: 1984–2001. Quarterly Journal of
Economics 119, 1383–1441.
Fan, J., Gijbels, I., 1996. Local Polynomial Modelling and its Applications. Chapman & Hall, London.
Hahn, J., Todd, P., Van Der Klaauw, W., 1999. Evaluating the effect of an anti discrimination law using a regression-discontinuity design.
NBER Working Paper No. 7131.
Hahn, J., Todd, P., Van Der Klaauw, W., 2001. Identification and estimation of treatment effects with a regression discontinuity design.
Econometrica 69, 201–209.
Härdle, W., 1990. Applied Nonparametric Regression. Cambridge University Press, Cambridge.
Heckman, J.J., Hotz, J., 1989. Alternative methods for evaluating the impact of training programs (with discussion). Journal of the
American Statistical Association 84, 862–874.
Holland, P., 1986. Statistics and causal inference (with discussion). Journal of the American Statistical Association 81, 945–970.
Imbens, G., 2004. Nonparametric estimation of average treatment effects under exogeneity: a review. Review of Economics and Statistics
86, 4–30.
Imbens, G., Angrist, J., 1994. Identification and estimation of local average treatment effects. Econometrica 61, 467–476.
Imbens, G., Rubin, D., 2007. Causal Inference: Statistical Methods for Estimating Causal Effects in Biomedical, Social, and Behavioral
Sciences. Cambridge University Press, Cambridge forthcoming.
Imbens, G., Van Der Klaauw, W., 1995. Evaluating the cost of conscription in The Netherlands. Journal of Business and Economic
Statistics 13, 72–80.
Imbens, G., Wooldridge, J., 2007. Recent developments in the econometrics of program evaluation. Unpublished Manuscript, Department
of Economics, Harvard University.
Jacob, B.A., Lefgren, L., 2004. Remedial education and student achievement: a regression-discontinuity analysis. Review of Economics
and Statistics 68, 226–244.
Lalive, R., 2007. How do extended benefits affect unemployment duration? A regression discontinuity approach. Journal of Econometrics,
this issue.
Lee, D.S., 2007. Randomized experiments from non-random selection in U.S. house elections. Journal of Econometrics, this issue.
Lee, D.S., Card, D., 2007. Regression discontinuity inference with specification error. Journal of Econometrics, this issue.
Lee, D.S., Moretti, E., Butler, M., 2004. Do voters affect or elect policies? Evidence from the U.S. house. Quarterly Journal of Economics
119, 807–859.
 ARTICLE IN PRESS
G.W. Imbens, T. Lemieux / Journal of Econometrics 142 (2008) 615–635 635

Lemieux, T., Milligan, K., 2007. Incentive effects of social assistance: a regression discontinuity approach. Journal of Econometrics, this
issue.
Li, Q., Racine, J., 2007. Nonparametric Econometrics. Princeton University Press, Princeton, NJ.
Ludwig, J., Miller, D., 2005. Does head start improve children’s life chances? Evidence from a regression discontinuity design. NBER
Working Paper No. 11702.
Ludwig, J., Miller, D., 2007. Does head start improve children’s life chances? Evidence from a regression discontinuity design. Quarterly
Journal of Economics 122 (1), 159–208.
Matsudaira, J., 2007. Mandatory summer school and student achievement. Journal of Econometrics, this issue.
McCrary, J., 2007. Testing for manipulation of the running variable in the regression discontinuity design. Journal of Econometrics, this
issue.
Pagan, A., Ullah, A., 1999. Nonparametric Econometrics. Cambridge University Press, Cambridge.
Porter, J., 2003. Estimation in the Regression Discontinuity Model. Mimeo. Department of Economics, University of Wisconsin. hhttp://
www.ssc.wisc.edu/jporter/reg_discont_2003.pdfi.
Rosenbaum, P., 1987. The role of a second control group in an observational study (with discussion). Statistical Science 2, 292–316.
Rosenbaum, P., Rubin, D., 1983. The central role of the propensity score in observational studies for causal effects. Biometrika 70, 41–55.
Rubin, D., 1974. Estimating causal effects of treatments in randomized and non-randomized studies. Journal of Educational Psychology
66, 688–701.
Sun, Y., 2005. Adaptive estimation of the regression discontinuity model. Unpublished Manuscript, Department of Economics, University
of California at San Diego.
Thistlewaite, D., Campbell, D., 1960. Regression-discontinuity analysis: an alternative to the ex-post facto experiment. Journal of
Educational Psychology 51, 309–317.
Trochim, W., 1984. Research Design for Program Evaluation: The Regression-Discontinuity Design. Sage Publications, Beverly Hills, CA.
Trochim, W., 2001. Regression-discontinuity design. In: Smelser, N.J., Baltes, P.B. (Eds.), International Encyclopedia of the Social and
Behavioral Sciences, vol. 19. North-Holland, Amsterdam, pp. 12940–12945.
Van Der Klaauw, W., 2002. Estimating the effect of financial aid offers on college enrollment: a regression-discontinuity approach.
International Economic Review 43, 1249–1287.