Regression Discontinuity Designs Using Covariates∗

Sebastian Calonico† Matias D. Cattaneo‡ Max H. Farrell§ Rocı́o Titiunik¶

March 31, 2016

Abstract

We study identification, estimation, and inference in Regression Discontinuity (RD) designs

when additional covariates are included in the estimation. Standard RD estimation and in-
ference is based on nonparametric local polynomial methods using two variables: the outcome
variable and the running variable that determines treatment assignment. Applied researchers of-
ten include additional “pre-intervention” covariates in their specifications to increase efficiency.
However, no results justifying covariate adjustment have been formally derived in the RD lit-
erature, leaving applied researchers with little practical guidance and leading to a proliferation
of ad-hoc methods that may result in invalid estimation and inference. We examine the prop-
erties of a local polynomial estimator that incorporates discrete and continuous covariates in
an additive separable, linear-in-parameters way and imposes a common (likely misspecified) co-
variate effect on both sides of the cutoff. Under intuitive, minimal assumptions, we show that
this covariate-adjusted RD estimator remains consistent for the standard RD treatment effect,
while also providing point estimation and inference improvements. In contrast, we show that
estimating a specification with interactions between treatment status and the covariates leads
to an estimator that is inconsistent in general. We present new asymptotic mean squared error
expansions, optimal bandwidth choices, optimal point estimators, robust nonparametric infer-
ence procedures based on bias-correction techniques, and heteroskedasticity-consistent standard
errors. Our results cover sharp, fuzzy, and kink RD designs, and we also discuss extensions
to clustered data. Finally, we present two empirical illustrations where we find 5% to 10%
reduction in confidence interval length, and an extensive simulation study. All methods are
implemented in companion R and Stata software packages.

Keywords: regression discontinuity, covariate adjustment, causal inference, local polynomial

methods, robust inference, bias correction, tuning parameter selection.

∗
We thank Stephane Bonhomme, David Drukker, Kosuke Imai, Michael Jansson, Lutz Kilian, Pat Kline, Xinwei
Ma, Andres Santos, and Gonzalo Vazquez-Bare for thoughtful comments and suggestions. Cattaneo gratefully ac-
knowledges financial support from the National Science Foundation through grants SES-1357561 and SES-1459931,
and Titiunik gratefully acknowledges financial support from the National Science Foundation through grant SES-
1357561.
†
Department of Economics, University of Miami.
‡
Department of Economics and Department of Statistics, University of Michigan.
§
Booth School of Business, University of Chicago.
¶
Department of Political Science, University of Michigan.
1 Introduction

The Regression Discontinuity (RD) design is widely used in Economics and other social, behavioral,

biomedical, and statistical sciences. Within the causal inference framework, this design is considered

among the most credible non-experimental strategies because it relies on relatively weak and easy-

to-interpret nonparametric identifying assumptions, which permit flexible and robust estimation

and inference for (local to the cutoff) treatment effects. The key feature of the design is the

existence of a score, index, or running variable for each unit in the sample, which determines

treatment assignment via hard-thresholding: all units whose score is above a known cutoff are

offered treatment, while all units whose score is below this cutoff are not. Identification, estimation,

and inference proceed by comparing the responses of units near the cutoff, taking those below

(comparison group) as counterfactuals to those above (treatment or intention-to-treat group). For

reviews see Imbens and Lemieux (2008), Lee and Lemieux (2010), Skovron and Titiunik (2016),

and references therein.

The most common approach to nonparametric identification in RD designs relies on continuity

assumptions. Under this approach, estimation of average treatment effects at the cutoff typically

relies on nonparametric local polynomial methods, where the unknown (but assumed-smooth) re-

gression function of the outcome variable given the score is flexibly approximated above and below

the cutoff, and then these estimates are used to assess whether there is a discontinuity in lev-

els, derivatives, or ratios thereof, at the cutoff. This discontinuity, if present, is understood as

the average response to the treatment, intention-to-treat, treatment effect on the treated, or local

average treatment effect, at the cutoff, depending on the specific setting and assumptions under

consideration.

A natural estimation strategy is to fit separate local polynomial regressions of the outcome on

the score above and below the cutoff, but in practice researchers often augment their models with

“pre-intervention” covariates in addition to the running variable. The main motivation behind this

practice is to increase the precision of the RD treatment effect estimator. In addition, covariates

are sometimes added with the goal of improving the plausibility of the RD design, though this

second motivation is much harder to justify because it rests on additional strong assumptions

(see Remark 1 below). The practice of including covariates in RD estimation has its roots in the

common analogy between RD designs and randomized experiments. Since the RD design is often

1
(formally or informally) thought of as a randomized experiment near the cutoff (Lee, 2008; Cattaneo

et al., 2015), and the inclusion of covariates is often used in the analysis of experiments to increase

precision, inclusion of covariates in RD estimation seems natural. However, despite covariate-

adjusted RD analysis being widespread in empirical practice, there is no existing justification for

using additional covariates for identification, estimation, or inference purposes, employing only

continuity/smoothness conditions at the cutoff. This has led to the proliferation of ad-hoc covariate-

adjustment practices that, at best, reduce the transparency of the estimation strategy and, at worst,

result in generally noncomparable (or even inconsistent) estimators.

We provide a set of results that formalize and justify covariate adjustment in RD designs,

and offer valid estimation and inference procedures. Following empirical practice, we augment the

standard RD framework in order to codify the inclusion of covariates. In particular, we study local

polynomial methods allowing for the inclusion of additional covariates in an additive separable,

linear-in-parameters way, which permits continuous, discrete, and mixed regressors and does not

require additional smoothing methods (e.g., no need for choosing other bandwidths or kernels).

This procedure for covariate-adjusted RD estimation covers linear model adjustments, which are

popular in applied work, and allows us to characterize not only the conditions under which the

inclusion of covariates is appropriate, but also the ways in which adjusting by covariates may

lead to inconsistent RD estimators. Thus, our formal results offer concrete guidelines for applied

researchers that were previously unavailable.

Under minimal smoothness assumptions, we show that the covariate-adjusted RD estimator

that imposes the same adjustment above and below the cutoff is consistent for the standard RD

treatment effect if a simple “zero RD treatment effect on covariates” condition holds. For example,

in the sharp RD design, the only requirement is that the covariates have equal conditional expecta-

tion from above and below at the cutoff, which is often conceived and presented as a falsification or

“placebo” test in RD empirical studies (see, e.g., Lee, 2008; Canay and Kamat, 2015, and references

therein). This requirement of “balanced” covariates at the cutoff, in the appropriate sense depend-

ing on the RD design considered, is the most natural and practically relevant sufficient condition

but, more generally, we are able to obtain necessary and sufficient conditions for consistency of the

covariate-adjusted RD estimator. We also discuss identification and consistency properties of an

alternative covariate-adjusted RD estimator that includes an interaction between treatment and

2
covariates, and show that this estimator is generally inconsistent for the standard RD parameter of

interest. We also characterize the (necessary and) sufficient conditions required for this alternative

estimator to be consistent, which are strong and unlikely to hold in empirical settings.

We offer a complete asymptotic analysis for the covariate-adjusted RD estimator, including novel

mean squared error (MSE) expansions, several MSE-optimal bandwidths and consistent data-driven

implementations thereof, MSE-optimal point estimators, and valid asymptotic inference, covering

all empirically relevant RD designs (sharp RD, kink RD, fuzzy RD, and fuzzy kink RD), with both

heteroskedastic and clustered data. These results have immediate practical use in any RD analysis

and aid in interpreting prior results. In particular, we characterize precisely the source of efficiency

gains obtained when using the covariate-adjusted RD estimator (see Remarks 2 and 3). Last but

not least, we provide new general purpose Stata and R packages that implement all our results—see

Calonico, Cattaneo, Farrell and Titiunik (2016b) and references therein for more details.

We illustrate our methods with two empirical applications. First, we employ the data of Ludwig

and Miller (2007) to re-analyze the effect of Head Start on child mortality, where we find that

including nine pre-intervention 1960 census covariates leads to an average reduction of confidence

interval length of about 5% to 10% relative to the case without covariate-adjustment. Second, we

use the data of Chay, McEwan and Urquiola (2005) on the effect of school improvements on test

scores, where we see a 3% to 5% reduction in confidence interval length when six region-indicator

covariates are included. Finally, we also discuss the findings from an extensive simulation study

investigating the finite-sample properties of our methods.

Our paper contributes to the large and still rapidly expanding methodological literature on RD

designs. Instead of giving a (likely incomplete) summary here, we defer to the review articles cited

in our opening paragraph for references and the references given throughout the manuscript. In

addition, our main results are also connected to the causal inference literature on covariate-adjusted

treatment effect estimation in randomized experiments. For a recent review see Imbens and Ru-

bin (2015) and references therein. In the specific context of RD designs, one recent strand of the

literature re-interprets the data as being “as good as randomized” within a small window around

the cutoff. This so-called “local randomization” RD approach requires conditions stronger than

continuity/smoothness of conditional expectations, but allows classical techniques and interpreta-

tions from randomized experiments to be brought to bear (Cattaneo et al., 2015, 2016). Although

3
we maintain the more standard continuity framework throughout this article, our main results

show that the large sample implications of a “local randomization” approach to RD designs carry

over for the particular case of covariate adjustment—the only requirement is a weak and intuitive

restriction on the additional “pre-intervention” covariates distribution included in the estimation.

The plan of presentation is as follows. The bulk of the paper is devoted to an in-depth treatment

of the sharp RD design, with a brief section discussing extensions to other popular RD settings.

Specifically, Section 2 introduces a framework for covariate-adjustment in sharp RD designs, while

Section 3 details important identification and interpretation issues. Section 4 gives a complete

analysis of nonparametric inference in sharp RD designs using covariates, including new MSE ex-

pansions, MSE-optimal estimators, valid inference based on robust bias-correction techniques, and

consistent standard errors. Section 5 discusses extensions to other RD designs. Section 6 presents

the results from the empirical illustrations and the simulation study, and Section 7 concludes. The

Appendix contains the main formulas concerning the sharp RD design, omitted from the main

text to ease the exposition. A supplemental appendix includes: (i) a thorough theoretical treat-

ment of all RD cases and extensions, including proofs of the results herein, (ii) a discussion on

implementation and other methodological details, and (iii) the complete set of Monte Carlo results.

2 Sharp RD Designs Using Covariates

The observed data is assumed to be a random sample (Yi , Ti , Xi , Z0i )0 , i = 1, 2, . . . , n, from a large

population. The key feature of any RD design is the presence of an observed continuous score or

running random variable Xi , which determines treatment assignment for each unit in the sample:

all units with Xi greater than a known threshold x̄ are assigned to the treatment group, while all

units with Xi < x̄ are assigned to the control group. In sharp RD designs, treatment compliance

is perfect and hence Ti = 1(Xi ≥ x̄) denotes treatment status. Using the potential outcomes
framework, the observed Yi is given by


Yi (0)
 if Ti = 0
Yi = Yi (0) · (1 − Ti ) + Yi (1) · Ti =

Yi (1)
 if Ti = 1,

4
where Yi (1) and Yi (0) denote the potential outcomes with and without treatment, respectively, for

each unit i in the sample. The parameter of interest is the average treatment effect at the cutoff:

τ = τ (x̄) = E[Yi (1) − Yi (0)|Xi = x̄].

Evaluation points of functions are dropped whenever possible throughout the paper. Hahn, Todd

and van der Klaauw (2001) gave precise, easy-to-interpret conditions for nonparametric identifi-

cation of the standard RD treatment effect τ , without additional covariates. The key substantive

identifying assumption is that E[Yi (t)|Xi = x], t ∈ {0, 1}, be continuous at the cutoff x = x̄.

The new feature studied in this paper is the presence of additional “pre-intervention”, “pre-

determined”, or “exogenous” covariates, collected in the random vector Zi ∈ Rd , which could be

continuous, discrete, or mixed. Without loss of generality, we assume



Zi (0)
 if Ti = 0
Zi = Zi (0) · (1 − Ti ) + Zi (1) · Ti =

Zi (1)
 if Ti = 1,

where Zi (1) and Zi (0) denote the (potential) covariates on either side of the threshold. In practice,

it is natural to assume that some features of the marginal distributions of Zi (1) and Zi (0) are

equal at the cutoff x̄ or, more extremely, that Zi (1) = Zi (0), which would match the definition of

a “pre-treatment” covariate in the context of randomized controlled trials.

A large portion of the literature on estimation and inference in RD designs focuses on nonpara-

metric local polynomial estimators. In practice, researchers first choose a neighborhood around the

cutoff, usually via a bandwidth choice, and then conduct local polynomial inference—that is, they

rely on linear regression fits using only units whose scores lay within that pre-selected neighborhood.

The standard RD treatment effect estimator is then given by

τ̂ (h) = e00 β̂ Y +,p (h) − e00 β̂ Y −,p (h),

where β̂ Y −,p (h) and β̂ Y +,p (h) correspond to the weighted least squares coefficients

 
n
 β̂ Y −,p (h)  X
  = arg min (Yi − r−,p (Xi − x̄)0 β − − r+,p (Xi − x̄)0 β + )2 Kh (Xi − x̄), (1)
β̂ Y +,p (h) β ,β
− + i=1

5
with β − , β + ∈ Rp+1 , r−,p (x) = 1(u < 0)(1, x, · · · , xp )0 , r+,p (x) = 1(u ≥ 0)(1, x, · · · , xp )0 , e0 the

(p + 1)-vector with a one in the first position and zeros in the rest, and Kh (u) = K(u/h)/h for

a kernel function K(·) and a positive bandwidth sequence h. The kernel and bandwidth serve to

localize the regression fit near the cutoff. We assume the following standard regularity conditions

for the kernel.

Assumption 1 (Kernel). k(·) : [0, 1] 7→ R is bounded and nonnegative, zero outside its support,

and positive and continuous on (0, 1). Set K(u) = 1(u < 0)k(−u) + 1(u ≥ 0)k(u).

The most popular choices of kernel are (i) the uniform kernel, giving equal weighting to obser-

vations Xi ∈ [x̄ − h, x̄ + h], and (ii) the triangular kernel that assigns linear down-weighting to the

same observations. The preferred choice of polynomial order is p = 1, which gives the standard

local-linear RD point estimator. The estimators β̂ Y −,p (h) and β̂ Y +,p (h) are, of course, numerically

equivalent to the coefficients that would be obtained from two separate weighted regressions, using

only observations on one side of the cutoff (with the same kernel and bandwidth). We set the

problem as a single joint least-squares linear regression fit to ease the upcoming comparisons with

the covariate-adjusted RD estimator.

While the standard estimator τ̂ (h) is popular in empirical work, and readily justified by local

smoothness assumptions, it is extremely common for empirical researchers to augment their speci-

fication with the additional covariates Zi . One way of introducing covariates leads to the following

estimator, which we call the covariate-adjusted RD estimator :

τ̃ (h) = e00 β̃ Y +,p (h) − e00 β̃ Y −,p (h),

where β̃ Y +,p (h) and β̃ Y −,p (h) are defined through

 
β̃
 Y −,p (h)  Xn
(Yi − r−,p (Xi − x̄)0 β − − r+,p (Xi − x̄)0 β + − Z0i γ)2 Kh (Xi − x̄),
 
 Y +,p (h)  = arg min (2)
 β̃ 
  β− ,β+ ,γ i=1
γ̃ Y,p (h)

where β − , β + ∈ Rp+1 and γ ∈ Rd . Throughout the paper and supplemental appendix we employ

the following notational convention whenever possible: a quantity with a tilde (θ̃, say) is estimated

with additional covariates, while a quantity with a hat (θ̂, say) is not; cf. (1) vs. (2).

6
 The estimator τ̃ (h) broadly captures the common empirical practice of first choosing a neigh-

borhood around the cutoff, and then conducting local “flexible” linear least-squares estimation

and inference with covariates. But our approach formalizes two restrictions in the way that the

additional covariates Zi enter the least-squares fit locally to the cutoff: (i) additive separabil-

ity between the basis expansion of the running variable and the additional covariates, and (ii) a

linear-in-parameters specification for these covariates. We avoid full nonparametric estimation over

(Xi , Z0i )0 ∈ R1+d , which would introduce d additional bandwidths and kernels, quickly leading to a

curse of dimensionality and hence rendering empirical application infeasible. Further, in practice,

Zi could include power expansions, interactions, and other “flexible” transformations of the origi-

nal set of covariates. This approach to RD covariate adjustment allows for any type of additional

regressors, including fixed effects or other discrete variables.

The typical motivation for using the covariate-adjusted RD estimator τ̃ (h) is to improve preci-

sion in estimating the RD treatment effect, τ , which arguably stems from least squares analysis of

randomized controlled trials. We build on this intuition and make precise the conditions required

for consistency of the covariate-adjusted RD estimator τ̃ (h) for τ . We also show that much more

stringent conditions are required if treatment-covariate interactions are included in the estimation

model.

Remark 1 (Covariates and Identification). A common misconception among some researchers

relying on continuity-based nonparametric identification results for RD designs is that including

additional (pre-intervention) covariates can enhance the plausibility of the design. Specifically, it is

sometimes claimed that even if E[Yi (t)|Xi = x], t ∈ {0, 1}, is not continuous at the cutoff, adding

covariates could restore nonparametric identification of the RD average treatment effect at the

cutoff. However, within the continuity-based RD framework, if E[Yi (t)|Xi = x, Zi (t)], t ∈ {0, 1},

is indeed continuous, then E[Yi (t)|Xi = x] = E[E[Yi (t)|Xi , Zi (t)]|Xi = x] will be continuous under

most reasonable assumptions. For example, suppose that Zi (0) and Zi (1) are binary (e.g., gender),

then E[Yi (t)|Xi = x] = E[Yi (t)|Xi = x, Zi (t) = 0]P[Zi (t) = 0|Xi = x] + E[Yi (t)|Xi = x, Zi (t) =

1]P[Zi (t) = 1|Xi = x] is a linear combination of assumed-continuous functions and hence must

be continuous as well. Thus, covariate adjustment does not solve identification problems when

E[Yi (t)|Xi = x] is discontinuous. On the other hand, adding covariates for identification purposes

can be rationalized and be useful within a local randomization framework for RD designs (e.g.,

7
Angrist and Rokkanen, 2015; Keele et al., 2015). Below we discuss in more detail the distinctions

between these frameworks (see also Cattaneo et al., 2016). 

2.1 Notation and Regularity Conditions

To make precise the difference in population parameters recovered with and without additional

covariates, and to analyze the asymptotic properties of the covariate-adjusted RD estimators, it is

useful to establish notation and state all the regularity conditions employed. This is done simul-

taneously in the assumption below, which is the only assumption imposed on the data generating

process.

Assumption 2 (Sharp RD Designs). For % ≥ p + 2 and all x ∈ [xl , xu ], where xl , xu ∈ R such that

xl < x̄ < xu :

(a) The Lebesgue density of Xi , denoted f (x), is continuous and bounded away from zero.

(b) µY − (x) := E[Yi (0)|Xi = x], µY + (x) := E[Yi (1)|Xi = x], µZ− (x) := E[Zi (0)|Xi = x],

µZ+ (x) := E[Zi (1)|Xi = x], E[Zi (0)Yi (0)|Xi = x], and E[Zi (1)Yi (1)|Xi = x] are % times con-

tinuously differentiable.

(c) V[Si (t)|Xi = x], with Si (t) := (Yi (t), Zi (t)0 )0 , t ∈ {0, 1}, are continuously differentiable and

invertible.

(d) E[|Si (t)|4 |Xi = x], t ∈ {0, 1}, are continuous, where | · | denotes the Euclidean norm.

Assumption 2 imposes standard continuity/smoothness assumptions common to all nonpara-

metric analyses of RD designs, properly enlarged to allow for the inclusion of additional covariates.

Indeed, if one simply ignores all statements involving these covariates, the conditions are exactly

those found in the RD literature.

The assumptions are placed only on features such as the mean and variance of the conditional

distributions given the running variable Xi alone. Importantly, Assumption 2 does not restrict in

any way the “long” conditional expectation E[Yi (t)|Xi , Zi (t)], t ∈ {0, 1}, which implies that our

methods allow for discrete, continuous, and mixed additional covariates, and do not require any

semiparametric or parametric modeling of this regression function. That is, as we discuss in more

detail below, we allow for complete misspecification of E[Yi (t)|Xi , Zi (t)], t ∈ {0, 1}, for fixed n, and

hence give a “best linear approximation” interpretation to the regression coefficients obtained in

(2).

8
 Assumption 2 is not intended to be minimal, but rather parsimonious and easily applicable to

nonparametric estimation and inference. Finally, all limits are taken as n → ∞, unless otherwise

noted.

3 Identification and Estimation in Sharp RD Designs Using Co-

variates

We now present the first main result of the paper, which connects and gives an interpretation to

the implicit estimand associated with the covariate-adjusted RD estimator.

Lemma 1 (Sharp RD with Covariates). Let Assumptions 1 and 2 hold. If nh → ∞ and h → 0,

then
 0
τ̃ (h) →P τ − µZ+ − µZ− γ Y ,

where

γ Y = (σ 2Z− + σ 2Z+ )−1 E (Zi (0) − µZ− (Xi )) Yi (0) + (Zi (1) − µZ+ (Xi )) Yi (1) Xi = x̄ ,
 

where σ 2Z− := V[Zi (0)|Xi = x] and σ 2Z+ := V[Zi (1)|Xi = x].

It is well known in the RD literature that, under the conditions of Lemma 1, τ̂ (h) →P τ . The

conclusion of this lemma gives a precise description of the probability limit of the covariate-adjusted

sharp RD estimator, when implemented according to (2). A similar result is discussed for all other

popular RD designs in Section 5.

Lemma 1 shows that this covariate-adjusted RD estimation approach is consistent for the stan-

dard RD treatment effect at the cutoff, τ = µY + − µY − , plus an additional term that depends

on the RD treatment effect at the cutoff for the additional covariates, τ Z := µZ+ − µZ− . It

follows that, given the smoothness conditions imposed in Assumption 2, a sufficient condition for

τ̃ (h) →P τ is that µZ+ = µZ− . This is weaker than assuming that the marginal distributions of

Zi (0) and Zi (1) are equal at the cutoff. In other words, µZ+ = µZ− is implied by, but does not

require that, P[Zki (0) ≤ z|Xi = x̄] = P[Zki (1) ≤ z|Xi = x̄] for all z and k = 1, 2, · · · , d, where

Zi (t) = [Z1i (t), Z2i (t), · · · , Zdi (t)]0 with t ∈ {0, 1}.

The typical motivation for including covariates in RD analyses is to gain precision in estimating

9
the RD treatment effect of interest, τ , which has its roots in the analysis of randomized experiments.

Even if implicitly, researchers employing covariates in RD designs assume some form of “local

randomization”, where units are thought to be assigned to treatment or control at random near the

cutoff: {Yi (0), Yi (1), Zi (0), Zi (1)} ⊥

⊥ Ti | Xi ∈ [x̄ − h, x̄ + h]. Such local randomization is discussed

intuitively by Lee (2008) and Lee and Lemieux (2010); more recently, Cattaneo et al. (2015, 2016)

discuss the stronger conditions, beyond continuity, required for the interpretation and valid analysis

of RD designs as local randomized experiments. (See also de la Cuesta and Imai, 2016, for a recent

discussion of the distinction between continuity and local randomization.) From this perspective,

“pre-intervention” covariates would satisfy Zi (0) = Zi (1) conditional on Xi ∈ [x̄ − h, x̄ + h], that

is, their distributions would be equal among control and treatment units near the cutoff.

In contrast, in this paper we do not assume a local randomization condition of any form, but

rather focus on continuity-based methods and hence (superpopulation) nonparametric identifica-

tion. In this setting, Lemma 1 shows that only continuity of the conditional expectations of the

additional covariates at the cutoff is the key condition required for consistency of the covariate-

adjusted RD estimator. In other words, whenever additional covariates satisfying µZ+ = µZ− are

included as in (2), the estimator τ̃ (h) will remain consistent for the standard RD estimand τ .

3.1 Identification and Estimation with Treatment Interaction

In addition to efficiency gains, another common motivation for examining covariates in randomized

experiments is to discover treatment effect heterogeneity, and covariate-adjusted linear regression

is a frequently used method for doing so. A potentially interesting extension of our work, and

in particular Lemma 1, would be to further augment the covariate-adjusted RD estimator τ̃ (h)

implemented as in (2) with interactions between rp (Xi − x̄) and Zi . This alternative estimation

method might be useful to assess treatment effect heterogeneity at the cutoff, as well as to provide

a more “flexible” approximation of the unknown conditional expectations in finite samples. While

such general approach is beyond the scope of this paper, we do discuss a special case of this idea

to illustrate the potential pitfalls of allowing for interactions in the local polynomial fits.

To be specific, we will examine what we call the treatment-interacted covariate-adjusted RD

estimator. This estimator, sometimes used in empirical work, is given by

η̌(h) = e00 β̌ Y +,p (h) − e00 β̌ Y −,p (h),

10
where now θ̌ p (h) = [β̌ Y −,p (h)0 , β̌ Y +,p (h)0 , γ̌ Y −,p (h)0 , γ̌ Y +,p (h)0 ]0 is computed by

n
X
min (Yi − r−,p (Xi − x̄)0 β − − r+,p (Xi − x̄)0 β + − Z0−,i γ − − Z0+,i γ + )2 Kh (Xi − x̄),
β − ,β + ,γ − ,γ +
i=1

where Z−,i = 1(Xi < x̄)Zi and Z+,i = 1(Xi ≥ x̄)Zi , and β− , β+ ∈ Rp+1 and γ − , γ + ∈ Rd .
In words, this alternative estimator fits a weighted least squares regression with full interactions

between treatment assignment and both the polynomial basis expansion of Xi and the additional

covariates Zi . Thus, θ̌ p (h) is numerically equivalent to fitting two separate weighted linear regres-

sions on each side of the cutoff, leading to [β̌ Y −,p (h)0 , γ̌ Y −,p (h)0 ] and [β̌ Y +,p (h)0 , γ̌ Y +,p (h)0 ]. We

present the estimation approach in a fully interacted version only for notation simplicity, so that

γ − and γ + have a symmetric interpretation.

As shown in the next lemma, including the treatment-covariate interaction has important impli-

cations for interpretation. The difference follows from the fact that including this interaction allows

γ − 6= γ + in the estimation, whereas the covariate-adjusted RD estimator of (2) forces equality.

Lemma 2 (Sharp RD with Covariates and Treatment Interaction). Let Assumptions 1 and 2 hold.

If nh → ∞ and h → 0, then

η̌(h) →P η := τ − µ0Z+ γ Y + − µ0Z− γ Y − ,

 

where

γ Y − = (σ 2Z− )−1 E (Zi (0) − µZ− (Xi )) Yi (0) Xi = x̄ ,

 

γ Y + = (σ 2Z+ )−1 E (Zi (1) − µZ+ (Xi )) Yi (1) Xi = x̄ .

 

The conclusion of this lemma defines a new RD parameter, η, recovered when additional covari-

ates interacted with treatment assignment are included linearly in the local polynomial estimation.

A similar result is established for all other RD designs in the supplement. This result gives a precise

and general interpretation to the probability limit of the interacted covariate-adjusted RD estima-

tor: η̌(h) is consistent for the standard RD average treatment effect at the cutoff, τ = µY + − µY − ,

plus an additional term which can be interpreted as the difference of the best linear approximations

at the cutoff of the unknown conditional expectations E[Yi (t)|Xi , Zi (0)], t ∈ {0, 1}, based on the

additional covariates included in the RD estimation. Alternatively, the “bias” due to the inclu-

11
sion of additional covariates interacted with treatment assignment, µ0Z+ γ Y + − µ0Z− γ Y − , can be

interpreted as the difference of the best linear predictions of Yi (t) on Zi (t), t ∈ {0, 1}, at the cutoff.

It follows that, when a treatment-covariate interaction is included in the estimation, a necessary

and sufficient condition for the resulting covariate-adjusted RD estimator to be consistent for the

standard RD treatment effect is that µ0Z+ γ Y + = µ0Z− γ Y − . This condition, however, is harder to

justify in practice than the condition required for the model without the interaction. In particular,

the previously sufficient condition µZ+ = µZ− (“covariate balance”) is now no longer sufficient

because one needs also to assume that γ Y + = γ Y − . The latter additional assumption can be

regarded as an “homogeneous partial effect of covariates on potential outcomes”.

The interacted covariate-adjusted estimand η can also be interpreted as a “partial effect” RD

estimator. To make this clear, consider the following example. Suppose that E[Yi (0)|Xi , Zi (0)] =

ξ− (Xi ) + Zi (0)0 δ Y − and E[Yi (1)|Xi , Zi (1)] = ξ+ (Xi ) + Zi (1)0 δ Y + near the cutoff. Then γ Y − = δ Y −

and γ Y + = δ Y + and η̌(h) →P η = ξ+ (x̄) − ξ− (x̄) 6= τ , and hence the interacted covariate-

adjusted RD estimator η̌(h) is consistent for a partial effect at the cutoff. In this example, the

additional condition required for η̌(h) →P τ becomes δ Y − = δ Y + , which in turn is implied by

E[Yi (t)|Xi , Zi (t)] = E[Yi (t)|Xi ] near the cutoff, though the latter is not required.

3.2 Practical Implications

Lemmas 1 and 2 not only give general, precise, and intuitive characterizations of the probability

limits of two popular covariate-adjusted RD estimators, but also have interesting implications for

the analysis and interpretation of RD designs using covariates. Most notably, the lemmas above

show the conditions under which a covariate-adjusted RD estimator is consistent for the standard

(causal) RD treatment effect of interest, τ , and, by implication, establish when estimators with and

without covariate adjustment can be compared to each other.

Since in most applications τ is the parameter of interest, comparing τ̂ (h) vs. τ̃ (h) requires the

assumption µZ+ = µZ− (Lemma 1), while comparing τ̂ (h) vs. η̌(h) requires µ0Z+ γ Y + = µ0Z− γ Y −

(Lemma 2). Adjusting for covariates in RD settings seems most useful when the estimand of interest

remains unchanged, in which case comparing precision becomes meaningful (as in Remarks 2 and

3). In applications, there is no a priori reason to blindly compare different estimators (τ̂ (h), τ̃ (h),

η̌(h)) without imposing (and, possibly, testing for) the underlying sufficient assumptions required

12
to retain the same target RD treatment effect of interest.

4 Inference in Sharp RD Designs using Covariates

Estimation and inference in RD designs using local polynomial methods without covariates (i.e.

using only Yi and Xi ) has been studied in great detail in recent years—see, among others, Hahn

et al. (2001), Porter (2003), Imbens and Kalyanaraman (2012), Calonico et al. (2014), Gelman

and Imbens (2014), Armstrong and Kolesar (2015), Kamat (2015), Calonico et al. (2016a), and

references therein. These papers give asymptotic MSE expansions, MSE-optimal point estimators,

data-driven bandwidth selection methods, asymptotically valid inference procedures based on bias-

correction and non-standard distributional approximations, and even valid Edgeworth expansions

to inform empirical practice.

We study the asymptotic properties of the covariate-adjusted RD estimator, τ̃ (h), building on

this literature. We assume that µZ+ = µZ− in order to maintain the same standard RD treat-

ment effect of interest (Lemma 1). We present new MSE expansions, several data-driven optimal

bandwidth selectors, valid distributional approximations based on bias-correction techniques, and

consistent standard errors for τ̃ (h). Analogous results for other RD designs are briefly discussed

in Section 5. A full treatment of all cases, including several other extensions, is given in the sup-

plemental appendix. All our methods are implemented in companion general purpose R and Stata

software packages (Calonico et al., 2016b).

To characterize the asymptotic properties of the covariate-adjusted RD estimator τ̃ (h), we rely

on the following representation (valid for each n):

τ̃ (h) = τ̂ (h) − τ̂ Z (h)0 γ̃ Y,p (h),

where τ̂ (h) and γ̃ Y,p (h) were defined above (see (1) and (2)), and τ̂ Z (h) is a d-dimensional vector

containing the standard RD treatment effect estimator for each covariate. In other words, each

element of τ̂ Z (h) is constructed using the corresponding covariate as outcome variable in (1). In the

appendix and supplemental appendix we give exact details. Using this partial-out representation,

13
it follows that    
 τ̂ (h) − τ  τ̂ (h) − τ 
τ̃ (h) − τ = s(h)0  0
=s   {1 + oP (1)}
τ̂ Z (h) τ̂ Z (h)

where s(h) = (1, γ̃ Y,p (h)0 )0 and s = (1, γ 0Y )0 , and because s(h) →P s using the results underlying

Lemma 1 (and, later, we will also use τ Z = 0).

Therefore, the asymptotic analysis proceeds by studying the (joint) large-sample properties of

the vector τ̂ S (h) := (τ̂ (h), τ̂ Z (h)0 )0 and then taking the linear combination s(h) or s, as appropriate.

Note that τ̂ S (h) →P τ S := (τ, τ 0Z )0 under the conditions in Lemma 1. In fact, most of the results

presented in this paper do not require the assumption τ Z = 0, though without this assumption

the parameter of interest changes, undermining the practical usefulness of the results. Finally, we

re-emphasize that our results do not impose any restrictions on the distribution of Yi (t)|Xi , Zi (t),

t ∈ {0, 1}, and impose instead minimal restrictions on the distributions of Yi (t), Zi (t)|Xi , t ∈ {0, 1}

(see Assumption 2). For example, we do not place any parametric or semiparametric assumption

on E[Yi (0)|Xi , Zi (0)] or E[Yi (1)|Xi , Zi (1)].

4.1 MSE Expansion and Optimal Point Estimators

We first establish a valid asymptotic MSE expansion for the covariate-adjusted RD estimator. This

expansion will aid in developing optimal bandwidth choices and MSE-optimal point estimators.

Furthermore, the bias expressions will be instrumental to develop valid inference procedures based

on robust bias-correction techniques. Let X = [X1 , X2 , · · · , Xn ]0 , and define

MSE[τ̃ (h)] = E[(s0 τ̂ S (h) − s0 τ S )2 |X] = (Bias[τ̃ (h)])2 + Var[τ̃ (h)],

where Bias[τ̃ (h)] := E[s0 τ̂ S (h) − s0 τ S |X] and Var[τ̃ (h)] := V[s0 τ̂ S (h)|X].

Theorem 1 (MSE Expansion). Let Assumptions 1 and 2 hold. If nh → ∞ and h → 0, then

1
MSE[τ̃ (h)] = h2(1+p) Bτ̃ (h)2 {1 + oP (1)} + Vτ̃ (h),
nh

where the precise expressions for all bias and variance terms are given in the appendix.

The bias and variance expressions in Theorem 1 are different from those available in the litera-

ture (Imbens and Kalyanaraman, 2012; Calonico et al., 2014) due to the presence of the additional

14
covariates Zi . As a consequence, MSE-optimal bandwidth selection and MSE-optimal point es-

timators in RD designs using covariates are different from their counterparts without covariates.

Bias-correction techniques and standard errors constructions will also be different, as discussed

below.

The leading bias and variance formulas in Theorem 1 are derived in pre-asymptotic form. For

the bias, the random term Bτ̃ (h) gives a pre-asymptotic stochastic approximation to the conditional

bias of the linearized estimator (hence the presence of the oP (1) term), whereas the variance term

Vτ̃ (h) is simply obtained by a conditional on X calculation for the linearized estimator. Calonico

et al. (2016a) prove, using valid Edgeworth expansions, that employing pre-asymptotic approxima-

tions when conducting asymptotic inference in nonparametrics can lead to superior performance.

Furthermore, fewer unknown features of the data generating process must be characterized and

estimated.

The main constants in Theorem 1 have a familiar form: the bias and variance are, respectively,

Bτ̃ (h) = Bτ̃ + (h) − Bτ̃ − (h) and Vτ̃ (h) = Vτ̃ − (h) + Vτ̃ + (h), where each component stems from

estimating the unknown regression function on one side of the cutoff. Here, the bias is entirely

due to estimating the unknown functions µY − (·) and µZ− (·) for the control group and µY + (·) and

µZ+ (·) for the treatment group. When the additional covariates are not included, these constants

reduce exactly to those already available in the literature. In the appendix, we also give the limiting

version of the bias and variance constants; that is, we characterize the fixed, real scalars Bτ̃ and Vτ̃

that satisfy (Bτ̃ (h), Vτ̃ (h)) →P (Bτ̃ , Vτ̃ ).

Assuming that Bτ̃ 6= 0, the MSE-optimal bandwidth choice for the covariate-adjusted RD

estimator τ̃ (h) is:

  1
1 Vτ̃ /n 3+2p
hτ̃ = .
2(1 + p) Bτ̃2

This choice can be used to construct a consistent and MSE-optimal covariate-adjusted sharp RD

point estimator: τ̃ (hτ̃ ) →P τ , provided that τ Z = 0. Note that d = dim(Zi ) does not impact the

rate of decay because we do not employ any nonparametric smoothing methods on the additional

covariates.

We address the issue of data-driven implementations of the new optimal bandwidth choices

further below, after discussing valid large sample inference.

Remark 2 (Asymptotic Efficiency). In addition to finite-sample efficiency considerations, which

15
are well known from the literature on linear least-squares, we can give a precise characterization of

the asymptotic efficiency gains from introducing additional covariates in the RD estimation. Using

the explicit formulas given in the appendix, it is easy to show that the asymptotic variance of the

covariate-adjusted estimator τ̃ (denoted by Vτ̃ ) is equal to the asymptotic variance of the standard

RD estimator τ̂ (denoted by Vτ ) plus a linear combination (based on γ Y ) of terms involving the

Cov[Yi (t), Zi (t)|Xi = x̄] and V[Zi (t)|Xi = x̄]. Therefore, τ̃ can be asymptotically more efficient

than τ̂ when the term 2Cov[Yi (t), Zi (t)|Xi = x̄]0 γ Y is negative and larger in absolute value than

γ 0Y V[Zi (t)|Xi = x̄]γ Y . 

Remark 3 (MSE-optimal Point Estimation). The results above also show that τ̃ (hτ̃ ) can be a

better point estimator in a MSE sense that its MSE-optimal counterpart without covariates, τ̂ (hτ̂ ),

where hτ̂ denotes the MSE-optimal bandwidth choice for the standard RD estimator τ̂ (Imbens and

Kalyanaraman, 2012; Calonico et al., 2014). Using the explicit formulas given in the appendix, it

is easy to give conditions such that MSE[τ̃ (hτ̃ )] < MSE[τ̂ (hτ̂ )] (both have the same rate of decay),

although this is not the main goal of our paper. We still recommend that τ̂ (hτ̂ ) be the benchmark

RD point estimator, and thus that researchers incorporate covariates to increase precision relative

to it, rather than to replace it. 

Remark 4 (Other Optimal Bandwidth Choices). In the supplemental appendix we discuss other

MSE-optimal bandwidth selectors based on the results underlying Theorem 1, which are specifically

tailored to one-sided and two-sided estimation problems in RD designs. Specifically, we present: (i)

separate MSE optimizations on either side of the cutoff, (ii) the MSE for the sum rather than the

difference of the one-sided estimators, and (iii) several regularized versions of the plug-in bandwidth

selectors. In all cases, the decay rate of these bandwidths matches the MSE-optimal choice, but

the exact leading constants differ, implying that any of these could be used to construct sharp

RD point estimators with an MSE-optimal convergence rate. These choices may be more stable in

finite samples or more robust to situations where the smoothing bias may be small. 

4.2 Asymptotic Distribution and Valid Inference

To develop valid asymptotic distributional approximations and inference procedures we employ

robust nonparametric bias-correction. It is by now well understood that inference based on large-

sample distribution theory using MSE-optimal bandwidths will suffer from a first order bias, leading

16
to invalid hypothesis testing and confidence intervals because of misspecification errors near the

cutoff. This local smoothing bias involves the bias term in Theorem 1, Bτ̃ (h), which can be esti-

mated and removed. Following recent ideas and results in Calonico et al. (2014, 2016a), we propose

robust bias-corrected distributional approximations for the covariate-adjusted RD estimator, and

discuss the associated inference procedures based on such approximations.

The bias terms of Theorem 1 are known up to a higher-order derivative of the unknown regres-

sion functions, µY − (·), µZ− (·), µY + (·), and µZ+ (·), all capturing the misspecification error intro-

duced by the local polynomial approximation. These objects can be estimated nonparametrically—

the complete details are available in the appendix (we replace s by s(h) for implementation). At

present, let us simply take as given the bias estimator B̃τ̃ (b) based on local polynomial techniques,

which depends on a preliminary bandwidth b → 0, possibly different from h, and on a polynomial

order q > p. Then, the bias-corrected covariate-adjusted sharp RD estimator is

τ̃ bc (h, b) = τ̃ (h) − h1+p B̃τ̃ (b) (3)

A particularly empirically useful choice is b = h, which is both allowed for by our asymptotic theory

and has some optimal properties (Calonico et al., 2016a). This bias correction approach is standard

in the literature (e.g., Fan and Gijbels, 1996, Section 4.4), and captures nicely “flexible” regression

adjustments to account for misspecification in finite samples (Calonico et al., 2014, Remark 7).

The key idea behind the robust bias-corrected distributional approximation is to employ an

estimator of the variability of τ̃ bc (h, b) for Studentization purposes, rather than an estimator of the

variability of τ̃ (h) only. Thus, the final missing ingredient before we can state our asymptotic Gaus-

sianity result is characterizing the (conditional) variance of the bias-corrected covariate-adjusted

RD estimator. Its fixed-n variability is easily characterized due to its approximate (conditional)

linearity, and is given by

Vτ̃bc (h, b) = [s0 ⊗ e00 Pbc 0 0 bc 0 0 0 bc 0 0 bc 0

−,p (h, b)]ΣS− [s ⊗ e0 P−,p (h, b)] + [s ⊗ e0 P+,p (h, b)]ΣS+ [s ⊗ e0 P+,p (h, b)] ,

where the (1 + p) × n matrices Pbc bc

−,p (h, b) and P+,p (h, b) can be computed directly from the data,

and the n(1 + d) × n(1 + d) matrices of variances and covariances ΣS− and ΣS+ are unknown.

Specifically, ΣS− = V[S(0)|X] and ΣS+ = V[S(1)|X], where S(0) = (Y(0)0 , vec(Z(0))0 )0 and S(1) =

17
(Y(1)0 , vec(Z(1))0 )0 , with Y(t) = [Y1 (t), Y2 (t), · · · , Yn (t)]0 and Z(t) = [Z1 (t), Z2 (t), · · · , Zn (t)]0 for

t ∈ {0, 1}. The appendix collects tedious details and specific formulas, including the exact form of

Pbc bc
−,p (h, b) and P+,p (h, b).

The (infeasible) variance formula Vτ̃bc (h, b) differs from that presented in Theorem 1, Vτ̃ (h),

because it also accounts for the leading additional variability injected by the bias estimation,

h1+p B̃τ̃ (b). By virtue of the variance formula being computed both conditionally and pre-asymptotic,

up to the linear combination term s, it involves only one unknown feature, ΣS− and ΣS+ , which

must be estimated, thereby simplifying implementation considerably.

To operationalize the variance formula we replace unknown quantities by plug-in estimators

thereof. The estimators need to account for the specific data structure at hand, such as het-

eroskedasticity and/or clustering. In particular, we discuss two type of plug-in variance estimators,

one based on a nearest neighbor (NN) approach and the other based on a plug-in residuals (PR)

approach, covering both conditional heteroskedasticity and clustered data. We defer the notation-

ally cumbersome details to the supplemental appendix, and instead we provide here only a brief

summary of the main ideas and results. The unknown matrices ΣS− and ΣS+ contain, under

conditional heteroskedasticity, diagonal submatrices with representative elements, respectively,

σY Zk −,i = Cov[Yi (0), Zki (0)|Xi ], σY Zk +,i = Cov[Yi (1), Zki (1)|Xi ]

for k = 1, 2, · · · , d. The feasible variance estimators are then constructed by replacing these un-

known objects with unbiased estimators thereof, as follows.

• NN Variance Estimation. Employing ideas in Muller and Stadtmuller (1987) and Abadie

and Imbens (2006, 2008), we replace σY Zk −,i and σY Zk +,i by, respectively,

  
J J
J  1 1
σ̂Y Zk −,i (J) = 1(Xi < x̄)
X X
Yi − Y`−,j (i)  Zki − Zk`−,j (i)  ,
J +1 J J
j=1 j=1

  
J J
J  1X 1
σ̂Y Zk +,i (J) = 1(Xi ≥ x̄)
X
Yi − Y`+,j (i)  Zki − Zk`+,j (i)  ,
J +1 J J
j=1 j=1

for k = 1, 2, · · · , d, and where `−,j (i) is the index of the j-th closest unit to unit i among

{Xi : Xi < x̄} and `+,j (i) is the index of the j-th closest unit to unit i among {Xi : Xi ≥ x̄},

18
 and J denotes a (fixed) the number of neighbors chosen. Replacing the non-zero entries of

ΣS− and ΣS+ in this fashion (which depend on the sampling structured assumed), and s by

s(h), we obtain the NN variance estimator of Vτ̃bc (h, b), denoted by V̌τ̃bc (h, b).

• PR Variance Estimation. This method applies ideas from least-squares methods; see Long

and Ervin (2000), MacKinnon (2012), and Cameron and Miller (2015) for review on variance

estimation in this context. We replace σY Zk −,i and σY Zk +,i by, respectively,

  
σ̂Y Zk −,q,i (h) = 1(Xi < x̄)ω−,i 0 0
Yi − rq (Xi − x̄) β̂ Y −,q (h) Zki − rq (Xi − x̄) β̂ Zk −,q (h) ,

  
σ̂Y Zk +,q,i (h) = 1(Xi ≥ x̄)ω+,i Yi − rq (Xi − x̄)0 β̂ Y +,q (h) Zki − rq (Xi − x̄)0 β̂ Zk +,q (h) ,

for k = 1, 2, · · · , d, and where β̂ V −,q (h) and β̂ V +,q (h) denote the q-th order local polynomial

fits using as outcome variable V ∈ {Y, Z1 , Z2 , · · · , Zd }, with bandwidth h, as described in

(1), and ω−,i and ω+,i denote finite-sample adjustments used to construct the HCk variance

estimators. See the supplement for details. Replacing the non-zero entries of ΣS− and ΣS+

in this fashion (which depend on the sampling structured assumed), and s by s(h), we obtain

the PR variance estimator of Vτ̃bc (h, b), denoted by V̂τ̃bc (h, b).

Putting together all the pieces, we obtain the following distributional approximation result.

Theorem 2 (Asymptotic Normality). Let the conditions of Theorem 1 hold, and τ Z = 0. If

√
nhh1+p min{h, bq−p } → 0 and lim (h/b) < ∞, then

τ̃ bc (h, b) − τ
T̃τ̃ = q →d N (0, 1).
1 bc
nh V τ̃ (h, b)

Furthermore, V̌τ̃bc (h, b)/Vτ̃bc (h, b) →P 1 and V̂τ̃bc (h, b)/Vτ̃bc (h, b) →P 1.

Theorem 2 provides valid inference in sharp RD designs using covariates. To our knowledge,

this is the first such result available in the literature for the covariate-adjusted RD estimation.

Extensions of this result to all other popular RD designs are discussed in Section 5 and the sup-

plemental appendix. Once bandwidths are chosen, asymptotically valid inference procedures are

easily constructed. For example, an approximately 95% robust bias-corrected covariate-adjusted

confidence interval for the RD treatment effect τ , using h = b and NN variance estimation, is given

19
by
 
1.96 1.96
q q
τ̃ bc (h, h) − √ · V̌τ̃bc (h, h) , τ̃ bc (h, h) + √ · V̌τ̃bc (h, h) .
nh nh

We explore the performance in finite samples of our proposed methods in Section 6.

Remark 5 (Clustered Data). Theorem 2 can also be established under clustered sampling. All

derivations and results remain valid, but the variance formulas will depend on the particular form of

clustering. In this case, asymptotics are conducted under the standard assumptions: (i) each unit i

belongs to exactly one of G clusters, and (ii) G → ∞ and Gh → ∞. See Cameron and Miller (2015)

for a review of cluster-robust inference, and Bartalotti and Brummet (2016) for a discussion in the

context of MSE-optimal bandwidth selection for sharp RD designs. This extension is conceptually

straightforward but notationally cumbersome, and is deferred to the supplement. Our companion

software in R and Stata also includes optional cluster-robust (i) bandwidth selection, (ii) MSE-

optimal point estimation, and (iii) robust bias-corrected inference. 

4.3 Data-driven Bandwidth Selection

We now discuss bandwidth selection briefly, leaving full details to the supplement (where we also

discuss several alternative bandwidth selectors as in Remark 4). Here we focus exclusively on two

main, distinct approaches: (i) the MSE-optimal choice derived previously (hη ), which can be used

to construct MSE-optimal RD point estimators, and (ii) a novel bandwidth selection approach

constructed to obtain the fastest decay of the coverage error rate (CER) of robust bias-corrected

confidence intervals (denoted hCER,η ), motivated by the valid Edgeworth expansions for RD inference.

One of the strengths of Theorem 2 is that the distributional approximation is valid under a

large set of tuning parameter choices (strictly more than would be possible without bias correction),

which in particular includes the MSE-optimal choice (which is not valid for standard procedures).

Assuming the bias is not zero, Theorem 1 can be used in the familiar way to construct feasible

MSE-optimal bandwidth choices. For bandwidths b → 0 and v → 0, these will be given by

" # 1
3+2p
1 Ṽτ̃ (v)/n
h̃τ̃ =
2(1 + p) B̃τ̃ (b)2

where the exact form of the bias estimator, B̃τ̃ (b), and variance estimator, Ṽτ̃ (v), are given in the

supplemental appendix. Heuristically, these estimators are formed as plug-in versions of the pre-

20
asymptotic formulas obtained in Theorem 1, following the previous discussion leading to Theorem 2.

In the supplemental appendix, we also show that these feasible versions of the optimal bandwidths

are consistent for their infeasible analogues; i.e., h̃τ̃ /hτ̃ →P 1.

A particularly attractive alternative to MSE-optimal bandwidth selection is to develop cov-

erage error optimal bandwidth choices. Following the results in Calonico et al. (2016a), we also

recommend the following plug-in bandwidth selector

p
− (3+p)(3+2p)
h̃CER,τ̃ = n · h̃τ̃ .

This bandwidth choice minimizes the coverage error rate for confidence intervals based on Theorem

2, and can be preferred for inference purposes.

5 Other RD designs

We extend our main results to cover other popular RD designs, including fuzzy, kink, and fuzzy

kink RD. Here we give a short overview of the main ideas, deferring all details to the supple-

mental appendix. There are two wrinkles to the standard sharp RD design discussed so far that

must be accounted for: ratios of estimands/estimators for fuzzy designs and derivatives in esti-

mands/estimators for kink designs.

5.1 Fuzzy RD Designs

The distinctive feature of fuzzy RD designs is that treatment compliance is imperfect. This implies

that Ti = Ti (0) · 1(Xi < x̄) + Ti (1) · 1(Xi ≥ x̄), that is, the treatment status Ti of each unit

i = 1, 2, · · · , n is no longer a deterministic function of the running variable Xi , but P[Ti = 1|Xi = x]

still changes discontinuously at the RD threshold level x̄. Here, Ti (0) and Ti (1) denote the two

potential treatment status for each unit i when, respectively, Xi < x̄ (not offered treatment) and

Xi ≥ x̄ (offered treatment).

To analyze the case of fuzzy RD designs, we first recycle notation for potential outcomes and

covariates as follows:

Yi (t) := Yi (0) · (1 − Ti (t)) + Yi (1) · Ti (t)

21
 Zi (t) := Zi (0) · (1 − Ti (t)) + Zi (1) · Ti (t)

for t = 0, 1. That is, in this setting, potential outcomes and covariates are interpreted as their

“reduced form” (or intention-to-treat) counterparts. Giving causal interpretation to covariate-

adjusted instrumental variable type estimators is delicate; see e.g. Abadie (2003) for more discus-

sion. Nonetheless, the above re-definitions enable us to use the same notation, assumptions, and

results, already given for the sharp RD design, taking the population target estimands as simply

the probability limits of the RD estimators.

The following assumption complements Assumption 2, now concerning the (potential) treatment

variables.

Assumption 3 (Fuzzy RD Designs). For % ≥ p + 2 and all x ∈ [xl , xu ], where xl , xu ∈ R such that

xl < x̄ < xu :

(a) µT − (x) := E[Ti (0)|Xi = x], µT + (x) := E[Ti (1)|Xi = x], E[Zi (0)Ti (0)|Xi = x], and E[Zi (1)Ti (1)|Xi =

x] are % times continuously differentiable.

(b) V[Fi (t)|Xi = x], with Fi (t) := (Yi (t), Ti (t), Zi (t)0 )0 , t ∈ {0, 1}, are continuously differentiable

and invertible.

(c) E[|Ti (t)|4 |Xi = x], t ∈ {0, 1}, are continuous.

(d) µT − (x) 6= µT + (x).

The standard fuzzy RD estimand is

τY
ς= , τY = µY + − µY − , τT = µT + − µT − ,
τT

where recall that we continue to omit the evaluation point x = x̄, and we have redefined the potential

outcomes and additional covariates to incorporate imperfect treatment compliance. Furthermore,

now τ has a subindex highlighting the outcome variable being considered (Y or T ), and hence

τ = τY by definition. See Hahn, Todd and van der Klaauw (2001) and Imbens and Lemieux (2008)

for further discussion on identification and interpretation of this estimand.

The standard estimator of ς, without covariate adjustment, is

τ̂Y (h)
ςˆ(h) = , τ̂V (h) = e00 β̂ V +,p (h) − e00 β̂ V −,p (h),
τ̂T (h)

22
with V ∈ {Y, T }, according to (1). Similarly, the covariate-adjusted fuzzy RD estimator is

τ̃Y (h)
ς˜(h) = , τ̃V (h) = e00 β̃ V +,p (h) − e00 β̂ V −,p (h),
τ̃T (h)

with V ∈ {Y, T }, according to (2). Our notation makes clear that the fuzzy RD estimators, with

or without additional covariates, are simply the ratio of two sharp RD estimators, with or without

covariates.

The properties of the standard fuzzy RD estimator ςˆ(h) were studied in great detail before, while

the covariate-adjusted fuzzy RD estimator ς˜(h) has not been studied in the literature before. With

these preliminaries, we can give the analogue of Lemma 1 for fuzzy RD designs using covariates.

Lemma 3 (Fuzzy RD with Covariates). Let Assumptions 1, 2, and 3 hold. If nh → ∞ and h → 0,

then
τY − [µZ+ − µZ− ]0 γ Y
ς˜(h) →P ,
τT − [µZ+ − µZ− ]0 γ T

where γ V = (σ 2Z− + σ 2Z+ )−1 E[(Zi (0) − µZ− (Xi ))Vi (0) + (Zi (1) − µZ+ (Xi ))Vi (1)|Xi = x̄] with

V ∈ {Y, T }.

Under the same conditions, when no additional covariates are included, it is well known that

ςˆ(h) →P ς. Thus, this lemma clearly shows that both probability limits will coincide under the same

sufficient condition as in the sharp RD design: µZ− = µZ+ . Therefore, at least asymptotically, a

(causal) interpretation for the probability limit of the covariate-adjusted fuzzy RD estimator can

be deduced from the corresponding (causal) interpretation for the probability limit of the standard

fuzzy RD estimator, whenever the condition µZ− = µZ+ holds.

Since the fuzzy RD estimators are constructed as a ratio of two sharp RD estimators, their

asymptotic properties can be characterized by studying the asymptotic properties of the corre-

sponding sharp RD estimators, which have already been analyzed in previous sections. Specifically,

the asymptotic properties of covariate-adjusted fuzzy RD estimator ζ̃(h) can be characterized by

employing the following linear approximation:

ς˜(h) − ς = fς˜0 (τ̃ (h) − τ ) + ς˜,

23
with      
1
τT  τ̃Y (h)   τY 
fς˜ =  , τ̃ (h) =  , τ = ,
 
− ττY2 τ̃T (h) τT
T

and where the term ς˜ is a quadratic (high-order) error. Therefore, it is sufficient to study the

asymptotic properties of the bivariate vector τ̃ (h) of covariate-adjusted sharp RD estimators, pro-

vided that ς˜ is asymptotically negligible relative to the linear approximation, which is proven in

the supplement. As before, while not necessary for most of our results, we continue to assume

that µZ− = µZ+ so the standard RD estimand is recovered by the covariate-adjusted fuzzy RD

estimator.

Employing the linear approximation and parallel results as those discussed above for the sharp

RD design (now also using Ti as outcome variable as appropriate), it is conceptually straightforward

to conduct inference in fuzzy RD designs with covariates. All the same results outlined in the previ-

ous section are established for this case: in the supplemental appendix we present MSE expansions,

MSE-optimal bandwidth, MSE-optimal point estimators, consistent bandwidth selectors, robust

bias-corrected distribution theory and consistent standard errors under either heteroskedasticity

or clustering, for the covariate-robust fuzzy RD estimator ς˜(h). We do not attempt to present

these results here because they are notationally cumbersome, with little new conceptual insight.

Nevertheless, all these results are implemented in the general purpose software packages for R and

Stata described in Calonico et al. (2016b).

5.2 Kink RD Designs

Our final extension concerns the so-called kink RD designs. See Card, Lee, Pei and Weber (2015)

for a discussion on identification and Calonico et al. (2014) for a discussion on estimation and

inference, both covering sharp and fuzzy settings without additional covariates. Dong and Lewbel

(2015) also study derivative estimation in RD designs, without additional covariates. We briefly

outline identification and consistency results when additional covariates are included in kink RD

estimation (i.e., derivative estimation at the cutoff), but relegate all other inference results to the

supplemental appendix.

To describe the estimands of interest in this context, let g (s) (x) = ∂ s g(x)/∂xs for any sufficiently

24
smooth function g(·). The standard sharp kink RD parameter is (proportional to)

(1) (1)
τY,1 = µY + − µY − ,

while the fuzzy kink RD parameter is

τY,1
ς1 =
τT,1
(1) (1)
where τT,1 = µT + − µT − . In the absence of additional covariates in the RD estimation, these RD

treatment effects are estimated by using the local polynomial plug-in estimators:

τ̂Y,1 (h)
τ̂Y,1 (h) = e01 β̂ Y +,p (h) − e01 β̂ Y −,p (h) and ςˆ1 (h) = ,
τ̂T,1 (h)

where e1 denote the conformable 2nd unit vector (i.e., e1 = (0, 1, 0, 0, · · · , 0)0 ). Therefore, the

covariate-adjusted kink RD estimators in sharp and fuzzy settings are

τ̃Y,1 (h) = e01 β̃ Y +,p (h) − e01 β̃ Y −,p (h)

and
τ̃Y,1 (h)
ς˜1 (h) = , τ̃V,1 (h) = e01 β̃ V +,p (h) − e01 β̃ V −,p (h), V ∈ {Y, T },
τ̃T,1 (h)

respectively. The following lemma gives our main identification and consistency results.

Lemma 4 (Kink RD with Covariates). Let Assumptions 1, 2, and 3 hold. If nh → ∞ and h → 0,

then
(1) (1)
τ̃Y,1 (h) →P τY,1 − [µZ+ − µZ− ]0 γ Y

and
(1) (1)
τY,1 − [µZ+ − µZ− ]0 γ Y
ς˜1 (h) →P (1) (1)
,
τT,1 − [µZ+ − µZ− ]0 γ T
(1) (1) (1) (1)
where γ Y and γ T are defined in Lemma 3, and recall that µZ− = µZ− (x̄) and µZ+ = µZ+ (x̄) with
(1) (1)
µZ− (x) = ∂µZ− (x)/∂x and µZ+ (x) = ∂µZ+ (x)/∂x.

As before, in this setting it is well known that τ̂Y,1 (h) →P τY,1 (sharp kink RD) and ςˆ1 (h) →P ς1

(fuzzy kink RD), formalizing once again that the estimand when covariates are included is in general

different from the standard kink RD estimand without covariates. In this case, a sufficient condition

25
 (1) (1)
for the estimands with and without covariates to agree is µZ+ = µZ− for both sharp and fuzzy

kink RD designs.

While the above results are in qualitative agreement with the sharp and fuzzy RD cases, and

therefore most conclusions transfer directly to kink RD designs, there is one interesting difference

concerning the sufficient conditions guaranteeing that both estimands coincide: a sufficient con-
(1) (1)
dition now requires µZ+ = µZ− . This requirement is not related to the typical falsification test

conducted in empirical work, that is, µZ+ = µZ− , but rather a different feature of the conditional

distributions of the additional covariates given the score—the first derivative of the regression func-

tion at the cutoff. Therefore, this finding suggests a new falsification test for empirical work in kink

RD designs: testing for a zero sharp kink RD treatment effect on “pre-intervention” covariates.

For example, this can be done using standard sharp kink RD treatment effect results, using each

covariate as outcome variable.

As before, inference results follow the same logic already discussed. Complete details are given

in the supplement and fully implemented in the R and Stata software described by Calonico et al.

(2016b).

6 Numerical Results

We now illustrate our methods empirically and present an extensive simulation study conducted to

assess the finite sample properties of the covariate-adjusted RD estimator and the associated large

sample inference procedures developed in this paper. To conserve space, we only discuss the main

findings and relegate details to the supplemental appendix.

6.1 Empirical illustrations

We first illustrate our methods in two empirical applications. First, we re-analyze the effect of

Head Start assistance on child mortality in the U.S., which was first studied by Ludwig and Miller

(2007). In this application, the unit of observation is the U.S. county, the treatment is receiving

technical assistance to apply for Head Start funds, and the running variable is the county-level

poverty index constructed in 1965. The RD design arises because the treatment was given only to

counties whose poverty index was x̄ = 59.1984 or above, a cutoff that was chosen to ensure that the

300th poorest counties received the treatment. The outcome of interest is the child mortality rate

26
(for children of age five to nine) due to causes affected by Head Start’s health services component.

Next, we revisit the effect of school improvements on student language test scores in Chile,

first studied by Chay, McEwan and Urquiola (2005). The unit of observation is the school, the

treatment is receiving the school improvement program P-900, which was assigned in 1990 based

on an index (the running variable) constructed from previous school-level test scores. The outcome

we study is school-level language score gain between 1988 and 1992.

We compare the standard RD estimator to the covariate-adjusted RD estimator employing

heteroskedasticity-robust nearest neighbor variance estimation for both applications. In the Head

Start application, the additional regressors Zi are nine county-level covariates from the pre-intervention

1960 U.S. Census: total population, percentage of black and urban population, and levels and per-

centages of population in three age groups (children age 3 to 5, children age 14 to 17, and adults

older than 25). In the education application, the additional covariates are seven binary variables in-

dicating the school’s region group (Chile’s 13 administrative regions were divided into seven groups,

with schools in each region group facing a different cutoff value).

Table 1 presents the main results. In each application’s panel in that table, the first row reports

the RD local-linear (p = 1) point estimate using the corresponding MSE-optimal bandwidth h

(depending on the column). The next three rows report 95% robust bias-corrected confidence

intervals, the percentage length change of the covariate-adjusted confidence interval relative to the

unadjusted one, and the p-value associated with a hypothesis of zero RD treatment effect. These

three rows appear twice, first when h for the RD point estimator and b for the bias estimator are

chosen separately (in this case, ρ = h/b is unrestricted), and then when b = h (in this case, ρ = 1).

Finally, the last two rows in each application’s panel report, respectively, the two bandwidths and

the number of observations to the left and to the right of the cutoff with Xi ∈ [x̄ − h, x̄ + h].

The columns in Table 1 correspond to different RD approaches. The first two columns employ

the MSE-optimal bandwidth without covariates—the first column reports inference results without

covariates while the second column presents covariate-adjusted inference. Thus, the second column

is intended to mimic a common practice among practitioners, who sometimes estimate the MSE-

optimal bandwidth without covariates and then include covariates in the estimation and inference

using the same observations (i.e., keeping the bandwidth choice fixed). It follows that, in the second

column, the h and b bandwidths used are the MSE-optimal bandwidths without covariates, and

27
therefore the point estimator in this second column is no longer MSE-optimal, since the optimal

bandwidths are not used; however, the confidence intervals and p-values are still valid because

the optimal bandwidths with and without covariates have the same rate of decay. The third

column reports covariate-adjusted RD estimation using the asymptotic approximations derived in

the previous sections; in this column, both bandwidths are chosen according to the MSE-optimal

formulas with covariates. In all cases, we use triangular kernel weights and nearest neighbor residual

estimates. Employing other kernels or variance estimators give very similar empirical results.

Our empirical findings are quite consistent across applications: employing covariate-adjusted

RD inference leads to precision improvements while the point estimators remain stable. In the

Head Start application, the point estimator ranges from −2.41 to −2.51, an effect that is statis-

tically different from zero at 5% significance level in all cases. As should be expected when the

additional covariates are truly pre-determined, including covariates does not substantially alter the

point estimates. (We also implemented “placebo tests” on the additional covariates and found,

as expected, no statistical evidence of RD treatment effects.) Including additional covariates in

this application can lead to sizable efficiency gains: for example, when both h and b are estimated

(ρ = h/b unrestricted), adding covariates within the MSE-optimal h without covariates (h = 6.81)

results in a 8.25% reduction in the length of the 95% confidence interval (column 2), as this confi-

dence interval shrinks from (−5.49, −0.10) to (−5.37, −0.45). The length of the confidence interval

is even shorter when both bandwidths are chosen optimally using covariates (column 3), one of the

novel results in this paper, with a length reduction of roughly 10%.

In the case of the education data, we also find that including additional covariates does not

affect the point estimators, while providing some efficiency improvements. In this case, the point

estimates range from 3.45 to 3.49, and the confidence interval length shrinks approximately 3% to

5% depending on the case considered.

Our empirical results suggest that including pre-intervention covariates can be empirically useful

in real RD applications, thereby illustrating the usefulness of the new methods developed in this

paper.

28
6.2 Simulation Evidence

We now illustrate our methods using simulated data. We consider four data generating processes

constructed using the data of Lee (2008): all parameters are obtained from the real data unless

explicitly noted otherwise. This model has been used extensively before, see Imbens and Kalya-

naraman (2012) and Calonico et al. (2014, 2016a), among many others. The additional covariate

included is previous democratic vote share, and the four models are distinguished by the importance

of this covariate: (i) in Model 1, the covariate is irrelevant; (ii) in Model 2 it enters the conditional

expectation of the potential outcomes E[Yi (t)|Xi = x, Zi (t)], t ∈ {0, 1}, according to the real data;

(iii) Model 3 takes Model 2 but sets the residual correlation between the outcome and covariate

to zero; (iv) Model 4 takes Model 2 but doubles the residual correlation between the outcome and

covariate equations. Note that Models 3 and 4 do not imply Cov[Yi (t), Zi (t)|Xi = x] = 0, t ∈ {0, 1}.

The constructions allowed E[Yi (t)|Xi = x, Zi (t)] to have different coefficients on each side of the

cutoff, while the conditional expectation of the potential covariates E[Zi (t)|Xi = x], t ∈ {0, 1},

were constructed assuming they are continuous at the cutoff (but still with different coefficients on

either side otherwise). Therefore, our covariate-adjusted RD estimator will be “misspecified” when

viewed as a local weighted least-square fit. The exact details of our Monte Carlo experiment are

given in the supplemental appendix to conserve space.

We use a sample of size n = 1, 000 and consider 5, 000 replications. We compare the standard

RD estimator (τ̂ ) and the covariate-adjusted RD estimator (τ̃ ), with both infeasible and data-driven

MSE-optimal and CER-optimal bandwidth choices. To analyze the performance of our inference

procedures, we report average bias of the point estimators and average coverage rate and interval

length of nominal 95% confidence intervals. In addition, we also explore the performance of our

data-driven bandwidth selectors by reporting some of their main statistical features, such as mean,

median, standard deviation, across the 5, 000 replications. We report only one table that presents

estimates using triangular kernel and nearest neighbor (NN) heteroskedasticity-robust variance

estimators; complete details and results are in the supplement.

The numerical results are given in Tables 2 and 3. All findings are highly consistent with our

large sample theory. Table 2 shows that including covariates can improve both MSE and interval

length, sometimes dramatically, and moreover, the gains are in line with our theory: the gains are

largest in Model 4 with the amplified residual correlation and least in Model 3 when that channel

29
is shut down. The results for Model 1 show that including an irrelevant covariate hardly changes

empirical results and conclusions. Finally, Table 3 shows that the data-driven bandwidth selectors

also work reasonably well.

7 Conclusion

We provided a formal framework for identification, estimation, and inference in RD designs when

covariates are included in the estimation. We augmented the standard local polynomial estimator

with covariates entering in an additive-separable, linear-in-parameters way. We showed that under

minimal additional smoothness assumptions, the resulting covariate-adjusted RD estimator remains

consistent for the standard RD treatment effect if the covariate adjustment is restricted to be

equivalent above and below the cutoff. Furthermore, this estimator can achieve substantial efficiency

gains relative to the unadjusted RD estimator. We also showed that relaxing the latter restriction

with the inclusion of treatment-covariate interactions leads to a point estimator that is not generally

consistent for the standard RD parameter of interest.

We also provided new MSE expansions, several optimal bandwidth choices and optimal point es-

timators, robust nonparametric inference procedures based on bias-correction, and heteroskedasticity-

consistent and cluster-robust standard errors. All these results were obtained for sharp, fuzzy, and

kink RD designs. Finally, we illustrated the practical implications of our results using two empiri-

cal applications and simulated data, and showed that including pre-intervention covariates in RD

designs can lead to useful improvements in precision. All the results presented in this paper are

implemented in companion general purpose R and Stata software packages.

8 Appendix: Sharp RD Design Main Formulas

We give a very succinct account of the main expressions for sharp RD designs, which were omitted

in the main paper to avoid overwhelming notation. A detailed treatment of this and all other RD

designs cases is given in the lengthy supplemental appendix.

Let Rp (h) = [(rp ((X1 − x̄)/h), · · · , rp ((Xn − x̄)/h))0 ] be the n × (1 + p) design matrix, and

K− (h) = diag(1(Xi < x̄)Kh (Xi − x̄) : i = 1, 2, · · · , n) and K+ (h) = diag(1(Xi ≥ x̄)Kh (Xi − x̄) :

i = 1, 2, · · · , n) be the n × n weighting matrices for control and treatment units, respectively. We

30
 (a) (a) (a) (a) (a) (a)
also define µS− := (µY − , µZ− 0 )0 , µS+ := (µY + , µZ+ 0 )0 , a ∈ Z+ , and σ 2S− := V[Si (0)|Xi = x̄] and

σ 2S+ := V[Si (1)|Xi = x̄], recall that Si (t) = (Yi (t), Zi (t))0 , t ∈ {0, 1}. Let eν denote a conformable

(1 + ν)-th unit vector. Finally, recall that s(h) = (1, −γ̃ Y (h))0 and s = (1, −γ Y )0 .

The pre-asymptotic bias Bτ̃ (h) = Bτ̃ + (h) − Bτ̃ − (h) and its asymptotic counterpart Bτ̃ := Bτ̃ + −

Bτ̃ − are characterized by

(p+1) (p+1)
s0 µS− s0 µS−
Bτ̃ − (h) := e00 Γ−1
−,p (h)ϑ−,p (h) →P Bτ̃ − := e00 ∆p,−
(p + 1)! (p + 1)!
(p+1) (p+1)
0 −1
s0 µS+ s0 µS+
Bτ̃ + (h) := e0 Γ+,p (h)ϑ+,p (h) →P Bτ̃ + := e00 ∆p,+
(p + 1)! (p + 1)!

where, with the (slightly abusive) notation vk = (v1k , v2k , · · · , vnk )0 , ιn = (1, · · · , 1)0 ∈ Rn , Γ−,p (h) =

Rp (h)0 K− (h)Rp (h)/n and ϑ−,p (h) = Rp (h)0 K− (h)(X − x̄ιn /h)p+1 /n, Γ+,p (h) and ϑ+,p (h) defined

analogously after replacing K− (h) with K+ (h), and

Z 0 −1 Z 0 
0 1+p
∆p,− := rp (u)rp (u) K(u)du rp (u)u K(u)du ,
−∞ −∞
Z ∞ −1 Z ∞ 
0 1+p
∆p,+ := rp (u)rp (u) K(u)du rp (u)u K(u)du .
0 0

The pre-asymptotic variance Vτ̃ (h) = Vτ̃ − (h) + Vτ̃ + (h) and its asymptotic counterpart Vτ̃ :=

Vτ̃ − + Vτ̃ + are characterized by

s0 σ 2S− s 0
Vτ̃ − (h) := [s0 ⊗ e00 P−,p (h)]ΣS− [s ⊗ P−,p (h)e0 ] →P Vτ̃ − := e0 Λp,− e0
f
s0 σ 2S+ s 0
Vτ̃ + (h) := [s0 ⊗ e00 P+,p (h)]ΣS+ [s ⊗ P+,p (h)e0 ] →P Vτ̃ + := e0 Λp,+ e0
f

√ √ √ −1 √
where P−,p (h) = hΓ−1 0
−,p (h)Rp (h) K− (h)/ n and P+,p (h) = hΓ+,p (h)Rp (h)0 K+ (h)/ n, and

Z 0 −1 Z 0  Z 0 −1
Λp,− := rp (u)rp (u)0 K(u)du rp (u)rp (u)0 K(u)2 du rp (u)rp (u)0 K(u)du ,
−∞ −∞ −∞
Z ∞ −1 Z ∞  Z ∞ −1
0 0 2 0
Λp,+ := rp (u)rp (u) K(u)du rp (u)rp (u) K(u) du rp (u)rp (u) K(u)du .
0 0 0

(p+1)
To construct pre-asymptotic estimates of the bias terms, we replace the only unknowns, µS−
(p+1)
and µS+ , by q-th order (p < q) local polynomial estimates thereof, using the preliminary band-

31
width b. This leads to the pre-asymptotic feasible bias estimate B̃τ̃ (b) := B̃τ̃ + (b) − B̃τ̃ − (b) with

(p+1) (p+1)
s(h)0 µ̃S−,q (b) s(h)0 µ̃S+,q (b)
B̃τ̃ − (b) := e00 Γ−1
−,p (h)ϑ−,p (h) and B̃τ̃ + (b) := e00 Γ−1
+,p (h)ϑ+,p (h)
(p + 1)! (p + 1)!

(p+1) (p+1)
where µ̃S−,q (b) and µ̃S+,q (b) collect the q-th order local polynomial estimates of the (p + 1)-th

derivatives using as outcomes each of the variables in Si = (Yi , Z0i )0 for control and treatment units,

that is, as in (1). Therefore, the bias-corrected covariate-adjusted sharp RD estimator is

1
τ̃ bc (h) = √ [s(h)0 ⊗ e00 (Pbc bc
+,p (h, b) − P−,p (h, b))]S,
nh

with S = (Y, vec(Z)0 )0 , Y = (Y1 , Y2 , · · · , Yn )0 , and

√  √
Pbc hΓ−1 0 1+p
ϑ−,p (h)e0p+1 Γ−1 0

−,p (h, b) = −,p (h) Rp (h) K− (h) − ρ −,q (b)Rq (b) K− (b) / n,

√  √
Pbc hΓ−1 0 1+p
ϑ+,p (h)e0p+1 Γ−1 0

+,p (h, b) = +,p (h) Rp (h) K+ (h) − ρ +,q (b)Rq (b) K+ (b) / n,

where P̃bc bc
−,p (h, b) and P̃−,p (h, b) are directly computable from observed data, given the choices of

bandwidth h and b, with ρ = h/b, and the choices of polynomial order p and q, with p < q.

The exact form of the (pre-asymptotic) heteroskedasticity-robust or cluster-robust variance

estimator follows directly from the formulas above. All other details such preliminary bandwidth

selection, plug-in data-driven MSE-optimal bandwidth estimation, and other extensions and results,

are given in the supplemental appendix.

References

Abadie, A. (2003), “Semiparametric Instrumental Variable Estimation of Treatment Response Mod-

els,” Journal of Econometrics, 113, 231–263.

Abadie, A., and Imbens, G. W. (2006), “Large Sample Properties of Matching Estimators for

Average Treatment Effects,” Econometrica, 74, 235–267.

Abadie, A., and Imbens, G. W. (2008), “Estimation of the Conditional Variance in Paired Experi-

ments,” Annales d’Economie et de Statistique, 175–187.

32
Angrist, J., and Rokkanen, M. (2015), “Wanna Get Away? Regression Discontinuity Estimation

of Exam School Effects Away from the Cutoff,” Journal of the American Statistical Association,

110, 1331–1344.

Armstrong, T. B., and Kolesar, M. (2015), “A Simple Adjustment for Bandwidth Snooping,”

arXiv:1412.0267.

Bartalotti, O., and Brummet, Q. (2016), “Regression Discontinuity Designs with Clustered Data:

Mean Square Error and Bandwidth Choice,” in Regression Discontinuity Designs: Theory and

Applications (Advances in Econometrics, volume 38), eds. M. D. Cattaneo and J. C. Escanciano,

Emerald Group Publishing, to appear.

Calonico, S., Cattaneo, M. D., and Farrell, M. H. (2016a), “On the Effect of Bias Estimation on

Coverage Accuracy in Nonparametric Inference,” arXiv:1508.02973.

Calonico, S., Cattaneo, M. D., Farrell, M. H., and Titiunik, R. (2016b), “rdrobust: Software for

Regression Discontinuity Designs,” working paper, University of Michigan.

Calonico, S., Cattaneo, M. D., and Titiunik, R. (2014), “Robust Nonparametric Confidence Inter-

vals for Regression-Discontinuity Designs,” Econometrica, 82, 2295–2326.

Cameron, A. C., and Miller, D. L. (2015), “A Practitioner’s Guide to Cluster-Robust Inference,”

Journal of Human Resources, 50, 317–372.

Canay, I. A., and Kamat, V. (2015), “Approximate Permutation Tests and Induced Order Statistics

in the Regression Discontinuity Design,” CeMMAP working paper CWP27/15.

Card, D., Lee, D. S., Pei, Z., and Weber, A. (2015), “Inference on Causal Effects in a Generalized

Regression Kink Design,” Econometrica, 83, 2453–2483.

Cattaneo, M. D., Frandsen, B., and Titiunik, R. (2015), “Randomization Inference in the Regression

Discontinuity Design: An Application to Party Advantages in the U.S. Senate,” Journal of Causal

Inference, 3, 1–24.

Cattaneo, M. D., Titiunik, R., and Vazquez-Bare, G. (2016), “Comparing Inference Approaches for

RD Designs: A Reexamination of the Effect of Head Start on Child Mortality,” working paper,

University of Michigan.

33
Chay, K. Y., McEwan, P. J., and Urquiola, M. (2005), “The Central Role of Noise in Evaluating

Interventions That Use Test Scores to Rank Schools,” American Economic Review, 95, 1237–

1258.

de la Cuesta, B., and Imai, K. (2016), “Misunderstandings about the Regression Discontinuity

Design in the Study of Close Elections,” Annual Review of Political Science, forthcoming, 19.

Dong, Y., and Lewbel, A. (2015), “Identifying the Effect of Changing the Policy Threshold in

Regression Discontinuity Models,” Review of Economics and Statistics, 97, 1081–1092.

Fan, J., and Gijbels, I. (1996), Local Polynomial Modelling and Its Applications, New York: Chap-

man & Hall/CRC.

Gelman, A., and Imbens, G. W. (2014), “Why High-Order Polynomials Should Not be Used in

Regression Discontinuity Designs,” NBER working paper 20405.

Hahn, J., Todd, P., and van der Klaauw, W. (2001), “Identification and Estimation of Treatment

Effects with a Regression-Discontinuity Design,” Econometrica, 69, 201–209.

Imbens, G., and Lemieux, T. (2008), “Regression Discontinuity Designs: A Guide to Practice,”

Journal of Econometrics, 142, 615–635.

Imbens, G. W., and Kalyanaraman, K. (2012), “Optimal Bandwidth Choice for the Regression

Discontinuity Estimator,” Review of Economic Studies, 79, 933–959.

Imbens, G. W., and Rubin, D. B. (2015), Causal Inference in Statistics, Social, and Biomedical

Sciences, Cambridge University Press.

Kamat, V. (2015), “On Nonparametric Inference in the Regression Discontinuity Design,”

arXiv:1505.06483.

Keele, L. J., Titiunik, R., and Zubizarreta, J. (2015), “Enhancing a Geographic Regression Discon-

tinuity Design Through Matching to Estimate the Effect of Ballot Initiatives on Voter Turnout,”

Journal of the Royal Statistical Society: Series A, 178, 223–239.

Lee, D. S. (2008), “Randomized Experiments from Non-random Selection in U.S. House Elections,”

Journal of Econometrics, 142, 675–697.

34
Lee, D. S., and Lemieux, T. (2010), “Regression Discontinuity Designs in Economics,” Journal of

Economic Literature, 48, 281–355.

Long, J. S., and Ervin, L. H. (2000), “Using Heteroscedasticity Consistent Standard Errors in the

Linear Regression Model,” The American Statistician, 54, 217–224.

Ludwig, J., and Miller, D. L. (2007), “Does Head Start Improve Children’s Life Chances? Evidence

from a Regression Discontinuity Design,” Quarterly Journal of Economics, 122, 159–208.

MacKinnon, J. G. (2012), “Thirty years of heteroskedasticity-robust inference,” in Recent Advances

and Future Directions in Causality, Prediction, and Specification Analysis, eds. X. Chen and N. R.

Swanson, Springer.

Muller, H.-G., and Stadtmuller, U. (1987), “Estimation of Heteroscedasticity in Regression Analy-

sis,” The Annals of Statistics, 15, 610–625.

Porter, J. (2003), “Estimation in the Regression Discontinuity Model,” working paper, University

of Wisconsin.

Skovron, C., and Titiunik, R. (2016), “A Practical Guide to Regression Discontinuity Designs in

Political Science,” working paper, University of Michigan.

35
 Table 1: Empirical Illustrations

MSE-optimal bandwidths: not using covariates using covariates

Standard Cov-adjusted Cov-adjusted
Head Start Data
RD treatment effect −2.41 −2.51 −2.47
Inference with ρ = h/b unrestricted
Robust 95% CI [ −5.46 , −0.10 ] [ −5.37 , −0.45 ] [ −5.21 , −0.37 ]
CI length change (%) −8.25 −9.76
Robust p-value 0.042 0.021 0.024
Inference with ρ = h/b = 1
Robust 95% CI [ −6.41 , −1.09 ] [ −6.64 , −1.46 ] [ −6.54 , −1.39 ]
CI length change (%) −2.86 −3.23
Robust p-value 0.006 0.002 0.003
h|b 6.81 | 10.72 6.81 | 10.72 6.98 | 11.64
n− | n+ 234 | 180 234 | 180 240 | 184
Education Data
RD treatment effect 3.45 3.42 3.49
Inference with ρ = h/b unrestricted
Robust 95% CI [ 1.53 , 6.19 ] [ 1.56 , 6.02 ] [ 1.61 , 6.21 ]
CI length change (%) −4.19 −1.31
Robust p-value 0.001 0.001 0.001
Inference with ρ = h/b = 1
Robust 95% CI [ 1.45 , 7.36 ] [ 1.21 , 6.80 ] [ 1.10 , 6.87 ]
CI length change (%) −5.43 −2.47
Robust p-value 0.004 0.005 0.007
h|b 3.62 | 6.64 3.62 | 6.64 3.36 | 6.11
n− | n+ 385 | 280 385 | 280 362 | 259

Notes:
(i) All estimates are computed using a triangular kernel and nearest neighbor heteroskedasticity-robust variance
estimators.
(ii) Columns under “Standard” and “Cov-adjusted” correspond to, respectively, standard and covariate-adjusted RD
estimation and inference methods, given a choice of bandwidths.
(iii) Bandwidths used (h and b) are data-driven MSE-optimal for either standard RD estimator or covariated-adjusted
RD estimator (depending on the group of columns). Specifically, in the first two columns the bandwidths are selected
to be MSE-optimal for τ̂ (standard RD estimation), while in the third column the bandwidths are selected to be
MSE-optimal for τ̃ (covariate-adjusted RD estimation).

36
 Table 2: Simulation Results (MSE, Bias, Empirical Coverage and Interval Length)

τ̂ τ̃ Change (%)
√ √ √
M SE Bias EC IL M SE Bias EC IL M SE Bias EC IL
Model 1
MSE-POP 0.045 0.012 0.938 0.199 0.045 0.012 0.934 0.198 0.2 0.1 −0.4 −0.6
MSE-EST 0.045 0.018 0.924 0.171 0.045 0.018 0.927 0.171 0.0 −1.0 0.3 −0.2
CER-POP 0.052 0.006 0.934 0.242 0.052 0.006 0.929 0.240 0.4 1.2 −0.5 −0.9
CER-EST 0.049 0.010 0.940 0.207 0.049 0.010 0.933 0.206 0.5 −1.5 −0.7 −0.5
Model 2
MSE-POP 0.047 0.013 0.935 0.213 0.041 0.008 0.941 0.185 −13.5 −33.6 0.6 −13.4
MSE-EST 0.048 0.017 0.929 0.188 0.041 0.011 0.932 0.163 −15.1 −34.8 0.3 −13.4
CER-POP 0.054 0.006 0.933 0.258 0.048 0.004 0.931 0.223 −11.7 −34.1 −0.2 −13.6
CER-EST 0.053 0.009 0.941 0.227 0.046 0.006 0.940 0.196 −13.1 −34.1 −0.2 −13.5
Model 3
MSE-POP 0.044 0.013 0.935 0.200 0.043 0.010 0.938 0.193 −3.3 −19.6 0.3 −3.5
MSE-EST 0.046 0.017 0.926 0.177 0.043 0.014 0.929 0.170 −5.5 −17.2 0.3 −4.0
CER-POP 0.051 0.006 0.933 0.243 0.050 0.005 0.930 0.234 −1.8 −20.9 −0.3 −3.8
CER-EST 0.050 0.009 0.939 0.213 0.048 0.008 0.939 0.205 −4.0 −16.8 0.0 −4.2
Model 4
MSE-POP 0.050 0.013 0.938 0.225 0.035 0.007 0.938 0.160 −29.3 −46.6 0.1 −28.8
MSE-EST 0.051 0.017 0.931 0.199 0.035 0.008 0.938 0.142 −30.5 −52.1 0.8 −28.4
CER-POP 0.058 0.006 0.934 0.273 0.042 0.003 0.926 0.194 −27.6 −43.6 −0.9 −29.0
CER-EST 0.056 0.009 0.942 0.240 0.040 0.005 0.934 0.171 −27.9 −51.2 −0.8 −28.5

Notes:
(i) All estimators are computed using the triangular kernel, NN variance estimation, and two bandwidths (h and b).
(ii) Columns
√ τ̂ and τ̃ correspond to, respectively, standard RD estimation and covariate-adjusted RD estimation;
columns “ M SE” report the square root of the mean square error of point estimator; columns “Bias” report average
bias relative to target population parameter; and columns “EC” and “IL” report, respectively, empirical coverage
and interval length of robust bias-corrected 95% confidence intervals.
(iii) Rows correspond to bandwidth method used to construct the estimator and inference procedures. Rows “MSE-
POP” and “MSE-EST” correspond to, respectively, procedures using infeasible population and feasible data-driven
MSE-optimal bandwidths (without or with covariate adjustment depending on the column). Rows “CER-POP” and
“CER-EST” correspond to, respectively, procedures using infeasible population and feasible data-driven CER-optimal
bandwidths (without or with covariate adjustment depending on the column).

37
 Table 3: Simulation Results (Data-Driven Bandwidth Selectors)

Pop. Min. 1st Qu. Median Mean 3rd Qu. Max. Std. Dev.
Model 1
ĥτ̂ 0.144 0.097 0.168 0.191 0.195 0.217 0.338 0.041
h̃τ̃ 0.144 0.094 0.166 0.189 0.194 0.217 0.338 0.040
Model 2
ĥτ̂ 0.156 0.092 0.170 0.193 0.198 0.222 0.335 0.041
h̃τ̃ 0.158 0.095 0.171 0.197 0.201 0.227 0.336 0.042
Model 3
ĥτ̂ 0.156 0.091 0.169 0.193 0.197 0.221 0.335 0.040
h̃τ̃ 0.154 0.095 0.170 0.194 0.198 0.223 0.334 0.041
Model 4
ĥτ̂ 0.156 0.093 0.170 0.194 0.198 0.223 0.334 0.041
h̃τ̃ 0.161 0.088 0.172 0.199 0.203 0.231 0.336 0.043

Notes:
(i) All estimators are computed using the triangular kernel, and NN variance estimation.
(ii) Column “Pop.” reports target population bandwidth, while the other columns report summary statistics of the
distribution of feasible data-driven estimated bandwidths.
(iii) Rows ĥτ̂ and h̃τ̃ corresponds to feasible data-driven MSE-optimal bandwidth selectors without and with covariate
adjustment, respectively.

38