Regression Discontinuity Designs in Stata

Matias D. Cattaneo
University of Michigan

July 30, 2015

Overview

Main goal: learn about treatment eect of policy or intervention.

If treatment randomization available, easy to estimate treatment eects.

If treatment randomization not available, turn to observational studies.

I Instrum ental variables.

I Selection on observables.

Regression discontinuity (RD) designs.

I Sim ple and ob jective. Requires little inform ation, if design available.
I M ight b e viewed as a lo cal random ized trial.
I Easy to falsify, easy to interpret.
I C areful: very local!
Overview of RD packages
https://sites.google.com/site/rdpackages

rdrobust package: estimation, inference and graphical presentation using local

polynomials, partitioning, and spacings estimators.
I rdrobust: RD inference (p oint estim ation and CI; classic, bias-corrected, robust).
I rdbwselect: bandwidth or window selection (IK, CV, CCT).
I rdplot: plots data (with optim al blo ck length).

rddensity package: discontinuity in density test at cuto (a.k.a. manipulation testing)

using novel local polynomial density estimator.
I rddensity: m anipulation testing using lo cal p olynom ial density estim ation.
I rdbwdensity: bandwidth or window selection.

rdlocrand package: covariate balance, binomial tests, randomization inference

methods (window selection & inference).
I rdrandinf: inference using random ization inference m etho ds.
I rdwinselect: falsication testing and window selection.
I rdsensitivity: treatm ent eect m o dels over grid of windows, CI inversion.
I rdrbounds: Rosenbaum b ounds.
Randomized Control Trials

Notation: (Yi (0); Yi (1); Xi ), i = 1; 2; : : : ; n.

Treatment: Ti 2 f0; 1g, Ti independent of (Yi (0); Yi (1); Xi ).

Data: (Yi ; Ti ; Xi ), i = 1; 2; : : : ; n, with

(
Yi (0) if Ti = 0
Yi =
Yi (1) if Ti = 1

Average Treatment Eect:

ATE = E[Yi (1) Yi (0)] = E[Yi jT = 1] E[Yi jT = 0]

Experimental Design.
Sharp RD design

Notation: (Yi (0); Yi (1); Xi ), i = 1; 2; : : : ; n, Xi continuous

Treatment: Ti 2 f0; 1g, Ti = 1(Xi x).

Data: (Yi ; Ti ; Xi ), i = 1; 2; : : : ; n, with

(
Yi (0) if Ti = 0
Yi =
Yi (1) if Ti = 1

Average Treatment Eect at the cuto:

SRD = E[Yi (1) Yi (0)jXi = x] = lim E[Yi jXi = x] lim E[Yi jXi = x]
x#x x"x

Quasi-Experimental Design: local randomization (more later)

 3
2
Outcome variable (Y)

0
1
0
1
2

0.6 0.4 0.2 0.0 0.2 0.4 0.6

Assignment variable (R)

 3
2
Outcome variable (Y)

0
1
0
1
2

0.6 0.4 0.2 0.0 0.2 0.4 0.6

Assignment variable (R)

 3
2
Outcome variable (Y)

0
1
0
1
2

0.6 0.4 0.2 0.0 0.2 0.4 0.6

Assignment variable (R)

 6
4
2
Outcome variable (Y)

0
0
2

Local Random Assignment

4
6

0.6 0.4 0.2 0.0 0.2 0.4 0.6

Assignment variable (R)

Empirical Illustration: Cattaneo, Frandsen & Titiunik (2015, JCI)

Problem: incumbency advantage (U.S. senate).

Data:
Yi = election outcome.
Ti = whether incumbent.
Xi = vote share previous election (x = 0).
Zi = covariates (demvoteshlag1, demvoteshlag2, dopen, etc.).

Potential outcomes:
Yi (0) = election outcome if had not been incumbent.
Yi (1) = election outcome if had been incumbent.

Causal Inference:
Yi (0) 6= Yi jTi = 0 and Yi (1) 6= Yi jTi = 1
Graphical and Falsication Methods

Always plot data: main advantage of RD designs!

Plot regression functions to assess treatment eect and validity.

Plot density of Xi for assessing validity; test for continuity at cuto and elsewhere.

Important: use also estimators that do not smooth-out data.

RD Plots (Calonico, Cattaneo & Titiunik, JASA):

I Two ingredients: (i) Sm o othed global p olynom ial t & (ii) binned discontinuous
lo cal-m eans t.
I Two goals: (i) detention of discontinuities, & (ii) representation of variability.
I Two tuning param eters:

F G lo b a l p o ly n o m ia l d e g re e (kn ).
F L o c a tio n (E S o r Q S ) a n d nu m b e r o f b in s (Jn ).
Manipulation Tests & Covariate Balance and Placebo Tests

Density tests near cuto:

I Idea: distribution of running variable should b e sim ilar at either side of cuto.

I M etho d 1: Histogram s & Binom ial count test.

I M etho d 2: Density Estim ator at b oundary.

F P re -b in n e d lo c a l p o ly n o m ia l m e th o d M c C ra ry (2 0 0 8 ).
F N e w tu n in g -p a ra m e te r-fre e m e th o d C a tta n e o , J a n sso n a n d M a (2 0 1 5 ).

Placebo tests on pre-determined/exogenous covariates.

I Idea: zero RD treatm ent eect for pre-determ ined/exogenous covariates.

I M ethods: global p olynom ial, lo cal p olynom ial, random ization-based.

Placebo tests on outcomes.

I Idea: zero RD treatm ent eect for outcom e at values other than cuto.
I M ethods: global p olynom ial, lo cal p olynom ial, random ization-based.
Estimation and Inference Methods

Global polynomial approach (not recommended).

Robust local polynomial inference methods.

I Bandwidth selection.

I Bias-correction.

I Condence intervals.

Local randomization and randomization inference methods.

I W indow selection.

I Estim ation and Inference m etho ds.

I Falsication, sensitivity and related m etho ds

Conventional Local-polynomial Approach

Idea: approximate regression functions for control and treatment units locally.

Local-linear estimator (w/ weights K( )):

hn Xi < x : x Xi hn :

Yi = + (Xi x) +" ;i Yi = + + (Xi x) + + "+;i

I Treatm ent eect (at the cuto): ^ SRD = ^ + ^

Can be estimated using linear models (w/ weights K( )):

Yi = + SRD Ti + (Xi x) 1 + Ti (Xi x) 1 + "i , hn Xi hn

Once hn chosen, inference is standard: weighted linear models.

I Details com ing up next.
Conventional Local-polynomial Approach
How to choose hn ?

Imbens & Kalyanaraman (2012, ReStud): optimal plug-in,

^ IK = C
h ^IK n 1=5

Calonico, Cattaneo & Titiunik (2014, ECMA): renement of IK

^ CCT = C
h ^CCT n 1=5

Ludwig & Miller (2007, QJE): cross-validation,

n
X
^ CV = arg min
h w(Xi ) (Yi ^ 1 (Xi ; h))2
h
i=1

Key idea: trade-o bias and variance of ^SRD (hn ). Heuristically:

" Bias(^SRD ) =) ^
#h and " Var(^SRD ) =) ^
"h
Local-Polynomial Methods: Bandwidth Selection
Two main methods: plug-in & cross-validation. Both MSE-optimal in some sense.

Imbens & Kalyanaraman (2012, ReStud): propose MSE-optimal rule,

1=5 1=5 Var(^SRD )

hMSE = CMSE n CMSE = C(K)
Bias(^SRD )2

I IK im plem entation: rst-generation plug-in rule.

I CCT im plem entation: second-generation plug-in rule.

I They dier in the way Var(^ SRD ) and Bias(^ SRD ) are estim ated.

Imbens & Kalyanaraman (2012, ReStud): discuss cross-validation approach,

n
X
^ CV = arg min CV (h) ,
h CV (h) = 1(X ;[ ] Xi X+;[ ] ) (Yi ^ (Xi ; h))2 ,
h>0
i=1

where
I ^ +;p (x; h) and ^ are lo cal p olynom ials estim ates.
;p (x; h)

I 2 (0; 1), X and X+;[ denote -th quantile of fXi : Xi < xg and fXi : Xi xg.
;[ ] ]

I Our im plem entation uses = 0:5; but this is a tuning param eter!
Conventional Approach to RD

Local-linear estimator (w/ weights K( )):

hn Xi < x : x Xi hn :

Yi = + (Xi x) +" ;i Yi = + + (Xi x) + + "+;i

I Treatm ent eect (at the cuto): ^ SRD = ^ + ^

Construct usual t-test. For H0 : SRD = 0,

^SRD ^+ ^
T^(hn ) = p = q d N (0; 1)
V^n ^ ^
V+;n + V ;n

95% Condence interval:

q
^ n) =
I(h ^SRD 1:96 V^n
Bias-Correction Approach to RD

Note well: for usual t-test,

^SRD
T^(hMSE ) = p d N (B; 1) 6= N (0; 1), B>0
V^n

I Bias B in RD estim ator captures curvature of regression functions.

^ n = 0:5 h
Undersmoothing/Small Bias Approach: Choose smaller hn ... Perhaps h ^ IK ?

=) Not clear guidance & power loss!

Bias-correction Approach:

^SRD B^n
T^bc (hn ; bn ) = p d N (0; 1)
^n
V
h p i
=) 95% Condence Interval: I^bc (hn ; bn ) = ^SRD ^n
B 1:96 ^n
V

How to choose bn ? Same ideas as before... ^bn = C

^ n 1=7
Robust Bias-Correction Approach to RD
Recall:
^ SRD ^ SRD ^n
B
bc
T^ (hn ) = p d N (0; 1) and T^ (hn ; bn ) = p d N (0; 1)
V^n ^n
V

I ^ n is constructed to estim ate leading bias B.

B

Robust approach:

^SRD B^n ^SRD Bn Bn B^n

T^bc (hn ; bn ) = p = p + p
^n
V ^n
V V^n
| {z } | {z }
d N (0;1) d N (0; )

Robust bias-corrected t-test:

^SRD B ^n ^SRD B ^n
T^rbc (hn ; bn ) = p = q d N (0; 1)
^n + W
V ^n ^
Vnbc

=) 95% Condence Interval:

q
I^rbc (hn ; bn ) = ^SRD ^n
B 1:96 ^n
V bc , ^n
V bc ^n + W
=V ^n
Local-Polynomial Methods: Robust Inference

Approach 1: Undersmoothing/Small Bias.

q
^ n) =
I(h ^SRD 1:96 V^n

Approach 2: Bias correction (not recommended).

q
I^bc (hn ; bn ) = ^SRD ^n
B 1:96 V^n

Approach 3: Robust Bias correction.

q
I^rbc (hn ; bn ) = ^SRD ^n
B 1:96 ^n + W
V ^n
Local-randomization approach and nite-sample inference

Popular approach: local-polynomial methods.

I Approxim ates regression function and relies on continuity assum ptions.

I Requires: choosing weights, bandwidth and p olynom ial order.

Alternative approach: local-randomization + randomization-inference

I Gives an alternative that can b e used as a robustness check.

I K ey assum ption: exists window W = [ hn ; hn ] around cuto ( hn < x < hn ) where

Ti indep endent of (Yi (0); Yi (1)) (for all Xi 2 W )

I In words: treatm ent is random ly assigned within W .

I Go o d news: if plausible, then RCT ideas/m etho ds apply.
I Not-so-good news: m ost plausible for very sm all windows (very few observations).
I One solution: em ploy sm all window but use random ization-inference m etho ds.
I Requires: choosing random ization rule, window and statistic.
Local-randomization approach and nite-sample inference

Recall key assumption: exists W = [ hn ; hn ] around cuto ( hn < x < hn ) where

Ti independent of (Yi (0); Yi (1)) (for all Xi 2 W )

How to choose window?

I Use balance tests on pre-determ ined/exogenous covariates.

I Very intuitive, easy to im plem ent.

How to conduct inference? Use randomization-inference methods.

1 Cho ose statistic of interest. E.g., t-stat for dierence-in-m eans.

2 Cho ose random ization rule. E.g., numb er of treatm ents and controls given.
3 Com pute nite-sam ple distribution of statistics by p ermuting treatm ent assignm ents.
Local-randomization approach and nite-sample inference

Do not forget to validate & falsify the empirical strategy.

1 Plot data to m ake sure lo cal-random ization is plausible.

2 Conduct placeb o tests.

(e.g., use pre-intervention outcom es or other covariates not used select W )

3 Do sensitivity analysis.

See Cattaneo, Frandsen and Titiunik (2015) for introduction.

See Cattaneo, Titiunik and Vazquez-Bare (2015) for further results and
implementation.