Regression Discontinuity Design

(RDD)
Patricio Valenzuela
 RDD Intuition
RDD is a quasi-experimental technique

• Unlike natural experiments, assignment to treatment and control is not

random

• We know the assignment rule influencing how people are assigned or

selected in to treatment

• There is a known cut-off in treatment assignment or in probability of

treatment receipt as a function of one or more continuous variables

• Recipients whose assignment variable is above the cutoff are assigned

to one group (e.g., treatment), those below assigned to the other (e.g.,
control).
 Some Nice Features of RDD
1. RDDs abound once you looked for them
– Program resources often allocated based on a formula with a cut-
off structure

2. RDD is intuitive and easily conveyed by a picture showing sharp

changes in treatment assignment and average outcomes around cut-
off value of assignment variable

3. There are several different ways to estimate the treatment effect, each
of which have credible causal interpretations
 Sharp RDD
In a Sharp RDD subjects assigned to or selected for treatment solely on
the basis of a cut-off value of an observed continuous variable, called the
assignment variable.

• Can be a single variable (e.g., Credit Score, income, accounting

variable)

• Or a function of a single variable, or a function of several variables

(e.g., average quarterly debt-to-ebitda ratio, sum of all household
expenditures)
 Sharp RDD Key Assumption
Assignment to treatment occurs through a known and measured
deterministic decision rule:

where x is the forcing variable and x′ the threshold.

Probability of Treatment Assignment in Sharp RDD
 Identifying Treatment Effects
Given the delineation of the data into treatment and control groups
by the assignment rule, a simple, albeit naive, approach to
estimation would be a comparison of sample averages.

(1)
ejemplo de las becas, y metodo sharp que es o 0 o 1 , en cambio fuzzy es continuo, por ejemplo como la tesis que explico
esa era un cambio de areas, por lo que es sharp.

where d = 1 for treatment observations and zero otherwise.

However, this specification assumes that treatment assignment d

and the error term u are uncorrelated so that assignment is as if it is
random with respect to potential outcomes.
 Identifying Treatment Effects
In the case of RDD, assignment is determined by a known rule that
ensures treatment assignment is correlated with the forcing variable, x, so
that d is almost surely correlated with u and OLS will not recover a
treatment effect of interest

In other words, the comparison of sample averages is confounded

by the forcing variables, net worth and current ratio.

One way to control for x is to include it in the regression as another

covariate:

(2)
 Identifying Treatment Effects
This approach is also unappealing because identification of the
parameters comes from all of the data, including those points that are far
from the discontinuity.

Yet, the variation on which RDD relies for proper identification of the
parameters is that occurring precisely at the discontinuity. This notion is
formalized in the second key assumption of sharp RDD, referred to as
the local continuity assumption.
 Sharp RDD Key Assumption # 2
Both potential outcomes, E(y(0)|x) and E(y(1)|x), are continuous in x at
x′. Equivalently, E(u|x) is continuous in x at x′.

Local continuity is a general assumption invoked in both sharp and fuzzy

RDD. As such, we do not preface this assumption with “Sharp,” as in the
previous assumption.
 Sharp RDD Key Assumption # 2

Assuming a positive density of x in a neighborhood containing the

threshold x′, local continuity implies that the limits of the
conditional expectation function around the threshold recover the
ATE at x′.

Taking the difference between the left and right limits in x of

equation (1) yields,
Conditional Expectation of Outcomes in Sharp RDD

La linea punteada son los no tratados,

por lo que ausmo su comportamiento

este salto puede ser continuo

 Sharp RDD Key Assumption # 2
• The local continuity assumption is that the conditional
expectations representing potential outcomes are smooth (i.e.,
continuous) around the threshold, as illustrated by the figure.
• What this continuity ensures is that the average outcome is
similar for subjects close to but on different sides of the
threshold.
• In other words, in the absence of treatment, outcomes would be
similar. However, the conditional expectation of the observed
outcome, E(y|x) is represented by the all solid line that is
discontinuous at the threshold, x′.
• Thus, continuity ensures that the only reason for different
outcomes around the threshold is the treatment.
 Fuzzy RDD
Assignment to treatment occurs in a stochastic manner where the
probability of assignment (propensity score function) has a known
discontinuity at x´

0  lim Pr  ti  1 | x   lim Pr  ti  1 | x   1
x  x x  x
Probability of Treatment Assignment in Fuzzy RDD

prob de quedar en la carrera si

tengo
 Fuzzy RDD Intuition
Fuzzy RDD is akin to:
• mis-assignment relative to the cut-off value in a sharp RDD
– Value of x near the cut-off appear in both treatment and control
groups
– Mis-assignment can occur if assignment is based on variables
observed by administrator but not evaluator
• random experiment with
– no-shows: treatment group members who do not receive
treatment, and
– cross-overs: control group members who do receive treatment

Practically speaking, imagine incentives to participate changing

discontinuously at cut-off
• But not powerful enough to move all subjects from non-participant to
participant status
 Fuzzy RDD Example
Decision to offer a scholarship based on:
• Continuous measure of academic ability (e.g., GRE) exceeds
given cut-off, and
• Subjective information (e.g., recommendation letters) observed
only by the evaluator

Does scholarship receipt impact academic achievement?

• Don’t compare recipients with non-recipients (even close to cut-
off) to estimate ATE  likely differ along unobservables related
to outcome (e.g., letters of rec)
• But, could compare average outcomes of all subjects, irrespective
of recipient status, just to the left and right of the cut-off…
 Identifying Treatment Effects

Maintaining the assumption of local continuity and a common

treatment effect,

This result implies that the treatment effect, common to the

population, β, is identified by
Graphical Analysis
 Internal and External Validity
• At best, Sharp and Fuzzy RDD estimate the average effect of the
sub-population with x close to x´
– Fuzzy RDD restricts this subpopulation even further to that of
the compliers with x close to x´

• Only with strong assumptions (e.g., homogenous treatment effects)

can we estimate the overall average treatment effect

• So, RDD have strong internal validity but weak external validity
 Density of Forcing Variable
(McCrary, 2008)
• Agents may manipulate forcing variable to self-select in/out of treatment
– Can, but not necessarily compromise identification
• Test for discontinuity in density of forcing variable
• Example: Beneficial job training program offered to agents with income <
x´. Concern, people will withhold labor to lower their income below the
cut-off to gain access to the program.

texto

• At a minimum, any discontinuity would need to be explained

 Estimation
Strictly speaking, we need to estimate boundary points of conditional
expectations. Recall ATE, under appropriate assumptions, in
– Sharp RDD:
lim E  yi | x   lim E  yi | x 
x  x x  x

– Fuzzy RDD:
lim E  y | x   lim E  y | x 
x  x x  x

lim E  t | x   lim E  t | x 
x  x x  x

With enough observations, we could focus on agents in a very small

interval around the cut-off and compare average outcomes for agents just
to the left and right of the cut-off
– Increasing the interval, increases the bias
The tradefoff researchers face when implementing a RDD is: bias versus
variance
 Parametric Estimation
An easier way to perform inference is to combine the data on both sides
of the threshold and estimate the following pooled regression:

where f and g are continuous functions. The treatment effect is β.

This approach maintains the functional form flexibility associated with

estimating two separate regressions by including the interaction term di ·
g(xi − x′).

This is an important feature since there is rarely a strong a priori rationale

for constraining the functional form to be the same on both sides of the
threshold.
 Fuzzy RDD
• In a fuzzy RDD, the above estimation approaches are typically
inappropriate.
• The estimation approaches discussed above will not recover unbiased
estimates of the treatment effect because of correlation between the
assignment variable di and ε.
• An easy solution to this problem is based on IV.
• We can take a similar approach here in solving the selection bias in the
assignment indicator, di, using the discontinuity as an instrument.
• Specifically, the probability of treatment can be written as,
 Fuzzy RDD
• Note, that the indicator T is not equal to di in the fuzzy RDD because
of misassignment or unobservables. Rather,

• where ω is a random error independent of x. Therefore, a fuzzy RDD

can be described by a two equation system:

• Estimation can be carried out with two stage least squares, where di is
the endogenous variable and Ti is the instrument.
• The standard exclusion restriction argument applies: Ti is only relevant
for outcomes, yi, through its impact on assignment, di.
 Sensitivity Analysis 1
• Check sensitivity of estimates to alternative specifications
– e.g., add higher order polynomials, vary bandwidth, etc.

• Restrict attention to subsample of observations close to the cut-off

– You can be more restrictive with the control function here since
the small distance will act as an instrument
– This reduces bias but also reduces efficiency
 Sensitivity Analysis 2
• Can subjects behavior invalidate the local continuity assumption?
– Can they exercise control over their values of the assignment
variable?
– Can administrators strategically choose what assignment variable
to use or which cut-off point to pick?
– Either can invalidate the comparability of subjects near the
threshold because of sorting of agents around the cut-off, where
those below may differ on average form those just above

• Continuity violated in the presence of other programs that use a

discontinuous assignment rule with the exact same assignment
variable and cut-off
 Sensitivity Analysis 3
• Even if agents or administrators (or both) exercise some control over
the forcing variable or cut-off position, continuity assumptions may
not be violated
– Lee (2008) shows that in Sharp RDD, as long as agents do not
have perfect control, continuity will be satisfied.
• i.e., there must be some independent random chance element
• Implies local conditional independence assumption will be
satisfied
• Manipulation will identify a weighted ATE

• Sorting undermines the causal interpretation of RDD only if sorting

is perfect
– Perhaps a break/discontinuity in the forcing variable (McCrary
(2008))
 Sensitivity Analysis 4
• Test for comparability of agents around the cut-off
– Visual test of covariates discussed earlier
– Repeat RDD using the characteristics as outcome variables (van
der Klaauw (2008))
– Finding a discontinuity does not necessarily invalidate the RDD
– Incorporate covariates, z, in the RDD, as additional controls
• This should only impact stat significance, not magnitude of
treatment effect
• Alternatively, regress the outcome variable on a vector of
controls and use the residuals in the RDD, instead of the
outcome itself

• This only addresses observables, not unobservables

 Sensitivity Analysis 5
• Falsification tests
– Test whether the treatment effect is zero when it should be
• e.g., at points away from the discontinuity
– Maybe data exists in a period where there was no program
– Test whether the actual cut-off fits the data better than near-by
cut-offs
• A spike in the log-likelihood at the actual relative to alternative
cut-off values can allay concerns that the found local
relationship was spurious
 Summary
• Sharp RDD
– Graph data: Average outcomes by forcing variable (discontinuity
at cut-off?)
– Estimate treatment effect: Use several methods for robustness
– Perform sensitivity analysis: Not just econometrics, think about
economics and potential concerns

• Fuzzy RDD
– Graph data: Average outcomes by forcing variable and
Pr(treatment)
– Estimate treatment effect: Use 2SLS and other methods for
robustness
– Perform sensitivity analysis: Not just econometrics, think about
economics and potential concerns