Regression Discontinuity

Dr. Muhammad Shahadat Hossain Siddiquee

Professor, Department of Economics
University of Dhaka
Regression Discontinuity
• In a nonexperimental setting, program eligibility rules can sometimes
be used as instruments for exogenously identifying participants and
nonparticipants.
• To establish comparability, one can use participants and
nonparticipants within a certain neighborhood of the eligibility
threshold as the relevant sample for estimating the treatment impact.
• Regression Discontinuity (RD) method allows observed as well as
unobserved heterogeneity to be accounted for.
• Although the cutoff or eligibility threshold can be defined
nonparametrically, the cutoff has in practice traditionally been defined
through an instrument.
Regression Discontinuity (Contd..)
• Discontinuities in program implementation, based on eligibility
criteria or other exogenous factors, can be very useful in
nonexperimental program evaluation.
• People above and below the threshold, assuming they are similar in
observed characteristics, can be distinguished in terms of outcomes.
• However, the samples across which to compare would have to be
sufficiently close to the eligibility cutoff to ensure comparability.
• Furthermore, unobserved heterogeneity may be a factor if people within
the eligible targeting range exhibit variation in actual take-up of the
program, leading to selection bias.
• In that case, eligible and noneligible samples close to the eligibility cutoff
would be taken to compare the average program effect.
Regression Discontinuity (Contd..)
• Discontinuity approaches are therefore similar to instrumental
variable (IV) methods because they introduce an exogenous variable
that is highly correlated with participation, albeit not akin to
participation.
• For example, Grameen Bank’s microcredit is targeted to households with
landholdings of less than half an acre; pension programs are targeted to
populations above a certain age; and scholarships are targeted to students
with high scores on standardized tests.
• By looking at a narrow band of units that are below and above the
cutoff point and comparing their outcomes, one can judge the program’s
impact because the households just below and above the threshold are
likely to be very similar to each other.
 Regression Discontinuity in Theory
• To model the effect of a particular program on individual outcomes yi
through an RD approach, one needs a variable si that determines
program eligibility (such as age, asset holdings, or the like) with an
eligibility cutoff of s∗.
• The estimating equation is yi  si   i , where individuals with si ≤ s∗,
for example, receive the program, and individuals with si > s∗ are not
eligible to participate.
• Individuals in a narrow band above and below s∗ need to be
“comparable” in that they would be expected to achieve similar
outcomes prior to program intervention.
Outcomes before Program Intervention: individuals below s∗ are considered poor,
and those above the threshold are considered nonpoor.
• If one assumes that limits exist on either side of the threshold s∗, the
impact estimator for an arbitrarily small ε > 0 around the threshold
would be the following
• E[yi|s∗ – ε] – E[yi|s∗ + ε] = E[βSi|s∗ – ε] – E[βSi|s∗ + ε]
• Taking the limit of both sides of above equation as ε → 0 would identify
β as the ratio of the difference in outcomes of individuals just
above and below the threshold, weighted by the difference in their
realizations of Si
Outcomes after program intervention as measured
by the discontinuity model
Regression Discontinuity in Theory
Steps in Applying the RD Approach
• Standard nonparametric regression (i.e., Nonparametric regression is a
category of regression analysis in which the predictor does not take a
predetermined form but is constructed according to information derived
from the data) can be used to estimate the treatment effect in either the
sharp or the fuzzy regression discontinuity setup.
• For a sharp discontinuity design, the treatment effect can be estimated by
a simple comparison of the mean outcomes of individuals to the left and
the right of the threshold. Specifically, local linear regressions on the
outcome y, given a set of covariates x , should be run for people on both
sides of the threshold, to estimate the difference y--y+
Steps in Applying the RD Approach (Contd..)
Steps in Applying the RD Approach (Contd..)

• For a fuzzy discontinuity design, a two-step process is needed.

• First of all, local linear regression can be applied on the outcome for
people on both sides of the threshold to determine the magnitude of the
difference (or discontinuity) for the outcome.
• Similarly, local linear regression can be applied to the treatment indicator
to arrive at a difference or discontinuity for the treatment indicator.
• The ratio of the outcome discontinuity to the treatment discontinuity is
the treatment effect for a fuzzy discontinuity design.
Steps in Applying the RD Approach (Contd..)
• Although impacts in a neighborhood of the cutoff point are nonparametrically
identified for discontinuity designs, the applied literature has more often used an
alternative parametric method in which the discontinuity in the eligibility criterion is
used as an IV for program placement.
• Graphing the predicted treatment effect also provides a useful contrast across eligible
and noneligible populations, as well as for those in a narrow band around the
threshold.
• Plotting the density of the variable determining eligibility around the threshold can
also help show whether the RD design is valid (that is, that members of the noneligible
sample do not ultimately become participants, which could happen, for example, if
they were aware of the threshold and adjusted their reported value of the eligibility
variable to qualify).
• Plotting the average values of the covariates around the threshold also can provide an
indication of specification problems, with either a sharp discontinuity or a fuzzy
discontinuity approach.
Variations of RD
• Numerous variations of RD designs are possible, including combinations
with randomized tie-breaking experiments around the threshold to create
stronger inferences. Depending on the nature of the eligibility rule (that
is, whether it is a function of a variable changing over time or from a
one-time intervention), panel or cross-section data can be used in RD
analysis.
• A tie-breaker randomization, for example, involves a situation where an
overlap occurs between treatment and control groups across the variable
determining eligibility for the program. In this situation, treatment would
be randomly assigned to observations in the area of overlap. Figure
below describes this situation.
Variations of RD: A Tie-Breaking Experiment
Variations of RD

• Another variant is where more than one cutoff point can be exploited to
compare treatment effects. The corresponding regression to estimate the
program impact, therefore, would include two treatment groups—one
corresponding to each discontinuity. Figure below describes this
situation.
Multiple Cutoff Points
Advantages and Disadvantages of the RD Approach
• The advantages of the RD method are: (a) it yields an unbiased estimate of
treatment effect at the discontinuity, (b) it can many times take advantage of
a known rule for assigning the benefit that is common in the designs of
social policy, and (c) a group of eligible households or individuals need not
be excluded from treatment.
• However, the concerns with RD are: (a) it produces local average treatment
effects (LATE) that are not always generalizable; (b) the effect is estimated at
the discontinuity, so, generally, fewer observations exist than in a randomized
experiment with the same sample size; and (c) that the specifi cation can be
sensitive to functional form, including nonlinear relationships and
interactions.
Advantages and Disadvantages of the RD
• One concern with the RD method is behavioral (Ravallion 2008). Program
officials may not always know precisely the eligibility criteria; hence,
behavioral responses to the program intervention may be confused with
actual targeting rules. Data collected prior to program intervention, in the
form of a baseline, for example, may help to clarify program design and
corresponding uptake.
• Another concern is that the exercise focuses only on individuals or units
closely situated around the threshold s∗. Whether this group is materially
interesting for the evaluator needs to be addressed; if program officials are
interested, for example, in identifying the effects of a program around a
geographic border and in determining whether the program should be
expanded across borders, the limited sample may not be as great a concern. A
similar example can be constructed about a poverty alleviation program
concerned with households whose status hovers near the poverty line.
Advantages and Disadvantages of the RD
• If the eligibility rules are not adhered to or change over time, the validity of the
discontinuity approach also needs to be examined more carefully.
• Robustness checks can be conducted to examine the validity of the discontinuity
design, including sudden changes in other control variables at the cutoff point.
• Examining the pattern in the variable determining eligibility can also be useful—
whether, for example, the average outcome exhibits jumps at values of the
variable other than the eligibility cutoff, as well as any discontinuities in the
conditional density of this variable.
• If the control data exhibit nonlinearities—for example, a steeper slope than the
treatment data—then a squared term for the selection variable can be added in
the outcome regression equation.
• Nonlinearities in the functional form can also be addressed by interacting the
selection variable with the cutoff point or, perhaps, by using shorter, denser
regression lines to capture narrower comparisons.