WEEK 7

Data Analysis and Applied Econometrics in Practice

 ECO372

Recap so far

¨ Linear Regression with Randomized design:

¤ A design which lends credence to the claim that the conditional independence assumption holds.
¤ Can be supported in the data by assessing whether observables differ across treatment.

¨ Linear Regression with Observational data:

¤ Harder to claim that the conditional independence assumption holds
¤ Can still look whether other predetermined observables differ across treatment
¤ Can control for these observables
¤ But still left needing to claim that additional unobservables are not leading to OVB.
n Sometimes a “natural experiment” or “quasi experiment” can bolster this claim.
n The 2021 Nobel Prize was awarded for such insights: https://www.nobelprize.org/prizes/lists/video-prize-
lectures-from-laureates-in-economic-sciences/

¨ We now turn to the different types of quasi experimental design

 ECO372

Regression Discontinuity Design

A type of quasi-experiment

General idea:
¨ Treatment depends on a cutoff / threshold (“discontinuity”) of the running variable
(or forcing variable)
¨ Observations on one side of cutoff are treated (D=1), and the other side control
(D=0)
¨ Assumption: observations close to the cutoff on both sides should be similar, except
assignment to treatment
 ECO372
Popularity in Economics Publications
Source: Cunningham’s Causal Inference (https://mixtape.scunning.com/index.html)

• Angrist, Joshua D., and Victor Lavy. 1999. “Using Maimonides’ Rule to Estimate the Effect of Class Size on Scholastic
Achievement.” Quarterly Journal of Economics 114 (2): 533–75.
• Black, Sandra E. 1999. “Do Better Schools Matter? Parental Valuation of Elementary Education.” Quarterly Journal of
Economics 114 (2): 577–99.
 ECO372

Assignment to Treatment depends on a cutoff

1 if 𝑅! ≥ 𝑘 0 if 𝑅! ≥ 𝑘
𝐷! = # or 𝐷! = #
0 if 𝑅! < 𝑘 1 if 𝑅! < 𝑘

Running variable Cutoff Treatment

𝑹𝒊 𝒌 𝑫𝒊

Age Minimum Legal Drinking Able to consume alcohol

age

Exam score Minimum Score required Entrance to med school

Votes 50% Being elected
Income Eligibility requirement Welfare program
Geographical position Boundary Policy that affects one region
 ECO372

Reminder: Potential Outcomes

¨ 𝑌%! : outcome for individual 𝑖 if not treated (𝐷! = 0)

¨ 𝑌&! : outcome for individual 𝑖 if treated (𝐷! = 1)

¨ Average Treatment Effect: 𝐸 𝑌&! − 𝑌%!

¨ But we either observe 𝑌! = 𝑌%! or 𝑌! = 𝑌&! NOT both for the same person

¨ Both 𝑌% and 𝑌& depend on many things, including running variable 𝑅

 ECO372

A linear regression model

¨ General model:
𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝑒
¨ The model for D=0:
𝑌 = 𝛼 + 𝛾𝑅 + 𝑒
¨ The model for D=1:
𝑌 = (𝛼 + 𝜌) + 𝛾𝑅 + 𝑒

¨ Can we graph this in Y-R space?

¤ Remember:
n Movement in anything measured in a graph is represented by movement along the line/curve and
n Movement in anything measured outside the graph leads to shifts in line/curve
n Movement in anything measured outside the graph that interacts with something measured in the graph
leads to shifts and pivots in the line/curve
 ECO372

𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝑒
𝐷=0 𝐷=1
Note: we can see
that treatment
depends on the 𝛾
variable R here.

There is a threshold
in R that determines Effect Size
treatment and we
observe no data 𝛾
where D=0 and R is 𝜌
above the threshold
(or D=1 and R is 𝛼
below the threshold)

These are key

elements of RDD.
 ECO372

Regression Discontinuity Design

𝑅 𝐷
Above/below cutoff
running treatment

𝑌
outcome

Running variable can have two effects on outcome:

¨ Assignment to treatment (necessary for RDD!)
¨ Direct effect on outcome, other than treatment
 ECO372

RDD in regression

¨ Basic RDD regression:

𝑌 = 𝛼 + 𝜌𝐷 + 𝑓 𝑅 + 𝜀

¨ 𝜌: effect of 𝐷 on 𝑌 (effect of interest)

¨ 𝑓 𝑅 controls for direct effect of 𝑅 on 𝑌

¤ Can be linear or not
 ECO372

Types of functional form for f(R)

¨ D is an intercept shifter alone

𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝑒&

¨ D is an intercept shifter and R has nonlinear effect modeled as a quadratic

¤ Remember polynomials from high school: y=ax2+bx+c
¤ Here is an online graphing tool to remind you of the look:
https://www.desmos.com/calculator/pfdwlq5qht

𝑌 = 𝛼 + 𝜌𝐷 + 𝛾& 𝑅 + 𝛾0 𝑅 0 + 𝑒0

¨ D is an intercept shifter and R has a differential linear effect by D

𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝛿𝐷 ∗ 𝑅 + 𝑒1
 ECO372

D is an intercept shifter alone: 𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝑒!

¨ Simulate this situation:

¤ 𝛼 = 34, 𝜌 = 10.5, 𝛾 = 0.5, 𝑅~𝑁 5,2 , 𝑒! ~𝑁(0,4), cutoff for D is 6
*RDD_simulate1
clear
capture log close
set obs 1000
set seed 7654321

* Generate running variable.

gen R = rnormal(5, 2)
replace R=0 if R < 0
drop if R > 10
sum R, det

* Set the cutoff at R=6. Treated if R > 6

gen D = 0
replace D = 1 if R > 6
gen R2 = R*R
gen Y = 34 + 10.5*D +0.5*R + rnormal(0, 4)
reg Y D R
predict Yhat
 ECO372

D is an intercept shifter alone: 𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝑒!

¨ Simulate this situation:

¤ 𝛼 = 34, 𝜌 = 10.5, 𝛾 = 0.5, 𝑅~𝑁 5,2 , 𝑒! ~𝑁(0,4), cutoff for D is 6
 ECO372

D is an intercept shifter alone: 𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝑒!

¨ Graph this situation:
¤ 𝛼 = 34, 𝜌 = 10.5, 𝛾 = 0.5, 𝑅~𝑁 5,2 , 𝑒" ~𝑁(0,4), cutoff for D is 6
scatter Y R if D==0, color(blue) msize(vsmall) || ///
scatter Y R if D==1, color(red) msize(vsmall) legend(off) xline(6, lstyle(foreground)) || ///
line Yhat R if D ==0, color(blue) sort || ///
line Yhat R if D ==1, color(red) sort xtitle("Running Variable (R)") ytitle("Outcome (Y)")
 ECO372
D is an intercept shifter and R is a quadratic: 𝑌 = 𝛼 + 𝜌𝐷 + 𝛾4 𝑅 + 𝛾5 𝑅 5 + 𝑒5

¨ Simulate this situation:

¤ 𝛼 = 34, 𝜌 = 10.5, 𝛾! = 4, 𝛾" = 0.2, 𝑅~𝑁 5,2 , 𝑒! ~𝑁(0,4), cutoff for D is 6
*RDD_simulate2
clear
capture log close
set obs 1000
set seed 7654321

* Generate running variable.

gen R = rnormal(5, 2)
replace R=0 if R < 0
drop if R > 10
sum R, det

* Set the cutoff at R=6. Treated if R > 6

gen D = 0
replace D = 1 if R > 6
gen R2 = R*R
gen Y = 34 + 10.5*D +4*R +.2*R2 + rnormal(0, 4)
reg Y D R R2
predict Yhat
 ECO372
D is an intercept shifter and R is a quadratic: 𝑌 = 𝛼 + 𝜌𝐷 + 𝛾4 𝑅 + 𝛾5 𝑅 5 + 𝑒5

¨ Simulate this situation:

¤ 𝛼 = 34, 𝜌 = 10.5, 𝛾! = 4, 𝛾" = 0.2, 𝑅~𝑁 5,2 , 𝑒! ~𝑁(0,4), cutoff for D is 6
 ECO372
D is an intercept shifter and R is a quadratic: 𝑌 = 𝛼 + 𝜌𝐷 + 𝛾4 𝑅 + 𝛾5 𝑅 5 + 𝑒5

¨ Simulate this situation:

¤ 𝛼 = 34, 𝜌 = 10.5, 𝛾! = 4, 𝛾" = 0.2, 𝑅~𝑁 5,2 , 𝑒! ~𝑁(0,4), cutoff for D is 6
scatter Y R if D==0, color(blue) msize(vsmall) || ///
scatter Y R if D==1, color(red) msize(vsmall) legend(off) xline(6, lstyle(foreground)) || ///
line Yhat R if D ==0, color(blue) sort || ///
line Yhat R if D ==1, color(red) sort xtitle("Running Variable (R)") ytitle("Outcome (Y)")
 ECO372
D is an intercept shifter and R has a differential linear effect by D:
𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝛿𝐷 ∗ 𝑅 + 𝑒!

¨ Simulate this situation:

¤ 𝛼 = 34, 𝜌 = 10.5, 𝛾 = 0.5, 𝛿 = −1, 𝑅~𝑁 5,2 , 𝑒! ~𝑁(0,4), cutoff for D is 6
*RDD_simulate3
clear
capture log close
set obs 1000
set seed 7654321

* Generate running variable.

gen R = rnormal(5, 2)
replace R=0 if R < 0
drop if R > 10
sum R, det

* Set the cutoff at R=6. Treated if R > 6

gen D = 0
replace D = 1 if R > 6
gen DR = D*R
gen Y = 34 + 10.5*D +0.5*R -1*DR + rnormal(0, 4)
reg Y D R DR
predict Yhat
 ECO372
D is an intercept shifter and R has a differential linear effect by D:
𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝛿𝐷 ∗ 𝑅 + 𝑒!

¨ Simulate this situation:

¤ 𝛼 = 34, 𝜌 = 10.5, 𝛾 = 0.5, 𝛿 = −1, 𝑅~𝑁 5,2 , 𝑒! ~𝑁(0,4), cutoff for D is 6
 ECO372
D is an intercept shifter and R has a differential linear effect by D:
𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝛿𝐷 ∗ 𝑅 + 𝑒!

¨ Simulate this situation:

¤ 𝛼 = 34, 𝜌 = 10.5, 𝛾 = 0.5, 𝛿 = −1, 𝑅~𝑁 5,2 , 𝑒! ~𝑁(0,4), cutoff for D is 6
scatter Y R if D==0, color(blue) msize(vsmall) || ///
scatter Y R if D==1, color(red) msize(vsmall) legend(off) xline(6, lstyle(foreground)) || ///
line Yhat R if D ==0, color(blue) sort || ///
line Yhat R if D ==1, color(red) sort xtitle("Running Variable (R)") ytitle("Outcome (Y)")
 ECO372

Centering around the cutoff point

¨ D is an intercept shifter alone

𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝑒&

¨ Repose with R centered around the cutoff K 𝑅? = 𝑅 − 𝑘

𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝑒&
𝑌 = 𝛼 + 𝜌𝐷 + 𝛾 𝑅? + 𝑘 + 𝑒&
𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅? + 𝛾𝑘 + 𝑒&
𝑌 = (𝛼 + 𝛾𝑘) + 𝜌𝐷 + 𝛾𝑅? + 𝑒&

¨ So, this changes only the meaning of intercept

¤ In the first model, the intercept is the average of Y when D and R are both 0
¤ In the second model, the intercept is the average of Y when D=0 and 𝑅 = 𝑅#
 ECO372

Centering around the cutoff point

¨ Simulate this situation:
¤ 𝛼 = 34, 𝜌 = 10.5, 𝛾 = 0.5, 𝑅~𝑁 5,2 , 𝑒! ~𝑁(0,4), cutoff for D is 6
*RDD_simulate4
clear
capture log close
set obs 1000
set seed 7654321

* Generate running variable.

gen R = rnormal(5, 2)
replace R=0 if R < 0
drop if R > 10
sum R, det

* Set the cutoff at R=6. Treated if R > 6

gen D = 0
replace D = 1 if R > 6
gen Rk = R-6
gen Y = 34 + 10.5*D +0.5*(Rk+6) + rnormal(0, 4)
reg Y D Rk
predict Yhat

scatter Y Rk if D==0, color(blue) msize(vsmall) || ///

scatter Y Rk if D==1, color(red) msize(vsmall) legend(off) xline(6, lstyle(foreground)) || ///
line Yhat Rk if D ==0, color(blue) sort || ///
line Yhat Rk if D ==1, color(red) sort xtitle("Running Variable realtive to cutoff (Rk)") ytitle("Outcome (Y)")
 ECO372

Centering around the cutoff point

𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝑒& 𝑌 = (𝛼 + 𝛾𝑘) + 𝜌𝐷 + 𝛾𝑅? + 𝑒&
 ECO372

Centering around the cutoff point

𝑌 = 𝛼 + 𝜌𝐷 + 𝛾𝑅 + 𝑒& 𝑌 = (𝛼 + 𝛾𝑘) + 𝜌𝐷 + 𝛾𝑅? + 𝑒&

𝛼 + .5 ∗ 6
𝛼 𝛼 + .5 ∗ 6
 ECO372

Example: MLDA

¨ Minimum drinking age 21 in most US states

¨ Q: Effect on mortality of young adults?

¨ Look at mortality as a function of age

¨ Treatment effect = jump in trend

 ECO372

MLDA revisited

¨ Running variable?
¤ 𝑎: age
¨ Cutoff?
¤ 𝑎$ : 21y, or 252 months
¨ Treatment?
¤ 𝐷: legal access to alcohol
¨ Outcomes?
¤ 𝑀: mortality
ECO372
 ECO372

Regression Discontinuity Design

𝑅 𝐷
Above/below cutoff
running treatment
Above/below 21
age legal alcohol

𝑌
outcome
mortality
 ECO372

MLDA RDD in regression

𝑀B = 𝛼 + 𝜌𝐷B + 𝛾 ⋅ (𝑎 − 252) + 𝑒C
¨ 𝑎: age in months
¤ 𝑎 − 252: months before or after turning 21
¨ 𝑀B : mortality rate for month 𝑎
¨ 𝐷B : dummy = 1 if running variable 𝑎 ≥ 𝑎%
¨ 𝛾 (“gamma”): natural trend of mortality with age (absent access to alcohol)
¤ Trend is the same before or after 21
¨ 𝜌 (“rho”): effect of having access to alcohol

Strictly speaking, could use 𝑎 instead of (𝑎 − 252), would just change intercept (as we’ve seen)
 ECO372

Regression Discontinuity Design

Restricting to observations around cutoff allows to avoid confounding problems

𝑅 𝐷
Above/below cutoff
running treatment
Above/below 21
age legal alcohol

𝑌
outcome
mortality 𝑈
Unobs.
confounders
 ECO372

Two assumptions for RDD

Assumption 1 (design):
Discontinuity in treatment at cutoff
¨ Treatment determined by which side of cutoff

Assumption 2 (identification):
Continuity of potential outcomes 𝒀𝟎 and 𝒀𝟏 at the cutoff
¨ Variation of CIA: Nothing that affects outcome is different on either side of cutoff,
other than treatment
¨ Assignment to treatment as if random (quasi-experiment)
ECO372
 ECO372

Regression versus RDD

¨ Regression: use controls to bring covariates to similar level

¤ Once we control for a confounder 𝑋% , treatment assignment is as good as random
¤ There will be units with the same values of the controls (matches) but with different treatment
status
¨ RDD extrapolates at the discontinuity
¤ There is no value of the running variable where we observe both treatment and control
observations.
¤ We need to be willing to extrapolate across the discontinuity
 ECO372

Local estimate – Bandwidth

Only keep observations in band around cutoff:

[21 − 𝑏; 21 + 𝑏]
¨ 𝑏 “bandwidth”

¨ RDD ASSUMPTION:
Average outcome for individuals marginally below the cutoff must
represent a valid counterfactual for the treated group just above
the cutoff.

Ø Only local estimate (LATE), at the discontinuity

¤ Don’t know effect of access to alcohol on mortality of 29 years
old
ECO372
 ECO372

How to choose bandwidth?

¨ Wider means:
¤ More data points (more precise)
¤ Introducing selection bias as we move away from cutoff (observations might differ in other ways
than just the running variable)

¨ ∃ algorithms for optimal bandwidth choice (e.g. Imbens & Kalyanaraman, 2012)

¨ More important point: check sensitivity of results to choice of bandwidth

 ECO372

Functional form of control

 ECO372

Regression Discontinuity Design

𝑅 𝐷
Above/below cutoff
running treatment

𝑌
outcome

Running variable can have two effects on outcome:

¨ Assignment to treatment (necessary for RDD!)
¨ Direct effect on outcome, other than treatment
 ECO372

Functional form of 𝑓(𝑅)

Cutoff age: 𝑎%
Linear 𝑓(𝑅) :

𝑀B = 𝛼 + 𝜌𝐷B + 𝛾(𝑎 − 𝑎% ) + 𝑒C

Can we have something more general?

𝑀B = 𝛼 + 𝜌𝐷B + 𝒇(𝑎 − 𝑎% ) + 𝑒C

Necessary:
Ø to isolate the “discontinuity”
Ø To avoid mistakes
 ECO372

Wrong functional form misidentifies causal effect

Estimated effect > 0

True CEF
Estimated CEF

Example: causal effect = 0, but estimated effect > 0

 ECO372

Different slopes on each side (interaction term)

𝑀B = 𝛼 + 𝜌𝐷B
+𝛾 𝑎 − 𝑎% + 𝛿 𝒂 − 𝒂𝟎 𝑫𝒂 + 𝑒C

Slope before cutoff 𝛾

Slope after cutoff 𝛾+𝛿

Ø Similar to heterogeneous effect of age (although that’s not what we are interested)
ECO372
ECO372
 ECO372

Quadratic fit

𝑀B = 𝛼 + 𝜌𝐷B
+𝛾& 𝑎 − 𝑎% + 𝜸𝟐 𝒂 − 𝒂𝟎 𝟐 + 𝑒C

Add curvature (𝛾0 )

Ø Similar to heterogeneous effect of age (although that’s not what we are interested)
ECO372
 ECO372

Quadratic fit + interaction terms

𝑀B = 𝛼 + 𝜌𝐷B
+𝛾& 𝑎 − 𝑎% + 𝛾0 𝑎 − 𝑎% 0
+𝛿& 𝑎 − 𝑎% 𝐷B + 𝛿0 𝑎 − 𝑎% 0 𝐷B + 𝑒B

Ø Different slopes before/after: 𝛾& versus 𝛾& + 𝛿&

Different curvatures before/after: 𝛾0 versus 𝛾0 + 𝛿0

¨ As usual with regressions, we can add extra controls to achieve CIA and decrease
SEs
ECO372
ECO372
 ECO372

How to get functional form right?

Flexibility allows better fit

¨ Less likely to mistake Treatment effect for non-linearity

But too much flexibility (high-order polynomial) :

¨ increases standard error (more parameters to fit with same data)
¨ May overweight observations very close to bandwidth and mis-estimate Treatment
Effect (Gelman & Imbens, 2019)

Ø Start simple and reduce bandwidth

 ECO372

Plotting is not enough: hard to see effects visually

https://twitter.com/KiraboJackson/status/1074062192037847040
 ECO372

Plotting is not enough: hard to see effects visually

Easier to detect by plotting averages by value of the running variable

 ECO372

Example: Incumbency effect

Lee (2008)
 ECO372

David S. Lee (2008)

David S. Lee (2008) “Randomized experiments from non-random selection in U.S. House elections”, Journal of
Econometrics vol. 142.

Q: In elections, are incumbent candidates at an advantage?

[Toronto 2018 municipal elections: only 2 councillor successfully challenged

incumbent.]

Data: history of US House of Representatives elections

 ECO372

If there is incumbency effect

¨ Democrat incumbent (Democrat won election 𝑡) makes Democrat score higher in

election (𝑡 + 1)
¨ Republican incumbent (Democrat lost election 𝑡) makes Democrat score lower in
election (𝑡 + 1)

Ø Incumbency effect would create positive correlation b/w Democrat vote shares in
elections 𝑡 and (𝑡 + 1)

Problem: some districts like Democrats, so high Democrats vote share follows high
Democrat vote shares, and some like Republicans (high Rs follow high Rs)
 ECO372

Estimating incumbency advantage using RDD

𝑅
Democrat wins running
Previous election Democrat
previous vote share

𝐷 𝑌
treatment ? outcome
Incumbent is Democrat vote
Democrat 𝑋 share
Taste for
Democratic
Party
 ECO372

Incumbency effect in Lee (2008)

Main estimation:
𝑉H,IJ& = 𝛼 + 𝜷𝐷H,I + 𝑓 𝑉H,I + 𝐷H,I ⋅ 𝑔 𝑉H,I + 𝜀H,IJ&

¨ 𝑉H,I and 𝑉H,IJ& : Democrat vote share this and next election
¨ 𝐷H,I = 1 if Democrat is incumbent
¨ 𝑓 𝑉H,I + 𝐷H,I ⋅ 𝑔 𝑉H,I : allows polynomial to be different on each side of threshold
 ECO372

Incumbency effect in Lee (2008)

𝑉H,IJ& = 𝛼 + 𝜷𝐷H,I + 𝑓 𝑉H,I + 𝐷H,I ⋅ 𝑔 𝑉H,I + 𝜀H,IJ&

(𝑓 + 𝑔) 𝑉#,%

𝑉<,=>4 𝑓 𝑉#,% - = 𝟎. 𝟎𝟕𝟕

𝜷

𝑫𝒅,𝒕 = 𝟎 𝑫𝒅,𝒕 = 𝟏

𝑉<,=
 ECO372

Checking identification assumption

Close to cutoff, no difference in taste for Democrats

 ECO372

Manipulation into treatment

Threat to identification
 ECO372

Can the running variable be manipulated?

¨ Can I ask my doctor to put me as slightly lower eyesight to be eligible for laser
surgery?
¨ Can I get a fake ID that says 21?

Conditional Independence Assumption likely violated

¨ Only certain people might manipulate their assignment, so assignment is not
uncorrelated with error term

⇨ Check distribution of running variable for bunching

Bunching is a sign of manipulation into treatment

Bunching at threshold No bunching

Distribution of the running variable

 ECO372

Manipulation into treatment (1)

When individuals manipulate to be on a specific side of cutoff, distribution of running

variable displays “bunching”:
¨ There exists statistical procedures to test for bunching (McCrary, 2008)

Violation of Conditional Independence Assumption

¨ Individuals above and below cutoff might differ in other ways than through 𝐷!
¨ Potential outcome functions 𝑌% and 𝑌& not continuous at threshold
 ECO372

Manipulation into treatment (2)

Also, check distribution of covariates around cutoff to check selection bias

¨ Can control for covariates but…

¨ Difference in observable characteristics (“selection on observables”) often sign of

differences in unobserved characteristics (violation of CIA)

⇨ Nothing can be done to mitigate selection bias

 ECO372

Fuzzy RD
 ECO372

Fuzzy RD

¨ Before, treatment deterministic

¤ Prob = 1 if 𝑅% ≥ 𝑘 and 0 otherwise

¨ Now probability/intensity of treatment increasing function of running variable 𝑅, with

a jump at cutoff 𝑘
¤ “Fuzzy”: treatment doesn’t cleanly switch on at 𝑅% = 𝑘

¨ Since the treatment doesn’t affect all units, jump at the cutoff in the outcome needs
to be rescaled by the jump at the cutoff in the probability of treatment
 ECO372

Sharp versus Fuzzy RDD

 ECO372

RDD Summary

¨ Probability of treatment discontinuous at cut-off

¨ Potential outcome continuous at cutoff
¤ Need to make sure to have the right functional form
¤ No manipulation into treatment!

Good reference: Imbens & Lemieux “Regression

discontinuity designs: A guide to practice” (2008, Journal of
Econometrics)