# Exam Schools, Ability, and the Eects of

Armative Action: Latent Factor Extrapolation in
the Regression Discontinuity Design
Miikka Rokkanen

Massachusetts Institute of Technology

Introduction

Latent Factor Modeling in a Sharp RD Design

Boston Exam Schools

Identication and Estimation

Extrapolation Results

Counterfactual Simulations

Placebo Experiments

Conclusions

Extrapolation Problem in the RD Design 𝑌
(0), 𝑌(1) 𝐸

[𝑌(1)|𝑅] 𝐸

[𝑌(1) − 𝑌(0)| 𝑅 = 𝑐] 𝐸

[𝑌(0)|𝑅] 𝑐 𝑅

Extrapolation Problem in the RD Design 𝑌(0). 𝑌(1) 𝐸[𝑌(1)| 𝑅 = 𝑟1 ] 𝐸[𝑌(1)|𝑅] 𝐸[𝑌(1)| 𝑅 = 𝑟0 ] 𝐸[𝑌(1) − 𝑌(0)| 𝑅 = 𝑟1 ] 𝐸[𝑌(1) − 𝑌(0)| 𝑅 = 𝑐] 𝐸[𝑌(1) − 𝑌(0)| 𝑅 = 𝑟0 ] 𝐸[𝑌(0)| 𝑅 = 𝑟1 ] 𝐸[𝑌(0)|𝑅] 𝐸[𝑌(0)| 𝑅 = 𝑟0 ] 𝑟0 𝑐 𝑟1 𝑅 .

Boston Exam Schools I Three selective high schools that are seen as the agship of the Boston public school system I RD estimates show little evidence of eects for marginal applicants (Abdulkadiroglu. Angrist & Pathak. forthcoming) I Is the lack of eects generalizable for inframarginal applicants? I Important question for discussion of armative action at selective school admissions .

This Paper 1. Simulate eects of introducing minority/socioeconomic preferences into exam school admissions I Both reforms increase average achievement among aected applicants . Estimate eects of Boston exam school attendance for full population of applicants I Achievement gains concentrated among lower-scoring applicants 3. Develop a latent factor-based approach to the extrapolation of treatment eects away from the cuto in RD I Nonparametric identication of treatment eects at any point in the running variable distribution 2.

Darolles. (Semi-)nonparametric instrumental variables models: Newey & Powell (2003). Chen & Kristensen (2007). Heckman & Schennach (2010). Pop-Eleches & Urquiola (2013).Related Literature 1. Florens & Renault (2011) 4. Blundell. Abdulkadiroglu. Card & Krueger (2005). Eects of attending selective middle/high schools: Jackson (2010). Cook & Wing (2013). Dong & Lewbel (2013) 2. Eects of armative action in school admissions: Arcidiacono (2005). Fan. Angrist & Pathak (forthcoming) 5. Evdokimov & White (2012) 3. Cunha. Latent factor and measurement error models: Hu & Schennach (2008). Extrapolation of treatment eects in RD: Angrist & Rokkanen (2013). Hinrichs (2012) .

Introduction Latent Factor Modeling in a Sharp RD Design Boston Exam Schools Identication and Estimation Extrapolation Results Counterfactual Simulations Placebo Experiments Conclusions .

Y (1)) ⊥⊥ R | θ I Example: selective school admission I I I R : entrance exam score θ : latent ability Y (0). Y (1): future achievement θ but not on . νR ) I Potential outcomes the noise in Y (0) and Y (1) depend on R: (Y (0) .Latent Factor Framework I Running variable R is a noisy measure of a latent factor θ: R = gR (θ .

Treatment Assignment in the Latent Factor Framework 𝜃 𝑙𝑜𝑤 𝑐 𝜃 ℎ𝑖𝑔ℎ 𝑅 .

Treatment Assignment in the Latent Factor Framework 𝜃 𝑙𝑜𝑤 + 𝜈𝑅𝑙𝑜𝑤 𝜃 𝑙𝑜𝑤 𝜃 ℎ𝑖𝑔ℎ + 𝜈𝑅𝑙𝑜𝑤 𝑐 ℎ𝑖𝑔ℎ 𝜃 𝑙𝑜𝑤 + 𝜈𝑅 𝜃 ℎ𝑖𝑔ℎ 𝜃 ℎ𝑖𝑔ℎ ℎ𝑖𝑔ℎ + 𝜈𝑅 𝑅 .

Extrapolation in the Latent Factor Framework 𝑌(0). 𝑌(1) 𝐸[𝑌(1)| 𝑅 = 𝑟1 ] 𝐸[𝑌(1)|𝑅] 𝐸[𝑌(1)| 𝑅 = 𝑟0 ] = 𝐸{𝐸[𝑌(1)| 𝜃]| 𝑅 = 𝑟0 } 𝐸[𝑌(1) − 𝑌(0)| 𝑅 = 𝑟1 ] 𝐸[𝑌(1) − 𝑌(0)| 𝑅 = 𝑐] 𝐸[𝑌(1) − 𝑌(0)| 𝑅 = 𝑟0 ] 𝐸[𝑌(0)| 𝑅 = 𝑟1 ] = 𝐸{𝐸[𝑌(0)| 𝜃]| 𝑅 = 𝑟1 } 𝐸[𝑌(0)|𝑅] 𝐸[𝑌(0)| 𝑅 = 𝑟0 ] 𝑟0 𝑐 𝑟1 𝑅 .

Object 1: Conditional Latent Factor Distribution 𝑓𝜃|𝑅 (𝜃|𝑟) 𝑓𝜃|𝑅 (𝜃|𝑟0 ) 𝑓𝜃|𝑅 (𝜃|𝑟1 ) 𝜃 .

𝑌(1) 𝐸[𝑌(1)|𝜃] 𝐸[𝑌(0)|𝜃] 𝜃 .Object 2: Latent Conditional Expectation Functions 𝑌(0).

M3 : baseline test scores Identication with less than three measures M = (M1 . M2 . νM2 ) M3 = gM3 (θ . M3 ) . and M3 is potentially discrete is a deterministic function of a subset of I Example: selective school admissions I I R = M1 : entrance exam score M2 . νM3 ) where I R M1 .Identication Using Multiple Noisy Measures I Suppose one has available three noisy measures of θ: M1 = gM1 (θ . νM1 ) M2 = gM2 (θ . M2 are continuous.

D I Literature: nonparametric instrumental variables models . and E [Y (1) | θ ] Identication of fθ . D] and fθ |M.Nonparametric Identication: Roadmap Goal: E [Y (1) − Y (0) | R] = E {E [Y (1) − Y (0) | θ ] | R} I Inputs: Step 1: fθ |R . E [Y (0) | θ ].M (and consequently fθ |R ) I Input: fM I Literature: nonclassical measurement error models Step 2: Identication of I Inputs: E [Y (0) | θ ] and E [Y (1) | θ ] E [Y | M.

D] and fθ |M. and E [Y (1) | θ ] Identication of fθ .M (and consequently fθ |R ) I Input: fM I Literature: nonclassical measurement error models Step 2: Identication of I Inputs: E [Y (0) | θ ] and E [Y (1) | θ ] E [Y | M. E [Y (0) | θ ].D I Literature: nonparametric instrumental variables models .Nonparametric Identication: Roadmap Goal: E [Y (1) − Y (0) | R] = E {E [Y (1) − Y (0) | θ ] | R} I Inputs: Step 1: fθ |R .

k = 1. 2. variances. 3 µM1 = 0.Parametric Illustration I Suppose the measurement model takes the form Mk = µMk + λMk θ + νMk . and     σ2 0 θ µθ θ 2    0   0 σνM νM1  1  ∼ N .   0 νM2  0   0 νM3 0 0 0 where     0 0  0 0     σν2M 0 2 0 σν2M 3 I Then the unknown parameters can be obtained from the means. and covariances of Details M . λM1 = 1.

2. 0 dier θ 6= θ 00 . fθ . 3.Identication of fθ .M identied from fM under the following assumptions (Hu & Schennach. . and M3 are jointly independent conditional on H such that θ ∈ Θ. fθ |M1 (θ | m1 ) and fM1 |M2 (m1 | m2 ) form (boundedly) complete families of distributions indexed by m1 ∈ M1 and m2 ∈ M2 . 2008): 1. For all for all over a set of strictly positive probability whenever 5.M fθ .M (θ . There exists a known functional   H fM1 |θ (· | θ ) = θ 4. θ ∈ Θ. Θ × M1 × M2 and some dominating All the corresponding marginal and conditional densities are also bounded. M1 .     0 00 0 00 θ . fM3 |θ m3 | θ and fM3 |θ m3 | θ θ. M2 . m) is bounded with respect to the product measure of the Lebesgue measure on measure µ on M3 .

M (and consequently fθ |R ) I Input: √ fM I Literature: nonclassical measurement error models Step 2: Identication of I Inputs: E [Y (0) | θ ] and E [Y (1) | θ ] E [Y | M. E [Y (0) | θ ]. and E [Y (1) | θ ] Identication of fθ .D I Literature: nonparametric instrumental variables models . D] and fθ |M.Nonparametric Identication: Roadmap Goal: E [Y (1) − Y (0) | R] = E {E [Y (1) − Y (0) | θ ] | R} I Inputs: Step 1: fθ |R .

Parametric Illustration I Suppose the latent outcome model takes the form E [Y (D) | θ ] = αD + βD θ . D = 0. Y (1)) ⊥⊥ M | θ I Then the unknown parameters can be obtained from the conditional means of the cuto Details Y and θ given M to the left and right of . 1 and (Y (0) .

< P [D = 1 | θ ] < 1 a. 0 and f θ |M.Identication of E [Y (0) | θ ] and E [Y (1) | θ ] E [Y (0) | θ ] and E [Y (1) | θ ] identied from E [Y | M. . 1 form (boundedly) 0 0 complete families of distributions indexed by m ∈ M and 1 1 m ∈M .s. D] and fθ |M.   f θ |M.D under the following assumptions: 1. 0 3.D θ | m1 .D θ | m0 . Y (1)) ⊥⊥ M | θ . (Y (0) . 2.

Goal:

E [Y (1) − Y (0) | R] = E {E [Y (1) − Y (0) | θ ] | R}

I Inputs:

Step 1:

fθ |R , E [Y (0) | θ ],

and

E [Y (1) | θ ]

Identication of fθ ,M (and consequently fθ |R )

I Input:

fM

I Literature: latent factor and measurement error models

Step 2:

Identication of

I Inputs:

E [Y (0) | θ ]

and

E [Y (1) | θ ]

E [Y | M, D] and fθ |M,D

I Literature: nonparametric instrumental variables models

Extensions

Extension 1: Latent Factor Modeling in a Fuzzy RD Design
Extension 2: Settings with Multiple Latent Factors

Introduction

Latent Factor Modeling in a Sharp RD Design

Boston Exam Schools

Identication and Estimation

Extrapolation Results

Counterfactual Simulations

Placebo Experiments

Conclusions

curriculum.Boston Exam Schools I Three selective public schools spanning grades 7-12 (new students admitted mainly for grades 7 and 9) I I I Boston Latin School Boston Latin Academy John D. resources I Each applicant receives at most one exam school oer I Most preferred school the applicant qualies for I School-specic RD experiments for sharp samples I I Running variable: rank based on GPA and ISEE scores Admissions cuto: lowest rank among admitted students DA Algorithm Sharp Sample . O'Bryant High School of Mathematics and Science I Exam schools dier considerably from traditional BPS I Peers. teachers.

Data and Sample I Data sources: I I I I Exam school application le BPS enrollment le MCAS le ACS 5-year summary le I Sample restrictions: I I I 7th grade applicants in 2000-2004 Enrolled in BPS in 6th grade Non-missing 4th grade MCAS scores and covariates I Outcome: high school MCAS composite score I Standardized average of 10th grade MCAS scores in English and Math Descriptive Statistics .

4 .4 .8 Enrollment Probability .2 Enrollment Probability .4 .6 .2 0 0 .Eects at the Admissions Cutos: First Stage −20 −10 0 10 Running Variable 20 .6 .6 .8 0 .2 Enrollment Probability .8 1 Latin School 1 Latin Academy 1 O’Bryant −20 −10 0 10 Running Variable 20 −20 −10 0 10 Running Variable 20 .

5 0 .5 Mean Score 1 1.5 −.5 0 −.5 Mean Score 1 1.5 2 Latin School 2 Latin Academy 2 O’Bryant −20 −10 0 10 Running Variable 20 −20 −10 0 10 Running Variable 20 .5 1 Mean Score .Eects at the Admissions Cutos: Reduced Form −20 Estimates −10 0 10 Running Variable 20 1.5 0 .5 −.

Introduction Latent Factor Modeling in a Sharp RD Design Boston Exam Schools Identication and Estimation Extrapolation Results Counterfactual Simulations Placebo Experiments Conclusions .

GPA. sociodemographic characteristics Measurement Scatterplots Measurement Correlations .Setup I Specify a model with two latent factors: I I English ability Math ability I School-specic running variables are functions of ISEE scores I I Reading and Verbal: Noisy measures of English ability Math and Quantitive: Noisy measures of Math ability I Data also contains 4th grade MCAS scores I I English: Noisy measure of English ability Math: Noisy measure of Math ability I Control for a set of additional covariates I Application year. application preferences.

X ] = αDs (z) X + βDEs (z) θE + βDMs (z) θM 0 E [Y (S (z)) | θ . X MkM | θ . X 0 µ θE X 0 µθM X #  . X ] = αY (S(z)) X + βYE (S(z)) θE + βYM(S(z)) θM I Inference: 5-step bootstrap . exp γM M + δM M θM "  |X MkE | θ .Estimation I Measurement model: Maximum Simulated Likelihood  θE θM ! σ θE θM σθ2E ∼ N σ θE θM σθ2M  2   0 ∼ N µM E X + λM E θE . exp γM E + δM E θE k k k k   2  0 ∼ N µM M X + λM M θM . k k k k I Latent outcome model: Method of Simulated Moments 0 E [Ds (z) | θ .

Introduction Latent Factor Modeling in a Sharp RD Design Boston Exam Schools Identication and Estimation Extrapolation Results Counterfactual Simulations Placebo Experiments Conclusions .

Preliminaries Figure: Distribution of English and Math ability Table: Factor loadings on ISEE and MCAS scores Table: Dependence of outcomes on English and Math ability .

Eects Away from the Admissions Cutos −80 −40 0 Running Variable Estimates 40 80 3 2 −1 0 1 Mean Score 2 −1 0 1 Mean Score 2 1 −1 0 Mean Score Latin School 3 Latin Academy 3 O’Bryant −80 −40 0 Running Variable 40 80 −80 −40 0 Running Variable 40 80 .

Eects Away from the Admissions Cutos −80 −40 0 Running Variable Estimates 40 80 3 2 −1 0 1 Mean Score 2 −1 0 1 Mean Score 2 1 −1 0 Mean Score Latin School 3 Latin Academy 3 O’Bryant −80 −40 0 Running Variable 40 80 −80 −40 0 Running Variable 40 80 .

Exam School vs Traditional BPS I Previous results about incremental eects of attending a better exam school I Latent factor model provides counterfactual predictions for all exam schools I This allows one to study the eects of attending a given exam school versus a traditional BPS .

024 (0.020) Latin Academy (2) 0. *** significant at 1% .064) LATE 0.058) 0.049) 0.021 (0.754*** (0.037) 0.Exam Schoolextrapolation_all vs Traditional BPS: All Applicants O'Bryant (1) 0.948*** (0.027 (0.016) Reduced Form 0.025 (0.968*** (0.052 (0.013) Latin School (3) 0.062) 0.704 * significant at 10%.049 (0.066) First Stage N 3. ** significant at 5%.

428*** (0.987*** 0.110) 0.927 .353*** (0.048) -0. *** significant at 1% Updated: 2014-01-13 1.053) N 1.300*** (0.053) -0.098) -0.026) Exam School Offer Latin Latin O'Bryant Academy School (4) (5) (6) 0.066) -0.094) -0.448*** (0.887*** 0.955*** (0.758*** 0.102) 0.325*** (0.030) (0.088) 0.435*** (0.052) LATE 0.005) Reduced Form 0.777 * significant at 10%.087) -0.Exam School vs Traditional BPS: By Oer Status extrapolation_atu_att First Stage No Exam School Offer Latin Latin O'Bryant Academy School (1) (2) (3) 0.348*** (0.925*** 0.376*** (0.353*** (0.334*** (0.017) (0.268*** (0.010) (0.105) 0.429*** (0.049) (0.987*** (0. ** significant at 5%.

088) 0.887*** 0.052) LATE 0.049) (0.017) (0.987*** 0.026) Exam School Offer Latin Latin O'Bryant Academy School (4) (5) (6) 0.268*** (0. ** significant at 5%.094) -0.053) N 1. *** significant at 1% Updated: 2014-01-13 1.955*** (0.053) -0.353*** (0.353*** (0.005) Reduced Form 0.098) -0.334*** (0.Exam School vs Traditional BPS: By Oer Status extrapolation_atu_att First Stage No Exam School Offer Latin Latin O'Bryant Academy School (1) (2) (3) 0.925*** 0.102) 0.066) -0.435*** (0.376*** (0.758*** 0.429*** (0.110) 0.105) 0.087) -0.777 * significant at 10%.987*** (0.448*** (0.927 .048) -0.348*** (0.428*** (0.010) (0.030) (0.325*** (0.300*** (0.

325*** (0.053) -0.268*** (0.098) -0.048) -0.030) (0.927 .758*** 0.094) -0.987*** (0.348*** (0.088) 0.053) N 1.955*** (0.887*** 0.987*** 0.777 * significant at 10%.087) -0.110) 0.010) (0.448*** (0.429*** (0.435*** (0.925*** 0.334*** (0.353*** (0.102) 0.376*** (0.026) Exam School Offer Latin Latin O'Bryant Academy School (4) (5) (6) 0.353*** (0.005) Reduced Form 0.049) (0.Exam School vs Traditional BPS: By Oer Status extrapolation_atu_att First Stage No Exam School Offer Latin Latin O'Bryant Academy School (1) (2) (3) 0.052) LATE 0. *** significant at 1% Updated: 2014-01-13 1.017) (0.300*** (0. ** significant at 5%.428*** (0.105) 0.066) -0.

Introduction Latent Factor Modeling in a Sharp RD Design Boston Exam Schools Identication and Estimation Extrapolation Results Counterfactual Simulations Placebo Experiments Conclusions .

teacher behavior. Introducing minority preferences I I 2. (Boston 1975-1998) 65% of seats assigned among all applicants 35% of seats assigned among blacks and Hispanics Introducing socioeconomic preferences I I (Chicago 2010-) 30% of seats assigned among all applicants 70% of seats assigned within four SES tiers I Exam school assignment of 27-35% of applicants aected I I I Table: Actual and counterfactual assignments Table: Counterfactual admissions cutos Table: Descriptives by counterfactual assignment I Caveats (outside the scope of this paper): I Application behavior.Counterfactual Simulations I Simulate eects of two admissions reforms: 1. peer eects .

044* (0.023) 223 0.016) 1.004) (0.074*** 0. ** significant at 5%.005) 737 876 901 SES Tier 4 (8) 0.007) (0.Eect of Amative Action on Average Achievement cfsim12_effects All Applicants Minority Preferences Applicant Group All NonApplicants Minority Minority (1) (2) (3) 0.024) 208 .068** (0.100 * significant at 10%.050*** (0.062*** 0.024** 0.015*** (0.017* (0.704 2.006) (0.704 0.005) 3. *** significant at 1% Updated: 2014-01-13 Socioeconomic Preferences Applicant Group SES SES SES Tier 1 Tier 2 Tier 3 (5) (6) (7) 0.086 1.019) (0.018* (0.011* 0.010) 1.025) (0.029) 404 0.011) (0.190 0.011) (0.046* 0.367 754 613 1.026 (0.054* (0.023*** 0.027*** 0.618 Affected Applicants All Applicants (4) 0.031) 265 0.009) 3.006 (0.

010) 1.Eect of Amative Action on Average Achievement cfsim12_effects All Applicants Minority Preferences Applicant Group All NonApplicants Minority Minority (1) (2) (3) 0.026 (0.007) (0.011) (0.618 Affected Applicants All Applicants (4) 0.074*** 0.018* (0.367 754 613 1.011) (0.086 1. ** significant at 5%.054* (0.190 0.016) 1.009) 3.006 (0.704 0.704 2.025) (0.044* (0.050*** (0.005) 3.031) 265 0. *** significant at 1% Updated: 2014-01-13 Socioeconomic Preferences Applicant Group SES SES SES Tier 1 Tier 2 Tier 3 (5) (6) (7) 0.006) (0.004) (0.005) 737 876 901 SES Tier 4 (8) 0.015*** (0.046* 0.017* (0.024** 0.024) 208 .023*** 0.011* 0.062*** 0.068** (0.029) 404 0.100 * significant at 10%.019) (0.027*** 0.023) 223 0.

*** significant at 1% Updated: 2014-01-13 Socioeconomic Preferences Applicant Group SES SES SES Tier 1 Tier 2 Tier 3 (5) (6) (7) 0.044* (0.023*** 0.068** (0.015*** (0.Eect of Amative Action on Average Achievement cfsim12_effects All Applicants Minority Preferences Applicant Group All NonApplicants Minority Minority (1) (2) (3) 0.704 2.190 0.024** 0.025) (0.006 (0.005) 3.074*** 0.011) (0.062*** 0.050*** (0.011) (0.010) 1.618 Affected Applicants All Applicants (4) 0.005) 737 876 901 SES Tier 4 (8) 0.017* (0.027*** 0.019) (0.007) (0.009) 3.016) 1.018* (0.704 0.100 * significant at 10%.031) 265 0.029) 404 0.046* 0.006) (0.023) 223 0.054* (0.367 754 613 1.086 1.024) 208 .011* 0. ** significant at 5%.004) (0.026 (0.

Introduction Latent Factor Modeling in a Sharp RD Design Boston Exam Schools Identication and Estimation Extrapolation Results Counterfactual Simulations Placebo Experiments Conclusions .

Split data on applicants with assignment Z = 0.Placebo Experiments Strategy: 1. 1. Repeat reduced form extrapolations to the left and right of the placebo cutos . 2. Re-estimate latent outcome models to the left and right of the placebo cutos 3. 3 in half based on the median of the running variable 2.

Placebo Extrapolation: Figures 2 0 25 50 −50 −25 0 Running Variable Running Variable Latin Academy Latin School 25 50 25 50 2 1 0 −1 −1 0 1 Mean Score 2 3 −25 3 −50 Mean Score 1 Mean Score −1 0 1 −1 0 Mean Score 2 3 O’Bryant 3 No Offer −50 −25 0 25 50 Running Variable −50 −25 0 Running Variable Fitted Extrapolated .

008 (0.054 -0.064) 1. ** significant at 5%.133 -0.056) (0. *** significant at 1% .041 (0.Placebo Extrapolation: Estimates placebo All Applicants Below Placebo Cutoff No Offer (1) -0.028 (0.080 (0.044 Placebo (0.133 (0.049) 368 Above 0.068) 563 Latin Academy (3) -0.020 0.099) 887 0.068) Cutoff 890 283 314 371 * significant at 10%.027 (0.777 O'Bryant (2) 0.057 (0.062) 625 Latin School (4) -0.027 (0.097) (0.045) 793 -0.102) (0.055) 311 0.061) 280 -0.

Introduction Latent Factor Modeling in a Sharp RD Design Boston Exam Schools Identication and Estimation Extrapolation Results Counterfactual Simulations Placebo Experiments Conclusions .

Conclusions I Develop a latent factor-based approach to the extrapolation of treatment eects away from the cuto in RD I Achievement gains from exam school attendance concentrated among lower-scoring applicants I Armative action predicted to increase average achievement among aected applicants I Latent factor-based extrapolation likely to be a promising approach also in various other RD designs .

Identication with Less than Three Measures I One noisy measure enough for the identication of fθ . νM1 . 2012) Back . Evdokimov & White.M if I I I M1 = θ + νM1 θ and νM1 independent either fθ or fνM1 known (Carroll & Hall.M if I I M1 = θ + νM1 . 1988. and νM2 jointly independent (Kotlarski. 1967. 2011) I Two noisy measures enough for the identication of fθ . Carrasco & Florens. M2 = θ + νM2 θ .

Mk ] = λMk σθ2 .Parametric Illustration: Details E [M1 ] = µθ E [Mk ] = µMk + λMk µθ . k = 2. 3 Cov [M1 . 3 k Back . M3 ] = λM2 λM3 σθ2 Var [M1 ] = σθ2 + σν2M 2 1 Var [Mk ] = λMk σθ + σν2M . k = 2. 3 Cov [M2 . k = 2.

D = 1 Back . D = 1     E Y | M = m21 . D = 0     E Y | M = m20 . D = 0     E Y | M = m11 . D = 1 = α1 + β1 E θ | M = m11 . D = 0 = α0 + β0 E θ | M = m10 .Parametric Illustration: Details     E Y | M = m10 . D = 1 = α1 + β1 E θ | M = m21 . D = 0 = α0 + β0 E θ | M = m20 .

Y (1) .s. 3.Extrapolation of LATE in Fuzzy RD Suppose 1. (Y (0) . P [D (1) > D (0) | θ ] > 0 a. D (0) .s. D (1)) ⊥⊥ R | θ 2. P [D (1) ≥ D (0) | θ ] = 1 a. Then E [Y (1) − Y (0) | D (1) > D (0) . R = r ] E {E [Y (D (1)) − Y (D (0)) | θ ] | R = r } = E {E [D (1) − D (0) | θ ] | R = r } for all Back r ∈R .

K . K 2 2 M3 = gW (θ1 . . . . θK : M1k M2k   = gM k θk . . νM k . . . M3 potentially discrete is a deterministic function of a subset of M = (M1 .Settings with Multiple Latent Factors I Suppose the data contains 2 × K factors +1 noisy measures of latent θ1 . . M3 ) and continuous. . . νM k . . θK . νM3 ) M1k I R M2k . . . K 1 1   = gM k θk . I Extending all of the identication results to this setting requires only slight modications Back . . k = 1. . . M2 . . k = 1. k = 1. . . . .

I Round k > 1: Applicants rejected in Round k −1 are considered for a seat in their next most preferred exam school.Deferred Acceptance Algorithm I Round 1: Applicants are considered for a seat in their most preferred exam school. The rest of the applicants are provisionally admitted. I The algorithm terminates once either each applicants are assigned an oer from one of the exam schools or all unmatched applicant are rejected by every exam school in their preference ordering. Each exam schools considers these applicants together with the provisionally admitted applicants from Round k −1 and rejects the lowest-ranking students in excess of its capacity. Back . Each exam schools rejects the lowest-ranking applicants in excess of its capacity. The rest of the students are provisionally admitted.

3. and she clears the admissions cuto. The applicant does not clear the admissions cuto for her 1st or 2nd choice. and she clears the admissions cuto. The applicant does not clear the admissions cuto for her 1st choice. and she clears the admissions cuto.Sharp Sample I Three ways for an applicant to be admitted to exam school s (given the school-specic cutos) 1. exam school s is her 3rd choice. 2. Exam school s is the applicant's 1st choice. exam school s is her 2nd choice. I This forms the basis for the denition of a sharp sample for each exam school I I Back Applicants who obtain an oer from exam school s i they rank higher than the school-specic cuto A given applicant can be in multiple sharp samples .

523 0.387 0.073 0.755 0.399 0.749 0.189 0.000 0.870 1.109 0.788 0.516 0.420 0.180 0.043 0.776 0.000 0.009 English 4 0.489 0.396 0.791 755 790 843 Updated: 2013-11-08 Back .716 0.259 0.822 0.Descriptive Statistics descriptives Exam School Assignment All All No BPS Applicants Offer O'Bryant Latin Latin Academy School (1) (2) (3) (4) (5) (6) Female 0.412 SPED 0.579 0.581 0.116 0.196 0.223 0.033 0.577 Black 0.749 0.004 Bilingual 0.545 0.353 0.081 FRPL 0.265 0.516 0.114 N 21.315 0.870 1.123 Hispanic 0.858 Math 4 0.212 1.094 5.499 LEP 0.179 2.206 0.275 2.073 0.251 0.227 0.451 0.009 0.006 0.064 0.

3 English 3 . −2 0 MCAS Math 4 2 4 −2 0 MCAS Math 4 2 4 . 2 4 2 1 −2 −1 0 ISEE Quantitative 2 1 ISEE Math −2 −1 0 2 1 0 −1 0 ISEE Quantitative 4 3 MCAS English 4 . 3 ISEE Verbal −2 −2 0 ISEE Verbal 2 1 −2 −1 0 ISEE Reading 2 1 0 ISEE Reading −1 −2 0 3 −2 ISEE Math .Measurement Scatterplots 2 4 3 2 1 −1 −2 −2 0 2 4 −2 0 2 MCAS English 4 Math .

2 0 −2 MCAS English 4 4 Measurement Scatterplots (cont'd) −2 0 2 MCAS Math 4 Back 4 .

598 0.713 Math 4 0.587 Math 0.000 Panel B: MCAS N Updated: 2013-11-08 Back 5.598 0.740 Quantitative 0.655 0.670 0.735 0.581 0.619 0.587 0.581 Verbal 0.570 1 0.631 0.713 1.621 0.617 0.740 0.655 0.621 0.845 0.179 .Measurement Correlations measurement_correlations ISEE MCAS Reading Verbal Math Quantitative English 4 Math 4 (1) (2) (3) (4) (5) (6) Panel A: ISEE Reading 1 0.670 0.631 0.617 0.619 1 0.718 0.570 0.735 1 0.718 English 4 0.845 1 0.

052) LATE 0.054) 1.018) Reduced 0.086 Form (0.055) (0.022 -0.044) (0.955*** 0.060 -0.475 1.781*** 0.018) (0.047 -0. ** significant at 5%.034) (0.Eects at therd_estimates_all Admissions Cutos: Estimates O'Bryant Latin Latin Academy School (1) (2) (3) First 0.046) (0. *** significant at 1% Back .089 (0.070) (0.964*** Stage (0.021 -0.999 907 N * significant at 10%.

6 Marginal Distribution of English Ability −2 0 2 English Ability 4 .4 .0 Marginal Density .2 .

1 Marginal Density .3 .2 .5 Marginal Distribution of Math Ability −2 0 2 Math Ability 4 .0 .4 .

−1 0 English Ability 1 2 3 4 Scatterplot of English and Math Ability −2 0 2 Math Ability Back 4 .

016) 0. *** significant at 1% Updated: 2014-01-13 Back 0.179 * significant at 10%.027) (0.033) θM θE 0.035) 1.Factor Loadings on ISEE and MCAS Scores factor_loadings ISEE Reading (1) θE 1.160*** (0.018) (0.020) θM MCAS Math Quantitative English 4 Math 4 (3) (4) (5) (6) Panel A: Factor Loading on Mean 1.124*** (0.028) Verbal (2) Panel B: Factor Loading on (Log) Standard Deviation 0.016 (0.019 -0.015) .013 (0.135*** 1.018) (0.017) N 5.180*** 1 (0.152*** 0. ** significant at 5%.081*** (0.119*** 1 (0.

008) Latin O'Bryant Academy (2) (3) Panel A: First Stage -0.057) 0.Dependencelatent_fs_rf of Outcomes on English and Math Ability No Offer (1) θE θM θE θM -0.604*** (0.346*** 0.446*** (0.062) 0.239*** (0.777 625 793 * significant at 10%.067) (0.044) 563 N 1.267*** (0.282*** 0.009) 0.012) 0.030) Latin School (4) 0.066) 0.003 (0.012 (0.059 0.102* (0. ** significant at 5%.070) (0.046) 0.027** (0.011) -0.056) Panel B: Reduced Form 0.053) -0.025 -0.158*** (0.012 (0.076) (0. *** significant at 1% Back .217*** (0.001 (0.061) (0.

068) 0.293*** (0.209*** (0.962*** (0.018) Reduced Form 0.340 * significant at 10%.760 3.858*** (0.021) Latin School (3) 0.252*** (0. ** significant at 5%.950*** (0.290*** (0.038) Latin Academy (2) 0.279*** (0.Estimates: Allrd_extrapolation_all Applicants First Stage O'Bryant (1) 0. *** significant at 1% .061) LATE 0.075) 0.081) 0.079) 0.199*** (0.240 2.064) N 2.

367*** (0.092) 0.091* (0.005 (0.079) -0.099* (0.054) LATE 0.892*** 0.075) -0.362*** (0.Estimates: Below/Above Admissions Cutos rd_extrapolation_cutoff First Stage Below Admissions Cutoff Latin Latin O'Bryant Academy School (1) (2) (3) 0.984*** 0.018) (0.021) Above Admissions Cutoff Latin Latin O'Bryant Academy School (4) (5) (6) 0.046) -0.749*** 0. *** significant at 1% Updated: 2014-01-13 Back .906*** (0.959*** (0.891*** 0.097) 0.089) 0.021 (0.027) (0.601 563 * significant at 10%.343*** (0.135 2.016) (0.006 (0.045) -0. ** significant at 5%.677 2.031) Reduced Form 0.102) 0.034) -0.281*** (0.293*** (0.050) (0.385*** (0.053) -0.028 (0.059) 625 739 N 1.

Educational attainment score I Each component is turned into a percentile and added up to get the socioeconomic index I BPS students divided into four tiers based on the quartiles of the socioeconomic index distribution Back .Socioeconomic Tiers I Census tracts are given a socioeconomic index based on 1. 4. 2. 3. Median family income Percent of households occupied by the owner Percent of families headed by a single parent Percent of households where a language other than English is spoken 5.

Observedcfsim12_assignments and Counterfactual Assignments Actual Assignment No Counterfactual Assignment Offer (1) Latin Latin O'Bryant Academy School (2) (3) (4) Panel A: Minority Preferences No Offer 2418 221 113 39 O'Bryant 280 129 133 213 Latin Academy 88 389 268 45 Latin School 5 16 276 546 Panel B: Socioeconomic Preferences No Offer 2579 159 39 O'Bryant 14 203 319 146 87 Latin Academy 9 106 403 272 Latin School 0 171 202 470 Back Updated: 2013-11-08 .

9 SES Tier 2 -6.1 -20.8 12.4 7.1 Notes: This table reports the differences between the actual Back .2 -17.1 SES Tier 3 -2.6 -11.5 -7.6 -31.1 -5.7 -16.8 Panel B: Socioeconomic Preferences SES Tier 1 -20.cfsim12_cutoffs Counterfactual Admissions Cutos Latin Latin O'Bryant Academy School (1) (2) (3) Panel A: Minority Preferences Minority -14.4 -26.9 Non-Minority 15.0 2.4 -32.1 SES Tier 4 8.

277 0.380 0.309 1.058 0.514 0.274 Hispanic 0.Descriptives by Counterfactual Assignment cfsim12_descriptives No Offer (1) Latin Latin O'Bryant Academy School (2) (3) (4) Panel A: Minority Preferences Female 0.556 LEP 0.673 0.781 Female 0.396 0.813 0.030 0.013 0.810 0.143 0.365 1.107 0.006 English 4 0.386 0.006 0.771 0.360 0.018 Bilingual 0.502 0.146 1.597 0.575 Black 0.009 0.220 0.744 N 2.073 0.116 0.981 1.391 SPED 0.182 0.203 Hispanic 0.588 0.359 0.073 0.423 0.289 1.572 Black 0.261 1.728 0.445 SPED 0.012 0.022 0.277 0.668 Math 4 0.120 FRPL 0.005 English 4 0.386 0.217 1.440 0.030 0.715 0.017 Bilingual 0.499 0.067 1.203 0.624 LEP 0.295 0.430 0.963 1.629 0.191 0.223 0.572 0.215 1.556 Math 4 0.391 0.791 755 790 843 Panel B: Minority Preferences Back Updated: 2013-11-08 .172 FRPL 0.