Analyzing Regression-Discontinuity Designs with Multiple Assignment Variables: A Comparative Study of Four Estimation Methods Vivian C. Wong Northwestern.

Slides:



Advertisements
Similar presentations
C82MST Statistical Methods 2 - Lecture 2 1 Overview of Lecture Variability and Averages The Normal Distribution Comparing Population Variances Experimental.
Advertisements

Review bootstrap and permutation
REGRESSION, IV, MATCHING Treatment effect Boualem RABTA Center for World Food Studies (SOW-VU) Vrije Universiteit - Amsterdam.
Regression Discontinuity. Basic Idea Sometimes whether something happens to you or not depends on your ‘score’ on a particular variable e.g –You get a.
ASSESSING RESPONSIVENESS OF HEALTH MEASUREMENTS. Link validity & reliability testing to purpose of the measure Some examples: In a diagnostic instrument,
T-tests continued.
Markov Chain Monte Carlo Convergence Diagnostics: A Comparative Review By Mary Kathryn Cowles and Bradley P. Carlin Presented by Yuting Qi 12/01/2006.
Regression-Discontinuity Design
Regression Discontinuity Design Thanks to Sandi Cleveland and Marc Shure (class of 2011) for some of these slides.
Introduction of Regression Discontinuity Design (RDD)
The World Bank Human Development Network Spanish Impact Evaluation Fund.
Lecture 9: One Way ANOVA Between Subjects
Regression Discontinuity (RD) Andrej Tusicisny, methodological reading group 2008.
Chapter 14 Simulation. Monte Carlo Process Statistical Analysis of Simulation Results Verification of the Simulation Model Computer Simulation with Excel.
One-way Between Groups Analysis of Variance
Today Concepts underlying inferential statistics
Multiple Linear Regression A method for analyzing the effects of several predictor variables concurrently. - Simultaneously - Stepwise Minimizing the squared.
Hypothesis Testing. Outline The Null Hypothesis The Null Hypothesis Type I and Type II Error Type I and Type II Error Using Statistics to test the Null.
Larysa Minzyuk - Felice Russo Department of Management and Economics - University of Salento (Lecce)
Analysis of Variance. ANOVA Probably the most popular analysis in psychology Why? Ease of implementation Allows for analysis of several groups at once.
Psy B07 Chapter 1Slide 1 ANALYSIS OF VARIANCE. Psy B07 Chapter 1Slide 2 t-test refresher  In chapter 7 we talked about analyses that could be conducted.
AM Recitation 2/10/11.
© 2013 Cengage Learning. All Rights Reserved. May not be scanned, copied or duplicated, or posted to a publicly accessible website, in whole or in part.
Non Experimental Design in Education Ummul Ruthbah.
Stats Lunch: Day 7 One-Way ANOVA. Basic Steps of Calculating an ANOVA M = 3 M = 6 M = 10 Remember, there are 2 ways to estimate pop. variance in ANOVA:
T tests comparing two means t tests comparing two means.
Copyright © 2012 Wolters Kluwer Health | Lippincott Williams & Wilkins Chapter 17 Inferential Statistics.
Chapter 21 Univariate Statistical Analysis © 2010 South-Western/Cengage Learning. All rights reserved. May not be scanned, copied or duplicated, or posted.
Education Research 250:205 Writing Chapter 3. Objectives Subjects Instrumentation Procedures Experimental Design Statistical Analysis  Displaying data.
Statistical Decision Making. Almost all problems in statistics can be formulated as a problem of making a decision. That is given some data observed from.
Slide 1 Estimating Performance Below the National Level Applying Simulation Methods to TIMSS Fourth Annual IES Research Conference Dan Sherman, Ph.D. American.
Slide 13-1 Copyright © 2004 Pearson Education, Inc.
Propensity Score Matching and Variations on the Balancing Test Wang-Sheng Lee Melbourne Institute of Applied Economic and Social Research The University.
Some terms Parametric data assumptions(more rigorous, so can make a better judgment) – Randomly drawn samples from normally distributed population – Homogenous.
Quantification of the non- parametric continuous BBNs with expert judgment Iwona Jagielska Msc. Applied Mathematics.
Statistical analysis Prepared and gathered by Alireza Yousefy(Ph.D)
Assumptions of value-added models for estimating school effects sean f reardon stephen w raudenbush april, 2008.
Estimating Causal Effects from Large Data Sets Using Propensity Scores Hal V. Barron, MD TICR 5/06.
Session III Regression discontinuity (RD) Christel Vermeersch LCSHD November 2006.
Inference and Inferential Statistics Methods of Educational Research EDU 660.
CAUSAL INFERENCE Presented by: Dan Dowhower Alysia Cohen H 615 Friday, October 4, 2013.
Using A Regression Discontinuity Design (RDD) to Measure Educational Effectiveness: Howard S. Bloom
AFRICA IMPACT EVALUATION INITIATIVE, AFTRL Africa Program for Education Impact Evaluation David Evans Impact Evaluation Cluster, AFTRL Slides by Paul J.
Educational Research Chapter 13 Inferential Statistics Gay, Mills, and Airasian 10 th Edition.
CROSS-VALIDATION AND MODEL SELECTION Many Slides are from: Dr. Thomas Jensen -Expedia.com and Prof. Olga Veksler - CS Learning and Computer Vision.
Statistics as a Tool A set of tools for collecting, organizing, presenting and analyzing numerical facts or observations.
Randomized Assignment Difference-in-Differences
Learning Objectives After this section, you should be able to: The Practice of Statistics, 5 th Edition1 DESCRIBE the shape, center, and spread of the.
Chapter 13 Understanding research results: statistical inference.
Rerandomization to Improve Covariate Balance in Randomized Experiments Kari Lock Harvard Statistics Advisor: Don Rubin 4/28/11.
HYPOTHESIS TESTING FOR DIFFERENCES BETWEEN MEANS AND BETWEEN PROPORTIONS.
Regression Discontinuity Design Case Study : National Evaluation of Early Reading First Peter Z. Schochet Decision Information Resources, Inc.
Chapter 7: Hypothesis Testing. Learning Objectives Describe the process of hypothesis testing Correctly state hypotheses Distinguish between one-tailed.
Educational Research Inferential Statistics Chapter th Chapter 12- 8th Gay and Airasian.
Alexander Spermann University of Freiburg, SS 2008 Matching and DiD 1 Overview of non- experimental approaches: Matching and Difference in Difference Estimators.
Chapter 22 Inferential Data Analysis: Part 2 PowerPoint presentation developed by: Jennifer L. Bellamy & Sarah E. Bledsoe.
NURS 306, Nursing Research Lisa Broughton, MSN, RN, CCRN RESEARCH STATISTICS.
Statistical Decision Making. Almost all problems in statistics can be formulated as a problem of making a decision. That is given some data observed from.
Hypothesis Testing.
Lezione di approfondimento su RDD (in inglese)
An Empirical Test of the Regression Discontinuity Design
Chapter 8 Experimental Design The nature of an experimental design
Research Design & Analysis II: Class 10
12 Inferential Analysis.
Stochastic Hydrology Hydrological Frequency Analysis (II) LMRD-based GOF tests Prof. Ke-Sheng Cheng Department of Bioenvironmental Systems Engineering.
Main Effects and Interaction Effects
12 Inferential Analysis.
Impact Evaluation Methods: Difference in difference & Matching
Evaluating Impacts: An Overview of Quantitative Methods
Explanation of slide: Logos, to show while the audience arrive.
Presentation transcript:

Analyzing Regression-Discontinuity Designs with Multiple Assignment Variables: A Comparative Study of Four Estimation Methods Vivian C. Wong Northwestern University

The Regression-Discontinuity Design (RDD) A visual depiction Comparison

RDD Visual Depiction ComparisonTreatment

RDD Visual Depiction ComparisonTreatment Counterfactual regression line Discontinuity, or treatment effect

Two Rationales for the Validity of the RDD 1.Selection process is completely known and can be modeled through a regression line of the assignment and outcome variables – Untreated portion of the AV serves as a counterfactual – Use parametric approach to estimate treatment effects 2.It is like an experiment around the cutoff – Observations just to the left and right of cutoff are exchangeable – Use non-parametric approach to estimate treatment effects

Required Assumptions for a Valid RD Design Assumptions 1.Probability of treatment receipt must be discontinuous at the cutoff 2.No discontinuity in potential outcomes in the cutoff (often referred to as the “continuity restriction”) Threats to Design Assumptions 1.Overrides to the cutoff (“fuzzy” discontinuity) – Can address by using the assignment variable and cutoff as an instrumental variable for treatment receipt 2.Manipulation of the assignment scores – No solution but can probe the data to assess whether manipulation occurred

RD in Recent Education Evaluation Studies Class size (Angrist & Lavy, 1999) State pre-kindergartens (Gormley & Phillips, 2005; Wong, Cook, Barnett, & Jung, 2008) Head Start (Ludwig & Miller, 2006)

Regression-Discontinuity Designs with Multiple Assignment Variables

Distribution of Units in an RD with Two Assignment Variables A visual depiction

Multivariate RDD with Two Assignment Variables A visual depiction

τRτR τMτM τ MRD Multivariate RDD with Two Assignment Variables Treatment effects estimated

The Average Treatment Effect along the Cutoff Frontier (τ MRD ) τ MRD is the weighted average of conditional expectations given the single frontiers F R and F M : Where G i is the average size of the discontinuity at the R and M cutoff frontiers, and f(r,m) is the joint density function for assignment variables R and M.

Frontier-specific Effect (τ R ) Where g(r, m) is the treatment function for the R frontier along the M assignment variable, and f r (r i = r c, m) is the conditional density function for the F R. To get the conditional expectation F R, we integrate the treatment function with the conditional density function along F R. Note that no weights are needed because there is no pooling of treatment effects across F R and F M. Average treatment effect for the M frontier is calculated in a similar with corresponding treatment and density functions.

Treatment Weights for τ MRD Weights w r and w m reflect the probabilities for observing a subject at the R- or M-frontiers However, note that weights are sensitive to the scaling and distribution of the assignment variables.

Requirements for a valid Multivariate RDD Similar to RD case with single assignment mechanism 1.A discontinuity in probability of treatment receipt across the frontier; 2.Continuity in the expectations of potential outcomes at F R and F M.

Recent Education Examples of RDDs with Multiple Assignment Variables College financial aid offer (van der Klaauw, 2002; Kane, 2003) Remedial education (Jacob & Lefgren, 2004a) Teacher professional development (Jacob & Lefgren, 2004b) High school exit exams (Martorell, 2005; Papay et al. 2010) No Child Left Behind (Gill et al., 2007)

Estimating Treatment Effects Four Proposed Approaches 1.Frontier approach 2.Centering approach 3.Univariate approach 4.IV Approach

Frontier Approach Estimates the discontinuity along each frontier simultaneously, and applies appropriate weights to obtain the overall effect. First, estimate the treatment function, which is the average size of the discontinuity along the cutoff frontiers using parametric, semi- parametric, or non-parametric approaches. Second, estimate the joint density function by using a bivariate kernel density estimator or by estimating conditional density functions for R and M separately for observations that lie within a narrow bandwidth around the frontier. Third, numerically integrate the product of treatment and joint density functions at the cutoff frontiers to obtain conditional expectations across both frontiers. Third, apply appropriate treatment weights to each discontinuity frontier. Estimates τ MRD, τ M, τ R

Multivariate RDD with Two Assignment Variables A visual depiction

Centering Approach Procedure allows researcher to address the “curse of dimensionality” issue by collapsing multiple assignment scores for unit i to a single assignment variable. First, for each unit i, center assignment variables r and m to their respective cutoffs, that is r i – r c and m i – m c. Second, choose the minimum centered value z i = min(r i – r c, m i – m c ) is chosen as the unit’s sole assignment score. Third, pool units and analyze as a standard RD design with z as the single assignment variable. Estimates τ MRD

Multivariate RDD with Two Assignment Variables A visual depiction

Univariate Approach Addresses dimensionality problem by estimating treatment effects for each frontier separately. First, exclude all observations with r values less than its respective cutoff (r c ), and choosing a single assignment variable (say, m) and cutoff (m c ). Second, estimate treatment effects by measuring size of discontinuity of the conditional outcomes at the cutoff for the designated assignment variable using parametric, semi-parametric, or non- parametric approaches Estimates τ R or τ M

Multivariate RDD with Two Assignment Variables A visual depiction

IV Approach (1) Rather than exclude observations assigned to treatment by alternative mechanisms, delegate these cases as “fuzzy” units. First, designate a treatment assignment mechanism serves as the instrument for treatment receipt. Second, estimate the local average treatment effect, which is the difference in conditional mean outcomes for treatment and comparison groups divided by the difference in treatment receipt rates for both groups within a neighborhood around the cutoff.

Multivariate RDD with Two Assignment Variables A visual depiction

IV Approach (2) Continuous potential outcomesDiscontinuous potential outcomes Estimates the local average treatment effect along the R cutoff. Requires continuous potential outcomes. Estimates τ R-IV or τ M-IV

Monte Carlo Study Wong, Steiner, and Cook (2010) examines the performance of the four approaches when the following factors are varied: 1.Complexity of the true response surface 2.Distribution and scale of the assignment variables 3.Methodological approach (frontier, centering, univariate, and IV) for analyzing MRDDs Simulations based on 500 replications with a sample size of 5,000 for each repetition.

Results from Simulation Study (1) In general, all four approaches replicated the theoretical true effects when their analytic assumptions are met The frontier approach produced unbiased effects for τ MRD, τ R, and τ M when the treatment function was correctly modeled The univariate approach produced unbiased effects for τ R and τ M when the functional form of the response function was correctly specified

Results from Simulation Study (2) The centering approach was prone to producing small but significant biases for τ MRD – Pooling units from different frontier increases heterogeneity in the outcome, which requires larger bandwidths for the nonparametric estimates and increases the complexity of the response function The IV approach produced biased effects when potential outcomes are discontinuous along either cutoff frontier – This is most likely to happen to when the different cutoffs result in heterogeneous treatments The frontier approach produced the most efficient effect estimates and the IV the least efficient (but this need not always be the case).

Implications for Practice Which approach to use? – Use univariate approach first for estimating τ R and τ M and to assess whether treatment effects are constant. – If treatment effects are constant, use frontier or centering approach for estimating τ MRD. Use frontier approach only if functional form of the response surface is known For the centering approach, may reduce heterogeneity in the outcome by using difference scores for the outcome – The IV approach is not recommended because we do not know when the potential outcomes are discontinuous and because of reduced efficiency

Vivian C. Wong Northwestern University Working draft of paper available at the Institute for Policy Research website (working paper WP-10-02) ers/wpabstracts10/wp1002.html

Extra Slides

MRDD with Two Assignment Variables R and M

Standardize or not standardize assignment variables? Scale-dependency of the joint density and weights at the cutoff frontier: By rescaling R such that the ratio of weights— represented by the ratio of the two areas along the frontier— changes. Implications for Practice (2)

In MRD designs, treatment contrast is limited to comparisons along the cutoff frontier But this may not be the treatment contrast of interest … Implications for Practice (3)