Issues in Selecting Covariates for Propensity Score Adjustment William R Shadish University of California, Merced.

Slides:



Advertisements
Similar presentations
Value Added in CPS. What is value added? A measure of the contribution of schooling to student performance Uses statistical techniques to isolate the.
Advertisements

Research Skills Workshop Designing a Project
Increasing your confidence that you really found what you think you found. Reliability and Validity.
Introduction to Regression with Measurement Error STA431: Spring 2015.
Regression Discontinuity Design William Shadish University of California, Merced.
Slides to accompany Weathington, Cunningham & Pittenger (2010), Chapter 4: An Overview of Empirical Methods 1.
Threats to Conclusion Validity. Low statistical power Low statistical power Violated assumptions of statistical tests Violated assumptions of statistical.
Who are the participants? Creating a Quality Sample 47:269: Research Methods I Dr. Leonard March 22, 2010.
Longitudinal Experiments Larry V. Hedges Northwestern University Prepared for the IES Summer Research Training Institute July 28, 2010.
Clustered or Multilevel Data
Using Covariates in Experiments: Design and Analysis STA 320 Design and Analysis of Causal Studies Dr. Kari Lock Morgan and Dr. Fan Li Department of Statistical.
Experimental Design The Gold Standard?.
Chong Ho Yu Department of Psychology, APU 362: Research Method.
Chapter 1: Introduction to Statistics
Chapter 11 Experimental Designs
S-005 Intervention research: True experiments and quasi- experiments.
Methodology Matters: Doing Research in the Behavioral and Social Sciences ICS 205 Ha Nguyen Chad Ata.
 Get out your homework and materials for notes!  Take-home quiz due!
A Randomized Experiment Comparing Random to Nonrandom Assignment William R Shadish University of California, Merced and M.H. Clark Southern Illinois University,
MSRP Year 1 (Preliminary) Impact Research for Better Schools RMC Corporation.
Controlling for Baseline
SW 983 Missing Data Treatment Most of the slides presented here are from the Modern Missing Data Methods, 2011, 5 day course presented by the KUCRMDA,
Can Mental Health Services Reduce Juvenile Justice Involvement? Non-Experimental Evidence E. Michael Foster School of Public Health, University of North.
One-Way Analysis of Covariance (ANCOVA)
1 Psych 5510/6510 Chapter 13 ANCOVA: Models with Continuous and Categorical Predictors Part 2: Controlling for Confounding Variables Spring, 2009.
 Research Design Part 2 Variability, Validity, Reliability.
Rerandomization to Improve Covariate Balance in Randomized Experiments Kari Lock Harvard Statistics Advisor: Don Rubin 4/28/11.
Patricia Gonzalez, OSEP June 14, The purpose of annual performance reporting is to demonstrate that IDEA funds are being used to improve or benefit.
Experimental Design Ragu, Nickola, Marina, & Shannon.
Research and Evaluation Methodology Program College of Education A comparison of methods for imputation of missing covariate data prior to propensity score.
William M. Trochim James P. Donnelly Kanika Arora 8 Introduction to Design.
© Yosa A. Alzuhdy - UNY © Yosa A. Alzuhdy – FBS-UNY 2b. HOW and WHY of RESEARCH Quantitative Research © Yosa A. Alzuhdy, M.Hum. English.
Looking for statistical twins
Multiple Regression.
Module 2 Research Strategies
Lurking inferential monsters
The Science of Social Psychology
Constructing Propensity score weighted and matched Samples Stacey L
Regression in Practice: Observational studies with controls for pretests work better than you think Shadish, W. R., Clark, M. H., & Steiner, P. M. (2008).
An Empirical Test of the Regression Discontinuity Design
William R. Shadish University of California, Merced
Experimental Design-Chapter 8
Classroom Assessment Validity And Bias in Assessment.
12 Inferential Analysis.
Module 02 Research Strategies.
PSYCH 610 Competitive Success/snaptutorial.com
PSYCH 610 Education for Service/snaptutorial.com.
Chapter Eight: Quantitative Methods
Review of Research Types
Reliability and Validity of Measurement
Introduction to Design
Research Methods in Behavior Change Programs
Multiple Regression.
Journalism 614: Experimental Methods
Cross Sectional Designs
Impact Evaluation Methods
Quasi-Experimental Design
RESEARCH METHODS Lecture 33
1/18/2019 1:17:10 AM1/18/2019 1:17:10 AM Discussion of “Strategies for Studying Educational Effectiveness” Mark Dynarski Society for Research on Educational.
12 Inferential Analysis.
Selecting the Right Predictors
Class 2: Evaluating Social Programs
Class 2: Evaluating Social Programs
Methods of Psychological Research
Methodological Evaluation of Experiments
Analysis of Covariance
Chapter 2: Research Methods
RESEARCH METHODS Lecture 33
Regression in Practice: Observational studies with controls for pretests work better than you think Shadish, W. R., Clark, M. H., & Steiner, P. M. (2008).
Chapter 3 Hernán & Robins Observational Studies
Presentation transcript:

Issues in Selecting Covariates for Propensity Score Adjustment William R Shadish University of California, Merced

Shadish, Clark & Steiner Randomly assigned participants to be in a randomized or nonrandomized experiment –Extensive pretesting on covariates Then tested whether we could reproduce the RE results by adjusting the NRE results. –Masking of NRE analyst from RE analyst Here is the design:

Random Assignment N = 445 Undergrad Psych Students Randomized Experiment N = 235 Randomly Assigned to Nonrandomized Experiment N = 210 Self-Selected into Mathematics Training N = 119 Mathematics Training N = 79 Vocabulary Training N = 131 Vocabulary Training N = 116 All Participants Post-tested on both Vocabulary and Mathematics Outcomes

Results Propensity score analysis could reduce bias in NRE estimates 58-96%, depending on the exact adjustment used. –Ordinary analysis of covariance did as well or better. –So did structural equation modeling –In fact, analytic method didn’t seem to matter much (By the way, this has been replicated in Germany) The quality of the measurement seemed to be the key:

Predictors of Convenience First hint of the importance of quality of measurement of predictors came from a sub-analysis of this study. If we limited our analysis to demographic predictors (age, gender, marital status, ethnicity) that are usually conveniently available, bias reduction was poor:

Mathematics outcome

Vocabulary Outcome

Exploring Covariate Sets 1 Imagine we split our 25 covariates into 5 concept domains –Demographics –Proxy Pretests –Topic Preference –Academic Achievement –Psychological Personality Traits Let’s explore how they relate to bias reduction 1 This section based on Steiner, Cook, Shadish & Clark (submitted).

Vocabulary Bias Remaining Depending on Which Covariate Sets Used Good bias reduction if you use most or all the covariates

Vocabulary Bias Remaining Depending on Which Covariate Sets Used But also good bias reduction using just a few of the “right” covariates.

Math Bias Remaining Depending on Which Covariate Sets Used Ditto for Math Good bias reduction if you use all covariates. But also good bias reduction with the “right”ones. In both math and vocab, topic preference and pretest were key.

Comments One strategy is to pick the “right” variables –But the best we can say to help researchers find them is they were among the most highly correlated with treatment and outcome An alternative is the “kitchen sink” strategy. –Measure as many covariates as possible and hope the key ones are in there. But does the kitchen sink need to include the “right” variables?

Vocabulary Bias Remaining using Individual Covariates For bias reduction in vocabulary outcome, the key individual variables were preference for math and vocabulary pretest scores.

Vocabulary Bias Remaining using Individual Covariates But bias reduction in vocabulary outcome was also good if you measured the right “domains” even if you didn’t have the “right” variables. Even if those domains were measured by variables that individually did not reduce bias very well.

Math Bias Remaining using Individual Covariates And the same was true for bias reduction in math outcome.

Comments So the best advice is to have the right variables. –But except for ensuring they are correlated with treatment and outcome, it is hard to know for sure. Alternative advice is the “kitchen sink” –With the “right” domains—possibly an easier task –With multiple measures in each domain even if none of the measures are the “right” ones.

Measurement Error 1 Some parts of the literature, like propensity score analysis, have mostly ignored the role of measurement error. We used the data from this study to simulate the effects of adding various amounts of measurement error to covariates. –2000 replications –Re-estimating PS’s each time –For three sets of All covariates Effective covariates Ineffective covariates 1 This section based on Steiner, Cook & Shadish (submitted).

Bias Reduction in Vocab Outcome Increasing measurement error decreases bias reduction for (a) all and (b) effective covariates, but has little impact on (c) ineffective covariates.

Bias Reduction in Math Outcome And the same is true for math outcome.

Some References Steiner, P.M., Cook, T.D., Shadish, W.R., & Clark, M.H. (under revision). The Importance of Covariate Selection in Controlling for Selection Bias in Observational Studies. Psychological Methods. Steiner, P.M., Cook, T.D., & Shadish, W.R. (under revision). On the Importance of Reliable Covariate Measurement in Selection Bias Adjustments Using Propensity Scores. Journal of Educational and Behavioral Statistics. Shadish, W.R., Clark, M.H., & Steiner, P.M. (2008). Can Nonrandomized Experiments Yield Accurate Answers? A Randomized Experiment Comparing Random to Nonrandom Assignment. Journal of the American Statistical Association, 103, Shadish, W.R., Clark, M.H., & Steiner, P.M. (2008). [Can Nonrandomized Experiments Yield Accurate Answers? A Randomized Experiment Comparing Random to Nonrandom Assignment]: Rejoinder. Journal of the American Statistical Association, 103, Shadish, W.R., & Cook, T.D. (2009). The Renaissance of Field Experimentation in Evaluating Interventions. Annual Review of Psychology, 60,

The End Acknowledgements: M.H. Clark (SIU), Peter Steiner (NU), Tom Cook (NU)