Addressing the Don’t Ask, Don’t Tell Practice in Observational Studies: Using Interviews to Understand the Assignment Mechanism Jordan H. Rickles Social Research Methodology Division Graduate School of Education & Information Studies University of California, Los Angeles SREE Annual Meeting | Washington, DC | March 6, 2010 Presentation Outline I. II. III. IV. V. VI. VII. VIII. IX. Overview of Study Brief Overview of Causal Inference The Assignment Mechanism Literature on Course Assignment Data and Methods Interview Findings Example at Three Schools Implications for Estimating Effects Conclusions Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 2 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions How many legs does a dog have if you call the tail a leg? Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 3 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Two Objectives Draw attention to the importance of studying the assignment mechanism Statistical inference for causal effects “requires the specification of a posited assignment mechanism describing the process by which treatments were assigned to units” (Rubin, 1991, p. 403). Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 4 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Two Objectives Draw attention to the importance of studying the assignment mechanism Statistical inference for causal effects “requires the specification of a posited assignment mechanism describing the process by which treatments were assigned to units” (Rubin, 1991, p. 403). “Knowledge of the selection process can significantly reduce selection bias provided the selection process is valid and reliably measured” (Cook, Shadish & Wong, 2008, p. 740). Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 5 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Two Objectives Draw attention to the importance of studying the assignment mechanism Statistical inference for causal effects “requires the specification of a posited assignment mechanism describing the process by which treatments were assigned to units” (Rubin, 1991, p. 403). “Knowledge of the selection process can significantly reduce selection bias provided the selection process is valid and reliably measured” (Cook, Shadish & Wong, 2008, p. 740). In randomized controlled trial and regression discontinuity designs assignment mechanism is usually known Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 6 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Two Objectives Draw attention to the importance of studying the assignment mechanism Statistical inference for causal effects “requires the specification of a posited assignment mechanism describing the process by which treatments were assigned to units” (Rubin, 1991, p. 403). “Knowledge of the selection process can significantly reduce selection bias provided the selection process is valid and reliably measured” (Cook, Shadish & Wong, 2008, p. 740). In randomized controlled trial and regression discontinuity designs assignment mechanism is usually known In observational study the assignment mechanism is usually unknown Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 7 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Two Objectives Provide example for using interviews to study the assignment mechanism Focus on study of taking algebra vs. pre-algebra in 8th grade Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 8 Causal Inference Overview | | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions The Potential Outcomes Framework Causal effects defined by potential outcomes at the unit of analysis δi = yt(i) – yc(i) Average Treatment Effect: E[δATE] = E[Yt] – E[Yc] Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 9 Causal Inference Overview | | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions The Potential Outcomes Framework Causal effects defined by potential outcomes at the unit of analysis δi = yt(i) – yc(i) Average Treatment Effect: E[δATE] = E[Yt] – E[Yc] Fundamental problem of causal inference (Holland, 1986) Can only observe yt for units assigned to the treatment and can only observe yc for units assigned to the control Under selection independence can estimate ATE as E[δATE]=E[Yt|D=1] – E[Yc|D=0] Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 10 Assignment Mechanism Overview | Causal Inference | | Literature | Data & Methods | Interview Findings | Example at Three Schools | Implications | Conclusions Selection Independence Treatment assignment is independent of the potential outcomes Holds for a randomized experiment Not likely to hold for an observational study where treatment assignment depends on factors (S) that are associated with potential outcomes Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 11 Assignment Mechanism Overview | Causal Inference | | Literature | Data & Methods | Interview Findings | Example at Three Schools | Implications | Conclusions Selection Independence Treatment assignment is independent of the potential outcomes Holds for a randomized experiment Not likely to hold for an observational study where treatment assignment depends on factors (S) that are associated with potential outcomes Assumption of strongly ignorable treatment assignment (Rosenbaum & Rubin, 1983) Treatment assignment is independent of the potential outcomes conditional on observed factors (S) Under conditional independence can estimate ATE as E[δATE]=E[Yt|S, D=1] – E[Yc|S, D=0] Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 12 Assignment Mechanism Overview | Causal Inference | | Literature | Data & Methods | Interview Findings | Example at Three Schools | Implications | Conclusions Validity of the assumption of strongly ignorable treatment assignment depends on the assignment mechanism If methods for conditioning on S accurately capture the true assignment mechanism then assumption holds Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 13 Assignment Mechanism Overview | Causal Inference | | Literature | Data & Methods | Interview Findings | Example at Three Schools | Implications | Conclusions Validity of the assumption of strongly ignorable treatment assignment depends on the assignment mechanism If methods for conditioning on S accurately capture the true assignment mechanism then assumption holds If methods for conditioning on S do not fully capture the true assignment mechanism then assumption may not hold Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 14 Assignment Mechanism Overview | Causal Inference | | Literature | Data & Methods | Interview Findings | Example at Three Schools | Implications | Conclusions Validity of the assumption of strongly ignorable treatment assignment depends on the assignment mechanism If methods for conditioning on S accurately capture the true assignment mechanism then assumption holds If methods for conditioning on S do not fully capture the true assignment mechanism then assumption may not hold How do we know if assumption holds? Short answer: we don’t. Better answer: we can investigate the assignment mechanism for a more informed determination Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 15 Literature Overview | Causal Inference | Assignment Mechanism | | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions What does the literature say about the assignment of 8th graders to algebra vs. pre-algebra? Need to rely on related literature on ability grouping, tracking, and high school course taking Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 16 Literature Overview | Causal Inference | Assignment Mechanism | | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions What does the literature say about the assignment of 8th graders to algebra vs. pre-algebra? Need to rely on related literature on ability grouping, tracking, and high school course taking Rational choice (or human capital) theory: students are matched with courses to efficiently accommodate differences in student ability Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 17 Literature Overview | Causal Inference | Assignment Mechanism | | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions What does the literature say about the assignment of 8th graders to algebra vs. pre-algebra? Need to rely on related literature on ability grouping, tracking, and high school course taking Rational choice (or human capital) theory: students are matched with courses to efficiently accommodate differences in student ability Nonacademic factors can mediate or constrain the optimality of the rational choice theory e.g., social class, student expectations, teacher & parent input, school assignment practices See Hallinan (1994) and Oakes, Gamoran & Page (1992) Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 18 Literature Overview | Causal Inference | Assignment Mechanism | | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions How have past researchers estimated the effect of 8th grade algebra? Mostly through regression-based modeling adjustments based on a variety of covariates OLS regression (Gamoran & Hannigan, 2000) Path analysis (Smith, 1996) Linear growth modeling (Ma, 2005) Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 19 Literature Overview | Causal Inference | Assignment Mechanism | | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions How have past researchers estimated the effect of 8th grade algebra? Mostly through regression-based modeling adjustments based on a variety of covariates OLS regression (Gamoran & Hannigan, 2000) Path analysis (Smith, 1996) Linear growth modeling (Ma, 2005) Studies generally find positive effect of 8th grade algebra on subsequent mathematics achievement Findings depend on assumption of strong ignorability Studies do not examine plausibility of this assumption Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 20 Data & Methods Overview | Causal Inference | Assignment Mechanism | Literature | Interview Findings | Example at Three Schools | Implications | Conclusions Data for estimation of causal effects Student-level data for a cohort of 2006-07 8th graders From large urban California school district Academic & demographic data from district’s administrative files, covering 2002-03 through 2007-08 school years Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 21 Data & Methods Overview | Causal Inference | Assignment Mechanism | Literature | Interview Findings | Example at Three Schools | Implications | Conclusions Data for estimation of causal effects Student-level data for a cohort of 2006-07 8th graders From large urban California school district Academic & demographic data from district’s administrative files, covering 2002-03 through 2007-08 school years Data to examine the assignment process 10 interviews of key school-level decision makers From middle schools with 8th graders in both algebra and prealgebra courses Interviews conducted in early 2009 Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 22 Data & Methods Overview | Causal Inference | Assignment Mechanism | Literature | Interview Findings | Example at Three Schools | Implications | Conclusions The interview protocol Part I: six semi-structured, open-ended questions about: The decision-making process The types of students typically assigned to algebra The general philosophy regarding 8th grade algebra Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 23 Data & Methods Overview | Causal Inference | Assignment Mechanism | Literature | Interview Findings | Example at Three Schools | Implications | Conclusions The interview protocol Part I: six semi-structured, open-ended questions about: The decision-making process The types of students typically assigned to algebra The general philosophy regarding 8th grade algebra Part II: questions about two student scenarios Designed to get more standardized summary of assignment mechanism across schools Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 24 Data & Methods Overview | Causal Inference | Assignment Mechanism | Literature | Interview Findings | Example at Three Schools | Implications | Conclusions The interview protocol Part I: six semi-structured, open-ended questions about: Part II: questions about two student scenarios The decision-making process The types of students typically assigned to algebra The general philosophy regarding 8th grade algebra Designed to get more standardized summary of assignment mechanism across schools Part III: rate importance of different information sources in assignment decision 5 category Likert scale Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 25 Data & Methods Overview | Causal Inference | Assignment Mechanism | Literature | Interview Findings | Example at Three Schools | Implications | Conclusions The interview protocol Part II, Scenario 1: Martin In 7th grade, Martin got a C in his first semester math class and a D the second semester. He received C’s and B’s in his other classes. He also received a mix of satisfactory and unsatisfactory marks for work habits and cooperation. In 6th grade, Martin’s math grades were a little higher, with a C the first semester and a B the second semester. Similarly, he scored Basic on the 6th grade math CST and Below Basic on the 7th grade math CST. You heard a couple of Martin’s 7th grade teachers mention that he started slipping behind and became more of a disruption in class as the year progressed. On a scale of 1 to 5, how likely is it that you would enroll Martin in Algebra in 8th grade? Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 26 Data & Methods Overview | Causal Inference | Assignment Mechanism | Literature | Interview Findings | Example at Three Schools | Implications | Conclusions The interview protocol Part II, Scenario 2: Maya Maya moved to California from Mexico during her 6th grade year and started attending this school in 7th grade. She is an English learner and is struggling to keep up in most of her classes. She received mostly D’s in 7th grade, but got a C in her second semester math class. Her work habits and cooperation marks are all satisfactory and there is no mention of any disciplinary problems in her records. She scored Far Below Basic on her 6th grade math and ELA CST tests but scored Basic on her 7th grade math test. You do not know much else about her except what is in her official record. On a scale of 1 to 5, how likely is it that you would enroll Maya in Algebra in 8th grade? Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 27 Data & Methods Overview | Causal Inference | Assignment Mechanism | Literature | Interview Findings | Example at Three Schools | Implications | Conclusions Analysis of interviews Coded each interview based on four assignment process characteristics Whether weight is given to objective (e.g., test scores) or subjective (e.g., teacher recommendations) criteria Whether decisions are primarily based on data systematically collected by the district (i.e., observable) or non-systematic data (i.e., unobservable). Whether well defined inclusion/exclusion decision rules are used or not Whether the school’s course placement philosophy is more protectionist or laissez faire Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 28 Data & Methods Overview | Causal Inference | Assignment Mechanism | Literature | Interview Findings | Example at Three Schools | Implications | Conclusions Analysis of interviews Coded each interview based on four assignment process characteristics Whether weight is given to objective (e.g., test scores) or subjective (e.g., teacher recommendations) criteria Whether decisions are primarily based on data systematically collected by the district (i.e., observable) or non-systematic data (i.e., unobservable). Whether well defined inclusion/exclusion decision rules are used or not Whether the school’s course placement philosophy is more protectionist or laissez faire Combined the codes with the close-ended responses for analysis Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 29 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Schools draw from a common battery of information sources in the decision-making process Teacher Recommendation 7th Grade Math Course Marks 7th Grade Math CST Individualized Education Plan (IEP) MDTP Algebra Readiness Test 6th Grade Math CST 6th Grade Math Course Marks School Attendance Record Parent Pref erence Student Pref erence 1 2 3 4 5 Importance of Inf ormation in Decision-Making Process (1=not at all important; 5=very important) Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 30 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Schools differ in weight given to specific information sources and the benchmarks (or cut-points) used to determine algebra placement Differences exemplified in responses to scenarios Favor Pre-Algebra Scenario 1 (Martin) n=5 Scenario 2 (Maya) 10% 20% Favor Algebra n=1 n=3 0% Uncertain n=4 n=4 30% 40% 50% n=3 60% 70% 80% 90% 100% Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 31 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Underlying philosophy associated with strictness of algebra placement benchmarks and decisions for “borderline” students Three schools had a protectionist philosophy Do not want to “program a kid for failure” “We don’t want students to be in a class where they’re going to struggle so much that they’re not going to be successful.” Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 32 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Underlying philosophy associated with strictness of algebra placement benchmarks and decisions for “borderline” students Three schools had a protectionist philosophy Do not want to “program a kid for failure” “We don’t want students to be in a class where they’re going to struggle so much that they’re not going to be successful.” Five schools had a laissez faire philosophy “Kids have a right to fail.” A student is “still better off failing the on-grade-level class than if he’d taken the other one.” Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 33 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Can exemplify school differences in the assignment mechanism by contrasting practices at three schools Haverbrook Middle School (HMS) Half of the 600 8th graders in algebra Teacher recommendations, math course grades & MDTP most important sources of information Soft benchmark of 70% on MDTP, C in 7th grade math class & Basic on CST Not strongly protectionist or laissez faire Martin in algebra, need more info to place Maya Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 34 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Can exemplify school differences in the assignment mechanism by contrasting practices at three schools Ogden Middle School (OMS) Half of the 900 8th graders in algebra Teacher recommendations, CST & weekly school-developed math quizzes most important sources of information No formal benchmarks but look for 60% correct on quizzes laissez faire philosophy Martin in pre-algebra, need more info to place Maya Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 35 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Can exemplify school differences in the assignment mechanism by contrasting practices at three schools Shelby Middle School (SMS) One-third of the 500 8th graders in algebra CST, math grades & MDTP most important sources of information Strong benchmarks of 80% on MDTP, Advanced on CST, B in 7th grade math class Protectionist philosophy Martin and Maya in pre-algebra Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 36 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Distributions of prior achievement for algebra and prealgebra students at the three schools are consistent with 7th Grade Mathematics Standardized Test 6th Grade Mathematics Standardized Test the general selection process described by each school Shelby Middle School PRE-ALG PRE-ALG ALGEBRA ALGEBRA 8th Grade Math Course 8th Grade Math Course Shelby Middle School Ogden Middle School PRE-ALG ALGEBRA Haverbrook Middle School Ogden Middle School PRE-ALG ALGEBRA Haverbrook Middle School PRE-ALG PRE-ALG ALGEBRA ALGEBRA 200 300 400 500 6th Grade Math CST Scale Score 600 200 300 400 500 600 7th Grade Math CST Scale Score Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 37 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Observed probability of algebra placement given CST performance and course grades also reflects school’s reported selection process Shelby Middle School (SMS) 7th Grade Math Course Grade 6th Grade Math CST Performance Level Example at Three Schools | Implications | Conclusions FBB A B C D F 1.00 . 0.00 0.00 0.00 0.00 to 0.24 0.25 to 0.49 BB 0.80 0.13 0.06 0.00 0.00 0.50 to 0.74 0.75 to 1.00 Basic 0.76 0.19 0.12 0.19 0.07 Pro 0.93 0.76 0.57 0.25 0.09 Adv 0.88 0.87 1.00 0.83 0.75 Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 38 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Observed probability of algebra placement given CST performance and course grades also reflects school’s reported selection process Ogden Middle School (OMS) 7th Grade Math Course Grade A B C D F 6th Grade Math CST Performance Level Example at Three Schools | Implications | Conclusions FBB . 0.00 0.16 0.10 0.18 0.00 to 0.24 0.25 to 0.49 BB 0.50 0.61 0.40 0.29 0.17 0.50 to 0.74 0.75 to 1.00 Basic 0.89 0.90 0.82 0.72 0.55 Pro 0.92 0.89 0.90 0.64 0.92 Adv 1.00 1.00 1.00 1.00 . Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 39 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Observed probability of algebra placement given CST performance and course grades also reflects school’s reported selection process Haverbrook Middle School (HMS) 7th Grade Math Course Grade A B C D F 6th Grade Math CST Performance Level Example at Three Schools | Implications | Conclusions FBB 0.20 0.43 0.30 0.19 0.19 0.00 to 0.24 0.25 to 0.49 BB 0.84 0.55 0.42 0.37 0.33 0.50 to 0.74 0.75 to 1.00 Basic 0.88 0.89 0.65 0.83 0.53 Pro 1.00 0.94 1.00 0.71 0.71 Adv 1.00 1.00 1.00 1.00 1.00 Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 40 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Investigating the assignment mechanism through interviews aides the causal effect research design Provides current & localized information about which factors are associated with treatment assignment What confounders should be controlled for? Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 41 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Investigating the assignment mechanism through interviews aides the causal effect research design Provides current & localized information about which factors are associated with treatment assignment What confounders should be controlled for? Provides information about how the factors are associated with treatment assignment How should confounders get included in a statistical model? Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 42 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Investigating the assignment mechanism through interviews aides the causal effect research design Provides current & localized information about which factors are associated with treatment assignment Provides information about how the factors are associated with treatment assignment What confounders should be controlled for? How should confounders get included in a statistical model? Provides information about heterogeneity in the assignment process across schools Can a single fixed-effects model control for confounders at all sites? Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 43 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Example: specifying a logistic regression model for assignment to algebra vs. pre-algebra Four different model specifications Model 1: naïve model with school homogeneity Model 2: informed model with school homogeneity Model 3: informed model with random intercept Model 4: informed model with random intercept & slopes Want to use predicted probabilities for propensity scorebased causal effect estimates, but need to use most appropriate propensity score model Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 44 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Different models result in different propensity scores For the average student … 0.80 SMS OMS HMS 0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00 Model 1 Model 2 Model 3 Model 4 Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 45 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Different models result in different propensity scores For a student like Martin … 0.80 SMS OMS HMS 0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00 Model 1 Model 2 Model 3 Model 4 Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 46 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Different models result in different propensity scores For a student like Maya … 0.80 SMS OMS HMS 0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00 Model 1 Model 2 Model 3 Model 4 Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 47 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Bias resulting for unobserved key factors should be tested through sensitivity analysis MDTP not observed Probably correlated with observed CST scores Teacher recommendations not observed Probably correlated with observed course grades Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 48 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions When the assignment mechanism is unknown, investigating the assignment process can help control for selection bias and communicate potential sources of bias Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 49 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions When the assignment mechanism is unknown, investigating the assignment process can help control for selection bias and communicate potential sources of bias An investigation of the assignment mechanism does not have to be extensive or resource intensive Short interviews conducted in a small sample of schools provides a lot of information Interviews could be part of a pilot study Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 50 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Complicated, heterogeneous assignment process found for 8th grade algebra probably applies to many major topics in education research Ability grouping, tracking & curricular intensity Intervention programs, tutoring services & school choice Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 51 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions Complicated, heterogeneous assignment process found for 8th grade algebra probably applies to many major topics in education research Ability grouping, tracking & curricular intensity Intervention programs, tutoring services & school choice Observational studies can play an important role in educational policy making But researchers must address the assignment mechanism complexities And honestly communicate those complexities for informed research consumption Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 52 Overview | Causal Inference | Assignment Mechanism | Literature | Data & Methods Interview Findings | Example at Three Schools | Implications | Conclusions How many legs does a dog have if you call the tail a leg? Four. Calling a tail a leg doesn’t make it a leg. - Abraham Lincoln Using Interviews to Understand the Assignment Mechanism | Rickles | SREE 2010 | Slide 53 Addressing the Don’t Ask, Don’t Tell Practice in Observational Studies: Using Interviews to Understand the Assignment Mechanism Jordan H. Rickles jrickles@ucla.edu Social Research Methodology Division Graduate School of Education & Information Studies University of California, Los Angeles SREE Annual Meeting | Washington, DC | March 6, 2010 Part of this research is made possible by a pre-doctoral advanced quantitative methodology training grant (#R305B080016) awarded to UCLA by the Institute of Education Sciences of the US Department of Education. The views expressed in this paper are the author’s alone and do not reflect the views/policies of the funding agencies or grantees.