Preliminary Findings Social and Character Development in Elementary School: The Effectiveness of the Making Choices Program Mark W. Fraser, PI, School of Social Work, UNC-Chapel Hill Steven H. Day, School of Social Work Shenyang Guo, School of Social Work Alan Ellis, Sheps Center and School of Social Work Roderick A. Rose, School of Social Work Maeda J. Galinsky, School of Social Work Kim Dadisman, Co-PI, Center for Developmental Science, UNC-CH Dylan Robertson, Center for Developmental Science Tom Farmer, School of Education, Pennsylvania State University This presentation was given at the School of Social Work, University of Maryland, Baltimore, MD, on April 9, 2009. Portions of this report were presented at the annual meeting of SACD Project grantees on June 13, 2008 in Washington, DC Agenda • Theoretical Bases and programs • Design and challenges • Analytic strategies • Analytic methods (skim – see slides) • Findings Acknowledgments • This project was support by a cooperative agreement (R305L030162) with the Institute of Education Sciences at the U.S. Department of Education (US DOE). Funding for the project was appropriated by the US DOE and the Centers for Disease Control and Prevention. • We thank Paul Rosenbaum (U Penn), Ben Hansen (U of Michigan), and Matthias Schonlau (Rand Corp) for their consultation on methodological issues related to this presentation. Teachers Talk about Making Choices • Changes in Classroom Atmosphere • Observable Differences in Student Behaviors • Measurable Academic Achievement Classroom Atmosphere “I noticed that the classroom started working more as one big group instead of individuals.” Gr.5 Sandy Grove Elementary, Hoke County Observable Behaviors “The students tend to be less critical of each other and more understanding of each other’s differences.” Gr. 5 Sandy Grove Elementary, Hoke County Academic Achievement “ The program uses excellent books to support the goals of being a good friend and not hurting others.… I use them during Language Arts time.” Gr. 4 Tommy’s Road Elementary, Wayne County Observable Behaviors … “It provided a way for students to put their feelings into words.” I am feeling really mad! Gr. 2, Bunn Elementary, Franklin County Academic Achievement “My students spend more time on task. They seem less distracted by annoying behavior.” Gr. 5 Scurlock Elementary, Hoke County Make the right choice! Children are actually stopping and thinking about making the right choices, and I have heard a lot of children say to themselves, “Make the right choice.” It is great to hear. Kdg. Bunn Elementary, Franklin County “It made a difference with teaching children how to deal with their feelings using better methods rather than having tantrums or hitting.” Kdg. Bunn Elementary, Franklin County Oh, boy! I need that Making Choices program. Classroom Atmosphere “This program provided a foundation on which we could build a classroom community.” Gr. 1 North Drive, Wayne County Children in my school need… …social skills to make friends and deal with interpersonal problems. …lessons that teach children respect toward others and responsibility for their own actions. …a program designed to reduce disruptive behavior and promote academic achievement. Making Choices Research Question: Does social and character education work? (i.e., is Making Choices effective?) Intervention Research Perspective: The Design and Development Approach 1. Specify the problem and develop a program theory 2. Create and revise program materials 3. Refine and confirm program components (sequential experimentation perspective) 4. Assess effectiveness in a variety of practice circumstances and settings 5. Disseminate findings and program materials Source: Fraser, M. W., Richman, J. M., Galinsky, M. J., & Day, S. H. (2009). Intervention research: Developing social programs. New York, NY: Oxford University Press. Social and Character Development in Childhood: A Risk and Resilience Orientation PROBLEM THEORY: Perspectives on Conduct Problems and Academic Achievement in Elementary School •Developmental risk perspective •Ecological theory •Social information processing theory Eco-Developmental Risk Cascade POINT: Risk factors for poor developmental outcomes vary over time. Lacking effective intervention, the potential for poor outcomes increases – and cascades – as function of complex bio-social processes. To promote positive outcomes, we must disrupt malleable risk mechanisms. Biological Risks Parenting Family-School Pre-School Climate Neighborhood School Readiness Processing Skills Parenting Family-School School Climate Neighborhood Increasingly Broad Repertoire of Potentially Damaging and Aggressive Behaviors Peer Rejection Academic Failure Parenting Family-School School Climate Neighborhood Cognitive Mediation Model (in Developmental Sciences) Biological Predisposition Biological Predisposition Parenting •Monitoring •Bonding Peers •Deviancy training •Contagion effect •False consensus effect Sociocultural Context •Stress/poverty •Racism •Street codes •Acute/chronic stress Mental Processes •Social knowledge •Scripts •Schema/skills Sociocultural Context •Stress/poverty •Racism •Street codes •Acute/chronic stress Conduct Problems •Conduct disorder •Fighting •Drug use Adapted from: Dodge, K. A., & Pettit, G. S. (2003, p. 351). A biopsychosocial model of the development of chronic conduct problems in adolescence. Developmental Psychology, 39(2), 349-371. Social Information Processing Theory: SIP Skills and Emotional Regulation as Malleable Mediators? Interpret social cues Encode social cues Assess outcomes State the problem Arousal, Emotions, Social Knowledge Select & enact the best solution(s) Set goal(s) Generate potential solutions Evaluate potential solutions Social Knowledge: Life experiences producing scripts, schemata, skills, and beliefs PROGRAM THEORY (specifies how a program is to work) Intervention Program Structure Such as: Targeting Unit: Classroom Entire school Other (after school, family) Curriculum Structure: Distinct activities Embedded in curriculum Activities to address SACD Goals Such as: Character education Violence prevention/peace promotion Social and emotional development Tolerance and diversity Risk prevention and health promotion Behavior management Social and Character Development: Prevention Model Behavior Social Emotional Competence (mediator) Attitudes about aggression Self-efficacy Empathy School Climate (mediator) School connectedness Victimization Feelings of safety at school Parent involvement Positive Behavior Responsible behavior Prosocial behavior Self-regulation Cooperation Negative behavior Aggression Minor delinquency Disruptive classroom behavior Academics Academic competence School engagement Grades Standardized test scores Moderating Factors Child Gender Socioeconomic status Race/ethnicity Risk status Prior test scores/grades Family Parenting practices Home atmosphere Community Community risk factors Social capital Program Fidelity Intensity and dosage School Activities to promote social and character development Organizational structure Social Development Model Perspective Engaged school behavior ... Focusing on work Paying attention Following instructions Instruction in social & emotional skills . . . Empathy Anger management Problem solving & impulse control School success Reductions in . . . Classroom disruptive behavior Anxiety Anger Opportunity to . . . Discuss and identify feelings Acquire language and communication skills Practice solving problems Observe models Social & emotional skills . . . Empathy Anger management Problem solving & impulse control Reductions in . . . Problem behavior Aggression Peer rejection Reinforcement and generalization of learning through . . . Naturally occurring opportunities in school Home discussion of materials Intervention Immediate outcomes Knock-on outcomes Outcomes PROGRAMS The Competence Support Program* Social Dynamics Training Social Skills Training for students for teachers Group randomization: Cohort 1: Hoke and Wayne Counties (10 schools randomized to 5 intervention; 5 control) Cohort 2: Franklin County (4 schools randomized to 2 intervention; 2 control) Classroom Behavior Management Training and Consultation for teachers *Developed by the program investigators, the intervention simultaneously focuses on the characteristics of children and on the classrooms in which they learn. The intervention combines three components. Program Elements • Making Choices: Skills Training curriculum for students in elementary school. In-service training introduced teachers to the risks of peer rejection and social isolation, including poor academic outcomes and conduct problems. Throughout the school year, teachers received consultation and support (2 times per month) in providing lessons designed to enhance children’s social information processing and other skills. As a part of the Standard Course of Study, the program was integrated into routine class instruction. • Classroom Behavior Management provided teacher consultation on classroom management strategies designed to strengthen engagement in instructional activities. • Social Dynamics Training provided teacher consultation on classroom contexts, social groupings, and interactional patterns that can be used to reinforce academic achievement and prosocial behavior. Theory of Change: Making Choices Random Assignment Core #1 Training the Teacher or Counselor Core #2 Application of Making Choices by Teacher or Counselor Characteristics of the Teacher or Counselor Treatment as Usual Control Condition Core #3 SIP skills of the Children in the School Characteristics of the Children and the Classroom Core #4 Impact on Social Engagement and Peer Rejection Core #5 Impact on Disruptive Behavior and Academics •Assess implementation of •Test the degree to which training the intervention is •Assess if teacher acquires delivered as intended, skills from training/supervision e.g., specific activities Note. In a randomized trial, you must figure out a way to measure each of the core elements. Make Program Manuals • From risk mechanisms, mediators, and logic models to the design of a program • Specifying program activities that target the malleable mediators and have cultural congruence • Example: Making Choices For a discussion of issues in the development and use of treatment manuals, see: Galinsky, M. J., Terzian, M. A., & Fraser, M. W. (2006). The art of group work practice with manualized curricula. Social Work with Groups, 29(1), 11-26. Warning: It is easy to under estimate the difficulty of developing a program manual. “That Sunk Feeling” If you start in the wrong place, it usually does not help to dig deeper! Source: Don Moyer, Harvard Business Review (October, 2004, p. 160) How to begin in the right place… Start with theory and research, plus practice experience… Develop a template for each lesson or session Grade 2 Recognizing Your Feelings Objectives: The learner will recognize that certain situations bring out feelings in all of us. Lesson 2 The learner will practice recognizing their own feelings. The learner will use personal experiences and knowledge to interpret written and oral messages. (SCS- LA 3.01) Overview The learner will write structured, informative presentations and narratives when given help with organization. (SCS- LA 4.08) Materials: Prep Materials Penguin Facts page, Response Sheets, Write About It worksheets A and B Introduction Review the idea that we all experience a variety of emotions and responses to emotions. Even when we experience the exact same situation, we may have different responses to the situation. Our responses to our feelings can cause us to do good things, but at times they can also cause us to do things that are not helpful. Activity 1 Answers Activity I: Pete the Penguin Using two columns, list on the board the emotions presented in Lesson 1 of the book, The Way I Feel. Column I- Emotions that Feel Good: happy, silly, excited, proud, or thankful Column II- Emotions that Don’t Feel Good: scared, sad, disappointed, bored, angry, or jealous Process Tip Introduce the students to Pete the Penguin using the penguin puppet. Pass out the Penguin Facts page and discuss the factual information about penguins. Explain to the students that Pete has experienced events that have brought out many different emotions. Sometimes his emotions feel good, but at other times they don’t feel very good at all. Review the emotions listed in the columns on the board. Then give each student four small pieces of paper (about the size of a note card). Read aloud the following events involving Pete the Penguin. After reading each event, ask the students, “How would you feel?” Give the students enough time to record their responses on one of Standard Course of Study Review Prop their blank pieces of paper. They can use the emotions on the board to express how they would feel or they may provide their own responses. Avoid labeling Scenarios Activity 2: Write About It! After you read each situation, collect a few responses randomly and read them aloud (so as not to bring attention to specific student responses). As you read through each response, discuss whether the event brought out a good feeling or a not-sogood feeling. The texts are ambiguous so that students can develop their own interpretations—not all students will feel the same way about each situation. Discuss the idea that everyone heard the same event, yet the feelings were different in many instances. Today Pete walked in the classroom. As he walked to his desk, Pete noticed Susan and Tony talking quietly and laughing. They both looked up at Pete and giggled. If you were Pete, how would you feel? When Pete was on the playground, he saw a group of students playing ball. He went to join them, and they told him he could play as soon as they started the next game. If you were Pete, how would you feel? At lunch, Pete was sitting next to Jermaine. Jermaine opened his lunch and Pete looked inside. All he saw was two cookies and a drink box. If you were Pete, how would you feel? Pete’s teacher told him he could play a game with Juan as soon as he finished his writing assignment. If you were Pete, how would you feel? After discussing the above events, ask the students how they recognize when they are feeling certain emotions. “What happens when you start to feel angry?” “Happy?” “Frustrated?” and so on. (Example response: When you are getting angry- you might get hot, start to shake, get tense, grit your teeth, etc.) Leave the list of emotions on the board to use in Activity II. Activity II: Write About It Give the students the Write About It page. On the top of the sheet, have students write about an event in their life that caused them to experience an emotion that made them feel good. On the bottom of the sheet they can write about an experience that caused an emotion that didn’t feel good. Each narrative should describe the emotion, what caused it, and how they responded to the emotion. Students can refer to the columns on the board to choose the emotions they want to write about. Share the following examples aloud or on a transparency: Example 1: Once I felt excited when I was going to my friends party. I knew I felt this way because I was smiling and jumping around. Develop all worksheets and artwork Activity II: Sheet A NAME: _________________ A good feeling: Once I felt ___________________ when ____________________________________ ____________________________________ ____________________________________ ____________________________________ I knew I felt this way because__________ ____________________________________ ____________________________________ Activity II: Sheet B NAME: ____________________ A not so good feeling: Once I felt _____________________ when ___________________________________ ___________________________________ ___________________________________ ___________________________________ I knew I felt this way because__________ ___________________________________ ___________________________________ “Pete the Penguin” Poster for Grade 2 Sample Lesson Activities from Making Choices Gr. 3 Lesson - Intentions SYMBOLS Grrrrr! MEAN FRIENDLY CAN’T TELL Intentions: Mean or Friendly? LAUGHS WITH YOU FRIENDLY HITS YOU SHARES WITH YOU MAKES A FACE AT YOU HUGS YOU IGNORES YOU FRIENDLY CAN’T TELL Grrrrr! TALKS ABOUT YOU BITES YOU HELPS YOU FRIENDLY MEAN Grade 3 Lesson 10 GOAL SETTING GOAL: Something a person wants or something a person wants to see happen. RELATIONSHIP GOAL: Goals that involve wanting to get along with another person. Grade 4 Lesson 6 Are these Relationship Goals? (thumbs up or thumbs down) • • • • • • • • • I want to make an “A” on my math test. I want to play more often with my friend. I want a new video game for my birthday. I want to eat out at a restaurant for dinner. I want to become friends with the new student. I want to join in the basketball game at recess. I want to sit with Jose on the bus. I want to be in the class play this fall. I want to stop getting upset when friends ignore me. GOAL SETTING • Set a relationship goal for these situations: I was playing basketball at recess with some friends. Terrell, who is not very good at basketball, asked if he could play with us. Set a Relationship Goal Denise just made me really upset. She tried to pick a fight with me by saying things that are not true. I am feeling angry with her right now. Set a Relationship Goal Yesterday, my mom gave me a really cool pen that writes in all different colors. When I brought it to school this morning, Stacey asked me if she could borrow it. Last time I let Stacey borrow something she lost it, but if I say no she might get angry with me. EVALUATION DESIGN: Cluster Randomized Trial with Ten Schools Randomly Assigned to Treatment (j=5) and Control (j=5) Conditions Cohort Design: Intervention provided in grades 3, 4, and 5 Prior Studies 1. Single-group qualitative trial of MC intervention (8th grade girls) 2. Two-group cluster randomized trial at classroom level in one middle school (6th grade only) 3. Two-group cluster randomized trial at classroom level in one school (3rd grade) 4. Two-group, MC+SF intervention randomized trial (11 sites, 3rd – 4th grade) 5. Cohort sequential study by classroom in two schools (3rd grade) 6. (Current) Two-group cluster randomized trail at 14 elementary schools SAMPLE Two Overlapping Samples Grade 3 n=571 Grade 4 n=557 Grade 3-4 Sample • 3rd and 4th graders • 10 schools • Any consented students on a 3rd or 4th grade roster • Change=entrants-leavers Grade 5 n=433 n=414 n=370 Grade 3-4-5 Sample • 3rd, 4th, and 5th graders • 9 schools* • Only consented students on a 5th grade roster • Change=addition of entrants * One treatment school was reorganized into a different building and dropped the program between 4th and 5th grade; students from that school were excluded from the 3-4-5 sample. Equivalence of Intervention and Control Groups on Selected Child, Family, and School Attributes: Grade 3 Cohort 1 N Total Control Intervention p-value Male child, % 618 47.7 50.0 45.3 .243 Age of child, years 618 7.92 7.94 7.89 .235 Black child, % White child, % 618 618 34.5 48.1 21.9 59.3 48.0 35.8 .001 .001 Hispanic child, % 618 10.0 13.7 6.1 .001 564 563 550 13.5 48.5 26.0 15.0 43.7 24.6 11.7 54.0 27.6 .259 .014 .416 School (aggregate, school level) Black, %* 10 41.3 28.4 51.4 .001 Free Lunch, %* Adequate Yearly Progress*# 10 10 44.7 81.9 37.4 84.7 50.7 79.0 .001 .068 Pupil/teacher ratio, mean* 10 16.3 16.2 16.4 .825 Child (analysis sample) Family (analysis sample) Primary caregiver, not a HS graduate , %+ Two biological parents not in household, %+ Income-to-needs < = 1, %+ Note. + MPR data from baseline child level data file. * NCES school level data (CCD 2003-2004) across all schools. # AYP Performance Composite score, Year 1 of the SACD study. Difference in School-Level Academic Performance: Percentage at Grade Level Cohort 1 AYP Performance Composite Percent of students achieving at grade level 90 85 80 75 Treatment Control 70 65 60 55 50 02-03 03-04 04-05 05-06 06-07 Test results for 2005-06 and 2006-07 are based on a revised accountability model and are not comparable to those from previous years. Equivalence of Intervention and Control Groups on Selected Site Specific Outcomes Grade 3 Sample (Pretest): Cohort 1 N Total Control Intervention p-value SIP Skill Level Assessment - Child Encoding 420 44.2 43.0 45.6 .149 Goal formulation Response decision making 412 410 67.6 66.0 65.5 64.1 70.1 68.3 .074 .225 Social contact Cognitive concentration Social competence 549 549 549 3.8 3.2 3.3 3.7 3.2 3.3 3.8 3.2 3.3 .121 .952 .945 Social Aggression 549 4.1 4.1 4.0 .183 Aggression 548 2.5 2.4 2.6 .095 Academic competence Popularity 548 548 5.1 4.9 5.1 4.9 5.1 4.8 .948 .798 Aggression 502 38.7 42.6 34.4 .152 Prosocial skills 502 82.8 84.8 80.6 .552 Carolina Child Checklist - Teacher Interpersonal Competence Scale Teacher Peer Interpersonal Assessment Sample sizes vary because pretest measures were collected from different respondents (teachers, students) at different times. SLA and Peer assessment pretest were collected from students at the end of 2nd grade. CCC and ICST were collected from teachers at the beginning of 3rd grade. SLA=Skill Level Assessment (SIP skill – HOME Scale adaptation by Dodge, 1980). CCC=Carolina Child Checklist (Macgowan et al. 2002 – Research on Social Work Practice). ICST-Interpersonal Competency Scale – Teacher (Xie et al., 2002, Social Development) Teacher and Classroom Characteristics by Intervention Status Characteristic Demographics White %+ Greater than a bachelor degree %+ Years of teaching, mean+ Years teaching at current school, mean+ Has regular/advanced teaching certificate, %+ SACD classroom activities N of SACD classroom strategies, mean+ Violence prevention hours, mean+ Social and emotional development hours, mean+ Classroom observations Number of feedback and structure exemplars in place, mean# Intervention Teachers (n = 21) Non Intervention Teachers (n = 23) Difference p-value 73.7 28.6 11.6 95.5 4.6 14.1 -11.8 24.0 -2.5 .051 .033 .385 4.1 8.0 -3.9 .040 85.0 90.9 -5.9 .566 14.8 7.8 13.7 2.7 1.1 5.1 .267 .039 10.9 2.9 8.0 .002 6.2 4.2 2.0 .019 Note. +MPR baseline data (spring). #Observations were conducted with the Classroom Observation Form (COF) by an intervention specialist blind to the treatment or control condition of each school. The COF assesses seven domains relevant to the intervention: daily routines, time and task management, consequences and follow-through, teaching alternative behaviors, communication and feedback, and group processes and peer support. Causation in Research Design: Randomization Is Supposed to Produce the Counterfactual Note. We let the control group serve as evidence for what would have happened counter to the fact of participation in intervention (the stat class). Randomization is supposed to create equivalence or balance between the intervention (taking the stat class) and control (not taking the stat class) groups. But it didn’t. On several observed and an unknown number of unobserved measures, the intervention and control group schools differ. Four evaluation challenges Selection Bias: Covariates are not balanced between treated and control groups Missing Data: No baseline data on enterers and lost data on leavers = constant churning of sample Rater Effects: Outcome ratings were made by the same teachers within grades, but different teachers over grades 3, 4, and 5 • Piecewise analyses – change scores within grade level Treatment Contamination/History: High intervention content in control schools* Note. Student Citizen Act (SL 2001-363) was passed into law by the Legislature in 2001. The Act required local boards of education to develop and incorporate character education instruction into standard curricula. Local boards of education began implementation in the 2002-2003 school year. MEASURES Site-Specific Outcomes • Skill Level Assessment Activity (SLA): Based on the Dodge Home Scale (1980), the SLA uses students’ responses to questions about hypothetical social situations. After viewing picture scenarios, students answer questions measuring different aspects of social information processing skill: encoding (α=.78), goal formulation (α=.76), and response decision making (α=.80). • Carolina Child Checklist (CCC): The CCC is a 35 item teacher questionnaire that yields factor scores on children’s behavior including social contact (α=.90), cognitive concentration (α=.97), social competence (α=.90), and social aggression (α=.91). • Interpersonal Competence Scale-Teacher (ICST): The ICST is an 18-item teacher questionnaire that yields factor scores on children’s behavior including aggression (α=.84), academic competence (α=.74), and popularity (α=.78). • Peer interpersonal assessments: Peer interpersonal assessments were used to examine classmates’ perceptions of participants’ social and behavioral characteristics including aggression (α=.92), prosocial skills (α=.84 ), and internalizing behavior (α=.67 ). Summary of Data Collection Occasions Time of year Cohort 1 (C1) Cohort 2 (C2) Grade 2 Wave 1 March 2003-04 2004-05 Wave 2 Sept 2004-05 2005-06 Grade 3 Wave 3 March 2004-05 2005-06 Wave 4 April 2004-05 2005-06 Wave 5 Sept 2005-06 2006-07 Grade 4 Wave 6 March 2005-06 2006-07 Wave 7 April 2005-06 2006-07 Grade 5 Wave 8 Wave 9 Sept April 2006-07 2006-07 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C1 C1 C2 C1 C2 C1 C2 C1 C1 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C1 C1 C1 C1 C1 Instrument (in-house data) Making Choices Student report (SLA) Friends and Groups Student report (Social Cognitive Maps) Peer groups Teacher report (PG) Carolina Child Checklist Teacher report (CCC) Interpersonal Competence Scale Teacher report (ICST) Peer Interpersonal Assessments Student report (PNOMS) Ratings of school adjustment Teacher report (TASS) C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 MPR data (across-site data) Student reported Teacher reported Parent reported C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C2 C1 C1 C1 C1 C2 C1 C2 C1 C2 C1 C1 C1 Minutes of Skills Training Instruction in 3rd and 4th Grades by Student Benchmark=1,140 minutes Below benchmark: 19% Above benchmark: 81% (overall n=571) ANALYTIC PROCEDURES Multiple Imputation of Missing Data (The imputation models employed both predictor and outcome variables, but the analysis models employed imputed missing values for predictor variables only). 50 imputations for each outcome variable Estimation of propensity scores using Generalized Boosted Modeling (gbm) -- aims to optimize balance on observed covariates between treated and control groups Heckman sample selection Model (Predictors of the selection equation are similar to the input of gbm) Dose (efficacy subset) analysis) using Abadie et al. Matching estimator Analytic Procedures: Flow Chart for Use of “Bias-Correcting” Statistical Methods Optimal pair matching using propensity scores estimated by gbm Piecewise change score HLM analysis using propensity score weighting (propensity scores estimated by gbm) Post-pair-matching with regression adjustment Optimal full matching using propensity scores estimated by gbm Post-full-matching with HodgesLehmann aligned rank test Procedures for multiple imputation of missing data Test for MCAR (Little, 1988) confirms models are not MCAR. Assumption of MCAR not required if imputation model is informed (i.e., data may be missing at random) A diagnostic stage identified models that resulted in 99% relative efficiency for all analysis variables. 50 simulations (copies of the raw data set) generated using MI. DVs and predictors both used in imputation; imputed DVs discarded after imputation (MID procedure; von Hippel, 2007). Missing Data Diagnostics: Proportion without Missing Data and Proportion of Missing Data Points General Procedure for Propensity Score Analysis Step 1: Logistic regression Dependent variable: log odds of receiving treatment Search an appropriate set of conditioning variables (boosted regression, etc.) Estimated propensity scores: predicted probability (p) or log[(1-p)/p]. Step 2:Analysis using propensity scores Analysis of weighted mean differences using kernel or local linear regression (difference-indifferences model of Heckman et al.) Step 2: Analysis using propensity scores: Multivariate analysis using propensity scores as weights Step 2: Matching Greedy match (nearest neighbor with or without calipers) Mahalanobis with or without propensity scores Optimal match (pair matching, matching with a variable number of controls, full matching) Step 3: Post-matching analysis Multivariate analysis based on matched sample Step 3: Post-matching analysis Stratification (subclassification) based on matched sample Estimating propensity scores Need relevant conditioning variables Obtain “best” logistic regression (i.e., best functional forms); however, no way to know Used Multiple Additive Regression Trees (MART) to run logistic regression. Rand Generalized Boosted Modeling (gbm): Aims for best balance on observed covariates between treated and controlled groups. Iteration stops when the sample average standardized absolute mean difference (ASAM) is minimized. Example of gbm output: Does gbm reduce the difference between treated and control schools? Point: After using gbm propensity score weights, all pretreatment differences are ns. STR=treatment group; LTR=control group; ASAM= average standardized absolute mean difference between treatment and control cases; pretreatment covariates: red solid diamonds= p-values before use of gbm weights (if below line then significant); black diamond outline = p-values after weights applied Predictors of the propensity score model ________________________________________________________________________________________________________________________ Outcome ___________________________________________________________________________ ICSTAGG ICSTACA ICSTINT CCCSCOM CCCPROS CCCEREG CCCRAGG Predictor Academ ic Social Aggression Com petence Internalizing Com petence Prosocial Em otion Regulation Relational Aggression White Hispanic Prim ary caregiver's education (years of schooling) Ratio of incom e to poverty threshold Prim ary caregiver full-tim e em ploym ent (part-tim e is referrence) Father's presence in fam ily: Yes (absence is reference) Age at baseline (year) Gender fem ale (m ale is reference) Race (Other is reference) African Am erican ICST-aggrestion at baseline ICST-academ ic com petence at baseline CCC-social com petence at baseline CCC-prosocial at baseline CCC-relational aggression at baseline ________________________________________________________________________________________________________________________ ICST-internalizing at baseline CCC-em ontion regulation at baseline Note. Predictors vary by outcome variable. Following the convention of propensity score analysis, we did not include predictors that are highly correlated with the outcome variable. Propensity score weighting When estimating the treatment effect, can use propensity scores as sampling weights. (Hirano & Imbens, 2001; McCaffrey et al., 2004; Rosenbaum, 1987) Suppose p is the propensity score of receiving treatment. Then: Average treatment effect for the treated (ATT): control weight = p/(1-p) treatment weight = 1 Average treatment effect for the population (ATE): control weight = 1/(1-p) treatment weight = 1/p Post-optimal-matching analysis For the matched sample created by optimal pair matching, regress pairwise differences in Y between treated and control cases on pairwise differences in X vector between treated and control cases (Rubin, 1979). In doing so, use the intercept of the regression to estimate the treatment effect and its p-value as a significance test. For matched sample created by optimal full matching or optimal variable matching, use the signed-rank test of Hodges and Lehmann (1962) to estimate the average treatment effects. Dose analysis using Matching estimator • The dose analysis evaluates the outcome difference between a dose group (i.e., low, benchmark, or high) and a comparison group using Matching estimator developed by Abadie et al. (2004). • Under the exogeneity assumption, this method imputes the missing potential outcome by using average outcomes for individuals with “similar” values on observed covariates. • The estimator uses the vector norm (i.e., ||x||v=(x’Vx)1/2 with positive definite matrix V) to calculate distances between one treated case and each of the matched multiple nontreated cases, and chooses the outcome of the nontreated case whose distance is the shortest among all as the predicted outcome for the treated case. Comparing model features _________________________________________________________________________________________________ Model The Model Controls for: ____________________________________ Level at which treatment was tested Multiple imputation of missing Rater's data effect Selection bias Clustering ____________________________________________________ ________ ________ ________ ________ ________ Piecewise change with propensity score weighting (ATE) School Yes Yes Yes Yes Piecewise change with propensity score weighting (ATT) School Yes Yes Yes Yes Optimal pair matching with regression adjustment School Yes Yes Yes Yes Optimal full matching with Hodges-Lehmann test Student Yes Yes Yes No Efficacy subset analysis using Matching estimator Student Yes Yes Yes No _________________________________________________________________________________________________ Note. Regression models include covariates age at baseline, female, black, white, latino, primary caregiver education, income-to-poverty ratio, primary caregiver fulltime employment, father in household, and midyear change in teacher. FINDINGS Findings: Treatment effects measured by changes in the 3rd Grade (g34) Outcome Variable Piecewise Change with Propensity Score (gbm) Weighting Model 1: Hypothetical ATE Sign Grade 3 Piecewise Change with Propensity Score (gbm) Weighting Model 2: ATT Grade 3 Change Score Using Change Optimal FullScore Using Matching Optimal Pair- (gbm) and Matching Hodges(gbm) and Lahmann Regression Aligned Adjustment Rank Test Grade 3 Grade 3 Approximate Sample Size Used in Analysis ≈571 ≈571 ≈542 ≈571 ICSTAGG - Aggression -.10 -.04 -.10 -.01 ICSTACA - Academic competence + .12+ .08 .11 -.08*** ICSTINT - Internalizing .14+ .13+ .14 .17 CCCSCOM - Social competence + -.22 -.25+ -.25 -.25 CCCPROS - Prosocial + -.25 + -.26 + -.25 -.19+ CCCEREG - Emotion regulation + -.16 -.18 -.20 -.24 CCCRAGG - Relational aggression .17 .21* .18 .28 _________________________________________________________________________________________________________ *** p <.001, ** p < .01, * p < .05, + p < .1, two-tailed. Findings: Treatment effects measured by changes in the 4th Grade (g34) Outcome Variable Piecewise Piecewise Change Change Change with with Score Using Propensity Propensity Optimal Score Score Pair(gbm) (gbm) Matching Weighting Weighting (gbm) and Model 1: Model 2: Regression Hypothetical ATE ATT Adjustment Sign Grade 4 Grade 4 Grade 4 Change Score Using Optimal FullMatching (gbm) and HodgesLahmann Aligned Rank Test Grade 4 Approximate Sample Size Used in Analysis ≈557 ≈557 ≈550 ≈557 ICSTAGG - Aggression -.13 -.14 -.14 -.17 ICSTACA - Academic competence + -.13+ -.10 -.11 -.08 ICSTINT - Internalizing -.00 .04 -.02 .18 CCCSCOM - Social competence + -.01 -.02 .06 .05 CCCPROS - Prosocial + -.00 -.03 .05 .07+ CCCEREG - Emotion regulation + -.00 -.01 .06 .03 CCCRAGG - Relational aggression -.12 -.12 -.13 -.06 ______________________________________________________________________________________________________ *** p <.001, ** p < .01, * p < .05, + p < .1, two-tailed. Findings: Treatment effects measured by changes in the 3rd Grade (g345) Outcome Variable Piecewise Change with Propensity Score (gbm) Weighting Model 1: Hypothetical ATE Sign Grade 3 Piecewise Change Change Change Score Score Using with Using Optimal FullPropensity Optimal Matching Score Pair(gbm) and (gbm) Matching HodgesWeighting (gbm) and Lahmann Model 2: Regression Aligned ATT Adjustment Rank Test Grade 3 Grade 3 Grade 3 Approximate Sample Size Used in Analysis ≈370 ≈370 ≈314 ≈370 ICSTAGG - Aggression -.15 -.12 -.12 -.08 ICSTACA - Academic competence + .15 .12 .03 .28 ICSTINT - Internalizing .09 .09 .13 .19+ CCCSCOM - Social competence + -.02 -.03 -.04 -.03 CCCPROS - Prosocial + -.03 -.05 -.04 -.01 CCCEREG - Emotion regulation + -.01 -.01 -.04 .03 CCCRAGG - Relational aggression .09 .09 .07 .07 _______________________________________________________________________________________________________ *** p <.001, ** p < .01, * p < .05, + p < .1, two-tailed. Findings: Treatment effects measured by changes in the 4th Grade (g345) Outcome Variable Piecewise Change with Propensity Score (gbm) Weighting Model 1: Hypothetical ATE Sign Grade 4 Piecewise Change with Propensity Score (gbm) Weighting Model 2: ATT Grade 4 Change Score Using Optimal Change FullScore Using Matching Optimal Pair- (gbm) and Matching Hodges(gbm) and Lahmann Regression Aligned Adjustment Rank Test Grade 4 Grade 4 Approximate Sample Size Used in Analysis ≈414 ≈414 ≈380 ≈414 ICSTAGG - Aggression -.17 -.21 -.18 -.21+ ICSTACA - Academic competence + -.08 -.08 -.06 -.13 ICSTINT - Internalizing -.07 -.03 -.02 .13 CCCSCOM - Social competence + .15+ .18* .20 .13 CCCPROS - Prosocial + .15 + .14 + .17 .11+ CCCEREG - Emotion regulation + .17+ .20+ .22 .21+ CCCRAGG - Relational aggression -.15 -.18 -.18 -.15+ _________________________________________________________________________________________________________ *** p <.001, ** p < .01, * p < .05, + p < .1, two-tailed. Findings: Treatment effects measured by changes in the 5th Grade (g345) Outcome Variable Piecewise Change with Propensity Score (gbm) Weighting Model 1: Hypothetical ATE Sign Grade 5 Piecewise Change with Propensity Score (gbm) Weighting Model 2: ATT Grade 5 Change Score Using Optimal Change FullScore Using Matching Optimal Pair- (gbm) and Matching Hodges(gbm) and Lahmann Regression Aligned Adjustment Rank Test Grade 5 Grade 5 Approximate Sample Size Used in Analysis ≈433 ≈433 ≈350 ≈433 ICSTAGG - Aggression -.08 -.08 -.12 -.01 ICSTACA - Academic competence + .20* .20* .16 .16 ICSTINT - Internalizing -.17+ -.18+ -.17 -.24+ CCCSCOM - Social competence + .27** .25** .29 .28* CCCPROS - Prosocial + .29*** .27*** .29 .28* CCCEREG - Emotion regulation + .24* .22* .27 .27+ CCCRAGG - Relational aggression -.20 -.22 -.24 -.24+ _______________________________________________________________________________________________________ *** p <.001, ** p < .01, * p < .05, + p < .1, two-tailed. Findings: Growth curve and dose models (g345) Low Confidence: Do Not Cite Growth Growth curve curve Dose (cum. Dose (cum. Dose (cum. (intervntn* (intervntn* minutes) minutes) minutes) Hypothetical months) months) Grade 3 Grade 4 Grade 5 Outcome Variable Sign ATE ATT ATE ATE ATE (9 months) (9 months) (8 hours) (8 hours) (8 hours) Approximate Sample Size Used in Analysis ≈472 ≈472 ≈370 ≈414 ≈433 ICSTAGG - Aggression -.01 -.02 -.00 -.09 .12 ICSTACA - Academic competence + .03 .03+ .03+ .08 -.03 ICSTINT - Internalizing -.13*** -.14*** .02 -.06 -.08 CCCSCOM - Social competence + .07** .06** .03 .10 -.05 CCCPROS - Prosocial + .06*** .05* .02 .07+ -.10 CCCEREG - Emotion regulation + .07*** .07*** .02 .07 -.10 CCCRAGG - Relational aggression -.05* -.05* .00 -.19** .22+ ____________________________________________________________________________________________________________ *** p <.001, ** p < .01, * p < .05, + p < .1, two-tailed. Findings of Efficacy Subset Analysis: Treatment effects measured by changes in the 3rd and 4th Grades (g34) __________________________________________________________________________________________________ Low Benchmark High Exposure Exposure Exposure Adequate (<900) (900-1044) (1045+) Exposure versus versus versus (240+) versus Hypothetical Comp. (0) Comp. (0) Comp. (0) Comp. (0) Outcome Variable Sign Grade 3 Grade 3 Grade 3 Grade 4 Approximate Sample Size Used in Analysis 343 372 446 545 ICSTAGG - Aggression .11 -.09 .01 -.25** ICSTACA - Academic competence + -.10 .10 .06 -.12 ICSTINT - Internalizing -.04 .10 .15 .09 CCCSCOM - Social competence + -.50*** -.17+ -.23* .08 CCCPROS - Prosocial + -.51*** -.20* -.24** .06 CCCEREG - Emotion regulation + -.39** -.21* -.16+ .08 CCCRAGG - Relational aggression .38*** .10 .21* -.22** __________________________________________________________________________________________________ *** p <.001, ** p < .01, * p < .05, + p < .1, two-tailed. Findings of Efficacy Subset Analysis: Treatment effects measured by changes in the 3rd, 4th, and 5th Grades (g345) Outcome Variable Hypothetical Sign Low Exposure (<900) versus Comp. (0) Grade 3 Benchmark Exposure (900-1044) versus Comp. (0) Grade 3 High Exposure Adequate Adequate (1045+) Exposure Exposure versus (240+) versus (240+) versus Comp. (0) Comp. (0) Comp. (0) Grade 3 Grade 4 Grade 5 Approximate Sample Size Used in Analysis 240 260 295 354 354 ICSTAGG - Aggression .16 -.08 -.10 -.30*** -.10 ICSTACA - Academic competence + .10 .14 .03 -.06 .10 ICSTINT - Internalizing -.17 .19 .14 .04 -.24* CCCSCOM - Social competence + -.28 -.08 -.02 .22* .29** CCCPROS - Prosocial + -.43* -.08 -.07 .20* .30** CCCEREG - Emotion regulation + -.27 -.09 .03 .23* .27** CCCRAGG - Relational aggression .36+ .05 .11 -.31** -.20+ __________________________________________________________________________________________________________ *** p <.001, ** p < .01, * p < .05, + p < .1, two-tailed. Findings of Efficacy Subset Analysis: Treatment effects measured by changes in the 4th & 5th Grades using subsets of Grade 3 exposure (g345) Outcome Variable Hypothetical Sign G3 Low Exposure (<900) versus Comp. (0) Grade 4 G3 Benchmark G3 High G3 Low Exposure Exposure Exposure (900-1044) (1045+) (<900) versus versus versus Comp. (0) Comp. (0) Comp. (0) Grade 4 Grade 4 Grade 5 G3 Benchmark G3 High Exposure Exposure (900-1044) (1045+) versus versus Comp. (0) Comp. (0) Grade 5 Grade 5 Approximate Sample Size Used in Analysis 221 246 285 234 252 288 ICSTAGG - Aggression -.09 -.45*** -.32** .34 -.05 -.26* ICSTACA - Academic competence + .61** -.09 -.22* .19 .17 .02 ICSTINT - Internalizing -.16 -.02 .12 -.50+ -.12 -.17 CCCSCOM - Social competence + .38* .09 .19* .27 .29* .28** CCCPROS - Prosocial + .36* .12 .15 .33 .25+ .28** CCCEREG - Emotion regulation + .39* .10 .22* .09 .29* .28** CCCRAGG - Relational aggression -.19 -.24+ -.32** .09 -.09 -.33** ______________________________________________________________________________________________________________________ *** p <.001, ** p < .01, * p < .05, + p < .1, two-tailed. Summary From different methods of analysis, a pattern of small, cumulative program effects emerges across grades 3, 4, and 5. These analyses exclude one poorly performing that was dissolved in third year of the study. Positive cumulative effects on: Social competence – including • Prosocial behavior and • Skill in regulating emotions Internalizing behavior Relational aggression By HLM and efficacy subsets, promising effects observed on: Academic competence Aggression Focuses on Program Development and Steps in Intervention Research Focuses on (Selection) Bias-Correction Statistical Methods For a description of Making Choices and copies of sample lessons, see http://ssw.unc.edu/jif/makingchoices/ References Abadie, A., Drukker, D., Herr, J. L., & Imbens, G. W. (2004). Implementing matching estimators for average treatment effects in Stata. The Stata Journal 4(3), 290-311. Fraser, M. W., Day, S. H., Galinsky, M. J., Hodges, V. G., & Smokowski, P. R. (2004). Conduct problems and peer rejection in childhood: A randomized trial of the Making Choices and Strong Families programs. Research on Social Work Practice, 14(5), 313-324. Fraser, M. W., Galinsky, M. J., Smokowski, P. R., Day, S. H., Terzian, M. A., Rose, R. A., & Guo, S. (2005).Social information-processing skills training to promote social competence and prevent aggressive behavior in third grade. Journal of Consulting and Clinical Psychology, 73(6), 1045-1055. Fraser, M. W., Nash, J. K., Galinsky, M. J., & Darwin, K. E. (2000). Making choices: Social problem-solving skills for children. Washington, DC: NASW Press. Fraser, M. W., Richman, J. M., Galinsky, M. J., & Day, S. H. (2009). Intervention research: Developing social programs. New York, NY: Oxford University Press. Galinsky, M. J., Terzian, M. A., & Fraser, M. W. (2006). The art of group work practice with manualized curricula. Social Work with Groups, 29(1), 11-26. Guo, S., & Fraser, M. W. (in press). Propensity score analysis: Statistical methods and applications. Thousand Oaks, CA: Sage Press. Hansen, B. B. (2007). Optmatch: Flexible, optimal matching for observational studies. R News, 7, 18-24. Heckman, J. J. (1979). Sample selection bias as a specification error. Econometrica, 47, 153-161. Heckman, J. J. (2005). The scientific model of causality. Sociological Methodology, 35, 1-97. Hirano, K., & Imbens, G. (2001). Estimation of causal effects using propensity score weighting: An application to data on right heart catheterization. Health Services and Outcomes Research Methodology, 2, 259-278. Hodges, J. L., & Lehmann, E. L. (1962). Rank methods for combination of independent experiments in the analysis of variances. Annals of Mathematical Statistics, 33, 482-497. McCaffrey, D. F., Ridgeway, G., & Morral, A. R. (2004). Propensity score estimation with boosted regression for evaluating causal effects in observational studies. Psychological Methods, 9, 403-425. Rosenbaum, P. (1987). Model-based direct adjustment. Journal of the American Statistical Association, 82, 387-394. Rosenbaum, P. (2002). Observational studies (2nd ed.). New York: Springer-Verlag. Rubin, D. B. (2008). For objective causal inference, design trumps analysis. The Annals of Applied Statistics, 2(3), 808-840. Rubin, D. B. (1979). Using multivariate matched sampling and regression adjustment to control bias in observational studies. Journal of the American Statistical Association,74(366), 318-328. Rubin, D. B. (1979). Using multivariate matched sampling and regression adjustment to control bias in observational studies. Journal of the American Statistical Association,74(366), 318-328. Von Hippel, P. T. (2007). Regression with missing Ys: An improved strategy for analyzing multiply imputed data. Sociological Methodology, 37(1), 83-117. Note. CCC=Carolina Child Checklist; ICST = Interpersonal Competency Scale - Teacher Four Stages in the Development of Program Manuals Integrated across the Five Steps in Intervention Research St eps in Int ervention Research Step 1: Specify Problem and Develop Program Theory Step 2: Create and Revise Program Materials Step 3: Refine and Confirm Program Components Step 4: Assess Effectiveness in Variety of Settings and Circumstances Step 5: Disseminate Findings and Program Materials Stage 1 Formulation Stage 2 Revision Stage 3 Differentiation Stage 4 Translation/Adaptation Source: Fraser, M. W., Richman, J. M., Galinsky, M. J., & Day, S. H. (2009). Intervention research: Developing social programs. New York, NY: Oxford University Press.