1. PRESENTATION AT THE 12 TH ANNUAL MARYLAND ASSESSMENT CONFERENCE COLLEGE PARK, MD OCTOBER 18, 2012 JOSEPH A. MARTINEAU JI ZENG MICHIGAN DEPARTMENT OF EDUCATION Borrowing the Strength of Unidimensional Scaling to Produce Multidimensional Educational Effectiveness Profiles

2. 2 Background Prior research showing that using unidimensional measures of multidimensional achievement constructs can distort value-added  Martineau, J. A. (2006). Distorting Value Added: The Use of Longitudinal, Vertically Scaled Student Achievement Data for Value-Added Accountability. Journal of Educational and Behavioral Statistics, 31(1), 35-62.  Construct irrelevant variance can become considerable in value-added measures when a construct is multidimensional, but is modeled in value- added as unidimensional.  Common misunderstanding is that if the multiple constructs are highly correlated, value-added should not be distorted.  Correct understanding is that if value-added on the multiple constructs is highly correlated, value-added should not be distorted

3. 3 Background Prior research showing that the choice of dimension/domain within construct changes value-added significantly  Lockwood, J.R et al. (2007). The Sensitivity of Value-Added Teacher Effect Estimates to Different Mathematics Achievement Measures. Journal of Educational Measurement, 44(1), 47-67.  Depending on choices made in value-added modeling, the correlation between teacher value-added on Procedures and Problem Solving ranged from 0.01 to 0.46.  This gives a surprisingly low correlation in value-added that indicates that at least in this situation, one needs to be concerned about modeling value- added in both dimensions rather than unidimensionally.  Only work I am aware of to date that has inspected inter-construct value- added correlations.

4. 4 Background Prior research showing that commonly used factor analytic techniques underestimate the number of dimensions in a multidimensional construct  Zeng, J. (2010). Development of a Hybrid Method for Dimensionality Identification Incorporating an Angle-Based Approach. Unpublished doctoral dissertation, University of Michigan.  Common dimensionality identifications procedures make the unwarranted assumption that all shared variance among indicator variables arise because the indicator variables measure the same construct (shared variance can also arise because the indicator variables are influenced by a common exogenous variable)  Because of this unwarranted assumption, commonly used dimensionality identification techniques underestimate the number of dimensions in a data set.

5. 5 Background Scaling constructs as multidimensional is a difficult task  Multidimensional Item Response Theory (MIRT) is time- consuming and costly to run  Replicating MIRT analyses can be challenging (there are multiple subjective decision points along the way)  Identifying the number of dimensions in MIRT can be challenging  Once the number of dimensions is identified, identifying which items load in which dimensions in MIRT can also be challenging  The factor analysis techniques underlying MIRT are techniques for data reduction, not dimension identification

6. 6 Background Short of resolving the considerable difficulties in analytically identifying dimensions within a construct (and replicating such analyses), can another approach be used? Propose using/trusting content experts’ identifications of dimensions within constructs (e.g., the divisions agreed upon by the writers of content standards) as the best currently available identification of dimensions, for example…  Within English language proficiency, producing reading, writing, listening, and speaking scales.  Within Mathematics, producing number & operations, algebra, geometry, measurement, and data analysis/statistics scales.

7. 7 Background However, separately scaling each dimension can also be difficult and costly compared to running a traditional unidimensional IRT calibration  Confirmatory MIRT  Bi-factor IRT model  Separate unidimensional calibration and year-to-year equating of each dimension score Another option:  Unidimensionally calibrate the total score  Unidimensionally equate the total score from year to year  Use (fixed) item parameters from the unidimensional calibration to create the multiple dimension scores as specified by content experts  Use of this method needs to be investigated Practical necessity for Smarter Balanced Assessment Consortium

8. 8 Purpose Investigate the feasibility and validity of relying on unidimensional total score calibration as a basis for creating multidimensional profile scores…  For reporting multidimensional student achievement scores  For reporting multidimensional value-added measures Investigate the impact of separate versus fixed calibration of multidimensional achievement scores in terms of impact on…  Student achievement scores  Value-added scores …as compared to the impact of other common decisions in scaling, outcome selection, and value-added modeling

9. 9 Methods Decisions Modeled in the Analyses  Psychometric decisions  Choice of psychometric model 1-PL vs. 3-PL PCM vs. GPCM  Estimation of sub-scores Separate calibration for each dimension vs. fixed calibration based on unidimensional parameters  Choice of outcome metric  Which sub-score is modeled  Value-added modeling decisions  Inclusion of demographics in models  Number of pre-test covariates (for covariate adjustment models)

10. 10 Methods Outcomes  Correlations in student achievement metrics compared across each psychometric choice and outcome choice  Correlations in value-added modeling compared across each choice  Classification consistency in value-added compared across each choice for  Three-category classification decisions Based on confidence intervals around point-estimates placing programs/schools into three categories: (1) above average, (2) statistically indistinguishable from the average, and (3) below average  Four-category classification decisions Based on sorting programs’/schools’ point estimates into quartiles, representing arbitrary cut points for classification

11. 11 Methods Data  Michigan English Language Proficiency Assessment (ELPA)  Level III (Grades 3-5)  3391 students each with 3 measurement occasions (10,173 total scores)  Measures  Total  Reading(domain)  Writing (domain)  Listening(domain)  Speaking(domain)  Calibrated the ELPA as a unidimensional measure using both 1- PL/Partial Credit Model and 3-PL/Generalized Partial Credit Model  Created domain scores both from fixed parameters from unidimensional calibration and in separate calibrations for each domain

12. 12 Methods Data  Michigan Educational Assessment Program (MEAP) Mathematics  Grades 7 and 8 (not on a vertical scale)  Over 110,000 students per grade  Measures  Total(using items from the two domains)  Number & Operations(domain)  Algebra (domain)  Calibrated the MEAP Math tests as unidimensional measures using both 1-PL and 3-PL models  Created domain scores both from fixed parameters from unidimensional calibration and in separate calibrations for each domain

13. 13 Methods

14. 14 Methods Value-added modeling the ELPA  VAMs were run in a fully-crossed design with…  All outcomes (R, W, L, S)  PCM- and GPCM-calibrated outcomes  Fixed and separately calibrated outcomes  With and without demographics in the VAMs  32 real-data applications across design factors

15. 15 Methods

16. 16 Methods Value-added modeling MEAP mathematics  VAMs were run in a fully-crossed design with…  Both outcomes (algebra and number & operations)  1-PL and 3-PL calibrated outcomes  Fixed and separately calibrated outcomes  With and without demographics  With either one or two pre-test covariates  32 real-data applications across design factors

17. 17 Results ELPA

18. 18 Results: ELPA Student-Level Outcomes Correlations across fixed vs. separate calibrations

19. 19 Results: ELPA Student-Level Outcomes Correlations across model choice (PCM vs. GPCM)

20. 20 Results: ELPA Student-Level Outcomes Correlations across content areas Low to moderate inter-dimension correlations However, Rasch dimensionality analysis from WINSTEPS identified the total score as a unidimensional score

21. 21 Results: ELPA Program District-Level Value-Added Outcomes Impact of fixed versus separate calibration

22. 22 Results: ELPA Program District-Level Value-Added Outcomes Correlations between Listening and Reading VA Min = 0.228, Max = 0.397 Mean = 0.322, SD = 0.037

23. 23 Results: ELPA Program District-Level Value-Added Outcomes Correlations between Listening and Writing VA Min = 0.342, Max = 0.420 Mean = 0.373, SD = 0.019

24. 24 Results: ELPA Program District-Level Value-Added Outcomes Correlations between Listening and Speaking VA Min = -0.005, Max = 0.108 Mean = 0.046, SD = 0.035

25. 25 Results: ELPA Program District-Level Value-Added Outcomes Correlations between Reading and Writing VA Min = 0.335, Max = 0.491 Mean = 0.412, SD = 0.047

26. 26 Results: ELPA Program District-Level Value-Added Outcomes Correlations between Reading and Speaking VA Min = 0.121, Max = 0.205 Mean = 0.151, SD = 0.026

27. 27 Results: ELPA Program District-Level Value-Added Outcomes Correlations between Speaking and Writing VA Min = 0.150, Max = 0.246 Mean = 0.199, SD = 0.029

28. 28 Results: ELPA Program District-Level Value-Added Outcomes Impact of choice of psychometric model

29. 29 Results: ELPA Program District-Level Value-Added Outcomes Impact of Including/Not Including Demographics

30. 30 Results MEAP Mathematics

31. 31 Results: MEAP Math Student-Level Outcomes Correlations among variables based on psychometric decisions

32. 32 Results: MEAP Math Student-Level Outcomes Very high correlations based on fixed versus separate calibrations

33. 33 Results: MEAP Math Student-Level Outcomes Very high correlations based on fixed versus separate calibrations

34. 34 Results: MEAP Math Student-Level Outcomes Not as high correlations based on 1-PL versus 3-PL calibrations

35. 35 Results: MEAP Math Student-Level Outcomes Moderate to high correlations across dimensions

36. 36 Results: MEAP Mathematics School-Level Value-Added Outcomes Impact of fixed versus separate calibration

37. 37 Results: MEAP Mathematics School-Level Value-Added Outcomes Impact of choice of outcome (Algebra vs. Number)

38. 38 Results: MEAP Mathematics School-Level Value-Added Outcomes Impact of choice of psychometric model

39. 39 Results: MEAP Mathematics School-Level Value-Added Outcomes Impact of Including/Not Including Demographics

40. 40 Results: MEAP Mathematics School-Level Value-Added Outcomes Impact of covarying on one vs. two pre-test scores

41. 41 Conclusions Practically important impacts on value-added metrics and value-added classifications  Choice of psychometric model  Including/not including demographics  Including/not including multiple pre-test values Prohibitive impacts on value-added metrics and value-added classifications  Choice of outcome (i.e., domain within construct) Practically negligible impacts on value-added metrics and value-added classifications  Separate versus fixed calibrations of domains within construct

42. 42 Conclusions, continued… Need to pay attention to modeling domains within constructs if constructs can reasonably be considered multidimensional Of the common psychometric and statistical modeling decisions one can make, the choice of which subscore to use as an outcome is the most influential Because subscores give different profiles of both student achievement and program/school value-added, each subscore should be modeled to the degree possible 4-category (i.e., quartile) classifications on value-added are appreciably impacted by every psychometric and statistical modeling choice evaluated here, but 3-category classifications are not  Discourage more than three categories  RTTT requires at least four categories

43. 43 Conclusions, continued… 3- vs. 4-category distinction is actually a proxy for  Statistical decision categories (3-categories)  Arbitrary cut point categories (4-categories) Can leverage unidimensional calibrations of multidimensional achievement scales to create multidimensional profiles of value-added  Except where using four categories of classifications

44. 44 Limitations Inductive reasoning  Results are likely to hold in similar circumstances  Still will need to investigate feasibility of using fixed parameters from unidimensional calibration for specific circumstances if those circumstances are high stakes  This is a proof of concept PCM and GPCM models were run using different software (WINSTEPS vs. PARSCALE)

45. 45 Contact Information Joseph A. Martineau, Ph.D.  Executive Director  Bureau of Assessment & Accountability  Michigan Department of Education  martineauj@michigan.gov martineauj@michigan.gov Ji Zeng, Ph.D.  Psychometrician  Bureau of Assessment & Accountability  Michigan Department of Education  zengj@michigan.gov zengj@michigan.gov