Report on the Stability of Student Growth Percentile (SGP) “Value-Added” Estimates

ShareTweet about this on TwitterShare on Facebook32Email this to someoneShare on Google+1Share on LinkedIn1Share on Reddit0

The Student Growth Percentiles (SGPs) model, which is loosely defined by value-added model (VAM) purists as a VAM, uses students’ level(s) of past performance to determine students’ normative growth over time, as compared to his/her peers. “SGPs describe the relative location of a student’s current score compared to the current scores of students with similar score histories” (Castellano & Ho, p. 89). Students are compared to themselves (i.e., students serve as their own controls) over time; therefore, the need to control for other variables (e.g., student demographics) is less necessary, although this is of debate. Nonetheless, the SGP model was developed as a “better” alternative to existing models, with the goal of providing clearer, more accessible, and more understandable results to both internal and external education stakeholders and consumers. For more information about the SGP please see prior posts here and here. See also an original source about the SGP here.

Related, in a study released last week, WestEd researchers conducted an “Analysis of the stability of teacher-level growth scores [derived] from the student growth percentile [SGP] model” in one, large school district in Nevada (n=370 teachers). The key finding they present is that “half or more of the variance in teacher scores from the [SGP] model is due to random or otherwise unstable sources rather than to reliable information that could predict future performance. Even when derived by averaging several years of teacher scores, effectiveness estimates are unlikely to provide a level of reliability desired in scores used for high-stakes decisions, such as tenure or dismissal. Thus, states may want to be cautious in using student growth percentile [SGP] scores for teacher evaluation.”

Most importantly, the evidence in this study should make us (continue to) question the extent to which “the learning of a teacher’s students in one year will [consistently] predict the learning of the teacher’s future students.” This is counter to the claims continuously made by VAM proponents, including folks like Thomas Kane — economics professor from Harvard University who directed the $45 million worth of Measures of Effective Teaching (MET) studies for the Bill & Melinda Gates Foundation. While faint signals of what we call predictive validity might be observed across VAMs, what folks like Kane overlook or avoid is that very often these faint signals do not remain constant over time. Accordingly, the extent to which we can make stable predictions is limited.

Worse is when folks falsely assume that said predictions will remain constant over time, and they make high-stakes decisions about teachers unaware of the lack of stability present, in typically 25-59% of teachers’ value-added (or in this case SGP) scores (estimates vary by study and by analyses using one to three years of data — see, for example, the studies detailed in Appendix A of this report; see also other research on this topic here, here, and here). Nonetheless, researchers in this study found that in mathematics, 50% of the variance in teachers’ value-added scores were attributable to differences among teachers, and the other 50% was random or unstable. In reading, 41% of the variance in teachers’ value-added scores were attributable to differences among teachers, and the other 59% was random or unstable.

In addition, using a 95% confidence interval (which is very common in educational statistics) researchers found that in mathematics, a teacher’s true score would span 48 points, “a margin of error that covers nearly half the 100 point score scale,” whereby “one would be 95 percent confident that the true math score of a teacher who received a score of 50 [would actually fall] between 26 and 74.” For reading, a teacher’s true score would span 44 points, whereby one would be 95 percent confident that the true reading score of a teacher who received a score of 50 would actually fall between 38 and 72. The stability of these scores would increase with three years of data, which has also been found by other researchers on this topic. However, they too have found that such error rates persist to an extent that still prohibits high-stakes decision making.

In more practical terms, what this also means is that a teacher who might be considered highly ineffective might be terminated, even though the following year (s)he could have been observed to be highly effective. Inversely, teachers who are awarded tenure might be observed as ineffective one, two, and/or three years following, not because their true level(s) of effectiveness change, but because of the error in the estimates that causes such instabilities to occur. Hence, examinations of the the stability of such estimates over time provides essential evidence of the validity, and in this case predictive validity, of the interpretations and uses of such scores over time. This is particularly pertinent when high-stakes decisions are to be based on (or in large part on) such scores, especially given some researchers are calling for reliability coefficients of .85 or higher to make such decisions (Haertel, 2013; Wasserman & Bracken, 2003).

In the end, researchers’ overall conclusion is that SGP-derived “growth scores alone may not be sufficiently stable to support high-stakes decisions.” Likewise, relying on the extant research on this topic, the overall conclusion can be broadened in that neither SGP- or VAM-based growth scores may be sufficiently stable to support high-stakes decisions. In other words, it is not just the SGP model that is yielding such issues with stability (or a lack thereof). Again, see the other literature in which researchers situated their findings in Appendix A. See also other similar studies here, here, and here.

Accordingly, those who read this report, and consequently seek to find a better or more stable model that yields more stable estimates, will unfortunately but likely fail in their search.

References:

Castellano, K. E., & Ho, A. D. (2013). A practitioner’s guide to growth models. Washington, DC: Council of Chief State School Officers.

Haertel, E. H. (2013). Reliability and validity of inferences about teachers based on student test scores (14th William H. Angoff Memorial Lecture). Princeton, NJ: Educational Testing Service (ETS).

Lash, A., Makkonen, R., Tran, L., & Huang, M. (2016). Analysis of the stability of teacher-level growth scores [derived] from the student growth percentile [SGP] model. (16–104). Washington, DC: U.S. Department of Education, Institute of Education Sciences, National Center for Education Evaluation and Regional Assistance, Regional Educational Laboratory West.

Wasserman, J. D., & Bracken, B. A. (2003). Psychometric characteristics of assessment procedures. In I. B. Weiner, J. R. Graham, & J. A. Naglieri (Eds.), Handbook of psychology:
Assessment psychology (pp. 43–66). Hoboken, NJ: John Wiley & Sons.

ShareTweet about this on TwitterShare on Facebook32Email this to someoneShare on Google+1Share on LinkedIn1Share on Reddit0

2 thoughts on “Report on the Stability of Student Growth Percentile (SGP) “Value-Added” Estimates

Leave a Reply

Your email address will not be published. Required fields are marked *