Data SGP – Reducing the Noise in SGP Estimates

The data sgp package offers an efficient method for organizing longitudinal (time dependent) student assessment data into statistical growth plots. It offers two formats that support this, WIDE where each case/row represents one student and LONG where time dependent variables span multiple rows per student; most SGP analyses require LONG data format whereas choosing between WIDE or LONG can often be a trade off between ease of preparation and speed of analysis – although you should generally be able to run analyses with either format; some higher level wrapper functions may only work properly when using LONG.

Though SGP measures are intended to offer fair assessments of student achievement, research indicates they may contain inherent errors due to imperfect standardized tests used and teachers being assigned according to student background characteristics (Bennett 2002; Akram & Erickson 2013; Lockwood & Castellano 2015). As a result, estimated SGP measures may contain noise that obscure their true measure of latent achievement traits; it’s essential that we know how SGP distributions vary with background variables so as to understand relationships between these measures and teacher effects.

Recent research revealed that SGP estimates obtained through typical teacher observations are far noisier than estimates obtained through value-added models that regress student test scores on teacher fixed effects, prior test score and student covariates. It also discovered that variation in expected true SGP across teachers is driven by individual teacher-student relationships rather than simply sorting or contextual effects; this represents a substantial source of bias when trying to interpret aggregated SGP as an indicator of teacher effectiveness and should therefore be minimized through alternative estimation methods.

In this article we present an approach for reducing noise in SGP estimates using student covariates that are measured continuously. Utilizing covariates makes SGP estimates more stable over time, and is an essential step towards making SGPs an objective and fair measure of student achievement. Additionally, this approach can easily be extended to other content areas and grade levels as well as additional covariates like attendance or socioeconomic status. This approach adds an alternative way of reducing noise in educational statistics and strengthens validity of educational policies based on standardized tests. These results indicate that current efforts to reduce measurement error for standardized tests should also encompass SGPs. It would also facilitate more meaningful interpretations of teacher effects and greater transparency about what causes student achievement. Researchers could then assess whether SGPs really do level the playing field between teachers, removing teacher-specific relationships. This will enable researchers to gauge whether improved standardized tests, SGP modeling techniques, or both, are needed in order to ensure SGPs provide fair measures of student progress.