What I Did
My contributions spanned the full research lifecycle — from data processing and methodology design, to analysis, visualization, and writing. My primary analytical focus was the demographic gap analysis and predictive modeling: I conducted achievement gap analyses across five dimensions — socioeconomic status, ethnicity, gender, English Language Learner status, and disability status — using z-score standardized scores to enable valid cross-year comparisons despite grading scale changes in 2018 and 2022. For binary factors I ran independent samples t-tests with Cohen's d effect sizes; for multi-category factors I applied one-way ANOVA and pairwise effect size comparisons. I then built a Ridge regression model using Scikit-learn to identify the relative predictive weight of each demographic variable while managing severe multicollinearity in the data (VIF > 10⁶).
Key Findings
English Proficiency, Disability, and SES Are the Strongest Predictors
The Ridge regression identified English proficiency, disability status, and socioeconomic status as the three strongest predictors of math performance — well ahead of ethnicity and gender once other variables were controlled for. Much of what appears to be an ethnicity gap in raw descriptive data is actually attributable to language access and economic disadvantage rather than ethnicity itself.
A Stable 1.8 Standard Deviation Economic Gap
Non-disadvantaged students consistently scored approximately 0.9 standard deviations above the annual average, while economically disadvantaged students scored 0.9 below — a stable 1.8 standard deviation gap across the entire decade. Effect sizes peaked at Cohen's d = 9.8 in 2018–2019, far exceeding the conventional 'large effect' threshold of 0.8, and remained substantial even as various policy initiatives came and went.
The ELL Advantage: Former ELL Students Outperform Peers
Students who had successfully exited ELL programs (Ever ELL) consistently outperformed not only current ELL students but also students who had never received ELL services — a pattern that held across the entire decade. This suggests that well-designed language acquisition support may confer cognitive benefits that extend beyond language itself, with implications for how schools think about early intervention and program exit criteria.
Disability Gaps Exceed National Benchmarks — Unchanged Over a Decade
Achievement gaps for students with disabilities were among the largest in the dataset, with Cohen's d values ranging from 6.44 to 16.72 — substantially exceeding national benchmarks — and showing no meaningful reduction over the decade despite various policy reforms.
Connection to Learning Design
For instructional designers and learning product teams, the practical implication of this work is less about the specific findings and more about the analytical approach. Large-scale behavioral data can reveal structural patterns that smaller studies miss — and those patterns should directly inform where intervention resources are targeted.
The same logic applies in organizational and product contexts. When a learning program shows uneven outcomes across participant groups, the instinct is often to look at content or delivery. But the more useful question is: which variables are actually driving the difference, and how much of the apparent gap disappears once confounding factors are controlled for? That is precisely what regression-based learning analytics enables — and it is the same framework used here at population scale. In EdTech product development, this kind of analysis can inform user segmentation, personalization decisions, and the prioritization of feature investment for underserved learner groups.