Research Portfolio

AI, Equity, and the Statistics Classroom

Four studies on how instruction, technology, and policy shape who succeeds in college mathematics, and a discipline of saying plainly where each claim stops.

Safaa Dabagh / MA in Statistics · Mathematics & Statistics Instructor
2 under review 2 preprints
The Through-Line

One question, studied from four angles

I study how instructional design, technology, and policy move student outcomes, especially for the students community colleges serve, using methods matched to the data and every claim scoped to what its design can support.

Pillar 01

AI as an instructor’s tool

Not a student shortcut, a way to raise the design of teaching and assessment.

Bloom’s analysisDistributional effects
Pillar 02

Equity & outcomes

Who moves, and where in the distribution, not just the average.

Placement reformThe floor effect
Pillar 03

Methodological range

Count models, panels, regularization, and honest ecological limits.

Negative-binomial panel
I
Placement & Equity Community College Review (SAGE)
Under Review

Rising Success, Persistent Gaps

When California placed students directly into transfer-level statistics, did outcomes rise or fall, and did they improve equitably across groups?

4269% Course success across the reform era, 2018 to 2025 (ρ = 0.61)
24pt White and Latinx success gap that held steady throughout
Method

867 enrollments, 17 semesters, one instructor held constant; logistic trend with semester-clustered SEs.

Key finding

Success rose for every racial/ethnic group, but broadened access alone did not close the gap.

Why it matters

Direct, student-level evidence in a policy debate usually fought with aggregate numbers.

Scoped as Correlational, a single-instructor design cannot separate the reform from the instructor’s own growth over the same years.

I
Placement & Equity · In detail Rising Success, Persistent Gaps
Under Review
Key results
  • Success 42% → 69% across eras (pre 42 → AB705 49 → post 59 → AB1705 69)
  • Odds of passing +10%/semester (OR 1.10, CI [1.07, 1.13]); Spearman ρ = 0.61
  • Every group rose: White 69→84, Latinx 44→61, Black 27→62, Asian 61→73%
  • White and Latinx gap held ~24 pts
Design & data
  • 867 enrollments, 852 students, 17 semesters, one instructor
  • Enrollment-level logistic regression, semester-clustered SEs; rank trend
What it shows

Direct placement was associated with rising, not eroding, success, but access alone did not close racial gaps.

Limitation

Single-instructor design, correlational; it cannot separate the reforms from the instructor's own growth.

II
AI in Teaching Teaching Statistics
Revise & Resubmit

AI-Assisted Assessment Design

Does designing exams with AI change the cognitive demand of what we ask students, and if so, in which direction?

1848% Higher-order (Analyse-level) items on the statistics final, with no drop in student performance
+30pt Shift toward reasoning over procedure, on a matched pre-AI vs. AI-assisted exam
Method

Bloom’s revised taxonomy; matched finals coded item-by-item across two courses, same instructor.

Key finding

AI raised reasoning demand in statistics and added scaffolding in algebra (mean 61.8→78.7%).

The framing

AI as amplification, acting on design intentions time had made impractical, not replacing judgment.

Replicable The paired-Bloom’s method transfers directly to other courses, instructors, and institutions.

II
AI in Teaching · In detail AI-Assisted Assessment Design
Revise & Resubmit
Key results
  • Statistics: Analyse items 18% → 48% (+30 pts); Apply 47% → 28% (−19)
  • Performance did not decline on the harder exam
  • Three Universal-Design features appeared only in the AI version
  • Algebra: profile stable (85% Apply); mean final 61.8% → 78.7%
Design & data
  • Bloom's revised taxonomy; matched pre-AI vs. AI-assisted finals, coded item by item
  • Two courses, same instructor (34 vs. 25, and 20 vs. 20 items)
What it shows

AI as an amplification mechanism, it lets instructors act on design intentions time had made impractical, not a replacement for judgment.

Method contribution

The paired-Bloom's design is replicable across courses, instructors, and institutions.

III
Distributional Effects EdArXiv preprint
Preprint

Raising the Floor, Holding the Bar

When an instructor redesigns a course with AI, who moves, and where in the grade distribution does the change happen?

+36pt Gain at the 10th percentile in algebra, a floor effect, with the top decile unchanged
315% Share of failing scores (<50%), after AI-assisted redesign
Method

Official gradebooks (N = 106); quantile decomposition; Holm family-wise error control across ten comparisons.

Key finding

Alignment decides value: AI-designed quizzes aligned to the exam stayed predictive; item-bank quizzes lost it.

Why it matters

Means hide the students open-access colleges most need to see. Companion to Paper II.

Discipline Effects reported distributionally and per-course before pooling, a reporting standard this literature still lacks.

III
Distributional Effects · In detail Raising the Floor, Holding the Bar
Preprint
Key results
  • Algebra final +16.9 pts (p = .004, Hedges g = 0.75), a floor effect
  • 10th percentile 22 → 58 (+36); share below 50% 31% → 5%; 90th ~unchanged
  • Statistics: performance maintained on a harder instrument
  • Alignment decides value: AI-aligned quizzes stayed predictive; item-bank quizzes lost it
Design & data
  • Official gradebooks, N = 106; four sections, two courses
  • Quantile decomposition; Holm across ten comparisons; bootstrap; ANCOVA
What it shows

Means hide who moved; AI-assisted design lifted the lowest performers most, and alignment governs whether formative work stays diagnostic.

Limitation

Small cells; unproctored scores inflated toward the ceiling. Reported per course before pooling.

IV
Methodological Range Research Square preprint
Preprint

Disadvantage, Not Policy or Illness

Which state characteristics track mass-shooting rates, gun policy, mental illness, or socioeconomic conditions?

1.45IRR Incidence per SD of poverty, the one predictor stable across every specification
1.04IRR Serious mental illness, no association (gun-law strength likewise, 1.11); LASSO kept only disadvantage
Method

50-state panel, 3,783 incidents; negative-binomial with CR2 cluster-robust SEs; LASSO + tree checks.

Key finding

The post-2020 surge was broad and sustained; its geography tracks disadvantage, not policy or illness.

Why it’s here

Shows range, count models, panel inference, regularization, triangulation.

Scoped as Ecological and observational, no causal or individual claims; sensitive to the “mass shooting” definition.

IV
Methodological Range · In detail Disadvantage & Mass-Shooting Rates
Preprint
Key results
  • 2021 peak IRR 2.02 [1.68 to 2.43]; elevated through 2024 (1.55)
  • Poverty IRR 1.45 [1.18 to 1.79], p = .006; inequality 1.44 (r = 0.50)
  • Gun-law 1.11, ownership 1.07, mental illness 1.04, none associated
  • LASSO retained only the socioeconomic measures
Design & data
  • 50-state panel, 350 state-years, 3,783 incidents (GVA, 2018 to 2024)
  • Negative-binomial + population offset + year FE + CR2 SEs; LASSO & tree checks
What it shows

The post-2020 surge was broad and sustained, and its geography tracks socioeconomic disadvantage, not gun policy or mental illness.

Limitation

Ecological and observational, no causal or individual claims, and sensitive to the “mass shooting” definition.

Where This Points

A research program, not four separate papers

Together they make the case: an emerging scholar of AI and equity in postsecondary education, grounded in rigorous quantitative method, and already moving toward doctoral work on AI-powered cognitive scaffolding in statistics education, with data in hand.

“I report what the design supports, and I say plainly where it stops.”

learnwithoutwalls.com dabagh_safaa@smc.edu Safaa Dabagh · 2026
1 / 11
Speaker notes

    ← → navigate  ·  N notes  ·  click a dot to jump