DIF_Dataset_Master_Comparison.csv

Dataset,Subject Domain,Individual Student Item Responses?,Response Detail (correct/incorrect vs. selected option),Question/Item Text Available?,Demographic/Grouping Variables Available?,Demographics Detail,Scale (students × items),DIF Feasibility Verdict,Download URL,Access Barrier,Key Limitation for DIF
EEDI (Diagnostic Questions / NeurIPS 2020),"Math (diagnostic MCQ, UK curriculum)",YES,Selected option (A/B/C/D) + correct/incorrect,YES — QuestionText + AnswerAText–AnswerDText,"YES — Gender, PremiumPupil (SES proxy)","Gender (M/F), PremiumPupil (binary SES proxy)",~118K students × ~27K items; ~20M interactions,✅ GOLD STANDARD — your current dataset,https://eedi.com/projects/neurips-education-challenge (or Kaggle: riiid),Free registration,"Only 2 demographic variables (gender, SES proxy); UK-specific"
"PISA (OECD, 2022 cycle)","Math, Reading, Science (15-year-olds, 80+ countries)",YES — individual student × item scored responses in STU_COG file,Scored 0/1/2 (correct/partial/incorrect). ❌ NOT raw selected option (A/B/C/D withheld for security),PARTIAL — only ~30-50% of items released in separate PDFs; must manually extract text and link by item ID. Not in data files.,"YES — rich demographics in STU_QQQ file, merge on CNTSTUID","Gender (ST004D01T), SES continuous (ESCS index), Immigration status (IMMIG), Language at home (LANGN), Country (CNT), Age, Grade, School type",~690K students × ~300 items per domain (matrix-sampled: each student sees ~50 items),"✅ STRONG for scored-response DIF (MH, logistic regression, IRT-DIF). ❌ Cannot do distractor-level DIF. ⚠️ Question text only for released subset.",https://www.oecd.org/pisa/data/2022database/,"None — free direct download (.sav, .sas7bdat, CSV)","No raw selected answers; question text only for ~30-50% of items (PDFs, not machine-readable); matrix-sampled booklet design (each student sees only ~50 of ~300 items)"
"TIMSS (IEA, 2019 cycle)","Math & Science (Grades 4 and 8, 60+ countries)",YES — individual student × item scored responses in Student Achievement files (ASA/BSA),Scored responses (correct/incorrect/partial). ❌ NOT raw selected option.,"PARTIAL — ~30-50% of items released in separate PDFs; item ID in PDF matches variable name (e.g., M031001). Must extract text manually. Not in data files.","YES — in Student Background files (ASG/BSG), merge on IDCNTRY+IDSTUD","Gender (ITSEX), Home resources/books (SES proxy), Language at home, Country, School",~300-500K students × ~150-200 items per subject (matrix-sampled: each student sees ~4 blocks),✅ STRONG for scored-response DIF. Same limitations as PISA re: question text and selected options.,https://timssandpirls.bc.edu/timss2019/international-database/,None — free download from IEA. IDB Analyzer tool (free) helps merge files.,Same as PISA: no raw selected answers; question text only for released subset via PDFs; booklet rotation design
"NAEP (NCES, US)","Reading, Math, Science (US, Grades 4/8/12)",❌ NO public individual student responses. Restricted-use data requires 3-6 month federal license application.,"Public: aggregate p-values only. Restricted: scored responses (0/1) per student, sparse BIB design (~25 items/student)",PARTIAL — ~20-30% released items on website with text. Not linkable to individual student responses in any public file.,"YES (restricted data only) — Race/ethnicity, Gender, NSLP (SES), ELL, IEP, Disability, State","Race/ethnicity (5 categories), Gender, NSLP lunch eligibility (SES), ELL status, IEP/disability — richest US demographic data BUT only in restricted files",Tens of thousands per assessment; ~150-300 items per subject (BIB: each student sees ~25),"❌ NOT USABLE without restricted-use license (3-6 month application, institutional IRB, secure data enclave). Even then, BIB design is sparse.",Public explorer: https://www.nationsreportcard.gov/nde/ | Restricted: https://nces.ed.gov/statprog/rudman/,"Public data = aggregates only. Restricted = multi-month federal application, IRB, secure facility required.","No public individual responses. Restricted data is sparse (BIB design). Released items not linkable to individual responses. Fundamentally a population-reporting system, not a research microdata release."
ASSISTments 2012-2013 (Data Mining Challenge),"Middle school math (US, Massachusetts)",YES — one row per student × problem attempt,Binary correct/incorrect. ❌ No selected answer option.,"❌ NO — only problem_id, skill_id, skill_name. No question text in any public ASSISTments release.","YES (partial) — Gender (~50% missing in 2009; better in 2012), School_id, District, State, Economically_Disadvantaged, ELL, IEP","Gender (M/F, may have missing values), School_id, Economically_Disadvantaged (school-level FRPL proxy), ELL status, IEP status",~19K students × ~26K problems; ~708K interactions,⚠️ USABLE for basic DIF (MH by gender or school group) but no question text limits interpretability. Gender missingness is a concern.,https://sites.google.com/site/assistmentsdata/datasets,None — free direct download,"No question text (cannot interpret WHY items show DIF). Gender column has missing values. Demographics are partially school-level (FRPL, ELL) not individual-level."
ASSISTments 2009-2010 (Skill Builder),"Middle school math (US, Massachusetts MCAS items)",YES — one row per student × problem attempt,Binary correct/incorrect. ❌ No selected answer option.,"❌ NO — only problem_id, skill_id, skill_name.",PARTIAL — Gender column exists but ~50% missing. School_id present.,"Gender (M/F, ~50% missing), School_id, Teacher_id",~4.2K students × ~26K problems; ~346K interactions,⚠️ WEAK — gender ~50% missing severely limits DIF. School-level DIF possible.,https://sites.google.com/site/assistmentsdata/datasets | HF: badranx/ASSISTments09,None — free direct download,Gender ~50% missing. No question text. Smaller scale than 2012 version.
"FoundationalASSIST (ASSISTments/WPI, 2026)","Grades 6-8 math (US, Common Core / Illustrative Mathematics)",YES — one row per student × problem; includes exact student answer text,Binary score (discrete_score 0/1) + answer_text (exact student response) + saw_answer + hint_count. For MC items: selected option identifiable.,"YES — full Problem Body (HTML/markup), Fill-in Options/Answers, MC Options/Answers in Problems.csv (3,395 problems)","❌ NO — only user_xid (anonymous student ID). No gender, race, SES, school, or any demographic variable.",NONE — only anonymized user_xid,"5,000 students × 3,395 problems; 1.7M interactions",❌ CANNOT DO STANDARD DIF — no demographic grouping variable. ✅ EXCELLENT for distractor analysis if you can obtain demographics externally (contact etrials@assistments.org).,https://huggingface.co/datasets/ASSISTments/FoundationalASSIST,"Gated — requires data agreement (CC-BY-NC-4.0, non-commercial, form on HF)",No demographic variables at all. Rich item content + distractor data but useless for focal/reference group DIF without external demographic linkage.
"EdNet (Riiid/Santa, Korea)",TOEIC English test prep (Korean students),YES — one row per student × question; includes selected answer (a/b/c/d) + correct answer,Selected answer option (user_answer: a/b/c/d) + correct_answer + is_correct boolean + elapsed_time,"❌ NO — questions.csv has only: question_id, bundle_id, correct_answer, part (1-7), tags (skill IDs). No question text (TOEIC items are copyrighted).",❌ NO — no demographic variables at all. Only platform metadata.,"NONE — no age, gender, race, SES, school, country",784K students × ~13K questions; 131M interactions (KT1 config),❌ CANNOT DO STANDARD DIF — no demographic grouping variable. Item-level response quality is excellent but demographics absent.,HF: https://huggingface.co/datasets/mgor/EDNet (~5.4GB) | GitHub: https://github.com/riiid/ednet,None — free download (HF Hub or GitHub),No demographics whatsoever. No question text (copyrighted). Korean students only. Massive scale is wasted for DIF purposes.
Junyi Academy (Taiwan),"K-12 Mathematics (Taiwanese curriculum, Traditional Chinese)",YES — one row per student × exercise attempt,Binary correct/incorrect + timestamps + exercise_id + skill tags,"❌ NO — only exercise IDs with skill tags (topic, area). No question text released.","❌ NO — no gender, age, grade, school, or any demographic variable in released data.",NONE — confirmed absent across all published papers using this dataset,72K students × 722 exercises; 16M interactions,❌ CANNOT DO DIF — no demographics and no question text.,https://pslcdatashop.web.cmu.edu/DatasetInfo?datasetId=1198,Free PSLC DataShop account required,No demographics. No question text. Content in Traditional Chinese. Previously misclaimed to have gender — it does not.
"OULAD (Open University, UK)",STEM and Humanities university courses,"❌ NO ITEM-LEVEL — only assessment-level scores (one score per TMA/CMA/Exam, not per question)",Single numeric score (0-100) per assessment. No individual question responses.,"❌ NO — no question text, no quiz content. Only VLE click counts on resources.",YES — rich demographics in studentInfo table,"Gender (M/F), Age band, Region (UK), Highest education, IMD band (deprivation/SES proxy), Disability (Y/N)","32,593 students × ~7 assessments per course-module; 22 courses","❌ CANNOT DO ITEM-LEVEL DIF — no individual question responses. Could only do assessment-level analysis (DIF at course-assessment level, not item level).",https://analyse.kmi.open.ac.uk/open_dataset,None — free direct download,No item-level responses (assessment-level only). No question text. Great demographics but wrong data granularity for DIF.
Duolingo SLAM (2018 Shared Task),"Language learning (Spanish↔English, French→English)",YES — token-level error prediction per exercise per learner,Per-word correct/incorrect (whether each word was correctly recalled). Not MCQ format.,YES — exercise text (sentences) available as part of the dataset,"❌ NO direct demographics — only days_in_course, UI language, client type","UI language (quasi-L1 grouping: Spanish vs French speaker), days_in_course, client — NO age/gender/race/SES",~6K learners; 2.6M exercises; 3 language tracks,⚠️ CREATIVE DIF POSSIBLE using L1 (Spanish vs French speaker) as quasi-demographic group. Not standard focal/reference design.,https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/8SWHNO,None — free download from Harvard Dataverse,No true demographic variables. Token-level responses (not MCQ). L1 as grouping is creative but non-standard for DIF.
KDD Cup 2010 (Algebra I / Bridge to Algebra),"Algebra (US, Carnegie Learning Cognitive Tutor)",YES — step-level correct/incorrect per student × problem-step,Binary correct first attempt (CFA). Hint/error counts.,"❌ NO — only problem_id, step_name, KC (skill) labels. No question text.","MINIMAL — School_id (2 schools only), Academic year (2005-06 or 2006-07)","School (2 schools = natural 2-group DIF), Cohort year",~575K rows; ~1.3K students from 2 schools,⚠️ VERY LIMITED — school-as-group DIF between 2 schools possible but weak. No individual demographics.,https://pslcdatashop.web.cmu.edu/KDDCup/,Free PSLC DataShop account,Only 2 schools for grouping. No individual demographics. No question text. Step-level (not item-level).