LEVANTE Longitudinal Analyses
Exploratory psychometrics of the core-task battery
The code and prose in this chapter were written by Claude Code (an AI coding agent) under human direction. They have not been line-by-line verified by a human and must not be published or cited as-is.
These are exploratory analyses of preliminary, in-progress LEVANTE data. Results have not been reviewed or validated and require checking by the LEVANTE Data Coordinating Center before any use.
LEVANTE longitudinal analyses
An exploratory analysis of the LEVANTE core-task battery — a sequence of reproducible notebooks that walk from raw data validation through construct structure to within-child variability. Data come from the per-site processed datasets on Redivis (pinned 2026-08 snapshot — the unified levante_data_latest is still at its June v1.2 release, so the loaders bind the per-site datasets directly; see chapter 00) via the levante / levantemodels packages; shared conventions, palettes, and loaders live in common.R. The 2026-08 update adds rural Colombia’s second wave (the project’s third substantial longitudinal cell) and four newly-collecting cross-sectional sites.
Each chapter computes its results inline from the data, so every number and figure is current with the code that produced it.
The arc
Longitudinal analyses walk from raw data validation through construct structure to within-child variability and network context.
Data & integrity — pull and clean the data (00); audit per-(site, task) integrity, adaptive/mode artifacts, and extreme scores (01).
Growth & construct structure — longitudinal traces and simple growth (02); the g-vs-domains structure and cross-site measurement invariance (03); emergent dimensions, age-differentiation, and the role of g with ROAR + MEFS as real measures (04); and how short the battery can be while still recovering the factor scores (05).
Within-child variability — a first pass at three consistency indices and whether they cohere (06); a deep dive on RT intra-individual variability — purification, cross-task structure, site/age, and longitudinal stability (07); and the accuracy companion using MSSD, and why nothing survives de-confounding (08).
Network context — enrollment ages and longitudinal coverage across the LEVANTE network (09).
Task deepdives & cross-task analyses
Per-task psychometric reports, plus cross-task analyses. Task deepdives cover each core task and the ROAR literacy measures: item structure and difficulty, DIF, answer-key forensics, and scoring audits — Hearts & Flowers, Memory, Same & Different (IRTree scoring models and recode audit), Math (plus item calibration and EAP-vs-ML scoring), Shape Rotation, Sentence Understanding (TROG), Vocabulary, Stories (ToM; one consolidated analysis on the corrected 2026-09 data, with the reality-check bug forensics as an appendix), ROAR-Word, ROAR-Phoneme/Sentence, and the ROAR-Sentence (SRE) scoring deepdive. Cross-task analyses collect the corrected longitudinal trajectories, the cross-language DIF batch, test–retest figures, and the training-vs-development comparison.
A few headline findings
- A column-order IRT scoring bug in an earlier data release mis-scored the CAT/guessing-floor items; found, validated, fixed upstream (data re-released as v1.2). With corrected scores, apparent longitudinal declines disappear.
- The core battery is a strong general factor with weakly-separable domains that differentiate with age; structure replicates across Germany, Bogotá, and Canada.
- The battery is highly compressible — administration time can be roughly halved while still recovering every construct at ≈ 0.9 determinacy.
- Within-child variability is elusive but not empty: RT variability carries a real, narrow speeded-task signal (cleanest after detrending item difficulty) whose composite reaches test–retest ≈ 0.63 with the 2026-08 rural retest data; accuracy variability, properly de-confounded, is essentially ability in disguise.
Reproducibility & provenance
The notebooks render from Redivis data (gitignored; access via levantemodels). common.R is sourced by every chapter. This book was assembled with Quarto; see the source repository.
This is internal, exploratory work of the LEVANTE Data Coordinating Center — please read the two banners above before using anything here.