Growing Minds

Where do you sit among 50,000 childhoods? A public data-education page, not a technical dashboard: build a rough profile from your own background, see it placed on the population as a range rather than a single number, then explore how the underlying factors relate to each other and to a synthetic measure of childhood IQ. Three acts — Where do I sit?, What's associated with this?, What can be changed? — each grounded in a short, plain-language note on the reasoning and research behind it, not just the chart itself.

Live dashboard: https://forageopen.github.io/growing-minds/

Dataset

Child IQ: Genes & Environment (Kaggle, sergionefedov) — 50,000 rows × 30 columns covering parental IQ and SES, prenatal factors (iodine status, smoking, alcohol, gestational age), early environment (home stimulation, books, screen time, nutrition), and two targets: continuous child_iq and binary high_cognitive_potential (top decile).

Validated at build time: no missing values, no duplicate child_ids, no malformed rows. See data/processed/validation_report.json for the current report and data/raw/data_dictionary.csv for the full column reference.

Repository layout

data/
  raw/                 source CSV + gzip + data dictionary (source of truth)
                        (personal research-reference notes also live here,
                        gitignored — see "Sources & citations" below)
  processed/            typed data dictionary, validation report
scripts/
  build-data.mjs        validates the raw CSV, emits dashboard-ready JSON,
                        fits the profile-builder regression
src/frontend/
  public/data/          precomputed aggregates the dashboard fetches (no
                         client-side parsing of the 50k-row CSV)
  src/
    charts/              hand-built SVG chart components
    components/          shared UI — cards, stat tiles, tables, tooltips,
                          the profile builder, and the long-form content
                          components (AcademicNote, ReadingList,
                          ProjectStructure)
    lib/                 scales, stats (KDE, multiple regression), the
                          profile scoring engine, data hooks
    styles/tokens.css     design tokens sourced from @atlaskit/tokens,
                          validated for categorical color accessibility
.github/workflows/deploy.yml   builds and publishes to GitHub Pages

Data pipeline — "optimize for parsing"

The dashboard never parses the raw CSV in the browser. scripts/build-data.mjs reads data/raw/child_iq_master.csv once, validates it, and writes small precomputed JSON files (overview, distributions, flynn_effect, by_category, by_binary, correlations, correlation_matrix, scatter_sample, dictionary) into src/frontend/public/data/. The full combined payload is well under 250 KB. Re-run it whenever the raw dataset changes:

node scripts/build-data.mjs

The CI workflow re-runs this script before every deploy, so the published dashboard always reflects data/raw/child_iq_master.csv.

Dashboard

Built with Vite + React + TypeScript. No charting library — every chart is a hand-rolled, accessible inline SVG component, chosen to avoid generic Excel-style bar/pie/donut charts in favor of forms suited to each question:

Chart	Form
Flynn effect over time	Line chart with a ±1 SD band
Predictor ranking	Diverging lollipop chart
Predictor inter-relationships	Correlation heatmap matrix
Variable shape	Smoothed density (KDE) curve
Category comparison	Cleveland dot plot vs. population mean
Risk-factor effect size	Dumbbell chart
Parent IQ vs. child IQ	Scatter plot with marginal distributions + trend line

Design system: skinned with Atlassian Design System tokens (@atlaskit/tokens) for surfaces, ink, and the 8-hue categorical chart palette, validated for CVD-safe adjacency and contrast against both light and dark chart surfaces. Icons from Lucide. Light/dark theme follows the OS setting with a manual toggle.

Profile builder — the scoring engine

The "where do I sit" number is a transparent composite index, computed entirely in the browser (src/frontend/src/lib/profile.ts) from whichever fields a visitor chooses to answer — nothing is sent anywhere.

• Categorical/binary fields contribute their dataset group-mean deviation from the population mean.
• Numeric fields contribute a standardized multiple-regression coefficient, not a raw correlation — coefficients are solved once at build time (scripts/build-data.mjs, Gauss-Jordan matrix inversion over the predictor correlation matrix) so that overlap between correlated predictors (e.g. maternal education, home stimulation, and nutrition all partly tracking the same socioeconomic status) is only counted once. A naive sum of each field's own correlation overstates the combined effect by roughly 38% on this dataset.
• Each field's z-score is clamped to ±4 SD before it's applied. Sliders are intentionally widened past the dataset's typical range so real answers outside the norm (e.g. multi-year breastfeeding) aren't clipped, but a linear model fit on this data has no business extrapolating a raw z-score in the double digits — the clamp keeps the wider slider UX honest without producing an implausible score.
• The index deliberately excludes parental IQ and other unmeasured traits — the strongest individual predictors in the dataset — so it will rarely reach the extremes a real IQ test could. This trade-off, and its cost, is disclosed on the page itself (see "Why variable selection matters").

Sources & citations

In-page research notes (AcademicNote) and the "Further reading" list (ReadingList.tsx) cite real, published sources for the claims made about population statistics, self-assessment, data visualization, and child development — linked via doi.org. Two personal research-reference documents used while drafting that copy live in data/raw/ but are gitignored and not published, since they contain mirror links alongside the citations; only the extracted citation metadata (author, year, title, venue, DOI) is used in the app.

License

Code: MIT, see LICENSE. Dataset: synthetic, released by its author on Kaggle for research/educational use — see the dataset page for terms.