StatLens

A free, browser-only statistics tool for the whole analysis loop — from raw data to a publication-quality figure.

→ Try it: ronnyallen.github.io/StatLens

StatLens runs a real Python scientific stack — NumPy, SciPy, statsmodels, pingouin, lifelines — inside your browser via Pyodide and WebAssembly. ~45 statistical tests, eight typed data table types, and 12 chart types with automatic significance brackets.

Nothing to install. No server. Your workbooks are saved to your own Google Drive.

Statistical correctness is the top priority; UX polish is second. 44 analyses are pinned to golden reference values from SciPy, statsmodels, pingouin, and R — and re-verified in CI on every push.

---

Table of contents

• Why
• Features
• What goes in which table
• Charts
• Coming soon
• How it works
• Correctness
• Privacy & data
• Code map
• Acknowledgements
• License

---

Why

Statistics software tends to be expensive, desktop-bound, or both. StatLens does the everyday work — put data in a typed table, run the right test, get a clean graph with significance brackets — in a browser tab, for free.

• Free and open. No licence, no seat count, no install.
• Actually correct. Every statistic is checked against a published reference implementation, not eyeballed. See Correctness.
• Yours. Your data goes to your Drive, not our server — there isn't one.
• Portable & Installable. Any machine with a modern browser. It functions as a Progressive Web App (PWA): you can install it to your device and, thanks to aggressive Service Worker caching, it loads instantly and runs offline!

---

Features

• 8 data-table types — XY, Column, Grouped, Contingency, Survival, Parts of Whole, Multiple Variables, Nested
• Spreadsheet grid with multi-cell selection, Excel copy/paste, and 50-step undo/redo
• Descriptive statistics with assumption checks — Shapiro-Wilk normality, Levene equal variance
• A test recommender that reads those assumptions — with manual override
• ~45 statistical tests (see the table guide)
• One- and two-tailed options for t-tests and Mann-Whitney
• Post-hoc tests — Tukey, Dunnett, Dunnett's T3, Games-Howell, Dunn, Bonferroni, Šídák, Holm-Šídák — with adjusted p-values and confidence intervals
• 12 chart types with significance brackets and asterisks
• Publication export — 600-DPI PNG (correct DPI metadata, embedded fonts) and SVG
• Google Drive save/load, light and dark themes
• Workspace organization — label workbooks with custom tags, assign them distinct colors, and apply multiple tags per workbook.
• Progressive Web App (PWA) — Install it as a standalone app with instant offline loading via background Service Worker caching.

An in-app About page carries the same guide, kept verified against the engine.

---

What goes in which table

The table type you pick determines which analyses are offered.

Table	Use when	Example	Tests	Post hoc
Column	Comparing groups on one measurement	Control / Drug A / Drug B tumour volumes	One-sample t, Unpaired t (Student's), Welch's t, Paired t, Mann-Whitney, Wilcoxon, Sign test, Kolmogorov-Smirnov, One-way ANOVA, Welch's ANOVA, Brown-Forsythe, Repeated-Measures ANOVA, Kruskal-Wallis, Friedman	Tukey, Dunnett (vs control), Dunnett's T3, Games-Howell, Dunn, Bonferroni, Šídák, Holm-Šídák
XY	One variable measured against another	Dose vs response; time vs signal	Pearson, Spearman, Simple linear regression, Nonlinear curve fitting (exponential, Michaelis-Menten, 4PL, Gaussian, polynomial, Boltzmann), Deming regression, Simple logistic regression, Area under curve, LOWESS, Spline, Smooth, Integrate, Differentiate	—
Grouped	Two factors at once	Drug × Dose on blood pressure	Two-way ANOVA (Type III), Repeated-Measures two-way, Mixed-effects, ART ANOVA (non-parametric)	Tukey on marginal means
Contingency	Counts in categories	Exposed / Unexposed × Case / Control	Chi-square (± Yates), Fisher's exact, McNemar's, Odds and risk ratios, Diagnostic (sensitivity / specificity)	—
Survival	Time until an event	Days to relapse by treatment arm	Kaplan-Meier, Log-rank (Mantel-Cox), Gehan-Breslow-Wilcoxon, Hazard ratios, Cox regression	Pairwise log-rank (Holm)
Parts of Whole	One whole split into parts	Cell counts per phenotype	Fraction of total, Chi-square goodness of fit, Binomial test	—
Multiple Variables	Many variables per subject	Age, BMI, dose, response per patient	Correlation matrix, Multiple linear regression, Multiple logistic regression, Poisson regression, PCA, Three-way ANOVA	—
Nested	Subsamples within groups	3 wells per animal, 4 animals per group	Nested t-test / nested one-way ANOVA	Tukey

Analyses report effect sizes (Cohen's d, eta², mean differences), confidence intervals, and degrees of freedom, alongside a Markdown results report.

---

Charts

Bar + error bars · Box & whisker · Violin · Raincloud · Scatter · Strip · Jitter · Beeswarm · Horizontal box · Range / dumbbell · CI forest · Kaplan-Meier step

All support significance brackets and asterisks, computed and laid out automatically. Error bars (SD / SEM / CI) and regression overlays with confidence bands. Light/dark theme, 15 bundled font families.

Export: 600-DPI PNG with correct DPI metadata and embedded fonts, or SVG.

---

Coming soon

• Excel / CSV export of data and results tables (today only graphs export)
• Guided Analysis wizard — a step-by-step walkthrough
• PDF / EPS graph export
• More chart types — grouped bar, connected / before-after line, nested plot, pie / donut, heatmap
• Local hard-drive file saving (offline storage)

---

How it works

StatLens is a 100% client-side single-page app. Nothing runs on a server.

┌──────────────────────── Browser tab ─────────────────────────┐
│                                                              │
│  React 19 + TypeScript (Vite 8)                              │
│    ├── ag-grid ............ spreadsheet data tables          │
│    ├── visx / d3 .......... chart rendering (SVG)            │
│    └── Radix + Tailwind ... UI (shadcn/ui)                   │
│                    │                                          │
│                    │  postMessage                             │
│                    ▼                                          │
│  Web Worker ──► Pyodide 0.26.1 (WebAssembly)                 │
│                   ├── numpy, pandas, scipy                    │
│                   ├── statsmodels, scikit-learn               │
│                   └── pingouin, scikit-posthocs, lifelines    │
│                   └── analysis_engine.py  (3.1k lines)        │
│                                                              │
└──────────────┬───────────────────────────────────────────────┘
               │  HTTPS
               ▼
   Google Drive (drive.file scope) ── your workbooks, your account

The parts that matter:

1. The stats engine is real Python. src/stats/analysis_engine.py (~3,100 lines) is the single source of statistical truth. It executes under Pyodide inside a Web Worker, so heavy analyses never block the UI.
2. Pyodide loads from a CDN at runtime, but caches locally. The Python runtime and scientific packages are not in the JS bundle. First load pulls tens of MB — the header shows Loading engine… → Engine ready. After that, the Service Worker caches everything locally, meaning the app loads instantly on subsequent visits (even offline) and updates silently in the background (StaleWhileRevalidate).
3. Optional packages install via micropip (pingouin, scikit-posthocs, lifelines) behind defensive imports — if one fails to install, the engine still loads and everything that doesn't need it still works.
4. No SharedArrayBuffer, so no COOP/COEP headers are required and the whole app can be served as plain static files.
5. Storage is Google Drive. Workbooks are JSON files named *.statlens, saved into a StatLens folder in your Drive.

Tech stack

Layer	Choice
Build	Vite 8 (Rolldown), TypeScript 6
UI	React 19, Tailwind CSS 3.4, Radix UI / shadcn/ui, lucide-react
Grid	ag-grid-community 35
Charts	visx 4, d3-scale / d3-shape / d3-array
Routing	react-router-dom 7
Validation	zod 4
Stats	Pyodide 0.26.1 → NumPy, pandas, SciPy, statsmodels, scikit-learn, pingouin, scikit-posthocs, lifelines
Auth	Google Identity Services
Lint	oxlint
Tests	vitest + a Python-side CI matrix

---

Correctness

The project's central rule, from Agents.md:

Statistical correctness is non-negotiable. Every statistic/test has a unit test against known reference values (SciPy/statsmodels/pingouin/lifelines docs, R, or published worked examples) with sensible tolerance. Never change an algorithm without updating tests + comments.

What that looks like in practice:

• golden_values.json — 44 analyses pinned to reference outputs
• reference_R/ — the R scripts that generated those expected values
• run_audit.test.ts — boots Pyodide and runs the real engine against the golden values, not a mock
• charts/__tests__/geometry.test.ts — chart geometry maths
• docs/AUDIT_MATRIX.md — per-analysis status: does it run, is it golden-tested, does it go through the engine path

Statistics CI

.github/workflows/stats-tests.yml runs on every push and PR, against a Python stack pinned to match what the Pyodide worker installs — so CI and the browser can't silently diverge:

• Golden-value verification — the reference numbers still hold
• Post-hoc contract matrix — including a Dunnett determinism guard
• UI ↔ engine string boundary contract — catches the class of bug where the UI asks for one test and the engine quietly runs another
• Typecheck — full tsc pass

An analysis without a reference test is considered unfinished.

---

Privacy & data

• There is no StatLens server. The app is static files; all computation happens in your browser.
• Your data goes to your own Google Drive, in a folder named StatLens, as *.statlens JSON files.
• The app requests the drive.file scope, which grants access only to files the app itself creates. It cannot read the rest of your Drive.
• Your access token is held in localStorage for the session and revoked on sign-out.
• The only third-party network calls are to Google (sign-in + Drive) and to the CDN/PyPI that serve the Python runtime and packages.

---

Code map

StatLens/
├── apps/web/src/
│   ├── App.tsx               # routes, header, engine bootstrap
│   ├── pages/AboutPage.tsx   # in-app about + usage guide
│   ├── stats/
│   │   ├── analysis_engine.py    # ★ all statistics live here (~3.1k lines)
│   │   ├── pyodide.worker2.ts    # Web Worker: boots Pyodide, runs the engine
│   │   ├── engine.ts             # main-thread client for the worker
│   │   └── __tests__/            # golden-value oracle tests
│   ├── charts/
│   │   ├── GraphEngine.tsx       # chart dispatch
│   │   ├── {Column,XY,Survival}Charts.tsx
│   │   ├── SignificanceLayer.tsx # significance bracket layout
│   │   └── geometry/             # pure, testable chart maths
│   ├── components/
│   │   ├── dashboard/            # workbook list
│   │   ├── workspace/            # grid, analyse panel, results, graph settings
│   │   └── ui/                   # shadcn/ui primitives
│   ├── data/                     # auth, Drive API, sheet templates
│   ├── lib/exportGraph.ts        # PNG / SVG export
│   └── types/workbook.ts         # zod schemas for the workbook format
├── apps/web/scratch/         # Python contract + verification scripts (run by CI)
├── docs/
│   ├── PRD.md                    # product requirements
│   ├── AUDIT_MATRIX.md           # every analysis + its runtime/golden status
│   └── *.md                      # statistical references (post hoc, curve fitting…)
├── tests/                        # Python-side oracle tests
└── Agents.md                     # project charter & engineering rules

---

Acknowledgements

Built on the shoulders of Pyodide, SciPy, statsmodels, pingouin, scikit-posthocs, lifelines, visx, and AG Grid.

Built with Claude and Antigravity.

License

MIT — free to use, modify, and distribute; just keep the copyright notice.