← Papers

Data Lab

Free-form notebook — experiment with data & AI

Data Lab

A Jupyter-style environment for data exploration, statistical analysis, and research design validation — powered by Claude AI.

Notebooks

Interactive code cells with AI assistance

Write

Python code cells with pandas, numpy, statsmodels, scikit-learn, scipy. Markdown cells for notes.

Run

Execute cells with one click. Output appears inline — regression tables, summary stats, print output.

Interpret

Add AI cells below any code cell. Claude reads your code and output, gives plain-language interpretation.

7 Templates

BlankExploratory Data AnalysisOLS RegressionPanel DiDTime SeriesInstrumental VariablesML Baseline

Workflow

Create notebookWrite & run codeAdd AI cellsLink paperPush to Draft

Simulation Experiments

Monte Carlo validation of research designs

Before collecting real data, validate your research design with Monte Carlo simulations. Claude generates and executes Python code (500-5000 iterations), then interprets the results. Available when creating a new notebook via the template picker.

RCT

Randomized Controlled Trial

Simulate a two-group experiment. Check statistical power, minimum sample size, Type I error rate, and effect estimate accuracy.

Am I powered enough to detect my expected treatment effect?

DiD

Difference-in-Differences

Panel with treatment/control groups, pre/post periods. Test parallel trends assumption and quantify bias if it fails.

Will my DiD estimate be biased if pre-trends aren't perfectly parallel?

Bias

Sample Bias Analysis

Model selection bias, survivorship bias, or differential attrition. Compare naive OLS against IPW-corrected estimates.

How bad is my attrition problem? Does IPW fix it?

Power

Power Analysis

Grid search over effect sizes and sample sizes. Find the smallest N that achieves 80% power for your minimum detectable effect.

What sample size do I need for my grant proposal?

SC

Synthetic Control

Generate multi-country panel data with one treated unit. Validate whether synthetic control recovers the true effect.

Can synthetic control work with my donor pool size?

Virtual World Builder

Generate realistic experimental datasets

Build a simulated population, design an intervention, configure a timeline — then generate a full panel dataset with realistic demographics, treatment effects, non-compliance, spillovers, and non-random attrition. Access it from the Simulation World tab above.

Step 1: Population

Set country, size, age/income distributions, urban fraction, baseline health scores.

Step 2: Intervention

Choose assignment (random/geographic/eligibility), effect size, compliance rate, spillover.

Step 3: Timeline

Baseline year, follow-up waves, intervention start, attrition rate, attrition bias.

Output

A long-format panel CSV with columns: id, year, treatment, age, female, urban, income, health_score, attrited, wave. Download as CSV, push to a notebook, or view a preview table directly in the UI.

AI Chat Sidebar

Context-aware research assistant

When working in a notebook, switch to the AI Chat tab in the right panel. The AI reads all cell outputs as context, so you can ask questions like:

Explain the regression outputIs my model correctly specified?Suggest a robustness checkHow do I interpret the p-value?What controls should I add?Is there multicollinearity?
or