
Interactive regressions
Coming with v1 โ runs in your browser via WebR
Status: preview placeholder. The full WebR-powered regression module ships together with the complete 3-LLM corpus (Sonnet โ, GPT-4.1-mini โ, Gemini Flash in progress). Today you can browse all scores in the Browse view and inspect per-manifesto evidence on each manifesto page.
What this page will let you do
Pick the sample, pick the score construction, pick the regression spec โ everything runs client-side in your browser, no server, no upload.
Inputs
- Country filter (multi-select; e.g. just Spain + Germany)
- Year range slider (1945 โ 2025)
- Model selection โ include any subset of the three LLM families
- Within-manifesto aggregation โ confidence-weighted mean (default), plain mean, median, or max-evidence weighting
- Between-model aggregation โ mean, median, leave-worst-out
- Outcome dimension โ any of the four liberalism dimensions or the overall index
- Specification โ FE choice (country, year, country ร year), SE clustering
Outputs
- The regression coefficient on
populism_overallwith 95 % CI - Sample size after filtering, Rยฒ within
- A coefficient plot showing how the answer moves when you flip one choice at a time
- A link to the specification curve for your current selection โ the distribution of coefficients across ~9,000 defensible specifications, visualised ร la Simonsohn et al. (2020) and Menkveld et al. (2024)
Why it matters
Without this view, a single point estimate hides researcher degrees of freedom. With it, you see all the answers the data could plausibly give โ and judge for yourself whether the headline holds up.
Specification-curve sketch โ placeholder
Once the full corpus is in, this section will show the live curve. Below is a stylised illustration of what it will look like (not from real data):
When does the live version go up?
The Gemini full-corpus run is the last piece. Status: ~3,327 manifestos, batch-checkpointed, ~5h wall clock. As soon as it lands, this page is rewired to a WebR module reading the live parquet from the Hugging Face dataset.