AN-107: Universe-scale coverage_class × sponsor interaction

Hypothesis: H4: Channel A contribution is larger where methodology flexibility is greater
Confidence: pending
Type: causal

Design

Sample: all-Brazil cand-poll panel (build/assemble/cand_poll.parquet) restricted to matched_share==1.0 and error.notna(), joined to full-universe coverage_class from pipelines/politica/build/llm/poll_coverage/
Specification: error ~ sponsored_by + opponent_sponsored + sponsored_by × coverage_class + log_sample_size + days_to_election polynomial; entity_effects=politico_id; other_effects=institute; cluster=muni_id. Plus per-coverage-class within-candidate β and a strict race × week FE variant.
Comparator: clean-comparator (sponsored vs poll_is_independent only) for the per-class and race-week variants
Cluster: muni_id
Notes: First universe-scale test of the Channel A coverage lever, replacing the n=200 pilot (AN-019/020). The coverage batch is the only methodology batch that landed at universe scale (sampling + operations batches cancelled per pipelines/politica/build/llm/batch/batch_state.json). 14,876 polls cached; ~57% are deterministic short-circuits (deferred / very_short / empty).

Script: source/analysis/an-107-coverage-x-sponsor-universe.py
Target: build/table/an-107-coverage-x-sponsor-universe.csv
Status: pending · 2026-06-18
Created: 2026-06-18

Question

The +7 pp sponsor effect (AN-001 / AN-002 / AN-093) is documented as an average across all sponsored polls. The Channel A prediction is that the slant is concentrated where the methodology gives sponsors room to slant without lying — specifically in coverage classes that exclude rural or peripheral sub-frames (urban_only, specific_neighborhoods) or that decline to commit at registration (deferred_complement, not_specified). AN-019/020 ran this on the n=200 methodology pilot; AN-107 runs it at universe scale now that the coverage extraction is fully cached (14,876 polls).

Two readings:

Channel A wins: β concentrates in opaque + slant-permissive classes; β in full_municipality shrinks toward 0. The slant lives in the lawful design lever.
Channel A null: β is constant across coverage classes. Coverage isn't the design lever doing the work — push the channel to the other levers (sampling / operations, both pilot-only) or to the Channel B residual.

Design

Sample: build/assemble/cand_poll.parquet, restricted to matched_share==1.0 and finite error, then merged with protocol → coverage_class from the LLM cache. Coverage class is collapsed to four interpretive buckets:

safe: full_municipality + urban_plus_selected_rural
slant_permissive: urban_only + specific_neighborhoods
opaque: deferred_complement + not_specified
other: residual (rare)

Three specifications:

Spec A (interaction, full panel): error ~ sponsored_by × coverage_bucket + opponent_sponsored + log_sample_size + days_to_election + days_to_election² with candidate FE + institute FE, clustered SE at muni_id.
Spec B (per-bucket β): same as Spec 2 above, run separately within each coverage_bucket, gives β by bucket with comparable SEs.
Spec C (strict timing): race × week FE on the clean-comparator sample, with sponsored_by × coverage_bucket.

Results

Pending — emits build/table/an-107-coverage-x-sponsor-universe.csv and build/figure/an-107-coverage-x-sponsor-universe.pdf.

Interpretation

Pending.

Follow-ups

Pending — see Step 4 of /next for the puzzles / extensions / blind-spots taxonomy.