The Wrong Unit of Uncertainty

Replication package for "The Wrong Unit of Uncertainty: Simultaneous Inference for Repeated Cross-National Surveys" (Kunwoo Park, Kookmin University; in preparation for submission to Political Analysis).

What this is

Claims about repeated cross-national surveys attach uncertainty to the wrong unit twice: a wave-pair mean contrast stands in for a persistent, distribution-wide trajectory, and an estimated distribution stands in for the latent one it samples. This package provides:

• within a country, a design-based simultaneous band over the whole response-distribution trajectory — a studentized sup-t band on wave-pair CDF differences from the stratified-PSU bootstrap, so an asymptotic guarantee — carrying a partially ordered claim family (pairwise, any-pair, net, persistent — persistent at the top);
• across countries, a finite-sample clustered conformal band, with the country as the exchangeable unit, plus a Bonferroni layer and a closed-testing lower bound on the across-country count;
• a non-identification theorem and survey-scale unreachability boundary for the design-aware (deconvolution) correction, with a provably selection-free deployed selector;
• two named reanalyses — ESS parliamentary trust (2002–2024) and the WVS Foa–Mounk deconsolidation battery (WVS Trend File, 1981–2022) — plus V-Dem cross-tabs and a same-items comparison with the Claassen latent panel.

Headline results (all regenerate from results/*.csv)

finding	where
Marginal readings flag 20/30 ESS countries; the hierarchy certifies net decline in 6, persistence in 1 (Greece)	e13, §7
Over the full 2002–2024 record, read off one joint band: persistence in 0/33, span erosion in 8, and 23/33 certifying both a decline and a recovery at one α	e50, §7
Closed testing across countries: with 90% simultaneous confidence at least 6 of 33 truly declined over their span, on each outcome — the across-country count itself now carries a guarantee	e56, §7
WVS: persistence cuts the wave-pair set 2.6–6.5×; 12 of the 13 descriptive multi-item core countries pass a valid ≥2-item partial-conjunction test (10/9 at design effects 1.5/2.0)	e26/e63/e65, §7
Deconvolution is non-identified without the design-noise law and unreachable at survey scale (K≥94 floor)	Thm 1, Prop 2, §2/§6
Robustness: RWY-rescaled bootstrap, WVS and joint-band design-effect sweeps, mode audit from the data's own mode variable, LORO exchangeability, null-imposed severity injection, window-matched Claassen — plus two withdrawn results with published diagnoses	e38–e53, Supplement

Quickstart

pip install -r requirements.txt && pip install -e .
python -m pytest tests/ -q          # contracts + claim ledger
python -m pcb.experiments.e28_wrong_unit_coverage   # Table 1, ~seconds

Using the method on your own data — Python:

from pcb import dapcb
fit = dapcb(cal_errors, v_cal, center, alpha=0.10)
fit.band, fit.selected_branch, fit.coverage_level, fit.target, fit.fallback_reason

or R (rpkg/dapcb, a pure-R port validated against the Python reference by golden tests to 1e-10; R CMD check passes with no ERROR — building the vignette additionally needs rmarkdown):

install.packages("rpkg/dapcb", repos = NULL, type = "source")
library(dapcb)
E <- as.matrix(read.csv("calibration_errors.csv", check.names = FALSE))
V <- as.matrix(read.csv("design_sds.csv", check.names = FALSE))
mu <- scan("target_center.csv", quiet = TRUE)
fit <- dapcb(E, V, mu, alpha = 0.10)
fit$selected_branch
fit$coverage_level
fit$target
fit$fallback_reason
cbind(lower = fit$lo, center = mu, upper = fit$hi)

E and V are matching K × T matrices: one row per exchangeable population and one column per stacked trajectory coordinate. mu is the length-T target center. The returned level always states which target it covers; if a gate blocks deconvolution, fallback_reason says why.

flowchart LR
  estimates[Population trajectory estimates] --> errors[Calibration errors E]
  design[Design uncertainty V] --> selector[Finite-K safe selector]
  errors --> selector
  selector --> band[Simultaneous band]
  band --> claims[Trajectory-level claims]

Loading

Reproduce

Start at REPLICATION.md — the replication analyst's entry point: environment, run order, runtimes, and the claim→artifact→test map (docs/REPLICATION_MAP.md). make deposit builds the curated, deterministic submission archive. Two tiers (details in docs/REPRODUCIBILITY.md):

• Tier 1 (no microdata): all simulation/theory/benchmark experiments and the V-Dem/Claassen public-data analyses run out of the box.
• Tier 2 (licensed microdata): the ESS/WVS/LAPOP reanalyses. Exact files, registration links, placement paths, and sha256 checksums are in docs/DATA_SOURCES.md; the ESS inputs (integrated files and all SDDF files) can also be fetched by script through the ESS Data Portal API with a registered user ID (scripts/fetch_ess_api.py, scripts/fetch_ess_sddf.py, scripts/build_ess_subset.py). The ESS certification, the WVS hierarchy, and the joint claim family reproduce the committed CSVs bit-identically (verified in independent environments, including from the API-built subset).

All runs use fixed seeds (pcb.util.det_seed) and are deterministic. Shell entry points: setup.sh, run_public.sh (Tier 1), run_restricted.sh (Tier 2, checks the licensed inputs first), run_all.sh. Fine-grained targets: make test, make tier1, make tier2, make figures, make paper.

Layout

paper/            LaTeX source + compiled PDFs (main, supplement, title page)
rpkg/dapcb/        R package: pure-R dapcb port, vignette, golden cross-language tests
pcb/
  inference/      clustered/population conformal, design_aware, safe selector
  data/           survey loaders: ESS, WVS trend file, LAPOP (schema audits)
  simulation/ theory/     generators and theory checks
  experiments/    e6–e65 (simulation arc, ESS/LAPOP/WVS, robustness, frontier,
                  prevalence; e43–e49, e51 superseded/withdrawn — see supplement)
  figures/        figure generators (write to figures/; tracked copies in paper/figures/)
tests/            contract tests (theorem<->code) plus a claim ledger pinning
                  every headline number in the paper to the CSV that licenses it
results/          precomputed result tables (CSV) — every paper number lives here
docs/             preregistrations, results write-ups, proofs, data sources, HANDOFF
configs/          frozen validation manifests (seeds, script hashes)

Data notice

/data/ is gitignored: the ESS, WVS, and LAPOP microdata are licensed by their providers and never committed. Download each from its provider and place per docs/DATA_SOURCES.md. Everything else — code, results, paper — is MIT-licensed (see LICENSE); the licensed survey data are not covered by that license.

Citation

See CITATION.cff. Until the paper is published, cite the package by its title and this repository.