Compensation application and tolerance toward wild boars

Executive result

Application is negatively associated with several management-related tolerance dimensions. The four questions do not form one interchangeable scale; the safety item behaves differently. The defensible conclusion is a qualified association, not proof that compensation causes lower tolerance.

Full HTML report

1. Data foundation and EDA

The authoritative workbook has a 625-row Input Data sheet and a 626-row Raw data sheet containing a header/response layer. All IDs match one-to-one. The duplicate Python workbook is byte-identical. The retained data contain 160 Baoxing and 465 Qingchuan respondents; reasons for excluding 49 initial interviews are not available.

Preparation choices

• Tol2 and Tol3 remain as stored; Mengxi confirmed they are already reverse-coded and raw Chinese responses verify the mapping.
• Missing descriptive-word zeros become missing. No nonexistent word is imputed.
• CE1/CE2 stay structurally missing for non-applicants. Percentage values above 100% remain unresolved and are excluded from primary preference models.
• Reporting-institution count is reconstructed from raw choices. Area units remain provisionally mu. Prevention non-use is separate from effectiveness among users.

Complete EDA tables cover distributions, missingness, category frequencies, correlations, county/application/park profiles, crop portfolios, raw question families, and descriptive words.

2. Measurement: EFA, CFA, and PCA

The measurement programme tests whether the proposed constructs are meaningful before using them in structural paths. EFA uses development-partition polychoric correlations, minimum-residual extraction, parallel analysis, and oblimin rotation. CFA uses ordinal WLSMV for theory models evaluated on a village-disjoint validation partition and the full sample.

Parallel analysis suggests four factors among eight items, but the item pool cannot support a defensible four-factor CFA with three indicators per factor. The two-factor EFA has an ultra-Heywood warning. CFA shows overlap between tolerance and the proposed cost factor, while formative institutional behaviors should not be forced through reflective HTMT thresholds. The historical HTMT 1.112 is therefore a warning to inspect indicator meaning, not a reason to merge constructs automatically.

3. Statistical analysis choices and outputs

Ordinal and score models

Each tolerance item is analyzed separately with cumulative-logit models. Adjustment families proceed from unadjusted to background, resources, and explanatory contemporaneous variables. Township-clustered uncertainty is primary; village clustering is sensitivity. Holm correction is applied within adjustment families.

Bayesian complement

Four cumulative-logit township-random-intercept models use weakly informative priors and posterior predictive checks. All four passed the prespecified convergence checks.

PLS-SEM and mechanisms

PLS-SEM compares baseline, mediation, extended, four-item, and IC1-only specifications. Formative blocks are evaluated with weights/VIF; reflective blocks with loadings and reliability; structural output includes R², f², SRMR, direct paths, indirect products, and the application-by-cost interaction.

4. Secondary analysis, prediction, and stress tests

The secondary programme uses applicant experience, agricultural dependence, prevention adoption/effectiveness, insurance preferences, reporting contacts, information channels, crop-specific damage, seasonality, and original descriptive words. Missingness and structural skips are retained explicitly.

Prediction and adjustment

Nested township-grouped cross-validation compares a mean benchmark, elastic net, GAM, random forest, and boosting. Fold-local preprocessing avoids leakage. Repeated township cross-fitting estimates an AIPW-style adjusted contrast; overlap and effective sample size are reported.

5. Specification curve and p-hacking safeguards

The specification curve varies prespecified outcomes, measurement summaries, adjustment families, economic transformations, anomaly handling, missingness sensitivity, and clustering. Exact duplicate inputs are deduplicated. Leave-one-township-out and county checks show influence and transport sensitivity.

Showing 24 rows; full CSV.

A dated decision register, complete result tables, explicit failures, multiplicity handling, and transparent outcome definitions make selective reporting auditable. Nonsignificant MGA differences are reported as no strong evidence of difference, not equivalence.

6. Reproduction

Rscript run_project.R --restore
Rscript run_project.R
Rscript run_project.R --render
Rscript run_project.R --verify

The generated manifest records source hashes, data hash, seed, session information, and slide counts. Validation passed for coding audit, respondent/application counts, bootstrap success, Bayesian diagnostics, and prediction outputs.

Data hash: 9fca74ec64e7cf5858b3c63b39ace1f9b891b7572114e531be9c7dcb30a92634. The source audit reports 625 input rows, 319 disagreements between CE0 and the obsolete sightings proxy, and 13 unresolved percentage anomalies.

All machine-readable outputs are under tables; model objects and logs are under models/logs. The complete narrative report contains the explicit answers to every Exploratory Contribution question.

7. Limitations and next clarification

This is a cross-sectional, nonrandom application comparison from two county-period combinations. Unknown sampling/exclusion mechanisms, incomplete payment timing, unresolved percentage and area-unit meanings, weak/overlapping reflective measurement, structural missingness, and potential unmeasured confounding limit causal claims. The most useful next step is to resolve these fieldwork questions with Mengxi and amend affected analyses transparently.