Executive assessment

This cross-sectional study contains 625 retained households from Baoxing and Qingchuan. Application for wildlife-damage compensation is associated with lower tolerance on several management-related measures, but the four questionnaire outcomes do not behave as one interchangeable scale. The safety question has little adjusted association with application.

The rebuilt PLS model estimates a direct standardized application path of -0.142 (95% county-stratified township-bootstrap interval [-0.193, -0.092]). The proposed indirect association through intangible costs is -0.025 [-0.065, 0.020]. Its interval includes zero. Neither this result nor the institutional pathway establishes mediation. The interaction is small and uncertain.

This supports a carefully qualified negative association, not proof that compensation reduces tolerance or that administrative fatigue explains it. PLS paths, ordinal log odds, raw-score contrasts, and outcome-SD contrasts are different quantities and must not be compared as if on the same scale.

The analysis also identifies useful secondary associations involving prevention and agricultural dependence. These are exploratory. Every tested specification, diagnostic, failed fit, and available result table is retained in the analysis outputs.

1. Background, design, and hypotheses

The study asks whether compensation application is associated with tolerance, whether intangible costs and institutional contact may link those variables, and whether the relationships withstand reasonable measurement and analytical choices. Mengxi’s wider dissertation concerns cultural perceptions, policy history, and governance of human–wild-boar conflict in China.

The Wildlife Tolerance Model motivates distinguishing conflict exposure, tangible losses, intangible costs, perceived ecological benefits, and tolerance. Compensation is a policy experience; institutional contact is an observed behavior; costs and attitudes describe lived experiences. Their conceptual roles should not be determined by whichever path is significant.

The retained sample has 160 Baoxing interviews from August 2024 and 465 Qingchuan interviews from March–April 2025. County is therefore confounded with survey period. This is not a before/after study. There is no untreated-policy control group: CE0 distinguishes respondents who report applying from those who do not, not randomly assigned policy exposure.

The primary hypotheses concern application associations with the four tolerance items. Direction is not required for an analysis to be successful. Indirect paths, the application-by-cost interaction, group differences, and secondary outcomes are explicitly exploratory. The decisions were frozen before this rebuild, but the dataset had previously been inspected; this is not preregistration of unseen data.

2. Data preparation and EDA

The input workbook is preserved. All 625 IDs match between its Input Data and Raw data sheets. The raw sheet itself contains retained interviews and earlier derived fields, not an untouched export. The exclusion reasons for 49 of the original 674 interviews remain unavailable.

Tol2 and Tol3 are already reversed; Mengxi confirmed this, and the raw Chinese responses agree. They are not reversed again. The primary rebuild restores one missing raw Tol1 answer, missing applicant timeliness/satisfaction answers, and missing word responses. Payment-experience items remain undefined for non-applicants. Original and derived values are linked in data/derived/transformation_log.csv.

Reporting institutions are reconstructed as an increasing count. Percentage answers above 100% are unresolved, not silently divided by ten. The primary preference analyses exclude those values; sensitivities retain them. Recorded area is provisionally labeled mu, without hectare conversion. Prevention non-use is distinct from poor effectiveness among users.

metric value
respondents 625.000
applicants 277.000
counties 2.000
townships 25.000
villages 178.000
inside_park 105.000
missing_village 1.000
invalid_minimum_ratio 9.000
invalid_desired_ratio 4.000
damage_exceeds_area 7.000
prevention_nonusers 83.000
alpha_four 0.338
alpha_three 0.502

There are 177 complete village keys plus one respondent-specific key for the missing village, nested in 25 county–township keys. No representativeness claim or design weights can be justified without the sampling protocol.

Application rates differ strongly by county (119/160 versus 158/465). The county-adjusted results consequently answer a different comparison from the pooled unadjusted difference. Economic variables are strongly skewed; log1p transformations retain zeros and reduce leverage. Damage greater than owned land is flagged, not automatically deleted, because reference periods could differ.

The four items concern non-lethal management, opposition to stricter population control, absence of safety threat, and acceptance of damage. They are substantively different. Strong floor/ceiling effects and very sparse extreme categories matter for ordinal estimation and factor analysis.

The EDA exports a numeric profile, category frequencies, county/application/park profiles, correlations, raw-field inventory, crop portfolio, species/seasonality/encounter frequencies, and original Chinese-word frequencies. These full machine-readable tables accompany this report rather than being reduced to selected significant comparisons.

3. Measurement structure: EFA, CFA, PCA

Development and validation

A fixed village-disjoint split balances county/application cell counts as closely as possible. Measurement exploration uses the development partition; theory models are evaluated in the validation partition unchanged. Full-sample estimates are then reported for precision. Prior inspection of this dataset means this remains internal validation, not independent confirmation.

Var1 Var2 Var3 Freq
Baoxing 0 development 21
Qingchuan 0 development 156
Baoxing 1 development 60
Qingchuan 1 development 79
Baoxing 0 validation 20
Qingchuan 0 validation 151
Baoxing 1 validation 59
Qingchuan 1 validation 79

EFA

EFA examines the four tolerance items and IC1 plus three word-sentiment items. It uses polychoric correlations, minimum-residual extraction, and oblimin rotation. Missing word responses are not replaced with artificial sentiments. Correlations use available pairs; this relies on a missingness assumption and does not solve selective word nonresponse. There are fewer complete respondents than the pairwise analysis count.

Sparse categories caused the initial continuity-corrected polychoric estimator to fail. The executed analysis uses uncorrected pairwise polychorics (correct=0, global=FALSE), with that numerical change logged. The parallel-analysis suggestion is examined against identification and loading structure, not accepted mechanically.

factors RMSR TLI RMSEA admissible development_n complete_n parallel_suggested
1 0.045 0.810 0.085 TRUE 316 238 4
2 0.029 0.930 0.051 FALSE 316 238 4
3 0.019 1.004 0.000 TRUE 316 238 4

Parallel analysis suggests four factors among only eight items. A four-factor structure with at least three indicators per factor is not supported by this item pool, so no such CFA is manufactured. The two-factor EFA has an ultra-Heywood solution; the three-factor solution’s better fit does not by itself establish interpretable constructs. This is substantive evidence against treating a neat factor solution as guaranteed.

item MR1 communality
tol1 0.524 0.274
tol2 0.763 0.582
tol3 0.041 0.002
tol4 0.541 0.292
ic1 -0.755 0.570
word1 -0.218 0.047
word2 -0.248 0.061
word3 -0.136 0.019
item MR3 MR1 MR2 communality
tol1 0.036 -0.035 0.805 0.606
tol2 -0.205 0.150 0.508 0.504
tol3 -0.150 -0.108 -0.021 0.017
tol4 0.000 0.999 -0.002 0.997
ic1 0.997 -0.004 -0.001 0.998
word1 -0.056 -0.112 -0.221 0.060
word2 -0.001 -0.212 -0.107 0.069
word3 0.282 0.207 -0.039 0.074

CFA and discriminant validity

Ordinal CFA uses WLSMV and available-pair information. The two-factor models separate tolerance from the proposed reflective intangible-cost block; one alternative excludes the safety item. A one-factor model tests empirical overlap. These are tests of measurement hypotheses, not permission to redefine formative constructs after seeing fit.

partition model n converged admissible cfi.scaled tli.scaled rmsea.scaled srmr
validation two_factor_four 309 TRUE TRUE 0.912 0.870 0.067 0.064
validation two_factor_three 309 TRUE TRUE 0.967 0.947 0.048 0.051
validation one_factor 309 TRUE TRUE 0.914 0.879 0.064 0.066
full two_factor_four 625 TRUE TRUE 0.932 0.900 0.061 0.054
full two_factor_three 625 TRUE TRUE 0.957 0.931 0.058 0.048
full one_factor 625 TRUE TRUE 0.933 0.907 0.059 0.054

The three-item tolerance version improves validation fit, but the combined model’s fit is not a standalone validation of a three-item scale. Factor loadings and factor correlations matter. The full-sample cost/tolerance factor correlation has large absolute magnitude; weak word-item loadings and overlapping attitudinal content make a clean separation doubtful. Factor signs are arbitrary; use loading orientation when interpreting the correlation.

lhs rhs std.all
262 Tolerance tol1 0.595
263 Tolerance tol2 0.808
264 Tolerance tol4 0.454
265 Intangible ic1 -0.643
266 Intangible word1 -0.247
267 Intangible word2 -0.200
268 Intangible word3 -0.214
model lhs rhs std.all ci.lower ci.upper
243 two_factor_four Tolerance Intangible 0.908 0.683 1.133
305 two_factor_three Tolerance Intangible 0.890 0.666 1.113
comparison htmt
Tolerance4 vs IC reflective hypothesis 1.074
Tolerance3 vs IC reflective hypothesis 0.927
Institutions vs IC (NOT applicable to formative blocks) NA

The attachment’s historical HTMT 1.112 is not reused as a current estimate. The corrected reflective tolerance-versus-cost diagnostics still indicate overlap. Institutional contact, however, is specified as a formative composite of behaviors; HTMT is not an appropriate blanket pass/fail criterion for that block. Neither merging institutional contact with sentiment nor calling it “governance fatigue” is justified by these questions alone.

PCA

PCA uses standardized observed items among complete cases. It describes variance, not latent causes; it is not PLS-SEM or an item-response model. PCA loadings and variance are supplied as an alternative description, not selected to maximize the compensation coefficient.

component variance cumulative
1 0.253 0.253
2 0.149 0.402
3 0.126 0.528
4 0.124 0.652
5 0.105 0.758
6 0.103 0.861
7 0.079 0.940
8 0.060 1.000

Measurement conclusion: retain the four observed outcomes as the primary family. Use the three-item and four-item scores transparently as secondary summaries. Treat the proposed reflective cost factor as uncertain, and keep institutional behaviors interpretable at item level. No single latent score should carry the whole scientific claim.

4. Primary models and uncertainty

The primary ordinal models report application associations for all four items, with Holm correction within each adjustment family. Background controls are county, park residence, age, gender, education categories, and household size. Resource adjustment adds log income, agricultural share, log land, livestock, and assets. Explanatory models additionally condition on simultaneous conflict, cost, benefit, reporting, and prevention variables; these should not be described as total causal effects.

Uncertainty clusters by township with a finite-cluster t reference. Village clustering is a sensitivity; only two counties are available, so county-clustered inference is not used.

outcome adjustment n estimate low high p_holm
tol1 unadjusted 624 -0.982 -1.409 -0.555 0.000
tol1 background 624 -0.634 -1.009 -0.259 0.008
tol1 resources 624 -0.668 -1.052 -0.283 0.006
tol1 explanatory 622 -0.469 -0.899 -0.038 0.083
tol2 unadjusted 625 -0.635 -1.039 -0.230 0.010
tol2 background 625 -0.484 -0.909 -0.060 0.081
tol2 resources 625 -0.561 -0.969 -0.152 0.027
tol2 explanatory 623 -0.517 -0.972 -0.062 0.083
tol3 unadjusted 625 0.396 0.114 0.677 0.016
tol3 background 625 0.066 -0.259 0.391 0.678
tol3 resources 625 0.100 -0.238 0.438 0.548
tol3 explanatory 623 0.225 -0.148 0.598 0.224
tol4 unadjusted 625 -0.002 -0.367 0.363 0.991
tol4 background 625 -0.304 -0.617 0.008 0.112
tol4 resources 625 -0.363 -0.644 -0.081 0.027
tol4 explanatory 623 -0.407 -0.721 -0.094 0.052

The strongest background-adjusted signal is less support for non-lethal management among applicants. The safety item has a different pattern. Adding resources strengthens some other item associations; adjusting simultaneous attitudes answers another question and can attenuate them.

Predicted category probabilities are in ordinal_probabilities.csv. These are more interpretable than odds ratios alone. The available partial proportional-odds diagnostics and models are reported as sensitivity analyses, not silently substituted to obtain favorable results.

outcome term LRT Pr(>Chi)
tol1 application 3.970 0.265
tol2 application 5.234 0.155
tol3 application 8.365 0.039
tol4 application 9.265 0.026

These likelihood-ratio diagnostics are conventional individual-level checks, not cluster-adjusted tests. Some items suggest nonparallel application slopes. A sparse-category partial model for Tol2 has a singular Hessian; its parameter table is retained for diagnosis but should not be interpreted as a reliable final estimate. The ordinal proportional-odds summaries therefore remain accompanied by category probabilities, partial-model diagnostics, and continuous-score sensitivity.

outcome adjustment estimate low high p
tolerance3 unadjusted -0.248 -0.406 -0.090 0.003
tolerance3 background -0.284 -0.452 -0.117 0.002
tolerance3 resources -0.325 -0.483 -0.168 0.000
tolerance3 explanatory -0.252 -0.377 -0.127 0.000
tolerance4 unadjusted -0.113 -0.285 0.059 0.188
tolerance4 background -0.232 -0.404 -0.061 0.010
tolerance4 resources -0.264 -0.434 -0.093 0.004
tolerance4 explanatory -0.169 -0.321 -0.017 0.031

GAMs assess nonlinear age/income/land relationships in the resource-adjusted score models. Their smooth-term and parametric tables are saved separately; their usual model-based p-values are not substitutes for the primary clustered inference.

Bayesian ordinal multilevel complement

Four cumulative-logit models use township random intercepts, background controls, standardized continuous predictors, Normal(0,1) coefficient priors, and exponential(1) random-intercept SD priors. Each uses four chains with 1,000 warmup and 1,000 retained iterations per chain. They make a random-effects assumption and are not a cure for unmeasured confounding or proportional-odds misspecification.

outcome n median low high prob_positive diagnostics_pass
tol1 624 -0.626 -0.978 -0.255 0.000 TRUE
tol2 624 -0.462 -0.818 -0.107 0.005 TRUE
tol3 624 0.144 -0.177 0.474 0.808 TRUE
tol4 624 -0.284 -0.590 0.037 0.044 TRUE
outcome max_rhat min_bulk_ess min_tail_ess divergences diagnostics_pass
tol1 1.005 1249.544 2062.815 0 TRUE
tol2 1.003 849.718 1655.986 0 TRUE
tol3 1.004 932.722 1765.343 0 TRUE
tol4 1.004 785.608 1023.325 0 TRUE

Posterior predictive figures are available for all four outcomes. Paired posterior category contrasts use the same draws under application 0/1. Passing MCMC diagnostics establishes computational adequacy, not causal validity or perfect model fit.

5. PLS-SEM, mechanisms, and moderation

The main PLS model contains a three-item reflective tolerance specification; formative intangible-cost, institutional-contact, and tangible-cost blocks; and observed application and ecological value. Standard PLS-PM is used without disattenuation. Thus path estimates describe standardized composite-score relationships; they are not error-free latent causal parameters. Equal-scoring/item-level analyses remain essential checks.

Five specifications are compared: baseline, mediation, extended conflict/prevention, four-item tolerance, and IC1-only. All use the same complete sample for this comparison. Formative weights, reflective loadings, VIF, AVE, reliability, f², R², and SRMR are exported. Interpret reflective reliability only for the reflective block; do not use it as a requirement for formative components.

model n admissible R2_Tol SRMR
baseline 622 TRUE 0.212 0.075
mediation 622 TRUE 0.213 0.091
extended 622 TRUE 0.234 0.074
four_items 622 TRUE 0.218 0.091
ic1_only 622 TRUE 0.187 0.092

Measurement and structural coefficients are re-estimated in every bootstrap. Whole townships are sampled within county, preserving the two-county design. Of 5,000 requested main-model replicates, 4999 are usable. Failures are retained in pls_bootstrap_replicates.csv. Both 2,000- and 5,000-replicate summaries are available for stability assessment.

parameter estimate low high p_centered successful
direct -0.142 -0.193 -0.092 0.000 4999
indirect_IC -0.025 -0.065 0.020 0.248 4999
indirect_PS 0.024 -0.002 0.049 0.062 4999
total -0.143 -0.194 -0.083 0.000 4999
IC_to_Tol -0.365 -0.440 -0.260 0.000 4999
PS_to_Tol 0.072 -0.006 0.148 0.070 4999
IB_to_Tol 0.186 0.113 0.262 0.000 4999
TC_to_Tol 0.060 -0.069 0.097 0.223 4999
interaction 0.100 -0.074 0.264 0.254 4999
interaction_f2 0.003 0.000 0.020 0.319 4999

The IC and institutional indirect paths both have intervals crossing zero. Their opposite signs nearly cancel in this specification. A significant direct application association does not establish why it occurs. Neither the measurement data nor temporal ordering establishes administrative fatigue.

The two-stage interaction re-estimates PLS scores in each bootstrap, then fits the application-by-centered-cost term. Its coefficient is 0.100 [-0.074, 0.264]; f² is approximately 0.003. It is an exploratory small, uncertain effect. The nonnegative f² percentile interval is not a significance test. The observed-score interaction provides a separate common-scale comparison.

term estimate se low high p n clusters f2
application:ic 0.053 0.047 -0.044 0.149 0.272 624 25 0.002

6. Multi-group analysis

Application defines the two groups; it is not included as a predictor within either group. Measurement blocks and algorithms are the same. MICOM’s conventional individual-permutation diagnostic uses 999 permutations; it is not a cluster-randomization test, so its p-values are supporting diagnostics. Full output is in logs/micom.txt.

The compositional-invariance tests do not reject, but some estimated score correlations are low and equality of means/variances does not generally hold. This is not proof of full invariance. Different group standard deviations can alter standardized path coefficients.

Group path differences below use 2,000 township-cluster bootstrap samples. Every interval includes zero. Interpret this as no strong evidence of these group differences, not proof of equality or support for moderation.

path difference low high successful
IC 0.144 -0.060 0.284 2000
PS -0.031 -0.291 0.199 2000
TC -0.129 -0.227 0.112 2000
IB -0.034 -0.171 0.106 2000

7. Prediction and cross-fitted adjustment

Prediction

Five outer township-disjoint folds evaluate prediction; three inner township folds tune elastic net, forest, and boosting. All imputation/scaling is learned from training rows. The score target is the raw three-item mean, not a globally standardized latent target. Features include directly observed attitudes, not precomputed full-data standardized IC scores. A mean-only benchmark establishes the baseline.

method n RMSE MAE R2
boosting 624 0.538 0.409 0.192
elastic_net 624 0.537 0.414 0.194
GAM 624 0.541 0.418 0.182
mean 624 0.598 0.452 -0.001
random_forest 624 0.538 0.406 0.190

test_county method n RMSE R2
Baoxing mean 160 0.679 0.000
Baoxing elastic_net 160 0.632 0.135
Baoxing random_forest 160 0.638 0.117
Qingchuan mean 464 0.567 0.000
Qingchuan elastic_net 464 0.524 0.147
Qingchuan random_forest 464 0.539 0.097

Out-of-fold performance is modest. Small differences among these models do not prove that relationships are linear. Held-out forest permutation importance is exported as change in MSE; it measures predictive reliance, not causal importance. County transfer checks use fixed hyperparameters and just two sites; they cannot establish broad transportability.

AIPW-style adjustment

Repeated five-fold township cross-fitting estimates separate applicant/non-applicant outcome models and application probabilities using fixed-hyperparameter forests. Each respondent’s nuisance predictions come from other townships. Adjustment uses background and resource variables, not simultaneous attitudes. Propensities are bounded at .01/.99; clipping counts, balance, and effective sample size are exported.

repetition estimate_raw low high min_propensity max_propensity clipped effective_sample_size
1 -0.198 -0.300 -0.097 0.122 0.871 0 547.407
2 -0.199 -0.296 -0.101 0.107 0.880 0 532.454
3 -0.193 -0.298 -0.088 0.124 0.868 0 536.464
4 -0.195 -0.297 -0.092 0.102 0.888 0 538.327
5 -0.216 -0.322 -0.111 0.089 0.868 0 535.990

Each repetition is a sensitivity to splitting, not an independent replication. Intervals cluster influence scores by township. These are raw 1–5 score contrasts. Their interpretation as treatment effects would require unverified consistency, exchangeability, temporal ordering, and overlap assumptions. No causal claim is made.

8. Broader exploratory contribution

analysis n estimate low high p_BH_exploratory
application_selection 625 0.087 -0.031 0.205 0.301
applicant_timeliness 275 0.013 -0.097 0.123 0.861
applicant_satisfaction 276 -0.018 -0.103 0.066 0.746
applicant_both_experiences 274 -0.043 -0.178 0.091 0.746
livelihood_dependence 624 0.250 0.103 0.398 0.031
damage_burden 624 -0.171 -0.349 0.008 0.169
income_loss_burden 605 -0.004 -0.022 0.013 0.746
prevention_adoption 625 1.134 0.358 1.910 0.035
prevention_effectiveness_users 541 0.132 0.041 0.222 0.035
insurance_positive_wtp 625 0.524 0.067 0.981 0.090
insurance_positive_amount 492 0.082 -0.135 0.298 0.746
minimum_ratio_validated 616 0.000 -0.032 0.033 0.986
minimum_ratio_retain_flagged 625 -0.107 -0.246 0.033 0.301
desired_ratio_validated 621 0.013 -0.039 0.064 0.746
desired_ratio_retain_flagged 625 0.015 -0.040 0.071 0.746
reporting_contacts 622 0.130 0.025 0.236 0.074
information_channels 622 -0.047 -0.138 0.044 0.562

The broad secondary family receives Benjamini–Hochberg adjustment. These estimates involve different outcomes/scales and must not be ranked by coefficient size. Notable avenues include agricultural dependence, prevention adoption, and effectiveness among users. Applicant satisfaction/timeliness associations are imprecise and do not establish the proposed mechanisms.

metric value
applicants 277
numeric_unambiguous_payouts_applicants 161
median_available_payout 60
timeliness_observed 275
satisfaction_observed 276
zero_income 19
loss_income_undefined 19
method users nonusers mean_effectiveness_users recorded_cost_n median_recorded_cost
A 248 377 2.153 55 500
B 391 234 2.026 115 200
C 339 286 2.909 3 0
D 56 569 2.893 0 NA
E 147 478 2.340 14 0
F 10 615 1.100 0 NA
G 6 619 1.333 0 NA

Prevention letters follow the raw coded fields, whose Chinese descriptions are retained in the full table. Monetary cost nonresponse is common. We do not calculate causal benefit–cost rankings from these incomplete, self-selected responses. Payout amounts are accepted only when the text is an unambiguous numeric response in exactly one amount field; payout/loss ratios are withheld until reference periods are confirmed.

The raw questionnaire supports crop-specific losses, seasonal descriptions, encounters, species, and original-word frequencies. Their complete tables are included. Seasonality is recalled timing, not a repeated-measures panel. Word-frequency counts do not establish validated psychological themes; human Chinese-language interpretation remains necessary.

9. Robustness, vibration of effects, and selective-reporting risk

The specification curve crosses prespecified outcomes, adjustment families, raw/log economic predictors, extreme-loss/reporting-contradiction handling, and complete-case versus median/mode covariate sensitivity. Identical model inputs are deduplicated before estimation. Township/village uncertainty variants are retained explicitly. Outcomes are displayed separately because they answer different questions.

Single median/mode imputation is only a limited sensitivity for the few incidental covariate gaps; it is not multiple imputation and does not account for imputation uncertainty. Structural payment nonresponse and missing words are not imputed.

outcome adjustment specifications min median max negative_fraction
tol1 background 6 -0.293 -0.286 -0.285 1
tol2 background 6 -0.258 -0.255 -0.252 1
tol3 background 6 0.036 0.039 0.049 0
tol4 background 6 -0.149 -0.140 -0.139 1
tolerance3 background 6 -0.296 -0.285 -0.284 1
tolerance4 background 6 -0.235 -0.232 -0.232 1
tol1 conflict 12 -0.251 -0.239 -0.230 1
tol2 conflict 12 -0.256 -0.245 -0.232 1
tol3 conflict 12 0.077 0.088 0.097 0
tol4 conflict 12 -0.153 -0.144 -0.134 1
tolerance3 conflict 12 -0.279 -0.267 -0.253 1
tolerance4 conflict 12 -0.197 -0.189 -0.184 1
tol1 explanatory 24 -0.178 -0.171 -0.162 1
tol2 explanatory 24 -0.231 -0.225 -0.214 1
tol3 explanatory 24 0.090 0.108 0.122 0
tol4 explanatory 24 -0.191 -0.182 -0.178 1
tolerance3 explanatory 24 -0.260 -0.254 -0.251 1
tolerance4 explanatory 24 -0.177 -0.171 -0.166 1
tol1 none 6 -0.426 -0.424 -0.420 1
tol2 none 6 -0.318 -0.308 -0.305 1
tol3 none 6 0.209 0.215 0.226 0
tol4 none 6 0.024 0.027 0.032 0
tolerance3 none 6 -0.260 -0.255 -0.248 1
tolerance4 none 6 -0.127 -0.113 -0.113 1
tol1 resources 12 -0.307 -0.301 -0.293 1
tol2 resources 12 -0.293 -0.280 -0.264 1
tol3 resources 12 0.036 0.047 0.055 0
tol4 resources 12 -0.186 -0.178 -0.168 1
tolerance3 resources 12 -0.333 -0.322 -0.309 1
tolerance4 resources 12 -0.266 -0.261 -0.253 1

These curves describe analytical sensitivity. They cannot establish that p-hacking never occurred, and the fraction negative/significant is not a posterior probability or an independent-replication count. Historical incorrect reverse coding is excluded from the set of reasonable primary specifications.

check minimum maximum
Leave-one-township-out application contrast -0.353 -0.289
county estimate low high n clusters
Baoxing -0.381 -0.786 0.024 160 5
Qingchuan -0.302 -0.494 -0.111 464 20

The omitted-confounding grid benchmarks how a hypothetical omitted variable could move the resource-adjusted linear estimate toward zero. It uses partial-R² values for outcome and application residuals. This is a point-estimate bias calculation, not a cluster-adjusted confidence bound, and no particular strength is asserted to be realistic.

partial_R2_y partial_R2_d bias estimate toward_zero
0.00 0.00 0.000 -0.195 -0.195
0.01 0.01 0.013 -0.195 -0.182
0.03 0.03 0.039 -0.195 -0.155
0.05 0.05 0.066 -0.195 -0.129
0.10 0.10 0.136 -0.195 -0.059
0.20 0.20 0.288 -0.195 0.093

10. Answers to the Exploratory Contribution

Requested question Defensible conclusion
Do the indicators form the expected constructs? Not cleanly. Sparse categories, weak word loadings, factor overlap, and weak safety-item coherence challenge a single tidy measurement model.
Remove, reassign, or merge indicators? Present observed outcomes as primary. Three-item tolerance is a transparent sensitivity. No automated significance-driven deletion or institution–sentiment merger.
Is policy support distinct from intangible costs? The institutional indicators measure contact/information behavior, not reflective fatigue. Evaluate their interpretation and formative construction before applying HTMT.
Are the proposed structural models defensible? They are useful exploratory association models. Compare baseline and extended versions; contemporaneous data do not identify path direction.
Does the indirect cost pathway hold? Its cluster-bootstrap interval includes zero in the rebuilt main model. Do not claim demonstrated mediation.
Does the interaction matter? Its f² is small and its interval crosses zero. Retain in exploratory/supplementary results, not as the central mechanism.
What does nonsignificant MGA mean? No strong evidence of differences on the tested paths, conditional on measurement comparability. Not equivalence.
Is the paradox robust? Several negative management-related associations persist. Conclusions depend on the tolerance dimension and adjustment; the safety item differs. “Negative association among applicants” is more precise than “policy failure.”
Has p-hacking been ruled out? No statistical test can certify that here. The decision register, complete outputs, honest validation, and deduplicated multiverse reduce selective-reporting risk.

11. Limitations and implementation boundaries

  • Cross-sectional, nonrandom application; no verified treatment timing, payment receipt, or pre-policy baseline.
  • Two counties observed in different periods; unknown sampling/exclusion processes limit generalizability.
  • Limited reflective measurement, overlapping attitudes, word nonresponse, and incomplete raw cleaning provenance.
  • Ordinal proportional-odds assumptions are imperfect for some items; sparse alternatives can be unstable.
  • Conventional CFA/MICOM diagnostics do not fully model the sampling design; clustered structural/association checks are complementary.
  • Exploratory analyses reuse this finite sample. Internal splits and cross-fitting do not turn the dataset into independent confirmatory evidence.
  • Some planned extensions require information not in the workbook: validated thematic coding, comparable payout/loss periods, sampling weights, and causal policy identification. They are explicitly withheld.

No outcome was selected because it produced the desired sign. Numerical warnings and earlier failed attempts remain in the event log. The archived legacy results are not overwritten. This release is suitable for methodological discussion; the report can be amended transparently after the presentation and Mengxi’s clarifications.

12. Reproduction and references

From the project root:

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

renv.lock records packages; run_manifest.json records source/data hashes and session information. The targets graph rebuilds changed stages. Source data, code, derived tables, and figures remain distinct. The PowerPoint and speaker notes read these same saved results.

Methodological references: