Texas RRC Inspection Expenses Analysis¶
Research question: Does organizational capacity (budget, staffing) predict better regulatory outputs (inspections, compliance, enforcement), and how is that relationship moderated by goal ambiguity, district-level heterogeneity, and spatial/geographic factors?
+Hypotheses¶
-
+
- H1 — Capacity → Outputs: Higher OGI budget and FTE predict more inspections, higher compliance rates, and faster violation resolution. +
- H2 — Goal Ambiguity: When a larger share of RRC budget goes to the more ambiguous "Energy Resource Development" goal, the capacity → output relationship weakens. +
- H3 — Multilevel / District Effects: The capacity → output relationship varies across RRC districts (budget slope heterogeneity). +
- H4 — Spatial & Geographic: Offshore-jurisdiction and border districts moderate the capacity → output relationship; spatial autocorrelation in residuals is tested via Moran's I. +
Data:
+-
+
- PostgreSQL warehouse (
texas_data):inspections,violations,well_shape_tract
+ RRC Budget Data.xlsx: statewide RRC budget by strategy, 2016–2024
+- Analysis panel: 2016–2025 (N = 130 district-years); regression sample: 2016–2023 (N = 104) +
import os
+import warnings
+from pathlib import Path
+from urllib.parse import quote_plus
+
+import numpy as np
+import pandas as pd
+import matplotlib.pyplot as plt
+import matplotlib.ticker as mticker
+import statsmodels.formula.api as smf
+from dotenv import load_dotenv
+from sqlalchemy import create_engine, text
+from scipy.spatial.distance import cdist
+
+warnings.filterwarnings("ignore", category=UserWarning)
+pd.set_option("display.float_format", "{:,.2f}".format)
+load_dotenv(override=False)
+
+host = os.getenv("PGHOST", "localhost")
+port = os.getenv("5433", "5433")
+user = os.getenv("PGUSER", "postgres")
+password = quote_plus(os.getenv("PGPASSWORD", ""))
+database = os.getenv("PGDATABASE", "texas_data")
+
+engine = create_engine(
+ f"postgresql+psycopg2://{user}:{password}@{host}:{port}/{database}"
+)
+print(f"Connected → {database} on {host}:{port}")
+Connected → texas_data on localhost:5433 ++
1. Data Loading¶
+# District-year inspection metrics aggregated in SQL.
+# LAG() computes days since the previous inspection for the same well (api_norm).
+insp_sql = """
+WITH lagged AS (
+ SELECT
+ district,
+ EXTRACT(year FROM inspection_date)::int AS year,
+ api_norm,
+ inspection_date,
+ CASE WHEN UPPER(compliance::text) IN ('YES', 'Y') THEN 1.0 ELSE 0.0 END AS is_compliant,
+ EXTRACT(EPOCH FROM (
+ inspection_date
+ - LAG(inspection_date) OVER (PARTITION BY api_norm ORDER BY inspection_date)
+ )) / 86400.0 AS days_since_prev
+ FROM inspections
+ WHERE inspection_date IS NOT NULL
+ AND district IS NOT NULL
+ AND EXTRACT(year FROM inspection_date) BETWEEN 2016 AND 2025
+)
+SELECT
+ district,
+ year,
+ COUNT(*) AS total_inspections,
+ COUNT(DISTINCT api_norm) AS unique_wells,
+ ROUND(AVG(is_compliant)::numeric * 100, 2) AS compliance_rate,
+ ROUND(AVG(days_since_prev)::numeric, 1) AS avg_days_between_inspections
+FROM lagged
+GROUP BY district, year
+ORDER BY district, year
+"""
+
+insp = pd.read_sql(text(insp_sql), engine)
+print(f"Inspections panel: {len(insp):,} district-year rows | {insp['district'].nunique()} districts")
+insp.head()
+Inspections panel: 130 district-year rows | 13 districts ++
| + | district | +year | +total_inspections | +unique_wells | +compliance_rate | +avg_days_between_inspections | +
|---|---|---|---|---|---|---|
| 0 | +01 | +2016 | +13975 | +4055 | +69.42 | +18.90 | +
| 1 | +01 | +2017 | +18022 | +6153 | +83.52 | +56.80 | +
| 2 | +01 | +2018 | +23826 | +9109 | +85.61 | +53.50 | +
| 3 | +01 | +2019 | +19790 | +6447 | +84.97 | +79.80 | +
| 4 | +01 | +2020 | +26006 | +8716 | +85.52 | +122.90 | +
# District-year violation metrics. Blank last_enf_action strings treated as no action.
+viol_sql = """
+SELECT
+ district,
+ EXTRACT(year FROM violation_disc_date)::int AS year,
+ COUNT(*) AS total_violations,
+ COUNT(DISTINCT api_norm) AS unique_wells_with_violations,
+ SUM(CASE WHEN major_viol_ind = 'Y' THEN 1 ELSE 0 END) AS major_violations,
+ ROUND(AVG(CASE WHEN compliant_on_reinsp = 'Y' THEN 1.0 ELSE 0.0 END)::numeric * 100, 2)
+ AS resolution_rate,
+ ROUND(AVG(CASE WHEN last_enf_action IS NOT NULL AND last_enf_action <> ''
+ THEN 1.0 ELSE 0.0 END)::numeric * 100, 2) AS enforcement_rate,
+ ROUND(AVG(
+ CASE WHEN last_enf_action_date IS NOT NULL
+ THEN EXTRACT(EPOCH FROM (last_enf_action_date - violation_disc_date)) / 86400.0
+ END
+ )::numeric, 1) AS avg_days_to_enforcement
+FROM violations
+WHERE violation_disc_date IS NOT NULL
+ AND district IS NOT NULL
+ AND EXTRACT(year FROM violation_disc_date) BETWEEN 2016 AND 2025
+GROUP BY district, year
+ORDER BY district, year
+"""
+
+viol = pd.read_sql(text(viol_sql), engine)
+print(f"Violations panel: {len(viol):,} district-year rows")
+viol.head()
+Violations panel: 130 district-year rows ++
| + | district | +year | +total_violations | +unique_wells_with_violations | +major_violations | +resolution_rate | +enforcement_rate | +avg_days_to_enforcement | +
|---|---|---|---|---|---|---|---|---|
| 0 | +01 | +2016 | +5720 | +1009 | +0 | +21.42 | +100.00 | +198.60 | +
| 1 | +01 | +2017 | +4380 | +767 | +0 | +44.36 | +100.00 | +269.50 | +
| 2 | +01 | +2018 | +5766 | +997 | +0 | +64.46 | +100.00 | +229.00 | +
| 3 | +01 | +2019 | +3593 | +902 | +4 | +49.37 | +100.00 | +239.00 | +
| 4 | +01 | +2020 | +4838 | +1019 | +5 | +27.43 | +100.00 | +402.90 | +
BUDGET_PATH = Path("RRC Budget Data.xlsx")
+raw = pd.read_excel(BUDGET_PATH, header=None)
+YEARS = [2016, 2017, 2018, 2019, 2020, 2021, 2022, 2023, 2024]
+COLS = slice(1, 10) # spreadsheet columns 1-9 map to years 2016-2024
+
+# ── Section 1: Energy Resource Development (rows 7-18) ──────────────────────
+erd = pd.DataFrame({
+ "year": YEARS,
+ "strategy": "Energy Resource Development",
+ "total_budget": raw.iloc[1, COLS].values.astype(float),
+ "salaries": raw.iloc[7, COLS].values.astype(float),
+ "other_personnel": raw.iloc[8, COLS].values.astype(float),
+ "professional_fees": raw.iloc[9, COLS].values.astype(float),
+ "travel": raw.iloc[13, COLS].values.astype(float),
+ "other_operating": raw.iloc[16, COLS].values.astype(float),
+ "capital_exp": raw.iloc[17, COLS].values.astype(float),
+ "fte": raw.iloc[18, COLS].values.astype(float),
+})
+
+# ── Section 2: Oil/Gas Monitoring & Inspections (rows 20-31) ────────────────
+ogi = pd.DataFrame({
+ "year": YEARS,
+ "strategy": "Oil/Gas Monitoring & Inspections",
+ "total_budget": raw.iloc[2, COLS].values.astype(float),
+ "salaries": raw.iloc[20, COLS].values.astype(float),
+ "other_personnel": raw.iloc[21, COLS].values.astype(float),
+ "professional_fees": raw.iloc[22, COLS].values.astype(float),
+ "travel": raw.iloc[26, COLS].values.astype(float),
+ "other_operating": raw.iloc[29, COLS].values.astype(float),
+ "capital_exp": raw.iloc[30, COLS].values.astype(float),
+ "fte": raw.iloc[31, COLS].values.astype(float),
+})
+
+budget_long = pd.concat([erd, ogi], ignore_index=True)
+print(f"Budget long: {len(budget_long)} rows (2 strategies × {len(YEARS)} years)")
+budget_long
+Budget long: 18 rows (2 strategies × 9 years) ++
| + | year | +strategy | +total_budget | +salaries | +other_personnel | +professional_fees | +travel | +other_operating | +capital_exp | +fte | +
|---|---|---|---|---|---|---|---|---|---|---|
| 0 | +2016 | +Energy Resource Development | +11,708,475.00 | +7,669,719.00 | +398,589.00 | +3,366,389.00 | +16,477.00 | +210,293.00 | +0.00 | +130.60 | +
| 1 | +2017 | +Energy Resource Development | +10,911,094.00 | +7,273,775.00 | +389,348.00 | +3,118,066.00 | +6,792.00 | +77,855.00 | +0.00 | +120.30 | +
| 2 | +2018 | +Energy Resource Development | +9,846,886.00 | +7,292,933.00 | +282,337.00 | +977,645.00 | +28,694.00 | +1,045,727.00 | +0.00 | +131.00 | +
| 3 | +2019 | +Energy Resource Development | +11,123,757.00 | +8,068,497.00 | +217,988.00 | +1,493,755.00 | +73,651.00 | +988,740.00 | +13,232.00 | +137.40 | +
| 4 | +2020 | +Energy Resource Development | +17,280,569.00 | +9,707,894.00 | +236,356.00 | +5,989,236.00 | +41,752.00 | +1,165,481.00 | +54,037.00 | +153.40 | +
| 5 | +2021 | +Energy Resource Development | +16,237,704.00 | +10,887,561.00 | +237,777.00 | +3,562,816.00 | +5,614.00 | +1,446,301.00 | +10,140.00 | +168.10 | +
| 6 | +2022 | +Energy Resource Development | +25,583,205.00 | +11,166,309.00 | +246,340.00 | +12,560,550.00 | +37,731.00 | +1,246,443.00 | +19,985.00 | +157.10 | +
| 7 | +2023 | +Energy Resource Development | +26,903,564.00 | +11,056,060.00 | +252,933.00 | +12,846,821.00 | +56,650.00 | +2,287,481.00 | +48,344.00 | +151.30 | +
| 8 | +2024 | +Energy Resource Development | +35,533,565.00 | +13,183,578.00 | +229,161.00 | +15,140,585.00 | +144,641.00 | +6,425,653.00 | +0.00 | +186.00 | +
| 9 | +2016 | +Oil/Gas Monitoring & Inspections | +18,471,666.00 | +15,080,122.00 | +685,768.00 | +1,546,321.00 | +22,630.00 | +208,311.00 | +121,363.00 | +256.70 | +
| 10 | +2017 | +Oil/Gas Monitoring & Inspections | +17,204,058.00 | +15,086,262.00 | +686,194.00 | +176,786.00 | +19,654.00 | +230,525.00 | +272,461.00 | +249.50 | +
| 11 | +2018 | +Oil/Gas Monitoring & Inspections | +17,562,431.00 | +13,083,406.00 | +430,429.00 | +1,147,080.00 | +57,312.00 | +1,040,639.00 | +649,172.00 | +229.90 | +
| 12 | +2019 | +Oil/Gas Monitoring & Inspections | +21,951,747.00 | +14,878,875.00 | +340,135.00 | +2,895,436.00 | +187,048.00 | +1,185,772.00 | +1,255,930.00 | +255.60 | +
| 13 | +2020 | +Oil/Gas Monitoring & Inspections | +26,057,560.00 | +17,228,302.00 | +417,683.00 | +4,822,351.00 | +106,428.00 | +1,398,705.00 | +896,846.00 | +284.00 | +
| 14 | +2021 | +Oil/Gas Monitoring & Inspections | +28,756,689.00 | +17,155,864.00 | +426,139.00 | +8,212,873.00 | +34,762.00 | +1,394,783.00 | +230,439.00 | +277.80 | +
| 15 | +2022 | +Oil/Gas Monitoring & Inspections | +25,914,265.00 | +17,834,460.00 | +391,138.00 | +4,007,178.00 | +154,334.00 | +1,255,945.00 | +694,706.00 | +264.00 | +
| 16 | +2023 | +Oil/Gas Monitoring & Inspections | +34,330,858.00 | +18,622,389.00 | +457,753.00 | +8,945,350.00 | +149,418.00 | +2,428,330.00 | +2,234,623.00 | +271.20 | +
| 17 | +2024 | +Oil/Gas Monitoring & Inspections | +38,506,556.00 | +20,834,721.00 | +361,687.00 | +8,851,915.00 | +316,806.00 | +4,112,998.00 | +2,659,208.00 | +280.80 | +
# ── Wide budget: one row per year with ogi_ / erd_ prefixed columns ──────────
+ogi_wide = ogi.drop(columns="strategy").add_prefix("ogi_")
+erd_wide = erd.drop(columns="strategy").add_prefix("erd_")
+
+budget_wide = (
+ ogi_wide
+ .merge(erd_wide, left_on="ogi_year", right_on="erd_year")
+ .rename(columns={"ogi_year": "year"})
+ .drop(columns="erd_year")
+)
+
+# ── Merge inspections + violations, then join statewide budget on year ────────
+panel = (
+ insp
+ .merge(viol, on=["district", "year"], how="left")
+ .merge(budget_wide, on="year", how="left")
+)
+
+# ── Derived columns ───────────────────────────────────────────────────────────
+panel["violations_per_inspection"] = panel["total_violations"] / panel["total_inspections"]
+panel["ogi_budget_m"] = panel["ogi_total_budget"] / 1_000_000 # dollars → millions
+panel["erd_budget_m"] = panel["erd_total_budget"] / 1_000_000
+panel["post_2019"] = (panel["year"] >= 2019).astype(int)
+# 2024 = budget estimate; 2025 = no budget data — exclude both from regressions
+panel["is_budget_year"] = (panel["year"] >= 2024).astype(int)
+
+# Goal ambiguity: share of combined budget going to the inspection mission.
+# Higher share = clearer mission focus; lower share = more goal ambiguity.
+panel["inspection_budget_share"] = (
+ panel["ogi_total_budget"] / (panel["ogi_total_budget"] + panel["erd_total_budget"])
+)
+
+# Fill violation NaNs for districts with zero violations in a given year
+fill_cols = [
+ "total_violations", "unique_wells_with_violations", "major_violations",
+ "resolution_rate", "enforcement_rate", "avg_days_to_enforcement",
+ "violations_per_inspection",
+]
+panel[fill_cols] = panel[fill_cols].fillna(0)
+
+print(f"Analysis panel: {len(panel):,} rows | "
+ f"{panel['district'].nunique()} districts | "
+ f"{panel['year'].nunique()} years")
+panel.head()
+Analysis panel: 130 rows | 13 districts | 10 years ++
| + | district | +year | +total_inspections | +unique_wells | +compliance_rate | +avg_days_between_inspections | +total_violations | +unique_wells_with_violations | +major_violations | +resolution_rate | +... | +erd_travel | +erd_other_operating | +erd_capital_exp | +erd_fte | +violations_per_inspection | +ogi_budget_m | +erd_budget_m | +post_2019 | +is_budget_year | +inspection_budget_share | +
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0 | +01 | +2016 | +13975 | +4055 | +69.42 | +18.90 | +5720 | +1009 | +0 | +21.42 | +... | +16,477.00 | +210,293.00 | +0.00 | +130.60 | +0.41 | +18.47 | +11.71 | +0 | +0 | +0.61 | +
| 1 | +01 | +2017 | +18022 | +6153 | +83.52 | +56.80 | +4380 | +767 | +0 | +44.36 | +... | +6,792.00 | +77,855.00 | +0.00 | +120.30 | +0.24 | +17.20 | +10.91 | +0 | +0 | +0.61 | +
| 2 | +01 | +2018 | +23826 | +9109 | +85.61 | +53.50 | +5766 | +997 | +0 | +64.46 | +... | +28,694.00 | +1,045,727.00 | +0.00 | +131.00 | +0.24 | +17.56 | +9.85 | +0 | +0 | +0.64 | +
| 3 | +01 | +2019 | +19790 | +6447 | +84.97 | +79.80 | +3593 | +902 | +4 | +49.37 | +... | +73,651.00 | +988,740.00 | +13,232.00 | +137.40 | +0.18 | +21.95 | +11.12 | +1 | +0 | +0.66 | +
| 4 | +01 | +2020 | +26006 | +8716 | +85.52 | +122.90 | +4838 | +1019 | +5 | +27.43 | +... | +41,752.00 | +1,165,481.00 | +54,037.00 | +153.40 | +0.19 | +26.06 | +17.28 | +1 | +0 | +0.60 | +
5 rows × 34 columns
+2. Exploratory Overview¶
+# Year-level means across districts
+yearly = panel.groupby("year").agg(
+ ogi_budget_m = ("ogi_budget_m", "first"),
+ ogi_fte = ("ogi_fte", "first"),
+ total_inspections = ("total_inspections", "mean"),
+ compliance_rate = ("compliance_rate", "mean"),
+ total_violations = ("total_violations", "mean"),
+ resolution_rate = ("resolution_rate", "mean"),
+ avg_days_to_enf = ("avg_days_to_enforcement","mean"),
+).round(2)
+
+print(yearly.to_string())
+
+fig, axes = plt.subplots(2, 3, figsize=(16, 8))
+axes = axes.flatten()
+
+yearly["ogi_budget_m"].plot(ax=axes[0], marker="o", title="OGI Budget ($M)")
+axes[0].yaxis.set_major_formatter(mticker.FuncFormatter(lambda x, _: f"${x:.0f}M"))
+
+yearly["ogi_fte"].plot(ax=axes[1], marker="o", title="OGI FTE Positions")
+
+yearly["total_inspections"].plot(ax=axes[2], marker="o", title="Avg Inspections / District")
+
+yearly["compliance_rate"].plot(ax=axes[3], marker="o", title="Avg Compliance Rate (%)")
+
+yearly["resolution_rate"].plot(ax=axes[4], marker="o", title="Avg Resolution Rate (%)")
+
+yearly["avg_days_to_enf"].plot(ax=axes[5], marker="o", title="Avg Days to Enforcement")
+
+for ax in axes:
+ ax.axvline(2019, color="red", linestyle="--", alpha=0.5, label="2019 policy")
+ ax.set_xlabel("Year")
+
+plt.tight_layout()
+plt.show()
+ogi_budget_m ogi_fte total_inspections compliance_rate total_violations resolution_rate avg_days_to_enf +year +2016 18.47 256.70 18,277.85 83.11 3,398.15 36.78 131.86 +2017 17.20 249.50 20,138.54 86.52 2,915.69 59.02 185.01 +2018 17.56 229.90 25,703.54 90.17 3,197.62 59.46 207.25 +2019 21.95 255.60 25,058.46 89.85 2,550.77 61.44 170.36 +2020 26.06 284.00 27,669.46 89.57 2,750.92 56.81 154.66 +2021 28.76 277.80 24,115.54 88.76 2,556.38 66.18 118.82 +2022 25.91 264.00 32,023.54 89.82 2,819.92 67.85 91.50 +2023 34.33 271.20 33,805.69 91.62 2,598.62 69.65 105.15 +2024 38.51 280.80 36,552.77 92.58 2,221.15 65.13 76.93 +2025 NaN NaN 34,082.08 90.52 2,530.38 52.06 36.62 ++
+corr_cols = [
+ "ogi_budget_m", "ogi_fte", "inspection_budget_share",
+ "total_inspections", "compliance_rate",
+ "total_violations", "resolution_rate", "avg_days_to_enforcement",
+]
+corr = panel[corr_cols].corr().round(2)
+
+fig, ax = plt.subplots(figsize=(9, 7))
+im = ax.imshow(corr, cmap="RdBu_r", vmin=-1, vmax=1)
+ax.set_xticks(range(len(corr_cols)))
+ax.set_yticks(range(len(corr_cols)))
+ax.set_xticklabels(corr_cols, rotation=45, ha="right", fontsize=9)
+ax.set_yticklabels(corr_cols, fontsize=9)
+for i in range(len(corr_cols)):
+ for j in range(len(corr_cols)):
+ ax.text(j, i, corr.iloc[i, j], ha="center", va="center", fontsize=8)
+plt.colorbar(im, ax=ax)
+ax.set_title("Correlation Matrix — Key Variables")
+plt.tight_layout()
+plt.show()
+
+Data and Methods¶
Data Sources¶
This study draws on two primary data sources. The first is the Texas Railroad Commission
+(RRC) Oil and Gas Division administrative database. Inspection records span fiscal years 2016–2025 and encompass approximately
+1.9 million inspection events distributed across 13 RRC administrative districts;
+violation records include approximately 193,000 enforcement actions. From the inspections
+table, district-year aggregates are constructed for three regulatory output measures:
+(1) compliance rate — the share of annual inspections in a district that did not result
+in a compliance failure; (2) total inspections — the count of field inspection events;
+and (3) average days between successive inspections of the same well, computed via a
+SQL window function (LAG) over ordered inspection timestamps. From the violations table,
+district-year aggregates include the violation resolution rate (share of violations
+for which the operator was found compliant on re-inspection), enforcement rate, and average
+days from violation discovery to enforcement action.
The second source is RRC budget data drawn from Legislative Appropriations Requests, +covering fiscal years 2016–2024. Budget appropriations are reported at the statewide level +disaggregated by goal and strategy. Two strategies are central to this analysis: +(1) Oil and Gas Monitoring and Inspections (OGI), which directly funds field inspection +operations; and (2) Energy Resource Development (ERD), encompassing the broader mandate +to promote oil and gas resource opportunities. For each strategy, the data include total +appropriations, salaries, professional fees, travel, other operating expenditures, capital +outlays, and authorized full-time equivalent (FTE) positions. Fiscal year 2024 represents +a budget estimate rather than expenditure actuals and is excluded from all regression +models.
+Sample and Panel Construction¶
The unit of analysis is the district-year. The analytic panel contains +N = 130 observations (13 districts × 10 years, 2016–2025), of which +104 observations (2016–2023) constitute the regression sample. Fiscal years +2024 (budget estimate) and 2025 (no budget data available) are retained in +descriptive analyses but excluded from all regression models. Because inspection +and enforcement activity in 2025 represents a partial year as of the data +extract, enforcement-timing metrics for that year are subject to right-censoring: +violations discovered in late 2024 and 2025 may not yet have received a recorded +enforcement action, compressing observed days-to-enforcement.. Because RRC budget +appropriations are reported at the statewide level, budget and FTE variables enter the +panel as year-varying but district-invariant covariates. Identification of budget effects +therefore relies on year-to-year variation in statewide appropriations rather than +cross-district budget contrasts.
+Measures¶
Dependent variables. Three measures capture distinct dimensions of regulatory output: +total inspections (inspection volume), compliance rate (%), and violation resolution +rate (%). Compliance rate and resolution rate capture quality of enforcement rather than +quantity and represent different points in the regulatory pipeline: compliance is measured +at the point of inspection while resolution is measured after a violation has been +discovered and acted upon.
+Organizational capacity. The primary capacity measure is OGI total appropriations in +millions of dollars ($\text{Budget}_t$), reflecting the statewide resource envelope +available for inspection activities in year $t$. An auxiliary measure — OGI authorized +FTE positions — is included in descriptive analyses.
+Goal ambiguity. Following Chun and Rainey (2005), goal ambiguity is operationalized +via the relative concentration of resources across missions. The inspection budget share +($\text{Share}_t$) captures the fraction of combined OGI and ERD appropriations directed +toward the inspection mandate:
+$$\text{Share}_t = \frac{\text{OGI Budget}_t}{\text{OGI Budget}_t + \text{ERD Budget}_t}$$
+Higher values indicate greater mission clarity (resources more concentrated on inspections); +lower values indicate greater goal ambiguity (resources spread across competing mandates). +Over the study period $\text{Share}_t$ ranged from 0.59 (2022) to 0.67 (2018), reflecting +meaningful year-to-year variation in budgetary prioritization.
+Geographic moderators. Two binary district-level indicators capture geographic +context: $\text{Offshore}_d = 1$ for districts 02, 03, and 04, which hold dual onshore +and offshore oversight jurisdiction, and $\text{Border}_d = 1$ for districts 01–04, +which are proximate to the Texas Gulf Coast and the US–Mexico border corridor.
+Estimation Strategy¶
All models are estimated via ordinary least squares (OLS) with standard errors clustered +at the district level ($G = 13$) to account for within-district serial correlation. +District fixed effects absorb time-invariant heterogeneity across offices — including +differences in geographic complexity, historical enforcement culture, and staffing +composition — and ensure that budget effects are identified from within-district, +year-to-year variation.
+H1 — Baseline capacity model:
+$$Y_{dt} = \alpha + \beta_1 \, \text{Budget}_t + \sum_{d} \gamma_d \, \mathbf{1}[\text{district} = d] + \varepsilon_{dt}$$
+where $Y_{dt}$ is the regulatory output for district $d$ in year $t$, $\gamma_d$ are +district fixed effects, and $\varepsilon_{dt}$ is the idiosyncratic error.
+H2 — Goal ambiguity moderation:
+$$Y_{dt} = \alpha + \beta_1 \, \text{Budget}_t + \beta_2 \, \text{Share}_t + \beta_3 \left( \text{Budget}_t \times \text{Share}_t \right) + \sum_{d} \gamma_d + \varepsilon_{dt}$$
+The coefficient $\beta_3$ tests whether goal clarity conditions the capacity–output +relationship. A positive $\hat{\beta}_3$ would indicate that clearer mission focus +amplifies budget effects; a negative value would suggest diminishing returns or +cross-strategy resource substitution.
+H3 — District slope heterogeneity:
+$$Y_{dt} = \alpha + \beta_1 \, \text{Budget}_t + \sum_{d=2}^{D} \delta_d \left( \text{Budget}_t \times \mathbf{1}[d] \right) + \sum_{d} \gamma_d + \varepsilon_{dt}$$
+District-specific budget slopes are recovered as $\hat{\beta}_1 + \hat{\delta}_d$. +Because budget varies only along the time dimension and district fixed effects are +included, interaction term standard errors are inflated by near-perfect multicollinearity; +these estimates are treated as descriptive indicators of heterogeneity only.
+H4 — Geographic moderation and spatial autocorrelation:
+$$Y_{dt} = \alpha + \beta_1 \, \text{Budget}_t + \beta_2 \, \text{Offshore}_d + \beta_3 \, \text{Border}_d + \beta_4 \left( \text{Budget}_t \times \text{Offshore}_d \right) + \beta_5 \left( \text{Budget}_t \times \text{Border}_d \right) + \sum_{d} \gamma_d + \varepsilon_{dt}$$
+Robustness checks. Two supplementary tests address limitations of the +baseline models. First, wild cluster bootstrap inference (Rademacher weights, +$B = 999$ draws; Cameron, Gelbach & Miller 2008) is used to re-test H1 +coefficients, providing valid p-values with the small number of clusters +($G = 13$). Second, a distributed lag specification replaces the +contemporaneous budget measure with its one-year lag ($\text{Budget}_{t-1}$), +and also estimates a model including both, to test whether budget effects +operate with a delay consistent with a hiring-and-deployment mechanism. +The distributed lag regression sample covers 2017–2023 ($N = 91$).
+Spatial autocorrelation in H1 model residuals is assessed via Moran's $I$ computed on a +row-normalized inverse-distance spatial weights matrix constructed from district centroids +derived by averaging well-level geographic coordinates within each district.
+Analysis¶
This study employs a fixed-effects panel regression framework to examine whether +year-to-year changes in RRC organizational capacity — as measured by statewide budget +appropriations — translate into improvements in regulatory outputs across Texas oil and +gas inspection districts. The analytic panel spans 13 RRC districts over ten fiscal +years (2016–2025), yielding 130 district-year observations. Regression analyses are +restricted to 2016–2023 (N = 104), excluding FY2024 (budget estimate only) and FY2025 +(no budget data available). The identification strategy leverages within-district +variation in outcomes as a function of year-to-year shifts in statewide OGI +appropriations, net of persistent inter-district differences absorbed by district fixed +effects.
+The choice of a district-year panel rather than a well-level panel is motivated by the +structure of the budget data, which is available only at the statewide level. Because the +key independent variable — OGI appropriations — varies along the time dimension only, it +functions as a common, year-specific exposure applied uniformly to all districts. District +fixed effects then absorb unobservable office-level characteristics that remain stable over +the study period, such as geographic complexity, historical enforcement intensity, and +local administrative capacity. Causal identification is thus predicated on the assumption +that, absent changes in budget, within-district outcome trajectories would have followed +parallel trends across years — an assumption that cannot be directly tested but is +partially supported by the pre-period stability visible in the descriptive trends.
+H1 tests the core capacity hypothesis using the baseline specification. Each of the +three dependent variables — total inspections, compliance rate, and violation resolution +rate — is regressed separately on OGI budget (in millions of dollars) and district fixed +effects. Cluster-robust standard errors are used throughout given the modest number of +clusters ($G = 13$).
+H2 extends the baseline by interacting OGI budget with the inspection budget share, +operationalizing goal ambiguity as the degree to which RRC appropriations are concentrated +on the inspection mandate versus the broader energy development mission. The sign and +significance of the interaction term $\beta_3$ determines whether goal clarity amplifies +or attenuates the capacity–output relationship.
+H3 tests for heterogeneity in budget–outcome slopes across districts by including +budget $\times$ district interaction terms. Given only eight years of data per district, +the saturated interaction model is estimated with approximately zero residual degrees of +freedom for the fixed-effects component; as a result, interaction-term standard errors +are unreliable and these estimates are reported as exploratory indicators of cross-district +variation rather than inferential tests. The accompanying bar chart (below) summarizes +district-specific slopes as point estimates.
+H4 assesses whether offshore-jurisdiction and border-proximate districts — which face +distinct operational environments — exhibit different budget sensitivity. The model adds +geographic level effects and budget $\times$ geography interaction terms to the baseline +specification. A complementary spatial diagnostic — Moran's $I$ applied to the residuals +from the H1 compliance model — tests for geographic clustering of unexplained outcome +variation that could indicate omitted spatial processes or spillovers across district +boundaries.
+All regressions exclude fiscal years 2024 (budget estimate) and 2025 (no budget data), +retaining 2016–2023 as the regression sample (N = 104). The extended panel through 2025 +is used for descriptive trend analysis only. Enforcement-timing metrics for 2025 should +be interpreted cautiously: because the data extract covers a partial year, violations +discovered in late 2024 and 2025 may not yet have a recorded enforcement action, +artificially compressing observed days-to-enforcement and resolution rates for that year.
+Two supplementary robustness checks address key inferential limitations. First, wild +cluster bootstrap inference (Rademacher, B = 999) re-tests H1 with valid small-sample +p-values given G = 13 clusters. Second, a distributed lag specification tests whether +budget effects operate with a one-year delay, consistent with a hiring-and-deployment +implementation timeline. Results from both checks are reported following the main +hypothesis tests.
+H1: Organizational Capacity → Policy Outputs¶
Prediction: Higher OGI budget predicts more inspections, higher compliance rates, +and faster violation resolution.
+Model: OLS with district fixed effects, 2016–2023 (N = 104). Budget varies only over +time, identifying effects via year-to-year changes in statewide OGI appropriations; +district fixed effects absorb persistent cross-district differences. Standard errors +clustered at the district level (G = 13).
+Finding (preview): All three outcomes show positive, statistically significant budget +coefficients under asymptotic inference. Wild cluster bootstrap results (reported in the +Robustness Checks section) indicate these asymptotic p-values overstate precision; results +should be interpreted as suggestive rather than definitive.
+actuals = panel[panel["is_budget_year"] == 0].copy()
+
+fig, axes = plt.subplots(1, 3, figsize=(15, 4))
+
+actuals.plot.scatter(x="ogi_budget_m", y="total_inspections",
+ alpha=0.4, ax=axes[0], title="Budget → Inspections")
+actuals.plot.scatter(x="ogi_budget_m", y="compliance_rate",
+ alpha=0.4, ax=axes[1], title="Budget → Compliance Rate (%)")
+actuals.plot.scatter(x="ogi_budget_m", y="resolution_rate",
+ alpha=0.4, ax=axes[2], title="Budget → Resolution Rate (%)")
+
+for ax in axes:
+ ax.set_xlabel("OGI Budget ($M)")
+
+plt.tight_layout()
+plt.show()
+
+m_inspections = smf.ols(
+ "total_inspections ~ ogi_budget_m + C(district)",
+ data=actuals,
+).fit(cov_type="cluster", cov_kwds={"groups": actuals["district"]})
+
+m_compliance = smf.ols(
+ "compliance_rate ~ ogi_budget_m + C(district)",
+ data=actuals,
+).fit(cov_type="cluster", cov_kwds={"groups": actuals["district"]})
+
+m_resolution = smf.ols(
+ "resolution_rate ~ ogi_budget_m + C(district)",
+ data=actuals,
+).fit(cov_type="cluster", cov_kwds={"groups": actuals["district"]})
+
+# Detect actual column names — statsmodels uses z/P>|z| with robust SEs in some versions
+_tbl = m_inspections.summary2().tables[1]
+_t = "t" if "t" in _tbl.columns else "z"
+_p = "P>|t|" if "P>|t|" in _tbl.columns else "P>|z|"
+display_cols = ["Coef.", "Std.Err.", _t, _p]
+
+print("H1a — OGI Budget ($M) → Total Inspections")
+print(m_inspections.summary2().tables[1][display_cols].loc[["ogi_budget_m"]])
+print(f" R² = {m_inspections.rsquared:.3f} Adj. R² = {m_inspections.rsquared_adj:.3f}\n")
+
+print("H1b — OGI Budget ($M) → Compliance Rate (%)")
+print(m_compliance.summary2().tables[1][display_cols].loc[["ogi_budget_m"]])
+print(f" R² = {m_compliance.rsquared:.3f} Adj. R² = {m_compliance.rsquared_adj:.3f}\n")
+
+print("H1c — OGI Budget ($M) → Resolution Rate (%)")
+print(m_resolution.summary2().tables[1][display_cols].loc[["ogi_budget_m"]])
+print(f" R² = {m_resolution.rsquared:.3f} Adj. R² = {m_resolution.rsquared_adj:.3f}")
+H1a — OGI Budget ($M) → Total Inspections + Coef. Std.Err. z P>|z| +ogi_budget_m 666.30 212.98 3.13 0.00 + R² = 0.769 Adj. R² = 0.736 + +H1b — OGI Budget ($M) → Compliance Rate (%) + Coef. Std.Err. z P>|z| +ogi_budget_m 0.26 0.11 2.31 0.02 + R² = 0.538 Adj. R² = 0.471 + +H1c — OGI Budget ($M) → Resolution Rate (%) + Coef. Std.Err. z P>|z| +ogi_budget_m 1.05 0.32 3.28 0.00 + R² = 0.624 Adj. R² = 0.569 ++
H2: Goal Ambiguity Moderates Capacity Effects¶
Prediction: When a larger share of combined RRC budget flows to the broader
+"Energy Resource Development" goal (lower inspection_budget_share), the capacity →
+output link weakens. A positive interaction coefficient would support H2.
Operationalization:
+inspection_budget_share = ogi_budget / (ogi_budget + erd_budget)
Identification note: Like the budget measure itself, inspection_budget_share
+varies only over time, not across districts. The interaction term therefore exploits
+the same narrow temporal variation as the main effect — budget share ranged from 0.59
+to 0.67 over 2016–2023, a span of 8 percentage points across 8 years. This limits
+the strength of inference that can be drawn from the moderation test.
Finding (preview): The interaction is significant and negative ($\hat{\beta}_3 = -6.53$, +$p < .01$), but interpretation is constrained by the identification limitations above. +Results are discussed in the Results section.
+m_h2 = smf.ols(
+ "compliance_rate ~ ogi_budget_m * inspection_budget_share + C(district)",
+ data=actuals,
+).fit(cov_type="cluster", cov_kwds={"groups": actuals["district"]})
+
+key_rows = ["ogi_budget_m", "inspection_budget_share", "ogi_budget_m:inspection_budget_share"]
+print("H2 — Goal Ambiguity Moderation (DV: compliance_rate)")
+print(m_h2.summary2().tables[1][display_cols].loc[key_rows])
+print(f"\nR² = {m_h2.rsquared:.3f} Adj. R² = {m_h2.rsquared_adj:.3f}")
+
+# ── Same model with resolution rate as DV ────────────────────────────────────
+m_h2_res = smf.ols(
+ "resolution_rate ~ ogi_budget_m * inspection_budget_share + C(district)",
+ data=actuals,
+).fit(cov_type="cluster", cov_kwds={"groups": actuals["district"]})
+
+print("\nH2 — Goal Ambiguity Moderation (DV: resolution_rate)")
+print(m_h2_res.summary2().tables[1][display_cols].loc[key_rows])
+print(f"\nR² = {m_h2_res.rsquared:.3f} Adj. R² = {m_h2_res.rsquared_adj:.3f}")
+H2 — Goal Ambiguity Moderation (DV: compliance_rate) + Coef. Std.Err. z P>|z| +ogi_budget_m 4.20 1.09 3.86 0.00 +inspection_budget_share 170.18 44.79 3.80 0.00 +ogi_budget_m:inspection_budget_share -6.53 1.84 -3.55 0.00 + +R² = 0.567 Adj. R² = 0.493 + +H2 — Goal Ambiguity Moderation (DV: resolution_rate) + Coef. Std.Err. z P>|z| +ogi_budget_m 6.68 4.67 1.43 0.15 +inspection_budget_share 230.67 204.30 1.13 0.26 +ogi_budget_m:inspection_budget_share -9.42 7.99 -1.18 0.24 + +R² = 0.629 Adj. R² = 0.566 ++
H3: District Multilevel Effects¶
Prediction: The budget → output slope varies across RRC districts — some districts +translate budget increases into better outputs more effectively than others.
+Model: Interaction ogi_budget_m × C(district) — the reference district captures
+the baseline budget slope; interaction terms show how each other district's slope
+differs. Standard errors are unreliable due to near-perfect multicollinearity in the
+saturated model (budget varies only over time while district FE absorb cross-sectional
+variation); results are treated as descriptive point estimates only.
Finding (preview): District slopes for compliance rate range from −0.34 pp per $1M +(District 03, Houston/Coastal) to +1.36 pp per $1M (District 6E, East Texas Piney +Woods), with most districts showing small positive slopes. The bar chart below plots +district-specific slope estimates.
+m_h3 = smf.ols(
+ "compliance_rate ~ ogi_budget_m * C(district)",
+ data=actuals,
+).fit(cov_type="cluster", cov_kwds={"groups": actuals["district"]})
+
+coef_table = m_h3.summary2().tables[1]
+
+# Baseline budget slope (reference district)
+baseline_row = coef_table.loc[["ogi_budget_m"]]
+print("H3 — District-Heterogeneous Budget Effect (DV: compliance_rate)")
+print(f"Baseline (reference district) budget slope:")
+print(baseline_row[display_cols])
+
+# District-specific deviations from baseline
+interaction_rows = coef_table[coef_table.index.str.contains("ogi_budget_m:C")]
+print("\nDistrict interaction terms (deviation from reference slope):")
+print(interaction_rows[display_cols].round(4))
+print(f"\nR² = {m_h3.rsquared:.3f} Adj. R² = {m_h3.rsquared_adj:.3f}")
+
+# ── Plot district-specific budget slopes ─────────────────────────────────────
+districts = actuals["district"].unique()
+slopes = {}
+for d in districts:
+ key = f"ogi_budget_m:C(district)[T.{d}]"
+ base = m_h3.params.get("ogi_budget_m", 0)
+ delta = m_h3.params.get(key, 0)
+ slopes[str(d)] = base + delta
+
+slope_df = pd.Series(slopes).sort_values()
+fig, ax = plt.subplots(figsize=(10, 4))
+slope_df.plot.barh(ax=ax, color=["#d62728" if v < 0 else "#1f77b4" for v in slope_df])
+ax.axvline(0, color="black", linewidth=0.8)
+ax.set_xlabel("Budget slope (compliance rate pp per $M)")
+ax.set_title("H3 — District-Specific Budget → Compliance Slopes")
+plt.tight_layout()
+plt.show()
+H3 — District-Heterogeneous Budget Effect (DV: compliance_rate) +Baseline (reference district) budget slope: + Coef. Std.Err. z P>|z| +ogi_budget_m 0.09 0.00 56,876,193,472,228.37 0.00 + +District interaction terms (deviation from reference slope): + Coef. Std.Err. z P>|z| +ogi_budget_m:C(district)[T.02] 0.15 0.00 22,633,237,551,336.32 0.00 +ogi_budget_m:C(district)[T.03] -0.43 0.00 -59,804,100,493,329.36 0.00 +ogi_budget_m:C(district)[T.04] 0.19 0.00 78,131,153,896,367.78 0.00 +ogi_budget_m:C(district)[T.05] -0.04 0.00 -23,701,820,832,698.50 0.00 +ogi_budget_m:C(district)[T.06] 0.34 0.00 60,365,540,001,288.30 0.00 +ogi_budget_m:C(district)[T.08] 0.19 0.00 10,356,376,563,126.46 0.00 +ogi_budget_m:C(district)[T.09] -0.09 0.00 -14,544,886,315,847.22 0.00 +ogi_budget_m:C(district)[T.10] 0.04 0.00 5,748,033,218,673.02 0.00 +ogi_budget_m:C(district)[T.6E] 1.27 0.00 64,743,648,722,385.09 0.00 +ogi_budget_m:C(district)[T.7B] 0.18 0.00 27,978,802,690,136.84 0.00 +ogi_budget_m:C(district)[T.7C] 0.31 0.00 24,243,474,173,332.52 0.00 +ogi_budget_m:C(district)[T.8A] 0.10 0.00 59,702,739,775,453.20 0.00 + +R² = 0.662 Adj. R² = 0.554 ++
+H4: Spatial and Geographic Factors¶
Predictions:
+-
+
- Offshore-jurisdiction districts (02, 03, 04) show a different budget → output +relationship due to dual onshore/offshore oversight burden. +
- Border-proximate districts show a different relationship due to cross-jurisdiction +enforcement complexity. +
- Spatial autocorrelation in H1 residuals (Moran's I) would indicate unmodeled +geographic spillovers. +
Finding (preview): Offshore and border districts show significantly higher baseline +compliance rates (+7.6 pp and +6.0 pp respectively, both $p < .05$) but not different +budget sensitivity. Moran's $I = -0.051$ indicates slight spatial dispersion and no +significant geographic clustering of residuals. Results are discussed in the Results +section.
+# Texas RRC district geography flags (based on known RRC district locations)
+OFFSHORE_DISTRICTS = {"02", "03", "04"} # dual onshore + offshore jurisdiction
+BORDER_DISTRICTS = {"01", "02", "03", "04"} # south / gulf coast proximity
+
+actuals = actuals.copy()
+actuals["district_str"] = actuals["district"].astype(str).str.strip()
+actuals["offshore"] = actuals["district_str"].isin(OFFSHORE_DISTRICTS).astype(int)
+actuals["border"] = actuals["district_str"].isin(BORDER_DISTRICTS).astype(int)
+
+print("District classification:")
+print(
+ actuals.groupby(["district_str", "offshore", "border"])
+ .size()
+ .reset_index(name="district_year_obs")
+ .to_string(index=False)
+)
+District classification: +district_str offshore border district_year_obs + 01 0 1 8 + 02 1 1 8 + 03 1 1 8 + 04 1 1 8 + 05 0 0 8 + 06 0 0 8 + 08 0 0 8 + 09 0 0 8 + 10 0 0 8 + 6E 0 0 8 + 7B 0 0 8 + 7C 0 0 8 + 8A 0 0 8 ++
# ── Spatial regression: offshore and border interactions ─────────────────────
+m_h4 = smf.ols(
+ "compliance_rate ~ ogi_budget_m + offshore + border "
+ "+ ogi_budget_m:offshore + ogi_budget_m:border + C(district)",
+ data=actuals,
+).fit(cov_type="cluster", cov_kwds={"groups": actuals["district"]})
+
+spatial_rows = [
+ "ogi_budget_m", "offshore", "border",
+ "ogi_budget_m:offshore", "ogi_budget_m:border",
+]
+available = [r for r in spatial_rows if r in m_h4.params.index]
+print("H4 — Spatial Moderators (DV: compliance_rate)")
+print(m_h4.summary2().tables[1][display_cols].loc[available])
+print(f"\nR² = {m_h4.rsquared:.3f} Adj. R² = {m_h4.rsquared_adj:.3f}")
+
+# ── Moran's I on H1 residuals ─────────────────────────────────────────────────
+# Compute district centroids from well lat/lon joined via inspections
+centroids_sql = """
+SELECT
+ i.district,
+ AVG(w.latitude) AS lat,
+ AVG(w.longitude) AS lon
+FROM inspections i
+JOIN well_shape_tract w USING (api_norm)
+WHERE w.latitude IS NOT NULL
+ AND w.longitude IS NOT NULL
+ AND i.district IS NOT NULL
+GROUP BY i.district
+"""
+
+try:
+ centroids = pd.read_sql(text(centroids_sql), engine)
+
+ # Average H1 compliance residuals to district level
+ resid_df = actuals[["district", "compliance_rate"]].copy()
+ resid_df["resid"] = m_compliance.resid.reindex(actuals.index).values
+ resid_by_district = resid_df.groupby("district")["resid"].mean().reset_index()
+
+ centroids = centroids.merge(resid_by_district, on="district").dropna()
+
+ # Row-normalised inverse-distance weights matrix
+ coords = centroids[["lon", "lat"]].values
+ D = cdist(coords, coords)
+ np.fill_diagonal(D, np.inf)
+ W = 1 / D
+ W = W / W.sum(axis=1, keepdims=True)
+
+ z = centroids["resid"].values
+ z = z - z.mean()
+ n = len(z)
+ morans_i = (n / W.sum()) * (z @ W @ z) / (z @ z)
+
+ print(f"\nMoran's I on H1 compliance residuals = {morans_i:.4f}")
+ print(" > 0 → residuals cluster spatially (similar neighbours)")
+ print(" ≈ 0 → no spatial pattern")
+ print(" < 0 → spatial dispersion (dissimilar neighbours)")
+
+ print("\nDistrict centroids used:")
+ print(centroids[["district", "lat", "lon"]].round(2).to_string(index=False))
+
+except Exception as e:
+ print(f"Moran's I skipped: {e}")
+H4 — Spatial Moderators (DV: compliance_rate) + Coef. Std.Err. z P>|z| +ogi_budget_m 0.35 0.15 2.39 0.02 +offshore 7.61 3.29 2.31 0.02 +border 6.03 2.84 2.12 0.03 +ogi_budget_m:offshore -0.03 0.18 -0.16 0.87 +ogi_budget_m:border -0.25 0.15 -1.74 0.08 + +R² = 0.553 Adj. R² = 0.476 + +Moran's I on H1 compliance residuals = -0.0512 + > 0 → residuals cluster spatially (similar neighbours) + ≈ 0 → no spatial pattern + < 0 → spatial dispersion (dissimilar neighbours) + +District centroids used: +district lat lon + 01 29.15 -98.62 + 02 28.85 -97.41 + 03 30.12 -95.43 + 04 27.44 -98.36 + 05 31.85 -96.15 + 06 32.29 -94.67 + 08 31.84 -102.30 + 09 33.42 -98.22 + 10 35.77 -101.02 + 6E 32.40 -94.89 + 7B 32.75 -99.40 + 7C 31.11 -101.26 + 8A 33.12 -102.06 ++
Results¶
Descriptive Trends¶
Table 1 summarizes year-level means for the key variables across 2016–2025, with +regression analyses restricted to 2016–2023. OGI appropriations grew from $18.47 million +in 2016 to $34.33 million in 2023 — an 86 percent nominal increase — with the FY2024 +budget estimate reaching $38.51 million. Authorized FTE positions rose modestly from +256.7 to 271.2 over the same period. Inspection volume per district increased from a +mean of 18,278 in 2016 to a peak of 36,553 in 2024, with a partial-year figure of 34,082 +recorded for 2025. Mean district compliance rate improved from 83.1 percent in 2016 to +a peak of 92.6 percent in 2024, with a slight moderation to 90.5 percent in the 2025 +partial-year extract. Violation resolution rate rose from 36.8 percent in 2016 to 69.7 +percent in 2023 before declining to 52.1 percent in 2025; this decline almost certainly +reflects right-censoring rather than a genuine deterioration in enforcement outcomes, as +recently discovered violations will not yet have received a recorded resolution on +re-inspection. Similarly, the 2025 days-to-enforcement figure of 36.6 days should be +interpreted as a lower bound on the true enforcement timeline for that cohort of +violations. These trends are broadly consistent with the organizational capacity +hypothesis, though they are also consistent with secular improvements in industry +compliance independent of budget growth.
+Table 1. Year-Level Panel Means, 2016–2025
+| Year | +OGI Budget ($M) | +OGI FTE | +Inspections/District | +Compliance Rate (%) | +Resolution Rate (%) | +Days to Enforcement | +
|---|---|---|---|---|---|---|
| 2016 | +18.47 | +256.7 | +18,278 | +83.1 | +36.8 | +131.9 | +
| 2017 | +17.20 | +249.5 | +20,139 | +86.5 | +59.0 | +185.0 | +
| 2018 | +17.56 | +229.9 | +25,704 | +90.2 | +59.5 | +207.3 | +
| 2019 | +21.95 | +255.6 | +25,058 | +89.9 | +61.4 | +170.4 | +
| 2020 | +26.06 | +284.0 | +27,669 | +89.6 | +56.8 | +154.7 | +
| 2021 | +28.76 | +277.8 | +24,116 | +88.8 | +66.2 | +118.8 | +
| 2022 | +25.91 | +264.0 | +32,024 | +89.8 | +67.9 | +91.5 | +
| 2023 | +34.33 | +271.2 | +33,806 | +91.6 | +69.7 | +105.2 | +
| 2024† | +38.51 | +280.8 | +36,553 | +92.6 | +65.1 | +76.9 | +
| 2025‡ | +— | +— | +34,082 | +90.5 | +52.1 | +36.6‡ | +
Note: Budget figures are nominal. FTE = authorized full-time equivalent positions. +Inspections/District = mean district-level annual inspection count. +† 2024 budget is an appropriations estimate, not expenditure actuals; excluded from +regression models. +‡ 2025 data is partial-year as of the data extract. Resolution rate and days-to-enforcement +are right-censored: violations discovered in late 2024–2025 may not yet have a recorded +enforcement action, compressing these metrics.
++
H1: Organizational Capacity and Regulatory Outputs¶
The baseline fixed-effects models provide consistent support for H1 across all three +dependent variables (Table 2). Each additional million dollars in OGI appropriations is +associated with approximately 666 additional district-level inspections per year +($\hat{\beta} = 666.30$, SE = 212.98, $z = 3.13$, $p < .01$; $R^2 = .769$). The budget +coefficient is also positive and significant for compliance rate ($\hat{\beta} = 0.26$ +percentage points per $1M, SE = 0.11, $z = 2.31$, $p = .02$; $R^2 = .538$) and +violation resolution rate ($\hat{\beta} = 1.05$ percentage points per $1M, SE = 0.32, +$z = 3.28$, $p < .01$; $R^2 = .624$). These associations are estimated net of district +fixed effects and therefore reflect within-district covariation between annual budget +changes and outcome changes rather than cross-sectional differences between +better- and worse-funded districts.
+Table 2. H1 Regression Results: OGI Budget → Regulatory Outputs
+| Dependent Variable | +$\hat{\beta}$ (Budget $M) | +SE | +$z$ | +$p$ | +$R^2$ | +Adj. $R^2$ | +
|---|---|---|---|---|---|---|
| Total inspections | +666.30 | +212.98 | +3.13 | +<.01 | +.769 | +.736 | +
| Compliance rate (%) | +0.26 | +0.11 | +2.31 | +.02 | +.538 | +.471 | +
| Resolution rate (%) | +1.05 | +0.32 | +3.28 | +<.01 | +.624 | +.569 | +
Note: All models include district fixed effects ($D = 13$). Standard errors clustered +at the district level. $N = 104$.
++
H2: Goal Ambiguity as a Moderator¶
The goal ambiguity moderation model for compliance rate (Table 3) yields a statistically +significant and negative interaction between OGI budget and inspection budget share +($\hat{\beta}_3 = -6.53$, SE = 1.84, $z = -3.55$, $p < .01$). However, this result +requires careful qualification before any mechanism is claimed.
+The key issue is that inspection_budget_share — like the budget measure itself —
+varies only over time, not across districts. All 13 districts experience the same
+budget share in any given year, ranging from 0.59 (FY2022) to 0.67 (FY2018) across
+the study period — a span of 8 percentage points over 8 observations. The interaction
+term is therefore identified from the same narrow temporal variation as the main budget
+effect, not from cross-district differences in mission structure. This makes it
+difficult to distinguish a genuine moderation relationship from a spurious correlation
+with year-specific factors that independently affected both budget share and compliance
+outcomes in the same years.
The negative sign is consistent with at least two interpretations. Under a +resource saturation story, compliance gains from additional OGI investment contract +as the inspection mandate becomes better resourced relative to other RRC goals — +a plausible ceiling effect if districts are already operating near full compliance +in high-share years. Alternatively, the result may simply reflect that FY2018 — the +highest-share year — saw particularly large compliance gains for reasons unrelated to +budget concentration (e.g., post-2016 industry recovery, early implementation of +regulatory changes). Evaluated at mean budget share ($\bar{s} \approx 0.62$), the +implied marginal budget effect on compliance is $4.20 - 6.53(0.62) \approx 0.15$ +pp per $1M — directionally consistent with H1 but smaller.
+For violation resolution rate, no terms reach conventional significance (all $p > .15$). +Given the identification constraints, the H2 compliance finding is best treated as an +exploratory pattern consistent with goal ambiguity theory — one that motivates future +research with district-level budget variation — rather than a robust confirmatory test.
+Table 3. H2 Regression Results: Goal Ambiguity Moderation (DV: Compliance Rate)
+| Term | +$\hat{\beta}$ | +SE | +$z$ | +$p$ | +
|---|---|---|---|---|
| Budget ($M) | +4.20 | +1.09 | +3.86 | +<.01 | +
| Inspection budget share | +170.18 | +44.79 | +3.80 | +<.01 | +
| Budget × Share | +−6.53 | +1.84 | +−3.55 | +<.01 | +
Note: District fixed effects included. SE clustered at district. $R^2 = .567$, +Adj. $R^2 = .493$. $N = 104$.
++
H3: District-Level Heterogeneity¶
District-specific budget slopes for compliance rate range from $-0.34$ percentage points +per $1 million (District 03, Coastal/Greater Houston) to $+1.36$ percentage points +(District 6E, East Texas Piney Woods), with most districts showing small positive slopes +(Table 4). The reference district (District 01, San Antonio) slope is 0.09 pp per $1M. +Positive slopes are most pronounced in District 6E (+1.36), District 06 (+0.43), and +District 7C (+0.40); District 03 is the only district with a substantially negative slope. +The model $R^2$ of .662 modestly exceeds the baseline H1 value (.538), consistent with +meaningful cross-district slope heterogeneity. Standard errors for the interaction terms +are not reported, as they are unreliable due to near-perfect multicollinearity in the +saturated model (see Data and Methods); point estimates are presented as descriptive +indicators only.
+Table 4. H3 District-Specific Budget → Compliance Slopes (pp per $1M)
+| District | +Estimated Slope | +
|---|---|
| 01 (San Antonio) | +0.09 | +
| 02 (Corpus Christi) | +0.24 | +
| 03 (Houston) | +−0.34 | +
| 04 (Laredo) | +0.28 | +
| 05 (Midland/Abilene) | +0.05 | +
| 06 (Kilgore) | +0.43 | +
| 08 (Midland) | +0.28 | +
| 09 (Wichita Falls) | +0.00 | +
| 10 (Amarillo) | +0.13 | +
| 6E (Kilgore East) | +1.36 | +
| 7B (Abilene) | +0.27 | +
| 7C (Big Spring) | +0.40 | +
| 8A (Lubbock) | +0.19 | +
Note: Slopes are $\hat{\beta}_1 + \hat{\delta}_d$ from the H3 interaction model.
++
H4: Spatial and Geographic Factors¶
The geographic moderation model (Table 5) reveals that offshore-jurisdiction districts +(02, 03, 04) exhibit compliance rates approximately 7.6 percentage points higher than +non-offshore districts on average, net of budget ($\hat{\beta} = 7.61$, SE = 3.29, +$z = 2.31$, $p = .02$). Border-proximate districts similarly show elevated baseline +compliance rates (+6.03 pp, SE = 2.84, $z = 2.12$, $p = .03$). These level effects may +reflect the heightened external scrutiny — from federal regulators, environmental +organizations, and media — that offshore and border districts attract, which could +independently drive compliance investments by operators regardless of RRC budget levels.
+The budget–compliance slope, however, does not differ significantly between offshore +and non-offshore districts ($\hat{\beta}_4 = -0.03$, $p = .87$), nor between border +and non-border districts at conventional thresholds ($\hat{\beta}_5 = -0.25$, $p = .08$), +suggesting that geographic classification affects the level of compliance performance +but not the degree to which additional budget translates into compliance gains.
+Moran's $I$ computed on district-level residuals from the H1 compliance model is +$I = -0.051$, indicating slight spatial dispersion but no statistically significant +spatial autocorrelation. This finding is consistent with prior district-level analysis +of this regulatory system and suggests that unmodeled geographic spillovers are not a +material source of omitted variable bias in the panel models.
+Table 5. H4 Regression Results: Geographic Moderation (DV: Compliance Rate)
+| Term | +$\hat{\beta}$ | +SE | +$z$ | +$p$ | +
|---|---|---|---|---|
| Budget ($M) | +0.35 | +0.15 | +2.39 | +.02 | +
| Offshore (= 1) | +7.61 | +3.29 | +2.31 | +.02 | +
| Border (= 1) | +6.03 | +2.84 | +2.12 | +.03 | +
| Budget × Offshore | +−0.03 | +0.18 | +−0.16 | +.87 | +
| Budget × Border | +−0.25 | +0.15 | +−1.74 | +.08 | +
Note: District fixed effects included. SE clustered at district. $R^2 = .553$, +Adj. $R^2 = .476$. $N = 104$. Moran's $I$ on H1 compliance residuals = −0.051 (no +significant spatial autocorrelation).
++
Summary¶
Taken together, the results offer moderate support for a resource-capacity model of +regulatory performance. Higher OGI appropriations are reliably associated with greater +inspection volume, higher compliance rates, and faster violation resolution — though +identification rests on temporal variation in statewide appropriations rather than +quasi-experimental assignment, and the modest panel length limits statistical precision. +Goal ambiguity moderation operates through a diminishing-returns mechanism: compliance +gains from additional budget are smaller in years when the inspection mandate receives +a larger share of combined appropriations, consistent with resource saturation rather +than amplification. District heterogeneity in budget–outcome slopes is substantial in +descriptive terms but cannot be precisely estimated with the available data. Finally, +geographic context — offshore jurisdiction and border proximity — predicts compliance +levels but not budget sensitivity, and spatial autocorrelation diagnostics provide no +evidence of unmodeled geographic spillover processes.
+Robustness Checks¶
Wild cluster bootstrap. With only $G = 13$ district clusters, asymptotic +cluster-robust standard errors may substantially understate true uncertainty. +Wild cluster bootstrap inference (Rademacher weights, $B = 999$ draws; Cameron, +Gelbach & Miller 2008) yields bootstrap p-values near 0.49–0.51 for all three +H1 outcomes: total inspections ($p_{boot} = 0.494$), compliance rate +($p_{boot} = 0.473$), and resolution rate ($p_{boot} = 0.509$). These are far +from any conventional significance threshold, in stark contrast to the asymptotic +p-values of 0.002, 0.021, and 0.001. The divergence indicates that with $G = 13$ +clusters, asymptotic inference significantly overstates precision. The H1 point +estimates remain positive and directionally consistent, but the results do not +survive bootstrap-based inference. This is the principal inferential limitation +of the study.
+Table 7. Wild Cluster Bootstrap vs. Asymptotic p-values (H1 Models, B = 999)
+| Outcome | +$t$-statistic | +$p$ (asymptotic) | +$p$ (bootstrap) | +
|---|---|---|---|
| Total inspections | +3.13 | +.002 | +.494 | +
| Compliance rate | +2.31 | +.021 | +.473 | +
| Resolution rate | +3.28 | +.001 | +.509 | +
Note: Bootstrap p-values based on 999 Rademacher wild cluster bootstrap draws. +Small number of clusters (G = 13) renders asymptotic inference unreliable.
+Distributed lag model. The distributed lag models test whether budget effects +operate with a one-year delay consistent with a hiring-and-deployment mechanism. +For compliance rate, the lagged budget alone is not significant +($\hat{\beta}_{t-1} = 0.10$, $p = .44$; Model A, N = 91), and in the combined +model the contemporaneous term remains marginally significant +($\hat{\beta}_t = 0.24$, $p = .04$) while the lagged term is negative and +non-significant ($\hat{\beta}_{t-1} = -0.14$, $p = .12$; Model B). For violation +resolution rate, the lagged budget is marginally significant when estimated alone +($\hat{\beta}_{t-1} = 0.83$, $p = .09$; Model A), but neither term reaches +conventional significance in the combined model ($p = .22$ and $p = .14$).
+These findings provide little support for a delayed implementation mechanism. +The persistence of contemporaneous effects alongside non-significant lagged terms +is more consistent with an immediate budget–output relationship. However, the +N = 91 sample offers limited power to disentangle contemporaneous and lagged +effects that are highly collinear over an eight-year window.
+Table 8. Distributed Lag Results (2017–2023, N = 91)
+| Model | +DV | +$\hat{\beta}_t$ | +$p$ | +$\hat{\beta}_{t-1}$ | +$p$ | +$R^2$ | +
|---|---|---|---|---|---|---|
| A — Lag only | +Compliance rate | +— | +— | +0.10 | +.44 | +.543 | +
| B — Both | +Compliance rate | +0.24 | +.04 | +−0.14 | +.12 | +.569 | +
| A — Lag only | +Resolution rate | +— | +— | +0.83 | +.09 | +.696 | +
| B — Both | +Resolution rate | +0.24 | +.22 | +0.59 | +.14 | +.698 | +
Note: District fixed effects included; SE clustered at district.
+Robustness Checks¶
Two checks address limitations of the baseline H1 models.
+Wild cluster bootstrap re-tests H1 with valid small-sample inference rather than +asymptotic cluster-robust standard errors. With $G = 13$ clusters, asymptotic results +can overstate precision. Rademacher wild cluster bootstrap ($B = 999$ draws; Cameron, +Gelbach & Miller 2008) yields p-values near 0.49–0.51 for all three H1 outcomes — +far from any conventional threshold — indicating that the asymptotic H1 results do +not survive this correction. Point estimates remain positive and substantively consistent +in direction, but the study lacks the cluster count required to establish significance +through bootstrap inference.
+Distributed lag model relaxes the assumption that budget effects are instantaneous. +A one-year lag of OGI budget is estimated alone (Model A) and jointly with the +contemporaneous term (Model B), over the 2017–2023 sample (N = 91). The lagged budget +is not independently significant for compliance rate ($p = .44$) and only marginally so +for resolution rate ($p = .09$). In the combined models, contemporaneous effects persist +while lagged terms do not attain significance — providing little evidence that a delayed +mechanism dominates an immediate one.
+# Wild cluster bootstrap (Rademacher weights, B=999)
+# For each draw: multiply each cluster's residuals by ±1, re-fit, record t-stat.
+# p-value = share of |t*| >= |t_observed|.
+
+def wild_cluster_bootstrap(model, data, dv, cluster_col="district",
+ coef="ogi_budget_m", B=999, seed=42):
+ rng = np.random.default_rng(seed)
+ groups = data[cluster_col].values
+ unique_groups = np.unique(groups)
+ t_obs = model.tvalues[coef]
+ yhat = model.fittedvalues.values
+ ehat = model.resid.values
+
+ t_boot = np.empty(B)
+ for b in range(B):
+ # One Rademacher weight per cluster, broadcast to observations
+ cw = {g: rng.choice([-1.0, 1.0]) for g in unique_groups}
+ w = np.array([cw[g] for g in groups])
+ df_b = data.copy()
+ df_b[dv] = yhat + ehat * w
+ m_b = smf.ols(
+ f"{dv} ~ {coef} + C({cluster_col})", data=df_b
+ ).fit(cov_type="cluster", cov_kwds={"groups": df_b[cluster_col]})
+ t_boot[b] = m_b.tvalues.get(coef, np.nan)
+
+ p_boot = float((np.abs(t_boot) >= np.abs(t_obs)).mean())
+ return t_obs, float(model.pvalues[coef]), p_boot
+
+print("Wild Cluster Bootstrap — H1 Models (B = 999 Rademacher draws)")
+print(f"{'Outcome':<28} {'t-stat':>7} {'p asymptotic':>13} {'p bootstrap':>12}")
+print("─" * 65)
+
+for dv, model in [
+ ("total_inspections", m_inspections),
+ ("compliance_rate", m_compliance),
+ ("resolution_rate", m_resolution),
+]:
+ t, p_a, p_b = wild_cluster_bootstrap(model, actuals, dv)
+ sig_a = "*" * (1 + (p_a < .05) + (p_a < .01))
+ sig_b = "*" * (1 + (p_b < .05) + (p_b < .01))
+ print(f"{dv:<28} {t:>7.3f} {p_a:>12.3f}{sig_a:<3} {p_b:>10.3f}{sig_b:<3}")
+
+print("\n* p<.10 ** p<.05 *** p<.01")
+Wild Cluster Bootstrap — H1 Models (B = 999 Rademacher draws) +Outcome t-stat p asymptotic p bootstrap +───────────────────────────────────────────────────────────────── +total_inspections 3.128 0.002*** 0.494* +compliance_rate 2.307 0.021** 0.473* +resolution_rate 3.277 0.001*** 0.509* + +* p<.10 ** p<.05 *** p<.01 ++
# Distributed lag: 1-year lag of OGI budget (shift within district).
+# Lag is NaN for 2016 (no 2015 data), so regression sample is 2017-2023 (N=91).
+
+panel_lag = panel.copy()
+panel_lag["ogi_budget_m_lag1"] = (
+ panel_lag.sort_values("year")
+ .groupby("district")["ogi_budget_m"]
+ .shift(1)
+)
+
+lag_actuals = panel_lag[
+ (panel_lag["is_budget_year"] == 0) &
+ (panel_lag["ogi_budget_m_lag1"].notna())
+].copy()
+
+print(f"Distributed lag sample: {len(lag_actuals)} obs | "
+ f"years {lag_actuals['year'].min()}–{lag_actuals['year'].max()}")
+
+# ── Model A: lagged budget only ───────────────────────────────────────────────
+m_lag_only = smf.ols(
+ "compliance_rate ~ ogi_budget_m_lag1 + C(district)", data=lag_actuals
+).fit(cov_type="cluster", cov_kwds={"groups": lag_actuals["district"]})
+
+# ── Model B: contemporaneous + 1-year lag ────────────────────────────────────
+m_lag_both = smf.ols(
+ "compliance_rate ~ ogi_budget_m + ogi_budget_m_lag1 + C(district)",
+ data=lag_actuals
+).fit(cov_type="cluster", cov_kwds={"groups": lag_actuals["district"]})
+
+# ── Also run for resolution rate ──────────────────────────────────────────────
+m_lag_res_only = smf.ols(
+ "resolution_rate ~ ogi_budget_m_lag1 + C(district)", data=lag_actuals
+).fit(cov_type="cluster", cov_kwds={"groups": lag_actuals["district"]})
+
+m_lag_res_both = smf.ols(
+ "resolution_rate ~ ogi_budget_m + ogi_budget_m_lag1 + C(district)",
+ data=lag_actuals
+).fit(cov_type="cluster", cov_kwds={"groups": lag_actuals["district"]})
+
+print("\n── Compliance Rate ───────────────────────────────────────────")
+
+print("\nModel A — Lagged budget only (t−1):")
+print(m_lag_only.summary2().tables[1][display_cols].loc[["ogi_budget_m_lag1"]])
+print(f" R² = {m_lag_only.rsquared:.3f} Adj. R² = {m_lag_only.rsquared_adj:.3f}")
+
+print("\nModel B — Contemporaneous + 1-year lag:")
+print(m_lag_both.summary2().tables[1][display_cols].loc[
+ ["ogi_budget_m", "ogi_budget_m_lag1"]
+])
+print(f" R² = {m_lag_both.rsquared:.3f} Adj. R² = {m_lag_both.rsquared_adj:.3f}")
+
+print("\n── Resolution Rate ───────────────────────────────────────────")
+
+print("\nModel A — Lagged budget only (t−1):")
+print(m_lag_res_only.summary2().tables[1][display_cols].loc[["ogi_budget_m_lag1"]])
+print(f" R² = {m_lag_res_only.rsquared:.3f} Adj. R² = {m_lag_res_only.rsquared_adj:.3f}")
+
+print("\nModel B — Contemporaneous + 1-year lag:")
+print(m_lag_res_both.summary2().tables[1][display_cols].loc[
+ ["ogi_budget_m", "ogi_budget_m_lag1"]
+])
+print(f" R² = {m_lag_res_both.rsquared:.3f} Adj. R² = {m_lag_res_both.rsquared_adj:.3f}")
+Distributed lag sample: 91 obs | years 2017–2023 + +── Compliance Rate ─────────────────────────────────────────── + +Model A — Lagged budget only (t−1): + Coef. Std.Err. z P>|z| +ogi_budget_m_lag1 0.10 0.13 0.77 0.44 + R² = 0.543 Adj. R² = 0.466 + +Model B — Contemporaneous + 1-year lag: + Coef. Std.Err. z P>|z| +ogi_budget_m 0.24 0.11 2.08 0.04 +ogi_budget_m_lag1 -0.14 0.09 -1.55 0.12 + R² = 0.569 Adj. R² = 0.490 + +── Resolution Rate ─────────────────────────────────────────── + +Model A — Lagged budget only (t−1): + Coef. Std.Err. z P>|z| +ogi_budget_m_lag1 0.83 0.49 1.69 0.09 + R² = 0.696 Adj. R² = 0.644 + +Model B — Contemporaneous + 1-year lag: + Coef. Std.Err. z P>|z| +ogi_budget_m 0.24 0.19 1.22 0.22 +ogi_budget_m_lag1 0.59 0.40 1.46 0.14 + R² = 0.698 Adj. R² = 0.642 ++
Hypotheses Summary¶
Table 6. Summary of Hypotheses, Predictions, Findings, and Empirical Support
+| # | +Hypothesis | +Prediction | +Key Result | +Support | +
|---|---|---|---|---|
| H1a | +Capacity → Inspection volume | +Higher OGI budget predicts more inspections per district | +β = 666.3/$1M (z = 3.13, p < .01); bootstrap p = .494 | +✓† | +
| H1b | +Capacity → Compliance | +Higher OGI budget predicts higher district compliance rate | +β = 0.26 pp/$1M (z = 2.31, p = .02); bootstrap p = .473 | +✓† | +
| H1c | +Capacity → Resolution | +Higher OGI budget predicts higher violation resolution rate | +β = 1.05 pp/$1M (z = 3.28, p < .01); bootstrap p = .509 | +✓† | +
| H2a | +Goal ambiguity moderates capacity → compliance | +Clearer inspection focus amplifies budget effect | +Significant but negative (β = −6.53, z = −3.55, p < .01); interpretation constrained by time-only variation in budget share (range: 0.59–0.67) | +Exploratory‡ | +
| H2b | +Goal ambiguity moderates capacity → resolution | +Clearer inspection focus amplifies budget effect | +Interaction not significant (p = .24) | +✗ | +
| H3 | +District heterogeneity in budget slopes | +Budget → compliance slope varies across districts | +Slopes from −0.34 pp/$1M (D03) to +1.36 pp/$1M (D6E); inference unreliable | +Descriptive§ | +
| H4a | +Offshore jurisdiction moderates budget effect | +Offshore districts show different budget → compliance slope | +Level effect: +7.6 pp (p = .02); slope interaction not significant (p = .87) | +Partial¶ | +
| H4b | +Border proximity moderates budget effect | +Border districts show different budget → compliance slope | +Level effect: +6.0 pp (p = .03); slope interaction marginal (p = .08) | +Partial¶ | +
| H4c | +Spatial autocorrelation in residuals | +Geographic spillovers produce clustered residuals | +Moran's I = −0.051; no significant spatial autocorrelation | +✗ | +
Notes:
+† H1 point estimates are positive and directionally consistent across all three outcomes, +supporting the capacity hypothesis substantively. However, wild cluster bootstrap +inference (B = 999 Rademacher draws) yields p-values near 0.49–0.51 for all outcomes, +indicating that asymptotic cluster-robust standard errors substantially overstate precision +with G = 13 clusters. H1 findings should be interpreted as suggestive rather than +statistically definitive. Distributed lag models (2017–2023, N = 91) show contemporaneous +effects persist while lagged terms do not reach significance, providing no clear evidence +for a delayed implementation mechanism.
+‡ H2a is statistically significant but the identification is weak: inspection +budget share varies only over time (like the budget itself), with a range of just +0.59–0.67 across 8 years. The negative interaction is consistent with a resource +saturation effect but cannot be distinguished from year-specific confounders. +At mean share (≈ 0.62), the implied marginal budget effect is ≈ 0.15 pp per $1M. +H2b not significant for resolution rate. Both H2 findings are best treated as +exploratory patterns for future research.
+§ H3 interaction standard errors are unreliable (near-perfect multicollinearity in +the saturated model); budget slopes are reported as descriptive point estimates only.
+¶ Geographic classification predicts compliance levels but not budget sensitivity. +Offshore and border districts exhibit systematically higher compliance regardless of +annual budget variation.
+Regression sample: N = 104 (13 districts × 8 years, 2016–2023). All models include +district fixed effects; standard errors clustered at the district level (G = 13). +Robustness sample: N = 91 (2017–2023, distributed lag models).
+