Falling resource revenue sharply cuts capital investment in Indonesia’s extractive districts — a 10% drop in revenue sharing reduces district capital expenditure by about 4.8% short-run and 7.8% long-run — while mining contractions raise unemployment, much of which is hidden by informal-sector absorption.

Main Finding

A 10% decline in natural-resource revenue sharing is associated with a 4.78% reduction in district capital expenditure (95% CI −5.62 to −3.94; p < 0.001), rising to a 7.76% long-run contraction. Capital spending absorbs a disproportionately large share of revenue shocks (35.4% of the shock) despite constituting only 18.2% of district spending. Separately, mining-output contractions are associated with higher open unemployment (reported coefficient −0.624; p < 0.001), an effect amplified by local extractive employment concentration and attenuated by workforce educational attainment — but measured unemployment captures only a small fraction of displaced workers (an estimated 17.8% would show up as unemployed under an 80% output decline). Authors conclude fiscal exposure is severe and measurable; labour exposure is severe but largely hidden.

Key Points

• Sample and scope: Balanced panel of 118 extractive-producing Indonesian districts (coal, oil, gas, metallic minerals), observed annually 2015–2024 (1,180 district-year observations). Producing districts are taken from the national 514-district budget frame.
• Main hypotheses tested:
  • H1: Revenue-sharing contractions → capital-expenditure contractions (supported).
  • H2: Own-source revenue and general transfers buffer but do not fully substitute for resource transfers (supported).
  • H3: Mining value-added contractions → higher open unemployment (supported).
  • H4: Extractive employment concentration amplifies unemployment response; higher education attenuates it (supported).
  • H5: Substantial state dependence in budgets and labour outcomes; long-run responses exceed short-run ones (supported).
  • H6: Fiscal dependency and labour readiness form separable dimensions (used descriptively to create a 4-quadrant readiness typology).
• Quantitative magnitudes:
  • 10% fall in revenue sharing → −4.78% capital-expenditure change short-run; −7.76% long-run.
  • Capital expenditure absorbs ~35.4% of a revenue shock though it is only 18.2% of spending.
  • Mining-output contraction linked to higher unemployment (estimated coefficient −0.624); measurement gap: large share of displaced employment may move into informal work and not be recorded as unemployment (only ~17.8% visible in the unemployment rate under extreme contraction).
• Policy recommendation: design transition transfers indexed to capital-expenditure incidence and improve district-level measurement/instrumentation of the informal labour margin.

Data & Methods

• Data sources:
  • Ministry of Finance (fiscal-transfer directorate): realised district revenue and expenditure outturns (oil-and-gas + mineral-and-coal revenue sharing, general allocation grant, own-source revenue, capital and personnel expenditure).
  • BPS (Statistics Indonesia): National Labour Force Survey (SAKERNAS) district-level aggregates (open unemployment rate, sectoral employment, educational attainment), and GRDP (constant 2010 prices).
  • Ministry of Energy and Mineral Resources: district-level commodity production and royalties (used to define producing districts).
• Variables (selected):
  • ln(CapEx): log of realised capital expenditure (constant prices) — outcome eq.1.
  • ln(DBH SDA): log of natural-resource revenue-sharing transfers — focal predictor eq.1.
  • Open unemployment rate — outcome eq.2.
  • ln(mining GRDP): log of mining & quarrying GRDP — focal predictor eq.2.
  • ln(PAD): log own-source revenue; ln(DAU): log general allocation grant; mining location quotient; upper-secondary workforce share; lagged outcomes for state dependence.
• Empirical strategy:
  • Two specifications: (a) two-way fixed-effects models with Driscoll–Kraay standard errors to account for cross-sectional dependence and serial correlation; (b) two-step System GMM (Arellano–Bover/Blundell–Bond) with Windmeijer finite-sample correction for endogenous dynamics and state dependence.
  • Authors compute short-run elasticities, long-run multipliers (accounting for lagged dependent variables), precision estimates, budget-incidence calculations (what share of revenue shock is borne by capital expenditure), and a four-quadrant district readiness typology.
• Design caveats and limitations (reported by authors):
  • The analysis reports within-district associations conditional on district and year fixed effects; it is explicitly not presented as a causal treatment effect.
  • Within-equation measurement dependence: all fiscal series originate from the same audited ledger (so correlated measurement error or accounting restatements could mechanically move regressors and outcome together); similarly, labour variables are from the same household survey (correlated sampling errors).
  • Panel is balanced and confined to producing districts; comparator (non-producing) group retained but not analyzed in published estimates.
  • Some archival gaps: rationale for sample window bounds (2015–2024) and handling of district boundary changes are not documented.
  • Survey precision varies across small districts; unemployment undercounts informal absorption of job losses.

Implications for AI Economics

• High-resolution policy modeling needs administrative + survey linkages: This study shows the value of integrating audited fiscal ledgers, production/royalty records, and household labour surveys at subnational resolution to quantify transition exposure. AI/ML models for policy should combine such heterogeneous administrative sources to predict local fiscal and labour impacts of decarbonisation or other structural shocks.
• Dynamic heterogeneity matters for forecasting: Strong state dependence in both fiscal and labour outcomes implies that forecasting and counterfactual simulation must model lagged dynamics explicitly (e.g., dynamic panel structures or sequence models). AI economists should incorporate temporal dependence (and corrected standard errors) when training models on panel data to avoid underestimating long-run impacts.
• Measurement gaps (informal sector) are crucial for model targets: A large fraction of displaced workers may migrate to informal employment and thus not appear in measured unemployment. For ML-driven policy targeting, augmenting survey data with alternative signals (mobile-phone mobility, nightlights, tax registries, transaction data) can help infer hidden labour-market transitions and improve targeting of retraining or cash transfers.
• Methodological caution for ML/causal work:
  • Beware common-source error: when predictors and outcomes come from the same administrative ledger or survey, correlations may reflect measurement mechanics rather than behaviour. Combine independent data sources or use techniques robust to shared measurement error.
  • Non-causal elasticities can still be policy-useful if framed correctly: the paper provides within-district elasticities (not causal estimates) that are actionable for sizing transfers and simulating budgets — AI models should likewise distinguish between predictive associations and causal effects, and incorporate domain constraints when generating policy recommendations.
• Policy-optimization and targeting applications:
  • Transfer design: authors recommend transfers indexed to capital-expenditure incidence. AI/optimization models could operationalize this by calculating district-specific transfer formulas using estimated elasticities, fiscal composition, and readiness typology — with explainability constraints for policymakers.
  • Readiness typology as a targeting input: the four-quadrant district classification (fiscal dependency vs. labour readiness) can be used as features in prioritization algorithms for conditional grants, retraining programs, or investment in diversification; combine with cost-effectiveness ML to allocate scarce transition finance.
• Research opportunities for AI economists:
  • Build improved proxies for informal absorption of job losses and incorporate them into micro-founded macro models of local adjustment.
  • Use heterogenous treatment-effect frameworks (causal forests, double/debiased ML) to estimate which district characteristics predict better outcomes under different transfer designs — to move from descriptive elasticity to conditional causal calibration.
  • Validate and extend findings with external high-frequency indicators (satellite, payments, web-scraped vacancy postings) to capture short-run labour and fiscal responses, enabling near-real-time monitoring of transition vulnerability.
  • Simulate decarbonisation scenarios combining dynamic panel estimates with agent-based or general-equilibrium models to quantify spillovers across districts and sectors, and to evaluate optimal sequencing of fiscal and labour instruments.
• Practical note for modelers: when training predictive or prescriptive models on similar panels, replicate techniques used here for inference on state dependence (include lagged outcomes), correct for cross-sectional dependence (robust SEs or appropriate cross-validation that respects clustering), and be explicit about whether your outputs are predictive associations or causal policy impacts.

Summary takeaway: The paper quantifies large, asymmetric fiscal exposure to resource-revenue declines at the district level and reveals substantial but partially hidden labour exposure due to informal absorption. For AI economics, this underscores the necessity of multi-source, dynamic models that handle measurement issues, explicitly model long-run state dependence, and produce explainable allocations for place-based transition policy.