Regional Disparities and Spillovers in Human Development: A Spatial Panel Analysis of Decentralised Indonesia

Averaged across Indonesia’s provinces, the Human Development Index (HDI) rose from 65.6 in 2010 to 73.6 in 2024, yet provincial disparities remain stark: Papua’s average over that period, 60.8, sits nearly nine points below the 33-province mean, while Jakarta and Yogyakarta have both passed 80 in recent years. Closing this within-country gap is a recurring commitment of Indonesia’s fiscal decentralisation policy and is central to its sustainable development agenda, which links human development to health, education, and reduced inequality [1]. Two very different stories could explain the narrowing: progress in one province might genuinely spread to its neighbours through rising income or public investment, or every province might simply be riding the same national wave. Distinguishing between them bears directly on how decentralisation and place-based development policy should be evaluated.

A substantial literature has examined the correlates of provincial and district-level human development in Indonesia, typically treating each region as independent of its neighbours [2-5]. A related strand studies how fiscal decentralisation shapes regional convergence using spatial methods, but at the district level and with a focus on economic growth rather than HDI directly [6, 7]. A smaller body looks explicitly at spatial dependence in HDI, though only at the single-province level and only as conference proceedings, both focused on Central Java rather than the full national panel [8, 9]. One question remains largely unexamined: whether human development itself genuinely spreads across provincial borders at the national level, once the shared national trend, which moves every province in the same direction and mechanically makes them look alike, is separated from spillovers proper.

This paper studies regional disparities and spillovers in human development for a balanced panel of all 33 Indonesian provinces from 2010 to 2024. We define each province’s neighbours by its three geographically nearest provincial centroids rather than shared land borders, a choice that suits Indonesia’s archipelagic geography, where six of the 33 provinces have no land neighbour at all. We also allow for unobserved, lasting differences between provinces. To separate the shared national trend from genuine cross-province relationships, we compare each province, in each year, to that year’s national average, letting us ask whether provinces are converging toward one another or merely moving in parallel.

Once the national trend is removed, the estimated spillover through HDI itself is small, negatively signed, and statistically indistinguishable from zero under both asymptotic and block-bootstrap inference, evidence against a direct HDI-to-HDI diffusion channel. What spreads more reliably is economic activity: a rise in a province’s own economic output is associated with a positive spillover to neighbouring provinces’ HDI that survives bootstrap resampling and remains significant, though unstable in magnitude, across alternative neighbour definitions. Own-province health expenditure tells a different story, carrying a robust negative association with own-province HDI. This is more consistent with compensatory allocation of health funds toward lagging provinces than with an adverse effect of spending: provinces with lower human development appear to receive larger remedial transfers.

We take particular care to distinguish genuine regional relationships from artefacts of the shared national trend. Before settling on our preferred approach, we estimated the same relationships without any trend adjustment and with calendar-year controls. The apparent spillover through HDI collapsed from strongly significant, when left uncorrected, to indistinguishable from zero under either trend-control strategy, a pattern unlikely to arise from noise. Two further checks support this reading: an independent reproduction using an unrelated statistical package matches our primary results closely, and a cross-check against an alternative estimator finds full agreement in the direction of all twelve shared relationships. We also subject every result to a resampling-based robustness check, flagging as fragile any finding that loses significance; the direct relationship between a province’s own economic output and its own HDI is one such case, which we do not treat as robust.

This paper contributes to sustainable regional development in Indonesia in three ways. First, we add an explicitly geographic dimension to the literature on determinants of Indonesian human development [2-5], which has generally treated provinces as independent, clarifying whether a province’s development is shaped only by its own conditions or also by its neighbours’. Second, we extend the spatial approach to the full 33-province national panel, with an explicit decomposition of direct and indirect effects on HDI, so that policymakers can distinguish which provinces benefit mainly from their own investments and which stand to gain from their neighbours’ growth, a distinction that matters for how decentralised spending is targeted; this builds on both the single-province HDI studies [8, 9] and the district-level convergence studies focused on economic growth [6, 7]. Third, and most consequential for policy, we show that Indonesia’s apparent convergence in human development across neighbouring provinces reflects a shared national trajectory rather than genuine cross-provincial diffusion: place-based policies presuming that HDI gains spread automatically to neighbours are not supported by the evidence, whereas regional economic integration offers a more credible channel through which one province’s progress can benefit its neighbours.

3.1 Data

Our data form a balanced panel of Indonesia’s 33 provinces observed annually from 2010 to 2024 (T = 15), yielding 495 province-year observations. The dependent variable is the provincial HDI, known domestically as the Indeks Pembangunan Manusia, a composite indicator of longevity, education, and standard of living compiled by Indonesia’s national statistics agency (BPS) following United Nations Development Programme (UNDP) methodology. Explanatory variables are the provincial Gini ratio, a measure of income inequality, and the natural logarithms of provincial government expenditure in four functional categories: economic affairs, health, education, and social protection, labelled in the tables ln(Economic affairs), ln(Health), ln(Education), and ln(Social protection). We additionally control for the natural log of provincial gross regional domestic product, ln(GRDP), the principal measure of a province’s overall resource base, plausibly correlated with both expenditure levels and human development outcomes.

Indonesia’s official province count rose from 33 to 38 during our sample period, with North Kalimantan splitting from East Kalimantan in 2012 and four new Papua provinces (South Papua, Central Papua, Highland Papua, and Southwest Papua) created in 2022–2023. Our panel nonetheless holds a fixed 33-province geography throughout 2010–2024 because the underlying BPS-derived series reports HDI and expenditure at the original 33-province level for its entire published history, folding the newer provinces’ figures back into their pre-split parents rather than reporting them separately. We inherit this boundary convention from the source series rather than re-aggregating independently, and match our spatial weight matrix to the same 33-province geography (Section 3.2). Expenditure variables are nominal rather than inflation- or population-adjusted; Section 5 discusses the implications.

No observations are lost to missing data: all 495 province-year cells have non-missing values on every variable, so the analytic sample equals the raw sample. All 33 provinces were also successfully matched to their corresponding administrative boundaries in the shapefile used to construct the spatial weight matrix (Section 3.2).

Table 1. Descriptive statistics

Table 1 shows substantial cross-provincial dispersion in HDI (range 54.5–83.1) alongside a comparatively narrow range in the Gini ratio (0.24–0.46), consistent with Indonesia’s pattern of large level differences in development outcomes across regions coexisting with more modest differences in within-province inequality. Figure 1 maps this dispersion, averaged over 2010–2024. The highest averages are Jakarta (79.7), Yogyakarta (78.6), East Kalimantan, Kepulauan Riau, Bali, Riau, and North Sulawesi; the lowest are Papua (60.8), West Papua, East Nusa Tenggara, West Sulawesi, West Kalimantan, West Nusa Tenggara, North Maluku, and Gorontalo.

1.png

Figure 1. Average Human Development Index (HDI) by province, 2010–2024

3.2 Spatial weight matrix

We construct the spatial weight matrix W from provincial centroids using k-nearest-neighbour (KNN) relations with k = 3, row-standardised. We prefer KNN to contiguity-based weights because Indonesia’s archipelagic geography leaves many provinces without a shared land border: under queen contiguity, six of the 33 provinces have no contiguous neighbour and the structure decomposes into eleven disconnected sub-graphs, an expected outcome where administrative adjacency does not track physical adjacency, rendering contiguity unsuitable as the primary weight matrix. We use KNN matrices with k = 5 and k = 8 for sensitivity analysis (Section 4.5). Scale matters: spatial decomposition of provincial sustainability indicators is highly sensitive to the administrative scale at which the underlying indicator is aggregated, a scale-mismatch problem documented most recently for Vietnam’s provincial indicator system [29].

3.3 Model specification

Answering whether a province’s human development is shaped by its neighbours, and not only its own circumstances, requires a model that lets both channels operate at once: a province’s HDI may respond directly to a neighbour’s HDI, and separately to a neighbour’s income, inequality, or public spending. We estimate a spatial Durbin model (SDM), which nests both the spatial autoregressive (SAR) and SEM specifications as special cases and allows spillovers through both the outcome and the explanatory variables [30, 31], consistent with broader developments in spatial panel data modelling [32]:

$\begin{aligned} & \mathrm{HDI}_{i t}=\alpha+\rho \sum_j W_{i j} \mathrm{HDI}_{j t}+\beta^{\prime X_{i t}} \, +\theta^{\prime} \sum_j W_{i j} X_{j t}+\mu_i+\varepsilon_{i t}\end{aligned}$           (1)

where, i indexes provinces (i = 1, …, 33), t indexes years (t = 2010, …, 2024), X_it is the vector of six covariates described in Section 3.1, W is the row-standardised KNN spatial weight matrix, ρ is the SAR coefficient, θ is the vector of spatial-lag (WX) coefficients, and μ_i is a province-specific random effect. The parameter ρ captures dependence through the outcome: a shock to HDI in one province propagates mechanically to its neighbours under the model's structure, while θ captures dependence through observable covariates. We do not interpret β and θ as marginal effects directly; instead, we report the decomposed direct, indirect, and total effects implied by the spatial multiplier (I − ρW)⁻¹, obtained via parametric Monte Carlo simulation (R = 1,000 draws) from the asymptotic covariance matrix of the estimated parameters [28]. The SDM is not specific to human-development applications: the general case for this specification, and for basing inference on indirect effects rather than raw coefficients, is well established in the wider applied spatial econometrics literature [33], and the specification suits any setting in which an outcome may respond to both a region's own covariates and its neighbours'.

3.4 Time-effect control

A two-way random-effects specification, absorbing both province- and year-specific heterogeneity, is the natural benchmark for a panel of this structure, but the maximum-likelihood spatial estimator we use (splm::spml in R) does not support two-way random effects for spatial models [34]. We therefore control for common national trends through cross-sectional yearly demeaning: before estimation, we subtract from HDI, from each covariate, and from each covariate’s spatial lag the cross-sectional (33-province) mean of that variable in the same year. For a row-standardised weight matrix, this operation commutes with spatial lagging, so demeaning before or after constructing WX yields identical results. It is algebraically equivalent to sweeping out a deterministic time fixed effect, a standard device when software does not support a formal two-way specification [35]. Separating a genuine common factor, such as a shared national trend, from cross-sectional dependence that reflects real unit-to-unit relationships is a recognised identification problem; conflating the two generates spurious evidence of dependence [36]. A fully interactive fixed-effects estimator would model such a factor directly rather than remove it by demeaning [37], but none is currently available for the specification required here. We therefore treat cross-sectional demeaning as an approximation rather than a formally equivalent two-way estimator, and return to this limitation in Section 5.

Cross-sectional demeaning carries an econometric cost worth making explicit. Because each year’s 33 province-level deviations from the cross-sectional mean sum to zero by construction, the demeaned variables cannot be independent across provinces even when the underlying data are: for N units demeaned each year, the expected pairwise correlation among purely idiosyncratic, mean-zero disturbances is mechanically −1/(N − 1), about −0.03 for our 33 provinces. This mechanical negative dependence is a standard consequence of within-year demeaning, not a peculiarity of our design, but has not previously been discussed in this literature: it could in principle bias the SAR coefficient ρ and the SEM coefficient λ, both estimated as negative, toward negative values. We doubt it is the primary explanation. At roughly −0.03, it is well under half the magnitude of the SDM (RE) estimate of ρ (−0.079) and about a tenth of the spatial Durbin error model (SDEM) (FE) estimate of λ (−0.320), so it cannot account for either coefficient’s full size even if it contributes to the sign. We nonetheless flag this as an open limitation rather than treat the negative sign of ρ and λ as fully settled; Section 5 returns to it.

3.5 Model selection and estimator choice

We select among SAR, SEM, and SDM specifications using robust Lagrange multiplier (LM) tests for spatial lag and SEM dependence [38]. We further test the common-factor restriction implied by the SEM specification, θ = −ρβ, via a likelihood-ratio test comparing the SDM and SEM log-likelihoods; rejection of this restriction indicates that the SDM cannot be simplified to a SEM [39].

We choose between random- and fixed-effects estimators using the spatial Hausman test, implemented separately for the spatial-lag and spatial-error branches [40]. Once cross-sectional demeaning is applied, the test rejects random effects for both branches, which would ordinarily favour fixed effects. We nonetheless report our preferred SDM under random effects, for a reason specific to the spatial-lag branch rather than a disregard of the Hausman result. Within-transformation maximum-likelihood estimation of a SAR parameter is asymptotically biased of order O(1/T) in panels with a fixed time dimension, an incidental-parameters problem that does not afflict the spatial-error specification, whose fixed-effects estimator remains consistent under the same asymptotics [41]. Because splm implements no bias-corrected estimator, a fixed-effects estimate of the spatial-lag branch would substitute a known small-T bias in ρ for the orthogonality violation the Hausman test detects, rather than resolving it. We therefore retain random effects for the SDM, and report the fixed-effects SDEM as a cross-check on the error branch, where no such asymmetry applies. We do not compare information criteria across the two, whose likelihoods are non-comparable: the fixed-effects estimator profiles out individual effects without counting them as parameters, while the random-effects estimator integrates over them via estimated variance components [42].

Rather than resting this choice on the incidental-parameters argument alone, we also estimate the naive fixed-effects SDM for the spatial-lag branch and compare it with our preferred random-effects specification (Table A1), so readers can judge the difference for themselves [41]. One specification issue deserves note: variance inflation factors on the spatially lagged regressors reach 9.1 for W × ln(Health), 10.5 for W × ln(Social protection), and 11.7 for W × ln(GRDP), so the spatial lags are correlated with one another, as is common when several covariates share the same neighbourhood structure. This inflates standard errors rather than biasing coefficients; we therefore read the individually insignificant spatial lags (W × ln(Health), W × ln(Education), and W × ln(Social protection)) as imprecisely estimated rather than established nulls, while the two we interpret, W × ln(GRDP) and W × ln(Economic affairs), remain significant.

3.6 Inference

Because maximum-likelihood inference in spatial panel models relies on asymptotic normality, a poor approximation with only 33 cross-sectional units, we complement asymptotic standard errors with a non-parametric block-year bootstrap. In each of 2,000 replications, we resample the fifteen years with replacement, rebuild the spatially lagged covariates, and re-estimate the full model; the resulting empirical distribution of each coefficient yields a bootstrap standard error, a continuity-corrected two-sided p-value, and a 95 percent percentile confidence interval, none relying on asymptotic normality. We test residual normality with the Shapiro–Wilk test and, where it is rejected, treat the bootstrap inference as the more reliable basis for judging significance.

We additionally assess robustness to the choice of spatial weight matrix by re-estimating the model with k = 5 and k = 8 nearest neighbours, and to the influence of individual provinces through a leave-one-province-out jackknife, in which the model, including the spatial weight matrix and spatially lagged covariates, is fully reconstructed and re-estimated 33 times, once with each province excluded. Because a jackknife of this kind flags only coefficients that change sign, we additionally re-estimate the full model with the single most influential province excluded and compare coefficients and p-values directly against the full-sample estimates, so that a coefficient that loses statistical support without reversing sign is not overlooked (Section 4.6).

Throughout, we interpret our estimates as associations rather than causal effects. Provincial government expenditure is plausibly endogenous to human development outcomes, both because higher-capacity provincial governments may systematically allocate spending differently and because central government transfers often respond to a province's development status, a form of reverse causality this design cannot rule out. We do not employ an instrumental-variables or GMM strategy for the spatial panel setting, and we discuss this limitation in Section 5.

This paper asked whether Indonesia’s regional disparities in human development are narrowing because progress genuinely spreads across provincial borders, or because every province is simply riding the same national trend. Studying all 33 provinces from 2010 to 2024, we separated the shared trend from real cross-province relationships, a distinction that determines the answer. Once it is drawn, the direct channel from one province’s human development to its neighbours’ is small, negatively signed, and statistically null under both inference methods, evidence against direct cross-province diffusion rather than for a compensatory pattern. What spreads more reliably is economic activity: a province’s economic scale (GRDP) carries a positive spillover to its neighbours’ HDI that is robust in sign and significance, though its magnitude (1.2 to 7.7) depends on how a neighbourhood is defined and on the exclusion of individual provinces. A province’s own health expenditure, by contrast, is negatively associated with its own outcomes, which we attribute to compensatory allocation toward lagging provinces rather than a harmful effect of spending. What looks like convergence among neighbours therefore tracks a shared national trajectory more than genuine diffusion, and place-based policy should not presume that human development gains spread automatically to neighbours through that channel.

These findings apply to Indonesia’s 33 provinces over 2010–2024, a period spanning substantial fiscal decentralisation reform and, toward its end, the COVID-19 pandemic. Whether the weakening of spatial clustering we document for 2022–2024 reflects a durable structural shift or a transient pandemic-recovery artefact cannot be determined from the present data, and is a natural direction for future work as more post-pandemic years become available.

For policy, our results offer a cautious lesson rather than a directive. Because the estimated direct spillover from neighbours’ human development is small, negatively signed, and indistinguishable from zero by both asymptotic and bootstrap standards, policies aimed at raising HDI in lagging provinces should not presume that gains will mechanically diffuse to neighbours through that channel. The one robust spillover we identify operates through economic scale, suggesting that regional economic integration, rather than proximity to high-HDI neighbours per se, is the more promising lever for indirect gains, consistent with Indonesia’s 2025–2029 national development plan, which frames strengthened local economic competitiveness rather than proximity-based diffusion as a principal instrument for narrowing regional disparities [49]. Our health-expenditure finding carries a complementary implication: because the negative association is more plausibly compensatory than harmful, it corroborates rather than undermines the plan’s strategy of directing larger transfers to fiscally weaker regions, and evaluations should not mistake that remedial targeting for evidence that health spending fails to work. We are conservative here because our design identifies associations, not causal effects: provincial expenditure is plausibly endogenous to development outcomes, and we lack a credible instrument or exogenous policy variation to sever this link.

Several caveats apply. First, our time-effect control, cross-sectional yearly demeaning, is a deterministic approximation to a formal two-way random-effects specification and, as Section 3.4 details, mechanically induces a small amount of negative cross-sectional dependence among the demeaned variables; we cannot show it is exactly equivalent to a two-way specification for the quasi-maximum-likelihood spatial estimator we use, and though we doubt this dependence is large enough to explain the negative SAR and SEM coefficients, we cannot rule out that it contributes to their sign. Researchers with software supporting genuine two-way random effects for spatial panels, or an interactive fixed-effects estimator that models the common factor directly [37], should treat our approach as second-best. Second, we cannot rule out reverse causality or omitted confounding in the expenditure–HDI relationship. Third, our jackknife shows Papua materially influences several coefficients; we retain it because excluding an outlier would itself bias inference about the disparities the paper studies. Fourth, we do not conduct an out-of-sample validation. Fifth, our expenditure variables are nominal and in levels rather than real, population-adjusted terms; demeaning nets out trends common to all provinces, including inflation, but cannot correct for cross-province differences in population or price levels, which may contribute to our expenditure coefficients. Sixth, our panel holds Indonesia’s geography fixed at 33 units throughout 2010–2024 by construction of the underlying BPS series (Section 3.1), even though the official province count rose over the period; we rely on that series’ boundary treatment rather than re-aggregating, and future work with district-level data could verify it.

Future work could extend this analysis in five directions. Three concern data and identification: an instrumental-variables or dynamic GMM strategy suited to spatial panels to move toward causal identification of the expenditure–HDI relationship; testing whether the apparent 2022–2024 weakening of spatial clustering persists as further post-pandemic data accumulate; and disaggregating the health-expenditure finding by sub-category (curative versus preventive spending) to distinguish a compensatory-allocation explanation from alternatives. Two concern specifications: a heterogeneous-coefficient spatial panel to test whether the province-level instability our jackknife detects reflects genuine coefficient heterogeneity rather than a single outlier; and replacing education and social-protection expenditure scale with measures of spending quality or targeting, which our design cannot distinguish from volume and which Section 4.8 identifies as the likely reason these covariates carry no detectable association with HDI.