This study generates data from 119 regencies and cities in Java as cross-section data. The selection of the 2015-2019 period was made due to the pre- global crisis and the decline in economic growth as the consequences of the Covid-19 pandemic. One of the prominent matters in the analysis of regional economic growth, according to (Lesage & Fischer, 2008), is describing variables included in the research. This study uses the Mankiw-Romer- Weil (MRW Model), which was initiated by (Mankiw Gregory et al., 1992). This model was derived from the neo- classical model with an addition of a factor, i.e., human capital. This, the MRW Model was used to describe all variables in this study. According to the Model, per-capita regional income growth is the main explanatory and dependent variable. Previous studies that used the same model are those by (Aspiansyah & Damayanti, 2019), (Fischer, 2011), (Nurjanna et al., 2020) and (Sun et al., 2017). There are at least two advantages of using the MRW model.MRW Model can detect the existence of convergence in regional economic growth according to the assumption of diminishing return to capital in rich regions and the catching-up process of poor regions (Mankiw, 2010). According to (Mankiw Gregory et al., 1992) and (Fischer, 2011), the β parameter of the initial per capita income is the indicator of the existence of convergence. This model can be extended according to the necessity of the analysis., Moreover, based on the purposes of this study, MRW Model adopts spatial effect and becomes MRW Spatial Model ((Ertur & Koch, 2007); (Fischer, 2011)). MRW Spatial Model’s problem statement is similar to the problem statement of this study, which is the direct effect from the region itself and the indirect effect of spatial spillover from other regions. Therefore, the economic growth of a region is not only affected by initial per capita income, physical capital investment, human capital, population growth, and total factor productivity from the region itself but also influenced by the initial per capita income, physical capital investment, human capital, population growth, total factor productivity, and economic growth of other regions. All variables and their operational definitions can be seen in Table 2.

Spatial regression models cannot be separated from the use of a spatial weight matrix. A spatial weights matrix is an n×n positive symmetric matrix W with element W_ij at location i, j for n locations. The values of Wij or the weights for each pair of locations are assigned by some preset rules which define the spatial correlation among locations and, therefore, determine the spatial autocorrelation statistics. By convention, W_ii or W_jj = 0 for the diagonal elements (Dubin, 2012).

According to (Lesage & Fischer, 2008), the travel time measures of distance reflect economic distance better since it may introduce additional realism to connectivity. The structure of the road networks, the presence of mountains, rivers, oceans, landlocked areas, national car, and lorry speed limits, as well as statutory rest periods for drivers, may lead to large differences between geodesic and drive time distances. The utilization of travel time as a spatial weight matrix has been performed by (Vidyattama, 2014), (Mustajab, 2009), and (Lesage & Fischer, 2008). The Excell Geocoding Tool in Microsoft Excell which is connected to Bing Maps Distance Matrix API was obtained to calculate the travel time between regions. The combination of longitude and latitude was used to detect the location of regions. However, the inverse distance matrix was utilized for denoting the regions’ correlation in travel time, i.e., Where dij is the time needed to travel from i to j.

Moran’s I was mainly used for global

autocorrelation analysis, which is defined as:

Where Ȳ is average of observations across regions, Wij is weights of spatial linkages between and , are the data of -th location variable ( = 1, 2, …., n), Yj are the data of i -th location variable ( = 1, 2,….,n). Moran’s I is a number that ranges from -1 to 1. If Moran’s I is greater than zero (I > 0), there is positive spatial autocorrelation, which indicates that the close observation locations have similarity (clustered). If Moran’s I is less than zero (I < 0), there is negative spatial autocorrelation, which indicates that the observations at an adjacent location tend to be different (dispersed) (Anselin, 1988).

For recognizing the specific location of spatial clusters or grouping and outliers, a local analysis of spatial autocorrelation is needed. According to (Anselin, 1995), local indicators of spatial association (LISA) is proposed to indicate the existence of any spatial autocorrelation at each unit which determines the spatial relationships among the units. The formula of LISA is as follows.

The result of Ii = 0 indicates that i has no spatial autocorrelation. Otherwise, the result of Ii ≠ 0 indicates that region has spatial autocorrelation. For determining the cluster value of region i and j, Moran Scatter Plot was used. Moran Scatter Plot is an illustration of the relationship between the value of the chosen attribute at each location unit depending on the average value of the same attribute at the neighboring locations. According to local Moran’s I Index (I), Moran Scatter Plot is dividen into 4 quadrants. Quadrant I is high-high (H-), which indicates high economic growth with high economic growth of neighbors, or clustering occurred; quadrant II is low-high (L-H), which indicates low economic growth with high economic growth of neighbors or the presence of spatial outlier; quadrant III is low-low (L-L), which indicates low economic growth with low economic growth of neighbors, or clustering occurred; and quadrant IV is high-low (H-L), which indicates high economic growth with low economic growth of neighbors or the presence of spatial outlier.

This study uses panel data. According to (Elhorst, 2014), the growing interest of econometric relationships based on spatial panel is caused by the increased availability of more data sets in which spatial units are followed over time. The selection of panel data model analysis such as fixed effect model, random effect model, or common effect must be considered. In the other words, before obtaining the accurate spatial panel data estimation results, it is necessary to select panel data regression (Elhorst, 2014). Hausman test was used for appropriately selecting the panel model in a spatial econometric. This test can be used to test the random effect model (REM) against the fixed effect model (FEM), which is more efficient than the common effect model (CEM). According to (Elhorst, 2014), the standard for spatial panels are fixed and random effect. If the result of the Hausman test is 0, random effect model must be used for the panel data model. If the result is not equal to 0, fixed effect model must be used. However, various similar previous studies by (Sun et al., 2017), (Arbia et al., 2005), (Vidyattama, 2014), (Endah Ervina & Jaya, 2018), (Wang et al., 2021), and (Álvarez & Barbero, 2016) used the fixed effect model for spatial panel data model estimation.

This study adopts Spatial Durbin Model (SDM) for the spatial econometric panel data model. According to (LeSage & Pace, 2009), the Spatial Durbin Model (SDM) is an extension of the Spatial Autoregressive (SAR) Model which adds spatial lag to its independent variables. Numerous past researchers such as (Fischer, 2011), (Aspiansyah & Damayanti, 2019), (Panzera & Postiglione, 2022), (Nurjanna et al., 2020), (Sun et al., 2017), dan (Álvarez & Barbero, 2016) found that, should the research on regional economic growth be conducted using MRW Model, the most theoretically appropriate regression type is the Spatial Durbin Model. This model enables the determination of a region’s economic growth with the help of the region’s independent variables, other region’s economic growth, and independent variables of other regions (Aspiansyah & Damayanti, 2019).

The estimation model for region using SDM

(Fischer, 2011) is as follows.

(1)

Where,

yit : n × 1 vector of the dependent variable of each i region at time t

yjt : n × 1 vector of depenent variable of each j region at time t

xit : independent variable of i region at time t

xjt : n × 1 vector of independent variable of each j region at time t

λ : the spatial autocorrelation coefficient

wij : i × j spatial weight matrix

β0 : intercept or constant parameter

βk : k × 1 vector of regression parameters

γk : k × 1 weighted vector of regression parameters

εi : n × 1 vectors of error terms at time t

The model estimation of SDM above is

modified to structural model (Elhorst, 2014):

Y = λWY + β0 + Xβ + WXλ  + ε

Where, λ, β, and γ are parameters from the dependent variable with spatial lag weighted, independent variable without spatial lag weighted, and independent variable without spatial lag weighted.

Based on equation 1, the proposed model, which is the combination of all variables of MRW Model, the model design of this study is as follows.

The distribution of this research’s data was explained using descriptive statistics. There are several items that can be statistically identified. The economic growth level of a regency, which is the response or dependent variable of this study, is proxied by per capita GRDP growth. The descriptive statistics was useful in identifying

that the average economic growth of regencies and cities in Java during the 2015-2019 period was 4.654%, with the highest economic growth being in Blora Regency in 2016 (23.016%) and the lowest economic growth being in Bangkalan Regency in 2015 (3.529%). Other statistical data of other variables such as mean, median, highest value, and lowest value ais presented in Table 3 below.

For exploring the spatial dependence on inter- regional economic growth among 119 regencies and cities in Java, statistical data over from 2015 to 2019 were employed. The data was gathered from Statistics Indonesia. In order to investigate the spatial dependence or spatial autocorrelation on inter-regional economic growth in Java in general, Moran’s I value was used. According to Table 3 below, the Moran’s I value for per capita region income growth can be indicated for per-year spatial dependence on inter-regional economic growth. To perform analyses on spatial dependence, spatial connectivity or spatial weight matrix across the 119 regencies and cities in Java was constructed. In this paper, travel time adopting inverse distance matrix was used for defining the neighborhood of each region. Then each region obtains 118 links or neighbors, and the total number of nonzero links is 14.042.

The measurement of the global spatial autocorrelation based on Moran’s I value and p-value (Table 3) shows the existence or significance of spatial dependence among the 119 regencies and cities in Java every year from 2015 to 2019. All Moran’s I value of all years fall between -1.0 and +1.0 or ≠0 with the p¬-value of < 0.05 (significance level of 5%), with an exception for 2018 whose p¬-value is < 0.1 or the significance level of 10%. Based on those values, there are positive spatial autocorrelations or spatial clustering among the 119 regencies and cities in Java from 2015 to 2019. However, the low Moran’s I value indicate the low inter-regional spatial autocorrelation/dependence or economic growth spillover. This low spillover occurred since all of the 119 regencies and cities in Java were involved together as an observation unit. Whereas, according to their economic activity, each regency and city have the clustering characteristic or spatial concentration. The rapid development of regions in Java is driven by economic activities in metropolitan areas and large cities. Regencies and cities with the lower distance from metropolitan areas enjoy more economic spillover effect (Pravitasari et al., n.d.). (Rahadini, 2018) even stated that the two largest manufacturing concentrations within Java are in Greater Jakarta and Greater Surabaya. They create bipolar patterns of economic activities from eastern Java to western Java along north Java corridor. This leads to differences in regional development between northern and southern regions of Java.

According to (Elhorst, 2014), before proceeding to data analysis for modeling the spatial panel data, it is important to determine the best model using the Hausman Test. The result of the Hausman Test is that the p-value =

5.79e-07. As the p-value < 0.05, the hypothesis is rejected. Hence, this study uses the Fixed Effect Model (FEM) in combination the with Spatial Durbin Model (SDM). The use of SDM model in this research accommodates dependencies on the inter-regional economic growth by the 2015-2019 panel data . Then, SDM-FE concept model was used to examine the determinant factors affecting the economic growth of a region. According to numerous previous researchers, the economic growth a region is influenced by the determinants of the region itself, by the determinants of other regions (spatial spillover), and by the economic growth of other regions (spatial spillover). The regression results are shown in Table 4 below.

Based on the results of the spatial regression estimation on the panel data in Table 4, the influence of all determinants of spatial dependence on inter-regional economic growth in Java can be interpreted. The results were obtained from the application of Fixed Effect of Spatial Durbin Model using splm package in R Studio. The estimation results show that the influences of the determinants from the region itself (i) – such as initial per capita income, physical capital investment, road infrastructure, population growth, and education except total factor productivity – are significant. This confirms the MRW theory of (Mankiw Gregory et al., 1992).However, the direction of the effects is different. As the directions of physical capital investment, road infrastructure, and education are positive, those of per capita income and population growth are negative.

First, the spatial autocorrelation λ (lambda) is negative and significant. It can be seen from the coefficient value of -0.96179 and the p-value of 0.007701 (significant at 0.01). The value of shows the existence of spatial dependence in the inter-regional economic growth in Java. This result indicates that all regions in Java cannot be treated as an independent observation unit and that, hence, the growth model among the regencies and cities in Java should explicitly account for this spatial dependence. This result supports the findings of (Aritenang, 2014), (Miranti & Mendez, 2022), (Aspiansyah & Damayanti, 2019),(Nurjanna et al., 2020), (Endah Ervina & Jaya, 2018), and (Laksono et al., 2018) that spatial dependence in inter-regional economic growth in Indonesia, both provincial and by districts, exists. However, (Takeda, 2013) concluded that there is no spatial regional economic dependence in Indonesia. The estimation results show that the coefficient of spatial economic growth (-0.96179) indicates negative spatial spillover effects among regions in Java. It also indicates that increased economic growth of regions around a regency or city in Java might decrease its current economic growth. According to (Laksono et al., 2018) and (Pasaribu et al., 2014) the empirical analysis on the spillover effect is important since the application of growth centers theories that has been done by developed countries and also by developing countries still raise pro-cons. Examination on spillover effect of growth centers in Kalimantan specifically has been done in order to reveal whether its contribution as a National Energy Stocks as mentioned in MP3EI program does not caused backwash effect for the surrounding areas. Early investigation of existence of spatial lag dependent between growth centers in Kalimantan and its surrounding areas was tested using Lagrange Multiplier Spatial Lag Dependent. It proved that output growth, labor growth, and investment growth which happened in the growth centers in Kalimantan significantly had negative spillover effect (backwash effect, the negative spillover effect is also called as the backwash effect. The labor movement from underdeveloped to more developed areas, unbalanced investment flows, trade and industry activities dominated by more developed regions, and the better accessibility in more developed regions are the causes of the negative spatial spillover effect (Laksono et al., 2018). This result is similar to the findings of (Aritenang, 2014), who studied selected districts in Indonesia, of (Laksono et al., 2018), who studied regencies and cities in East Java, and of (Vidyattama, 2014), who studied the matter at a provincial level. (Vidyattama, 2014) argued that the negative spatial spillover was caused by several developed areas like Jakarta, which have absorbed the financial and human capital resources from other regions.

Second, the (β) coefficient of initial per capita income is -0.46798 with the p-value 0.006427. The negative sign of the initial per capita income specifically indicates convergencies among the 119 regencies and cities in Java. This finding confirms the regional economic growth concept of MRW model suggested by (Mankiw Gregory et al., 1992). The identified convergencies point out the process of economic disparity reduction across the Javanese regencies and cities. This finding is similar to the finding of (Aspiansyah & Damayanti, 2019) that convergencies exist in the economic growth of Indonesian provinces and the finding of (Miranti & Mendez, 2022), who studied all regencies and cities in Indonesia. The spatial lag coefficient of the initial per capita income (γ) is 4.419659 with the p-value of 0.058734. This significant and positive impact implies that the increase of initial per capita income in a certain regency or city tend to increase the economic growth of its neighboring regencies and/or cities. In other words, positive spatial spillover occurs. This finding supports the result of (Aspiansyah & Damayanti, 2019), (Álvarez & Barbero, 2016), and (Sun et al., 2017). Hence, when a region neighbors another richer region, the former region will benefit from the latter region’s economic activity due to intensive commercial relations between the two (Álvarez & Barbero, 2016).

Third, the (β) coefficient of physical capital

investment is 0.350384 with the p-value of 0.000110. The positive sign of the physical capital investment indicates the positive influence on the regencies and cities’ own economic growth. In other words, it was proven that the larger the initial per capita income, the higher the economic growth of the regencies and cities in Java. This result supports the finding of (Fischer, 2011), which adopted (Mankiw Gregory et al., 1992) in describing the source of economic growth from physical capital, labor, and human capital. Further, the estimation of the spatial lag of physical capital investment (γ) resulted in the coefficient of -0.6389 and the p-value of 0.630491, indicating no spillover effect. Thus, the increase or the decrease in the physical capital investment of regencies and cities has no tendency of either increasing or decreasing the economic growth of other regencies and cities. (Aspiansyah & Damayanti, 2019) and (Miranti & Mendez, 2022) also did not find any spatial dependence effect of physical capital investment on regional economic growth in Indonesia. According to (Aspiansyah & Damayanti, 2019) the absence of the spillover effect of physical capital investment on the regional economic growth was caused by the limited physical capital of a region.

Fourth, the (β) coefficient of road infrastructure is 0.2997 with the p-value 0.067215. The positive sign of the road infrastructure indicates the positive influence on the regencies and cities’ own economic growth. In other words, the better the road infrastructure of a region, the higher its economic growth through accessibility. (Vidyattama, 2014) also found that road infrastructure is a positive determinant for regional economic in Indonesia. (Arbués et al., 2014) examined that road infrastructure is an important determinant for accessibility to enhance 47 provinces in Spain. Meanwhile, the spatial lag estimation for road infrastructure (γ) resulted in the coefficient of 2.992056 and the p-value of 0.139161, indicating no spillover effect. Thus, the increase or decrease in the road infrastructure of regencies and cities has no tendency of either increasing or decreasing the economic growth of other regencies and cities. This result contradicts the findings of (Vidyattama, 2014) and (Arbués et al., 2014). (Rahadini, 2018) argued that the bipolar development from Jakarta to Surabaya, in which goods and services are transported, has transformed some regions in Central Java into traversed areas. Thus, Central Java was divided into three regions according to certain characteristics.

Fifth, the (β) coefficient of population growth is -0.7177 with the p-value of 1.791e-13. The negative sign for the population growth indicates the negative influence on the regencies and cities’ own economic growth. In other words, it was proven that the larger the population growth, the lower the economic growth of the regencies and cities in Java. The significant and negative influence of population growth supports the findings of (Ertur & Koch, 2007), (Aspiansyah & Damayanti, 2019), (Álvarez & Barbero, 2016), and (Sun et al., 2017). According to (Aspiansyah & Damayanti, 2019), high population growth leads to a low proportion of labor force, which decrease per capita income. Meanwhile, the spatial lag estimation of population growth (γ) resulted in the coefficient of 1.960392 and the p-value of 0.12078, indicating no significant spillover effect. The absence of spatial lag of population growth indicates that the increase or decrease in the population growth of regencies and cities has no tendency of either increasing or decreasing the economic growth of other regencies and cities. This result varies from the findings of previous research in general. (Álvarez & Barbero, 2016) concluded no spillover effect of population growth between provinces in Spain. (Supartoyo et al., n.d.) argued that the insignificant spatial effect of population growth on economic growth was caused by low quality of population for the current economic activities. Large populations with low quality burdens the economy, not enhancing it. According to the theories, economic growth is determined by population growth rate. However, population growth does not necessarily give a positive contribution to economic growth.

Sixth, the (β) coefficient of education, i.e., the proxy of human capital investment, is 0.012776 with the p-value of 0.008465. The positive sign of the education indicates the positive influence on the regencies and cities’ own economic growth. In other words, it was proven that the better the education, the higher the economic growth of the regencies and cities in Java. The significant and positive influence of education supports the findings of (Aspiansyah & Damayanti, 2019), (Fischer, 2011), (Nurjanna et al., 2020), and (Álvarez & Barbero, 2016). This result is also in line with the theory of MRW Model. Increasing investment in human capital through education will improve production, which in turn will increase economic output. However, this result contradicts the finding of (Miranti & Mendez, 2022) that education has no influence on human capital and on regional economy in Indonesia. Meanwhile, the spatial lag estimation of education (γ) resulted in the coefficient of -0.12644 and the p-value of 0.098367 (significant at 0.1), indicating significant spillover effects. The existence of spatial lag of human capital investment indicates that the increase of human capital investment in regencies and cities tend to decrease the economic growth of other regencies and cities. This result is similar to the findings of (Álvarez & Barbero, 2016) and (Miranti & Mendez, 2022)that the increasing economic growth through spatial spillover of human capital investment is caused by the migration of low or skilled labor between regions. The migration of population is caused by various factors, generally economic activity. However, this result contradicts the findings of (Aspiansyah & Damayanti, 2019) and (Fischer, 2011).

Seventh, the (β) coefficient and the p-value of total factor productivity (TFP) indicate no influence on the regencies and cities’ own economic growth. The results of the spatial lag estimation on the total factor productivity indicate no spatial spillover effect, which means that the increase or decrease in total factor productivity in regencies and cities has not tendency of affecting the economic growth of other regencies and cities. This result differs from the findings of (Ertur & Koch, 2007) and (Fischer, 2011), whose research uses the MRW model of (Mankiw Gregory et al., 1992). The absence of total factor productivity among regencies and cities in Java indicates that there is no maximization in the driving factors of TFP such as level of openness, R&D investment, industrial structure, government expenditure, and human capital (Zeng et al., 2022).

After examining the role of spatial dependence in the inter-regional economic growth of 119 regencies and cities in Java, it can be inferred that negative spillover effects or backwash effect occurred (see Table 4). This is an indication that development activities in Java as a whole have not been synergistic. Therefore, development policies from government need to be carried out. These policies are needed to balance the development across regions and to decrease the backwash effect in the spatial growth process because such effect can cause divergence and worsen the inequality in the economic growth among regions (Barro & Sala-I-Martin, 1992).

In the Indonesian context, strategies to reduce inequality and to distribute growth among regions have been made, but the policy implementation evaluation must follow. In detail, the development strategy for each region has been written in the National Medium-Term Development Plan (RPJMN), the most recent being the 2020-2024 RPJMN. There are two types of development policy approaches for each district and city in Java: the growth approach and the equity approach. The growth approach is the development of growth centers through regional potentials that can increase added value, foreign exchange, employment opportunities, and even distribution of economic growth for the next five years. The equity approach prioritizes the development of hinterlands, which are around growth centers, as well as of backward regions. The two approaches are applied to maintain the growth momentum of developed regions and increase equity in less developed regions. The application of the two approaches is supported by (Hardjoko et al., 2021) and (Sakti, 2005), who mentioned pro-growth (pro-concentration) and pro-equity (pro-dispersion) regional development. The combination of these two policies is the right collaboration to increase convergency and reduce economic disparities between regions. In addition to the implementation of the RPJMN, the opening of toll-road gates in each region must also be evaluated because its impact varies across regions. (Ardiyono et al., n.d.) found that the revenues of Small and Medium-Sized Enterprises (SMEs) around hotels and restaurants along the non- toll road of Pantura in Subang, Indramayu and Brebes has decreased along with the reduction of traffic. Similarly, (Dwiputri et al., 2022) found a negative impact of the opening of the Pandaan- Malang toll road on the income of the people living around the toll-road exits.

The authors would like to express their gratitude and heartfelt thanks to the Indonesia Endowment Funds for Education (Lembaga Pengelola Dana Pendidikan or LPDP) from the Ministry of Finance of the Republic of Indonesia for granting financial support for the completion of this research.