Regional economic structure and heterogeneous effects of monetary policy: evidence from Indonesian provinces

The importance of the industrial structure across regions accentuates the relevance of the interest rate channel of monetary transmission. The effect of a change in the policy interest rate is more significant in regions where the share of interest-sensitive industries in the economy is relatively higher (see, e.g., Taylor 1995; Mishkin 1996). This is because some industries are more interest-sensitive than others. While classifying more interest sensitive industries is an open question, some studies categorize durable and investment goods producers and other highly capital-intensive industries as interest sensitive (Carlino and DeFina 1999).

The operation of the credit channel of monetary policy transmission is premised on the assumption of imperfect substitution of retail bank deposits and loans. Most borrowers cannot access financing sources other than banks. Transmission through the credit channel implies that monetary policy works by affecting bank assets, i.e., loans, in addition to interest rates (Bernanke and Blinder 1992). For example, loosening monetary policy raises the price of collateral and increases banks’ equity capital or net worth, allowing them to access more financing sources. As a result, the amount of bank loans available rises since bank reserves and deposits increase, and this eventually affects aggregate demand by accelerating investment and consumption. The responses to monetary policy might differ from one bank to another, depending on the components through which the credit channel operates: the bank’s balance sheet, lending capacity, and capital. However, the different reaction of all the credit channel components across individual banks is reflected in the total amount of bank credit disbursed as a financing source in the economy for either consumption or investment. Empirical evidence suggests that the credit channel amplifies the effects of monetary policy on output and prices more forcefully if bank credit is more important in an economy, indicated by a higher ratio of total bank credit to gross domestic product (GDP) (see, e.g., Dornbusch et al. 1998; Cecchetti et al. 1999; Mihov 2001).

The importance of regional export intensity for the differential effects of monetary policy is still debated given the mixed empirical evidence. On the one hand, some studies argue that regions with more export-intensive industries appear to be more responsive to monetary policy change. When contractionary policy is implemented, currency appreciation might push export products’ prices to rise, reducing domestic firms’ competitiveness and dragging down demand for exports. By the same logic, currency appreciation might lead to a surge in imports due to the relatively cheaper price of imported goods and smaller markups. The combination of these two effects stemming from currency appreciation eventually worsens net exports, decreasing exports and increasing imports, and might push the domestic price level downward due to the income effect (Hayo and Uhlenbrock 2000). On the other hand, evidence from different studies supports the view that monetary policy effects are less pronounced in highly export-intensive regions (Ber et al. 2001). The central argument of this view is based mainly on the evidence for a supply-side effect that works via the domestic credit market under the free movement of international capital. When domestic financing becomes more expensive due to monetary tightening, exporting firms can access alternative financing from the foreign currency market, from foreign financial institutions or their affiliated parties, and thus do not necessarily reduce their investment. Moreover, exporting firms’ financial conditions are less exposed to domestic interest rates since their revenue depends largely on external market conditions.

In the context of monetary policy transmission, labor market rigidity expresses how frequently nominal wages can be adjusted in a given circumstance. As suggested by the baseline new Keynesian business cycle model, when nominal wages can be adjusted less frequently, the response of firms’ marginal costs to a monetary policy shock is minimal, which is thus reflected in a weaker inflation response (see, for example Galí 2015). In practice, a group of empirical studies supports this view with different model specifications to explain the precise mechanisms (see, e.g., Zanetti 2007; Christoffel and Kuester 2008; Lechthaler et al. 2010). In general, the studies find that inflation responses to a monetary policy shock are more muted in economies with more rigid labor markets.

We begin by assuming that \({Y}_{t}\) is an M × 1 vector of observed economic variables and F_t is an F × 1 vector of unobserved or latent factors that contain most of the relevant economic information. The FAVAR model assumes that the dynamic relationship of \({Y}_{t}\) and \({F}_{t}\) jointly follows a VAR process (Bernanke et al. 2005), that is:where Φ(L) is a finite-order polynomial in the lag operator and may include a priori constraints such as those found in the structural VAR literature, and \({v}_{t}\) is the error term with zero mean and covariance matrix. One could refer to as standard VAR in \({Y}_{t}\) when the terms of Φ(L), that connect \({Y}_{t}\) to \({F}_{t-1},\) are all zero.

Bernanke et al. (2005) assume that the non-observed factors \({F}_{t}\) and the observable variables \({Y}_{t}\) can be related to the informational series in \({X}_{t}\) as expressed in the following equation:where \({\Lambda }^{f}\) is the N × K loading matrix, \({\Lambda }^{y}\) is the N × K matrix of coefficients, and \({\varepsilon }_{t}\) is the N × K error vector with a zero mean.

We cannot directly estimate Eq. 1 since the F_t factors are not observed. Therefore, we apply a two-step approach with principal component analysis (PCA), following Bernanke et al. (2005). The first step involves estimating \(\widehat{{F}_{t}^{1}}, \widehat{{F}_{t}^{2}} ,\dots , \widehat{{F}_{t}^{n}}\) which are the first principal components obtained from each group of a series carrying the largest eigenvalue in each category. In the context of our study, the factors summarize all information on inflation dynamics that is excluded from\({Y}_{t}\). In the second step, we use these factors within the VAR framework, as represented in Eq. 1, to estimate Φ(L). Specifically, following the lead of Bernanke and Boivin (2003) and Bagliano and Morana (2009), before conducting PCA, we first group the variables under consideration into two categories that represent two different factors affecting inflation: (1) domestic financial system conditions and (2) global economic and financial conditions. We normalize all series to make the values of each series in the data have zero mean and unit variance to reduce the bias toward high variances caused by the multiple measurement units of the series used in PCA. Table 4 in the Appendix contains the list of indicators used in the estimations of the factors.

Then, to preserve the model’s economic interpretation, we create a structural form of the FAVAR model. Following Ridhwan et al. (2014), we also define a vector of exogenous variables represented by a factor that contains major global economic information, including the industrial production index and price dynamics in the US and China and the international price indices of oil and other commodities.

The corresponding structural model of our FAVAR can therefore be written in the following reduced form:where \({Y}_{t}\) is an N × 1 matrix of endogenous variables, \({Z}_{t}\) is a vector of exogenous variables, \(A(L)\) and \(B(L)\) are polynomial matrices, and \({\mu }_{t}\) is a vector of reduced-form disturbances.

We further use the above structural FAVAR (SFAVAR) specification to model economic activity in the 34 Indonesian provinces with a set of endogenous variables for province i as follows:where \({\mathrm{GDP}}_{t}^{\mathrm{nat}}\) is aggregate national output, \({\mathrm{INF}}_{t}^{\mathrm{nat}}\) is aggregate inflation, \({\mathrm{POL}}_{t}\) is an interest rate variable representing the monetary policy stance, \({\mathrm{NER}}_{t}\) is the nominal exchange rate, \({\mathrm{FIN}}_{t}\) is a financial factor, and \({\mathrm{GDP}}_{it}\) and \({\mathrm{INF}}_{it}\) are regional output and inflation in province i, respectively.

Finally, as Christiano et al. (1999) point out, the order of the structural equation variables is critical in determining the identification criterion. Therefore, the restrictions are carefully defined by following Bernanke et al. (2005), who propose the assumption that “slow-moving” variables at the national level [in our case, national GDP (GDP) and INF] do not respond contemporaneously to unanticipated changes in monetary policy while “fast-moving” variables at the national level (in our case, FIN) are allowed to respond contemporaneously to policy shocks. Overall, the identification restriction in our setting refers to economic theories that exclusively limit the contemporaneous structural characteristics as follows:where  ∗  indicates unrestricted parameters.

Following the standard literature on monetary policy transmission analysis using FAVAR models, we use various national and regional variables. In total, we exploit data on 45 macroeconomic variables at monthly frequency from January 2010 to December 2019. The endogenous variables at the national level include the nominal policy rate (POL), the GDP growth rate (GDP), CPI inflation (INF), the nominal exchange rate expressed in IDR/USD (NER), and the financial factor (FIN), while the external factor (EXT) denotes exogenous variables.⁶ The nominal policy rate (POL) represents the monetary policy stance of the central bank; the nominal exchange rate (NER) and financial factor (FIN) are intermediate variables through which monetary policy is transmitted to the whole economy; both the GDP growth rate (GDP) and CPI inflation (INF) are the final variables affected by monetary policy changes. At the regional level, we include the growth rate of provincial GDP and CPI inflation. Following Wimanda et al. (2011), we convert the GDP series from quarterly to monthly frequency by using the quadratic-match-average interpolation technique. In this setting, the GDP growth rate is the percentage change in GDP at time t from t − 12 according to the constant price in 2010. CPI at the province level is computed as the weighted average of CPI at the city level with this formula:where \({\mathrm{CPI}}_{jt}\) is the CPI of province j at time t, \({\mathrm{CPI}}_{it}\) is the CPI of city i at time t, \({N}_{j}\) is a set of indices for city located in province j, and \({\omega }_{i2012}\) is the consumption weight of city i based on the 2012 cost of living survey (the “Survei Biaya Hidup” in Indonesian). Then, the CPI inflation at the province level is the percentage change of CPI over twelve months. Data for POL, NER, and INF are collected from Bank Indonesia and the Indonesian Central Bureau of Statistics, respectively.

The estimated coefficients and factors in the SFAVAR model are shown in Table 5 in the Appendix.¹² The coefficients of GDP growth, inflation, and the nominal exchange rate are consistent with expectations from theory. Both GDP growth and inflation decline following monetary policy shocks in the previous period, while the nominal exchange rate appreciates.

We also report the impulse response functions in Fig. 2. The responses are generally of the expected sign and magnitude. The immediate responses to a hike in nominal policy rates are the appreciation of nominal exchange rate and the decrease of financial factor. Then, GDP growth rate gradually falls. However, the initial consequence of rate of inflation against a nominal interest hike is a jump of rate of inflation, which is not consistent to monetary theory but has been observed in many studies on monetary transmission mechanism called “Price Puzzle”.

Although there is a price puzzle observed here, the response to monetary tightening is eventually followed by a drop in inflation pressure from the seventh month onward as the market expects slowing demand. The existence of a price puzzle in monetary transmission, that is, a rise in prices in the short term due to positive shocks from the monetary policy contraction, has also been evidenced in other countries, as documented by Ramey (2016).¹³ For the case of Indonesia, Kusmiarso et al. (2002), Harahap et al. (2013), and Nurliana et al. (2016) also show a price puzzle in monetary transmission.

After implementing the SFAVAR at the national level, we proceed with our analysis at the province level as expressed in Eq. 4 to capture the contemporaneous response of inflation in each province to a common monetary policy shock. In Fig. 3, we plot the impulse responses from the SFAVAR analysis at both the national and province levels. All graphs share the same scale on the vertical axes to facilitate comparability. The figure demonstrates evidence of heterogeneous effects of a common monetary shock across Indonesian regions, where provinces on the Java and Bali Islands have very similar responses to the national response.

Based on the impulse response of each province, we construct the measurement of the regional inflation response to monetary policy shocks. We include both the magnitude and length to gauge two measurements of a regional inflation response to monetary policy. The first measure is computed by multiplying the peak of the impulse response with the corresponding time taken to reach the peak, as represented by Eq. 7:

We call this measure the monetary policy impact. The second measure is monetary policy efficiency, where we divide the peak of the impulse response by the corresponding time taken to reach the peak, as shown in Eq. 8:¹⁴

Next, in Fig. 4, we visualize the geographical distribution of the monetary policy impact on regional inflation by using quantile classification. We observe an expected pattern from this mapping: all provinces in Java (except Jakarta) show a larger than median impact. This island is home to large manufacturing industries, with the share of the manufacturing sector on the island accounting for nearly 70% of national manufacturing industry output (based on real GDP in 2019). Furthermore, the size of bank lending on the island contributes 62% to total national bank lending (2019), implying a significant role of these two regions in the national banking industry. Similar patterns are also observed in Ridhwan et al. (2014), where manufacturing regions in Java and Sumatra are found to be more sensitive to monetary policy shocks.

Up to this point, we have shown our findings from the SFAVAR analysis of differences in regional responses following monetary policy actions. Next, we examine regional factors that explain the variation in the reaction of regional inflation to monetary policy changes. More specifically, our goal is to answer the following question: to what extent do regional structural features explain the heterogeneity in regional inflation responses to monetary policy shocks? To achieve this goal, we implement a cross-sectional ordinary least squares (OLS) estimation as follows:where the dependent variable V is the measure of monetary policy impact and monetary policy efficiency, respectively, derived from the estimated impulse responses based on the previous FAVAR model, \({\beta }_{0}\) is the intercept of the model, \({Z}_{t}\) corresponds to the jth explanatory variable of the model representing regional economic structures, and ε is the random error with expectation 0 and variance \({\sigma }^{2}\).

The existing literature provides a good background for the selection of regional features that can explain the different responses to monetary policy actions. For example, Arnold and Vrugt (2002, 2004) point out the importance of the industrial composition to explain the regional transmission of monetary policy in the Netherlands and Germany. For the case of Greece, Anagnostou and Papadamou (2016) find that regions’ sensitivity to interest rate shocks is related to regional characteristics such as the share of particular industries in GDP. Finally, Ridhwan et al. (2014) also mention that industrial composition explains the variation in the regional impacts of monetary policy changes, corroborating the pertinence of the interest rate and credit channels for monetary policy transmission in Indonesia.

Referring to the previous literature and discussion in Sect. 3, we consider several indicators representing the regional economic structure that might explain regional heterogeneity in the response of regional inflation to monetary policy shocks. The first two variables are the share of manufacturing to GDP and the share of mining sector to GDP. Commonly used in previous studies, these two variables are used to capture regional differences in capital-intensive industries or industrial mixes. As mentioned earlier, these variables reflect the way that monetary policy is transmitted via the interest rate channel. The third and fourth variables are the bank lending to GDP ratio and the export share in GDP, respectively. While the former reflects the relative importance of bank lending to the provincial economy, the latter demonstrates regional openness. According to the literature, these variables have information on how the credit and exchange rate channels of monetary policy operate at the regional level. As shown in Table 1, these four initial variables represent different regional economic structures that potentially give rise to the heterogeneity in regional responses to monetary policy. We also include the manufacturing industry’s annual growth rate to capture the importance of manufacturing in the regional economy. Moreover, recent studies have shown the relevance of investigating the spatial dependence of regional inflation (see, e.g., Yesilyurt and Elhorst 2014; Aginta 2020b). Thus, to capture the role of spatial externalities, we include the spatial lag of inflation, that is, neighboring provinces’ inflation rates. More precisely, we compute the spatial lag of the inflation rate by applying a k-nearest neighbors (KNN) weight matrix (W). The neighbors of a province are the four nearest provinces based on the Euclidean distance.¹⁵ Finally, to capture labor market flexibility, we use the standard deviation of the regional unemployment rate, which reflects the labor market’s sensitivity to macroeconomic shocks.

Table 2 shows the results of the regression on monetary policy impacts. Similar to the findings from previous studies, the results reveal that four variables related to the regional economic structure are statistically significant in affecting the regional inflation response to monetary policy shocks (see Model 1). First, both the share of manufacturing to GDP and the mining sector’s share to GDP have a positive effect on the regional impact of monetary policy. In the literature on monetary policy transmission, this finding supports the view that regional differences in industrial composition play an important role in the differential effects of monetary policy. The finding also suggests that the interest rate channel of monetary policy in Indonesia works reasonably well. Second, we find that the bank lending to GDP ratio positively affects the impact of monetary policy on regional inflation. As discussed earlier in Sect. 3, this result confirms the relevance of a credit channel in monetary policy transmission in Indonesia. Third, the estimated coefficient of the export share to GDP is negative, suggesting that the impact of monetary policy is lesser in export-dependent regions. This finding supports the view that the effects of monetary policy are less pronounced in highly export-intensive regions (Ber et al. 2001). Exporting firms’ real investment does not necessarily decline in the face of a higher real interest rate from tightening monetary policy. Exporting firms can access foreign currency credit because they are usually part of a larger trade network and are thus more likely to have contacts and a reputation in foreign credit markets. Additionally, they can obtain trade credit from suppliers or customers abroad. Therefore, when the domestic real interest rate is high, exporting firms may continuously finance their investment by raising foreign currency loans (or foreign currency-denominated credit). For this reason, exporting firms’ income might not be negatively affected by tighter monetary policy. This stable income of exporting firms leads to more persistent aggregate demand in regions that are more export intensive. As a result, the impact of monetary policy on inflation is lower in export-intensive regions.

In Models 2, 3, and 4, we include more variables to show the robustness of the four regional economic structure variables. All models show the consistency of the estimates of the four variables, in both sign and magnitude, suggesting their robustness. Moreover, the improvement in the R-squared to approximately 55% in Models 3 and 4 reflects the importance of our additional variables in explaining the sources of regional heterogeneity in monetary policy impacts. For instance, we find that manufacturing growth has a positive effect on the monetary policy impact, re-emphasizing the role of interest-sensitive industries mentioned before. Additionally, we reveal the importance of geographical externalities, which is new evidence that has never been uncovered in previous studies. Consistent with our expectation, we find a significant negative effect of rate of inflation in neighboring provinces on the impact of monetary policy. This finding indicates that the impact of monetary policy is smaller when the province is spatially surrounded by the neighboring provinces which suffer from high inflation. This finding further implies spatial monetary transmission mechanism where price shock transmits from one province to another. Therefore, the higher the rate of inflation in neighboring provinces, the smaller the impact of monetary policy on that province. We also found as shown by Model 5 that the coefficient of inflation of the province is also statistically significant and negative. This negative sign can be explained by the notion of inflation inertia (Fuhrer and Moore 1995; Mankiw 2001). Inertia in the inflationary process means that it takes longer to bring inflation down to the target if output costs, such as wages, are not easily adjusted due to high inflation in the past. Therefore, if inflation displays inertia, it is natural to expect that the monetary policy impact on inflation is smaller when the inflation rate is high. Wimanda et al. (2011) find significant inertia when analyzing Indonesia’s inflation by using a hybrid version of the new Keynesian Phillips curve (NKPC).

Finally, despite the coefficient having the predicted sign, we find no substantial effects of labor market flexibility on regional monetary policy impact. We propose two possible reasons for this finding: first, the standard deviation of unemployment as a measure of labor market flexibility is far from perfect, and second, even if the measurement is adequate, labor market conditions in Indonesia are not significantly diverse among regions, in contrast to those in the cross-country samples used in previous studies. In summary, our findings favor the literature that emphasizes the role of the regional economic structure in the heterogeneity of regional responses to monetary shocks.

In the second regression, we run another cross-provincial regression with the efficiency of monetary policy as a dependent variable. We separate the efficiency of monetary policy from the total impact of monetary policy to consider the diverse consequences of monetary policy observed in Fig. 3. The figure shows that it takes almost 36 months for the impulse response to die out in some provinces while it takes only 12 to 15 months to do so in some other provinces. It also shows that the impulse response quickly peaks and dies out faster in some provinces but slowly peaks and dies out more gradually in some other provinces. The diversity of the shape of the impulse response functions indicates that the determinants of the speed or efficiency of monetary policy transmission could be different from those of their total impact. For this reason, the efficiency of monetary policy is defined as the maximum value of the impulse response divided by the number of months that it takes for the impulse response to peak.

The regression outcome is shown in Table 3. The R-squared, 61.3% for Model 1 and approximately 60% for others, is fairly good since this is a cross-sectional regression. We examine six variables combined with some other variables to show the robustness of these six variables. The outcome shows different determinants from those of the outcomes of the regression on the total impact of monetary policy discussed before. First, foreign imports, the share of foreign imports divided by provincial GDP, are significant with a positive sign. Monetary policy affects the exchange rate and therefore the prices of imported goods. Monetary policy seems to have a quick effect on inflation in provinces where imported goods account for a large share of consumption. This finding suggests that the exchange rate channel is effective in monetary policy transmission in Indonesia.

Secondly, on the other hand, the shares of investment to GDP and share of manufacturing sector to GDP are statistically significant but negative with respect to monetary policy efficiency. Possible reason for a negative sign is that manufacturing sector share to GDP and investment share to GDP are both sensitive to bank lending. Therefore, monetary policy affects manufacturing firms and investment through a bank lending channel. An interest rate hike has a large impact on manufacturing sector and investment spending. However, the impact of an interest hike is slowly priced-in to an increase in bank lending rate. This is because banks usually adjust their lending rates—following the change in monetary policy—only when borrowers such as manufacturing firms apply for new loans for new investments. In addition, it takes time for banks to undertake credit assessment and finalize the new loan contracts. Consequently, the impact of monetary policy is large but it takes time to materialize the full impact when the province has a large share of manufacturing sector and investment share. In other words, the effects of monetary policy on the real economy are slower in provinces with high manufacturing and investment spending shares to GDP, as shown by the negative signs of the coefficients.

The third finding is that provincial housing prices are statistically significant and positive. This finding does not indicate a causal relationship but shows a positive association between the efficiency of monetary policy and housing prices. Theory implies a causal link not from monetary policy efficiency to housing prices but rather the other way around; i.e., monetary policy has a quick effect in provinces where housing prices are rising since the housing sector is heavily credit intensive and its credit disbursement is quick. This explanation is also true in relation to the statistically significant and positive sign of the construction industry share to GDP.

Apart from the findings above, the regression outcome shows that monetary policy is very slow when GDP growth fluctuates and the share of domestic exports in GDP is high, both indicated by a statistically significant negative sign. Monetary policy may not have a quick impact in provinces where the economy experiences large fluctuations and therefore is subject to a high degree of uncertainty. However, it is hard to find a possible reason for the statistically significant and negative sign for the domestic export share to GDP. One of the possible reasons is that bank credit is not as sensitive for those companies that are expanding exports because an increase in domestic export sales reduces the companies’ risk premium and improves their creditworthiness.