1 Introduction

A large literature has studied the economic consequences of democratization, but we know much less about the conditions that determine whether and to what extent a transition to democracy is successful. The existing body of research has focused on a country’s level of economic development as potential driver and finds mixed results (Rodrik & Wacziarg, 2005; Aghion et al., 2008; Acemoglu et al., 2019). Fortunato and Panizza (2015) and Acemoglu et al. (2019) further show that a better educated population increases the success of democratization. An entirely different literature which does not study political system changes finds that national leaders matter for economic growth (Jones & Olken, 2005), and that educated leaders generate higher growth than others (Besley et al., 2011). Given this evidence and a widespread belief that leaders matter particularly during challenging economic or political times, it is surprising that no study has asked how the characteristics of newly elected democratic leaders shape the economic success of nascent democracies. We identify leader education as a crucial determinant of the economic impact of democratization and also pin down underlying channels, thereby addressing a long-standing gap in the political leader literature highlighted by Besley et al. (2011): “The exact mechanism at work in explaining how leadership matters remains opaque. And one unresolved issue is to understand which growth-related policies are affected by leaders.” (p. 219).

We focus on Indonesia, which became the world’s third largest democracy after the fall of President Suharto in 1998 following more than 30 years of autocratic rule. A unique feature of Indonesia’s transition to democracy is that at the sub-national district level, the last mayor that had been appointed by the Suharto regime (“Suharto mayor”, henceforth) was allowed to finish his or her five-year term before a new mayor was democratically elected. At the moment of Suharto’s resignation the remaining time until the end of term of the Suharto mayor varied by district and is unrelated to district characteristics and trends, as Martinez-Bravo et al. (2017) and additional evidence in this paper show. This implies staggered and exogenous timing of democratization over the period 1999–2003, which we exploit via a difference-in-difference specification at the sub-national level. Thereby we improve identification relative to the existing democracy and growth literature which typically studies data on multiple countries where democratization is a result of country-specific and potentially unobserved characteristics. Leader education might still reflect other leader attributes or local factors, but we show that a wide range of such characteristics, as well as tests for pre-treatment trends, for selection on unobservables, and for close elections do not explain away the important role of leader education in shaping local policies and growth.

We find robust evidence that manufacturing-sector economic outcomes after democratization are worse in districts where the democratic leader does not have a college degree, irrespective of the last autocratic leader’s education level. Manufacturing represents 25% of Indonesian GDP and has been targeted as the principal growth engine by the national governmentFootnote 1, similar to many other developing countries. As mechanisms we identify increased taxation, less provision of physical infrastructure and more corruption under non-college educated mayors. While such mayors might simply be elected for having other priorities than supporting local manufacturing, we do not find that they spend more on items such as family welfare, health, housing, environment, religion, or education.

Data on manufacturing come from the annual census of manufacturing plants with 20 or more employees. These panel data allow us to study the impact of democratization on a given plant relative to ‘counterfactual plants’ in the same four-digit industry, province and year in districts that did not yet democratize, thereby refining identification compared to the existing literature which has focused on aggregate data such as national GDP. We find that in districts where the democratic mayor has no college degree, employment of incumbent manufacturing plants drops by 5% in the first few years after the mayor election. We also show that this effect is not only relative but also absolute, thus reflecting actual lay-offs. When the democratic mayor does have a college degree this negative impact is entirely offset, such that democratization has no effect over our sample period. We find no impact of democratization when we do not condition on mayor education, and observe similar patterns for plant revenue and total factor productivity.

Once democracy is established, leader education appears to lose relevance: the education level of the second democratic mayor does not have a statistically significant impact on local manufacturing, no matter the education level of the first democratic mayor.Footnote 2 We also find that the negative employment effects under first democratic mayors without college education equally hold when the last Suharto mayor has no college degree, such that the local mayor education level is unchanged. These additional findings show that leader education matters particularly as a country democratizes, and perhaps more generally during times of political or institutional change, which is a novel result in the literature.

Although the timing of democratization is exogenous, a potential concern is that democratic leader education is endogenous to district characteristics and developments or other leader attributes. In this regard we show that among a comprehensive set of mayor- and district characteristics that might determine the impact of democratization, the only robust driver of democratic mayor education is the education level of the Suharto mayor.Footnote 3 Our results are robust to including this and all of our other controls, which rules out for example that low leader education is simply a result of democratization paired with limited education of the local population.Footnote 4 We also show that (1) prior to democratization, manufacturing employment exhibits common trends across districts that later elect a college-educated mayor and those that do not; and (2) our results are robust to evaluating the impact of democratization and mayor education relative to a restricted set of control districts that elect a mayor with the same education level later, where unobserved factors are likely more similar. These two findings clearly speak against the presence of confounding developments at the local level. Finally, we apply the method of Oster (2019) and analyze close elections of the second democratic mayors (vote share data for the first democratic mayors are unavailable) to corroborate that the threat of endogenous leader education is very limited in our setting, if at all present.

We identify several channels through which local manufacturing performs worse under non-college educated leaders. First, using plant-level data on annual payments of indirect taxes, fees and levies, we find that the local tax incidence on manufacturing generally increases after democratization, but it increases twice as much under mayors without college education. However, we do not find evidence that total or social welfare expenditure increases, suggesting that not all extra revenue benefits the district. We also show that large plants, exporters and capital-intensive plants experience both a larger rise in the tax incidence and a larger drop in employment under mayors without a college degree. The taxation channel thus provides one reason why on average we do not find positive effects of democratization, contrary to recent cross-country level evidence (Acemoglu et al., 2019). In Sect. 2 we discuss the roots of this detrimental channel, which may be partly attributed to a specific design of democratization that has been adopted in other countries as well.

Second, we use longitudinal survey data to highlight another mechanism: after democratization the local business community perceives a significant deterioration of both availability and quality of local physical infrastructure, and the effect is driven by districts with non-college educated mayors. Combined with our taxation results, these findings suggest that after democratization less educated leaders enact worse policies at a higher cost to local manufacturing.

Third, we identify corruption as a mechanism: a novel mayor-level dataset that we hand-collect reveals a negative and statistically significant correlation between a democratic mayor’s education level and involvement in official corruption cases. All documented mechanisms are highly relevant because taxation, infrastructure and corruption are often cited among the most serious business constraints in developing and emerging economies, including Indonesia. The mechanisms are also internally consistent: more corruption can explain how a larger increase in taxation is accompanied by less infrastructure provision and no additional expenditure on other items. In the context of democratization, our results on mechanisms are not only in line with the finding of Keefer (2007) that young democracies engage in excessive rent seeking, under-provide public goods and are relatively corrupt, but they also highlight (insufficient) education as a driver of these issues.

Finally, a question that follows from both our results and those of related studies is through which underlying channels education affects leader behaviour. Based on evidence from different strands of the broader education literature, we discuss that a heavier tax burden, the neglect of infrastructure, and corruption by less educated mayors likely reflect a weaker understanding of the underlying costs and thus less competence. Moreover, these issues may be explained by different beliefs and values or other factors such as a more myopic attitude of less educated leaders towards their career.

1.1 Related Literature

We build on a large body of work that analyzes the impact of democracy on growth and finds mixed results overall (Helliwell, 1994; Barro, 1996; Tavares & Wacziarg, 2001; Rodrik & Wacziarg, 2005; Persson & Tabellini, 2006; Papaioannou & Siourounis, 2008; Doucouliagos & Ulubaşoğlu, 2008; Bates et al., 2012; Murtin & Wacziarg, 2014; Madsen et al., 2015; Acemoglu et al., 2019). We contribute to this cross-country literature by improving identification via our subnational approach that exploits random timing of democratization and by showing that the economic success of a nascent democracy depends on the education level of the newly elected leader.Footnote 5 By studying the transition from the last Suharto mayor to the first democratic mayor, we also contribute to a scarce literature on the immediate and short-run effects of democratization (Rodrik & Wacziarg, 2005; Acemoglu et al., 2019).

Our paper also adds to a small quantitative literature analyzing the Indonesian democratization process. Martinez-Bravo et al. (2017) show that the longer the Suharto mayor stays in power, the worse are governance outcomes after democratization, which is attributed to elite capture. Hallward-Driemeier et al. (2021) find that the disruption of political connections to Suharto due to democratization leads to more competition in manufacturing sectors disproportionately exposed to cronyism.Footnote 6 While these papers also exploit the staggered nature of the democratic transition, they do not shed light on the fundamental role of leader characteristics in shaping the outcomes of democratization.

Beyond democratization, our paper relates to a literature studying the effect of political leader education on diverse (socio-)economic outcomes (Dreher et al., 2009; Besley et al., 2011; Carnes & Lupu, 2016; Martinez-Bravo, 2017; He & Wang, 2017; Pertuze et al., 2019; Brown, 2020; Lahoti & Sahoo, 2020; François et al., 2020). Our contribution is to both study growth effects of leader education and identify underlying mechanisms, and to highlight the importance of leader education during a political transition. More broadly, our paper relates to a literature on leaders and growth which does not focus on leader education (Jones & Olken, 2005; Yao & Zhang, 2015; Easterly & Pennings, 2020) and to studies analyzing the effect of CEO education (Chevalier & Ellison, 1999; Bertrand & Schoar, 2003; Beber & Fabbri, 2012; Miller et al., 2015; King et al., 2016). Finally, we add to a body of work highlighting the importance of political leader characteristics other than education, such as gender (Chattopadhyay & Duflo, 2004; Clots-Figueras, 2011, 2012; Brollo & Troiano, 2016), nativeness (Hodler & Raschky, 2014), age (Yao & Zhang, 2015; Alesina et al., 2019), previous occupation (Dreher et al., 2009; Beach & Jones, 2016; Neumeier, 2018), prior experience in office (Freier & Thomasius, 2016), and heroic credentials (Cagé et al., 2021).

The remainder of the paper is structured as follows. Section 2 discusses the context of democratization in Indonesia, Sect. 3 data and key variables and Sect. 4 our empirical strategy. Section 5 presents our results and Sect. 6 concludes.

2 Background

President Suharto’s regime lasted from 1965 to 1998 and was characterized by tight control of Indonesian citizens and opposition parties. Following the Asian financial crisis and the disclosure of several corruption cases, Suharto was forced to step down on 21 May 1998 amid nationwide protests. A transitional government led by Suharto’s vice president Habibie assumed power and set the scene for the first free democratic elections since 1955, held on 7 June 1999. The main opposition party PDI-P clearly won these elections, followed by Suharto’s party Golkar which continued to represent the autocratic style of his regime and served as a pool for former members of the military and the bureaucracy (Hadiz, 2010). Besides the national parliament and president also provincial and district parliaments were elected, and the new district parliament (DPRD) was responsible for electing a new district mayor.Footnote 7 However, this (indirect) democratic mayor election only took place once the last mayor that had been appointed by the Suharto regime finished his or her five-year term.Footnote 8 This creates variation in the timing of first democratic mayor elections: in districts where the last Suharto mayor was appointed in the second half of 1994, the local parliament could elect the mayor within months after the legislative election of 1999, while in other districts the Suharto mayor could stay in office until as late as the beginning of 2003. Starting from 2005, mayors were directly elected by the district population once the five-year term of the incumbent had expired. Both before and after the fall of Suharto, mayors have been entitled to serve a maximum of two terms. Some Suharto mayors could therefore be elected as first democratic mayor, which happened in nine districts in our sample.

The mayor position entails a considerable amount of authority, in particular over local policies, regulations and the district budget (Martinez-Bravo et al., 2017). Although Law 22/1999 grants the local parliament the right to disapprove the district budget and regulations proposed by the mayor, and to reject the mayor’s annual accountability speech, this has not occurred frequently in practice (Hofman & Kaiser, 2006). In line, Von Luebke (2009) finds that mayors rather than citizen groups or local parliaments tend to initiate policy.Footnote 9 Mayors have thus been the main driver of local governance outcomes after the fall of Suharto. For these reasons, we adopt the notion of Martinez-Bravo et al. (2017) that democratization at the local level was triggered by the mayor election rather than the 1999 legislative elections.

The process of democratization was accompanied by decentralization, which Indonesia implemented nationwide on 1 January 2001. The country thereby pursued a similar strategy as several other developing nations across Latin America, South Asia and sub-Saharan Africa, particularly during the “third wave” of democratization after the 1980s. The motivation is that decentralization has the potential to promote democracy, participation, and empowerment at the local level (Kulipossa, 2004). The Indonesian decentralization laws transferred a substantial amount of power from the central government to the districts, largely bypassing the provincial level (see e.g. Jones, 2004).Footnote 10 This empowered local parliaments but also strengthened the mayor position, for example in the field of public goods provision (Hofman & Kaiser, 2006).

In post-decentralization Indonesia, the principal source of revenue for districts are non-earmarked transfers from the central government. The largest transfer (“DAU” = General Allocation Fund) is allocated based on local population, area, poverty rate, and other factors, and is set at 25% of central government domestic revenue in total (Brodjonegoro, 2004). The larger scope of action for mayors and the discretion over the use of transfers implies that decentralization is a key ingredient in creating a link between democratic mayor characteristics and the local success of democratization. We therefore design our empirical strategy so that our coefficients capture the impact of local democratization conditional on decentralization being in place. Our approach also isolates the impact of democratization from direct effects of the implementation of decentralization in 2001 (see Sect. 4).

While allowing discretionary use, the predominance of central government transfers as source of revenue also reflects that the fiscal decentralization law 25/1999 “continues the reluctance to give local governments any meaningful ability to raise local revenue” (Brodjonegoro, 2004, p. 129). Indeed, the official locally derived revenue (“PAD”) made up less than 10% of the local budget for 87% of districts in 2002 (Brodjonegoro, 2004). Many district governments have expressed their dissatisfaction about too low funding to promote regional development, especially in relation to new infrastructure provision (Brodjonegoro, 2009). Led by the powerful democratic mayor, local governments have frequently used this perceived lack of funding to justify the introduction of new local taxes and levies, which have been described as “illegal and disruptive” (Brodjonegoro, 2009, p. 207) and “distorting” (Ray, 2009, p. 151).Footnote 11 This matters particularly for manufacturing since “the easiest targets for these new additional revenues are unfortunately the local businesses that seem to be powerless against this challenge” (Brodjonegoro, 2009, p. 207). Similarly, Hofman et al. (2009) highlight the “high relative importance of political factors” for the local business climate and point out that “manufacturing in particular is prone to illegal levies, either by government officials or the surrounding community” (p. 110). “Illegal exactions” are in turn the most commonly cited factor that negatively affects the local business climate in a 2002 survey of companies (see Ray, 2009, p. 164).Footnote 12 The business community has further listed policy uncertainty, “demands by inexperienced local governments empowered by decentralization”, and corruption as serious constraints (Dhume, 2004, p. 66).Footnote 13

Our findings confirm the view of the above-discussed literature that the new and illegal exactions had a detrimental impact, rather than help stimulate local development: while manufacturing plant-level payments on indirect taxes, fees and levies rise with democratization, we observe no increase in total development expenditure or relevant sub-categories at the district level. This suggests that at least parts of the extra tax revenue served to increase mayors’ personal rent. This interpretation is consistent with the analysis of Lewis (2003), which “offers no support for the contention that regional governments create new taxes and charges because they lack fiscal capacity.” (p. 187; see also Hadiz, 2010). Most importantly, we contribute to this discussion by highlighting that a key local driver of excessive taxation, corruption, and insufficient focus on infrastructure is the education level of the first democratic mayor. This link echoes in a 2003 statement of Indonesia’s minister for Administrative and Bureaucratic Reform, where he argues that the majority of civil servants are “under-educated” and less than half “know what they are doing and do their jobs properly” (Webber, 2006, p. 408).

3 Data

3.1 Main variables and data sources

Our key data ingredients are information on the district-specific timing of the first democratic mayor election, mayor education level data, and plant-level manufacturing data. Table 1 reports descriptive statistics.

We obtain information on election timing and mayor education from the data repository of Monica Martinez-Bravo and Andreas Stegmann (Martinez-Bravo & Stegmann, 2018). The source distinguishes the education categories ‘Less than Bachelor degree’, ‘Bachelor degree’, ‘Master degree’ and ‘PhD degree’. We compute a dummy variable College Degree which equals one if the democratic mayor holds at least a bachelor degree and zero otherwise. Law 22/1999 requires mayors to have completed junior high school (Sekolah Menengah Pertama), which implies that all democratic mayors in our sample have at least nine years of schooling. The first democratic mayor has a college degree in 79% of districts in our final sample, while the last Suharto mayor has a college degree in 63% of districts.Footnote 14 These numbers are consistent with the result of Besley and Reynal-Querol (2011) that at the aggregate (country) level, democratization leads to an increase in leader education levels. We also use the data from Martinez-Bravo and Stegmann (2018) to control for democratic mayors’ age, gender, birth district, a dummy indicating prior work in the private sector, political party affiliation, and the education level of the last Suharto mayor, and exploit data on the field of study of college-educated democratic mayors. Selected data points on some variables are missing, but we are mostly able to fill the gaps through other sources.Footnote 15 We do not have information on vote shares in the first democratic mayor elections, but we use such data for the second democratic mayor elections in a robustness check (see Table OA11 in the Online Appendix (OA)).

The annual census of manufacturing plants (IBS) is collected and compiled by the Statistical Agency of Indonesia (Badan Pusat Statistik (BPS)) and has produced a panel of manufacturing plants that employ at least 20 employees in the particular year. We use mainly employment but also revenue, total factor productivity, and investment to measure performance. For our analysis of mechanisms, we employ data on plants’ reported payments of indirect taxes, fees and levies, and a proxy for bribe payments. We further use plants’ district location and sector information, which we translate into the ISIC Rev. 3.1 classification.Footnote 16

To analyze additional mechanisms we use data from the Regional Autonomy Watch KPPOD, which has conducted annual surveys in slightly varying sub-samples of districts across Indonesia from 2001 onwards. This effort has generated district-level data on the availability and quality of local physical infrastructure such as streets or telephone service, and on local institutional quality such as the consistency of regulations or law enforcement, as perceived by the local business community.Footnote 17 Data on institutions are collected through surveying local business actors and consulting a panel of experts, while for infrastructure KPPOD complements these sources with actual availability and quality data collected by the BPS. Data for the period 2002–2004 constitute a panel, which we exploit in our analysis.

We also hand-collect a novel dataset on mayor-level corruption cases using data from the Corruption Eradication Commission (KPK), the watchdog Indonesia Corruption Watch (ICW), and additional sources. For each democratic mayor in our sample, this dataset informs us whether the mayor was involved in an official corruption case (which is true for 50% of mayors) and which stage the case has reached (research, investigations, taken to court, convicted, or acquitted). We detail this dataset and the institutional background in Section OA4.3.

Finally, we collect data on additional district-level variables from different sources, specifically GDP per capita, population, population density, education of the working age population, 1999 election outcomes, religious fractionalization, city versus rural district status, and public expenditure items (see Section OA4).

3.2 Sample of districts, plants and years

We choose the time interval from 2000 to 2004 as our sample period. Thereby we analyze the transition from the last Suharto mayor to the first democratic mayor conditional on Suharto and the transitional government being out of power, and thus against the background of a constant national political setting. Starting in 2000 also ensures that 1999 election outcomes are predetermined controls rather than outcomes or endogenous variables. Since we focus on the first democratic mayor, we drop the year 2004 for districts where the second democratic mayor is elected in 2004.

The starting point of our district selection process is the set of 297 districts that existed at the end of 1997, and thus shortly before the fall of Suharto. First, we drop the five districts comprising the capital city of Jakarta, due to missing data and their different nature (see footnote 7). Following Martinez-Bravo et al. (2017) we then drop remaining districts that may endanger our identification strategy or conceptually do not allow to estimate our effect of interest, which is the impact of the direct transition from the Suharto mayor to the democratic mayor. Both issues apply to districts that split between the fall of Suharto and 2004; we therefore exclude the 87 districts that were involved in a district split (either as “parent” or “child”) over 1998–2004.Footnote 18 In 65 other districts, the Suharto mayor’s term expired between the fall of Suharto in May 1998 and the local legislative elections in June 1999, which implies that the Suharto mayor’s successor was selected by the transitional government. Since we can only speculate about the nature of these appointments, we exclude these districts from our sample. We further drop eight districts for which we do not know whether the mayor is selected by the transitional government or the 1999-elected local parliament. In 19 of the remaining districts, an interim mayor was installed to serve for a period of up to around one year between the last Suharto mayor and the first democratic mayor. Since the underlying reasons are unclear but appear district-specific and may represent confounding factors, we drop these districts as well. Based on the same reasoning, we exclude five districts in which the last Suharto mayor stepped down before the end of his or her five-year term and another four districts where the first democratic mayor stepped down prematurely within our sample period. Missing data on one district brings us to a set of 103 districts, of which 26 are cities and 77 are rural districts. Since two of these 103 districts do not have medium- or large-scale manufacturing over 2000–2004, and due to data availability and the chosen fixed effect structure in our specifications, our regressions include at most 96 districts.Footnote 19

Table 1 Summary statistics

4 Empirical strategy

We set up a difference-in-difference (DiD) specification with staggered treatment across space, exploiting that local mayor elections occurred in different years across Indonesian districts. Specifically, our empirical model is the following:

$$\begin{aligned} ln(Y_{ijkpt})&= \beta _1 PostElec_{kt} + \beta _2 [PostElec_{kt} \times CollegeDegree_k] \nonumber \\&\quad +\, \beta _3 ElecYear_{kt} + \beta _4 [PostElec_{kt} \times X_k] \nonumber \\&\quad +\, \mu _i + \omega _{jt} + \delta _{pt} + \epsilon _{ijkpt} \end{aligned}$$
(1)

where \(Y_{ijkpt}\) is outcome variable Y of manufacturing plant i in four-digit ISIC Rev. 3.1 industry j in district k in province p at time t. \(PostElec_{kt}\) equals one in the years after the democratic mayor election and zero otherwise; and \(ElecYear_{kt}\) equals one in the mayor election year and zero otherwise, and is mainly included to clearly separate the pre- and post-election period, given that manufacturing plant data are annual while elections happen throughout the year. \(CollegeDegree_{k}\) is a dummy that takes one if the democratic mayor in district k has a college degree and zero otherwise. \(X_{k}\) is a vector of mayor- and district-level control variables that are measured at the beginning of our sample period if they vary over time and are described further below. \(\mu _i\) are plant fixed effects, which also nest district fixed effects since we drop plants for which the census records two or more districts as location over our sample period.Footnote 20 These fixed effects control for (1) unobserved and time-invariant factors that influence the education level of the first democratic mayor and local manufacturing characteristics; and (2) any difference in manufacturing characteristics across the groups of districts that differ in terms of the democratization year.Footnote 21\(\omega _{jt}\) are four-digit industry-times-year fixed effects and \(\delta _{pt}\) are province-times-year fixed effects. These fixed effects control for example for the fact that Indonesia decentralizes in 2001 and the possibility that decentralization has a different impact across industries or provinces in Indonesia. We cluster standard errors at the district level.

\(\beta _1\) captures the effect of the democratic election of a mayor without college education, while \(\beta _2\) captures the differential impact of democratization when the newly elected mayor does have a college degree. Given our fixed effects structure, the effects captured by \(\beta _1\) and \(\beta _2\) are relative to plants in the same four-digit industry, province and year. In the case of \(\beta _1\) these ‘counterfactual plants’ are located in districts that did not yet democratize, while for \(\beta _2\) they are located in democratized districts in which the democratic mayor has no college degree. Such that \(\beta _1\) and \(\beta _2\) indicate effects conditional on decentralization being in place rather than (weighted) average effects across the pre- and post-decentralization period, we drop the year 2000 for the five districts where the mayor election occurred in 1999.

There are three identifying assumptions that must hold such that \(\beta _1\) and \(\beta _2\) are unbiased estimators of the described effects. The first is that the timing of the democratic mayor election is as good as randomly assigned across the districts in our sample. Athey and Imbens (2022) show that given random treatment timing in a staggered DiD setting, the standard DiD estimator is an unbiased estimator of a weighted average causal effect. Under the additional assumption of no anticipation effects—which we show to be valid in Table OA10—this average effect is conceptually meaningful, as all individual effects involve switching from not being treated to being treated. The random timing assumption is plausible for several reasons. In all districts in our sample, the timing of the first democratic mayor election is determined by the term end of the last Suharto mayor. This term end is a function of the timing of previous mayor terms, which in turn is determined by different accumulations of early term ends since the latter part of the Dutch colonial period, be it for health or other reasons. Based on this setting, Martinez-Bravo et al. (2017) conclude that the appointment timing of the last Suharto mayors—which determines the election timing of the first democratic mayors in our sample – is plausibly as good as randomly assigned. As supporting evidence, the authors show that the appointment timing of the last Suharto mayor is uncorrelated with the level of a wide range of district-level variables (see their Appendix-B Table 3). We complement these findings by showing that there is no correlation between the election year of the first democratic mayor and the level and growth rate of manufacturing outcomes at the district level prior to Suharto’s fall (see Table OA8). Furthermore, we corroborate the validity of the first identification assumption by showing that prior to democratization, manufacturing employment exhibited parallel trends across districts with different democratic mayor election years (see Table OA9). Finally, our results are robust to applying the estimator of De Chaisemartin and d’Haultfœuille (2020), which is preferred if there are both heterogeneous treatment effects and the timing of the democratic mayor election is not as good as randomly assigned (see Table OA10).

The second identification assumption is that conditional on our controls, democratic mayor education is exogenous to time-varying factors that impact local manufacturing. If democratic mayor education is solely determined by the composition of the local parliament elected in 1999, then this assumption is valid because the election results are a time-invariant factor captured by district fixed effects. If there are unobserved variables that affect mayor education even conditional on the 1999 election results, then these may be at least partly captured by the included province-times-year fixed effects and/or industry-times-year fixed effects. More importantly, we show that prior to democratization, manufacturing employment exhibits parallel trends across districts that later elect a college graduate as first democratic mayor and those that do not (see Table OA9)—which provides direct empirical support for the assumption’s validity. Event study graphs (see Figure OA1) complement these results by illustrating that both in districts that democratize under a college-educated mayor and those that do not, there are no significant trends in employment prior to democratization. We also show that the estimated impact of the second democratic mayors’ education level is no different when we focus on close elections (as discussed, vote share data for the first mayor elections are unavailable), which provides indirect support for the unbiasedness of our main results (see Table OA11). Our results are further robust to evaluating the impact of democratization and mayor education relative to a restricted set of control districts that elect a mayor with the same education level later, where unobserved factors are likely more similar (see Table OA10). Finally, we apply the recent method of Oster (2019), which evaluates robustness to omitted variable bias by analyzing the relative movement of the treatment coefficient and R-squared upon the inclusion of controls, and obtain reassuring results (see Section OA2). All these findings underpin the validity of the second identification assumption.

The third identification assumption is that conditional on the controls in vector \(X_k\), democratic mayor education is exogenous to (time-varying or fixed) variables that determine the impact of democratization on local manufacturing. We therefore include an extensive set of variables into \(X_k\), which are motivated by the existing literature and the Indonesian context. Democratic mayor-level controls are gender, age, and dummies indicating whether the mayor (1) works in the private sector pre-election; (2) is born in the district; and (3) is member of the Golkar party, respectively. We also control for whether the last Suharto mayor has a college degree. District-level controls are GDP per capita, average education level of the local working age population, population, population density, religious fractionalization, a city dummy, political competition in the local 1999-parliament (measured via a Herfindahl-Hirschman index using 1999 election vote shares), and a dummy indicating whether Golkar wins the 1999 elections. To avoid simultaneity and to make sure that these controls are predetermined (see “bad control problem”, Angrist and Pischke, 2008), we measure time-varying variables at the beginning of our sample period. Table OA12 shows that among the mentioned controls, only Suharto mayor education significantly and consistently correlates with democratic mayor education across different specifications. The vector \(X_{k}\) therefore includes only this variable in our baseline specification, while in robustness checks we add a separate interaction with all controls (see Tables OA13–OA15). Given our rich set of controls, the result that most of them do not correlate with democratic mayor education and do not affect the local success of democratization, and the discussed evidence based on vote share data and the method of Oster (2019), we are confident that the third identification assumption holds as well.

5 Results

5.1 Democratization, mayor education and manufacturing outcomes

To analyze real effects of democratization and democratic mayor education in the manufacturing sector, we estimate Eq. (1) for the number of employees, revenue, total factor productivity, investment, and the wage bill divided by the number of employees as dependent variables. Our main focus is on employment, which we analyze in Table 2. In column 1 we estimate Eq. (1) without the interaction terms, and find that the average impact of democratization on manufacturing employment is not significantly different from zero. However, column 2 shows very different results depending on the education level of the newly elected mayor. The marginal effects at the bottom of column 2 show that employment is unaffected in districts with college-educated mayors, while employment significantly drops by around 5% after the election of mayors that do not have a college degree.Footnote 22 The results are highly robust to controlling for any potential effects of Suharto mayor education after democratization (column 3).Footnote 23 In column 4 we test whether the effect of democratic mayor education depends on the Suharto mayor’s education level.Footnote 24 The results show that this is not the case: the election of a college-educated mayor has no employment effects both when the Suharto mayor has a college degree and when he or she does not (see the bottom two marginal effects in column 4), and the negative effect of electing a non-college graduate is large (see first marginal effect) and not significantly different (see third coefficient) when the last Suharto mayor is also not college-educated. This shows that our main results do not merely reflect the effect of a change in leader education irrespective of democratization. We explore this finding further by analyzing the effect of the second democratic mayor’s education level on manufacturing employment. We do so over the period 2004-2009, thus after all districts elected their first democratic mayor and before districts elected their third democratic mayor. The results are reported in Table OA6 and show no statistically significant change in manufacturing employment as a non-college graduate (or a college graduate) is elected as second democratic mayor, irrespective of the first democratic mayor’s education level. Since in 39 out of 76 districts in this sample the first democratic mayor is re-elected for a second term, we cannot rule out that the absence of significant results is due to limited variation in mayor education during this second period; however, taken together with the results in column 4 of Table 2, these findings suggest that leader education matters particularly as a country democratizes (and decentralizes), and perhaps more generally during times of political or institutional change. This is a novel result in the literature.

In Table 3 we study other manufacturing outcomes. Manufacturing revenue (columns 1–2) and total factor productivity (columns 3–4) significantly fall after democratization if the democratic mayor has no college degree, while these variables are unaffected if the mayor does have a college degree. The magnitude of the revenue reduction is strikingly high at around 15%. Plant investment (columns 5–6) does not significantly change with democratization, and there is no heterogeneity with respect to democratic mayor education. While speculative, the absence of a significant investment reduction under non-college educated mayors might be explained by survey evidence that “uncertainty in doing business locally has been increasing since 1999” (Brodjonegoro, 2004, p. 130), thus already before democratization in most districts. The election of the new mayor might have decreased this uncertainty and thereby stimulated investment, while the negative effects underlying our results on employment, revenue, or TFP might have offset such a positive impact. The results of columns 7–8 of Table 3 suggest that if anything, the wage bill divided by the number of employees falls rather than rises after democratization.

Table 2 Democratization, mayor education and manufacturing employment
Table 3 Democratization, mayor education and other manufacturing outcomes

5.1.1 Relative versus absolute effects

Since \(\beta _1\) captures relative effects (see Sect. 4), our results are not informative on whether employment actually declines after the election of a non-college educated mayor or if employment growth remains positive but is reduced. To investigate this, we take the sample of column 3 in Table 2, keep districts with democratic mayors without college education, compute the average log employment at the plant level before and after the mayor’s election, take the difference of the two numbers and generate the mean across all 1,318 plants. This mean equals -0.051, which clearly indicates that employment falls also in an absolute sense.

5.1.2 Time dimension of effects

In Fig. 1 we employ event study regressions to analyze the time dimension of the effects on employment, revenue and TFP. We extend Eq. (1) with one lead and two lagged dummies relative to the year of democratization: two years before, one year after, and two or more years after, such that the estimated effects are relative to two excluded periods (one and three years before). This is necessary because all of our districts are treated eventually, see Borusyak and Jaravel (2017).Footnote 25 The graphs show that democratization under non-college educated mayors has an immediate impact, and that the effects increase over time and are thus persistent over our sample period.

Fig. 1
figure 1

Timing of effects. Notes The graphs are based on event study regressions. We estimate the following specification for the sample of districts with a college-educated democratic mayor (left panel) and the sample of districts without (right): \(ln(Y_{ijkpt}) = \beta _0 + \beta _1 ElecYear_{k,-2} + \beta _2 ElecYear_{k,0} + \beta _3 ElecYear_{k,1} + \beta _4 ElecYear_{k,\ge 2} + \mu _i + \omega _{jt} + \delta _{pt} + \epsilon _{ijkpt}\), where e.g. \(ElecYear_{k,-2}\) is a (lead) dummy that equals one if in district k the democratic mayor election occurs two years later. Dots indicate point estimates and lines indicate 90% confidence intervals based on standard errors clustered at the district level. The sample period is 2000–2004.

5.2 Mechanisms

What causes the drop in manufacturing performance under democratic mayors without college education, and why does it not occur under college graduates?

5.2.1 An increasing tax incidence

Given the high relevance of local taxes, fees and levies for doing business after democratization and decentralization (see Sect. 2), we start by analyzing the plant census variable “expenditure on indirect taxes”. It includes sales taxes, fees for business permits, the building and land tax (PBB), road use tax (SWP3D), import duties, custom fees, and other levies, except income and personal taxes. Given this broad definition the variable likely provides an accurate representation of the overall incidence of local taxes, fees and levies on manufacturing, and we simply refer to the variable as “indirect taxes”, “taxes” or “taxation” in the following.Footnote 26 The results displayed in the fourth panel of Fig. 1 and in column 1 of Table 4 show that after democratization, manufacturing plants pay significantly more indirect taxes per rupiah of value added. Column 2 of Table 4 reveals that the increase is significantly larger under non-college educated mayors. The magnitude of the effect under such mayors (a two percentage-point rise) is very large, considering that the average ratio of indirect tax payments to value added equals 3 percent (see Table 1). The results therefore provide a plausible explanation for the decline in employment and other real outcomes.Footnote 27

In order to test the robustness of this conclusion we also analyze different sub-samples of manufacturing plants. Table OA1 focuses on employment and companion Table OA2 on indirect taxes. The results are very reassuring: larger, exporting, and capital-intensive plants face greater employment cuts that are accompanied by higher tax increases, while other plants that face small employment changes also experience small deviations in taxation.Footnote 28

5.2.2 Decreasing quality of infrastructure

Mayors can also influence the provision and maintenance of local physical infrastructure, which in turn is important for manufacturing. We therefore regress the log of a district-level score of general infrastructure provided by KPPOD for the years 2002–2004 on democratization (see columns 3–4 of Table 4). We adjust Eq. (1) to the more aggregate nature of the data: we drop industry-times-year fixed effects and replace plant fixed effects by district fixed effects, but continue to include province-times-year fixed effects. Column 3 shows that the combination of availability and quality of local physical infrastructure significantly decreases after the election of the democratic mayor. This matches the numerous Indonesian news reports on deteriorating infrastructure and a lack of attention by local governments to improve the quality of public service delivery during the democratization process (Brodjonegoro, 2009). Column 4 shows that the negative impact is driven by mayors without a college degree. Since a depreciation of public infrastructure increases the cost of producing and/or transporting goods, this result likely provides an additional explanation for the poor performance of manufacturing plants under non-college educated democratic mayors.

In Panel I of Table OA3 we deepen our analysis by studying the individual components of infrastructure. Our results continue to hold for infrastructure availability and quality separately, and elements that may deteriorate or improve relatively fast such as “quality of telephone service” are affected more by democratization and mayor education. This is intuitive given the relatively short period of analysis.

5.2.3 Total expenditure and spending on other public goods

Are college-educated mayors better able to generate funding from higher levels of government, which enables them to spend more on infrastructure and implies a smaller need for local taxes? We do not find empirical support for this hypothesis: public expenditure by college graduates is not significantly higher (see Table OA4).

It is also possible that mayors without a college degree are simply elected for having promised policies that focus on other areas than supporting the local manufacturing sector. However, Table OA4 also shows that non-college educated mayors do not spend relatively more on local development.Footnote 29 Furthermore, in Table OA5 we analyze subcategories of district-level development expenditure and do not obtain evidence that mayors without college education spend more on non-business items such as family welfare, health, housing, environment, religion, or education. The result that large increases in indirect taxes, fees and levies under non-college educated mayors are not accompanied by more government spending is consistent with the hypothesis that these mayors are more corrupt. We investigate this potential link in the next subsection.

5.2.4 Local institutions and corruption

Having a democratic mayor with lower educational attainment could be related to worsening institutions and corruption, which may also affect the business environment.

In columns 5–6 of Table 4 we report the district-level effect of democratization and democratic mayor education on institutional quality over the period 2002–2004, as measured by KPPOD. The coefficients’ signs point in the same direction as our real outcome and infrastructure results, but they are not statistically significant. In Panel II of Table OA3 we study the individual components of institutional quality and find similar results: variables such as “consistency of regulations” and overall “law certainty” appear to score higher under college-educated mayors—possibly indicating that bureaucracies under such mayors write clearer rules—but the coefficients are largely insignificant.

We can also use the more granular plant-level data that are available for more years to study “gifts, donations and the like” (hadiah, sumbangan dan sejenisnya), which has been interpreted as a proxy for bribe payments.Footnote 30 The results in columns 7–8 indicate that democratization and democratic mayor education do not affect plant-level expenditure on gifts and donations per rupiah of value added.

Gifts and donations are at best an indirect indicator of corrupt activities by the local democratic mayor because such expenses are also a choice variable of the plant (Fisman & Svensson, 2007; Vial & Hanoteau, 2010), and plants might for example require some time to understand the susceptibility of a new mayor to bribes. Moreover, a newly elected mayor may be not corrupt, but in the short run be unable to detect and limit bribes that have long been extorted by Suharto officials inherited from the old regime. For these reasons, we hand-collect a novel dataset on mayor-level corruption involvement (see Section OA4.3 for details). In Table 5 we regress indicators for whether individual democratic mayors are cited in an official corruption case and the outcome of the case on mayor-level characteristics, including education. The results show that mayors without a college degree are significantly more often researched, investigated, declared defendant, and convicted of corruption (see columns 1–6). These results might partly reflect that mayors with a college degree are more able to hide corrupt activities, prevent a corruption case, or block a case from moving forward; however, columns 7 and 8 show that among mayors for which at least research on potential corruption is conducted, college-educated mayor cases are not more likely to be closed early in the process. The mayor-level corruption evidence is therefore overall consistent with the negative effects of democratization on local manufacturing under non-college educated mayors, and can explain the co-existence of higher taxation and worse infrastructure under such mayors.

Table 4 Mechanisms
Table 5 Mechanisms (continued): Mayor-level corruption involvement

5.2.5 College-degree field of study

In Table OA7 we test whether our findings on the different manufacturing outcomes are driven by a particular type of college degree (i.e. field of study). The results provide some indication that democratic mayors with a degree in the area of political science, administration, and government are better able to promote employment and revenue, but the evidence is less clear for TFP and indirect taxes. This suggests that the effects we find generally hold across all college degrees.

5.2.6 Deeper mechanisms: education and mayor behavior

An important question is why non-college educated mayors raise taxation by more, neglect infrastructure, and are more corrupt, thereby harming local manufacturing. Krcmaric et al. (2020) lay out four mechanisms through which biographical characteristics might affect political leader behavior: (1) competence and skill, (2) material interests, (3) beliefs and values, and (4) other’s perceptions. The authors point out that “education, particularly university-level education, is commonly considered a key formative experience that affects outcomes through all four mechanisms” (p. 137). Better education can thus lead to better governance through each of these channels. However, empirical evidence on these or other mechanisms is scarce, and often focuses on specific degrees rather than a college degree more generally (Flores et al., 2013; Nelson, 2017). We therefore discuss a broader education literature that goes beyond leaders, which sheds light on why leaders’ characteristics may affect taxation policy, infrastructure provision, and corruption combined.

Along the competence and skill channel, more educated mayors may better understand the detrimental effect of excessive taxation, worsening infrastructure and corruption on the manufacturing sector. While it is difficult to test these hypotheses directly, they receive support from a literature showing that education leads to higher cognitive ability—not only in the short term (Brinch & Galloway, 2012; Carlsson et al., 2015) but also decades later at an older age (Banks & Mazzonna, 2012). The abundant evidence that educated individuals earn higher wages and produce better economic outcomes further corroborates that education and competence are positively related.

Higher skills might also be reflected in the ability to assemble a better team around oneself. We test this hypothesis in our setting by studying the correlation between the education level of the first democratic mayor and the education level of the average civil servant in the same district at the time.Footnote 31 While mayors do not have full control over the local appointment of civil servants, the decentralization laws did grant them extensive rights to make decisions on the careers of village heads and other civil servants (Martinez-Bravo, 2014). However, the (unreported) results show no statistically significant correlation between mayor and civil servant education and the coefficient sign is not constant across different specifications. This is consistent with the hypothesis that “team assembly effects” are not particularly relevant.

Skill-based explanations may also interact with mayors’ material interests: based on a better understanding of the involved costs, more educated mayors might refrain from corruption, excessive taxation, or neglecting infrastructure with the goal of raising aggregate output and thereby also personal income. In a similar vein, college-educated mayors may implement growth-enhancing policies in order to increase their chance of re-election, especially if they have longer time horizons in mind than less educated mayors. The latter appears plausible since Warner and Pleeter (2001) and Falk et al. (2018) show that more educated individuals in the US and around the world are more patient, and Jung et al. (2021) even demonstrate a causal effect of education on patience using Indonesian data. Another material interest channel could be that college-educated mayors are less corrupt because money plays a smaller role in motivating them to run for office in the first place, not least because education is typically linked to wealthier backgrounds (Björklund & Salvanes, 2011).

Regarding beliefs and values, more educated mayors may for example be more altruistic, and therefore aim at larger aggregate output in order to increase the district population’s income. This hypothesis receives support from studies showing that conditional on personal income and other factors, education is positively correlated with charitable giving (Forbes & Zampelli, 2013), unconditional helping behavior (Westlake et al., 2019), and social engagement such as community service (Helliwell & Putnam, 2007). College-educated mayors might also refrain from corruption because they have more integrity, but studies analysing the correlation between education and honesty (Abeler et al., 2014; Hübler et al., 2018) do not support this hypothesis. Finally, more educated mayors might be less corrupt because college graduates typically belong to higher social classes in the Indonesian society (Booth, 2021), in which the detection of corruption involvement likely carries higher non-monetary costs—not least because social status itself is a function of what Tirole (1996) calls “collective reputations” (Galiani & Weinschelbaum, 2013).

5.3 Robustness checks

We perform and discuss a large range of robustness checks in the Online Appendix (see Section OA2), some of which we already mention in the discussion of our key identification assumptions in Sect. 4. For reasons of space and relevance, we mainly focus on manufacturing employment as outcome variable in these exercises. We start by discussing robustness checks that test the validity of our first identification assumption (Tables OA8, OA9, OA10) and then move on to checks that address the second (Tables OA9, OA10, OA11, and Figure OA1) and the third identification assumption (Tables OA11–OA16). We conclude by presenting robustness checks that address other potential concerns such as district splits and sample selection bias (Tables OA8, OA16, OA17). Our results are robust to this battery of tests.

6 Conclusion

We provide novel evidence that the education level of newly elected democratic leaders crucially affects the economic success of democratization at the local level. In terms of economic outcomes we focus on the manufacturing sector, a key growth engine particularly for developing and emerging economies for which we have highly granular plant-level panel data. Our results show that in Indonesian districts where the democratic mayor has a college degree, democratization has no effect on manufacturing performance, while the impact is significantly negative under mayors without a college degree. For identification, we exploit the unique feature that in Indonesia democratization exogenously occurred at different times at the sub-national district level over the period 1999–2003. Thereby we also improve identification relative to a large literature on the effects of democracy that uses cross-country data. We also pin down mechanisms: non-college educated mayors increase local taxes, fees and levies by more than mayors with a college degree, and also invest less in infrastructure. While it could be that mayors without a college degree have different priorities than supporting manufacturing, we find no evidence that they rather support items such as family welfare, health, housing, environment, religion, or education. Instead, we find that non-college educated mayors are more likely to be involved in corruption cases. The education level of the local leader is thus closely related to good governance.

Additional findings indicate that more leader education is most beneficial during a democratic transition, and perhaps more generally during times of political or institutional change. Overall, our study thereby makes an important contribution to both the literature on democracy and growth and the literature on the effect of political leaders on economic outcomes. In terms of policy, our results suggest that a college degree requirement for political leaders—which exists in Turkey, Azerbaijan, or Kenya, and is hotly debated in IndiaFootnote 32—are most useful during a democratization period. Via achieving better economic outcomes such as employment, leader education may also make democratic transitions more durable. Our results therefore contain important lessons for other countries that have or will transition to democracy, particularly for developing countries where weak governance and infrastructure constraints are more prevalent.