Family Matters in a Meritocracy: Networks, Civil Service Exams, and Officialdom in the Joseon Dynasty

Sok Chul Hong† Christopher Paik‡ Yangkeun Yun§

July 2019

Abstract How do family networks influence social mobility in a meritocracy? Climbing the ladder of success may be fraught with nepotism and corruption, especially in monarchies where connections can trump talent. A merit-based selection of government officials in such context may serve as a remedy to curb these negative outcomes. In this paper, we investigate the effects of family networks on successfully obtaining official positions during the Joseon Dynasty from 1392 to 1897 CE. The Korean kingdom implemented literary examinations intended to fill central official positions based on merit. Its comprehensive records on family ties, exam results and official positions span over 503 years, longer than any other such data under a single dynasty in the world to our knowledge, and offer researchers a unique opportunity to investigate the efficacy of merit- based selections of political elites under a monarchy. We use an individually linked database of successful candidates and their family members from the literary examination rosters and official position information. We find that those from more connected predecessors in the network had significantly higher likelihoods of obtaining high-level rank positions after passing the exams, even when conditioning on age and performance at the examination. In light of the persistent family network influence, we evaluate the efficacy of meritocratic selection of political elites under a monarchy, and changing relevance of family networks as reflective of state performance over time.

 This study was financially supported by Center for Distributive Justice at Seoul National University (Grant Number: 0405-20180014) † Professor, Department of Economics, Seoul National University; [email protected] ‡ Assistant Professor, Division of Social Science, New York University Abu Dhabi; [email protected] § Ph.D. student, Department of Economics, University of Connecticut; [email protected] 1

1. Introduction Throughout history, monarchies have existed as common forms of governance. Hierarchical order and class divisions have often characterized their structures, in which social mobility was limited and representation in government was reserved for only the powerful few with connections. In order to combat nepotism and corruption from within, the rulers sought out ways to select government officials based on merit. An examination system that screened talented and capable candidates would, in theory, serve as a remedy to curb these negative outcomes. Similar to the civil service examination system in China, the Joseon Dynasty (Korean dynastic kingdom, 1392-1897 CE) was a centralized bureaucratic state which gave opportunities for officialdom to those who succeeded in Joseon’s merit-based examination. Those who passed gwageo, the Civil Service Examination, formed the ruling class. The examination system was the most significant means of recruiting officials for major central and provincial government posts (Wagner, 1974). Among the different types of exams under gwageo, mungwa (the literary examination) in particular was the most selective and accordingly prized as ensuring the elite status in society.1 Did the merit-based selection process actually work? Family connections, particularly in the form of prestigious lineage of mungwa passers, could strongly influence the career paths of elites. According to Kyŏngguk taejŏn (the National Code) and its subsequent laws, a variety of promotion standards were applied (Kim, 2017). The assignment of official position was apparently based on not only the competence of specific candidates but also recommendations from other officials as well as the king, particularly for high-level officials. In other words, the road to eventual officialdom was likely met with subjective factors such as family connections and king’s favors, in addition to scholarly ability (Won, 2007; Kim, 2017). Passing mungwa thus might not have been a sufficient condition to obtain a high-ranking position in the court. In this paper, we investigate the effects of family networks in obtaining official positions

1 Gwageo were composed of four categories: (1) mungwa (literary examination); (2) mugwa (military examination); (3) japgwa (technical examination); and (4) saengwon jinsa (classics and literary licentiate examination). We focus on those who succeeded in mungwa because they represented the ruling class of Joseon as major state officials. Successful candidates of saengwon jinsa became qualified to enroll at Sungkyunkwan, the National Confucian Academy, which trained students for mungwa. They could alternatively be appointed as ninth-ranked junior officials, which were the lowest positions among the court officials. Those who passed mugwa (military examination) or japgwa (technical examination) were regarded as lower-class officials. Table A1 in Appendix A provides the official ranking system during the Joseon Dynasty.

2 during the Joseon Dynasty. We use an individually linked database of successful mungwa passers and their family members from both mungwa rosters and appointment records of government officials. We find that those who are central, i.e. having more connected predecessors in the network, had significantly higher likelihoods of getting high-level rank positions after passing the exams, even when conditioning on age and performance at the examination. Specifically, our centrality score for each successful mungwa passer considers whether the ancestors themselves also passed mungwa. Because only the successful candidates appeared on the exam rosters, the ancestors had records of their own only if they passed the exam. The successful candidates with more ancestors who passed the exams themselves thus would have more connections, and become central by definition. We capture this score by using the eigenvector centrality measure, which accounts for the number of each candidate’s ties as well as the connections of the ties themselves (Bonacich, 1972, 1987; Jackson, 2010). The measure allows one to capture how those connected to the candidate are themselves influential, and is often used to assess prestige and popularity (Cruz et al. 2017; Jackson, 2010). Under the Confucian tradition that dominated the morals of society, scholarly achievement was of utmost importance for the elites in Korea. In our case, a higher score would indicate a more academic, and thus prestigious and influential, family connection. Throughout five centuries of rule in which the merit-based exam was in effect, we find that family connection was a key factor in selection into officialdom. The estimates from our preferred specifications, which controls for family clan, king in rule, pre-exam social status, exam type, year-of-birth, and residence fixed effects among other, suggest that one standard deviation increase in the natural log of eigenvector centrality is associated with approximately a 4 percent point increase in the likelihood of being high-level officials.2 We find this result for the positions that are higher than or equal to the upper senior third rank, the upper echelon of political elites in Joseon. The case of Joseon is of interest for its commonalities with other monarchies, but also for its unique context and comprehensive historical records. Nepotism among political elites likely exists in any governance structure, and arguably remains more prevalent in non-democratic settings where political representation and accountability are limited. Throughout history, monarchies represented the most common types of government. They adopted and institutionalized mechanisms to curb the influence of family connections and sought to recruit talent, while still

2 The probability that a candidate reached high-level officials was 56% in our sample.

3 maintaining strict social hierarchies. One of these institutional inventions was the merit-based selection of government officials. First found in Chinese dynasties, other Asian countries including Korea and Vietnam adopted similar practices. In a broader context, exam-based civil servant selections were also later found in the British government and other European states, as well as the United States. Among the monarchies that adopted the examination system, the Joseon dynasty stands out as an invaluable case study offering comprehensive records of exam passers, their eventual career paths as well as family connections spanning over five centuries. For example, in addition to information on the history of passing exams and the ranks of official positions the ancestors had obtained, the data allow us to analyze family networks through both marriages as well as patrilineal connections.3 Marriage is an important underlying mechanism that influences social status, as it plays a key role in preserving status-relevant family groups (Clark, 2014; Shiue, 2016). We have information on the marriage relations of each candidate who passed mungwa, with profile information of the candidate’s maternal grandfather and father-in-law. To our knowledge, these records comprise the world’s longest continual data of such kind under single dynasty. Our paper is also the first study examining the eventual political ascension of successful examination candidates, and the importance of family networks that could compromise the principles of meritocracy.4 Our work relates to several strands of the literature. First, this paper contributes to the literature on the institutional selection process of political candidates. Dal Bò et al. (2017) for example show that democracy can produce competent and socially representative politicians, while Cruz et al. (2017) document that family connections still matter for electoral outcomes in a democracy, as they facilitate relationships of political exchange. Several works in the literature also look at the authoritarian context to find that official appointment is heavily determined by power hierarchies and loyalty concerns; the leaders tend to hire mediocre and loyal, non-

3 By mandate, each mungwa taker filled out information on his father, paternal grandfather and great-grandfather, as well as information on the maternal grandfather and father-in-law as well as foster father, if any. Surviving records only have these information on the candidates who passed the exam.

4 Even though many authors studying the Chinese contexts explore the civil examination system, to our best knowledge, no research analyzes entry into office for the successful candidates. See, for example, Kracke (1947), Ho (1962), Hymes (1986), Jiang and Kung (2016), Bai and Jia (2016).

4 threatening candidates for positions in the bureaucracy (Zakharov 2016; Egorov and Sonin 2011; Reuter and Robertson 2012). Our paper also focuses on political candidates and outcomes, but expands the scope of the literature by looking at the monarchy system commonly found in history as opposed to the contemporary democracy context. Our paper also relates to the literature on social mobility and its long-term implications. Some of the existing research find significant inter-generational persistence in wealth (Clark and Cummins, 2015), education (Clark and Cummins, 2014; Shiue, 2016), as well as exam success (Wagner, 1974; Hao and Clark, 2012; Jiang and Kung, 2016) due to predecessors’ social class standings.5 Our study provides further support for intergenerational persistence of socio-economic status, by showing that having connected predecessors indeed have an important role in obtaining the high-level court positions for the next generation, even in the presence of meritocratic institutions (i.e. mungwa). Finally, our paper contributes to studies on Korean political economy and economic history which, despite offering important insights with rich context and data, have been relatively underexplored in the broader literature. Taking a full advantage of the rich set of available historical records, we combine each candidate’s exam ranking, family connections and eventual ascension to officialdom for 4,227 individuals from 1396 to 1894, and cover essentially the entirety of the Joseon dynasty period (1392-1897). This exercise provides us with a complex narrative on how meritocratic institutions worked to increase social mobility under a monarchy. We organize this paper as follows. In Section 2, we summarize the historical background of the examination system and official rank positions in the court during the Joseon Dynasty. In Section 3, we introduce our data sources and key variables. Section 4 presents the results of baseline and alternative specifications. We discuss how do estimates vary in relation to the extent of family networks increasing over time in Section 5. Section 6 concludes.

5 Wagner (1974) points out that successful candidates of mungwa were released from about 750 family clans, with the 21 leading clans producing over 40% while 560 extremely minor clans producing only 10%. This may indicate that those who came from major family clans might have been in a favorable position to pass the grueling civil examination. However, there is a debate about whether the family clan was a group that had a single identity and cohesiveness. For instance, Jeonju Yi clan which produced 870 successful candidates (5.74% of total) was a complex group of 123 fractions divided in the late Joseon Dynasty (Baek, 2017). Thus, we need to inspect more narrow unit by focusing on the micro-level relationships among individuals, not on how certain paternal blood groups produced political elites.

5

2. Historical Background 2.1. Literary Examination Passing mungwa brought personal glory and honor to the family, as it was the official gateway to public officialdom in the Joseon era.6 However, passing mungwa was as difficult as finding a needle in the haystack. It was so competitive that it took 10-15 years on average to pass the exam and the average passing age was 34.3.7 Men who entered government service by passing mungwa thus could be expected to serve at important posts in principal departments (Lee, 1994). Mungwa was divided into regular exams and irregular exams. Siknyeonsi, the regular exam, was triennial, and the irregular exams included jeunggwangsi (augmented exam), byulsi (special exam), alsungsi (memorial exam of royal visitation to the Confucian hall), and so on. Though jeunggwangsi was one of the irregular exams, we regard it as a regular exam from now on since it was more similar to the triennial exam.8 Figure 1 summarizes the structure and selection process of mungwa. 9 The regular examination was implemented in the order of chosi (initial examination), hoesi (metropolitan examination), jeonsi (palace examination), and the confirmation by the king. Both chosi and hoesi consisted of chojang (first-round test), jungjang (second-round test) and jongjang (final test). These exams were meant to mainly test the applicant’s ability in writing composition and knowledge of the Confucian texts based on the Five Classics and the Four Books (Lee, 2003). A total of 240 successful candidates, including those from provinces (150 in total), the capital (40 from Hanyang, presently Seoul), and Sungkyunkwan (the National Confucian Academy) (50) were chosen in chosi, and they assembled in the capital for hoesi (Lee, 2008). Only 33 successful

6 Eumseo were positions specifically reserved for the merit subjects who did not have to take mungwa. However, these were also positions with limited opportunities for promotions and were generally considered inferior to those obtained through mungwa (Paik, 2014).

7 This indicates that studying for the literary exam might have been prohibitively costly since the average lifespan was estimated only 40 during the Joseon Dynasty (Paik, 2014).

8 Both siknyeonsi and jeunggwansi consisted of the three separate examinations and selected top 33 candidates. Since the regular exams and irregular exams have different characteristics regarding the process and purpose of the test, we control for different types of exams in our empirical analysis.

9 Irregular exams varied in their structures and processes, thus we only provide those of regular ones.

6 candidates were selected to compete in jeonsi to determine the respective rankings. At jeonsi, candidates wrote essays on a subject suggested by the king, and their administrative and political competence were evaluated accordingly. Both the examiners and the king determined the rankings of these final candidates (Won, 2019).10

[Figure 1 Here]

2.2. Official Position Assignment and Promotion Government officials in the Joseon Dynasty consisted of nine ranks (see Table A1 of the Appendix). Each rank was divided into jeong (senior) and jong (junior), and the posts above the sixth rank junior official were subdivided into sanggye (upper) and hagye (lower), for a total of 30 ranks. High-ranking officials above or equal to the third rank senior official title were collectively called dangsanggwan (palace-ascendable officials). The remainders, called danghagwan (palace- downward officials), comprised chamsanggwan (mid-level officials) who were higher than or equal to the sixth rank junior officials, and chamhagwan (low-level officials) who were lower than the sixth rank junior officials. Dangsanggwan officials were the ministers authorized to participate in discussions or parties with the king at palace halls (Yi, 2015). They were given important rights to vote on the administration, to recommend other officials, and to direct the military (Cha, 2002). On the other hand, chamsanggwan officials were in charge of the central administration as well as local government and implementation of duties, with possibilities of promotion to dangsanggwan positions (Cha, 2012). However, promotions from the low-level to the mid-level, and from the mid-level to the high-level were difficult to achieve because of the entailed rigorous screening processes (Lee, 1994). Table 1 shows that the initial placement of successful candidates for the court was based on the official status of the candidates (if any) and the final grades in jeonsi. According to Kyŏngguk Taejŏn (the National Code), the candidates at the time of passing the exam fit into one of the

10 The ranking by examiners were sometimes changed by the king. For example, Sung Jin who passed jeonsi as a third-place in 1465 was nominated as a first-place since King Sejo was highly impressed after reading his essay. In theory, the exam grading was done anonymously; however, there are debates among historians on whether the anonymity rule was always implemented or not.

7 following categories: saengwon (classics licentiate), jinsa (literary licentiate), yuhak (Confucian student); and those already holding official titles. Candidates placed in gapgwa (first-division) were able to become court officials immediately.11 On the other hand, those placed in eulgwa (second-division) and byunggwa (third-division) were not guaranteed actual posts, and only received official ranks; they were assigned as temporary officials until positions became available (Won, 2007). Therefore, there was a huge gap in the career path of civil servants according to whether being placed in gapgwa in jeonsi or not. Similarly, those already holding official titles acquired immediate promotion, but the opportunity also differed by the final grade in jeonsi.12

[Table 1 Here]

3. Prestige: Inheritance of Political Power

We apply and extend a simple model of Dal Bó et al. (2009). Let 푦푖 be a successful candidate

푖’s political power and 푘푖 be the amount of political capital available to him. Suppose that successful candidate 푖 has a predecessor (hereafter referred to as 푖’s father), whose amount of political power and capital are defined as 푦푖,1 and 푘푖,1, respectively. Assume that candidate 푖’s political capital 푘푖 is a linear function of father’s political capital and political power. That is, the political capital of the candidate 푖 is determined as follows

푘푖 = 훼푘푖,1 + 훽푦푖,1 where 훼 and 훽 are scalars. As in Dal Bó et al. (2009), we define political capital as any personal characteristic that is inherited within family and influences on political power, from raw talent and overall competence to human capital to name recognition. We assume that candidate 푖’s political power depends on the political capital

푦푖 = 훾푘푖 + 푣푖

11 Jangwon (first-rank candidate) were appointed as the sixth rank junior officials, and the second- and third- rank candidates (non-jangwon gapgwa candidates) were assigned as the seventh rank senior officials.

12 If mid-level officials passed mungwa, they were guaranteed to have promoted positions though they were not placed in gapgwa in jeonsi (Won, 2007).

8 where 훾 is a positive scalar and 푣푖 is a random shock. Thus, we can rewrite

푦푖 = 훾훼푘푖,1 + 훾훽푦푖,1 + 푣푖

From this equation, which is similar to the one derived by Dal Bó et al. (2009) in page 121, we find the effect of father’s political power and capital on candidate 푖’s political power. Since father’s political capital also receives influences from his predecessors’ (hereafter referred to as 푖’s grandfather) political capital (푘푖,2) and political power (푦푖,2), the process becomes

푘푖,1 = 훼푘푖,2 + 훽푦푖,2 and candidate 푖’s political power can be replaced to

2 푦푖 = 훾훼 푘푖,2 + 훾훼훽푦푖,2 + 훾훽푦푖,1 + 푣푖

In a similar manner, political power and capital of candidate 푖’s great-grandfather are inherited to the political capital of candidate 푖’s grandfather, and so on. Therefore, we get

3 2 푦푖 = 훾훼 푘푖,3 + 훾훼 훽푦푖,3 + 훾훼훽푦푖,2 + 훾훽푦푖,1 + 푣푖 … 푡 푡 푙−1 = 훾훼 푘푖,푡 + 훾훽 ∑ 훼 푦푖,푙 + 푣푖 푙=1 where 푡 denotes a generation distance from candidate 푖 to each ancestor. As 푡 → ∞ and 훼 is small enough, our expression for candidate 푖 ’s political power simplifies to

∞ 푙−1 푦푖 = 훾훽 ∑ 훼 푦푖,푙 + 푣푖 푙=1

∞ 푙−1 Notice that ∑푙=1 훼 푦푖,푙 is proportional to Katz prestige, which was proposed as an index of 푛 푙 status by Katz (1953), for successful candidate 푖 where 푦푖,푙 = ∑푗=1(퐴 )푗푖 and 퐴 is a real-

9 valued 푛 × 푛 adjacency matrix representing the network constructed by family members. 13 Therefore, our conceptual framework predicts that successful candidates with higher Katz prestige are in a better position to attain political power. In matrix terms, we can get the linear system of Katz prestige 푥 by multiplying 훼 to what ∞ 푙−1 ′ 푙 we derived, ∑푙=1 훼 (퐴 ) ퟏ, as follows

∞ ∞ 푥 = ∑ 훼푙(퐴′)푙 ퟏ = [∑ 훼푙(퐴′)푙 − 퐼] ퟏ = [(퐼 − 훼퐴′)−1 − 퐼]ퟏ 푙=1 푙=0 where 퐴′ is the transpose of the adjacency matrix 퐴, 퐼 is a 푛 × 푛 identity matrix, and ퟏ is a column vector of ones.14 The column sums of 퐴 (i.e. 퐴′ퟏ) give the numbers of direct influences made by family members (father, maternal grandfather, father-in-law, or foster father in our context). Similarly, the column sums of 퐴푙 (i.e. (퐴′)푙ퟏ) give the numbers of length-푙 influences from ancestors (Katz, 1953). The decay or attenuation factor 훼 gives higher weights to influences of shorter generation distance.15 Hence, it is a way of looking at all of the inheritances from ancestors to decedents in the family network and weighting them by generation distance (Jackson, 2010).

4. Data

13 According to Wasserman and Faust (1994), the terms “centrality” and “prestige” can be used separately when quantifying the importance or prominence of a node in a network: “centrality” focuses on evaluating a node without considering directionality (i.e. undirected networks), whereas “prestige” focuses on evaluating a node according to the ties that the node is receiving (i.e. in-edges in directed networks). Katz (1953) actually used the term “status”, but we use the term “prestige” following the convention of later works which are using “prestige” as a more general concept (Jackson, 2010; Wasserman and Faust, 1994).

14 Katz prestige is almost identical and perfectly correlated to Bonacich centrality, which was introduced by Bonacich (1987) as a direct extension of Katz prestige, and alpha centrality, which was suggested by Lloyd and Bonacich (2001) to solve the problem of eigenvector centrality in asymmetric networks. Please refer to Appendix B for the relationships between eigenvector centrality, alpha centrality, and Katz prestige. Also, when the generation distance (푡) is finite, Katz prestige becomes similar to the diffusion centrality which was proposed by Banerjee et al. (2013). However, the diffusion centrality focuses on the powers of out-edge, which is opposite from our setting, and moreover, we assume that a family tree can be extended infinitely. Thus, we do not apply the diffusion centrality in our study.

15 We have to make sure that 퐼 − 훼퐴′ is invertible, otherwise the linear system has no solution. Therefore, the decay factor 훼 should be chosen between 0 and 1/휆, where 휆 is the largest eigenvalue of adjacency matrix 퐴.

10

4.1. Sources Our empirical analysis is based on data drawn from two sources. The first is Jae-Ok Lee’s data that link individuals appearing in the Academy of Korean Studies (AKS)’s digitized mungwa rosters (Lee, 2018).16 The mungwa bangmok, or literary examination rosters, from AKS contains lists of all the successful candidates qualified for jeonsi taken at the palace over different time periods and exams. These digitized rosters include the following information for each candidate: name; post or title at the time of the examination; year of birth; ranking at jeonsi; father, paternal grandfather and paternal great-grandfather, maternal grandfather, and father-in-law; family clan; place of residence; and brief career highlights.17 Lee’s dataset links candidates with others based on their ancestral lineages. There are 47,308 nodes (14,634 successful candidates and 32,674 their family members) which altogether construct networks with 49,229 ties.18 Specifically, each node is directly linked to another through four types of ties (father, foster father, maternal grandfather, and father-in-law). In our network analysis, we also capture additional ties with other members of extended families, as long as they passed mungwa and shared common ancestors. For example, a candidate’s uncle would be in the network if he passed the exam; we could then match his grandfather as the same as the candidate’s great-grandfather. Figure 2 describes the partial networks that are directly or indirectly linked to Sa-Ahn Kang,

16 Jae-Ok Lee is a research fellow at the Academy of Korean Studies in charge of the digitized mungwa records. His data linking individuals based on family connections are available at http://dh.aks.ac.kr/~sonamu5/wiki.

17 Each exam taker had to fill out the application form requesting information on the three paternal ancestors, maternal grandfather, and father-in-law to enroll the exam. From these forms, we have information on the candidates’ family backgrounds at the time of the examination, and before they started their careers as government officials. The construction of the database initially started with the “Civil Examination Rosters Project (Munkwa Project)” by Edward W. Wagner and Jun-ho Song (Song, 2010). They were dedicated to digitizing mungwa rosters for about 40 years. The Academy of Korean Studies has since expanded the data and made them available to the public after taking over the project (Wagner, 1974; Lee, 2018). The mungwa database is available at the Academy of Korean Studies (AKS)’ Historical Figures Comprehensive Information System (http://people.aks.ac.kr/index.aks).

18 The adjacency matrix 퐴 of our family network (graph), with respect to all nodes, becomes the 47,308 × 47,308 zero-one matrix (i.e. unweighted graph) with its (푖, 푗)th entry,

1, if 푖 contributes to 푗′s status 푎 = { 푖푗 0, otherwise

We check the robustness of weighted graph in the analysis of Table 3.

11 who passed mungwa in 1542, in the dataset.19 For example, Kang’s biological father was On Kang, foster father was Ho Kang, grandfather was Yeong-Suk Kang, great grandfather was Hyeong Kang, maternal grandfather was Sik Park, and father-in-law was Gan Im. The names in bold represent those who passed mungwa themselves.

[Figure 2 Here]

Mungwa bangmok provides only limited information about the official positions of successful candidates after they pass mungwa. The rosters record only one of each successful candidate’s official positions of during one’s entire career. Thus, it is not possible to identify whether the recorded positions in mungwa bangmok are each successful candidate’s highest position. Furthermore, from the middle of the 18th century onward, records on candidates’ official positions are missing entirely altogether. To bridge this gap, we complement the information for officialdom by using a document containing a list of officials, titled Cheongsungo, or the Reference for the Uncorrupted Selection for High Officials. The document contains information on over 40,000 officials and is the most comprehensive data of all over the Joseon Dynasty. 20 We match individuals who appear on this list to those on Mungwa bangmok to investigate who was assigned which position after they passed the exam.21 Of the 14,638 successful candidates who passed mungwa during the dynasty, 5,738 officials are matched with those on the official list.22

19 Kang was initially placed in the ninth rank senior position, and later reached the third rank lower senior position. There are many other men who are linked to Sa-Ahn Kang indirectly; given the space constraint, we only present the linkages directly coming from Kang’s closest ancestors.

20 The Academy of Korean Studies provides the digitized database. There are 196 positions and we can identify 120 positions’ official ranks from the junior ninth rank to the first senior upper, altogether comprising 30 ranks as described in Section 2.2.

21 There are 15,151 successful candidates in Mungwa bangmok. Among them, 503 successful candidates passed mungwa more than once, so in total there are 14,638 unique individuals who passed the exam. Using the Universal Content Identifier (UCI) system created by the Academy of Korean Studies to code each individual mentioned in historical documents throughout the Dynasty, we exactly match individuals in Mungwa bangmok and Cheongsungo, at the same time also checking Lee’s network data for any duplicate entries and identical names.

22 The remainders, which are not matched between two sources, include both those who never obtained official positions after passing mungwa and those who obtained some positions but are missing in Cheongsungo. We use the matched sample as our base dataset, but later check the results by including the unmatched sample as a robustness test in Table 3 (NOT YET DECIDED TO INCLUDE THIS PART).

12

4.2. Variables Our main dependent variable is an indicator equal to one if the successful candidate’s highest official position was higher than or equal to the third rank upper senior position. As in Section 2.2, this was the dividing rank between those classified low- and mid-levels vs. high-level (palace- ascendable), the latter with authority to participate in discussions with the king at the palace halls (see Table A1 in the Appendix). It was of paramount importance for the government officials in the Joseon Dynasty to move from the mid-level to the high-level in order to obtain considerable political power. Therefore, the indicator of dangsanggwan is the variable that captures political power each candidate attained or not. Graphing the full family networks of all the candidates in our dataset (47,308 nodes with 49,229 ties), we evaluate how centrally connected each individual is in his extended family network. As discussed in Section 3, we use Katz prestige, which captures how well each individual has influences from other important people, rather than simply considering the number of links (Katz, 1953; Bonacich, 1972, 1987; Jackson, 2010).23 Using this measure, central individuals (or nodes) are those with connections from other well-positioned predecessors (Katz, 1953; Bonacich, 1972, 1987; Jackson, 2010; Cruz et al., 2017). This is our measure of family connection and is the main variable of interest. In our network setting, well-positioned family members for an individual candidate are those who themselves pass mungwa, have their own family members recorded in mungwa bangmok and thus provide more ties for the candidate. The family members who appear in mungwa bangmok not because they pass mungwa but because of family connection, on the other hand, do not have additional family records in mungwa bangmok and thus contribute less to the centrality of the candidate. To measure the political prestige inherited within a family, we exploit the Katz prestige for

23 For the details on network centrality measures, please refer to Appendix B and Jackson (2010). There are various centrality measures including degree centrality, betweenness centrality, closeness centrality, eigenvector centrality, alpha centrality, and PageRank. Each captures theoretically different statistics. In a broad category, degree centrality, eigenvector centrality, alpha centrality, Katz prestige, and PageRank centrality can be classified as degree-based centrality measures, on the other hand, betweenness centrality and closeness centrality are shortest-path based centrality measures (Freeman, 1972; Meghanathan, 2015). Since we focus on the power and prestige of predecessors (ancestors) and their influences to decedents rather than the distance from decedents to predecessors, exploiting degree-based centralities can be a right strategy in our context. We summarize the relationship between degree-based centralities in Appendix B.

13 directed graphs as shown in Figure 2 instead of undirected ones.24 Each candidate’s success in the court would have been influenced by his predecessors, rather than vice versa.25 A directed graph thus emerges as the appropriate framework to work with, especially since we want to predict the candidate’s eventual position assignment. Our prediction is that successful candidates with higher Katz prestige would have been in a better position to become dangsanggwan (high-level officials) as discussed in Section 3. Figure 3 shows the partial networks that correspond with Figure 2, in which the size of the circle is proportional to the node’s Katz prestige.26 We set the decay factor, 훼, as 0.3 in our basic estimation and check other decay factors (0.1 and 0.5) in the robustness test.

[Figure 3 Here]

There are clear advantages to using the network measure for overall family connections, rather than simply looking at whether ancestors passed mungwa or not. We are interested in maximizing the use of family data from the exam records and measuring the level of prestige that came with it in a quantifiable way. Using the network approach, we are not only able to identify the ties that individuals have, but also weigh each connection based on the ties that these connections themselves have. Furthermore, we are able to include multiple indirect connections outside the family members on the candidate’s records, and assess their contribution to the candidate’s overall level of connectedness. Finally, our key mechanism explaining the role of family connection on political careers is the level of prestige perceived by others and its subsequent

24 The Katz prestige works for directed networks in a more novel way than in undirected networks (Jackson, 2010).

25 We exclude information on the official positions of ancestors in a candidate’s network, and instead calculate our centrality measure based only on whether the ancestors pass mungwa or not. It is difficult to imagine that a successful candidate may influence the likelihood of his father passing mungwa, especially given the age gap and the years of preparation it takes. On the other hand, a father’s position in the court may indeed change during his lifetime by his son passing mungwa. We also assume here that since a typical marriage in Joseon was customarily pre-arranged by the parents rather then decided by the individual, the directed relationship with the father-in-law is also appropriate (as opposed to an undirected relationship in which the individual at the time of taking the exam somehow may affect the likelihood of the father-in-law’s success in the court).

26 Figure A1 in Appendix A describes the sample family network of our dataset. To simplify the graph, we restricted the nodes that have more than or equal to 5 ties including in- and out-edges. In this network, the number of nodes is 2,729 (5.71% of total) and that of ties is 2,391 (4.86% of total), respectively. In the graph, it is easy to check that those who achieved high-level rank positions (black dots) also tended have bigger circles (i.e. to be more centrally connected) than those without (gray dots).

14 influence, rather than the cultural transmission of scholarly aptitude passed down the families.27 For capturing the prestige mechanism, we argue that the network-based centrality measure, especially Katz prestige, is more suitable. In addition to the family connection variable, we also consider the candidate’s age upon passing hoesi (the metropolitan examination) and the final grade in jeonsi (the palace examination). These two variables together control for the candidate’s overall level of competence, and also reflect the level of “family human capital” that provided the know-how in passing the exam (Jiang and Kung, 2016). First, the age upon passing hoesi reflects the applicants’ ability in writing compositions and memorizing codified knowledge of the Confucian classics (Lee, 2003). The more competent candidates would be more likely to pass the exam at a younger age (Marsh, 1961; Jiang and Kung, 2016). Younger successful candidates, in turn, would have more opportunities to go up in the ranks than their counterparts. Second, the final grade in jeonsi captures the knowledge pertaining to statecraft beyond Confucian studies. This final exam tested on both administrative and political expertise, and was meant to screen those for senior appointments in the government (Jiang and Kung, 2016). The topic of the exam was open and allowed for highly subjective answers from candidates, who often had to give their opinions on thorny issues that the king himself confronted. As discussed in Section 2.2, the achievement of gapgwa in jeonsi was a starting point to become an important court official in a later political career (Table 1).28 Table A2 in the Appendix shows the summary statistics of the relevant variables.

5. Family Network, Prestige, and Political Power

27 In the Chinese civil service examination context, Jiang and Kung (2016) for example interpret the father’s educational attainment as a proxy for a family’s human capital. However, looking at only the influences of immediate ancestors is more close to degree centrality which is likely to fail to capture influences coming from longer distance. We also test this measure in our regression analysis (Table 2).

28 To compare the final rankings of candidates in jeonsi across time and exam, Jiang and Kung (2016) standardize the ranks as the following:

total number of passers in a given exam − actual ranking

total number of passers in a given exam with the value ranging in [0,1). However, as shown in Table 1, the association between the final ranking in jeonsi and official position assignment was step-wise rather than linear. Thus, we exploit the final “grade” (gapgwa, eulgwa, or byunggwa) instead of ranking in this study.

15

5.1. Baseline Estimation We start with simple graphical evidence of network effect in Figure 4. We compare the proportions of those who achieved dangsanggwan positions by the final grade in jeonsi and across Katz prestige. To construct this binned scatter plot, we first divide Katz prestige into ten equal sized-groups (deciles) and then plot the means of the y-axis variable within each bin against the mean value of Katz prestige (z-score) within each bin. The first thing we can find here is that candidates placed in gapgwa generally show a higher proportion of dangsanggwan within the same group of prestige. Also, candidates with higher Katz prestige were likely to enjoy a higher likelihood of being dangsanggwan in all grades. Finally, and most importantly, even though candidates started their career as eulgwa or byunggwa finalists, if they came from prestigious family (i.e. high Katz prestige), they acquired a higher proportion of dangsanggwan in average than those who were placed in gapgwa with low prestige. What we discover is that even after scoring high on the final stage of mungwa and in spite of subsequent advantages in the initial appointment, the eventual career ascension could have been reversed by family background.

[Figure 4 Here]

We now explore to quantify the impact of network connections on the successful candidate’s likelihood of obtaining high-level official positions. To do so we estimate a linear probability model of the following form:

′ ′ 푦푖 = 훼 + 훽푥푖 + 퐶푖 Γ + 푍푖 H + 휀푖 (1) where 푦푖 is an indicator equal to one if individual 푖’s highest official position was higher than or equal to the third rank upper senior position (i.e. belonging to the group of high-level rank officials).

푥푖 denotes the Katz prestige of individual 푖. 퐶푖 is a vector of competence controls including candidate 푖’s indicators of the final grade in jeonsi as well as the age upon passing hoesi. 푍푖 is a vector of identifiers including candidate 푖’s family clan, king in rule at the time of the exam, pre- exam status (Confucian student, classics licentiate, literary licentiate, or court officer), type of exam (regular exams and irregular exams), year-of-birth, and place-of-residence before the exam. Table 2 summarizes the baseline estimation results across different model specifications, reporting the coefficients and their standard errors clustered by family clans. In each specification,

16 we control for the family clan, king in rule, pre-exam status, year-of-birth, and exam type fixed effects. The estimation results strongly suggest that candidates with high prestige scores tend to have higher probabilities of becoming high-level officials. The coefficient estimate reported in column (1) for example indicates that one standard deviation increase in Katz prestige is associated with approximately 6 percentage points increase in the likelihood of becoming dangsanggwan. The estimates, reported in column (2), show that this effect is robust under within variations of the final grade in jeonsi. We additively control for the age upon hoesi in column (3) confirming the small decrease of the effect of prestige. Column (4), which is our preferred specification, adds the residence fixed effects to rule out the varying spatial impact of regional characteristics, suggesting that the likelihood of becoming dangsanggwan is higher by approximately 3.9 percent points with one standard deviation increase in the prestige score. For our prestige measure, we standardize the score so that one unit increase in the measure is equivalent to one standard deviation increase. In other words, if one were to take the minimum prestige score in our sample and increase it to the maximum (about an increase of 6.8 standard deviations), the likelihood of becoming dangsanggwan would increase by 26.5 (= 6.8 × 3.9) percentage points. This magnitude amounts to approximately 48% of the sample average of 0.55. Overall, the estimated result is robust across different regression specifications: Katz prestige, final grade in jeonsi, and age upon passing exam are highly associated with the probability of becoming dangsanggwan positions. Further, we closely follow Jiang and Kung (2016)’s approach in columns (5)-(7) by additionally controlling indicators for whether the ancestor passed mungwa. The findings yield the robust results, in line with our main findings even when considering the family’s immediate human capital.

[Table 2 Here]

5.2. Robustness Checks We first check how our results vary across different cutoffs for our outcome variable. If prestige inherited within family really mattered for eventual political power, the impact would have been greater especially when obtaining a high-level position (third rank upper senior position or higher) because of the political powers and privileges that it entailed. We test this hypothesis with the same approach as the baseline estimation. Specifically, we use equation (1) but replace the dependent variable with different cut-offs. We report the coefficient estimates for Katz prestige 17 in panel (a), estimates for the gapgwa dummy in panel (b), and estimates for the age upon passing hoesi in panel (c) with their 95 percent confidence intervals in Figure 5.29 We observe several important trends emerging from Figure 5. First, the estimated coefficients of Katz prestige are relatively stable across different cut-offs in rankings until US3 (third rank upper senior position or higher). The results from this specification can be interpreted as a placebo experiment to our baseline. Second, we find that from the third rank upper senior position cutoff, the estimates rapidly deviate in magnitude. The effect on position assignment for low- and mid- level positions appears systematically different from the high-level positions. Third, the magnitudes in gapgwa dummy are quite consistent throughout different cutoffs but show a peak in the third rank upper senior position cutoff. This may also indicate the importance of statecraft competence and initial position at attaining power. Finally, we observe that the age effect becomes larger in magnitude and statistically significant for higher-rank positions. Competence in the form of learning codified knowledge and early success in exam appears to matter.

[Figure 5 Here]

In Table 3, we present a number of further tests that include 1) other regression models; 2) other decay factors; and 3) different weighting for family link types. First, one may argue that the linear probability model can produce inaccurate estimates (Johnston and Dinardo, 1997; Wooldridge, 2002). Columns (2) and (3) show the results of logit and probit regressions, respectively. To make them comparable using the probability scale, we report the average marginal effects across the values of each variable. Compared to the baseline result in column (1), we find that the results are strikingly similar. These show that our baseline linear probability estimation does not cause bias. Next, we use different decay factors of Katz prestige replacing the base value (훼 = 0.3) to 0.1 in column (4) and 0.5 in column (5) to identify whether a particular decay factor causes biases to the baseline estimation.30 The estimated coefficients in columns (4) and (5), which are similar

29 As an example, when the dependent variable is the dummy for obtaining the fourth rank junior position or higher, we plot the estimation results at J4 on the x-axis in each panel. The results plotted on the third rank upper senior position or higher are marked as US3, and are the same as those in column (4) in Table 2.

30 Small values of decay factor (훼) heavily weight the local structure, whereas large values take more into account the position of individuals in the structure as a whole (Bonacich, 1987).

18 to those of the baseline estimation, can rule out those concerns. From column (6) to column (8), we test whether different types of family ties cause inaccurate estimation. Specifically, in our network, there are four types of ties (father, foster father, maternal grandfather, and father-in-law), but each type is based on different generation distances and different characteristics. For instance, a link from father is the one-generation length and a link from maternal grandfather is the two-generation length, but our network structure treats them equal. In addition, the influence of father may be quite different from that of father-in-law or foster father. Hence, we need to reflect the distinct features of ties by using different edge weights. In column (6), we replace the edge weights of maternal grandfather from 1 to 1/2. Additionally, we replace the edge weights of father-in-law from 1 to 1/2 in column (7) and the edge weights of foster father from 1 to 0 in column (8), respectively. In other words, we give fewer weights to the influences from maternal grandfather and father-in-law, and no weights from foster father.31 The results in columns (6)-(8) dispel such worries.

[Table 3 Here]

6. Family Network and Political Stability Our results suggest that successful candidates with higher prestige through family networks had significantly higher likelihoods of getting high-level official positions after passing their exams. In this section, we examine whether the changing importance of family networks and exam performances in officialdom reflects macro trends in the political stability of the Joseon Dynasty. If family connections mattered as much or even more than success in mungwa for high-level official positions, one could argue that the Dynasty suffered from weakening principles of meritocracy and plausibly poor state performance. One of the variables that we can compare against key variables is the number of exiled officials over the 500 years across 26 kings (please refer to Table A3 in Appendix A for the list of Joseon monarchs). Figure 6 shows the yearly number

31 The alternative weighting method 3 implies that the adjacency matrix 퐴 is constructed with its (푖, 푗)th entry

1, if 푖 contributes to 푗′s status as a father 1/2, if 푖 contributes to 푗′s status as a maternal grandfather f(x) = { 1/2, if 푖 contributes to 푗′s status as a fatherinlaw 0, otherwise

19 of exiles (black dots) and political events including Literati Purges and Treason Cases (vertical lines), using the data of exiled officials during the Dynasty from Hong et al. (2019). It is straightforward that the number of exiles abruptly increased during politically unstable times. Thus, we interpret this variable as a proxy for political instability in the state.

[Figure 6 Here]

In Figure 7, the x-axis represents the natural log of exile numbers per year during each king’s reign. To construct the graph, we first residualize each key independent variable (Katz prestige, gapgwa dummy, and passing age) with respect to exam year fixed effects to capture year-specific common shocks impacting the entire candidates and secular trends in each variable.32 We then take average separately by dangsanggwan group and non-dangsanggwan group, and check how the differences between the two groups evolve against exile numbers by each king’s rule. The dashed lines show the best linear fit. The coefficients show the estimated slope of the best-fit line with p-values. The results show that the difference in Katz prestige between those who achieved dangsanggwan positions and those who did not become larger as the number of exiles increases. However, panel (b) indicates that the performance in jeonsi becomes less important when political environments are unstable. This circumstantial evidence suggests that political instability was the highest when family connections mattered the most in explaining the discrepancy between high and low-ranking officials rather than meritocratic merits.

[Figure 7 Here]

7. Conclusion In this paper, we have explored Joseon’s civil service examination as an important institutional

32 Specifically, Katz prestige, gapgwa dummy, and passing age are projected on to dummies for passing year of mungwa, i.e. we run the following regression:

푥푖 = 훼 + 훽푦푒푎푟푖 + 휀푖

We then compute the residual (푟푖) using the following equation:

̂ 푟푖 = 푥푖 − 푥̂푖 = 푥푖 − 훽푦푒푎푟푖 − 훼̂

20 feature of the monarchy. In particular, mungwa was set in place in order to recruit a talented pool of candidates for positions in the court, providing a channel of social mobility and preventing nepotism and corruption that jeopardized the performance of the state. While similar civil service exam systems were historically adopted in other countries and across different time periods, the surviving records from the Joseon Dynasty stand out for their comprehensive information on the candidates’ families and official positions obtained, as well as coverage over five centuries of rule under a single dynasty. We find that even with the adoption of the meritocratic recruiting system, family connections under the monarchy continued to matter in determining the candidate’s eventual career path in the court. More centrally connected candidates obtained higher positions in the court after passing mungwa, controlling for measures of competence (age upon taking the final examination and score ranking in the exam), type of family clan, pre-exam status, exam types and periods, place of residence, and year of birth. Performance in mungwa, on the other hand, had a tenuous effect on the likelihood of obtaining a high-level position. Younger candidates passing mungwa did eventually find their way to high-level positions in the court more easily relative to older counterparts, but scoring high in jeonsi (final palace exam) only mattered when considering a much broader pool of candidates without verified records on their final positions. While we do not necessarily observe a trend toward weakening meritocracy near the demise of the Joseon Dynasty, we do find that periods of political instability and poor economic performance coincide in time with those during which being centrally connected heavily influenced appointment into officialdom. These observations provide suggestive evidence that changing relevance of family connections for official positions does reflect the state performance at the time. As an important feature of the government institution, the exam system certainly fostered scholarly traditions and creation of educated political elites. The efficacy of mungwa system in selecting competent officials in the court, however, appears to have been limited by persistent effects of family connections.

21

Appendix A. Figures and Tables

[Table A1 Here]

[Table A2 Here]

[Table A3 Here]

[Figure A1 Here]

Appendix B. Degree-based Centrality Measures In this section, we briefly summarize degree-based centrality measures. We first explain the concept of degree centrality and show how it can be extended to eigenvector centrality. However, there is a shortcoming of the eigenvector centrality for an asymmetric network. Hence we discuss how it can be fixed using alpha centrality and how it relates to Katz prestige. We also state the reason why we do not consider PageRank centrality in our study. Finally, we provide a simple example of a network to compare degree centrality and Katz prestige. Let a graph (N, A) consist of a set of nodes, N = {1, …, 푛}, and a real-valued 푛 × 푛 adjacency matrix, 퐴, where 푎푖푗 represents that node 푖 contributes to node 푗. That is, the adjacency matrix 퐴 is a zero-one square matrix with its (푖, 푗)th entry,

1, if node 푖 gives a (one way) link to node 푗 푎 = { 푖푗 0, otherwise

Since we focus on a directed graph, 푎푖푗 ≠ 푎푗푖 in our study. Indegree centrality assigns one point for every link a node receives. Thus, we can obtain 푖푛 indegree centrality of node 푖, 푥푖 , by simply calculating

푖푛 푥푖 = ∑ 푎푗푖 푗

In a similar way, outdegree centrality of node 푖 becomes

표푢푡 푥푖 = ∑ 푎푖푗 푗

22

However, degree centrality is easily missing important information of a network by considering all nodes are equivalent. Some are more relevant than others, and endorsements from important nodes need to be counted more. To reflect this aspect, eigenvector centrality, proposed by Bonacich (1972), considers the centrality of neighboring nodes. Let a 푛 × 1 vector 푥푒 denote the eigenvector centrality associated with a network (adjacency matrix) 퐴. The key idea of eigenvector centrality is that the centrality of a node is proportional to the sum of the centrality of its neighbors (Bonacich, 1972). Thus, we write

푒 푒 휆푥푖 = ∑ 푎푗푖 푥푗 푗 where 휆 is a proportionality factor. In terms of matrix notation,

휆푥푒 = 퐴′푥푒 (B1) and

(퐴′ − 휆퐼)푥푒 = 0 where 퐼 denotes the 푛 × 푛 identity matrix. In order for this equation to have a non-zero solution 푥푒, it must be that 퐴′ − 휆퐼 is a singular (or non-invertible) matrix. In other words,

푑푒푡(퐴′ − 휆퐼) = 0 where 푑푒푡(·) indicates determinant. Therefore, 푥푒 is the left eigenvector of 퐴 (or right eigenvector of 퐴′), which corresponds to the in-edges in the graph, and 휆 is its corresponding eigenvalue. The standard convention is to look for the eigenvector associated with the largest eigenvalue (dominant eigenvalue).33 A problem of eigenvector centrality occurs in directed networks, especially when some positions are unchosen. Only nodes in a strongly connected component of two or more vertices can have a positive centrality value, which makes eigenvector centrality useless in asymmetric

33 According to the Perron-Frobenius Theorem, the largest eigenvalue of any nonnegative matrix is real-valued, and its corresponding eigenvector is nonnegative (Jackson, 2010).

23 networks.34 As a solution to this problem, Lloyd and Bonacich (2001) suggest alpha centrality (푥퐴) by assigning a certain minimum score to each node. Let 푒 be a parameter of base value or exogenous sources of status and replace the equations (B1) with the following new equation

푥퐴 = 훼퐴′푥퐴 + 푒ퟏ where ퟏ is a column vector of ones. The parameter 훼 reflects the relative importance of endogenous versus exogenous factors in 1 the determination of centrality. The solution of this equation exists only if 훼 < where 휆 is the 휆 largest eigenvalue of 퐴. If 훼 → 0+, then alpha centrality reduces to degree centrality. On the other

1 hand, if 훼 → ( ) and 훽 = 0, then it becomes similar to eigenvector centrality.35 The matrix 휆 solution of this equation becomes

푥퐴 = (퐼 − 훼퐴′)−1푒ퟏ where 퐼 denotes the 푛 × 푛 identity matrix. This measure of centrality is almost identical to Katz prestige (푥), and the relationship between the two measures can be derived as follows

푥퐴 = [(퐼 − 훼퐴′)−1]푒ퟏ = [(퐼 − 훼퐴′)−1 − 퐼 + 퐼]푒ퟏ = 푒푥 + 푒ퟏ

Therefore, we can find that alpha centrality is simply an affine transformation of Katz centrality. In a practical manner, the exogenous source of status (푒) is usually normalized to one since it just scales scores (Bloch et al., 2017; Jackson, 2010). In this case, the two measures differ only by one (Lloyd and Bonacich, 2001). There is a common concern about Katz prestige in that if a node with high prestige gives links to many other nodes, all targeted nodes are likely to get high prestige. PageRank centrality is an adjustment of this issue by giving less weight for the influences from higher outdegree nodes. However, in our context, it is unlikely that ancestors give shared (penalized) influences to

34 Our family network structure is asymmetric, thus theoretically, we cannot apply eigenvector centrality.

35 The degree centrality measures the immediate local influence and the eigenvector centrality measures the global influence within the network. On the other hand, alpha centrality (and Katz prestige) covers both the local and global influence based on the damping factor (Cruz et al., 2017; Zhan et al., 2017).

24 decedents because of many out-edges. For example, assume that an individual has a (politically) powerful father and many brothers. When evaluating his centrality, it is not suitable to reduce the power of his father because of many brothers. Thus, there is no reason to penalize the link from a high-outdegree source node. For this reason, PageRank centrality is not the right measure for our study. Now, we show an example of a directed network (N, g) which consists of a set of nodes, 푁 =

{1,2, … ,10}, and a 10 × 10 adjacency matrix, 퐴, where 푎푖푗 denotes the relation from 푖 to 푗.

The adjacency matrix 퐴 is represented in Table B1. For example, 푎31 = 1 and 푎13 = 0 mean that node 3 gives an influence to node 1, but not vice versa.

[Table B1 Here]

We calculated scores of each measure in Table B2 and graphically represented the network with corresponding scores in Figure B1.

[Table B2 Here]

We need to note that a node receiving many links does not necessarily have high Katz prestige. Also, a node with high Katz prestige is not necessarily linked to many other nodes. In Table B2, node 3 and node 4 have the largest scores of indegree centrality (= 3), but they have second and third scores of Katz prestige (when 훼 = 0.3), respectively, because they receive links from less central (prestigious) nodes (Figure B1).

[Figure B1 Here]

25

References Baek, Kwang Ryeol (2017) “The concept and analysis method of ‘Kinship Network’ of t he ruling elite in the late Joseon Dynasty,” Society and History, Vol. 114, pp. 7-6 0. Bai, Ying and Ruixue Jia (2016) “Elite recruitment and political stability: the impact of t he abolition of China’s civil service exam,” Econometrica, Vol. 84, pp. 677-733. Banerjee, Abhijit, Chandrasekhar, Arun G., Duflo, Esther, and Jackson, Matthew O. (201 3) “The diffusion of microfinance,” Science, Vol. 341, pp. 363-370. Bloch, Francis, Jackson, Matthew O., and Tebaldi, Pietro (2017) “Centrality measures in networks,” Available at SSRN 2749124. Bonacich, Phillip (1972) “Factoring and weighting approaches to status scores and clique identification,” Journal of Mathematical Sociology, Vol. 2, pp. 113-120. _____ (1987) “Power and centrality: A family of measures,” American Journal of Sociol ogy, Vol. 92, pp. 1170-1182. Bonacich, Phillip and Lloyd, Paulette (2001) “Eigenvector-like measures of centrality for asymmetric relations,” Social Networks, Vol. 23, pp. 191-201. Cha, Mihee (2002) “The promotion to the sixth grade-officials’ duties of the passers of c ivilian service examination in the former 16th century,” The Choson Dynasty Hist ory Association, Vol. 20, pp. 79-106. _____ (2012) “The promotion to the sixth grade-officials’ duties of the passers of civilia n service examination in the 17th and the former 18th centuries,” The Society for Studies of Korean History, Vol. 47, pp. 75-108. Clark, Gregory (2014) The son also rises: Surnames and the history of social mobility, Vol. 49, Princeton, New Jersey: Princeton University Press. Clark, Gregory and Neil Cummins (2014) “Surnames and social mobility in England, 117 0-2012,” Human Nature, Vol. 25, pp. 517-537. _____ (2015) “Intergenerational wealth mobility in England, 1858-2012: Surnames and so cial mobility,” The Economic Journal, Vol. 125, pp. 61-85. Cruz, Cesi, Julien Labonne, and Pablo Querubin (2017) “Politician family networks and e lectoral outcomes: Evidence from the Philippines,” American Economic Review, Vo l. 107, pp. 3006-37. 26

Dal Bò, Ernesto, Frederico Finan, Olle Folke, Torsten Persson, and Johanna Rickne (201 7) “Who becomes a politician?” The Quarterly Journal of Economics, Vol. 132, p p. 1877-1914. Dal Bó, Ernesto, Dal Bó, Pedro, and Snyder, Jason (2009) “Political dynasties,” The Revi ew of Economic Studies, Vol. 76, pp. 115-142. Egorov, Georgy and Konstantin Sonin (2011) “Dictators and Their Viziers: Endogenizing the Loyalty-competence Trade-Off,” Journal of the European Economic Associatio n, Vol. 9, pp. 903-930. Freeman, Linton C. (1979) “Centrality in Networks: I. Conceptual Clarifications,” Social Networks, Vol. 1, pp. 215-239. Hao, Yu and Gregory Clark (2012) “Social mobility in China, 1645-2012: A surname stu dy,” Working paper. Ho, Ping-ti (1962) The Ladder of Success in Imperial China: Aspects of Social Mobility, 1368-1911, New York: Columbia University Press. Hong, Sok Chul, Christopher Paik, and Yangkeun Yun (2019) “Spreading knowledge in exile,” Working Paper. Hymes, Robert (1986) Statesmen and Gentlemen: The Elite of Fu-Chou Chiang-Hsi, in N orthern and Southern Sung, Cambridge (England): Cambridge University Press. Jackson, Matthew O. (2010) Social and Economic Networks, Princeton, NJ: Princeton uni versity press. Jiang, Qin and James Kai-Sing Kung (2016) “Social mobility in pre-industrial China: Rec onsidering the ‘Ladder of Success’ hypothesis,” Working Paper. Johnston, Jack, and John DiNardo (1997) Econometric Methods (Fourth edition), New Yo rk, NY: McGraw-Hill. Katz, Leo (1953) “A new status index derived from sociometric analysis,” Psychometrika, Vol. 18, pp. 39-43. Kim, Yul (2017) “A study on the determinants of bureaucrats’ promotion in the Joseon Dynasty,” Korean Public Administration Quarterly, Vol. 29, pp. 637-657. Kracke, Edward A (1947) “Family vs. merit in Chinese civil service examinations under the empire,” Harvard Journal of Asiatic Studies, Vol. 10, pp. 103-123. Lee, Jae-Ok (2018) “Digital humanities, and applications of the successful exam passers l 27

ist,” The Study of the Eastern Classic, Vol. 70, pp. 303-345. Lee, Nam-Hee (2008) “The state examination system, its lights and shadows,” Issues in East-Asian Philosophy, pp. 117-136. Lee, Peter Hacksoo (2003) A history of Korean literature, Cambridge (England): Cambrid ge University Press. Lee, Sung-moo (1994) “Education system and civil service examination,” in Jon-hee Lee ed. History of Korea, Gwacheon: National Institute of Korean History, pp. 279-36 8. Li, Sŏng-Mu (2000) “The history of the factional strifes in Chosun Dynasty,” The Journ al of Korean History, Vol. 111, pp. 241-247. Marsh, Robert M (1961) “Formal organization and promotion in a pre-industrial society,” American Sociological Review, pp. 547-556. Meghanathan, N (2015) Correlation Coefficient Analysis of Centrality Metrics for Comple x Network Graphs, Intelligent Systems in Cybernetics and Automation Theory, Swit zerland: Springer International Publishing AG. Paik, Christopher (2014) “Does lineage matter? A study of ancestral influence on educati onal attainment in Korea,” European Review of Economic History, Vol. 18, pp. 43 3-451. Reuter, Ora John and Graeme B. Robertson (2012) “Subnational Appointments in Authori tarian Regimes: Evidence from Russian Gubernatorial Appointments,” Journal of P olitics, Vol. 74, pp.1023-1037. Shiue, Carol H (2016) “Social mobility in the long run: An analysis with five linked ge nerations in China, 1300-1900,” Working paper. Song, Man-ho (2010) “Professor Edward Wagner and Professor Jun-ho Song, who set a milestone in the pursuit of the ruling class in the Joseon Dynasty,” Civil Educatio n of Korean History, pp. 217-232. Wagner, Edward W (1974) “The ladder of success in Yi dynasty Korea,” Occasional Pa pers on Korea, pp. 1-8. Wasserman, Stanley, and Katherine Faust (1994) Social network analysis: Methods and a pplications, Cambridge (England): Cambridge University Press. Won, Chang-ae (2007) “Assignment of successful candidates of the civil service examinat 28

ion to regular official positions in the Choseon Dynasty,” The Choson Dynasty Hi story Association, Vol. 43, pp. 5-50. _____ (2019) “Description for civil service examination rosters,” URL: http://people.aks.a c.kr/front/haeje/haeje.aks?h=MN, accessed on 4th January, 2019. Wooldridge, Jeffrey M. (2010). Econometric analysis of cross section and panel data, Ca mbridge, MA: MIT press. Yi, Hyang-woo (2015) Koran Royal Palace: Changdeokgung, Seoul: Inmunwalk. Zakharov, Alexei V (2016) “The loyalty-competence tradeoff in dictatorships and outside options for subordinatesy,” Journal of Politics, Vol. 78, pp. 457-466. Zhan, J., Gurung, S., and Parsa, S. P. K (2017) “Identification of top-K nodes in large networks using Katz centrality,” Journal of Big Data, Vol. 4.

29

Figures and Tables

Figure 1. The structure of mungwa in Joseon Dynasty

Notes: The regional quotas at chosi were proportional to population in each province. A total of 240 candidates were selected from 20 in Gyeonggi Province, 15 in Gangwon Province, 15 in Pyeongan Province, 25 in Chungcheong Province, 25 in Jeolla Province, 30 in Gyeongsang Province, 10 in Hwanghae Province, and 10 in Hamgyeong Province (Lee, 2008). The local quotas were not applied to hoesi and jeonsi.

30

Table 1. Initial placement based on previous status and results of palace examination Official rank (position) assignment Ranking in palace exam Number Without official position With official position Gapgwa (First division) 1st rank (jangwon) 1 Junior 6th rank position 4 ranks promoted 2nd and 3rd rank 2 Senior 7th rank position 3 ranks promoted Eulgwa (Second division) 7 Junior 8th rank 2 ranks promoted Byunggwa (Third division) 23 Senior 9th rank 1 rank promoted Notes: This table is re-tabulated from the tables of Lee (1994). “Without previous position” means the status upon taking the examination was either saengwon (classics licentiate), jinsa (literary licentiate), yuhak (Confucian student). “With previous official position” denotes those already holding official titles. Candidates of the second and third divisions without previous official position were not guaranteed a post but only received an official rank (not a real official position) and had to wait as temporary officials until one became vacant. If mid-level officials passed mungwa, they were guaranteed to have promoted positions though they were not placed in gapgwa in jeonsi (Won, 2007).

31

Figure 2. Sa-Ahn Kang and related figures

Notes: This family network shows the nodes and ties related to Sa-Ahn Kang. We describe the partial networks that are directly or indirectly linked to Sa-Ahn Kang in our dataset. There are many other men who are linked to Sa-Ahn Kang indirectly. However, due to the limitation of space, we restricted figures to the extent that does not hurt our efforts to help for the understanding of network structure. Bold represents men who passed the literary examination, having subsequent information on the family. The reason Gan Im who was a father-in-law of Sa-Ahn Kang has a tie with Yu-Gyeom Im even though Gan Im was not a successful candidate, which means has no information about family, is that the family information comes from Gwang Im who passed the literary exam in 1624 and was a grandson of Gan Im. That is, in the rosters, there are information that Gwang Im was a son of Ye-Shin Im, a grandson of Gan Im, a great grandson of Yu-Gyeom Im, and so on. In this specific example, the number of nodes is 24 and the number of ties is 25. The relationships between figures are in parenthesis.

32

Figure 3. Sa-Ahn Kang and related figures with scores of Katz prestige

Notes: This family network shows the nodes and ties related to Sa-Ahn Kang. We describe the partial networks that are directly or indirectly linked to Sa-Ahn Kang in our dataset, corresponding with Figure 2. The size of the circle is proportional to scores of Katz prestige.

33

Figure 4. Proportion of dangsanggwan by grade in jeonsi and Katz prestige

Notes: This figure is a binned scatter plot indicating the proportion of dangsanggwan (high-level officials). To construct this binned scatter plot, we first divide Katz prestige into ten equal sized-groups (deciles) and plot the means of the y-axis variable within each bin against the mean value of Katz prestige (z-score) within each bin.

34

Table 2. Effects of family networks on high-level official position: Basic estimation (1) (2) (3) (4) Baseline (5) (6) (7) Network measure Katz prestige 0.0623*** 0.0620*** 0.0549*** 0.0387*** 0.0325*** 0.0292*** 0.0258*** (0.0093) (0.0093) (0.0092) (0.0092) (0.0095) (0.0096) (0.0099) Performance in jeonsi (Reference: Byunggwa) Gapgwa 0.0836*** 0.0852*** 0.0909*** 0.0896*** 0.0894*** 0.0894*** (0.0243) (0.0247) (0.0237) (0.0238) (0.0238) (0.0238) Eulgwa 0.0130 0.0101 0.0083 0.0074 0.0077 0.0079 (0.0223) (0.0231) (0.0210) (0.0207) (0.0207) (0.0207) Performance until hoesi Passing age -0.0705*** -0.0600*** -0.0576*** -0.0576*** -0.0576*** (0.0132) (0.0136) (0.0137) (0.0137) (0.0137) Ancestors’ achievement dummies in mungwa Father, grandfather, 0.0484*** 0.0486*** 0.0487*** and great grandfather (0.0153) (0.0153) (0.0153) Father-in-law 0.0181 0.0198 (0.0192) (0.0191) Maternal grandfather 0.0158 (0.0155) Pre-exam status (Reference: Confucian student) Classics licentiate 0.0652*** 0.0631*** 0.0681*** 0.0323 0.0313 0.0313 0.0312 (0.0232) (0.0234) (0.0233) (0.0271) (0.0272) (0.0272) (0.0273) Literary licentiate 0.0940*** 0.0900*** 0.0976*** 0.0567** 0.0547** 0.0542** 0.0539** (0.0215) (0.0214) (0.0211) (0.0245) (0.0245) (0.0245) (0.0245) Previous official holder 0.1499*** 0.1456*** 0.1708*** 0.1282*** 0.1254*** 0.1245*** 0.1244*** (0.0170) (0.0172) (0.0172) (0.0205) (0.0206) (0.0205) (0.0205) Exam type (Reference: Regular exam) Irregular exam 0.0540*** 0.0543*** 0.0563*** 0.0446*** 0.0435*** 0.0433*** 0.0431*** (0.0153) (0.0154) (0.0157) (0.0156) (0.0156) (0.0156) (0.0158) Additional controls King in rule        Family clan        Year of birth        Residence     Observations 5,179 5,179 5,179 5,179 5,179 5,179 5,179

35

Notes: We conducted the regressions in equation (1) across different model specifications. The dependent variable is a dummy variable whether highest official position was higher than or equal to the third rank upper senior position (i.e. belonging to the group of high-level rank officials). Katz prestige and passing age are standardized to be mean zero and standard deviation one. In column (1), the specification includes the pre-exam status, exam type, period of king in rule, family clan, and year-of-birth fixed effects. We additively contain grade in jeonsi in column (2) and age upon passing hoesi in column (3), respectively. Column (4), which is our preferred specification, adds residence fixed effects to rule out the varying spatial impact of the regional characteristics. Family clan fixed effects include the dummies of 319 family clans. King fixed effects include the dummies of 26 kings during the Joseon Dynasty. Exam type fixed effects include the regular exams (siknyeonsi and jeunggwangsi) and the irregular exams (byulsi, alsungsi, jungsi, and so on). Year-of-birth fixed effects include each successful candidate's birth year dummies from 1363 to 1878. Residence fixed effects include 202 district-level (Gun-Hyun) region dummies. Each cell reports the estimated coefficients and their standard errors clustered on family clans in parenthesis. A single asterisk denotes statistical significance at the 90% level of confidence, double 95%, and triple 99%.

36

Figure 5. Alternative cut-offs of high-level official position

Notes: We repeat the same estimation with the baseline by replacing the dependent variable with the different cut-offs. We report the coefficient estimates for the Katz prestige in panel (a), estimates for gapgwa in jeonsi in panel (b), and estimates for the age upon passing hoesi in panel (c) with their 95 percent confidence intervals.

37

Table 2. Effects of family networks on high-level official position: Alternative estimation Alternative model Alternative decay factor Alternative edge weight Baseline Logit Probit α=.1 α=.5 Method 1 Method 2 Method 3 (1) (2) (3) (4) (5) (6) (7) (8) Network measure Katz prestige 0.0387*** 0.0384*** 0.0388*** 0.0313*** 0.0338*** 0.0342*** 0.0260*** 0.0410*** (0.0092) (0.0105) (0.0085) (0.0099) (0.0073) (0.0094) (0.0090) (0.0109) Performance in jeonsi (Reference: Byunggwa) Gapgwa 0.0909*** 0.0953*** 0.0924*** 0.0900*** 0.0928*** 0.0904*** 0.0911*** 0.0914*** (0.0237) (0.0333) (0.0223) (0.0238) (0.0237) (0.0238) (0.0238) (0.0237) Eulgwa 0.0083 0.0071 0.0060 0.0087 0.0082 0.0087 0.0086 0.0088 (0.0210) (0.0188) (0.0187) (0.0210) (0.0209) (0.0209) (0.0210) (0.0209) Performance until hoesi Passing age -0.0518*** -0.0523*** -0.0544*** -0.0524*** -0.0532*** -0.0548*** -0.0516*** -0.0515*** (0.0118) (0.0140) (0.0119) (0.0117) (0.0119) (0.0119) (0.0120) (0.0110) Observations 5,179 4,633 4,633 5,179 5,179 5,179 5,179 5,179 Notes: We conducted the baseline regressions using alternative specifications. Column (1) presents the baseline results, which is reported in column (4) of Table 2, for comparison purpose. In columns (2) and (3), we estimate using logit and probit model instead of linear probability model. To make them comparable using probability scale, we report the average marginal effects across values of each variable. Columns (4) and (5) show the results of different decay factor in Katz prestige. From column (6) to column (8), we use different edge weights according to the kind of ties. In the weighting method 1, we replace the edge weights from maternal grandfather to 1/2. Additionally, we replace the edge weights from father-in-law to 1/2 in column (7) and the edge weights from foster father to 0 in column (8), respectively. In other words, we give less weight to the influences from maternal grandfather and father-in-law, and no weight from foster father. Each cell reports the estimated coefficients and their standard errors clustered on family clans in parenthesis. A single asterisk denotes statistical significance at the 90% level of confidence, double 95%, and triple 99%.

38

Figure 6. Political stability and exile

80

60

40

Number

20

0

1390

1410

1430

1450

1470

1490

1510

1530

1550

1570

1590

1610

1630

1650

1670

1690

1710

1730

1750

1770

1790

1810

1830

1850

1870

1890

Notes: Each dot depicts the number of exiles by year. Vertical lines are political events including Literati Purges and Treason Cases.

39

Figure 7. Relationship between political stability, political power, and key independent variables

Notes: The numbers of plots refer to each king periods presented in Table A3 in Appendix A. We first residualize each key independent variable (Katz prestige, gapgwa dummy, and passing age) with respect to exam year fixed effects to capture year-specific common shocks impacting the entire candidates and secular trends in each variable. We then take average separately by dangsanggwan group and non-dangsanggwan group, and check how the differences between two groups evolve against exile numbers by each king’s rule. The x-axis shows the natural log of the number of exile per year by king in rule as a proxy for political stability. The dashed lines show the best linear fit. The coefficients show the estimated slope of the best-fit line with p-values.

40

Table A1. Rank system of the government officials Category Rank and Sub-rank Dangsanggwan (high-level officials) Upper 1st Senior Lower Upper 1st Junior Lower Upper 2nd Senior Lower Upper 2nd Junior Lower Upper 3rd Senior Chamsanggwan (mid-level officials) Lower Upper 3rd Junior Lower Upper 4th Senior Lower Upper 4th Junior Lower Upper 5th Senior Lower Upper 5th Junior Lower Upper 6th Senior Lower Upper 6th Junior Lower Chamhagwan (low-level officials) 7th Senior 7th Junior 8th Senior 8th Junior 9th Senior 9th Junior Notes: Dangsanggwan (high-level officials) were defined as the ministers of upper senior third or higher ranks, collectively known as ‘palace-ascendable officials’ They were given important rights to vote on the administration, to recommend other officials, and to direct the military of the relevant officials (Cha, 2002). Officials from lower junior sixth rank to the lower senior third rank were called ‘chamsanggwan (mid-level officials)’ and they were in the charge of central administration and local government.

41

Table A2. Summary statistics Total Dangsanggwan vs. Non-dangsanggwan Mean SD Min Max Mean Difference P-value Proportion of dangsanggwan 0.556 0.497 0 1 1 0 Network measure Katz prestige 2.287 0.587 1.000 4.857 2.373 2.179 0.194*** 0.000 Performance in jeonsi Gapgwa 0.117 0.321 0 1 0.137 0.092 0.045*** 0.000 Eulgwa 0.200 0.400 0 1 0.208 0.190 0.018 0.100 Byunggwa 0.683 0.465 0 1 0.654 0.718 -0.064*** 0.000 Performance in hoesi Passing age 33.169 9.176 13 82 32.698 33.757 -1.059*** 0.000 Pre-exam status Confucian student 0.275 0.447 0 1 0.225 0.338 -0.113*** 0.000 Classics Licentiate 0.147 0.354 0 1 0.143 0.151 -0.008 0.407 Literary Licentiate 0.198 0.398 0 1 0.204 0.190 0.014 0.198 Previous official holder 0.380 0.486 0 1 0.428 0.321 0.107*** 0.000 Irregular exam 0.526 0.499 0 1 0.544 0.503 0.040*** 0.004 Lived in Seoul before exam 0.604 0.489 0 1 0.674 0.515 0.159*** 0.000 Observations 5,179 5,179 5,179 5,179 2,879 2,300 5,179 Notes: This table reports descriptive statistics on variables related to empirical analysis.

42

Table A3. List of Joseon monarchs Number Temple name Period of reign Name of king 1 Taejo 1392–1398 Yi Seong-gye 2 Jeongjong 1398–1400 Yi Bang-gwa 3 Taejong 1400–1418 Yi Bang-won 4 Sejong 1418–1450 Yi Do 5 Munjong 1450–1452 Yi Hyang 6 Danjong 1452–1455 Yi Hong-wi 7 Sejo 1455–1468 Yi Yu 8 Yejong 1468–1469 Yi Gwang 9 Seongjong 1469–1494 Yi Hyeol 10 Yeonsangun 1494–1506 Yi Yung 11 Jungjong 1506–1544 Yi Yeok 12 Injong 1544–1545 Yi Ho 13 Myeongjong 1545–1567 Yi Hwan 14 Seonjo 1567–1608 Yi Yeon 15 Gwanghaegun1 1608–1623 Yi Hon 16 Injo 1623–1649 Yi Jong 17 Hyojong 1649–1659 Yi Ho 18 Hyeonjong 1659–1674 Yi Yeon 19 Sukjong 1674–1720 Yi Sun 20 Gyeongjong 1720–1724 Yi Yun 21 Yeongjo 1724–1776 Yi Geum 22 Jeongjo 1776–1800 Yi San 23 Sunjo 1800–1834 Yi Gong 24 Heonjong 1834–1849 Yi Hwan 25 Cheoljong 1849–1863 Yi Byeon 26 Gojong 1863–1897 Yi Myeong-bok 27 Sunjong 1907–1910 Yi Yu Notes: This table shows the list of twenty-seven kings of the Joseon Dynasty.

43

Figure A1. Sample of family networks having more than or equal to 5 ties

Notes: This figure shows sample family networks of those who reached dangsanggwan, or high-level official, (black dots) and those who did not (gray dots). The size of the circle is proportional to Katz prestige. To simplify the graph, we restricted the nodes that have more than or equal to 5 ties. In this network, the number of nodes is 2,729 (5.71% of total) and that of ties is 2,391 (4.86% of total), respectively. We represented this network using ForceAtlas layout (see https://gephi.org/users/tutorial-layouts/).

44

Table B1. Adjacency matrix of a sample network Node ID 1 2 3 4 5 6 7 8 9 10 1 0 0 0 0 0 0 0 0 0 0 2 0 0 0 0 0 0 0 0 1 0 3 1 0 0 0 0 0 0 0 0 0 4 1 0 0 0 0 0 0 0 0 0 5 0 0 1 0 0 0 0 0 0 0 6 0 0 1 0 0 0 0 0 0 0 7 0 1 0 1 0 0 0 0 0 0 8 0 1 1 1 0 0 0 0 0 0 9 0 0 0 0 0 0 1 0 0 0 10 0 0 0 1 0 0 0 0 0 0 Notes: This table shows the adjacency matrix of sample network. For example, a31 = 1 and a13 = 0 mean that node 3 gives an influence on node 1, but not vice versa.

45

Table B2. Example of network and centrality measures Degree centrality Katz prestige Node ID In-edge Out-edge α=.1 α=.3 α=.5 1 2 0 0.261 1.181 2.786 2 2 1 0.211 0.737 1.571 3 3 1 0.300 0.900 1.500 4 3 1 0.311 1.037 2.071 5 0 1 0.000 0.000 0.000 6 0 1 0.000 0.000 0.000 7 1 2 0.112 0.456 1.143 8 0 3 0.000 0.000 0.000 9 1 1 0.121 0.521 1.286 10 0 1 0.000 0.000 0.000 Notes: This table shows each centrality measure corresponding network in Table B1.

46

Figure B1. Graphical representation of sample network with scores (a) Indegree centrality (b) Katz prestige (α=.3)

Notes: Each network graph corresponds to the sample network of Table B1. The numbers in each circle denote the nodes’ id. The size of circles is proportional to scores of each measure.

47