Bulletin of Data Analysis of Japanese Classification Society
Online ISSN : 2434-3382
Print ISSN : 2186-4195
General Trust in a Changing Society: The Development of Interpersonal Trust between 1978 and 2013 in Japan
(Findings from the Japanese National Character Survey)
Wolfgang JagodzinskiHermann DuelmerYusuke InagakiTadahiko Maeda
Author information

2019 Volume 8 Issue 1 Pages 25-46


Yamagishi & Yamagishi は,アメリカと比べて日本では一般的信頼が低いことを示したうえで,日本社会で広く見られる長期的なコミットメント関係が,一般的信頼の欠如をもたらしたと考察した.

これに対して本研究は,「日本人の国民性調査」の二次分析を通じて,日本におけるコミットメント関係への選好が過去35 年ほどの間に一般的信頼の水準に強い影響を及ぼしたという仮説に疑問を投げかけるものである.

なお,分析からは,一般的信頼と以下のような事柄との関連性が示唆された. ①一般的信頼は「年齢」によって弱い正の影響を受けるが,「教育」によって強い正の影響を受ける.したがって,教育機会の拡大と政治的関心の広がりによる「認知動員」の上昇や日本社会の高齢化の過程は,一般的信頼の向上をもたらした可能性がある. ②都市化の影響については,依然としてあいまいなままであった. ③「性別」について、初期の調査時点では女性の一般的信頼は低かったが,1990 年頃になるとこのような性差は消失した.

ただし上記の変数は,いわゆる「失われた10 年」や2008 年の世界的な経済危機の影響によって生じたと考えられる,1990~2008 年の一般的信頼の低下を説明するものではなかった.

要 旨

More than two decades ago, Yamagishi and Yamagishi demonstrated in a seminal article that trust was markedly lower in Japan than in the US. They tentatively attributed the lack of trust to the prevalence of committed long-term relationships in the Japanese economy. These relationships, they argued, assure cooperation because defection is costly and can be effectively sanctioned by the partner. This study’s secondary analysis of the Japanese National Character Survey casts doubt on the assumption that the Japanese preference for committed relationships has had a substantial impact on the level of general trust during the last 35 years, as other factors seem to be more relevant. General trust is weakly positively affected by age but substantially positively affected by education. Processes of cognitive mobilization and the aging of the Japanese society may, therefore, lead to an increase in general trust. The effects of urbanization remain ambiguous. Women were less trustful in the early surveys, but around 1990 gender differences disappeared. All together these variables cannot explain the decline in general trust between 1990 and 2008, which we tentatively attribute to deteriorating economic conditions during the “lost decade” and the worldwide economic crisis in 2008, which seem to have had a particularly strong impact on the youngest cohorts.

1. Introduction1

International comparative survey research has reported astonishing differences between countries and cultures—differences in life satisfaction and happiness, values, or social capital, for example. These differences are often traced back to long-lasting properties of the political or economic system, such as a pluralistic democracy, a welfare state, or market economies. The seminal paper of Yamagishi and Yamagishi (1994) is a good example of this research tradition. The authors observe that general trust in Japan is significantly lower than in the US and suggest that differences in the structure of business relationships are the decisive determinant. While short-term relationships with frequently changing partners prevail in the US, the Japanese prefer long-term, committed relationships, which are safeguards against defection. The reduction in transaction costs reached in the open markets of the US through generalized trust is reached in Japan by committed relationships. While committed relationships give assurance, they do not promote general trust.

One is inclined to agree with the message that long-term relationships play an important role in Japan, not only in the domain of business relationships but in the labor market as well. Life-long employment and “inbred management” still seem to be typical features of the Japanese economic system. However, can this factor really explain the development of interpersonal trust during the last decades? Citizens have often substantially changed their attitudes and general orientations within relatively stable institutional settings. Putnam’s (2000) analysis of the declining social capital in the US during the last decades is a good example. Though the democratic and economic institutions are still functioning and companies nowadays invest considerable resources into building and maintaining trust relationships with their customers, interpersonal trust steadily declines. Accordingly, Putnam does not attribute the spread of distrust to institutional change but to factors like the increase in commuting time or hours of TV-watching.

These considerations lead to two research questions. First, can differences in the role of committed relationships really explain cross-country differences in generalized interpersonal trust between the US and Japan? And second: are there other than institutional factors which influence the development of trust in Japan? Before we clarify these questions, we briefly outline our conceptualization and operationalization of general interpersonal trust in Section 2. The first research question is addressed in Section 3. It cannot be answered by means of a strict empirical test because, just as Yamagishi and Yamagishi, we lack a reliable measure of the degree of committed relationships in business and economy. Rather, by using the General Social Survey US and the Japanese National Character Survey, we compare the development of general interpersonal trust over a period of more than 30 years and discuss whether the observed pattern in both countries can be plausibly attributed to differences in committed business relationships. Other fundamental changes may have affected the development of general interpersonal trust in Japan. We discuss the following in Section 4: Old (1) and new (2) differences between urban and rural areas, the rise of education (3), the increasing participation of women in the labor market (4), the aging of the Japanese society (5), intergenerational change (6), and the economic development between 1978 and 2013 (7). Our considerations will be summarized in a set of hypotheses and tested in Section 5, first in a pure micro-level model, and second in a model which includes micro and macro variables. The results are assessed and qualified in the concluding remarks (Section 6).

2. The concept and measurement of general interpersonal trust

2.1. The concept

Trust is a useful concept for Web 4.0, the web of things2 because it can be applied to all kinds of objects, beginning with trust in our old car and ending with trust in the stability of a currency. It has become an interdisciplinary concept (cf. Fetchenhauer et al., 2013) which, due to space limitations, cannot be fully elaborated here. Instead, we briefly outline concepts that explicitly or implicitly interpret trust in humans as (1) a mathematical function of (2) the subjective probability that a (3) class of persons P (4) will perform a class of conditioned actions A. We call the person who gives trust the trustor, and the person who receives trust the trustee or trusted person. Expression (1) clarifies that we require a mathematical function because the term “a function of” is ambiguously used in the human sciences, while expression (2) further articulates our assumption that the trustor assigns a subjective probability to the outcome, that is, the actions A of P. This probability will often deviate from empirically estimated probabilities.

The degree of generality of trust is determined by expressions (3) and (4) above. If Jack has full trust that Peter will pay back his loan tomorrow at 5 pm in Jack’s department, the class of persons (3) consists only of Peter, the class of actions (4) reduces to the repayment of money, and time and place of the action are precisely determined. This is of course by no means a complete description of all relevant conditions. Peter may not be able or willing to repay Jack for many reasons. Trustors rarely have complete information about the incentives, gains, and losses that are relevant in a particular situation. Generalizations from former experiences, analogies, and heuristics must then help crudely estimate the respective probability.

Trust can first be generalized with regard to the class of people P. We can have not only trust in a single person but also in our parents and family, in neighbors, community citizens, fellow Japanese, or all human beings. Those who equate generalized trust to trust in all humans may classify all other forms as particularized trust (Uslaner, 2002, 2008). We will not follow this practice, not so much because we believe that truly generalized trust should include living creatures but rather because (a) levels of generalization within the broad category of particularized trust are completely ignored and (b) only a minority of people will have unqualified trust in all humans. We only distinguish levels of generalization. The higher the level, the more categories of people that are included. If we follow Tajfel and Turner (1979), categorization promotes the distinction between in- and outgroups and generates a differential between ingroup and outgroup trust. Accordingly, it can be expected that interpersonal trust decreases with the level of generalization as long as ingroup/outgroup thinking prevails.

While generalizations with regard to the class P have been widely discussed in recent years (cf. Delhey, Newton, and Welzel, 2011; Sturgis and Smith, 2010; Tang, 2016), generalizations with regard to class A have received astonishingly little attention. It is true that when operationalizing interpersonal trust, the question arises whether fair, honest, or helpful behaviors are expressions of trustworthiness and are therefore suitable as indicators of trust but this is exclusively seen as a measurement issue. The broader problem is possibly ignored because trustworthiness is predominantly seen as a personal trait and generalized trust as a positive view of other humans. This is a simplistic perspective, however. Traits become manifest in behaviors and we usually, of course, have behaviors in mind when we say that most people can be trusted. Which kinds of behavior are considered depends partly on the context of communications. In business we usually think of economic transactions; in legal matters of the fairness of judges; in political matters possibly of the confidential handling of information; and in discussions about school teachers of the care for entrusted children. These are domain-specific interpretations of trust, which in turn can be further specified. Our judgment on trust may be based on a typical or more unusual situation where people act under heavy pressure. There are various possibilities for generalizing behaviors and situations.

To summarize so far, there are different ways to generalize trust over not only people but also over actions and situations. Whatever level of generalization is chosen, the trustor will estimate the probability that a class of actions is performed by the group of trusted persons. This subjective probability is a belief, not a norm or world view. Even if the foundation of the belief may have already been laid in childhood and youth, it needs continuous confirmation in later life. In this sense, generalized trust is a cumulative life experience.

We leave open whether generalized trust alternatively can be and should be fruitfully conceptualized as moral trust (Uslaner, 2002, 2008). We also do not distinguish between generalized trust and assurance (Yamagishi and Yamagishi, 1994) because the trustor rarely can manipulate the incentive structure in such a way as to obtain assurance. If the Japanese really do differ from North Americans in the level of generalized trust, there can be several reasons: the Japanese may be more risk-averse, they may generalize trust in different ways, or they may derive their subjective probability from different information than North Americans.

2.2. The measurement of generalized interpersonal trust

Among the many scales that have been deployed to measure generalized trust, an item of Rosenberg’s (1956) faith-in-people or misanthropy scale has become the most prominent. Simplicity and parsimony make the item attractive for survey research. The trust item reads:

Some people say that most people can be trusted. Others say you can’t be too careful in your dealings with people. How do you feel about it?

The wording of the question, which we call the Rosenberg T-item for the sake of brevity, can vary slightly and the response scales3 can vary greatly. The Japanese National Character Survey (JNCS) introduced this and two other items of the misanthropy scale4 in 1978 with a dichotomized response scale5. With the exception of the 1988 survey, the item is replicated every five years in the K-type6 questionnaire of the survey. Thus we analyze seven datasets of the JNCS. The English translation of the item can be found in the codebooks, for instance in the Codebook for the 2013 survey (Nakamura, Yoshino, Maeda, Inagaki, and Shibai, 2017).

Similar to questions on values, life satisfaction, general political attitudes, and so on, the Rosenberg T-item addresses the topic in a very general form. It neither refers to particular groups nor to any kind of behavior but simply asks whether most people can be trusted. Given the short response time in an interview, such questions can mentally only be processed by our system 1 (Kahneman, 2011), which allows for quick reactions. As long as system 1 operates in all respondents in precisely the same way, we may hope to get comparable answers. Unfortunately, that is not the case. Sturgis and Smith (2010) have shown that only a minority of their British sample thinks of people in general when asked about “most people.” Delhey, Newton, and Welzel (2011) explicitly measured trust in different groups, beginning with trust in members of the family, and ending with trust in people of other nations. They show that ingroup trust is higher than outgroup trust and that the effect of both kinds of trust on the Rosenberg T-item varies among countries. In the few Asian countries included in their study7, ingroup trust had a much stronger impact on the Rosenberg T-item than outgroup trust.

The question of what kind of trust behavior respondents are referring to has not yet been investigated. Neither can we examine this question with the data at hand nor can we find out how respondents in our data sets interpret the expression “most people.” Rather we must make the untestable assumption that all respondents frame the Rosenberg T-question in the same way as far as trusted people and trust-related behaviors are concerned. We have to exclude, for instance, that American respondents interpret “most people” as “my neighbors and people I know,” while Japanese respondents also take into consideration the people they meet for the first time when answering the Rosenberg T-question. We also assume that trust-related behaviors are generalized in the same way by all respondents and that all also base their judgment on economic behavior. They all would agree, for example, that thieves, crooks, or defalcators are not trustworthy. Consistent with this assumption, empirical studies have shown that trust is dependent on income and social status.

On balance, we do not know for sure on which level of generalization trust is measured. We know, however, that respondents refer to a larger group of people and also take a broader range of behaviors into consideration when confronted with the Rosenberg T-question. For this reason, it seems to be justified to speak of generalized interpersonal trust, or, for the sake of brevity, general trust.

3. Change in general trust in Japan and the US

Surveys from the eighties and early nineties consistently showed that general trust in the US was much higher than in Japan (Yamagishi and Yamagishi, 1994 with further references). But is that still true considering the fact that general trust in the US has declined? A crisis of confidence in social and political American institutions had already been diagnosed by Lipset and Schneider (1983) in the early eighties, but Putnam (2000) analyzed the erosion of social capital in the US on a much broader basis.

To answer this first question, we briefly compare the percentages of trustors in the General Social Survey US8 (GSS-US) and in the JNCS. The JNCS began in 1953, but a question on trust was not included until 1978. The GSS-US has included a question on trust since the early seventies, and the survey is carried out more frequently. Thus, the two time series in Figure 1 differ in length and number of data points. They also include slightly different versions of the Rosenberg T-item. In contrast to the JNCS-item already described, the GSS-US question version has a middle category (“it depends on the situation”) between the two opposite poles of trust (“most people can be trusted”) and distrust (“you can’t be too careful”). To make the items more comparable, we remove the middle category in two different ways: a) The optimistic variant assigns all “middle responses” to trust, assuming, in other words, that the respective respondents would choose “most people can be trusted” if only two categories were offered. This leads to a higher percentage of interpersonal trustand is depicted as a dashed line in Figure 1. The corresponding data points are marked as upward directed triangles. They represent, so to speak, the upper bound of trust in the US. b) The cautious, conservative variant assigns all “middle responses” to distrust and assumes that the respective respondents would choose the no-trust alternative (“you can’t be too careful”) if only two alternatives were offered. The percentages are displayed as downwarddirected triangles, which are connected by a dotted line, and represent the lower bound of trust in the US. The Japanese percentages are represented by circles, which are connected by a bold line.

Figure 1 confirms the decline in trust in the US. It also shows that trust in Japan increased considerably between 1978 and 1993. In the latter year, the Japanese percentage even exceeded the upper-bound percentages in the GSS-US surveys of the same and the following year. After that peak, the Japanese percentages decline and remain slightly below the cautious estimates for the US, but then in 2013, the Japanese level of interpersonal trust again rises above the optimistic estimate for the US. On balance, we can hardly say that the level of generalized interpersonal trust is still significantly lower in Japan than in the US.

Figure 1.

Development of trust between 1970 and 2015 in Japan and the US.

4. Determinants of the Japanese change in trust

How can the Japanese change in trust be explained? The transformation of committed relationships can hardly be responsible for the increase in trust before 1993 and its decline 31 General Trust in a Changing Society: The Development of Interpersonal Trust between 1978 and 2013 in Japan (Findings from the Japanese National Character Survey) afterward. It is true that at the beginning of the eighties, the so-called “freeters”9 were identified as a new type of Japanese workers. In the beginning, freeters were sometimes seen in the media as individualists who had deliberately chosen temporary jobs for the purpose of self-realization and independence (cf. Hommerich, 2008). Even if this type of freeters had existed before the economic crises10, this would not indicate a broader interest in temporary employment or a general renunciation to long-term committed relationships in Japan. Freeters were on the fringes of the Japanese labor market and, during the economic crises, temporary employment was rarely chosen deliberately but was forced by the economic conditions. In 2016, the governor of the Bank of Japan stated: “. . . With the Japanese economy entering a period of stagnant growth in the second half of the 1990s, the economic situation underwent dramatic changes. Nevertheless, despite shifts in the use of non-regular workers, the employment and wage system for regular workers remained virtually the same.”11

We have no quantitative information on long-term committed relationships, and we also lack a number of other interesting variables. It has been shown in several studies that the poor are less trustful than the rich (cf. Delhey and Newton 2003 for further references), but our dataset has no measures of income, objective social class, or social status. The self-rated social status is included in only three of the K-surveys, and a question on happiness is included only once. Measures of anxiety and job- and life-satisfaction are completely lacking. We must restrict our analysis to the influence of urbanization, education, gender, age, generation, and macro-economic conditions.

(1) Parochial rural areas

Urbanization might be related to general trust. Due to a reform of the administrative districts in Japan in the 2000s, the change in the percentage of the residents in the so-called rural areas (“Gun” areas)12 does not accurately reflect the degree of urbanization. Nevertheless, the de-population of rural areas has become a serious problem not only in Japan but in many countries around the world. “Rural” often connotes strong family ties, small but dense networks of personally known people, and high social control. Communities of this type will be called “parochial rural communities.” The smaller the circle of trustees, the larger the outgroup of distrusted people. Parochial trustors who only trust their family members perceive themselves as surrounded by a large group of hostile people. High ingroup trust is associated with to low outgroup trust13. Under the assumption that respondents take hostile outgroups into consideration14 when answering the Rosenberg T-question, we tentatively derive the first hypothesis: General trust is lower in rural areas (H1a).

One may point out that parochial communities have largely disappeared in advanced society. Modern forms of transportation and communication have broken up formerly closed communities and tightly connected them to the surrounding world. In contrast, if some closed communities still exist, they certainly do not coincide with the geographical tiering of statistical bureaus. Many rural cities and towns have become bedroom communities of larger cities15 so that the distinction between rural and urban is increasingly blurred. In other words, our indicators of community size no longer distinguish between parochial and other rural areas. A possibility is that we still find the theoretically expected difference in earlier but not in later surveys (H1b).

(2) Large Cities

It has been argued that large cities offer many more opportunities for positive contacts with strangers (Allport, 1958 [1954]; Pettigrew, 1998), which would determine the spread of general trust. Glaeser, however, suspects a nonmonotonic relationship between trust and population density: “At low levels of density, social networks cannot exist because transportation costs between people are too high. At high levels of density, individuals do not form bonds because they are too mobile and find it too easy to take advantage of one another and just move on” (Glaeser 2000: 130). He shows that indicators of trust are lower in US cities with a population above 1 million than in other American cities. According to the first view, large cities should display higher levels of trust; according to Glaeser, lower levels. We conclude that the effect of large cities should be different from zero (H2 – two-tailed test).

(3) Education

It is widely assumed that general trust increases with education (cf., for instance, Uslaner, 2002; Glaeser, Laibson, Scheinkman, and Soutter, 2000; Sturgis and Smith, 2010). This may be true in open societies. However, one can easily imagine conditions under which exactly the opposite effect will occur. If education is religiously controlled, it may foster distrust towards disbelievers and highly educated may be less trustful. The same may hold in authoritarian societies. It is, therefore, not surprising that Delhey and Newton (2003) report mixed findings. There are various arguments, however, why education might be conducive to general trust under the rule of law in an advanced democracy. First, high schools and universities in these societies might propagate universal human values and cut down on ingroup-outgroup thinking. Second, highly educated people have better skills for analyzing the incentive structure and personal traits of their partner. Third, they also have better means of controlling other people either by persuasion or setting appropriate incentives. Fourth, highly educated people are usually economically better off and have comparably less to lose. Therefore, they can take higher risks, that is, express higher levels of trust. On balance, we expect education to have a strong positive impact on general trust in Japan (H3).

(4) Gender

While the variables mentioned above display clear trends, the proportion of female respondents has not changed very much in the JNCS. As in other surveys, women are slightly overrepresented in the JNCS: 56% of the respondents were women in 1978, and 54% in 201316. As the composition of men and women remains the same, composition effects on general trust cannot be expected. Suppose, however, that women have been more trustful than men in the past and that in the course of history men have gradually adapted to the higher trust levels of women. Then we would make two observations: the level of trust in the whole society would increase, and the gender effect would disappear. There are various hypotheses about the relationship between gender and general trust (cf. Romano, Balliet, Yamagishi, and Liu, 2017; Ermisch, Gambetta, Laurie, Siedler, and Uhrig, 2009). We do not assume that women are always more trustful than men. On the contrary, in societies where the public sphere is reserved to men and women have to take care of children and households, men have better chances to make new contacts. As long as these contacts are non-aversive, men should display higher levels of general trust. This situation may entirely change with increasing female work participation if women work predominantly in the service sector and have more contacts with new customers. Japan has probably belonged for a very long time to the first type of society where women were less trustful than men. This may have changed during the last decades, but we do not know when. We, therefore, make the first hypothesis about gender, namely, that women and men do not differ in general trust (H4a—two-tailed test). If there is any difference, we assume that women were less trustful than men only in the early surveys (H4b).

(5) Age

Among the socio-demographic trends that have occurred during the last decades, the aging of Japanese society is a central topic. This change is also visible in our data. While the average age is 43.3 years in the 1978 survey, it has increased by about ten years (= 52.9) in 2013. Thus, if general trust increases with age, this change in the composition of the population would result in a steadily growing proportion of high trustors.

Very different kinds of relationships between age and trust have been reported in the literature (see Delhey and Newton, 2003 with further references; Clark and Eisenstein, 2013; Ermisch et al., 2009; Glaeser et al., 2000; Robinson and Jackson, 2001; Sturgis and Smith, 2010). Whether age is positively related to general trust depends on our life experience. In a secure environment and under the rule of law, people may gradually learn in adulthood that promises are more often kept than broken. Obedience of laws may also increase with age. In Japan, crime rates are very low and Japanese, after World War II, should have made the experience on average that most other people are reliable. Accordingly, we expect a positive relationship between age and general trust in Japan (H5). We leave open whether marginal trust diminishes in the later stages of life so that the relationship is only monotonically increasing but not linear17.

(6) Generations and cohorts

In this study, we assume that World War II and the rise of Japan to one of the most advanced societies have resulted in three Japanese generations. The most dramatic historical event, which in our view had lasting effects on a huge group of people, was the end of World War Ⅱ. After the war, the world was no longer the same. The defeated authoritarian systems had propagated nationalism, ingroup trust, and parochial altruism but had also taught that the country was surrounded by enemies who could not be trusted. We distinguish between

 • a broadly defined prewar generation, born before 1939, which experienced the end of the war in youth or adulthood,

 • a generation of people born shortly before, during, or after the Japanese defeat in World War Ⅱ (born between 1939 and 1948), which we call the war generation for the sake of brevity, and

 • a postwar generation, born between 1949 and 1968 who experienced the rise of Japan as one of the most advanced societies during their formative years.

We expect that the prewar generation displays the lowest level of general trust (H6a) because it suffered most from the defeat and the following occupation. Members of this generation often maintained the old authoritarian world view and had difficulties adapting to the new open society. Members of the postwar generation, by contrast, should display the highest level of general trust (H6b) because the rise of Japan to one of the most advanced societies in the world took place during their childhood and youth. This development probably increased their self-esteem and general trust. Members of the war generation were raised under the influence of the old traditional worldview but also experienced the functioning of democracies and open markets. They should display an intermediate level of general trust (H6c).

The end of the phenomenal rise of Japan was reached with the breakdown of the so-called bubble economy in the nineties—a decade that is often called the “lost decade” in Japan. It seems clear that this event and the subsequent economic crises let to a decline in general trust in all age groups. We are less sure, however, whether the development had lasting effects on the younger Japanese people, who grew up under worsening economic conditions. We are, in other words, insecure about the emergence of new generations after 1968 and speak instead of cohorts which, in contrast to generations, need not to be imprinted by unique historical experiences. We, therefore, distinguish between cohorts18:

 • the cohort born between 1969–78 which experienced the lost decade and the following economic crisis in their late twenties and thirties, and

 • the cohort born after 1978 which was fully struck by the economic crises during the lost decade and afterward.

If the lost decade and the worldwide economic crisis of 2008 had a lasting effect on the 1969–78 cohort, the members should display a lower level of trust than the postwar generation (H6d). The members of the youngest cohort experienced the economic crises in their formative years. Therefore, it seems a little bit more likely that they will remain more distrustful in later life (H6e).

(7) Economic security

Our survey sequence ends in 2013. Therefore, we cannot definitely determine whether the economic development around the turn of the century has an enduring impact on the In economic crises, goods become scarce, and the less people have, the more cautious and careful they are in handling their resources. It is, therefore, near at hand that general trust declines under deteriorating economic conditions. We will use two economic macroindicators to investigate this relationship: the Nikkei index, which reflects the breakdown of the bubble economy in the 1990s fairly well, and the youth unemployment rate, which was a topic of many discussions in Japan. General trust should be positively affected by the Nikkei index (H7a) and negatively by the youth unemployment rate (H7b). The hypotheses are summarized in Table 1.

Table1. Summary of the hypotheses.

5. Empirical analyses

Our dependent variable is the dichotomous Rosenberg T-item. The codebook includes an undefined category “other” which we define as a missing value. From the community size variable, we create two independent variables, RuralCom and LargeCom, to test whether people in rural areas have less general trust on average and whether residents of large cities differ from other residents on average. We collapse the two lower levels of education into the baseline category and specify the codebook categories of secondary education and tertiary education as Educ_2 (secondary education) and Educ_3 (tertiary education). For testing the effects of gender and age, we use the variables Female and Age. Generations and cohorts are specified as dummy variables: War2 represents the war generation, Postwar the postwar generation, Coh69-78 the cohort born between 1969 and 1978, and Coh79+ the youngest cohort born after 1978. The baseline is the Prewar generation. In the second part, we introduce the logarithmically transformed Nikkei 225 index (Ln(Nikkei)) and the unemployment rate of young people of age 15 to 24 (UnEmp 15-24). The operationalization of all variables will be reported as an additional material available on the Web site of one of the authors. (https://www.ism.ac.jp/˜maeda/)

5.1. Binary logistic regression analysis on the individual level

We first check the individual-level relationships with the pooled data set of all seven surveys. In this way, we can clarify whether the effects of independent variables remain constant throughout the whole period. If they change in a theoretically plausible way, we can specify appropriate interaction terms in the final model. The estimation in Table 2 performs, so to speak, a screening of the effects and prepares the specification of the final two-level model. Note that the cohort 1969–78 participated for the first time in survey 1993, and the cohort 1979+ in JNCS 2003. In a preliminary step, it turned out that the cohort 1979+ did not differ from the prewar generation in general trust, so we dropped this variable from all our models.

There are several possible ways to check whether the variables have the theoretically expected effect and whether these effects remain stable over time. We apply the ANCOVA approach (Steenbergen and Jones, 2002; Snijders and Bosker, 2012) to the pooled data set, which allows the estimation of not only the main effects across all surveys but also the deviations from these effects in each survey. In this way, we can recognize whether the effects are stable over time. It has the additional advantage that we control for possible heteroscedasticity across the surveys as time contexts in which the response behavior is embedded. These goals are reached by a) fixing a baseline survey, b) representing each of the remaining surveys by an effect-coded variable19, and c) creating interaction effects by multiplying each independent variable X by these effect-coded variables for the survey year.

The estimates20 of the ANCOVA model are reported in Table 2, with the main effects in the first column and the first order interaction effects with the survey years in the following columns. All significance levels refer to two-tailed tests. If the main effects are significant in a two-tailed test, they would of course also be significant in a one-tailed test. Two-tailed tests are appropriate for all interaction effects because we do not expect deviations from the main effects in the first step. The specified interaction terms allow for unspecified short-term period effects on general trust in all surveys (baseline: 2013). By and large, the results are in line with our theoretical expectations. General trust seems to be on average significantly lower in rural communities (H1a: b = −.128) but not in larger cities because the respective coefficient (H2: b = .006) is not significant. People with secondary education (H3: b = .479) and, more so, those with tertiary education (H3: b = .982) are on average markedly more trustful than people with primary education. Contrary to hypothesis H4a, women are significantly less trustful than men on average (H4a: b = −.116). Age has a significantly positive effect on general trust (H5: b = .011). The generation and cohort effects are in the predicted rank order (H6a-6e): General trust is lowest in the prewar generation (baseline), highest in the postwar generation (Postwar: b = .454), and on an intermediate level in the war generation (War2: b = .328). The war generation is closer to the postwar generation, but the difference is still significant on the 5% level in a one-sided test21. There is an indication that general trust declines in the two younger cohorts. Japanese born between 1969 and 1978 already display a lower level of general trust on average than the postwar generation (Coh69-78: b = .386). The decline is most dramatic among those born after 1978 because the respective cohort effect does not become significant at all—which means that the youngest cohort does not differ from the prewar generation in terms of the general trust.

Table2. Screening of the relationships between general trust and the independent variables: Regression coefficients and significance levels of the binary logistic regression of the pooled JNCS (ANCOVA-Model; N = 11,093; listwise deletion of missing data).


The model has been estimated with SPSS (Version 25)

The interaction terms indicate deviations from the main effects in the respective survey.

Significance levels refer to two-tailed-tests:

1) p < 0.01;

2) p < 0.05;

ns not significant;

∗) For the main effects of these variables two-tailed hypotheses have been stated:

List of abbreviations (For (0,1)-coded dummy variables D, the name of D = 1 is displayed)

RuralCom: Rural communities (test of H1a & H1b);

LargeCom: Cities with more than 500,000 residents (test of modified H2);

Educ_2: Secondary education;

Educ_3: Tertiary education (test of H3);

Female: Respondent is a woman (test of H4a)

Age: Age of the respondent (test of H5)

War2: Generation born between 1939 and 1948 (test of H6a-H6e);

Postwar: Generation born 1949 to 1968 (postwar generation) (test of H6a-H6e);

Coh69-78: Cohort born between 1969 and 1978 (test of H6a-H6e);

Most interaction effects in the middle of the Table are insignificant. Nevertheless, we must scrutinize them because they may indicate change over time. For instance, the first two interaction effects of RuralCom are negative. If we argued in line with hypothesis H1b that rural-urban differences were only valid until the mid-eighties and specified the main effect accordingly, the latter would increase in magnitude. On the other hand, we would obtain a similar result if we specified an effect of RuralCom only for 2003 and 2008 because these two interaction effects are also negative. We tentatively conclude that general trust is still lower in parochial rural areas but that the administrative definition of “rural” is not suited to select this type of community accurately. It, therefore, depends on chance whether a sufficiently large number of parochial rural communities are sampled in a survey. On balance, we stay with the specification of hypothesis H1a but keep in mind that the administrative criterion of rural is suboptimal for the purpose of our empirical study.

General trust is not lower in larger cities (LargeCom). The respective coefficient (b = .006) is not significant. In 2013, however, we estimated a significant effect. We cannot entirely exclude that social conditions in large cities have recently deteriorated. We, therefore, modify H2 and assume that residents in large cities have become less trustful in 2013.

As far as education is concerned, we observe deviations from the main effects in either direction, but none of these effects becomes significant.

The results for females are not unexpected. It turns out that women were more distrustful than men before 1993, but since then the gender difference has disappeared. This can probably be attributed to the increasing work participation of women in the service sector. The strong positive interaction effects in 1993 and 2003 do not prove that women have already become more trustful than men because they reflect deviations from the negative main effect. If we subtracted the main effect from these interaction effects, non-significant positive effects would remain. Accordingly, we specify H4b with regard to the time-point of change: Women were less trustful than men in the seventies and eighties, but gender differences have disappeared since the nineties. The new variable for estimating the effect is Female 78&83.

It is somewhat surprising that all three generation effects seem to melt down in the crisis year of 2008 and remain on a somewhat lower level in the last survey. In 2008, the negative deviations of the postwar generation (−.326) and the cohort 1969–78 (−.490) even became significant at the 5% level though generation effects should be stable over time. This may indicate that the model overestimates generation effects. Apparently, long-lasting differences between generations that were built up in times of prosperity do not persist in severe crisis. We expect to obtain more accurately estimated—and therefore lower—generation and cohort effects in our final model, which includes economic variables.

5.2. The final model with economic macro-variables

Most of the independent variables identified so far lead us to expect a long-term increase in general trust. It is true that the impact of gender is ambiguous because women may have adapted to the higher trust level of men or vice versa. However, ongoing urbanization will gradually erode parochial rural communities and with them the low general trust in these communities. Whether this process is counterbalanced by declining trust in larger cities, which we observe for the first time in 2013, remains to be seen. Age has a positive effect on general trust, and, as the Japanese society gets older, general trust should gradually increase; the same is true for education. Highly educated people are more trustful, and, as the level of education is continuously increasing, general trust should also increase.

Two findings in Table 2 point in the opposite direction: Japanese born between 1969 and 1978 (Coh69-78) are somewhat, and Japanese born after 1978 are markedly less trustful than the postwar generation. But why is that so? We have tentatively assumed that primarily economic factors have lowered the level of general trust in Japan—the breakdown of the bubble economy in the nineties and, according to Japanese standards, the high youth unemployment rate during the last decades. We, therefore, have included the ln-transformed Nikkei-Index22 and the youth unemployment rate as two macro-indicators in the model in Table 3.

This model has a much smaller number of independent variables than the ANCOVA model, but the overall fit has only marginally decreased by about one-half percent. The pseudo-R2 now reach 5.9% (respectively 3.4% or 4.2% — cf. Table 3). The explanatory power of the socio-demographic variable is fairly low. Note that we have modified the hypotheses about the influence of large cities (LargeCom) and gender (Female) and re-specified our variables accordingly (cf. notes below Table 3). The new variables have highly significant effects, and Female 78&83 has the theoretically expected sign. By contrast with Table 2, the effect of the postwar generation has diminished and that of the cohort 1969–78 (Coh69-78) does not even become significant. Apparently, the cohort does not differ from the youngest one and both display a similar level of trust as the prewar generation if we control for the other variables. Members of these two groups with the same education, of the same age, and so on would not differ on average in general trust. Actually, the two cohorts are composed of many more people with tertiary education than the prewar generation so they will also be more trustful on average if we do not control for other variables and only compare the percentages of trustful people in the prewar generation and the two other generations. The effects of all other individual-level variables only marginally differ from those in Table 2 and therefore confirm our hypotheses to the same extent as the main effects in Table 2.

Table3. Final model with micro- and macro-variables (N = 11,093): Binary logistic regression of the pooled JNCS (listwise deletion of missing data).


The model with robust standard error has been estimated with STATA (Version 14); Abbreviations in the column entries: b: Logistic regression coefficient; S.E.: Robust standard error; z: z-statistic; eb and OR: odds ratio; 95% CI of eb: 95% confidence interval of the odds ratios; Lower: lower bound, Upper: upper bound of the 95% confidence interval


LargeCom 2013: respondent was interviewed in 2013 and lived in a large city (test of H1b);

Female 78&83: respondent is female and interviewed in 1978 or 1983 (test of H4b);

Ln(Nikkei): Nikkei-225 index ‘close’ price for that year, logarithmically transformed (test of H7a);

(from Nikkei Indexes (https://indexes.nikkei.co.jp/en/nkave/archives/data))

UnEmp 15-24: Unemployment rate of age group 15 to 24 (test of H7b);

Year 2008: the interview took place in 2008;)

See notes to Table 2 with regard to other variables.

The two economic variables confirm hypothesis H7. As predicted, the Nikkei index is positively related (H7a), and the youth unemployment rate is negatively related (H7b) to general trust. The impact of the two variables is not very strong. An increase in the unemployment rate of 1 unit, for example, would decrease the actual level of general trust by about 5.3% (= 0.947 −1 = −0.053).

The dummy variable for the year 2008 becomes significant. So far, we have no convincing interpretation of this effect. Apparently, the crisis of 2008 led to a sharper decline in general trust than one would expect from the independent variables in our model. Factors such as the Akihabara massacre23 or the discovery of the pension record falsification by the Social Insurance Agency staff, which both occurred in 2008, may have produced additional distrust but we are not able to empirically demonstrate the influence of these or other factors.

6. Conclusions

The JNCS has offered the unique opportunity to investigate the development of general interpersonal trust for a period of more than 35 years. In the early nineties of the last century, the conjecture was wide-spread that long-lasting structural differences in the economies or the political systems of countries might explain differences in social capital and its components, in particular, general interpersonal trust. The development in the US has shown that functioning institutions obviously do not prevent a long-term decline in general trust. Our own analysis also illustrates that general trust in Japan undergoes long-term changes, too, and that the development is relatively independent of institutional arrangements like the dominance in the economy of committed long-term relationships. More specifically, we have shown that, according to the JNCS, Japan had already reached US-American levels of trust at a point in time when other studies reported significant differences between the two countries. A closer examination of the JNCS data suggests that trust in Japan has increased until the early nineties and declined afterward until 2008, particularly in the younger cohorts.

Age and education have positive effects on general trust. People with tertiary education are most trustful, followed by people with secondary education. Though the level of education has impressively increased in Japan, it has not led to a substantial increase in general trust. The deteriorating conditions between 1993 and 2008 have prevented such a development. Whether they have left lasting imprints on the young malleable cohorts can still not be definitely settled. It depends on whether the general trust has already been internalized as a stable orientation, or whether younger people are still sensitive to influences of the social environment.

We have found slightly more distrust in rural areas, but the effect is very weak and not consistent across surveys—presumably, because the administrative criterion of “rural” does not only include parochial rural communities. Contrary to our expectations, distrust is more widespread in larger cities in 2013 but not earlier. Whether this indicates, a new development has yet to be seen. Consistent with our theoretical expectations, we have found a negative effect of gender only in the early surveys. In the seventies and eighties of the last century, women were slightly less trustful than men, but this effect disappears in later surveys.

A caveat needs to be mentioned: Most of our coefficients in Table 3 are highly significant. This, of course, does not mean that they are substantially important. On the contrary, the relatively low goodness of fit statistics indicates that our model has little explanatory power. Under these conditions, it is relatively easy to specify alternative models with different sets of independent variables that display more or less the same goodness of fit. This is particularly true for the economic macro-variables, which are measured at only seven points in time. Cultural factors may explain the observed change in general trust equally well. Our theoretical conjecture is that the economic conditions around the turn of the century weakened general trust in all generations and cohorts, and at the same time generated enduring low levels of general trust in the youngest, still malleable cohorts. The economic conditions produced, so to speak, longer-lasting period effects and persistent differences between the two younger cohorts and the postwar generation. There is no guarantee, however, that the differences from the postwar generation will persist after a profound recovery of the economy. It remains possible that the two youngest cohorts were temporarily struck more heavily by the economic crises and will display similar levels of general trust when the economic conditions sustainably improve. This must be investigated in future surveys. On balance, we have specified a model that best fits our theoretical expectations. There may be equally fitting and theoretically convincing models, but those who argue in this direction have the burden of empirical proof.

There are other topics for further research. An empirically unsolved problem is the framing of the Japanese trust question. We do not know which kind of behaviors respondents imagine when they say that most people can be trusted. We also have assumed throughout our study that the majority of the respondents frame the question in the same way and interpret “most people” as “most Japanese.” In a globalized world, one might object, mainly younger respondents cannot avoid taking foreign people into consideration, such as competitors from China or from other countries in the world. Accordingly, the frame of reference of the trust question is gradually extended to less trusted outgroups. This alone may explain the decline in trust, which we have attributed to economic changes in Japan. Apart from the improvement of trust measurement, we should also place more emphasis on the determinants of general trust and include not only socio-economic variables such as income or socio-economic status in our surveys but also values and psychological traits. This is the only way to improve the predictive power of our models, which currently still is deplorably low.

脚 注
1   We are indebted to three anonymous reviewers for helpful comments.

2   The basic idea behindWeb 4.0 is a symbiotic interaction between humans and machines. As a consequence, humans may also increasingly develop trust relations to intelligent machines.

3   The dichotomies have been changed to 4-point-items (International Social Survey Program) or 11-point items (European Social Survey) in some of the international survey programs, but the European Values Study and the World Values Survey have kept at least two of the three items as dichotomies. For a comparison of the two different measurement instruments, see Jagodzinski and Manabe (2004).

4   Together with the first item, these two items [people are more inclined to help others; people will take advantage of you] are sometimes used as multiple indicators of generalized interpersonal trust. Their inter-correlations are so low in Japan however, that they cannot be used for that purpose. We therefore rely exclusively on the T-item, which comes closest to what we intend to analyze.

5   We define the response category “Other” as a missing value.

6   The JNCS was originally carried out as a single survey. Since 1973 (the fifth survey) two surveys are conducted in parallel, with the K-type and the M-type questionnaire. While the K-type questionnaire includes larger numbers of repeated items since the first survey, the M-type questionnaire includes larger number of new items.

7   Japan was not included.

8   They can be downloaded at https://gssdataexplorer.norc.org/variables/441/vshow.

9   “Freeter” is an expression for a person who takes a series of casual jobs, excluding housewives and students in Japan. “Freeter” is not an occupation but a form of employment. It can also be described as “permanent part-time workers” or “underemployed workers”.

10   10According to an article of Colin Joyce in the Los Angeles Times of September 21, 2003, Toshie Ikenaga, director of the Cabinet Office’s Quality of Life bureau, has explained: “From the 1980s to the mid-1990s, most people chose to become freeters for the purpose of living their lives according to their own interests. But now many have no choice because of the difficult job market. As the economy worsened, people who became freeters in the ’90s found that they cannot escape and cannot acquire job skills. Being a freeter was once a stage, but it is possibly becoming a condition.”

11   Haruhiko Kuroda, https://www.boj.or.jp/en/announcements/press/koen_2016/data/ko161021b.pdf

12   “Gun” (“Machi” and “Mura”) areas as defined by the Japanese Statistical Office.

13   The constellation fits the portrayal of a parochial culture (Almond and Verba, 1963) or of Banfield’s (1958) description of the narrow-minded citizens of a Southern Italian village.

14   This assumption was possibly not met in Putnam’s study because he arrives at the opposite conclusion: “Residents of small towns and rural areas are more altruistic, honest, and trusting than other Americans” (Putnam, 2000: 205).

15   Putnam deals with that problem under the heading of suburbanization.

16   In 2017, the estimated percentage of women is somewhat above 51.5% of the population, due to the higher life expectancy of women. There are still more males than females born in Japan (https://www.cia.gov/library/publications/the-world-factbook/geos/ja.html).

17   An inverted u-shaped relationship seems possible if feelings of insecurity increase with age and would make people more cautious. So far, there is no empirical evidence that these processes really take place in Japan.

18   The word cohort indicates that we have aggregated several years of birth into larger groups without having theoretically convincing reasons for choosing exactly these time intervals. The distinction between generation and cohort expresses our confidence in the classification. We are convinced that the prewar and the postwar generation fundamentally differ in values and experiences though it may still be discussed whether the borders of the birth intervals have been optimally chosen. It seems also plausible that the war generation, socialized by members of the prewar generation but experiencing the dawn of a new era, holds a sandwich position between the prewar and postwar generation. We are much less sure, however, whether the cohorts 1969–78 and the cohort 1979+ differ in values and experiences. In cohort analysis, cohorts are often defined in such a way that the cohort width is equal to the time interval between two surveys. This has good statistical reasons but is rarely theoretically grounded. By using the word “cohort” we want to indicate that these classifications have so far an explorative status. Of course, there are other uses of the term “cohort” in the literature. (cf., for instance, Ryder, 1965).

19   Note that in the present context the expression ‘baseline survey’ has nothing to do with the first or original survey but refers exclusively to the effect coding of the survey year. The baseline survey can be deliberately chosen. The effect-coded variable Dy = −1, if the interview has taken place in the year of the baseline survey, Dy = +1, if the interview took place in survey year y, and Dy = 0 otherwise. We have chosen the seventh survey (survey year 2013) as the baseline. As a consequence, the intercept b07 is equal to the negative sum of the intercepts of the other six surveys (b07 = −Σb0i) and the interaction effect of a variable X with survey year 2013 is: (bX7 = −ΣDy ∗ bXi). We have estimated the ANCOVA-model a second time with the baseline year 1998. In this way we can get the significance levels for the 2013 survey which are reported in the last column of Table 2.

20   The option listwise deletion has been chosen for estimation. A case is excluded from the analysis if one or several variables have missing values. Accordingly, the total number of cases in the pooled data set is larger than the number of cases in the estimation.

21   For performing the significant test either the war generation or the postwar generation have to be chosen a baseline. If the war generation is chosen as a baseline, the effect of the postwar generation is b = 0.125 and the significance level is 0.065 in a two-sided test. It is therefore < 0.05 in a one-sided test.

22   We use a ln-transformation for reducing the effects of outliers.

23   The Akihabara massacre (in Japanese: Akihabara Torima Jiken) was an incident of mass murder that took place on 8 June 2008, in the Akihabara shopping quarter, Tokyo. Seven innocent people were killed and also ten people were injured in this incident. It gave a lot of shock to Japanese public confidence in what is considered a society safe from violent crime.

© 2019 Japanese Classification Society