Tài liệu The Effect of Social Trust and Economic Growth pdf

Aarhus School of Business, Aarhus University

Master of Science in International Economic Consulting

Master Thesis
The Effect of Social Trust and Economic Growth





Author: Lena Pfister
Academic Supervisor: Christian Bjørnskov
September 2010
Abstract:

In recent years, social trust has gained in importance within social science, especially
in an economic growth context. The thesis examines social trust as a potential
determinant of economic growth using a panel data set including 116 countries over
a time span from 1950 until 2005. The findings suggest a strong association between
social trust and economic growth and stay robust throughout a Jackknife exercise
and an extreme bound analysis implying that it is unlikely that these results are driven
by outliers or omitted variables. Reverse causation is ruled out by adopting an
instrumental variable approach. Further findings suggest that the effect of social trust
on economic performance depends, however, on the development level of the
country. Moreover, the analysis provides further evidence that human capital and
legal quality are indirect links through which social trust has an economic effect.
Finally, the thesis gives insight on the individual characteristics of the respondents
that answer the trust question in the affirmative.
Index

I


Index

Index I

Figures and tables II

0 Introduction 1

1 Overview of the literature 3
1.1 Social Trust 3
1.2 Social trust and growth 5
1.2.1 Direct effects of social trust on economic performance 5
1.2.2 Indirect effects of social trust on economic performance 6
1.3 Who trusts others? 8

2 Methodology and Data 10
2.1 Data description 10
2.2 Methodology 16
2.2.1 Basic assumptions 16
2.2.2 Random effects 19
2.2.3 Logit estimation 20

3 Econometric Analysis 23
3.1 Social trust and economic growth 23
3.2 Extreme Bound Analysis 29
3.3 Instrumental variables 31

3.4 Jackknife exercise 33
3.5 Divided sample 34

4 Transmission channels 38
4.1 Human capital 38
4.2 Legal quality 40

5 Determinants of trust 44
5.1 GSS 46
5.3 WVS 50

6 Conclusion 54

Appendix 60

Figures and tables

II


Figures and tables

Figure 1: Social Trust and Log GDP
per capita
1950 23
Figure 2: Social Trust and Log GDP
per capita
2005 24

Table 1: correlation matrix PWT 25

Table 2: PWT regression 1 26
Table 3: PWT regression 2 27
Table 4: EBA 30
Table 5: Sagran Hansen statistic 32
Table 6: Instrumental variables 32
Table 7: Jackknife exercise 34
Table 8: Divided sample 36
Table 9: Human capital 39
Table 10: Human capital, divided sample 39
Table 11: Legal quality 41
Table 12: Legal quality, divided sample 43
Table 13: Overview GSS, WVS 1 44
Table 14: Overview GSS, WVS 2 45
Table 15: Overview GSS, WVS 3 46
Table 16: correlation matrix, GSS 47
Table 17: Social trust, GSS 48
Table 18: correlation matrix, WVS 50
Table 19: Social trust, WVS 51
Table 20: Social trust, WVS, USA/Canada 53
0 Introduction

1


0 Introduction

The notion of social capital started to develop throughout the 20
th
century. It did not,
however, have its breakthrough until 1993 when Robert Putnam published “Making

Democracy Work: Civic Traditions in Modern Italy”. In his book, Putnam investigates
different regions in Italy with the same institutions and governmental structure and
tries to explain why there are nevertheless huge disparities in economic performance
between Northern and Southern Italy. His findings suggest that the economic
disparities are due to different endowments of social capital in the two regions.
Putnam’s work appeared to be a starting shot for social scientists to explore the topic,
since it subsequently enjoyed a surge in popularity. Today a wide literature can be
found and known journals like the “American Economic Review” and “Quarterly
Journal of Economics” publish articles on social capital. It was in the latter that Knack
and Keefer published their paper “Does Social Capital have an Economic Payoff?” in
1997. They were the first to examine different features of social capital separately in
a standard empirical growth framework, and they provided proof that social trust in
particular is positively associated with economic performance. Subsequently, more
papers have been written on the issue, however, relatively few compared to the size
of the social capital literature. Moreover, research within this topic has mainly been
performed on the basis of cross sectional data.
The aim of this thesis is to obtain further insight into the relationship between social
trust and economic performance and thus to contribute to a deeper understanding of
economic growth and social trust. To achieve this goal the analysis is based on a
panel dataset and thus more comprehensive than previous studies. The questions
investigated are:
1. Does Social trust influence economic growth? If so, how and to what extend?
2. Who does trust others?
The thesis is structured as follows: The first chapter gives an overview of the existing
literature and the current state of research. The concept of social trust is introduced
in detail and its measurement is discussed. Subsequently, an overview of social trust
within an economic growth context is given and direct and indirect links through
which social trust might have an economic effect are presented. Finally, the literature
on who trusts others is explored.
0 Introduction

2


The second chapter describes the data used for the analysis. Furthermore, the
methodology applied is amplified. In this context the random effects model, which is
applied in the third and fourth chapter and the logit model, which is applied in the fifth
chapter are introduced. In the third chapter the effect of social trust on economic
growth is investigated. Additionally, to verify the robustness of the results, an extreme
bound analysis and a Jackknife exercise are conducted. Moreover, an instrumental
variable approach is applied to control for endogeneity. The fourth chapter further
examines the association between social trust and economic growth. To shed more
light on how social trust might influence economic growth, the relationship between
social trust and human capital and social trust and legal quality as two potential
transmission channels are analysed. In the fifth chapter individual level data is
applied to find out more about the characteristics of the trusting citizen. Moreover, it
is investigated if the grandparents’ trust levels still influence their grandchildrens’ trust
today. This could give more information about the stability of trust. The analysis of
this chapter aims to get a deeper understanding of social trust, which could be useful
for policy makers. In the conclusion, which is presented in the final chapter, the
results of the analysis are summed up and evaluated.

1 Overview of the literature

3


1 Overview of the literature

In 1993 Robert Putnam published the book “Making Democracy Work: Civic

Traditions in Modern Italy”. He looks into the question why some democratic
governments succeed and why others fail and aims at to contribute to the
understanding of the performance of democratic institutions. He concludes that their
success is based on their endowment of social capital, which he defines as “features
of social organization, such as trust, norms, and networks that can improve the
efficiency of society by facilitating coordinated actions” (Putnam, 1993: 167). Even
though there is no consistent definition for social capital, Putnam’s is the one most
referred to.
Today social capital is well established as a determinant of growth. So much that the
World Bank started a “Social Capital Initiative” in 1996 with the goal of further
investigating the formation of social capital and its impacts on project effectiveness
and development (World Bank, 2010).
However, there have also been discussions about whether to treat social capital as
an entity since its different features have different effects (Bjørnskov, 2006b). Several
papers have shown that the three pillars of social capital, namely trust, cooperative
norms and associations within groups have different or no effect on economic growth,
whereas social trust appears to be the most promising candidate (Knack and Keefer,
1997; Newton, 1999; Whiteley, 2000).
The first chapter is structured as follows: In section 1.1 the concept of social trust is
introduced in detail and the way to measure it is discussed. Subsequently, an
overview of social trust within an economic growth context is given in section 1.2.

1.1 Social Trust

When the concept of social trust gained popularity within social science there were
several discussions about what it consist of and how it can be measured. It became
apparent that it is very important to distinguish between two kinds of trust
(generalized and particularized trust). Social trust is mainly defined as generalized
trust, i.e. it measures how much people trust others about whom they possess no
information. This is opposing to particularized trust or reputation that is based on trust,

1 Overview of the literature

4


which in turn originates from previous experience or information obtained about
others. How important this differentiation is shows the paper from Alesina et al. (2009)
where they confirm Banfield’s (1958) theory of “amoral familiarsm” and show that
there is a negative association between trust within the family and generalized trust.
In recent years the following question has shown to be a good measure of social trust:
“Generally speaking, would you say that most people can be trusted, or that you can’t
be too careful in dealing with people?“ (trust question).This question originates from
the German political scientist Elisabeth Noelle-Neumann who formulated it in 1948. It
was adapted by for instance the General Social Survey (GSS) in 1972, the World
Value Survey (WVS) in 1981 and the Barometers. These are today the main sources
for researchers within this area and the number of countries for which the data is
available rises every year.
Yet, the validity of the trust question was questioned on several accounts. Glaeser et
al. (2000) conducted an experiment with 189 students of the introductory economics
course at Harvard University to investigate the validity of survey questions about
hard-to-measure characteristics like trust and trustworthiness. First the students had
to answer survey questions about trusting attitudes and trusting behaviour.
Subsequently, trust and trustworthiness were measured by experimental behaviour
by playing the Berg et al. (1995) “trust game”. Finally the two results were compared.
They concluded that “standard attitudinal survey questions about trust predict
trustworthy behaviour in our experiments much better than they predict trusting
behaviour” (Glaeser et al., 2000). However, there was a severe problem with the
setup of the experiment. They allowed students who knew each other to play
together. Their reasoning was that non-random pairing procedure generates more
variation in social connections. This had however the consequence that particularized

trust rather than social trust was measured.
In a recent study, Ostrom et al. (2009) repeat the experiment but ensure that the
game is played anonymously. They find that “the response to the survey question
regarding trust is highly significant and in the expected direction [positively
correlated]”, which confirms the validity of the trust question. Another study by
Sapienza et al. (2007) gets to the same result under the condition that the stakes of
the game are sufficiently high. By the same token, they show that trust and
trustworthiness are strongly related. In the trust game they find that “players
1 Overview of the literature

5


extrapolate their opponent’s behavior from their own”, which means that people who
trust others are also more trustworthy.
In general, an increasing support for the trust question as a measure for generalized
trust, trustworthiness as well as a proxy for economically relevant beliefs can be
observed. The latter is covered in the next section.

1.2 Social trust and growth

Already Putnam associates social capital with economic performance. However,
Knack and Keefer (1997) were the first to examine different features of social capital
separately in a standard empirical growth framework (Bjørnskov, 2009a). In a cross
section of 29 countries they show that social trust and civic norms are positively
associated with economic performance whereas being a member of a network shows
no effect. Zak and Knack (2001) found, in conformance with Knack and Keefer’s
results that social trust is positively associated with economic growth, when they
repeated the study with a larger sample of 41 countries. Additionally, they show that
this relationship is causal, i.e. social trust promotes economic growth and is not just a

consequence of it. Bjørnskov (2009a) provides an overview of the existing literature
about social trust and economic growth in the “Handbook of Social Capital”. He gives
an overview of several studies that find a positive association, which differs, however,
in size. Bjørnskov calculates the average effect in these studies and concludes that
an increase of trust by 10 percentage points increases the annual GDP growth rate
by approximately half a percentage point.
The following sections introduce direct and indirect effects social trust might have on
economic activity.

1.2.1 Direct effects of social trust on economic performance

As Knack and Keefer (1997) point out, in an economy many commercial transactions
are determined by mutual trust between two or more parties. This can be a
transaction between parties where goods or services are supplied in exchange for
future payments. As within a company, managers have to trust their employees. In
general, the principle agent problem arises when there is asymmetric or incomplete
1 Overview of the literature

6


information. Thus, need for trust rises with the extent to which the performed task is
not monitorable, which is mainly the case for highly educated employees. Also,
investment and saving decisions are dependent on the assurance of banks and
governments that these assets are protected. The same is valid for the trust of
agents that laws and rights will be abided or otherwise enforced like for example
property rights.
As outlined before, trust and trustworthiness are highly correlated (Sapienza et al.,
2007). This means that the likelihood of dishonest behaviour in commercial
transaction is lower in a high trust society since less money for protection is needed.

Contracts for example do not have to be as specified and cover every contingency
(La Porta et al., 1997). The probability of legal disputes decreases and therewith
reduces the deadweight burdens of enforcing and policing agreements (Whiteley,
2000). Besides, managers can save money on monitoring their employees since they
are more reliable in a high trust society. Moreover, Zak and Knack (2001) show that
social trust and the investment rate are positively correlated. They find that
investment/GDP share rises by nearly one percentage point for each seven
percentage point increase in trust. This might be due to two reasons. Firstly, property
rights are better protected in high trust societies. This increases the return to
investment in innovation and hence incentives to invest in new products. Additionally,
entrepreneurs have to direct less of their resources to protect themselves from
possible dishonest behaviour of their employees and business associates. Secondly,
since trust and trustworthiness create a safer investment climate, agents tend to
invest more and choose investment projects with a longer time horizon, which might
appear too risky in a low trust society. Long term financing is essential for the
infrastructure industry, which in turn is an important driver of economic growth.
In a nutshell, social trust decreases the costs of economic transactions and increases
the investment rate and therewith enhances the economy with a larger capability for
production.

1.2.2 Indirect effects of social trust on economic performance

In addition to the direct influence of social trust on economic growth elaborated
above, social trust might have a positive effect on the quality of institution and thus
indirectly on economic performance (Knack and Keefer, 1997; Whiteley, 2000).
1 Overview of the literature

7

The definition of institution varies across the literature. Here Acemoglu’s definition is
applied, who decides between economic institutions, “which correspond to taxes, the
security of property rights, contracting institutions, entry barriers, and other economic
arrangements” and political institutions, “which correspond to the rules and
regulations affecting political decision making, including checks and balances against
presidents, prime ministers, or dictators, as well as methods of aggregating the
different opinions of individuals in the society “(Acemoglu, 2009, p. 782).
Social trust seems to have influence on both kinds of institutions. It has shown to
improve the quality of formal institutions, legal quality and bureaucratic efficiency
(Knack and Keefer, 1997; Knack, 2002; Bjørnskov, 2006a; Bjørnskov, 2009b), which
in turn influences the security of property rights. As explained in the paragraph above,
the probability of legal disputes decreases in a high trust society. However, if they do
occur laws can be enforced more easily. This leads as well to a safer investment
environment with higher returns and hence to a higher investment rate. Trust has
also shown to lower the corruption rate. This is due to the fact that the necessary
bribe increases with higher trust levels (Bjørnskov, in press).
Furthermore, social trust increases voter turnout and political participation. Knack
(1992) finds that the probability of voting increases by 8.6% if the person responds
positively to the trust question. This has the consequence that politicians have to
account for their actions.
Bjørnskov (in press) tries to get to the bottom the relationship between social trust
and governance. He investigates if the association is due to the political
responsiveness to the demands of the voters or due to a higher supply of honest
politicians. His findings suggest that social trust affects economic-judicial governance
but not electoral institutions.
Also, social trust has shown to be influential on the accumulation of human capital.
Coleman (1988) was one of the first to establish this association. He finds that social
capital increases the accumulation of human capital through the family and the
society. Coleman argues that a higher endowment of social capital eases the transfer
of human capital from the parents to their children. Moreover, he finds that families

with multiple social relations are more likely to be rewarded and sanctioned for their
behaviour and hence to comply with norms and trustworthiness. This decreases the
chance of their children dropping out of school. Coleman’s study was one of the first
to combine these two kinds of capital, yet, it his study is limited to particularized trust.
1 Overview of the literature

8


Knack and Keefer (1997) find in their regressions that social trust influences human
capital accumulation positively. They base this on the idea that the return to human
capital increases in a high trust society, since there will be a higher weight on
educational credentials than on other factors that signal trustworthiness.
Bjørnskov (2009b) formalizes and explores this mechanism further and shows that
the association is causal. He points out that this is especially true for highly educated
employees as their tasks are often hard to monitor. Also, he stresses that hiring costs
in the labour market decrease in a high trust society, since employers rather hire and
trust than closely screen the applicants. Moreover, Bjørnskov points out that a
country needs to have a certain level of technological sophistication before this
mechanism starts to work, since a demand for highly educated workers has to arise
in the first place. In summary, lower costs associated with hiring educated employees
and lower monitoring costs for employers increase the demand for educated worker
in a high trust society and, hence, accelerate the accumulation of human capital.
This association is also detected by others studies but interpreted in a different way.
Alesina et al. (2002) find that trust is positively correlated with education and income.
They suggest that this can be due to the fact that individuals that are successful on a
professional level are more likely to trust others. Glaeser et al. (2000) find in their
sample that trust is higher among well-educated people for which they provide two
possible explanations. Firstly, educated individuals mainly socialize with other
educated that are more trustworthy. Hence, their expectation of trustworthiness gets

in general confirmed. Secondly, their education enhances them with a higher level of
social skills and a higher status, which gives them a better opportunity to reward and
punish others.

1.3 Who trusts others?

After social trust gained popularity within social science, several relationships
between social trust and other variables were investigated. After finding out that high
trust levels enhance for example economic growth and increase institutional quality,
other questions came to light. Where does social trust come from? What are people
who trust others about whom they posses no information about like? And what, in
turn, are those that do not trust others like? Can social trust be created?
1 Overview of the literature

9


Several studies find associations between social trust and individual characteristics
like gender, age, race, income and education (Alesina et al., 2002; Demaris et al.,
1994).
There are mainly two different ideas about how a person’s trust level is determined.
One states that if a person trusts or not depends on his surroundings and
experiences he makes in his everyday life. So this would mean that highly educated
people with a higher income are more trusting because they have more positive
experiences in life and have it easier, which in turn creates higher trust level (Glaeser
et al., 2000; Alesina et al. 2002).
The other concept states that social trust is something that we learn from our parents
and that is more or less determined after having obtained it or not if no traumatic
experience occurs. As a consequence, trust levels in countries are more or less
stable over time and thus can not be actively increased by policies. This means that

trusting people do not become trusting because they experience positive things.
They experience positive things because they are trusting. So people who are more
prone to trusting others have more positive experiences in life and are more likely to
succeed in educational institutions and later on in their professional lives (Uslaner,
2008; Bjørnskov, 2007).
2 Methodology and Data

10


2 Methodology and Data

2.1 Data description

The dataset that serves as a foundation for this thesis is an unbalanced panel. It
includes data of 116 countries from 1950 until 2005. It is presented for averages of
five years' sub-periods.
One of the variables of main interest is the measure of social trust. The social trust
scores are measured in how many percent of the population of a country answer the
trust question in the affirmative. They are obtained from the five waves of the World
Values Survey (WVS) conducted between 1981 and 2007. The WVS obtains data
from 97 societies covering 88% of the world population. Its goal is to capture
changing values and beliefs and hence provide a base for researchers to study the
impact of these on for example economic or political development. The survey
started out as the European Values Survey. That is why in the older waves data from
European and developed countries in general is dominating. However, the number of
countries investigated increased with the number of waves and most recent waves
cover a variety of countries providing a range from very poor to very rich countries.
To attain a bigger sample further trust scores are taken from the Afro Barometer,
Asian and East Asian Barometers, LatinoBarometer and the Danish Social Capital

Project. The basic assumption is that social trust is stable over time. Bjørnskov (2007)
points out that this is based on the idea that a person’s trust is established at a young
age and therewith often a reflection of the parents’ trust through socialization.
Furthermore, it can be expected that an equilibrium like this is self-enforcing since
trust and trustworthiness are highly correlated, meaning that an individual’s
expectations get a reality check and are likely to alter if others do not behave as
expected. He repeats the stability exercise introduced by Volken (2002) and finds
that social trust appears to fluctuate around stable levels. Recently this gets further
confirmation by studies that show that second and third generation immigrants’ trust
levels correlate with those of their ancestors. One of the first to show this was
Uslaner (2008) who came to the conclusion that “where your ancestors came from
matters more for trust than who your neighbours are now (p.7)”.
2 Methodology and Data

11


On average there are about three trust observations per country, whereas there are
more observations for richer than for poorer countries. In the sample the average of
all available observations is used and assumed to be the same for all years. The trust
scores range from a low of 3.4% in Cape Verde to a high of 64.3% in Sweden with
the average trust being 25.5%. A full list A1 of all the countries and their trust scores
can be found in the appendix. Other control variables that are taken from the WVS
are education, religiosity and income of the respondent. These variables are applied
in the analysis about the determinants of trust. Education is measured on a scale
from 1 to 8 reaching from “inadequately completed elementary education” to
“university with degree/higher education”. Income is measured on a scale from 1 to
10 with 10 being the highest income. Religiosity is a dummy variable that takes on
the value 1 if the respondent considers religion as an important part of his life and 0
otherwise.

Another source for trust scores is the general social survey (GSS), which is
conducted for the United States of America by the National Opinion Research Center
(NORC) of the University of Chicago on an annual base. It conducts basic scientific
research on the structure and development of American society by asking
demographic, behavioural, and attitudinal questions, plus topics of special interest. It
covers the years between 1972 and 2008 with only a few exceptions. The trust
scores are measured the same way as in the WVS. However, the GSS is applied for
individual level analysis, so social trust serves as a dependent variable taking on the
value 1 if the respondent trusts others and 0 otherwise. Moreover, other individual
characteristics are obtained from the WVS, i.e. education, income, religion and skin
colour. Education is measured in the highest year of school completed ranging from 0
to 20 with and average of 13 years. Income is measured in family income and divided
into 12 subgroups where the lowest income is less than 1,000 USD and the highest
more that 25,000 USD. About 5% of the respondents refused to answer the question.
Religiosity is measured in how often the respondent attends religious services. There
are nine different answers to choose from ranging from to “never” to “more than once
a week”. The dummy on skin colour indicates if the person is black or not. About 13%
of the respondents consider themselves as black, which is representative with the
American population. Also, age cohorts are generated dividing the respondents into 7
groups of equal size to control for the time period they were born in.
2 Methodology and Data

12


Another important source for some of the key variables is the Penn World Table
(PWT), which is a set of national accounts economic time series. The variables are
provided in a common set of prices and currencies, which allows for a direct
comparison of the different countries. The newest version 6.3 covers 189 countries
over a time period from 1950 until 2007 with 2005 as base year. As a measure of

economic performance, the real GDP
per capita
is used. Several variables that have
shown to be determinants of growth are used to isolate the effect of social trust on
growth. The PWT table provides data on the consumption, government and
investment share of GDP. As an indicator for the openness of a country, the exports
plus the imports are divided by GDP. This is applied as a measure on how much a
country is exposed to international competition and thus how big the incentives are to
increase productivity. Moreover, the population of the countries measured in 1000s is
included in the dataset.
Two of the control variables are from the Major Episodes of Politcal Violence (MEPV)
data set, which is compiled by Monty G. Marshall the director of the Center for
Systemic Peace. It covers major armed conflicts in the world over the period 1946-
2009, which are characterized in different types, e.g. civil or ethnic conflict, inter-state
or intra-state. Furthermore, the magnitude of the conflicts on the directly-affected
society or societies is evaluated on a scale of 0 (smallest) to 9 (greatest) for the civil
conflicts and on a scale from 0 (smallest) to 6 (greatest) for the international conflicts.
They serve as an indicator if economic activity was limited due to violent conflicts.
Further control variables are obtained from the World Bank’s World Development
Indicators (WDI), which cover more than 900 economic, social and environmental
indicators for 210 economies between 1960 and today. Fertility rate and population
density are used as potential determinants of growth. The former is measured in
births per woman ranging from 0.93 in Hong Kong in 2005 and 8.25 in Rwanda in
1980 with an average of 3.76. The population density is measured in people per
square km with a minimum of 0.66 in Mongolia in 1965, a maximum of 6484.81 in
Hong Kong in 2005 and an average of 176.15.
As a measure of human capital, data from the Barro Lee dataset is employed. It is
available for 142 countries from 1950 to 1995 with projections to 2000, disaggregated
by sex and by 5-year age intervals. Moreover, it accounts for the distribution of
educational attainment in the population in, for most instances, six categories: no

formal education, incomplete primary, complete primary, first cycle of secondary,
2 Methodology and Data

13


secondary cycle of secondary, and tertiary (Barro and Lee, 2001). The values for
complete primary range from 0.4% in Benin in 1970 to 67.1% in the UK in 1960. For
complete secondary from 0% in Zimbabwe in 1970 to 47.5% in Austria in 1980.
Another factor that might determ economic growth is inequality. The gini coefficient is
obtained from the World Income Inequality Database, which is a compilation of
several gini coefficient sources, e.g. World Bank’s Deininger and Square database. It
includes 5313 observations from 159 regions or countries reaching more than 100
years back. It is, however, fragmentary. To make the coefficients comparable, only
those calculated on the basis of gross income and consumption based on a
representative sample covering all of the population are used. Moreover, a score of
6.6 is added to the consumption-based gini coefficient to make them comparable to
the income-based (Deininger and Squire, 1996). The coefficient ranges from 18.7 in
the Czechoslovakia in 1990 to 80.5 in Namibia in 1995.
Also, a measure for economic freedom from the Fraser Institute is applied. The
Fraser Institute publishes an annual with a measure of the degree of economic
freedom in 141 nations. It is constructed by using forty-two data points in five broad
areas: 1 Size of Government: Expenditures, Taxes, and Enterprises; 2 Legal
Structure and Security of Property Rights; 3 Access to Sound Money; 4 Freedom to
Trade Internationally; and 5 Regulation of Credit, Labor, and Business. The rating
takes on values between 0 and 10, with a higher rating indicating a greater degree of
economic freedom. The minimum value in the sample is 2.3 in Nicaragua in 1985, the
maximum value 9.08 in Honk Kong in1995 and a mean of 6.06. Moreover, the
measure for legal structure and security of property rights is applied by itself.
The dataset contains five regional dummies for the analysis on the country level,

namely Europe, Asia, Africa, South America and North America to control for effects
specific to regional differences. For the analysis on the individual level based on the
WVS six regional dummies are included, i.e. Africa, Asia and the Pacific, Latin
America, post-communist countries, the Middle East and North America. They are
binary variables taking on the value 1 if a country is from the respective region and 0
otherwise. Furthermore it includes year dummies to allow for different intercepts
across years and to capture the aggregate time effects.
Finally, the dataset includes several variables that are candidates for instrumental
variables, fulfilling the condition to be correlated with social trust but not with a
measure of economic growth. Bjørnskov (2007) suggests that social trust is positively
2 Methodology and Data

14


associated with a country being a monarchy, which might be due to the following
factors. Having a royal family is something that the whole population has in common
despite their social background, which provides a national feeling and a strong sign
of unity. It also might endow with social and political stability, since it is not as
temporary as presidencies. Furthermore, Bjørnskov points out that high trust
countries like for example the Scandinavian countries and the Netherlands have a
rather peaceful political history and that their royal families have been highly
accessible to the public throughout history. In Denmark, neither a leading politician
nor a royal family member has been assassinated since 1284, which has been a
different story for many other European countries. This moreover indicates that social
trust is rather stable over time.
A first look at our data provides more evidence for this theory. The average trust
score of the 18 countries being monarchies in the sample is 40.5% whereas the
average for the remaining countries is 22.8%. It should be pointed out that the high
average for monarchies is especially due to the trust scores of the three

Scandinavian countries, which are all above 60%. However, also most of the other
monarchies have a trust score above average compared to other countries in the
same region.
Also, Bjørnskov (2007) introduces the post-communist (postcom) dummy variable,
which takes on the value 1 if the respective country has been a Central or Eastern
European communist state and 0 if otherwise. The idea that the postcom dummy and
social trust are correlated is based on the dictatorship theory of Paldam and
Svendson (2001). Most communist countries had a suppressive government with an
intelligence apparatus with an enormous number of informants among the population.
With that kind of surveillance and potential danger to be discredited it seems only
natural to only trust people, who are close to you and that you have a lot of
information about. Thus, Paldam and Svendson reason that communism lead to
deterioration of social trust in East and Central Europe. In favour of this theory is the
example of Germany. When Germany was divided, Eastern Germany’ had a
repressive government that easily punished people that were accused to be political
dissidents and its secret service, the Ministerium für Staatssicherheit (Stasi),
observed its population in an unprecedented way. During its existence about 624.000
unofficial employers were engaged (Müller-Enbergs, 2008). This part of history
2 Methodology and Data

15


reflects in the trust scores today. While Germany as a whole has trust score of 37.7%
Eastern Germany has only a trust score of 25.8%.
Another candidate for an instrumental variable is the pronoun-drop dummy variable
following Tabellini (2008). It takes on the value 1 if the personal pronoun in a
language can be dropped and 0 otherwise. In some languages the personal pronoun
is only used when the speaker intends to emphasize who is doing something but is
generally left out. In these cases the verb is normally conjugated so that it reflects the

grammatical person. The idea is that in cultures where the language allows the
pronoun-drop, the importance of the individual and his rights are relatively low
(Kashima and Kashima, 1998), which in turn leads to a low trust level.
Finally, the average temperature of the coldest month of the year is applied. There is
a consensus that norms develop the best where their payoff is the highest. This
means that social trust has most likely developed best in countries were it had the
highest payoff. In countries where the winters are very hard, farmers were more
dependent on each other in form of collective action and mutual insurance. Today
agriculture is far more advanced and not as dependent on the climate anymore,
furthermore, the share of the population being involved in jobs that are related to the
weather has shrunken to a minimum. However, trust levels seem to be stable over
time and to have survived the technological revolution that made trust less important
in the question of survival. This would be an explanation of the north-south divide in
trust levels in Europe, since the winters become milder with approaching the south.



2 Methodology and Data

16


2.2 Methodology
The following paragraph describes the properties of panel data and the possibilities
and problems that arise in that context. Furthermore, the estimation method of
random effects is elaborated. Finally, the logit estimation model is introduced. The
paragraph is based on Wooldridge (2002, 2006) if not stated otherwise.
2.2.1 Basic assumptions
The starting point for the analysis is a multiple linear regression model as follows:
Y

t
=β
0
+ β
1
x
t1
+ β
2
x
t2
+ … + β
k
x
tk
+ u
t
, t=1, 2, …, n (3.1)
where y is the dependent variable and x
1
, x
2
, …, x
k
are the independent variables
that determine y. The error term denoted as u includes the factors that affect y other
than x
1
, x
2

, …, x
k
. It can include omitted variables and measurement errors. The
independent variables can be observed within the sample, whereas the goal is to
estimate the coefficients β
1
, β
2
, …, β
k.

A panel data set contains a time series for each cross-sectional unit in the data set.
This means that the same units, which can be individuals, cities, firms, countries and
so on, are followed over a certain time period. It gets obvious from equation 3.1 that
β is the same in each time period whereas x can but does not have to change over
time. An i subscript can be added to refer to specific observations within the sample
and a t subscript refers to different time periods.
For the remaining methodology part the equations are written in vector form.
Equation 3.1 can now be expressed as:
y
t
= x
t
β + u
t
, (3.2)
where the vector x
t
=(1, x
t1

, …, x
tk
)

is defined as a 1 x (k+1) vector for each t and

β=( β
0
, β
1
, …, β
k.
) is defined as a (k+1) x 1 vector for all parameters.
To ensure consistent estimators for pooled OLS the following assumptions are
sufficient:
2 Methodology and Data

17


Assumption 1
E (x
t
’u
t
) = 0, t=1, 2, … , T. (3.3)
Assumption 1 holds if the error term u is uncorrelated with the independent variables
in the respective time period. It does, however, not imply information about the
association between x
s

and u
t
for s≠t. In this case we speak of contemporaneous
exogeneity instead of strict exogeneity, where the error term at each time is
uncorrelated with the independent variables in each time period. The latter condition
is required to obtain unbiased estimators.
Assumption 2
Rank
[
]
=
∑
=
(3.4)
Assumption 2 holds if there are no perfect linear dependencies among the
independent variables. This does not rule out correlation between the independent
variables in general, only that one of the variables is an exact linear combination of
the other variable.
Assumption 3
σ=
t=1, 2, … , T, where )(
22
t
uE=
σ
for all t (3.5a)
= , t≠s, t, s=1, 2, … , T (3.5b)
For assumption 3.5a to hold the errors have to be homeskedastic. This means that
Var(u
t

|x) cannot depend on x and that Var(u
t
) has to be constant over time. 3.5b
holds if the conditional covariance of the errors is zero across different time periods.
However, for panel data it can not be expected that the observations are
independently distributed across time. Often, there are time-constant unobserved
attributes of the units that cause problems within the estimators, called the
unobserved effect c. These can be individual attributes, company attributes,
geographical location and so on, depending on the kind of observation unit. As
before, the idiosyncratic error u includes unobserved factors that change over time.
2 Methodology and Data

18


Based on Equation 3.2 the unobserved effects model for a randomly drawn cross
section observation i takes on the following form:
y
t
= x
it
β + v
it
(3.6)
where
v
it
= c
i
+ u

it
, (3.7)
Again, it can be seen from the notation that x can change across i and t even though
it does not have to. u changes across i and t, whereas c changes only over t.
By definition, the conditional expected value of u given x is 0
E (u
t
|x
t
,c) = 0, t=1, 2, … , T (3.8)
This implies that
E (x
'
t
u
t
) = 0, t=1, 2, … , T (3.9)
However, 3.8 says nothing about the correlation between c and x. If E(x
'
t
c) ≠0 the
OLS estimators would be biased and inconsistent. But even if E(x
'
t
c) =0 another
problem occurs. When running a pooled OLS regression the fact is ignored that c is a
part of the composite error term in each time period. Since c is constant over time it
leads to serial correlation in the error term and thus to incorrect standard errors.
There are several methods that can be applied to deal with this problem. The two
most common ones for estimating unobserved effects panel data models are fixed

effects estimation (FE) and random effects estimation (RE).
The main difference between these two methods is the basic assumption about the
relationship between the unobserved effect c and the explanatory variables. The
unobserved effect is called a random effect when there is no correlation and fixed
effect when those two are correlated. Or expressed in a formula:
Random effects
Cov(x
it
,c
i
) = 0, t=1, 2, … , T (3.10)
2 Methodology and Data

19


Fixed effects
Cov(x
it
,c
i
) ≠ 0, t=1, 2, … , T (3.11)
However, the FE has one major disadvantage over the RE. It does not allow for
variables that are constant over time since the estimator subtracts the time averages
from the corresponding variable. As one of the key variables in the following analysis
is constant over time, namely social trust, FE can not be applied. Thus, only the RE
method is described further in the following paragraph.
2.2.2 Random effects
Using equation 3.6 as a starting point, RE imposes the following assumptions to
obtain unbiased and consistent estimators

Assumption 1
E (u
it
|x
i
,c
i
) = 0, t=1, 2, … , T (3.12a)
E (c
i
|x
i
,) = E (c
i
) = 0, t=1, 2, … , T (3.12b)
where x
i
≡(x
i1
, x
i2
, … , x
iT
)
Assumption 3.12a holds if x
i
is strictly exogenous conditional on the unobserved
effect c
i
. This implies that x

is
has no partial effect on y
it
for s≠t, once x
it
and c
i
are
controlled for. Hence, the difference to pooled OLS is that the expected value of y
it

does not only depend on the explanatory variables but also on the unobserved effect.
Assumption 3.11b holds if c
i
and x
i
are orthogonal.
Equation 3.6 expressed for all time periods takes on the following form:
y
i
= X
i
β + v
i
(3.13)
The unconditional variance matrix of v
i
is defined as
Ω≡E(v
i

,v
'
i
) (3.14)
To obtain consistent estimators assumption 2 has to hold
2 Methodology and Data

20


Assumption 2
rank E(X
'
i
Ω
-1
X
i
) (3.15)
The first two assumptions are sufficient to obtain consistent estimators. However, the
serial correlation in the error term is not taking care of. Therefore, Assumption 3
takes the unobserved effects structure of v
it
into account
Assumption 3
E(u
i
u
'
i

|x
i
,c
i
)=σ
2
u
I
T
(3.16a)
E(c
2
i
|x
i
)= σ
2
c
(3.16b)
Assumption 3a implies that the idiosyncratic errors u
it
have a constant unconditional
variance across t and that they are serially uncorrelated. Assumption 3b is a
homoskedasticity assumption on the unobserved effect. Based on this we can
assume that the variance matrix v
i
conditional on x
i
is constant
E(v

i
v
'
i
|x
i
)=E(v
i
v
'
i
) (3.17)
Furthermore, Ω can be derived, which takes on a special random effects structure:
Ω=σ
2
u
I
T
+ σ
2
c
j
T
j
'
T
(3.18)
It is used to calculate the random effects estimator:
∑∑
=

−
=
−−
=
(3.19)
which is applied in the analysis in chapter 3 and 4.

2.2.3 Logit estimation

In this section the logit model is introduced. Contrary to the fixed effect model, the
logit model is designed for limited dependent variables. The range of values these
2 Methodology and Data

21


variables can take on is restricted in a fundamental way. This section only deals with
the case of binary independent variables, which implies that the independent variable
can only take on two values. In most of the cases these are 0 and 1. The binary
response models are about the response probability that takes on the following form:

P(y=1|x)=G(β
0
+β
1
x
1
+ … + β
k
x

k
)=G(β
0
+xβ), (3.20)
where 0<G(z)<1 for all real numbers z

In the logit model G takes on the logistic function, which ensures that P is strictly
between 0 and 1:

G(z)=exp(z)/[1+exp(z)]=Λ(z) (3.21)

The interest in this case does not lie in the effect of x
j
on y but on the effect of x
j
on
the response probability P(y=1|x). Since y rarely has a clear unit of measurement,
the maginitude of β
j
is not very useful. To obtain the effect of x
j
on P(y=1|x) = p(x), its
partial derivative has to be calculated:

=
∂
∂
g(β
0
+xβ) β

j
, where g(z)
dz
dG
≡
(z) (3.22)
g is a probability density function, since G is the cumulative distribution function of a
continuous random variable. Also, G )(
⋅
being a logistic function has a strictly positive
slope. Therefore it follows that g(z)>0, which has the consequence that the sign of
the partial effect is determined by the sign β
j
, This means that the direction of the
effect of x
j
can be obtained before calculating the partial derivative.
To get the results, a maximum likelihood estimation is applied. Therefore, the density
of y
i
given x
i
is required:

F(y|x
i
;β)=[G(x
i
β)]
y

[1- G(x
i
β)]
1-y
, y=0,1 (3.23)

The log-likelihood function is derived by taking the log of function 3.23:

i
l
( β)=y
i
log)[G(x
i
β)]+(1-y
i
)log[1-G(x
i
β)] (3.24)