The impact of foreign banks on private credit: The role of
home-host country cultural distance
University of Groningen Faculty of Economics and Business
Msc Thesis International Economics and Business
Name student: Dirk Brandwijk Student ID number: S2568764
Student e-mail: d.m.brandwijk@student.rug.nl Date: June 9, 2015
1 ABSTRACT
This paper analyzes the effect of home-host country cultural distance on the relation between foreign bank presence and private credit extended by banks in an economy. It is argued that ‘liability of foreignness’ will increase a foreign bank’s dependence on hard information customers, and that this dependence will be exacerbated by cultural distance. The study finds no strong evidence that cultural distance negatively affects the relation between foreign bank presence and private credit. Perhaps other types of distance can explain this relation better, or foreign banks face distance constraints regardless of their country of origin.
TABLE OF CONTENTS
INTRODUCTION ... 1
LITERATURE REVIEW ... 2
Private credit ... 2
Foreign banks and private credit ... 3
Cultural distance ... 4
Cultural distance and foreign bank lending... 5
METHODOLOGY ... 6
Methods in other literature ... 6
A first look at the model ... 7
The sample ... 8
Dependent variable ... 8
Main independent variables ... 9
Foreign bank presence ... 9
Cultural distance ... 10
Control variables ... 13
RESULTS ... 14
Trends in the data ... 14
Data description... 16
Empirical results ... 19
Robustness checks ... 28
Threshold effects ... 33
Discussion of the results ... 37
CONCLUSION ... 39
BIBLIOGARPHY ... 41
1 INTRODUCTION
In the past decades, globalization has increased at an unprecedented pace, also in the financial sector. International capital flows and cross-border banking activities have surged dramatically. Many banks have used this opportunity to expand and have established presence in foreign markets. However, this increasing globalization has also been accompanied by a number of financial crises, particularly in developing countries. This has led to increasing concern about the effect that the entry of foreign banks can have on development of the financial sector and financial stability.
Much of the research on the conditions under which foreign banks bring a positive contribution, has focused on a limited amount of countries and/or by looking at bank-level data (e.g. Gormley, 2005; Mian, 2006). The downside of this approach is that it is difficult to account for heterogeneity between banks and individual countries, making it harder to draw strong conclusions about the conditions under which a negative or positive relation holds. Cleassens and Van Horen (2013) tried to improve this by creating a database of over 5000 banks worldwide, which provides data on the home country of foreign banks. They show that accounting for this heterogeneity is indeed relevant, for example by showing that larger geographical distance to the home countries of foreign banks, negatively affects the amount of private credit in the economy. Private credit is seen as an important indicator of financial development (King and Levine, 1993). The aim of this research is to build on the work of Cleassens and Van Horen (2013) by analyzing the following research question:
What is the role of home-host country cultural distance in the relation between foreign bank presence and total private credit extended by banks in an economy?
2 Different types of customers can be identified when it comes to bank lending. There are hard information customers that can easily be screened and monitored using hard information such as credit scores, or information about previous borrowing activities of the customer. Customers on which this information is not available will have more trouble obtaining loans, these are soft information customers. For this type of customers, banks will have to apply relational contracting to ensure their repayment. This paper is based on the argument that the foreign banks face ‘liability of foreignness’ which makes it harder for them to screen soft information customers, making them rely more on hard information customers. As cultural distance between the home- and host country of foreign firms has been shown to exacerbate liability of foreignness, it is hypothesized that cultural distance will exacerbate a negative effect or moderate a positive effect of foreign bank presence on private credit extended by banks.
By constructing a weighted average of the cultural distance of foreign banks in a country, this paper analyzes whether these constraints related to cultural distance are indeed visible at the country level. The analyses do not provide any robust evidence for this. No strong evidence has been found for a negative relation between foreign bank presence and private credit. There is also no significant evidence that this relation is negatively affected by cultural distance. It might be that other types of distance are better able to explain the negative effect of distance in the relation between foreign banks and private credit.
This report will have the following structure. First of all, the paper will discuss previous literature and provide further motivation for the main hypothesis. After that a description of the methodology and the data used to test this hypothesis will be given, followed by a presentation of the results. After a more in-depth discussion of the results there will be a conclusion including the limitations of the research and suggestions for future research.
LITERATURE REVIEW Private credit
3 private sector on economic growth, and they find that growth is often higher in countries with more lending to the private sector (King and Levine, 1993; Levine et al., 2000). For that reason, this paper will focus on private credit. This appears to be the most important indicator of financial development.
Foreign banks and private credit
4 advantage compared to domestic banks. In general, Detriache et al. (2008) conclude that total lending may improve with foreign bank entry, but that this is far from certain.
The results above refer primarily to developing countries, because the urgency for financial development is higher here than for developed countries. Therefore it has been researched more extensively. However, Cleassens and van Horen (2013) do a cross country study and they find a negative effect for countries overall. However when they limit the sample to only developed countries they get a positive but insignificant effect, while the effect is negative and significant for developing countries. Although they do not provide an explanation for their results related to developed countries, the analysis of Cleasssen and van Horen (2013) does show that foreign bank participation is much lower in developed countries. This might be an explanation as informational disadvantages related to the foreign status can more easily be compensated by operations of domestic banks. In addition, Detriache et al. (2002) argue that developed countries are likely to be better equipped when it comes to availability of creditor information, reducing the amount of soft information customers. This would mean the foreign banks are likely to face less distance constraints. Distance constraints and informational disadvantages related to foreign status, so called ‘liability of foreignness’ will be the focal point in the next two sections.
Cultural distance
Cultural distance can be defined as the extent to which the shared norms and values in one country differ from those in another (Drogendijk and Slangen, 2006; Hofstede, 2001; Kogut and Singh, 1988). A field of international business in which cultural distance has widely been studied is entry-mode choices of multinational enterprises (henceforth MNEs). Cultural distance is expected to increase entry costs, decrease operational benefits and hinder a firm’s ability to transfer competencies (Gomez-Mejia and Palich, 1997). Under similar reasoning, many studies have suggested that large cultural distance will likely lead to a strategy with the lowest possible resource commitments (e.g. Kogut and Singh, 1988). Although there are also studies suggesting the opposite (e.g. Drogendijk and Slangen, 2006; Chen and Hu 2002), all conclusions are based on the argument that cultural distance between the home- and host country makes it difficult for an MNE to operate internationally.
5 2013). Cultural distance complicates operations of an MNE as it will become harder to deal with these conflicting pressures (Kostova and Roth, 2002). As these conflicting pressures can prevent MNEs from adapting to the local context, they often operate less efficiently than their local competitors (Xu and Shenkar, 2002). This indicates that foreign firms have to deal with a ‘liability of foreignness’ (henceforth LOF) (Nachum, 2003). Consequently, the degree of LOF can be seen as a function of home-host cultural distance (Beugelsdijk et al., 2013). For the remainder of this paper, home-host cultural distance refers to the cultural distance between the home- and the host country of a foreign bank.
Cultural distance and foreign bank lending
Before discussing the existing theory it is important to note the type of distance that were are dealing with. Mian (2006) says that cultural distance in banking activity can play out at two levels: (1) The distance between the CEO in the home country and the loan officer operating in the host country. (2) The distance between the loan officer and the borrower. In this paper the focus will be on the first definition (Arrow 1 in Figure 1).
The discussion earlier suggests that subsidiaries of foreign banks in developing countries will have limits with respect to whom they finance. Mian (2006) suggests that the origin of the foreign banks could play a significant role in this. This study finds that foreign banks shy away from lending to ‘soft information’ customers which will require relational contracting. These are generally smaller firms seeking a loan for the first time. This is argued to come primarily from distance constraints faced by foreign banks. For the Pakistani market, he finds that geographical and cultural distance provide important explanations for the differences in lending recovery and renegotiation between foreign and domestic banks. Cultural distance even seems to be more important than the local enforcement problems, which is contrary to what was argued by Genanietti and Ongena (2002) in the previous section. Stein (2002) confirms that banks with foreign subsidiaries that have a more centralized structure, depend on more discretion in their lending activities and are therefore less likely to lend to informationally opaque firms. In line with this reasoning, Mian (2006) says that an environment with a different
CEO
(Controlling Shareholder) Loan Officer Borrower
Home country 1 Host country 2 Host country
6 corporate culture, legal institutions or regulatory framework might lead to an increase in asymmetry of information and will therefore make it more difficult for the CEO to develop policies that are specifically tailored to the needs of the host country, again making the bank rely more on hard information such as credit scores. From this reasoning we can imply that the LOF of foreign banks causes them rely on hard information, forcing them to limit lending to smaller firms. Foreign banks can therefore be said to have an efficiency disadvantage compared to domestic banks.
Agarwal and Hauswald (2010) state that geographical proximity to the source of information can encourage the collection of firm-specific subjective intelligence by loan officers and therefore make the bank rely less on hard information. This shows that small distance can erode certain informational disadvantages of foreign banks. Hahn (2014) studies cross-border lending activity of Austrian banks to Eastern European countries and finds that banks from regions which are culturally closer, are more involved in cross-border lending. This is about cross-border lending, in other words Arrow 2 in Figure 1 (only in this case the loan officer resides in the home country). However, these findings do show that geographical proximity does not always have to imply cultural proximity, something that is assumed by Mian (2006). Consequently, the informational issues that arise from cultural distance lead to the following hypothesis:
Hypothesis 1: Large cultural distance between the location of the headquarters (home country) and the subsidiaries of foreign banks (host country) will have a negative effect on the relation between foreign bank presence and total private credit extended by banks in the host country.
METHODOLOGY
Methods in other literature
7 are also taken at earlier dates and smoothed out over several years to limit measurement errors. Cleassens and Van Horen (2013) use the same method, only they do not smooth out any independent variables, only the dependent variable. However, both acknowledge that there are limitations to a cross-sectional analysis, for example omission of important variables or the fact that market share of foreign banks can be endogenous to financial development of the country and hence to private credit. For this reason, Detriache et al. (2008) also do a panel analysis over a period of five years, using a fixed effects model. The possibility of using fixed effects can help control for unobserved heterogeneity between countries. In addition they do an analysis using the system GMM estimator developed by Arrellano and Bover (1995), this test uses lagged levels of the series as instruments for the pre-determined and endogenous variables in first differences and lagged differences of the dependent variable as instruments for equations in levels. Moreover, they do an instrumental variable estimation using the share of large international banks in the former colonizing country of the host country, potential market size measured by population and a measure of distance based on language, as instrumental variables.
When it comes to measuring the effect of distance, most studies use bank level data, only Cleassens and Van Horen (2013) measure it at the country level. Mian (2006) uses bank level data from Pakistan and just like Cleassens and Van Horen (2013), measures the effect of distance by using an interaction term between the foreign banks presence variable and the variable for distance.
A first look at the model
8 also be further explained below. All other variables are the control variables. Table 11 in the appendix provides a full description of all the abbreviations.
𝐿𝑛(𝑃𝐶𝐺)𝑖𝑡 = 𝛽0 + 𝛽1𝐹𝐵𝑃𝑖𝑡+ 𝛽2𝐶𝐷𝑖𝑡+ 𝛽3𝐹𝐵𝑃 × 𝐶𝐷𝑖𝑡+ 𝛽4𝐶𝐼𝑖𝑡+ 𝛽5𝐸𝑇𝑖𝑡+
𝛽6𝐺𝐷𝑃𝑝𝐶𝐺𝑖𝑡 + 𝛽7𝐿𝑛(𝐺𝐷𝑃𝑝𝐶)𝑖𝑡+ 𝛽8𝐼𝑁𝐹𝑖𝑡+ 𝛽9𝐿𝑛(𝐴𝑉𝐴)𝑖𝑡+
𝛽10𝐿𝑛(𝐺𝐸𝑋𝑃)𝑖𝑡+ 𝛽11𝐿𝑛(𝑇𝐼)𝑖𝑡+ 𝑇𝑖𝑡+ 𝜀𝑖𝑡
(1)
The sample
A total number of 77 countries has been analyzed over the time period 2004-2009. The sample includes countries of different income groups which helps to implement a more comprehensive breakdown. In the Cleassens and Van Horen (2013) database, countries are classified as OECD if they are core OECD members. Then there are ‘other high-income countries’ which are classified by the World Bank as high-income countries in 2000, but which are not OECD members. Emerging markets are all countries reported in the Standard & Poor’s Emerging Market and Frontier Market Indexes and that were not classified as high-income countries in 2000. All other countries are seen as developing countries. Table 9 in the appendix shows the full list of countries used and their classifications.
Dependent variable
The dependent variable will be domestic credit to the private sector by banks, as a percentage of GDP. This has been obtained from the World Development Indicators (henceforth WDI) of The World Bank. It is described as financial resources provided to the private sector by depository corporations (except central banks) through for example loans and purchases of non-equity securities that establish a claim for repayment (World Bank, 2015).
9 Main independent variables
Foreign bank presence
The presence of foreign banks can be measured in two ways. First of all there is the percentage of foreign banks among total banks in the country. This indictor is nice for analyzing trends in the globalization of banks and patterns of cross-border banking activity. However, to assess the influence of foreign banks on the economy of the host country it might better to also take into account their position relative to domestic banks. This brings us to the second measure which is the share of foreign bank assets as a percentage of total bank assets. This relative measurement of foreign bank presence will be used in this study. Using their own database of banks, Cleassens and Van Horen (2013) constructed their own measurement of foreign bank market shares for each host country. As this is the same database used to calculate the aggregate cultural distance for this study (discussed in the next section), it is likely subject to the same missing data and definitions. As a result it is likely to be the best source for this research. Because asset data is limited in Bankscope for some countries, the database only gives a percentage if asset data is available for more than 60% of all banks in a country (Cleassens and van Horen, 2013)1.
A limitation for using this data in a panel-analysis that is pointed out by Detriache et al. (2008) is that Bankscope data has increased its comprehensibility over time. This could cast doubt on whether all years are really comparable. However, they point out that this should not matter if the broadening coverage does not disproportionally effect foreign banks over domestic banks or vice versa. Even if this is the case then there would still be no bias if the error does not significantly differ across countries with different growth rates of private credit. They mention that it is likely not an issue for their study and they analyze an earlier time period (1996-2002). Cleassens and Van Horen (2013) argue that they did not include data for this period as asset data was simply too limited to construct a reliable dataset. This could be an argument that this possible error applies less to this analysis than it did for Detriache et al. (2008). Nevertheless it is a limitation of applying this data in a panel setting.
Another downside of using this data is that foreign bank assets include private credit, which is also included in the dependent variable of the model. This brings the chance that both
1 The public database only provides values as integers, without decimals. This is not a problem in general but
10 variables are actually measuring the same thing, which could affect the estimates. This problem is not addressed in Cleassens and Van Horen (2013) and Detriache et al. (2008), which apply the same method. However, the correlation matrix in Table 2 shows that the two variables only have a correlation of 25%, indicating that it is likely not problematic. Nevertheless, it is a weakness of this research and also of previous studies applying the same measure.
Cultural distance
Many measures of cultural distance have been applied, however the one that seems most popular is Hofstede’s (1980) measures of national culture (Drogendijk and Slangen, 2006). Hofstede (1980) analyzed survey data that was obtained between 1967 and 1973 from 117,000 IBM employees in 40 different countries. Using this analysis he came to four different dimensions of national culture: power distance, uncertainty avoidance, masculinity and individualism. Power distance refers to the extent to which people accept that power and status are distributed unequally while uncertainty avoidance refers to the extent to which people are threatened by uncertain situations. Individualism is about the degree to which society focusses on the role of the individual instead of the group (collectivism). Finally masculinity is about how much society focusses on more traditional masculine values such as competitiveness, assertiveness, ambition and high earnings as opposed to more feminine values such as nurturing and helping others (Hofstede, 1980). Although this work has been widely criticized (e.g. Schwartz, 1994) there are numerous studies which have shown that the framework of Hofstede (1980) is still one of the better measures available and that it is too early to dismiss it completely (e.g. Drogendijk and Slangen, 2006). The most important advantage of Hofstede is that it is available for a large number of countries, including both developing and developed countries. That makes the measure particularly interesting for this study. For these reasons, this research will focus on using the Hofstede measure. The amount of countries for which the measure is available has increased over time, the latest version from Hofstede (2015) will be used.
11 This brings them to the following index:
𝐶𝑢𝑙𝑡𝑢𝑟𝑎𝑙 𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑗𝐻= ∑{(𝐼𝑖𝑗− 𝐼𝑖𝐻) 2 /𝑉𝑖}/4 4 𝑖=1 (2)
where 𝐶𝑢𝑙𝑡𝑢𝑟𝑎𝑙 𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑗𝐻 is the cultural distance between a MNE’s home country j and host country H, 𝐼𝑖𝑗 is country j’s score on the ith cultural dimension of Hofstede, and 𝐼𝑖𝐻 the score of the host country on that dimension, and 𝑉𝑖 is the variance of the scores of the respective dimension. Many studies have used this measure (e.g. Brouthers and Brouthers, 2000; Harzing, 2002; Vermeulen and Barkema, 2001; Drogendijk and Slangen, 2006).
However, there is also a Euclidian index as a measure of cultural distance, which is used in some studies as well (e.g. Drogendijk and Slangen, 2006; Brouthers and Brouthers, 2001; Vermeulen and Barkema, 2001). The difference of this measure compared to the Kogut and Singh index is that it does not assume that the differences in the scores on each of the dimensions are equally important, as it computes the distances in a four-dimensional space by taking the square root of the squared differences in the scores (Drogendijk and Slangen, 2006):
𝐶𝑢𝑙𝑡𝑢𝑟𝑎𝑙 𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒𝑗𝐻= √∑{(𝐼𝑖𝑗− 𝐼𝑖𝐻) 2 /𝑉𝑖} 4 𝑖=1 (3)
As discussed in the literature review, these indexes and cultural dimensions are used primarily in studies about entry mode choices and performance of MNEs. To my knowledge it has not yet been applied to lending decisions of banks, which could be used as a point of critique for using them in this study. Although it is not possible to fully nullify this argument, it is worth noting that some of the studies on entry mode choices actually use banks as their sample (e.g. Nachum, 2013). While this does not say anything about lending behavior specifically, it does show that general behavior of banks can likely be explained using this method.
12 Weighing the average on bank size is relevant as Hutzschenreuter & Voll (2008) argue that a firm is expected to obtain more benefits, or incur more of the costs associated with cultural distance if it is more involved in a country. Therefore it is assumed that larger foreign banks will have more impact on the domestic banking industry. Using similar reasoning, De Jong and Van Houten (2014) introduce a modified version of the Kogut and Singh index, in which the cultural distance is weighted on the amount of subsidiaries that a company has in the respective country.
The new weighted average measure of cultural distance is shown in equation 4. 𝐶𝐷𝐻 is the weighted average cultural distance from the home country of all foreign banks located in the respective host country. 𝑁 is the number of foreign banks in the country, 𝑎𝑏𝐻 is the amount of assets held by foreign bank 𝑏 in the host country H and 𝐹𝐴𝐻 is the total number of foreign bank assets in the host country H. The final part of the equation is the Kogut and Singh index. It shows the cultural distance between the country in which the headquarter of the respective foreign bank is located (home country), and the location in which the subsidiary is located (host country). 𝐶𝐷𝐻 = ∑( 𝑎𝑏𝐻 𝐹𝐴𝐻 ) 𝑁 𝑏=1 ∑{(𝐼𝑖𝑗− 𝐼𝑖𝐻) 2 /𝑉𝑖}/4 4 𝑖=1 (4)
A challenge in this approach is that asset data from Bankscope is limited for some countries, especially before 2004. Therefore the time period 2004–2009 is used. To further deal with this limitation, a country or year is only included if asset information on more than 60% of foreign banks is available. If asset data is missing for a bank in a particular year while information is available for other years, the assets in the closest year are used as a weight.
13 for, as well as the dates on which banks became active or inactive. Finally, a bank was determined as foreign owned when 50% or more of the shares were held by foreigners. The country with the highest percentage of shares was labelled as the home country.
For this research, only foreign banks were used from countries for which the Hofstede (2015) measures were available. In some cases, Hofstede data was available for the host country, but not the home country, or the home country was unavailable. In this case, the same benchmark applies, a country or year is only included if the Kogut and Singh index can be calculated for at least 60% of foreign banks.
Control variables
14 likely require more private credit so it is expected to have a positive effect, while high government expenditure will likely lead to a better infrastructure and investment climate, also leading to a positive effect. Figures for government expenditure and total investment were both obtained from the World Economic Outlook Database of the IMF2.
RESULTS
Trends in the data
Before turning to the regression analysis it is useful to first look at some trends in the data, especially for the main variables. One of the first remarks that was made in this paper was that insufficient extension of private credit can be a constraint to the development of a country. Figure 2 provides evidence for this remark as the amount of private credit provided by banks is significantly higher in the high income countries. It is evident that there is a credit constraint in lower income countries and that there is a reason to further analyze what causes this.
Figure 3 shows the trend for the market share of foreign banks in the dataset. Overall, the trend of increasing financial globalization is visible especially for developing countries. For emerging markets the average trend is less positive. When looking deeper into the data for individual countries, it can be argued that this is likely caused by high differences between countries. Among emerging markets there are some countries which have very high shares (e.g. Croatia around 90%) and some with very low shares (e.g. China around 1%). This makes it harder to observe an average trend. When comparing Figure 2 and 3, it can be concluded that countries with high foreign bank shares have on average lower levels of private credit.
Finally, Figure 4 shows the average cultural distance for the different types of countries. It is hard to draw conclusions from the trend line however it is obvious that cultural distance is more likely to be a constraint for developing countries and emerging markets than for high income countries. This could show that foreign banks in developing countries indeed suffer from distance constraints as suggested by Mian (2006).
15 0 20 40 60 80 100 120 140 160 2004 2005 2006 2007 2008 2009
PRIVATE CREDIT BY BANKS (% OF GDP)
High income countries
Emerging markets + Developing countries Emerging markets Developing countries 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2004 2005 2006 2007 2008 2009
AVERAGE CULTURAL DISTANCE
High income countries
Emerging markets + Developing countries
Emerging markets
Developing countries
Figure 2: Trends in private credit by banks (% of GDP) 4
20 25 30 35 40 45 50 55 60 2004 2005 2006 2007 2008 2009
FOREIGN BANKS ASSETS AS % OF TOTAL
DOMESTIC BANK ASSETS
High income countries
Emerging markets + Developing countries
Emerging markets
Developing countries
Figure 4: Trends in average cultural distance of foreign banks 2
16 Data description
Table 1 shows the summary statistics and Table 2 shows the correlations between all the variables. Definitions of the abbreviations can be found in Table 11 in the appendix. What stands out in Table 1 is the high standard deviations. These will not necessarily cause problems, but it can be seen as an indication that outliers are present in the data. The goal of this study is to analyze an average effect and outliers could influence the regression as they might not represent average conditions. In this case outliers could for example be caused by extraordinary policies implemented by certain countries. If these are included in the regression, the results might not be representative. Presence of outliers will be studied more in-depth later on. Table 2 gives a first look at correlations between the independent variables and the dependent variable. Although correlations do not infer causality, it can be noted that the signs of correlations with private credit are as expected, except for GDP per capita growth. Another point worth noting is the negative correlation between GDP per capita growth and GDP per capita. This might seem odd at first, but it could be explained by saying that countries with lower GDP experience more growth, which makes sense. When looking at the main independent variables of cultural distance and foreign bank presence, there is no real evidence for high correlations with the other variables. However, Table 2 does show evidence for collinearity between some of the control variables as well between the variable for foreign bank presence and the interaction term. Possible presence of collinearity will be further explored below.
Table 1: Descriptive statistics
Variable Obs Mean Std. Dev. Min Max
17 Table 2: Correlation matrix
As mentioned above, the descriptive statistics give a serious reason to look for outliers in the data, in particular for the dependent variable. This can be done by plotting the dependent variable against the main independent variables of CD and FBP, as well as the interaction term between the two. A negative overall trend can be observed in all three cases (Figure 5, Panel A, B and C) but values above 150 seem to be deviate a lot from the trend in at least two of the three cases. What stands out is that all of these countries are classified in the dataset as high income countries. This indicates that it might be relevant to do additional analyses testing developed and developing countries separately. To truly know whether these are outliers, it might also be relevant to plot the dependent variable against the GDP per capita, as shown in Panel D. There are big outliers on the right side of the graph. These are countries with high levels of GDP per capita, but relatively low levels of private credit. This confirms the earlier finding that it is relevant to further explore the existence of outliers and influential points.
One way to deal with outliers is transforming the dependent variable by taking its natural log. As described below this will be done anyways in order to bring the distribution of the combined error terms closer to normal. In addition, Cook’s distance test will be performed to see if any observations should be deleted. An observation can be said to be influential if its removal significantly changes the estimates. Influence does not only depend on whether the point is an outlier, but also on its leverage. The distance measure by Cook (1977) does not only look at whether a point is an outlier, but also at the influence that each observation has on
18 the regression estimates. Performing the test on the main regression using the full sample leads to the deletion of 23 observations3.
In addition to a linear curve showing the fitted values, the scatter plots also show Lowess curves which provide a smoother analysis of the fitted values point-by-point. This can help visualize possible nonlinearities in the relation. For Panel A, B and C there seems to be some evidence for a non-linear relationship. For these variables, a log-linear, log-log or linear-log model might fit the data better.
3 Threshold of Cook’s D > 4/N
Panel A Panel B
Panel C Panel D
19 Empirical results
In line with Detriache et al. (2008), this study will use panel data. With panel data there are three types of regressions that can be considered: a pooled model, a fixed effects model and a random effects model. In a pooled model, all country data is pooled together and therefore it does not account for heterogeneity between countries. Although the model in this study controls for some differences between countries, it is unlikely that all differences are captured, and therefore many important differences are likely to be missed. The fixed effects model does account for this heterogeneity by giving a different constant to all individuals, while the slope coefficients remain the same. The random effects model is similar, although this also takes into account that all individuals are randomly chosen. The intercept of the random effects model has a fixed part which contains the population average, and a non-fixed part that captures random individual differences from the population average (Hill et al., 2012).
To determine the model appropriate for this study, several tests have been performed. First of all, a Lagrange multiplier (LM) test was performed to look for presence of random effects (and hence individual heterogeneity). This test was significant so it can be concluded that individual differences are present4. After that, a Hausman test was performed to test the suitability of the random effects model over that of the fixed effects model. This test can check whether there is no correlation between the random effects and the regressors, making the random effects model appropriate for estimation. For the overall sample, the Hausman test points towards a fixed effects model.5 The random- and fixed effects model depend on a couple of assumptions which need to be met. First of all the combined errors need to have zero mean and a constant variance. In other words, they need to be normally distributed. This is checked using a kernel density plot as shown in Figure 6. Panel A shows the density plot for the original linear fixed effects model. It is evident that the errors are not normally distributed. The most common way to solve this is by transformation of the dependent- or independent variables. The transformed model takes the log of the dependent variable (private credit) and of all control variables that do not have any negative values in their untransformed form, with the exception of creditor information and enforcement time. This leads to the plot shown in Panel B. The problem is still not completely solved as several normality test still lead to the rejection of the null-hypothesis that the distribution is normal6. This does not mean that the coefficients will no longer be the best linear unbiased estimators. Still, it does imply that the confidence intervals
4 Chibar2(01) = 643.22, p-value = 0.0000 5 Chi2(16)=41.80 p-value=0.0004
20 and significance tests might be less reliable. A solution could be increasing the amount of observations (Hill et al., 2012). Unfortunately this is not possible in the current situation.
Transformation of variables is not without consequences. Transforming the dependent variable to its natural log means that the interpretation of all independent variables changes. Overall the consequences might not be too big in this case as the scatter plot already shows some evidence for non-linearities in the data. A log-linear model with a negative sign has a shape similar to the smoothed line in Figure 2 (Panels A, B and C).
In addition, one has to consider the presence of autocorrelation. If autocorrelation is present, it means that present values of a particular variable are influenced by its previous values, which could lead to a bias in the estimates. To detect possible autocorrelation, the Wooldridge (2002) test is used. As shown by Drukker (2003), this test has good size and power properties in reasonably sized samples. The test provides evidence for autocorrelation and this will be controlled for by using panel-robust standard errors7. However, robust standard errors only control for correlation over time within individuals (countries) but not between different individuals. Again, this can result in inaccurate standard errors, possibly leading to false rejections of the null-hypothesis. As the sample consists of countries from different continents and income groups, these countries are likely to be exposed to similar shocks related to certain points in time. An example could be that they are affected by the same financial crises in particular years. Roodman (2006) says that to remove such universal time-related shocks from the error term, it is always relevant to include time dummies. The time period analyzed in this
7 F = 119.433, p-value = 0.0000
Figure 6: Kernel Density plots of combined error terms for the transformed model
21 paper includes years in which many countries across the world were suffering from a banking crisis, making the time dummies of extra importance here.
Finally, it is necessary to check for multicollinearity. The goal of this study is to find the isolated effects of different variables. If variables are highly correlated, then this might lead to false inferences about the estimates as a result of inaccurate standard errors. The correlation matrix above already gave an indication that correlations between pairs of variables seemed to be a problem for some control variables and for the interaction term. However, to analyze the issue more in-depth a variance inflation matrix can provide more information. Table 3 shows the VIF matrix for the regression using the full sample. The threshold for the VIF scores is very arbitrary. For example Hair et al. (1998) impose a threshold of 10, while Heiberger and Holland (2004) argue that 5 is more appropriate. The values that stand out immediately are those for GDP per capita and agriculture value-added, all other control variables are well below 5. Solutions to the problem include increasing the amount of observations or deleting one of the correlated variables. As noted before, the first option is not possible. Deleting the variables is a possibility however including the correlated variables in the regression did not seem to affect the main independent variables, and resulted in a higher explanatory power of the model. For this reason they have not been deleted. As a consequence, any significant or insignificant effects for these variables should be interpreted with care. In the end the goal is to make inferences about the main independent variables, making the accuracy of the control variables of less importance.
22 Table 3: VIF table
Variable VIF 1/VIF Variable VIF 1/VIF
FBP 4.75 0.210535 Ln(GDPpC) 5.82 0.171946 CD 2.56 0.390777 INF 1.47 0.68003 FBP x CD 6.38 0.156782 Ln(AVA) 4.15 0.241179 CI 1.33 0.750386 Ln(GEXP) 1.74 0.575749 ET 1.58 0.633329 Ln(TI) 1.37 0.72894 GDPpCG 2.09 0.478538
Table 4 shows the regression results for the fixed effects model, time dummies are included but not reported. All control variables have their expected signs, except for GDP per capita growth. It was predicted that high growth is often stimulated by an increase in private credit. However the output shows a negative sign. An explanation could be that growth is higher in countries at lower levels of development and that in these countries, less private credit is extended relative to their more developed counterparts. When considering the main independent variables, an overall negative but insignificant effect is found for the asset share of foreign banks. The magnitude of the effect changes as more control variables are added, but because the effect is insignificant it is not possible to draw conclusions from it. The estimate for cultural distance is negative but not significant, however the coefficient of cultural distance on itself is not relevant as its effect depends on the relative position of foreign banks in the market. The interaction variable is therefore more important and this one has a positive sign, which is the opposite of what was predicted. To interpret the interaction term more specifically, and to truly test Hypotheses 1, the marginal effect of foreign bank presence conditional on cultural distance will have to be analyzed, as suggested by Brambor et al. (2005). This marginal effect is calculated as follows:
𝛿𝐿𝑛(𝑃𝐶𝐺)
𝛿𝐹𝐵𝑃 = 𝛽1+ 𝛽3𝐶𝐷 (7)
Using this formula, the marginal effect can be calculated at different values of cultural distance. The expectation is that the marginal effect decreases and becomes more significant as cultural distance increases.
23 formulated in Hypothesis 1 would be that the marginal effect decreases as the cultural distance increases. The marginal effect is negative (but insignificant) when cultural distance ranges from 0 to 3, but after that it turns positive and is still insignificant. As not all positive and negative point estimates are significant, it cannot be concluded with certainty that the positive slope of the curve is actually significant. When interpreting Table 7, it is important to take into account the descriptive statistics in Table 1. The mean value of CD is around 1.70 at which there is no significant negative effect of FBP. The only values that are close to significance (at the 10% level) are those in the range of 0 to 1, which is only represented by around 20% of the sample. As there are no significant results at any value of CD, there is no real evidence that the marginal effect of FBP is negative for certain values of cultural distance.
Table 4: Fixed effects estimation result for full sample
(1) (2) (3) (4) VARIABLES Ln(PCG) FBP -0.00151 -0.00110 -0.000902 -0.00274 (0.00178) (0.00150) (0.00118) (0.00167) CD -0.0418 (0.0375) FBP x CD 0.000874 (0.000765) ET -0.000106 -0.000297 -0.000296 -0.000256 (0.000411) (0.000195) (0.000189) (0.000195) CI 0.0166 0.00873 0.00813 0.00832 (0.0134) (0.0108) (0.0101) (0.0100) GDPpCG -0.00905*** -0.00954** -0.00916** (0.00318) (0.00396) (0.00407) Ln(GDPpC) 0.603** 0.504* 0.525* (0.300) (0.270) (0.264) INF -0.0151*** -0.0157*** -0.0155*** (0.00230) (0.00214) (0.00211) Ln(AVA) -0.227** -0.228** (0.111) (0.113) Ln(GEXP) 0.0586 0.0782 (0.166) (0.162) Ln(TI) 0.211* 0.209 (0.125) (0.130) Constant 4.174*** -1.451 -1.028 -1.236 (0.269) (2.908) (2.634) (2.518) Observations 373 373 373 373 R-squared 0.499 0.671 0.642 0.675 Adj. R-squared 0.488 0.659 0.631 0.660 Number of countries 74 74 74 74
Note: Time-dummies included but not reported
24 The literature analyzed above has shown a lot of evidence for differential effects between developed and less developed countries. Since the sample used in this study is sufficient for separation into these different groups, further analyses are possible. Table 5 shows the results for the model including only less developed countries. Less developed countries are those listed in the appendix as emerging markets or developing countries. The type of model has been determined using the same steps as in the previous analysis8.
The sign for foreign bank presence is negative in all regressions but the effect is not significant in the first three columns. We cannot say with certainty that any overall negative effect of foreign banks on private credit is driven by less developed countries. Compared to the regressions for the overall sample, the variable for agriculture value added is no longer significant. What is again noteworthy is the positive sign for the interaction variable, which is now even significant at the 10% level. As both the coefficients for CD and FBP are negative and significant, this would imply that CD positively affects the relation between FBP and private credit, which is contrary to what is stated in Hypothesis 1. Figure 8 shows that the slope of the marginal effect line is even steeper than shown in Figure 7 for the complete sample. Again, it cannot be concluded that this slope is actually significantly different from zero. The
8 LM-test for country heterogeneity: chibar(01): 323.45, p-value:0.0000; Hausman-test for comparing random-
and fixed effects model: Chi2 = 35.77, p-value = 0.0019 ; Wooldridge (2002) autocorrelation test: F = 104.216, p-value = 0.0000. Multicollinearity has increased due to reduction of the sample size, no corrections are made for the same reasons stated earlier. Shapiro–Wilk test rejects normality of combined errors at 1% level, Jarque-Bera test does not reject null hypothesis of normality. Cook’s distance test led to deletion of 11 observations. Figure 7: Marginal effects of foreign bank presence on private credit,
25 mean value of CD for less developed countries is 1.87 at which there is no significant, negative effect. The values of CD that are closest to significance are again those in the range of 0 to 1, which represent only 20% of the less developed countries in the data. Consequently, there is no strong evidence that cultural distance can explain a negative effect of foreign bank presence on total private credit by banks in less developed countries.
Table 5: Fixed effects estimation results for less developed countries
(1) (2) (3) (4) VARIABLES Ln(PCG) FBP -0.00106 -0.000628 -0.000314 -0.00377* (0.00223) (0.00189) (0.00165) (0.00224) CD -0.0879* (0.0448) FBP x CD 0.00170* (0.000912)
ET 0.000168 3.42e-05 1.68e-05 7.13e-05
(0.000538) (0.000287) (0.000279) (0.000282) CI 0.0126 0.00550 0.00578 0.00601 (0.0153) (0.0119) (0.0120) (0.0117) GDPpCG -0.0121** -0.0137** -0.0127** (0.00513) (0.00611) (0.00622) Ln(GDPpC) 1.084*** 0.798** 0.846** (0.373) (0.371) (0.323) INF -0.00669* -0.00749** -0.00715** (0.00373) (0.00360) (0.00349) Ln(AVA) -0.166 -0.159 (0.191) (0.192) Ln(GEXP) 0.0230 0.0563 (0.216) (0.200) Ln(TI) 0.262*** 0.244** (0.0969) (0.0998) Constant 3.616*** -6.131* -4.082 -4.460 (0.373) (3.489) (3.520) (3.011) Observations 255 255 255 255 R-squared 0.508 0.615 0.640 0.653 Adj. R-squared 0.492 0.598 0.619 0.629 Number of countries 50 50 50 50
Note: Time-dummies included but not reported
26 Finally, Table 6 shows the result for developed countries9. The Hausman test indicated that a random effects model was more appropriate for this estimation, which is noteworthy as most previous literature uses fixed effects (e.g. Detriache et al., 2008)10. An extra regression has been added as a robustness check because the VIF values showed some evidence for correlation between the interaction term and two of the control variables. Regression 5 excludes those control variables and it seems that the interaction term is not effected too much. The effect for the market share of foreign banks is now positive, but not significant. It can be noted that the variable for enforcement time is now significant, indicating the negative effect in the overall result in Table 4 is most likely driven by developed countries, where enforcement time seems to matter more. Remarkable once again is the result for the interaction effect, which is now negative. When studying the marginal effects in Figure 9, there is a downward sloping marginal effect which is in line with what was predicted in Hypothesis 1 (again, it cannot be said whether this slope is significant). The mean value of CD for developed countries is 1.38 at which no significant effect is observed. In addition, the slope is not very steep and none of the effects are actually statistically significant.
9 LM-test for country heterogeneity: chibar(01): 186.52, p-value:0.0000; Hausman-test for comparing random-
and fixed effects model: Chi2 = 10.95, p-value = 0.6899; Wooldridge (2002) autocorrelation test: F = 473.321, p-value = 0.0000; Multicollinearity has increased due to reduction of the sample size, there is now also
correlation between some control variables and the interaction effect. To control for this, an additional regression is done excluding these controls. Both the Shapiro–Wilk test and Jarque-Bera test reject the null hypothesis of normality at the 1% level.
10 It is worth noting that the results and conclusions for the main independent variables are not much different
when a fixed effects model is used.
27 Table 6: Random effects estimation results for developed countries
(1) (2) (3) (4) (4) VARIABLES Ln(PCG) FBP 0.000815 0.00154 0.000571 0.00234 0.00359 (0.000896) (0.00104) (0.000953) (0.00256) (0.00238) CD 0.0651 0.0859* (0.0461) (0.0467) FBP x CD -0.000695 -0.000664 (0.00274) (0.00224) CI 0.0310 0.0130 0.0192 0.0138 0.00500 (0.0522) (0.0485) (0.0444) (0.0450) (0.0499) ET -0.000433** -0.000522* -0.000470* -0.000577* -0.000657* (0.000193) (0.000293) (0.000263) (0.000314) (0.000373) GDPpCG -0.0123*** -0.00844** -0.00777* -0.00907** (0.00407) (0.00399) (0.00426) (0.00409) Ln(GDPpC) -0.386 -0.335 -0.288 -0.455* (0.287) (0.222) (0.263) (0.271) INF -0.0114** -0.00729 -0.00621 -0.00848 (0.00511) (0.00536) (0.00484) (0.00548) Ln(AVA) -0.152*** -0.140*** (0.0517) (0.0494) Ln(GEXP) 0.218 (0.219) Ln(TI) -0.0458 -0.0675 -0.138 (0.227) (0.209) (0.201) Constant 4.946*** 9.109*** 8.558*** 7.430** 10.23*** (0.281) (3.069) (2.627) (3.662) (2.896) Observations 122 122 122 122 122 Number countries 24 24 24 24 24 Overall R-squared 0.313 0.266 0.311 0.292 0.199
Note: Time-dummies included but not reported
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
28
Robustness checks
The analysis above does not provide any strong evidence for the negative effect of average home-host cultural distance that was predicted in Hypothesis 1. One possible explanation for this could be the accuracy of the Hofstede dimensions as a measure of cultural distance. As mentioned before, Hofstede’s work has been criticized extensively for being outdated and for being composed from survey data that only looked at one company (Sondergard, 1994; Engelen et al., 2008). Therefore, in line with Beugelsdijk et al. (2013) the values of the Global Leadership and Organizational Behavior Effectiveness (GLOBE) study are included as a check for robustness. This study includes the dimensions of Hofstede and in addition uses dimensions such as assertiveness, future orientation, gender egalitarianism, humane orientation and performance orientation. The study can be said to be more extensive and more recent, involving 17,300 managers from 62 nations between 1993 and 2004 (House et al., 2004). A downside of this data is that it is available for less countries than the Hofstede measure. This means that the sample size is lower, but also that the aggregate home-host cultural distance sometimes has to be calculated using a smaller number of foreign banks, making the overall estimate less accurate than when calculated using the Hofstede measure. In addition, the method of Drogendijk and Slangen (2006) is used by calculating cultural distance using the Euclidian index in Equation 3, instead of the Kogut and Singh index. To sum up, an aggregate cultural distance of foreign banks in the country is composed using the Kogut and Singh index for the GLOBE values. Moreover, an analysis will be done using the Euclidian index instead of the Kogut & Singh index, using both Hofstede and GLOBE values.
Table 10 in the appendix shows the correlations between the different measures and shows that it is indeed relevant to check the robustness of the findings as the correlations between the GLOBE and the Hofstede measures are not very high. Table 12 in the appendix shows the results for the main independent variables and the marginal effect plots for the respective measures of cultural distance. Again a Hausman test was performed for each regression do decide between a fixed- and random effects model.
29 the Euclidian index, there are also positive effects that are significant at the 10% level. For this measure, the mean value is 3.48 which is just outside this range of 10% significance. Furthermore, those observations that are within the 10% significance range again only represent around 30% of the sample, indicating that no strong inferences can be made from these results. Consequently, the insignificant effects seem robust across the different measures. Secondly, it can be argued that the model is suffering from simultaneity bias as a result of reversed causality. This means that the dependent variable and one or more independent variables are simultaneously determined. The main suspect in this model with regard to simultaneity bias is foreign bank presence. There could be exogenous reasons for why banks enter a country with low financial development, for example because of stronger growth prospects. Alternatively, business prospects may be low in countries with low financial development, making banks more prone to enter more financially developed countries (Detriache et al., 2008). In this case, endogeniety of foreign bank presence will likely mean that the interaction term of the variable with the average home-host cultural distance is endogenous as well. When performing a Davidson-Mackinnon test, the null-hypothesis that foreign bank presence is exogenous is not rejected11. A Hausman test provided more evidence for endogeneity, almost rejecting the null hypothesis of exogeneity at the 10% level12. Nevertheless, because there is a theoretical justification, it can never hurt to check whether the estimate of foreign bank presence changes when using a different estimation model. This is done by using a fixed effects instrumental variable regression model. Coming up with instruments for macro-data is difficult. Detriache et al. (2008) do a system GMM estimation. Although this is not the same as a fixed effect instrumental variable regression, it does show the relevance of using a lagged value of the endogenous variable as an instrument. This lagged value is likely correlated with its present value and only indirectly affects private credit in the current year, through its effect on private credit in the preceding year. Therefore a one-year lag will be used. In addition, population size is used as an instrument. Detriache et al. (2008) argue that this is a valid instrument as global banks operating in many markets can better diversify country specific risk than their domestic counterparts, this is particularly valuable in smaller countries. In addition, smaller market size means less investment is needed to obtain a large
11 F=.4219, p-value = 0.5166.
12 The Hausman test consists of a first-stage least squares regression where the potentially endogenous variable
FBP is regressed on all exogenous independent variables and the instruments. The residuals are then saved and entered into the regular pooled OLS-regression. A t-test on the coefficient of the residuals shows the
30 share of the market. To test the strength of the instruments a Cragg-Donald Wald F-test was used and the validity of the instruments has been tested using a Sargan-test13. The Sargan test led to rejection of the null-hypothesis that both instruments are valid, at the 10% level. This casts doubt on the validity of the instruments and could indicate that the instrumental variable estimator is inconsistent. Nevertheless, theoretically these are the best instruments available at this point. As it is complex to instrument for the interaction term, the regression is only done without controlling for cultural distance. If the estimate for foreign bank presence does not seem to be affected much compared to the initial regression, it is simply assumed that the interaction term is likely not affected much either. Especially when considering their 80% correlation.
Table 7: Fixed effects instrumental variable estimation results for overall sample
(1) (2) (3) VARIABLES Ln(PCG) FBP -0.00105 -0.00147 -0.00125 (0.00238) (0.00205) (0.00195) ET 7.73e-05 -0.000131 -7.90e-05 (0.000236) (0.000206) (0.000196) CI 0.00971 0.00400 0.00228 (0.00775) (0.00697) (0.00656) GDPpCG -0.00547** -0.00607** (0.00248) (0.00257) Ln(GDPpC) 0.489*** 0.436*** (0.159) (0.156) INF -0.0146*** -0.0160*** (0.00195) (0.00188) Ln(AVA) -0.294*** (0.0742) Ln(GEXP) 0.0344 (0.101) Ln(TI) 0.218*** (0.0622) Constant 4.057*** -0.421 -0.314 (0.172) (1.524) (1.529) Observations 305 305 305 Number countries 74 74 74 Overall R-squared 0.0514 0.639 0.625
Note: Time-dummies included but not reported Robust standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
13 Cragg-Donald Wald F-test: F = 32.483, which is above the Stock-Yogo (2005) critical value for 10%
31 The results of the fixed effects instrumental variable estimation results are shown in Table 7. The estimation can only be done for the overall sample because the instruments become even less valid when using only developed- or less developed countries. This is likely due to the reduction in sample size. When looking at the variable for foreign bank presence it can be seen that it is slightly higher in magnitude but still not significant. The estimates would have been slightly understated in the initial fixed effects estimation. However, if the model suffers from simultaneity bias, the impact seems to be limited. For that reason it is assumed that the interaction term will likely not be affected much either. It is concluded that the unexpected results can probably not be explained by possible endogeneity.
An additional explanation for the unexpected results could be the calculation of the aggregate cultural distance, which is an asset weighted average of the home-host country cultural distance for all foreign banks operating in the country. However, as already mentioned in the literature review, cultural distance between the home- and host country of a firm can lead to lower commitment strategies or can affect performance of firms relative to their domestic counterparts. In other words, the assets used to calculate the aggregate home-host country cultural distance of foreign banks in a country, could actually be determined by the cultural distance itself. Foreign banks with large cultural distance between the home- and the host country would then automatically have lower asset shares, leading to a convergence of aggregate foreign bank cultural distance values between countries. Some evidence can be provided that this is not likely to be the case. First of all, Figure 4 already showed that there are differences in aggregate cultural distance between countries from different income groups, providing an indication that values differ per country and that no convergence has taken place. In addition, the descriptive statistics in Table 1 show that the minimum and maximum value of the cultural distance data differ significantly, and the standard deviation is quite high as well. Finally, one can look if there is a high negative correlation between home-host cultural distance of a foreign bank and its asset share among total foreign banks. This can be checked by looking at the bank level data that was used to construct the aggregate home-host cultural distance of foreign banks in a country. Figure 11 in the appendix shows this relation in a scatter plot and although there are some outliers on the left side of the graph, the overall correlation is negative but very low (-0.03). This is another indication that the estimates are unlikely to be affected by the theory that large home-host cultural distance can lead to lower asset shares.
32 be tested by running the regression again using a lagged value of foreign bank presence and cultural distance. A one year lag is used simply because the short time period of the data does not allow for further lags, as this would reduce the sample size too much. Table 13 in the appendix shows the results for the overall sample. It is evident that there are no significant changes in the estimates and their significance. When looking at Figure 12 in the appendix, no significant differences are found either.
The instrumental variable estimation also used a lagged value of the independent variable however there is a difference between the IV- estimation and the method used in Table 13. In the IV-estimation, there is the assumption that foreign bank presence in a particular year (𝐹𝐵𝑃𝑡), will have an effect on financial development in that same year (𝑃𝐶𝐺𝑡). The IV-estimation tested the robustness of this finding by using an instrument that is correlated with 𝐹𝐵𝑃𝑡 but only affects 𝑃𝐶𝐺𝑡 indirectly. A lagged value (𝐹𝐵𝑃𝑡−1) is not assumed to have a direct effect on 𝑃𝐶𝐺𝑡, but it is likely to be correlated with 𝐹𝐵𝑃𝑡 and indirectly affects 𝑃𝐶𝐺𝑡, through 𝑃𝐶𝐺𝑡−1. The regression in Table 13 is different in the sense that it is assumed that the direct relation does not run from 𝐹𝐵𝑃𝑡 to 𝑃𝐶𝐺𝑡, but from 𝐹𝐵𝑃𝑡−1 to 𝑃𝐶𝐺𝑡.
A final remark that can be made about the analysis is that the sample is split up into different groups, leading to a smaller amount of observations. This could provide an explanation for why the negative sign of foreign bank presence is not significant for less developed countries. An additional method that can be used to pool the results for different groups is the one used by Blonigen and Wang (2004). They use a dummy variable for less developed countries (LDC=1) and interact this with every independent variable to allow all coefficients to differ per group. Leading to the model shown in Equation 6. It does not show the control variables, but for these the exact same method applies. What the equation shows is that we now have a triple interaction effect, making it possible to estimate the marginal effect of foreign bank presence, conditional on cultural distance, for two different income groups.
𝐿𝑛(𝑃𝐶𝐺)𝑖𝑡 = 𝛽0 + 𝛽1𝐹𝐵𝑃𝑖𝑡+ 𝛽2(𝐹𝐵𝑃 ∗ 𝐿𝐷𝐶) 𝑖𝑡+ 𝛽3𝐶𝐷𝑖𝑡+ 𝛽4(𝐶𝐷 ∗ 𝐿𝐷𝐶)𝑖𝑡+
𝛽5(𝐹𝐵𝑃 ∗ 𝐶𝐷)𝑖𝑡+ 𝛽6(𝐹𝐵𝑃 ∗ 𝐶𝐷𝑖𝑡∗ 𝐿𝐷𝐶) + 𝑇𝑖𝑡+ 𝜀𝑖𝑡 + …..
(6)
The marginal effect of FBP is shown in Equation 7. The marginal effects of the different groups can now differ as 𝛽2 and 𝛽6 will be 0 when LDC=0
𝛿𝐿𝑛(𝑃𝐶𝐺)
33 A downside of this method is that the dummy variable is time-invariant, meaning that the model cannot be estimated using a fixed effects model. Therefore a random effects model is used. The overall results are reported in Table 14 in the appendix. Table 15 in the appendix shows the average marginal effect of foreign bank presence for each group. The effect is again negative for less developed countries and positive for developed countries. However, none of the effects are significant making it difficult to conclude that a negative effect is more likely to be present in less developed countries. Nevertheless, some inferences could still be made from the marginal effect of the dummy variable. Figure 13 shows the marginal effect of the dummy variable LDC for different values of foreign bank presence. Because we are dealing with a dummy variable, the marginal effect should now be interpreted differently than in the previous parts. The graph shows that the difference from the base group (developed countries) moves more in the negative direction as foreign bank presence increases. However, the confidence intervals are very wide meaning that one needs to be careful when it comes to interpreting the slope. The result could confirm the earlier guess that the negative effect of foreign banks on financial development is likely driven by less developed countries, but again the evidence is not very strong. Figure 14 shows the marginal effects of foreign bank presence conditional on cultural distance for both developed- and less developed countries. It can be concluded that both curves are similar compared to the marginal effect curves in Figures 8 and 9 and none of the results are significant. This bring us to the conclusion that the initial estimation results are robust across different methods of controlling for groups, in the sense that they do not lead to more support for Hypothesis 1.
Threshold effects
34 information such as credit scores, and soft information customers which can be monitored less easily and therefore rely more on relational contracting. If the amount of foreign and domestic banks is in equilibrium, the foreign banks can take all the hard information customers, for which their LOF forms less of a constraint. This will then free up capital at domestic banks to serve more soft information customers, making the overall effect of foreign banks beneficial for the country. If this reasoning is extended to this particular research, it can be argued that cultural distance does not matter when this equilibrium is present, as foreign banks will then only deal with hard information customers. Cultural distance will only be relevant once a certain threshold of foreign bank presence is reached where domestic banks can no longer absorb all soft information customers. At that point foreign banks will have to start serving these customers and the required relational contracting will be easier for foreign banks that have a small cultural distance between their country of origin and the respective host country.
In order to find some evidence for a possible threshold effect, an additional regression is done with a dummy variable that takes the value of 1 if a country has a foreign bank asset share of more than 40%14. The regression is the same as the Blonigen and Wang (2004) method used in the previous section. The threshold effect is only analyzed for the overall sample and less developed countries, as the foreign bank asset shares for developed countries are less dispersed with approximately 80% of the sample having a foreign bank asset share below 30%, making it very hard to get a reliable estimate from comparing two groups. As the negative effect of foreign bank presence seems to be driven by less developed countries, it should not be a problem that developed countries are not analyzed.
Table 16 in the appendix shows the regression results for the overall sample and for the less developed countries. However, the most important results are displayed in Table 8 and Figure 10 below. Table 8 shows the average marginal effects of foreign bank presence for the countries below and above the threshold of 40%15. The negative and significant effect for countries below the threshold shows that the negative effect of foreign banks on private credit is actually driven by the countries where foreign banks have a lower market share. This confirms the finding of Cleassens and Van Horen (2013) who argue that in countries where their market share is small, foreign banks tend to be niche players, targeting only specific customers. Therefore they contribute less to financial development.
14 This was randomly chosen
35 When looking at the results for the overall sample, it can be said that, on average, a 1 percentage point increase in foreign bank presence, leads to a reduction in private credit by banks of around 0.6% (1% for less developed countries). It important to emphasize that this is about the marginal effect of foreign bank presence and that this does not mean that countries below the threshold have on average less private credit that those above the threshold. It means that on average an incremental increase in foreign banks relative to domestic banks will have a more negative effect in countries below the threshold. Consequently, it does not change the argument of Mian (2006) and Detriache et al. (2008) that in countries where foreign banks have higher market share, these foreign banks have to serve more soft information customers, making the negative role of cultural distance of higher magnitude. It could be that the insignificant effects turn significant when cultural distance is accounted for. Figure 10 tries to provide some evidence for this as it compares the marginal effect of foreign bank presence for both groups, conditional on average home-host cultural distance.
Table 8: Average marginal effect of FBP for countries with FBP<=40 and FBP>40 Overall Sample
dy/dx Std. Err. z P>z [95% Conf. Interval]
FBP<=40 -0.00644 0.003518 -1.83 0.067 -0.01333 0.000459
FBP>40 0.000651 0.001998 0.33 0.744 -0.00326 0.004566
Less Developed Countries
dy/dx Std. Err. z P>z [95% Conf. Interval]
FBP<=40 -0.01067 0.004094 -2.61 0.009 -0.0187 -0.00265
FBP>40 -0.0017 0.001723 -0.99 0.323 -0.00508 0.001674
Figure 10: Marginal effects of FBP conditional on cultural distance
