Securitization, Financial Development and Economic Growth 1

Transcription

1 Securitization, Financial Development and Economic Growth 1 This Draft: December 2012 Abstract: We analyze the impact of securitization technology on long-run growth performances and the economic growth experience during the financial crisis in a crosscountry perspective. We modify the main model linking financial development and economic growth including the securitization technology as a dummy variable. The dynamic panel approach suggests that there is some evidence for a marginal positive impact of securitization on economic growth in the long run, but it is not consistent throughout the methodologies. This weak positive effect is not through increasing the loan supply lending channel- but rather other aspects of securitization such as ameliorations in risk sharing. The positive impact of securitization is also observed during the initial period of the recent financial crisis: the cross-sectional evidence suggests that in 2007 countries with securitization technology perform better than other countries, but once whole crisis period included the relationship loses its significance. JEL Classification: O16; O40; G01; G21 Keywords: Financial Development; Securitization; Financial Crisis; Economic Growth 1 We thank Wolf Wagner and Burak Uras for valuable comments and suggestions. 1

2 I. INTRODUCTION Over-the-counter (OTC) derivatives have come to play an exceptionally important role in our financial system and in our economy. These instruments allow users to unbundle risks and allocate them to the investors most willing and able to assume them. A growing number of financial and nonfinancial institutions have embraced derivatives as an integral part of their risk capital allocation and profit maximization. In particular, the profitability of derivative products has been a major factor in the significant gain in the finance industry's share of American corporate output during the past decade--a reflection of their value to nonfinancial industry. Indeed, this value added from derivatives itself derives from their ability to enhance the process of wealth creation throughout our economy. Alan Greenspan 2, 2000 I have found very little evidence that vast amounts of innovation in financial markets in recent years have had a visible effect on the productivity of the economy. it was quite good in the 1980s without credit-default swaps and without securitization and without CDOs. I do not know if something happened that suddenly made these innovations essential for growth. In fact, we had greater speed of growth and particularly did not put the whole economy at risk of collapse. That is the main concern that I think we all need to have. Paul Volcker 3, 2009 The last financial crisis hit the global economy very badly and it put the macrofinancial linkages on the top of the world leaders agenda. Many economists agree that this crisis is one of the worst economic crises in the history; the former governor of Federal Reserve Greenspan even claims that the crisis is worse than the Great Depression of 1930s. The crisis started with the burst of the housing bubble in US, which is mainly created by financial innovation, namely securitization of subprime mortgages via complex financial instruments (Duca et al., 2010). Many banks had to write down huge amounts of assets from their loan books, which in turn led to a credit crunch and some bankruptcies or bailouts -such as Lehman Brothers and Northern Rock. The 2 Greenspan, A., Over-the-counter derivatives, testimony of the US Federal Reserve Bank Chairman, before the Committee on Agriculture, Nutrition and Forestry, United States Senate, February Volcker, P., Discussion at the Wall Street Journal s Future of Finance Initiative, December. 2

3 securitization technology is one of the most important financial innovations in the last decades. It enabled financial institutions to pool and repackage illiquid loans or similar financial instruments into liquid securities and then to sell these securities to different kinds of investors, such as other banks, insurance companies, pension funds, and mutual funds. Its impact on macroeconomy, however, is a controversial issue, which can be summarized in two completely opposite views of former Federal Reserve governors given at the beginning. Securitization had a big impact on the financial sector long before the recent financial crisis. Altunbas et al. (2009) even argue that the role of banks changed dramatically from originate and hold to originate, repackage and sell. Having by far the highest amount of securitization, the United States is the leading country in the securitization trend, in which only mortgage-related securitizations are included, demonstrate American dominance in the securities industry. Nevertheless, many other countries both developed and emerging economies- joined the securitization trend and there has been a dramatic increase in new entrants to the market, together with a greater variety of assets being securitized (Estrella, 2002; Lejot et al. 2008). Moreover, this securitization trend became global not only through origination of various securities, but also through foreign investment or in other words cross-border trading of these instruments (Figure 1). Foreign activity in US-originated securities increased dramatically especially after 2000-, which can provide a possible explanation to contagion of recent financial turmoil. A crucial point is that, a financial institution does not have to have the securitization technology to participate in securities markets. For example, a Bolivian bank or pension fund can buy securities in American or European security markets and share the default risk and the involved income of this security. Before the crisis securitization has seen by many economists as a bless -rather than a curse leading to the global turmoil- as it creates extra liquidity to financial markets, which in turn stabilizes the credit supply (Loutskina and Strahan, 2009), enables new profit channels for financial institutions and enhances the global allocation of risks (Jobst, 2005). There are many micro level studies about the securitization patterns of financial institutions. The transformation of the financial system through securitization technology is not a simple phenomenon in terms of its relation to the real economy. Securitization 3

4 had an impact on macro-financial linkages via conduct of monetary policy and interest rates transmission not only for developed countries but also for emerging markets (Goswami et al., 2009). Its effect on economic growth, however, is not studied yet and it lacks both a theoretical and empirical background. There is a large literature about the causal relationship between financial development and economic growth, but this literature did not pay enough attention to financial innovation and specifically securitization. I will try to fill this gap in the literature by connecting securitization to economic growth performance in a cross-country perspective. In this paper, regarding securitization as a part of financial development I try to distinguish its impact on economic growth. A traditional financial development indicator is also kept in the analysis to control the effect of securitization through conventional means such as loan supply. Two different approaches are used; one is a long-term dynamic panel approach to see the effects of securitization technology on economic growth. The other one is about the impact of securitization on economic performance of countries during the recent global financial crisis. According to the analysis, securitization technology has a weak positive effect on countries long-term economic growth rates, but this implication is not robust to different specifications. Moreover, this relationship seems to hold also for the initial period of recent financial crisis: countries with securitization technology perform better than other countries at the beginning of the crisis, namely in the year When we consider the whole crisis period or the rather serious part of the crisis , we cannot find any significant difference among two groups. Moreover, there is also no significant impact of securitization technology on the growth performances for the year 2006, which was a boom year. These findings suggest that securitization technology is especially useful to deal with a crisis, but when the crisis becomes serious -or systemic- this positive impact cannot be observed. These findings may seem counterintuitive as securitization is claimed to be one of the most important causes of the global financial turmoil. It is crucial to note that the analysis does not say much about the formation of the crisis. It indicates, however, that securitized countries allocate their risk better or in other words transfer their risk to nonsecuritizing countries- and thus perform better than the countries without that technology at least at the beginning of the crisis. This supposedly better risk allocation does not 4

5 create any significant difference, however, when the crisis proved to be very deep threatening the system as a whole. The basic contribution of this paper would be a better understanding of securitization phenomenon and its impact on countries growth performances in a cross-country perspective. The model used is a modification of the main stream model of financial development and economic growth literature. The remainder of this paper is organized as follows: chapter 2 provides a literature review of the relevant areas of economic research. Various microeconomic studies about securitization are presented together with financial development and economic growth literature, as we try to link these lines of research. The model and data is introduced in chapter 3. This chapter also includes a detailed presentation of the econometric methodology used in the dynamic panel and cross-sectional crisis regressions. The empirical analysis is done in chapter 4 with a discussion of the results. Chapter 5 concludes with a discussion of limitations of the approach and future research. II. LITERATURE REVIEW a. Securitization and Financial System Two seminal papers, Diamond and Dybvig (1983) and Diamond (1984), define two basic functions of banking, namely liquidity transformation and delegated monitoring. These two functions are part of basic functions of financial system summarized by Levine (1997), where other functions are optimal allocation of capital, mobilization of savings and ease of exchange. Financial innovation is continuously transforming the financial system and its relation to real economy, but it is not clear whether this transformation is beneficial for the system as a whole (Rajan, 2006). The securitization technology, as one of the most influential financial innovations, had a dramatic impact on financial system and its functions. Specifically, banks, which were originator and holder of illiquid loans, became originator and distributor of these loans. Loutskina and Strahan (2009) find loan supply s dependence on the financial condition of the lenders is reduced by securitization. Furthermore, Keys et al. (2010) demonstrate the adverse effects of securitization on screening incentives of lenders in their subprime loans. There is also some empirical evidence of lower credit standards associated with securitization, which is also in line with worse screening theory (Dell Ariccia et al., 2009). These findings indicate a fundamental change in banks role in the financial 5

6 system, which may also have macro consequences such as economic growth, monetary policy changes or even financial crisis. Why are the banks and other financial institutions securitizing? The main incentives for banks to securitize or to change to the new originate-to-distribute model- are regulatory capital arbitrage, which lowers capital costs of banks and is not possible with Basel II regulation, gaining extra liquidity and more efficient risk sharing (Bannier and Hänsel, 2007). The expectation from securitization is new profit opportunities, lower cost of funding and better bank performance. The empirical findings, however, are rather mixed. In one hand, Panetta and Pozzolo (2010), for instance, find that the results of securitization are ex-post in line with the expectations in a cross-country bank level analysis. Again, using individual bank data Affinito et al. (2010) find that banks once they securitize have higher profits and lower bad loans. On the other hand, in their study with US bank data and propensity score matching technique, Casu et al. (2009) conclude that first-time securitizing banks would have comparable costs of funding, credit risk and profitability if they would not securitize. A crucial point is the complexity of these financial instruments. Creating a high fixed cost to originate securities, this complexity is a barrier to enter the securitization market (Panetta and Pozzolo, 2010), but there is no effective barriers to buy these highly sophisticated securities and participate the market as a buyer rather than originator. The complexity of securities -caused by the wrong risk assessment and valuation by credit rating agencies- lead to global misallocation of credit (Levine, 2010). The positive incentives for securitization led some financial institutions to use the technology too aggressively and the complexity of this process contributed to the recent collapse of financial system (Caprio et al, 2010). Another interesting line of research is concerned with macroeconomic linkages of securitization. The focus of this approach has been the effect of securitization on monetary policy transmission. It has been well documented that the securitization phenomenon influences the monetary policy transmission via liquidity and credit channel 4 (Altunbas et al., 2009; Estrella, 2002). Goswami et al. (2009) argue that the relationship of monetary policy transmission and real economy has become much more complex and for national authorities it is now more difficult to control the economy. 4 Or so called, bank lending channel. 6

7 They also point out that the impact of interest changes on real output is affected by financial market deepening only through securitization. In a macro perspective securitization can also affect economic growth; this issue will be discussed in the next section. b. Financial Development and Economic Growth The financial sector development and the economic performance of countries drew a lot of attention for decades and economists tried to establish causal linkages between financial development and economic growth 5. There are various theories about the effects of financial development on macroeconomic performance i.e. economic growth, capital accumulation and technological change- and numerous empirical studies using different econometric techniques try to reveal the relationship between financial sector and the economy both with a macro and micro perspective (Beck, 2008). In their cross-sectional analysis King and Levine (1993) demonstrate strong association between financial development and various economic growth variables with a cross-country perspective. More recently the literature deals with the chronic endogeneity problem with instrumental variable estimation or dynamic panel data approach. Levine, Loayza and Beck 6 (2000) use legal origin of 71 countries to instrument their financial development and show that exogenous variation in financial development leads to higher economic growth for the years They also reach the same conclusion using a dynamic panel data estimator. These dynamic panel data estimators are also used for within country studies for example using provinces as individuals (Hasan et al., 2009; Anwar and Nguyen, 2009). There are also influential papers establishing a casual relation going from financial development to economic growth with country or industry specific analysis. For example, Rajan and Zingales (1998) show that lower cost of external finance fosters firm growth and new firm formation, which indicates the microeconomic roots of financial development and economic growth theory. An alternative approach is using time series analysis -with the traditional VAR (vector autoregression) or VEC (vector error correction) models (e.g. Arestis and Demetriades, 1997)- or more recent panel estimations with longer time dimensions and panel unit root and cointegration tests 5 For a thorough review of the issue Levine s (2005) book chapter on finance and growth. 6 Referred as LLB (2000) from here onwards. 7

8 (e.g. Christopoulos and Tsionas, 2004). Nevertheless, there are also some papers claiming financial development is following the economic success (e.g. Kar and Pentecost, 2000). One of the most crucial issues in this literature is how to measure the financial development. In their seminal paper, King and Levine (1993) suggested four different measures: The first one is the traditional financial depth indicator, which is liquid liabilities of financial system over GDP. The rationale behind this measure is that the deeper financial system is the more financial services and practices are provided and applied. The second indicator is composition of financial system between central banks and deposit money banks. The theory suggests that banks provide better financial functions than the central banks and thus the higher involvement of deposit money banks the more developed is the financial system. The other two indicators deal with allocation and size of credits to private sector, more specifically domestic credit to non-financial private enterprises over GDP or over total domestic credit. There are also other indicators suggested by the literature such as public ownership of banks (La Porta et al., 2002). It is important to note that none of these indicators are perfect as they cannot reflect financial services such risk management, information processing and reduction in transaction costs thoroughly. The level of analysis, the econometric methodology used and the identification strategy followed are all crucial elements in this line of research due to prevailing endogeneity issues. It may, however, be even more important to understand the theory behind the macro-financial linkages. Obviously, the financial system is not static and it evolves with financial innovation just as the real economy evolves with technological innovation. Thus, economists try to model financial innovation s real impact on macroeconomic phenomena, such as macroeconomic volatility or recessions (Gai et al., 2008 ; Dynan et al., 2006). Still, the impact of securitization on economic growth is not dealt with in the literature. We try to establish this link between securitization and economic growth in the next chapter, where we include a securitization technology dummy in the standard financial development and economic growth model. Moreover, it is also not obvious how securitization influenced the growth performance of countries during the recent financial crisis, which is partially caused by securitization, in a comparative cross-country perspective. A cross-country analysis of the recent financial 8

9 crisis is appropriate as developed and developing economies have many aspects in common both before and during the crisis (Reinhart and Rogoff, 2008). The theory suggests that securitization should mitigate the impact of a negative shock as it mitigates the constraints of financial institutions (Loutskina and Strahan, 2009) or possibly transfer the risk to other parties. Yet, it may obviously not hold with a systemic crisis, in which all economies are hit by the crisis. The effects of securitization are analyzed in a crosscountry perspective in the next chapter. III. THE MODEL, DATA AND ECONOMETRIC METHODOLOGY a. The Model We set up the model following the main financial development and growth literature mentioned in the previous chapter. To establish a causal link between financial development and long-run rates of per capita GDP growth, one has to control for various variables. The securitization phenomenon is a new factor influencing the relationship between finance and macroeconomic growth. Creating extra liquidity for financial system securitization stabilizes and increases the loan supply and thus affects economic growth through the conventional financial development, which is controlled. I also add the variable of interest, securitization, to the model. This variable will absorb other factors, through which securitization might influence economic growth. One of these factors can, for instance, be better risk allocation and diversification opportunities, which are not well captured by conventional financial development indicators. Some seminal studies already explained theoretically and empirically how better risk sharing can increase economic growth (Acemoglu and Zilibotti, 1997; Athanasoulis and van Wincoop, 2000; Obsfeldt, 1994). The model becomes: where is real per capita GDP growth, is a financial intermediary development indicator, is a dummy variable, which takes the value 1 if the country did securitization before and thus has the securitization technology. I assume that countries or banks- are using the securitization technology optimally, so the crucial point is not how much they securitize but whether they can securitize or not. The 9

10 conditioning set includes control variables widely used in the finance-growth models and controls for remaining factors influencing economic growth. There are two conditioning sets, the basic conditioning set and the full conditioning set. The basic set incorporates only the initial levels of economic wealth (GDP per capita) and the educational attainment together with a constant. These variables capture possible convergence effects and human capital effects, respectively. The full set, on the other hand, is more comprehensive as it controls for government expenditures, inflation and openness of the economies together with the two controls of the basic set. Government expenditures capture the government s relative economic activity in the economy, whereas inflation is used as a macroeconomic stability indicator and openness tries to capture the growth benefits of participating to the global economy. A negative sign for initial wealth levels will indicate a convergence pattern. Moreover, positive effects of human capital and openness are expected, whereas higher inflation meaning macroeconomic instability- and higher government involvement are expected to have negative effects on long-term economic growth. In the last decades there have been enormous amounts of securitization of various illiquid assets from student loans to credit card debts. This badly supervised process claimed to be one of the leading causes of the recent financial crisis. To our knowledge, the assessment of the impact of securitization on economic growth during the recent financial crisis has not been done. The model we use to evaluate securitization s impact during the financial crisis is very similar to the finance-growth model discussed above. This short-term model, however, uses a cross-sectional approach only and thus it may suffer from endogeneity, which is addressed in more detail in the next section. The interpretation of the control variables changes also due to the short-term nature of the setting. All the explanatory variables are from the year 2005, which is the last year of our securitization dataset and a boom year before the crisis, whereas the dependent variable is the average or yearly real per capita GDP growth rates for the years before and during the crisis. These real per capita GDP growth rates are taken as a proxy for the economic growth experiences of countries during the relevant period. I also add a control for institutional quality in this model to be able to cope with omitted variable problem better. The crisis model becomes: 10

11 This model cannot explain the cause or the dynamics of the financial crisis; the mere purpose is to associate the countries situations before the crisis with their growth performances during the relevant time periods given that the crisis occurred. Again, the variable of interest will be securitization and its expected impact is ambiguous. Securitization is expected to reduce the negative impact of economic crisis thanks to better risk allocation or lessened credit constraints. Nevertheless, in a systemic crisis like the recent economic crisis-, which affects every country, securitization technology may harm economic growth of securitizing countries -for instance- through international linkages, bankruptcies or credit crunches. b. Data i. Panel Data The panel consists of data for around 70 countries, which are listed in the appendix Table A, and covers the period Note that, these countries are not selected arbitrarily, but the set is almost identical with the LLB (2000) paper. We use nonoverlapping four years averaged data, which provides 5 observations per country ( , , , and ). The panel is strongly balanced and only a few observations are missing. For the dynamic panel regression per capita GDP in constant 2000 US dollar is used, the data is taken from World Bank s World Development Indicator database (WDI). This variable is transformed using natural logarithm, so including its lagged value as an explanatory variable leads to growth rates as we will discuss in the next section. The variable of interest is securitization, which is one of the explanatory variables. For the securitization data we use the ABS Database, which includes securities collateralized by assets of some kind and rated by at least one major rating agency. The main types are public and private asset-backed securities, mortgage-backed securities and collateralized debt obligations (CDO). It excludes, however, some important types of 11

12 securitizations such as commercial mortgage-backed securities 7 or Fannie Mae and Freddie Mac issues. Thus, we will not use the amounts of securitization; instead we create a dummy variable representing the securitization technology. Once there is an issuance origination of a security- in a country, I assume that this country has the securitization technology from that period onwards 8. The years of first securitizations are given at the Table A in the appendix. This dummy variable is used as a proxy to the securitization activity of the countries and in around 30% of the observations securitization technology is present. From other explanatory variables financial intermediary development indicator, initial wealth, education and inflation are taken from WDI database, whereas openness indicator and government involvement variables are from Penn World Table 6.1. As financial intermediary development indicator we use three different variables, namely domestic credit to private sector over GDP, liquid liabilities over GDP 9 or M3 over GDP- and domestic credit provided by banking sector over GDP. The last two measures are used in the robustness checks. The initial values of natural logarithm of per capita GDP and average total schooling over 15 in years are used as initial wealth and education controls. Openness indicator is total trade exports plus imports- over GDP in constant prices and government share in real GDP is used to proxy government s involvement in the economy. Note that all the ratios are in percentages. Also note that there are a few outlier observations in inflation variable countries with hyperinflation. These observations are excluded from the regressions, but their inclusion does not affect the results. The summary statistics and correlations are provided in the appendix, Table B and Table D respectively. ii. Cross-sectional Crisis Data As we discussed previously, the model for the crisis regressions is very similar to the dynamic panel model. An important difference is that the dependent variable is not per capita GDP level but the average or yearly growth rate of real per capita GDP for specified period in percentages and they are from WDI. The definitions and sources of other variables are the same and they are for the year As for the securitization 7 Commercial MBS are very important type of securitization and their issuance is higher than ABS (assetbacked securities). 8 Including the period of securitization. 9 With this variable number of observations decrease dramatically, it is only used for robustness check. 12

13 technology dummy we assume that the non-securitizers did not enter the market till the crisis, which can be seen as a reasonable assumption as even though they would securitize they would need some years to be able to use the technology effectively. As the institutional control variable I use the rule of law index from Kaufmann indicators. Again the summary statistics and pairwise correlations for the cross-sectional crisis data are provided in the appendix, Table C and Table E respectively. c. Econometric Methodology i. Dynamic Panel Estimators The literature mainly starts with a cross-sectional estimator, which will be skipped in my analysis. The reason for that is the nature of the variable of interest, namely securitization technology as a dummy variable. Unlike traditional financial development indicators, securitization technology cannot be used in a cross-sectional analysis via averaging. It can be used, however, as a dummy variable in such a way that securitization dummy gets one for the countries, which used securitization at any point between the year 1985 and This is a very unreliable way of modeling securitization, as it does not differentiate -for instance- between USA, who has the technology in the beginning of the period, and Chile securitizing at Such huge measurement error will bias the results of cross-sectional regressions and even though such a model is estimated, it would not provide precise results 10. Instead, we begin the analysis using a widely used dynamic panel approach, where the subscript is used to identify countries and for time periods. The typical Arellano-Bond dynamic equation is as follows:, (1) where is the dependent variable, is the lagged dependent variable, represents the explanatory variables, and the error terms. The equation also includes as the country specific fixed-effects. The growth literature uses this estimator smartly. Using logarithm of real per capita GDP as, is subtracted from both sides:, (2) 10 Indeed, in the cross-sectional regressions using OLS or IV estimations securitization dummy is extremely insignificant. The results are not provided in the paper, but available upon request. 13

14 where the log difference in the left-hand side of the equation is the average real per capita GDP growth, is set of explanatory variables including the variable of interest securitization dummy-, and represent error terms. Moreover, the equation also includes, a set of period dummies to control time-specific factors. In short, equation (1) will be estimated, but equation (2) will be used to interpret the growth rates. This approach is superior to cross-sectional analysis in many aspects (Beck, 2008). First of all, the dynamic panel estimator can control for possible endogeneity of all explanatory variables instead of, for example, only financial development indicator in cross-sectional studies. Using country-fixed effects mitigate the omitted variable problem. Secondly, these estimators have a higher explanatory power including the lagged dependent variable, which controls for possible business cycles. Estimation with lagged dependent variables will be highly biased as will be correlated to. Moreover, compared to cross-sectional regression dynamic panels use a higher number of observations as it exploits both time-series and cross-sectional dimensions of data. Another advantage is that external instruments, which are sometimes very difficult to find, are not a necessity as these estimators can use internal instruments, such as lagged levels or lagged differences of variables. All in all, these dynamic panel estimators are standard account to estimate large N, fixed T datasets and is generally superior to OLS estimates. Following LLB (2000) we take all the explanatory variables as predetermined meaning that current values of these variables can be correlated with post and current error terms but not with future error terms. Another crucial assumption for dynamic panel estimators is that the error terms are not serially correlated. Just with these assumptions a consistent GMM estimator can be derived from the first difference of equation (1). First differencing helps to get rid of the country specific fixed-effects:, (3) 14

15 The difference GMM estimator (Arellano and Bond, 1991) uses lagged levels of the explanatory variables as instruments, thus the moment conditions used in this estimator becomes: (4) (5) There may be, however, an identification problem with the variable of interest, securitization technology, which is a binary variable. The lagged values of securitization technology may not be a good instrument for the first difference of this variable, as for each country the first difference takes the value 1 only once for the countries, who starts securitization at some point and all the other values are 0. Note that, however, when I run an auxiliary regression for the first stage, lagged level of securitization seems to be a significant determinant of the first difference of securitization technology. To assure a good instrumentation we add an exogenous instrument to the estimation, namely population. The argument is that population is exogenous and it has only an indirect effect on growth rates through securitization. The countries with higher populations tend to have larger financial institutions simply because these countries will have larger economies and thus more people willing to put their money to deposit accounts or more credit users to pool. Panetta and Pozzolo (2010) argue that there are fixed costs of originating complex securitization deals and larger banks have an advantage to afford these costs. They, indeed, find that larger banks are more likely to securitize their assets. Haensel and Krahnen (2007) confirm this finding for Europe by telling the typical originator in their sample is a large financial institution. Loutskina (2005) provides two additional explanations for larger banks securitization tendency. Firstly, large banks have better relationships with investment banks, which can link them to derivative markets or in other words provide them with securitization technology. This can reduce cost of securitization and tempt large banks to securitize. Secondly, larger banks are advantaged as they have their own diversified and large loan pool, which can be securitized, whereas smaller banks have to attract other financial institutions to 15

16 diversify their loan pool enough to securitize. All in all, countries with higher population tend to have larger financial institutions and larger pool of eligible customers, and thus, they have a larger probability to use securitization technology. Here we assume that scale effects are non-existent, which is also supported by empirical studies (Jones, 2005). Population is used as an instrument by some macro studies. For example, Boone (1996) used logarithm of population to instrument aid flows. He also points out the economic success of many small countries and the economies of scale argument for economic growth is not supported empirically and thus, exclusion restrictions hold. So I use population as an external instrument, which adds the following moment condition:, (6) where represent population variable. The difference estimator above has various drawbacks. First of all, it only exploits the time series dimension of data as it uses first differences, but not the cross-country dimension, which can be a crucial part of the analysis. One distinct flaw specific to my model with securitization technology is that by first differencing the difference estimator just identifies the observations, where countries acquire the securitization technology. All period securitizers or non-securitizers do not add anything to the estimation of securitization technology coefficient. In other words, it leads to waste of information. Furthermore, Beck (2008) states that when the right-hand side variables are persistent over time the lagged values become weak instruments deteriorating the asymptotic properties of the difference estimator. The system GMM estimator (Arellano and Bover, 1995) can deal with these drawbacks exploiting the information in levels. The original regression in levels is added to the system together with the first differenced regression and lagged first differences are used as instruments. To be consistent this estimator needs an extra assumption, namely lagged first differences of right-hand side variables should be orthogonal to country specific fixed-effects. Thus, the additional moment conditions become:, (7) 16

17 , (8), (9) Another issue relevant for both difference and system estimators- is whether to use one-step or two-step GMM estimators. Two-step estimator uses optimal weighting matrix, which is estimated by a first-step imposing no restrictions on the covariance matrix of the error terms i.e. using identity matrix as weighting matrix in the first-stepand gives the most efficient estimator (Verbeek 2008). On the other hand, imposing some restrictions on the error term, namely homoskedasticity and no autocorrelation assumptions, the optimal GMM estimator can also be estimated in one-step, which can be advisable in small samples according to Verbeek (2008). Roodman (2009) states researchers are used to report the results of both estimators as the standard errors of the two-step estimator has a downward bias, which is reduced through Windmeijer (2005) correction, which adjusts the covariance matrix for finite samples. Although the literature in general uses the two-step GMM estimation we will report both for illustrative purposes. Nevertheless, the two-step system GMM estimator with Windmeijer (2005) correction seems to be the best estimation method with lower bias and standard errors and time-invariant regressors can also be included. There are three main tests to determine the consistency of dynamic GMM estimations. The first test is the Sargan/Hansen overidentifying restrictions test with the null hypothesis that instruments are exogenous. If the null hypothesis is not rejected, it means that the instruments are valid. Note that Sargan OIR test is not robust to heteroskedasticity, and thus, for two-step optimal GMM estimations Hansen J statistic is used, which is robust but sensitive to the number of instruments (Roodman, 2009). Other test is to check possible serial correlation in the error terms. The null hypothesis of this Arellano-Bond test for autocorrelation is no autocorrelation in differenced residuals. The test concerning the first-order process is expected to reject the null hypothesis even if there is no autocorrelation in the original error terms as by construction both and have. The second-order test in first differences is the 17

18 crucial one as it detects autocorrelation in levels. The third test is a difference-in- Sargan test to check the orthogonality of a subset of instruments for example the external instrument population- with the null hypothesis that these additional instruments are valid (Beck, 2008). ii. Cross-sectional Crisis Estimation The model we use for the crisis regressions is very similar to the dynamic panel regressions. A cross-sectional approach is used instead of the panel approach and it comes with a higher possibility of an endogeneity problem 11. Securitization technology may well be endogenous and there may be unobserved factors, which affect both economic growth performances and securitization tendencies of countries. Although the exogeneity of securitization technology is an extremely strong assumption, the fact that we use lagged explanatory variables should mitigate endogeneity problem and thus this analysis can be a good start to understand the impact of securitization during the financial crisis. The cross-sectional equation is as follows:, (10) where represents individual country, which has just one observation, stands for the real per capita GDP growth rates, includes the exogenous control variables introduced with some lag 12, stands for the error terms and is the possibly endogenous securitization variable. The possible omitted variable bias is derived in the appendix, Section b. Instrumental variables estimation can mitigate the omitted variable problems or even other problems such as measurement error. Exogenous variation in the instruments will identify the exogenous variation of the endogenous variable, securitization technology dummy. Both two-stage least squares (TSLS) and GMM can be used, but GMM is more efficient as it uses the optimal weighting matrix as we discussed in the previous section. Together with population -also used in the dynamic panel estimations- we use log of biggest bank s assets of a country as our second instrument. The argument to use this 11 Reverse causality is not a big issue here, as securitization technology variable is from 2005 as explained in the next section. Omitted variable bias or measurement error problems are more important. 12 All the controls are from the year 2005, for which we have the securitization data. 18

19 variable as an additional instrument is the same with population and discussed above. As robustness check we also use absolute latitude and log GDP levels as instruments. In the literature, latitude is used as an instrument for financial development (McCaig and Stengos, 2005). The argument is that absolute latitude through colonization strategy a la Acemoglu or through climate shaping behavior of societies affects securitization tendency of countries, but not the growth rates directly. Just like in the previous section, there are tests for the validity and weakness of instruments. The details of these tests will be provided the next chapter, when we specify the estimation method and the model. The moment conditions for the GMM estimator come from: (11) where represents a vector of instrumental variables. A relevant way to increase the precision of the estimation is to impose more structure, namely use a nonlinear approach to estimate the first-stage (Cameron and Trivedi, 2009). As the endogenous variable, securitization technology, is a binary variable, a linear probability model in the first stage may not be an appropriate one. Instead, we impose treatment-effects model, which takes securitization amount as an unobserved latent variable 13,, determining the binary securitization dummy variable as 0 or 1. The model becomes:, (12), (13), (14) That way, probit regression can be used for the first stage with and. The most important drawback of this approach is that the diagnostic possibilities for good instruments are limited. Furthermore, extra structure imposed may lead to a possible misspecification error. The estimator becomes inconsistent in the case of heteroskedasticity, but using the two-step estimator we get a consistent estimator. 13 This unobserved variable can for example be the amount of securitization, which is originated optimally meaning that it allocates the risk in an optimum way and has a right valued underlying asset. 19

20 IV. EMPIRICAL ANALYSIS a. The Dynamic Panel Model i. Specification Tests To confirm the validity of the identification strategy, specification tests for this kind of dynamic panel data estimations should be analyzed (Arelano and Bond, 1991). The specified estimators require serially uncorrelated error terms for consistent estimation. We use the Arellano-Bond test for autocorrelation with the null hypothesis of no autocorrelation (both for first and second order). The first-order autocorrelation tests for some specifications reject the null hypothesis as expected 14. None of the second order tests, which are more important, are rejected in any conventional significance levels meaning that and are serially uncorrelated and there is no serial correlation in the original error terms. The second specification test is an overidentifying restrictions test for the validity of the instruments, which is called Sargan/Hansen test (Beck, 2008). In one-step estimators error terms assumed to be homoskedastic so that Sargan OIR test will be consistent- and the Sargan test p-values are reported. The instruments in the difference estimator pass the test whereas the null hypothesis is rejected for system estimator, suggesting that instruments are not valid for this estimator. Using two-step estimators, which are robust to panel autocorrelation and heteroskedasticity, is advised by the literature. Indeed, the instruments in the two-step estimation pass the OIR test. Note, however, Hansen J statistic is used for this test instead of the Sargan statistics, because it is robust. It may be also important that the regressions with the basic conditioning set have lower Hansen test p-values. This may indicate that the full conditioning set specifications are better in terms of instrumentation of predetermined variables. Another explanation can be that the test became weaker as the number of instruments increased (Bowsher, 2002). Nevertheless, Roodman s (2009) rule of thumb to limit the number of instruments to the number of groups is almost met, as only two specifications have higher number of instruments 15. The last test, which is not reported in the table 1, is differencein-sargan test to check the validity of additional instruments. Again, except one-step system estimator all the specifications pass the test supporting the model and additional instrument. Thus, we conclude that the one-step system estimator seems problematic as 14 As discussed in the previous chapter: both includes. 15 In regressions 6 and 8, 83 instruments are used. It can be argued that the difference between number of instruments and number groups is not dramatic. 20

21 the instruments are not working well. It can well be that the homoskedasticity assumption does not hold leading to inconsistency of Sargan test. Thus, rejection of OIR tests does not necessarily mean that the specifications are bad, but one has to be cautious with the results. On the other hand, the two-step estimators seem to work reasonably well. I will use the two-step system estimator, which is suggested by the literature, as my baseline regression. ii. Results and Interpretation The results of the dynamic panel data estimations are given table 1 below. The dynamic panel estimates using various specifications and econometric methods cannot establish a robust link between securitization technology and economic growth. The sign of the securitization technology s coefficient is positive and rather significant throughout the dynamic system GMM regressions, but dynamic difference GMM estimator cannot find a significant impact of securitization technology. This difference is not very surprising as the difference estimator only uses time-series dimension and thus loose quite some information. Although the two GMM estimators should be asymptotically the same, they can differ in small samples as we discussed in the previous chapter. Two estimators also diverge in some control variables. Education seems to have a counterintuitive negative impact on growth in difference estimator, whereas this impact is positive in the system estimator as expected. Indicating the importance of macroeconomic stability inflation seems to have a significant negative impact as expected. The coefficients of initial per capita GDP levels are negative indicating to convergence but insignificant. Government expenditures are also insignificant with a counterintuitive sign. Openness, on the other hand, seems to have a positive impact. An interesting pattern observed in the estimates is that compared to the difference estimator the system estimator finds a higher impact of securitization technology with relatively higher significance levels. The coefficients of system estimator range from 0.03 to meaning a 3-4.5% increase in economic growth 16, which is economically significant since per capita GDP growth average is only 1.4% for non-securitized observations. On the other hand, the difference estimator indicates there is no significant impact securitization on growth. Instead the difference estimator finds a significant positive 16 Due to the fact that the model is log linear, one can calculate the marginal effects by multiplying the relevant coefficient by

22 impact of financial intermediary development indicator besides the securitization technology-, which is not present with the system estimator. According to difference estimator, one percentage increase in domestic credit to private sector over GDP leads to an increase around 0.15% in annual per capita GDP growth, which is economically significant. Low and middle income country average of domestic credit to private sector the financial intermediary development indicator- for the years 1985 to 2005 is 48%. For this period, India has an average of 28%, whereas Argentina s average is only 19%. If these two countries would reach the average their growth rates will increase by 3% and 4.5% respectively and such increases are indeed dramatic compared to their realized growth performances, 4% and 1% -respectively. These regressions are not robust enough to reach a firm conclusion about the impact of securitization; the difference estimators do not show any significant relationship between the securitization technology and economic growth. The dynamic one-step system estimator, on the other hand, finds a significant positive impact of securitization technology on economic growth. These one-step estimates, however, are not consistent due to possible weakness of instruments or heteroskedasticity. The dynamic two-step system GMM estimator should provide consistent and the most efficient estimates. Indeed, the coefficients of securitization are lower than the one-step estimators and not significant in any conventional significance level. The two-step system estimator with the basic conditioning set cannot find a significant impact of securitization. In the dynamic two-step system regression with the full conditioning set (regression(8)), one-sided Wald test rejects the null hypothesis of a negative or zero effect of securitization technology ( ) in favor of positive impact ( ) at the 0.1 significance level 17. Thus, one can argue that securitization does not have a significant impact on economic growth in a cross country perspective for the years But if there is an effect of securitization, it would be a positive one. iii. Sensitivity Analysis In this section, some robustness tests will be evaluated to the specifications mentioned in the previous section. The two-step system estimator will be used as the baseline regression, since it is most widely used in the literature as the most efficient 17 To calculate p-values of one-sided Wald test, I divide two-sided p-values by 2. The calculated p-values are (=0.187/2). 22

23 estimator- and the Table 2 containing the results can be found below. The main issues we touch upon are the selection of number of instruments, possible non-linearity of explanatory variables, exclusion of the traditional financial development indicator, effects of alternative financial development indicators, use of different averaging time spans and inclusion of institutions as a control. As Roodman (2009) suggests the choice of instruments can change both the results of the regression and the tests for validity of the instruments. Hansen OIR test is robust but it gets weaker as the number of instruments increase, thus it is important to use different number of lags as a robustness check. We reduce number of instruments by using only second lag of endogenous variables and lagged differences but not deeper lags. The number of instruments decreases from 83 to 41, the Hansen OIR test is rejected again still indicating the validity of instruments but its p-value is much lower suggesting possible weakness of the test. Securitization technology seems to be insignificant compared to baseline regression, which can be caused by worse instrumentation. Another issue with the instrumentation is the use of population as an exogenous instrumental variable. We estimate the baseline setting without population used as an instrument. The results look quite similar but including population as an instrument seems to increase the coefficient and significance of securitization dummy. It supports the use of population to solve identification problem of securitization technology and reduce a possible downward bias 18 on the coefficient of securitization. Another issue is possible non-linear effects of explanatory variables. Using natural logarithms of explanatory variables 19 we estimate the models again. The securitization technology dummy, again, is insignificant with lower coefficients. Inflation also becomes insignificant indicating non-linear relationship with economic growth. There is also a regression (5), which excludes financial development indicator. The results are very similar to the baseline regression. This may indicate that the impact of securitization on economic growth through lending channel that is an increase in loan supply related to securitization technology- is not an important one or the securitization technology 18 The reason of a possible downward bias on securitization technology is discussed in detail in next section for crisis regression. 19 Education and institutions variables are not transformed. Inflation is transformed by first adding one and then taking natural logarithm. 23

24 dummy cannot absorb this relationship. We also include institutions in linear and nonlinear setting (regressions (4) and (8)). Assuming that institutions are persistent and changes only very slowly we use the average Kaufman indicator for the rule of law in the last decade as the institutions variable, which has a significant positive impact in both regressions. In both regressions controlling for institutions increased the coefficient and significance of securitization technology compared to the regressions without institutions. Moreover, in the log-linear regression (4) inflation becomes insignificant; it can possibly indicate that inflation not only proxies macroeconomic stability but also institutional elements. We also estimate the baseline regressions (9)-(11) using alternative measures of financial development indicator, namely domestic credit provided by banking sector and liquid liabilities. We have liquid liabilities variable for only 40 countries, so it is also a good robustness check with a smaller sample of countries. The decrease in the number of countries, however, leads overfitting and one has to decrease number instruments to reduce bias (Roodman, 2009). Indeed, regressions with lower number of instruments result in more precise estimation for both alternative measures. The regressions with the reduced number of instruments and alternative financial development indicator find a positive significant impact of securitization technology around 4-6% in 10% significance level. Finally, we use an alternative panel dataset, in which yearly data is averaged over non-overlapping 5 year periods instead of 4 years. The panel again covers the period between 1985 and 2004, but now the time dimension includes 4 periods instead of 5. The literature traditionally use 5 year averages, but we preferred 4 years so that we will be able to use Arellano-Bond second order autocorrelation test to test the basic assumption of dynamic panel models. The result of regression (12) seems quite similar to the baseline regression using 4 year averages. The coefficient of securitization technology is lower, but the significance level is still low enough to reject the one-sided test with the null hypothesis of positive or no impact of securitization. To conclude, sensitivity analysis indicates that the positive impact of securitization technology is not consistent, and in some specifications securitization technology seems to have no significant impact on growth rates. 24

25 b. The Crisis Regressions i. Specification Tests The regression results are discussed in the next section and the identification strategy is discussed in section II. Here, different specifications will be evaluated and compared using various specification tests for instrumental variables estimation to decide whether the instruments are valid. Three IV regressions are reported with different explanatory variable sets together with two Treatment regressions. Robust standard errors are used in each regression allowing for arbitrary heteroskedasticity. Indeed, tests for heteroskedasticity reject the null hypothesis that the disturbance is homoskedastic in all the IV regressions 20. First of all, we check whether the models in IV estimations are identified using a Lagrange Multiplier test, where the null hypothesis that the equation is underidentified (Baum et al. 2010). All three tests are rejected at the 1% significance level indicating that the models are identified. Due to the fact that the model is overidentified in terms of instruments, it is also possible to use the overidentifying restrictions (OIR) tests for the validity of the instruments. The Sargan-Hansen overidentification tests 21 have the null hypothesis that the instruments are valid that is, uncorrelated with the error term. None of these tests are rejected indicating the validity of the instruments. Although the instruments seem relevant and valid, they are not necessarily strong instruments. If the exogenous variation in the instruments cannot identify the variation in the endogenous variable securitization technology dummy here-, then asymptotic theory will not work properly with finite-sample distributions. There are several ways to detect a possible weak instrument problem. The simplest method is to check the pairwise correlations between the endogenous regressor and the instruments. The pairwise correlations are given in Table 3. Population and absolute latitude do not have not particularly high correlation with securitization technology dummy, but they are not too low to indicate a weak instruments problem. On the other hand, log assets of the biggest bank in the system and log GDP level are highly correlated with the securitization technology. All candidate instruments are significantly and 20 ivhettest command is used to test the homoskedasticity of disturbances. The test results are not reported but they are available upon request. 21 With the efficient GMM estimator OIR test statistic is Hansen s J statistic, whereas for 2SLS it is a Sargan s statistic. 25

26 positively correlated to our endogenous variable as expected. Another conventional way to diagnose the weak instruments is analyzing the first stage regressions. The results of the first stage regressions with our main two instruments, namely log population and log assets of the biggest bank, are given in the table 4 below. It seems that population has a significant positive impact on securitization technology as argued; whereas biggest bank s assets has a negative counterintuitive impact suggesting it may not be an appropriate instrument note, however, when included as the only instrument its coefficient turns to positive and highly significant. One other way to interpret the first stage regressions is to assess the F statistic for the joint significance of the key instruments in the first stage. Staiger and Stock (1997) suggested the rule of thumb that the instruments are weak if the F statistic is less than 10. IV regressions meet this rule of thumb indicating strong instrumentations (the first one is a bit below 10). The first stage results from probit regressions, on the other hand, provide rather low joint significance for excluded instruments, which may be indicating weak instrumentation. Furthermore, the partial of excluded instruments are around 0.31 indicating our excluded instruments only, explain almost one third of the variation in securitization technology after controlling for other regressors. Another weak identification test is proposed by Stock and Yogo (2005), where the null hypothesis is that the instruments are weak can lead to size distortions of Wald tests on the significance of coefficients. The null hypothesis cannot be rejected at 10% tolerance level, however it is rejected for 20% level for full specifications, which will allow to see the impact of securitization with onesided tests as in the dynamic panel part. Although Stock and Yogo s (2005) weak identification test provide some evidence for the weakness of the instruments, the theory does not apply exactly as it assumes independently identically distributed error terms and does not allow for heteroskedasticity. More discussion about weak instruments issue can be found in sensitivity analysis section below. It seems that there is some evidence of a downward bias to the OLS estimates of securitization technology, but this evidence is not strong as the endogeneity tests 22 cannot reject the null hypothesis that securitization technology is exogenous. As for the treatment regressions, the treatment-effects literature 22 These tests are defined as the difference of two Sargan-Hansen statistics: one for the equation with the smaller set of instruments, where securitization technology is treated endogenous, and one for the equation with the larger set of instruments, where it is treated as exogenous (Baum et al., 2010). 26

27 relies upon typical IV approach 23, which is discussed already, to decide on the validity of instruments (Yörük, 2009). Assuming that the instruments are valid, treatment-effect regressions are providing more precise estimation. Note that we use two-step treatment effects estimates, which are robust to heteroskedasticity. ii. Results and Interpretation Table 5 summarizes the regression results for various time periods during the financial crisis and before. The results suggest a positive impact of securitization technology on economic growth performance during the first year of the financial crisis, namely This finding is robust between different methods and specifications. There is no significant relationship for the other sub-periods of the crisis. The IV regressions with average growth provide rather marginal results, which can be used only with one-sided tests as in the dynamic panel regressions and most probably driven by the year For the whole crisis period or the last two years of the crisis, there seems to be no significant relationship between securitization and economic performance. Note, however, almost all the coefficients of securitization turn to negative, but stay insignificant. We observe the impact of securitization technology controlling for other factors from the year 2005 only for 2007, which is the closest time period to It may then be the case that these variables from 2005 are not persistent enough to have an impact on the later years of the crisis. To check such a possibility we regress the same variables on the growth performance in 2006 and find no significant relationship. Thus, the better performance of these economies in the beginning of the crisis can be explained by the better ability of the countries with securitization technology to respond to the crisis. This analysis suggests that the securitization technology improves economic performance in the bad times, where better risk allocation helps countries to respond to the bad economic conditions or may be downturns of their business cycles. It does not much of an impact during the good times, which make sense as we control the traditional mechanism working through increased loan supply, and again no impact if the crisis proves to be too big leading a systemic depression. Turning to the regression results for the year 2007 presented on table 6-, according to OLS regressions securitization technology leads to around 1.5-2% better growth 23 Other approaches used to validate the instruments are matched data, which will not be discussed due to data limitations, or alternative instruments, which is done in sensitivity analysis part (Yörük, 2009). 27

28 performance in the beginning of the crisis. This effect is significant at 10% significance level. In the IV GMM regressions securitization seems to be significant and has a higher coefficient compared to OLS regressions indicating a downward bias in OLS regressions. The countries, which have the securitization technology, had 3-3.5% higher growth performance compared to other countries without that technology. The treatment-effects approach confirms the findings and finds a significant positive impact around 2-2.5% on economic growth, which is higher than OLS and lower than IV-GMM estimations. Also note that, the treatment regressions have higher precision compared to other regressions as expected. The reason that OLS estimates are downward biased can be explained by a possible omitted variable bias. Equation (15) below, which is derived in the appendix, indicates that OLS estimator will not be the same with true parameter but includes a bias: (15) where is the securitization technology dummy, is the unobserved omitted variable and is the relation, for which we are looking for. Here, represents the association between the omitted variable and endogenous variable securitization technology- and is the impact of the omitted variable on the dependent variable- crisis growth. Both relationships are influential on determining the direction of the bias. A possible omitted variable is the systemic risk exposure of a national economy as a whole, which can have a positive correlation with the securitization technology as risk seeking countries try harder to get the technology but affects economic growth negatively. Then, is positive and will be negative leading to a negative bias of the OLS estimator. We assume that the effect of securitization technology is not through fostering excessive risk taking or credit expansion, which is partially controlled by financial development indicator- but through better risk allocation. As for other control variables, government expenditure seems to have rather high significance levels and a negative impact on economic growth as expected. Higher government involvement in the economy leads to bad crisis performance, which is intuitive, as government involvement can be related to bad public finances and excessive 28

29 sovereign debt worsening the economic performance. We also find a negative impact of education on economic performance in some regressions. This finding is highly counterintuitive and may be caused by the fact that the crisis stemmed from developed countries, which have higher education values. iii. Sensitivity Analysis For the sensitivity analysis, we estimate different models using various methods. TSLS approach instead of GMM provides virtually the same results with lower significance levels as expected. Excluding financial development increases the significance levels in each specification. This indicates that the effect of lending channel on economic growth is well absorbed by securitization technology dummy during the first year of the financial crisis. Using an alternative measure for financial development domestic credit provided by banking sector- does not change the results and actually increase the significance levels dramatically throughout different specifications. We also vary the external instruments; instead of population the size of the economy log GDP level- is used as an instrument and log biggest bank s assets are replaced by absolute latitude. None of these changes in instruments do affect the conclusion. When I use population as the only instrument the first stage F statistics become much larger than the rule of thumb value 10, and is around 20. This indicates that population, which is highly significant throughout the first stage regressions, is a strong instrument. The analysis with a conditional approach, in which p-values and confidence intervals are with asymptotically correct size even with weak instruments-, also supports population as a strong instrument (Cameron and Trivedi, 2009). All the size-corrected tests, namely conditional likelihood, Anderson-Rubin and Lagrange multiplier tests, provide exactly the same confidence interval, which is a bit wider than the conventional confidence interval. When we use population as the only instrument, we cannot test whether the instrument are valid as the system is just identified. In any case, using just population to instrument the securitization technology does not change the conclusions. V. CONCLUSION In this paper, we analyze the effects of securitization technology on long-run growth performances and the economic growth experience during the financial crisis in a crosscountry perspective. We modify the main model linking financial development and economic growth to include the securitization technology as a dummy variable. The 29

30 dynamic panel approach suggests that there may be a positive impact of securitization on economic growth, but it is not robust. This weak positive effect is not through increasing the loan supply lending channel- but rather other aspects of securitization such as ameliorations in risk sharing. This relationship seems to hold also at the beginning of the recent financial crisis: countries with securitization technology perform better than other countries in the year This impact, however, cannot be observed for the crisis period as a whole. The positive impact at the beginning of the crisis seems counterintuitive as securitization is claimed to be one of the most important causes of the global financial turmoil. The analysis does not say much about the formation of the crisis, but it indicates securitized countries allocate their risk better perhaps transferring their risk to nonsecuritizing countries- and thus they are able to respond to the crisis better than the countries without that technology during the crisis. Once the crisis proves to be systemic, it does not matter whether a country has securitization technology or not and everyone seems to be in trouble. There are also some limitations of the approach we use. To model securitization we used a dummy variable, representing securitization technology. The amount of securitization, however, can also make a difference. Too much securitization, for example, can be harmful for the economy, in the long-run or during crisis period. Another problem is that the use of dummy variable of the data, which is averaged over 4 years, leads to some measurement error by construction. We cannot differentiate between countries, which start securitization in the beginning of the period and in the end of the period. Furthermore, securitization is a rather new phenomenon, so its long-run impact may not be fully established yet. Data covering a longer period can provide more reliable robust results. Moreover, the analysis cannot explain how exactly securitization can influence economic growth other than increasing loan supply. More sophisticated theory is needed to explain this relationship. Future research about this issue should try to solve these problems with using more detailed data and modeling securitization in a more sophisticated way, which provides deeper explanation to its relationship with macroeconomy. Using population or the absolute size of financial institutions as an instrument to identify securitization seems to be a very promising approach, which can be used in the future research. 30

31 A. APPENDIX a. Summary and Regression Tables Table A. Countries in the Sample and First Securitization First Securitization First Securitization First Securitization Argentina 1995 Guyana - Paraguay ' - Australia' 1996 Haiti ' - Peru 1995 Austria ' 2000 Honduras - Philippines 1996 Bangladesh' - Iceland *' - Portugal ' 1998 Barbados * - India - Senegal - Belgium ' 1996 Ireland 1994 Sierra Leone ' - Bolivia' - Israel ' 2005 South Africa 1997 Brazil 1995 Italy 1995 Spain ' 1986 Canada 1995 Jamaica ' 2000 Sri Lanka - Chile 1997 Japan 1996 Sweden 2000 Colombia 1996 Kenya - Switzerland 1997 Congo, Dem. Rep.' - Korea, Republic of 1996 Syria ' - Costa Rica 2002 Liberia *' - Thailand ' 1996 Cyprus - Malaysia ' 1995 Togo - Denmark ' 2000 Malta ' - Trinidad &Tobago - Dominican Republic 1997 Mauritius - United Kingdom ' 1995 Ecuador - Mexico 1992 Uruguay - El Salvador 1997 Nepal ' - US 1985 Fiji ' - Netherlands ' 1996 Venezuela 1992 Finland 1995 New Zealand ' 1996 Zimbabwe * - France 1995 Niger - Germany* 1996 Norway *' - Ghana* ' - Pakistan 1997 Greece ' 2000 Panama ' 1997 Guatemala - Papua New Guinea ' - Notes: Countries with ' are not in M3 over GDP dataset. Countries with * not in the crisis regressions. (-) if no securitization is in the dataset. 31

32 Table B. Summary Statistics for the Panel Data VARIABLES Obs Mean Std. Dev. Min Max ln(gdp per capita) Domestic Credit over GDP Securitization Technology Inflation Total years of Education Initial GDP per capita Trade over GDP Government Share Rule of Law M3 over GDP Domestic Credit by Banking Sector ln(population) Absolute Latitude Table C. Summary Statistics for the Cross Sectional Crisis Data VARIABLES Obs. Mean Std. Dev. Min Max Ave. Growth Ave. Growth Ave. Growth Growth Growth Inflation Securitization Technology ln(gdp per capita) Education Trade over GDP Government Share Domestic Credit over GDP Domestic Credit by Banking Sector Rule of Law ln(population) ln(gdp) Absolute Latitude ln(biggest Bank Asset)

33 Table D. Pairwise Correlations of the Panel Data Variables lngdppc domcre sectech inf edutot inigdppc openk gov ruleoflaw m3ovgdp dombank lnpop abslat ln(gdp per capita) 1 Domestic Credit over GDP Securitization Technology Inflation Total years of Education Initial GDP per capita Trade over GDP Government Share Rule of Law M3 over GDP Domestic Credit by Banking Sector ln(population) Absolute Latitude Table E. Pairwise Correlations of the Crisis Data Variables crgrowth inf sectech lngdppc edu2005 openk gov domcre dombank realint ruleoflaw lnpop lngdp abslat Crisis Growth 1 Inflation Securitization Technology ln(gdp per capita) Education Trade over GDP Government Share Domestic Credit over GDP Domestic Credit by Banking Sector Real interest rate Rule of Law ln(population) ln(gdp) Absolute Latitude

34 VARIABLES Table 1 Dynamic Panel Estimations using GMM (1) (2) (3) (4) (5) (6) (7) (8) Basic Conditioning Set Full Conditioning Set Difference System Dif-2step Sys-2step Difference System Dif-2step Sys-2step Dependent Variable: ln(gdp per capita) Lagged ln(gdp per capita) *** *** * *** *** *** ** *** (0.000) (0.000) (0.066) (0.000) (0.000) (0.000) (0.016) (0.000) Domestic Credit over GDP * *** *** (0.062) (0.367) (0.276) (0.698) (0.002) (0.339) (0.001) (0.862) Securitization technology *** *** (0.893) (0.001) (0.808) (0.283) (0.836) (0.000) (0.753) (0.187) Inflation -0.01*** -0.01*** -0.01** -0.01*** -0.01*** -0.01*** -0.01*** -0.01*** (0.001) (0.000) (0.013) (0.000) (0.002) (0.000) (0.000) (0.000) Education * * * ** (0.080) (0.076) (0.135) (0.124) (0.152) (0.071) (0.142) (0.044) Initial GDP per capita (0.397) (0.808) (0.221) (0.943) Openness ** *** (0.036) (0.000) (0.170) (0.295) Government expenditures (0.382) (0.117) (0.938) (0.444) Observations Number of Countries No of Instruments Arellano-Bond AR(1) test p-value Arellano-Bond AR(2) test p-value Hansen/Sargent OIR Test p-value Notes: Robust P-values are reported in parentheses for the two-step estimators. *** p<0.01, ** p<0.05, * p<0.1. In all regressions time fixed effects are included as dummies. Inflation coefficient multiplied by 100 is reported. Reported OIR tests are evaluated according to Sargan statictics for one-step estimations and Hansen J statistics for two-step estimations. 34

35 VARIABLES Lagged ln(gdp per capita) Financial Development Indicator Securitization technology Inflation Education Initial GDP per capita Openness Government expenditures Institutions Indicator Table 2- Dynamic two-step system GMM estimator with different specifications (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Nonlinear liabilities Credit by Average Liquid Domestic 5 year System Nonlinear incl. by banks Non-linear Dom. Cre. Baseline Less Incl. Without w/o System instruments Institutions Fin. Dev. less over GDP banks (1985- Population Model institutions less instr. instr. less instr. over GDP 2004) *** *** *** *** 1.127*** 1.049*** 0.995*** 1.004*** *** *** 1.138*** 1.182*** (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.959) (0.718) (0.485) (0.534) (0.994) (0.716) (0.561) (0.561) (0.292) (0.252) (0.957) * * (0.187) (0.241) (0.948) (0.107) (0.265) (0.840) (0.565) (0.397) (0.054) (0.230) (0.066) (0.175) *** *** *** *** *** *** *** *** (0.000) (0.000) (0.000) (0.255) (0.000) (0.583) (0.785) (0.501) (0.000) (0.000) (0.000) (0.000) ** * (0.044) (0.175) (0.459) (0.707) (0.486) (0.053) (0.804) (0.214) (0.650) (0.112) (0.833) (0.911) ** ** (0.943) (0.998) (0.173) (0.979) (0.036) (0.575) (0.979) (0.610) (0.493) (0.594) (0.042) (0.273) * (0.295) (0.198) (0.420) (0.088) (0.330) (0.642) (0.269) (0.294) (0.131) (0.351) (0.233) (0.401) * (0.444) (0.583) (0.147) (0.882) (0.274) (0.655) (0.062) (0.685) (0.283) (0.952) (0.249) (0.686) ** 0.075* (0.030) (0.076) Observations Number of Countries No of Instruments Arellano-Bond AR(1) test Arellano-Bond AR(2) test Hansen OIR test Notes: Robust P-values are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1. In all regressions time fixed effects are included as dummies. Inflation coefficient multiplied by 10 is reported. Domestic Credit to Private Sector or Liquid Liabilities or Domestic Credit provided by banking sector over GDP are used as financial development indicator. Log-linear models are reported with 4 decimals, whereas log-log models are reported with 3 decimal precision. 35

36 Table 3- Pairwise Correlations of the Instruments Securitization Dummy Securitization Dummy 1 ln(biggest Bank Asset) ln(biggest Bank Asset) 0.680*** 1 Ln(GDP) Ln(GDP) 0.738*** 0.905*** 1 Ln(Population) Ln(Population) 0.374*** 0.350*** 0.633*** 1 Absolute Latitude Absolute Latitude 0.409*** 0.748*** 0.613*** Notes: *** p<0.01, ** p<0.05, * p<0.1 Table 4- First Stage Regressions of IV and Treatment Estimators VARIABLES IV GMM Without Institutions Institutions Treatment without Institutions Institutions ln(population) 0.159*** 0.151*** 3.878** 4.290** (0.001) (0.004) (0.025) (0.022) ln(biggest Bank's Assets) (0.811) (0.989) (0.109) (0.125) Inflation * (0.436) (0.525) (0.137) (0.072) ln(gdp per capita) 0.260*** 0.267*** 4.733** 6.246** (0.002) (0.002) (0.031) (0.034) Openness 0.002* 0.002* (0.092) (0.092) (0.365) (0.474) Government expenditures (0.626) (0.646) (0.103) (0.260) Education (0.586) (0.534) (0.182) (0.194) Domestic Credit over GDP (0.130) (0.292) (0.333) (0.176) Rule of Law (0.498) (0.137) Constant *** *** ** ** (0.000) (0.000) (0.014) (0.016) Number of Observations Joint significance of the model Joint significance of excluded instruments P-value Partial for excl. instruments Notes: The reported test statistics are chi-squared for probit regression and f statistics for IV regression. 36

37 Table 5. Securitization Technology in 2005 on Growth Performance in Selected Periods Average Growth in the years: Variable of interest: (1) (2) (3) (4) (5) (6) (7) (8) (9) OLS Basic Set OLS Full Set OLS Without Financial Dev. OLS including Institutions IV Basic Set IV Full Set IV Full Set incl. Inst. Treatment Treatment incl. Inst Securitization Technology Securitization Technology Securitization Technology 2007 Securitization Technology 2006 Securitization Technology (0.498) (0.541) (0.717) (0.519) (0.886) (0.883) (0.927) (0.871) (0.552) (0.316) (0.286) (0.389) (0.281) (0.986) (0.785) (0.778) (0.493) (0.27) (0.569) (0.612) (0.413) (0.634) (0.125) (0.134) (0.182) (0.193) (0.369) 1.662* 1.612* 1.816* 1.601* 3.098* 3.508** 3.391* 2.541*** 2.083** (0.091) (0.096) (0.051) (0.096) (0.061) (0.041) (0.061) (0.01) (0.025) (0.576) (0.651) (0.428) (0.683) (0.218) (0.228) (0.389) (0.178) (0.204) Notes: Robust p-values in parentheses. *** p<0.01, ** p<0.05, * p<0.1 37

38 Table 6 OLS, IV and Treatment Regressions for the economic growth in the Beginning of recent financial crisis (Year 2007) VARIABLES (1) (2) (3) (4) (5) (6) (7) (8) (9) OLS Full OLS Without IV Basic IV Full IV Full Set Treatment Set Financial Dev. Set Set incl. Inst. OLS Basic Set OLS including Institutions Treatment incl. Inst. Securitization Technology 1.662* 1.612* 1.816* 1.601* 3.098* 3.508** 3.391* 2.541*** 2.083** (0.091) (0.096) (0.051) (0.096) (0.061) (0.041) (0.061) (0.01) (0.025) Domestic Credit over GDP (0.135) (0.147) (0.217) (0.292) (0.37) (0.376) (0.206) (0.248) Inflation (0.38) (0.339) (0.221) (0.367) (0.324) (0.319) (0.29) (0.277) (0.282) ln(gdp per capita) (0.284) (0.343) (0.789) (0.287) (0.626) (0.877) (0.879) (0.596) (0.464) Education * * * (0.288) (0.208) (0.19) (0.228) (0.181) (0.087) (0.111) (0.055) (0.062) Openness (0.536) (0.585) (0.543) (0.299) (0.294) (0.419) (0.464) Government expenditures * * (0.131) (0.113) (0.128) (0.177) (0.136) (0.083) (0.077) Rule of Law (0.875) (0.99) (0.899) Constant * * (0.755) (0.192) (0.055) (0.279) (0.45) (0.073) (0.189) (0.145) (0.294) Number of Observations Adjusted Rsquared Significance of the model (p-value) F-statistic of excluded instruments (1 st Stage) OIR Test p-value Endogeneity Test pvalue Underidentification Test Notes: Chi-squared statistic is reported for treatment regressions, instead of f-statistics of the first stage regressions. For treatment regressions endogeneity test is a Likelihood ratio test of independent equations with the null hypothesis of exogeneity of endogenous variable. Robust p-values in parentheses. *** p<0.01, ** p<0.05, * p<0.1 38

39 b. Econometric Derivation Let the cross-sectional equation be as follows:, (1) where represents individual country, which has just one observation, stands for the average crisis real per capita GDP growth, includes the exogenous control variables, stands for the error terms and is the possibly endogenous securitization variable. If the model suffers from omitted variable bias and endogeneity an unobserved variable for example - has to be introduced to represent the true model:, (2) And an auxiliary regression for the association between and :, (3) So substituting equation (3) in to equation (2), the OLS estimator estimates:, (4) which can also be rewritten as:, (5) Unless or the OLS estimator will be biased: 39

40 Figure 1 Cross-Border Trading of Securities Notes: In billion US$. Taken from Securities Industry and Financial Markets Association (SIFMA). (A) The amount for 2004 has been annualized from nine month data. Source: US Treasury 40