Life-cycle effects in small business finance

This paper studies the life-cycle profiles of small firms’ cost and use of credit using a panel of Finnish firms. The choice of method matters for the conclusions drawn about the relationship between firm age and financing costs; the cross-sectional age profiles of financing costs are hump-shaped and consistent with hold-up theories, whereas methods that control for cohort fixed effects demonstrate that the financing costs decrease monotonically as the firms mature. The life-cycle profiles of the use of credit also indicate that firms are more dependent on financial intermediaries in the early periods of their lives. Furthermore, the cohorts born during recessions pay higher financing costs and use smaller amounts of bank loans, even after their creditworthiness is controlled for. The recession cohort effect appears to be more related to the experience of starting-up the firm in the recession than to the CEOs growing up in a recession during their early adulthood.

M A N U S C R I P T 4 possible to identify all these effects in the same model without some restrictions (see, e.g., Hall et al. 2005). The existence of unobserved firm-specific heterogeneity results in another problem; firm fixed effects remove the cohort effects but do not eliminate the problem of identifying the age and time effects simultaneously (Hall et al. 2005). All these issues would have to be tackled to identify the life-cycle profiles of the cost and use of credit. This study takes several steps in this direction and supplements the scarce corporate finance literature on this largely unexplored issue.
Whether there are significant time and cohort effects in the cost and use of small business loans is a policy-relevant issue that would benefit from further empirical research.
Time effects could arise from fluctuations in the macroeconomic and financial environment, while there remains a question whether such effects would affect each age group equally. Holmström and Tirole (1997) predict that poorly capitalized firms, such as start-ups, are most hurt by credit tightening. The empirical evidence indeed suggests that bank-dependent firms are most affected by the tightening of monetary policy and by negative shocks faced by the banking sector (e.g., Gertler and Gilchrist 1994;Kroszner et al. 2007;Dell'Ariccia et al. 2008;Khwaja and Mian 2008;Chava and Purnanandam 2011). The literature on the micro foundations of credit cycles also suggests that financial market imperfections could have significant real effects; shocks to collateral values and their interaction with credit limits could affect borrower net worth and result in large and persistent fluctuations in the output and asset prices (Bernanke and Gertler 1989;Kiyotaki and Moore 1997).
Where could cohort-specific effects arise in small business finance? Building on the analogue from the labor economics literature (see, e.g., Kahn 2010;Oyer 2006), firms established during weak economic times could be perceived as a different quality than otherwise identical firms born during stronger times. The corporate finance literature also A C C E P T E D M A N U S C R I P T 5 highlights the adverse effects suffered by bank-dependent borrowers who lose their banking relationships or become otherwise credit-constrained during recessions and financial crises (e.g., Slovin et al. 1993;Kashyap et al. 1994;Kroszner et al. 2007). Corporate managers who started their businesses during a recession may also have less faith in financial markets. Graham and Narasimham (2004) find that publicly listed U.S. firms that experienced the U.S.
Great Depression use less leverage in the 1940s than other firms.  suggest that the CEOs of publicly listed U.S. firms who grew up during the Great Depression lean excessively towards internal finance. Schoar and Zuo (2016) observe in their sample of publicly listed U.S. firms that CEOs who started their careers during recessions use more conservative management approaches, including the lower use of leverage. Malmendier and Nagel (2011) provide complementary evidence suggesting that macroeconomic shocks faced earlier in life could affect the financial risk taking of individuals (see also Knüpfer, Rantapuska and Sarvimäki 2016). Giuliano and Spilimbergo (2014) study the impressionable years hypothesis of social psychology, which suggests that economic and political beliefs are formed during the early adulthood and change only slowly after this critical age. They indeed find that the experience of a recession during the critical years of early adulthood has a longlasting effect on the beliefs and preferences of individuals.
The development of the financial markets and improvements in the informational environment could also be potential sources of cohort effects. For instance, the younger cohorts may benefit from improved bank screening technologies, such as credit scoring and the better availability of high-quality credit information (cf. Petersen and Rajan 2002). The improved availability of borrower-specific information from credit bureaus and credit rating agencies could reduce adverse selection, lower the informational rents banks can extract from borrowers, and improve borrower discipline Pagano 1993, 2000; A C C E P T E D M A N U S C R I P T 6 Pagano 19976 Pagano , 2000. The development of financial markets would generally predict the availability of lower cost external finance for firms (Rajan and Zingales 1998).
This study analyzes the life-cycle profiles of financing costs and the use of credit using a large register-based panel of Finnish firms. This new dataset covers the period 1999-2013 and provides a longer and more recent study period than used in the previous studies. The firms in the sample are on average very smallmost of them are micro firmsand thus provide an effective testing ground for the theories of asymmetric information. In addition, the Finnish financial system is bank-based, an institutional setup that provides a good comparison point to the studies on more market-orientated financial systems, including the U.S system. The study also differentiates itself from the previous literature by paying careful attention to disentangling age, period, and cohort effects. This identification problem has been largely ignored in the previous corporate finance studies, which have relied on crosssectional data and short panels. The current study utilizes a number of alternative methods to overcome the identification problem. In particular, the life-cycle profiles estimated from the cross-sectional data and models are compared to more appropriate methods that control for cohort or firm fixed effects. An important feature of the current dataset is that it also includes the widely used commercial credit scores of the firms in the dataset. Thus, the observed creditworthiness of the firms can be controlled in the analysis among other key variables.
The findings of the paper suggest that the choice of the method affects the conclusions drawn about the relationship between the firm age and the financing costs. The crosssectional age profiles of financing costs are hump-shaped and consistent with hold-up theories. In contrast, the regressions that control for the cohort or firm fixed effects suggest that the financing costs decrease monotonically as the firms mature, in line with the prediction of Diamond (1989). The findings suggest that these differences in the age profiles A C C E P T E D M A N U S C R I P T 7 relate to cohort effects. After controlling for the cohort or firm fixed effects that are essential in capturing the firm-specific heterogeneity, the baseline findings point more towards the reputational theories than towards the hold-up theories as an explanation for the life-cycle profiles of financing costs. Moreover, the age profiles of the use of credit indicate that firms are more dependent on financial intermediaries in the early periods of their lives.
A few main findings are made about the cohort effects. First, the younger cohorts face lower costs of credit than the older cohorts. While the source of this cohort effect was not formally tested, the longer-term trend of decreasing cohort-specific financing costs would generally appear to be consistent with the hypothesis about the improvements in the financial system and the information environment. Second, the findings suggest that cohorts born in recessions, particularly the Finnish Great Depression and the banking crisis of the 1990s and the more recent international financial crisis, face higher financing costs and use a smaller amount of bank loans in a persistent fashion. This effect is robust to controlling for the observed creditworthiness of the firms with commercial credit scores. The recession-born firm effect is larger for younger CEOs, in line with the prediction that macroeconomic shocks have a more significant effect on young individuals. However, the recession cohort effect appears to be more related to the experience of starting-up the firm in the recession than to the CEOs growing up in a recession during their early adulthood. Overall, these findings suggest that recessions and periods of financial instability could have a lasting impact on the perceived riskiness of the firms and their use of external finance in the future.
The remainder of the paper is organized as follows: Section 2 presents the dataset.
Section 3 provides an overview of the empirical methods. Sections 4-6 present the empirical results, and Section 7 concludes.

Data sources
The dataset used in this study consists of a register-based panel of Finnish firms from the period 1999-2013. The panel design is unbalanced and therefore allows firms to enter into and exit from the sample (e.g., because of bankruptcy) during the study horizon. The dataset consists of financial statements and related data compiled from official sources by Asiakastieto ltd, an information provider of firm and credit data in Finland. The financial statement data originate from the Finnish Trade Register, an official register of Finnish firms.
The dataset also contains the commercial credit scores and associated credit ratings of the firms computed by Asiakastieto. Several macroeconomic variables were matched to the dataset, including the aggregate country-level unemployment rates, GDP growth, house prices, and consumer prices, which were obtained from the databases of Statistics Finland.
The Finnish government bond yields were obtained from the database of the Bank of Finland.
The estimation sample is restricted to non-farm and non-financial corporations. This restriction helps avoid issues such as differences in the accounting practices from affecting the results. 2 The sample is restricted to small businesses by removing firm-year observations exceeding the EU-level small and medium-sized firm thresholds in terms of employment 2 The majority of the observations (i.e., about 98%) in the original data belong to corporations. The following industries are dropped from the sample: Agriculture, forestry and fishing; Financial and insurance activities; Electricity, gas, steam and air conditioning supply; Water supply; sewerage, waste management and remediation activities; Activities of membership organizations; Operation of dwellings and residential real estate; Management activities of holding companies; Public administration and defense; compulsory social security; Activities of extraterritorial organizations and bodies; Industry unknown.
A C C E P T E D M A N U S C R I P T 9 (less than 250 persons), net sales (50 million euros) or total assets (43 million euros). The estimation sample concentrates on the cohorts born between the periods 1960-2012. Because the informational asymmetries are likely to be the most relevant for relatively young firms, the firms older than 40 years are dropped from the sample. This helps to control for the additional noise caused by the relatively few firm observations among the older firms in the age distribution. 3 Firm observations with negative total assets are dropped from the sample.
The CEO characteristics focus on working-age CEOs younger (or not older) than 74 years.

Variable definitions
The study analyzes the life-cycle profiles of small firms' cost and use of credit. These measures are computed from the financial statement data. Financing costs are measured as financial expenses at period t divided by the average outstanding interest-bearing debt between periods t-1 and t. 4 Bank debt is a ratio of outstanding debt from financial institutions scaled by total assets at period t. 5 3 In the robustness tests, start-up firms born potentially because of mergers or spinoffs were removed from the sample based on a mechanical rule of dropping firms with net sales larger than or equal to the 99 th percentile of the start-up firm distribution, which includes the firms of age one year or less.
This did not materially change the estimated life-cycle profiles. 4 The financing cost measure is computed only for the observations in which the financial expenses are positive and non-zero. Indeed, it would be conceptually problematic to evaluate the effects of firm age on the financing costs if the firms have no financial expenses (see, e.g., Hyytinen and Pajarinen 2007). 5 For convenience, loans obtained from financial institutions are referred to as bank financing in the text. This seems a reasonable shortcut definition because bank loans represent a major fraction of The previous empirical banking literature suggests that age, size, and type of business are three key determinants of firms that rely on banking relationships: In particular, younger, smaller, and less transparent firms that have more intangible assets are more difficult to screen successfully (see, e.g., Freixas and Rochet 2008, 105). Firm age is a key variable of interest in this study. It is also a measure that has been considered as a good proxy for informational asymmetries in the previous corporate finance literature (e.g., Beck et al. 2006;Hyytinen and Pajarinen 2008;Hyytinen and Väänänen 2006). The firm age is calculated by subtracting the year of birth from the current year. The year of birth is defined as the year the firm was registered in the Finnish Trade Register. The current year refers to the year of the financial statement. 6 Firm size is proxied with ln(Sales), which is a natural logarithm of net sales at t-1. Tangibility is a proxy for collateralizable assets and defined as a ratio of fixed to total assets at t-1. Profitability is measured as earnings before interest, taxes, depreciation, and amortization divided by total assets at t-1. Credit score measures the observed creditworthiness of the firms (i.e., the probability of default) at t-1 computed by Asiakastieto.
This commercial credit score is defined at the interval 3-100, where low values indicate high creditworthiness and high values indicate low creditworthiness. The credit scores for 2006 are not available in the data because of changes in the dataset; therefore, values from the previous period are used in that particular year. If the values from the previous period are not the financing obtained from financial institutions in Finland (see, e.g., Business Financing Survey 2009). Note that this measure can generally also cover financing from other financial institutions, including special financing institutions. Furthermore, the non-use of bank loans does not necessarily indicate supply side constraints, simply because some firms may have no demand for bank loans. 6 In some instances, the method of calculating the firm age resulted in negative ages, in which case, the observation was dropped.
A C C E P T E D M A N U S C R I P T 11 available, the values from the following period are used instead to avoid losing startup firms. 7 Recession-born is a dummy that takes a value equal to one if the firm is born during a period of negative real GDP growth (years 1991, 1992, 1993, 2009, and 2012) and zero otherwise.
The information on defaults and bankruptcies is available in the data since 2005. Default measures the number of missed debt payments at t-1. Bankruptcy is an indicator for bankruptcy applications made by the debtor or creditors at t.
The entrepreneur-specific information is obtained from the social security numbers of the CEOs of the firms. The entrepreneur's age is controlled with ln(CEO age), which is a natural logarithm of the age of the CEO. It is computed as a difference between the current year and the birth year obtained from the social security number. The gender of the entrepreneur is controlled with an indicator Female, taking a value one for female CEOs, and zero otherwise. The gender is obtained from the first three numbers of the final part of the social security number, where odd and even numbers define male and female, respectively.
Recession-grown CEO is an indicator for CEOs who experienced a recession during the impressionable years of their early adulthood (i.e., between 18 and 25 years). 8 The social security numbers are available in the data since 2003, although not for all entrepreneurs. 9 7 The transition matrices of the credit ratings and the serial correlations of the credit scores indicate that there is considerable persistence in the creditworthiness of the firms. This suggests that the above approach provides a reasonable solution for the missing data issue. 8 Since the study focuses on working-age CEO cohorts whose early adulthood took place in the postwar period, the recession years for this measure are the same as for the firm-level recession-born indicator (i.e., the first negative post-war GDP growth took place in 1991; see, e.g., Hjerppe 1989Hjerppe , 2010. That is, the CEO cohorts born in 1966-1975 and 1984-1995 are defined as recession-grown The industry-specific characteristics are controlled using two-digit-level industry dummies. There was a change in the industry classifications in 2008 that affects the classifications used in the data. Specifically, firms that existed in the earlier periods but did not exist anymore in 2008 were classified using the previous standard industrial classification (SIC) version. In the following analysis, industries are classified using SIC 2008 when available and using SIC 2002 in the other instances. A dummy for the firms classified using SIC 2002 is included in the regressions to take into account the scale differences in the different versions of the classifications. Regional dummies measured at the two-digit zipcode level based on the firms' addresses are also included in the regressions. Petersen and Rajan (1995) argue that the life-cycle profiles of financing costs could differ between competitive and non-competitive markets because monopolistic banks may be able to subsidize younger firms. It is also worth taking into account that firms in certain areas of the country are eligible for more government subsidies than others, which could be reflected in the financing costs. The regional fixed effects provide a way to control for the fixed regional characteristics, including these local credit market characteristics.
The macroeconomic control variables used in some of the specifications are defined as follows: Unemployment measures the country-level unemployment rates at period t. Term spread measures the difference in the yields of the Finnish government bonds of the maturity CEOs for the Finnish Great Depression and the more recent recession that accompanied the international financial crisis, respectively, in line with the impressionable years hypothesis. 9 One potential reason for missing social security numbers (besides not having a formal CEO) could be foreign CEOs, who do not have a Finnish social security number. In a few thousand instances the social security numbers in the data only have a birth date and not the final part defining the gender.
A C C E P T E D M A N U S C R I P T 13 of ten and five years at period t, respectively. House prices growth measures the growth of the house prices index defined in natural logarithms between periods t-1 and t. CPI growth measures the growth of consumer prices defined in natural logarithms between periods t-1 and t. GDP growth measures the growth of gross domestic product defined in natural logarithms between periods t-1 and t. The time fixed effects used in other panel specifications provide an alternative way to control for the macroeconomic conditions of the period.
To avoid issues related to large outliers, some of the key variables are trimmed or winsorized as follows: Financing costs are trimmed at the 5 th and 95 th percentile of the distribution because of the large outliers typical for this kind of data. 10 In this type of measure, outliers could arise, for example, because of large changes in the amount of outstanding debt near the end of the period that are not reflected in the financial expenses accrued over the year (see, e.g., Bernhardsen and Larsen 2003;Kim et al. 2012). 11 Bank debt is trimmed at the values below zero and above one. Profitability and Tangibility are winsorized at the 1st and 10 This drops out (erroneous) negative values, some unrealistically small (but positive) values and very large values. Note that this trimming does not remove any zero values, which are already removed in the process of forming the variable. The trimming percentiles as such are more conservative than the ones used by Kim et al. (2012), who remove observations outside the 10 th and 90 th percentiles. 11 In the robustness tests, Financing costs were alternatively trimmed at the 1 st and 99 th percentile of the distribution, which resulted in similar albeit noisier profiles of financing costs. However, this alternative trimming strategy cannot completely deal with some unreasonably large values despite setting the value of interest-bearing debt to missing when its value was non-positive (zero) in either of the two subsequent periods used in the computation of the measure. It remains possible that the financial expenses could reflect foreign exchange losses, while this issue is probably less acute among domestically orientated small businesses.
A C C E P T E D M A N U S C R I P T 14 99th percentiles to ensure that outliers in the control variables are not confounding the results.
These control variables are winsorized rather than trimmed to avoid any unnecessary loss of observations.

Descriptive statistics
The panel statistics are provided in table A1 in the appendix. They indicate that the number of firms covered by the data has increased over the years. The descriptive statistics are The average ratio of fixed assets to total assets in the balance sheet is 0.255. The profitability measure indicates that the average return on assets before interest, taxes, depreciation and amortization is 17.4%. The average credit score suggests that the firms are on average rated as A+ (i.e., "Satisfactory+") on the seven-step rating scale 12 The financing cost sample has a lower number of observations than the bank debt sample particularly for the reason that many firms do not report positive (non-zero) financial expenses, indicating that they tend to rely on other forms of finance than interest-bearing debt.

Identification of age, period, and cohort effects
The fundamental problem of identifying age, period, and cohort effects is well known in the economics literature, but the issue has been left almost unaddressed in the corporate finance literature. 13 The identification problem is stated as follows: because there is a linear 13 Sakai et al. (2010) provide a short discussion about the issue. Otherwise, the identification problem has been largely ignored in the empirical corporate finance research. Notably, Petersen and Rajan (1995, 419) claim that they can identify the age effects in cross-sectional data under certain assumptions, namely the stationarity of the survival process of firms.
A C C E P T E D M A N U S C R I P T 17 relationship between the age, period, and cohort effects (based on the identity: ), it is not possible to identify all of them in the same model without some restrictions (see, e.g., Hall et al. 2005). This makes it difficult to evaluate the life-cycle profiles of small firms' cost and use of credit. Indeed, the modeling and identification of such relationships is complicated for an obvious reason: it is impossible to observe two firms (or entrepreneurs) at the same point in time that have the same age but who are born at different periods (cf. Hall et al. 2005). This is problematic because the otherwise identical firms that belong to different cohorts could face very different economic environments. This problem could be acute, for instance, if the stage of the business cycle during which the firm is born has persistent effects on the firm for the rest of the periods. The identification problem is equally complicated if the younger cohorts face fundamentally different financial environments than the older cohorts. Such cohort-specific differences could arise because of certain factors, including differences in the availability of credit information, developments in bank screening methods and general developments in the financial system.
The previous economic literature suggests several solutions to the identification problem in various other contexts. Deaton and Paxson (1992) and Attanasio (1998) identify the life-cycle effects as follows: they use a polynomial of age or age dummies, together with cohort effects, and normalize the time dummies to sum to zero and to be orthogonal to a linear time trend (see also Deaton 1997). Hall et al. (2005) analyze the identification problem related to the life-cycle effects in another context and discuss various approaches for addressing the issue, such as testing which effects are present and constraining some of the cohort, time or age dummies to have equal effects in the same dimension. They also highlight the problems that arise in the presence of unobserved firm-specific heterogeneity; for example, including firm fixed effects removes the cohort effects and renders some of the cohort-based approaches unavailable. However, the firm fixed effects do not eliminate the problem of identifying the age and period effects simultaneously (cf. Hall et al. 2005).
In the context of corporate finance, Sakai et al. (2010) argue that the empirical approach suggested by Deaton and Paxton (1992), Attanasio (1998) and Deaton (1997) could result in the unstable age profiles of financing costs in short panels. They, in turn, focus on analyzing the slope of the age profile of financing costs and control the year effects using a prime lending rate. However, their study does not analyze other aspects of life-cycle effects in small business finance, such as the use of credit, nor does it provide measurements of the magnitude of the age or cohort effects. The current study aims to overcome these shortages by building on the alternative methods suggested in the earlier cohort literature.

Estimation of life-cycle profiles
The empirical analysis of the study proceeds as follows: First, the age profiles of the cost and use of credit are estimated from yearly cross-sections. This method has been a common practice in the previous corporate finance studies that have often used cross-sectional data because of the limitations of the survey datasets (see, e.g., Rajan 1994, 1995).
This study investigates whether the cross-sectional age profiles are stable over time and whether there are significant biases in the cross-sectional estimates in comparison to the estimates obtained from other methods. This analysis should help to consider the relevance of cross-sectional age profiles in comparison to the profiles obtained from more appropriate cohort methods.
Second, several alternative identification assumptions are considered in the analysis of the life-cycle profiles, building on the earlier suggested cohort methods and full panel dataset.
Consider a general age, period, and cohort effects model (adopted from Hall et al. 2005) assuming the additively separable nature of the equation defined as follows: where measures the cost or use of credit, is a constant, is the cohort effect, is the period effect, is the age effect, is a vector of control variables, is an error term and ; ; t and index firms, cohorts, time periods, and ages, respectively. The estimation of the above model requires that the indicator variables are estimated relative to their reference values. This is implemented by imposing nullity on the coefficients , , and that measure the first cohort, period, and age, respectively. This, however, does not remove the collinearity between the age, period, and cohort effects because the variables in the equation are not linearly independent (Hall et al. 2005). Consider the following modifications on equation (1) that allow the identification of the model based on the several alternative identification assumptions: In the baseline case, the models that contain age dummies (and controls) together with either time or cohort dummies are compared to each other (cf. Heathcote et al. 2005). This comparison provides a useful starting point for the analysis and evaluates the relevance of the time versus cohort effects. These first two models are defined in more detail as follows: The first model includes time dummies but leaves out all the cohort dummies (i.e., is dropped from equation (1)). That is, this model assumes that there are no cohort effects and treats the dataset as a pooled cross-section. The time dummies included in the regressions control for the period-specific effects that might arise, e.g., because of macroeconomic or financial factors such as the market level of interest rates or changes in the supply of credit.
The second model includes cohort dummies but no time dummies (i.e., is dropped from equation (1)). This model accounts for the possibility of the existence of cohort effects but assumes away any time effects, in contrast to the earlier model. The comparison between these first two models provides an informal evaluation about whether the time or cohort dimension is a more important factor influencing the age profiles of the cost and use of credit.
However, these baseline models could as such provide an unsatisfactory solution to the identification problem of the age, period, and cohort effects; that is, failure to control one of these distinct dimensions (period or cohort) could result in spurious findings (Mason et al. 1973). Because of this, the baseline models are compared to other models that aim to identify age, period, and cohort effects in several alternative ways.
In the third model, the identification is achieved by aggregating the cohorts into groups by grouping the cohorts at the four-year level. The grouping of single-year cohorts in this way overcomes the fundamental identification problem (see, e.g., Hall et al. 2005;Levin and Stephan 1991). Hall et al. (2005) note that the grouping of cohorts is equivalent to obtaining the identification of the age effect by comparing closely adjacent ages to each other and assuming that they come from the same cohort. They note, however, that the grouping of cohorts at multi-year intervals may be a less satisfactory solution than utilizing a priori information about the cohorts or time periods in the identification of the models, as suggested by Rodgers (1982). In the current study, special attention is paid to make sure that the cohort groups are natural and match some key macroeconomic and financial regimes observed, for instance, during the Finnish Great Depression and the banking crisis of the 1990s.
In the fourth model, the time effects are controlled by replacing the time dummies with macroeconomic control variables following the suggestion of Rodgers (1982) (see also Hall et al. 2005;Gourinchas and Parker 2002). Specifically, Rodgers (1982)  In the fifth model, the identification is obtained by constraining two time dummy coefficients equal to each other. This approach builds on the suggestion of Mason et al. (1973), who note that it is possible to identify the three sets of dummy variables for age, period, and cohort by setting two coefficients equal to each other in the same dimension (see also Hall et al. 2005). In the current study, this approach is implemented by dropping both the first and last time dummies from the model (i.e., setting both and to zero in equation (1)). This allows for the inclusion of the single-year cohort dummies into the model. Recall that the first cohort dummy is dropped from the model to avoid the dummy variable trap because of the constant term.
In the sixth model, the cohort effects are replaced with firm fixed effects in equation (1). The identification is obtained by setting two time dummies equal to each other.
As was done earlier, the first and last time dummies are dropped from the model, which allows the identification of the model. 14 In each specification, the standard errors of the panel data models are adjusted for the firm-level clustering. The cross-sectional regressions based on yearly cross-sections use heteroskedasticity robust standard errors.

Cross-sectional analysis
The  variables are defined as follows: Age, modeled as a third-order polynomial, is the age of the firm defined as the years since the initial incorporation at t. Ln(Sales) is a natural logarithm of net sales at t-1. Tangibility is a ratio of fixed to total assets at t-1. Profitability is EBITDA divided by total assets at t-1. Credit score measures the observed creditworthiness of the firms (i.e., the probability of default) at the scale 3-100 (scaled by dividing by 100), where higher values mean lower creditworthiness. The models also include industry dummies measured at the two-digit level, regional dummies measured at the two-digit zip-code level, and a constant. N is the number of observations.

Cohort analyses
The summary of the models that analyze the life-cycle profiles of financing costs using the full dataset and alternative identification assumptions is provided in table 3. Firm age is modeled using dummies for each age in each specification. Note that the time fixed effects and macro controls absorb the overall level of market interest rates. The fitted age profiles obtained from the models are presented in figure 2. The pooled panel model (model 1) that controls the time fixed effects provides a hump-shaped age profile, which is very similar to the ones observed in the cross-sectional data. Model 2 replaces the time fixed effects with cohort fixed effects, which results in a downward-sloping age profile. Both models 1 and 2 overcome the fundamental identification problem by assuming away one of the dimensions, i.e., time or cohort effects. However, this assumption could result in biased findings if the ignored distinct dimension remains important for the age profiles.
The next models address the identification problem in the following alternative ways: The control variable estimates seem sensible and provide statistically highly significant findings in most cases. Larger firms pay lower financing costs. Profitability, however, shows positive and negative relationship with financing costs in specifications (1)- (5) and (6), respectively. Firms with more tangible assets in their balance sheet face lower financing costs.
Firms with lower credit quality, as indicated by their credit scores, pay more for their credit. Financing costs is financial expenses divided by the average interest-bearing debt between t and t-1. The independent variables are defined as follows: Age, modeled using dummies for each age, is the age of the firm defined as the years since the initial incorporation at t. Ln(Sales) is a natural logarithm of net sales at t-1. Tangibility is a ratio of fixed to total assets at t-1. Profitability is EBITDA divided by total assets at t-1. Credit score measures the observed creditworthiness of the firms (i.e., the probability of default) at the scale 3-100 (scaled by dividing by 100), where higher values mean lower creditworthiness. All the models include a constant. The table also reports whether the firm, cohort,

A C C E P T E D M
A N U S C R I P T 28 industry, region, and time fixed effects, and macro controls, are included in the models. NT is the number of firm-year observations. Rho measures the intra-class error correlation. R2 stands for R-squared. Standard errors clustered at the firm level are reported in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01. In addition, two-digit-level industry dummies and two-digit zip-code-level regional dummies are included in specifications 1-5. The models with the following additional controls are estimated: 1) time fixed effects 2) cohort fixed effects 3) cohort fixed effects (birth years grouped at the four-year level) and time fixed effects 4) cohort fixed effects and macro controls 5) cohort and time fixed effects 6) firm and time fixed effects. The two-period hold-up models of Sharpe (1990), Rajan (1992) and von Thadden (2004) focus on banks' private information and do not concentrate explicitly on firm age. Ioannidou and Ongena (2010) also suggest that the hold-up problem could arise each time after the firms switch banks.
However, concentration on the firm age rather than other proxies of banks' private information (for example, relationship length) should make little difference in the current context. First, firm age is a well-reasoned proxy for asymmetric information (see, e.g., Hyytinen and Pajarinen 2008). Second, even an assumption that the firms would on average borrow only from one bank does not seem unreasonable. Niskanen and Niskanen (2000) utilize cross-sectional Finnish survey data and provide evidence that the average number of firms' banking relationships (including non-borrowers) is 0.85.
The recent survey results support the view that the majority of the firms have few, and in many cases one, bank relationships. Almost 80% of the micro firms and about 50% of the small firms that responded to the survey have only one main lending bank (Business Financing Survey 2012).

A C C E P T E D M
A N U S C R I P T 32 firm age and financing costs. 21 Finally, while the baseline results are perhaps more supportive of the predictions of Diamond (1989) compared to the hold-up theories, the findings from the more recent subsamples are not quite as clear-cut in distinguishing between the theories. 22 The comparison between the age profiles obtained from the cross-sectional models and the models that control for cohort or firm fixed effects suggests that the differences between the profiles are related to cohort effects. The following analysis examines the cohort effects The differences between the predictions from the alternative models are not economically trivial.
The spread between the financing costs of one-and ten-year-old firms is about -0.5 and 0.8 percentage points in models (1) and (4), respectively. The mean interest-bearing debt of about 385 000 euros suggests that one-year-old firms pay about 2000 euros lower and 3000 euros higher financing costs per annum than ten-year-old in models (1) and (4), respectively.

22
The Finnish banking sector is highly concentrated having three major banking groups. Both the reputational and hold-up theories could provide reasonable descriptions of the markets. The models that control for the cohort or firm fixed effects, which are essential in capturing the unobserved firm heterogeneity, point more in favor of reputational theories, even if the recent sample periods suggest a less clear-cut distinction. It remains an issue worth considering whether changes in the banking environment in more recent periods could explain some differences related to the sample periods. 23 The alternative (unreported) cohort profiles, including those averaged over all firm ages and using macro controls instead of time dummies, provide similarly shaped profiles.  The model includes firm age dummies, single-year cohort dummies and the following controls: ln(Sales), Tangibility, Profitability, Credit score, two-digit-level industry dummies, two-digit zip-code-level regional dummies and time dummies.

A C C E P T E D M A N U S C R I P T
While this analysis does not formally analyze the source of these cohort effects, the previous literature suggests some potential explanations for the findings. The earlier international literature, including a study by Jappelli and Pagano (2000), suggests that the availability of

A C C E P T E D M A N U S C R I P T
34 borrower-specific credit information has improved over time because of the birth of credit bureaus and credit rating agencies. Such an improved information environment could generally reduce the adverse selection, lower the informational rents banks can extract from borrowers, and improve borrower discipline Pagano 1993, 2000;Pagano 1997, 2000). Moreover, in the U.S. context, Petersen and Rajan (2002) suggest that in the banking sector, technological innovations, such as credit scoring, have improved the availability of credit for more distant firms. In general, the financial development would reduce the cost of the external finance available to firms (Rajan and Zingales 1998). These predictions from the previous literature appear to be consistent with the observed trend of decreasing cohort-specific financing costs. 24

Cross-sectional analysis
The cross-sectional estimates of the life-cycle profiles of bank debt, scaled by total assets, are show themselves as cohort effects, say, in the form of sticky old credit terms (e.g., higher interest rates).

A C C E P T E D M A N U S C R I P T
35 indicate that the age profiles of bank debt are s-shaped in the earlier periods, and the downward-sloping relationship is more pronounced in the later years.  1950 0.1932 0.1900 0.1943 The table shows the estimates for the bank debt obtained from the separate yearly cross-sections from the 1999-2013 period. The dependent variable Bank debt is a ratio of outstanding loans from financial institutions divided by total assets at t. The independent variables are defined as follows: Age, modeled as a third-order polynomial, is the age of the firm defined as the years since the initial incorporation at t.
Ln(Sales) is a natural logarithm of net sales at t-1. Tangibility is a ratio of fixed to total assets at t-1. Profitability is EBITDA divided by total assets at t-1. Credit score measures the observed creditworthiness of the firms (i.e., the probability of default) at the scale 3-100 (scaled by dividing by 100), where higher values mean lower creditworthiness. The models also include industry dummies measured at the two-digit level, regional dummies measured at the two-digit zip-code level, and a constant. N is the number of observations. R2 stands for R-squared.

Cohort analyses
The The control variable estimates are in line with the expectations. Larger firms and firms with more tangible assets use more bank debt. More-profitable firms and firms of higher observed creditworthiness use less bank debt. The findings seem consistent with the hypothesis that borrowers with lower credit ratings are more dependent on the monitoring provided by banks, as predicted by Diamond (1991). Bank debt is a ratio of outstanding loans from financial institutions divided by total assets at t. The independent variables are defined as follows: Age, modeled using dummies for each age, is the age of the firm defined as the years since the initial incorporation at t. Ln(Sales) is a natural logarithm of net sales at t-1. Tangibility is a ratio of fixed to total assets at t-1. Profitability is EBITDA divided by total assets at t-1.
Credit score measures the observed creditworthiness of the firms (i.e., the probability of default) at the scale 3-100 (scaled by dividing by 100), where higher values mean lower creditworthiness. All the models include a constant. The table also reports whether the firm, cohort, industry, region and time fixed effects, and macro controls, are included in the models. NT is the number of firm-year observations. Rho measures the intra-class error correlation. R2 stands for R-squared. Standard errors clustered at the firm level are reported in parentheses: * p < 0.10, ** p < 0.05, *** p < 0.01.  two-digit-level industry dummies, two-digit zip-code-level regional dummies and time dummies.

Recession cohorts
This section provides a further evaluation of the sources of the cohort effects. The following analysis studies whether the firm cohorts born or the CEO cohorts grown during severe recessions show persistent differences in their costs and use of credit. The focus of the analysis is on the two alternative recession cohort indicators. The first measure is the recession-born dummy, which takes a value equal to one if the firm was born during recession, and zero otherwise. The second measure is the recession-grown CEO dummy, which takes a value equal to one if the CEO experienced a recession during the impressionable years of early adulthood (i.e., between 18 and 25 years). Recessions are

A C C E P T E D M A N U S C R I P T
42 defined in this context as a year of negative real GDP growth. Based on this definition, the years 1991, 1992, 1993, 2009, and 2012 are defined as recession years. The real GDP contracted during these years by 5.9%, 3.3%, 0.7%, 8.3% and 1.4%, respectively, according to data from Statistics Finland. These economic contractions reflect two major financial crises that are described in more detail below.
Finland suffered a great depression and banking crisis in the 1990s after the collapse of the Soviet Unionits trade partnerand the boom and bust followed by the liberalization of the Finnish financial markets. The resulting economic contraction in Finland during the period 1991-1993 turned out to be the deepest contraction experienced by an industrialized country since the 1930s (see, e.g., Gorodnichenko et al. 2012). Honkapohja and Koskela (1999) document that trade with the Soviet Union collapsed almost overnight by 70% in 1991.
Their analysis suggests that financial factors were a key propagation mechanism for the crisis.
After the revaluation of the currency in 1989, Finland had to defend its currency peg from speculative attacks, which kept real interest rates high and short-term rates volatile. The household and firm sectors had become highly indebted because of rapid lending growth in the boom period. Moreover, a large fraction of the corporate borrowing was in foreign currency terms. The hard currency policy was eventually abandoned, which resulted in the depreciation of the Finnish markka in 1991 and 1992 after the forced devaluation and floatation of the currency, respectively. The asset price collapse and the corporate bankruptcies resulted in a banking crisis. Real house prices had risen rapidly in the boom period, only to collapse from the top observed at the end of 1980s to approximately half of their previous value after the financial crisis that accompanied the depression (see, e.g., Taken together, these economic contractions and the accompanying financial crises provide an effective testing ground for analyzing the effects of negative shocks faced by the real economy and the banking sector on the firms established during that period.
Because the recession-born dummy is time-invariant, the firm fixed effects must be dropped in the following models. However, the cohort fixed effects based on the aggregated cohorts, in which the single-year cohorts are grouped at the four-year level, are included in the models. This grouping follows the same approach as used in the earlier analysis. In the current analysis, such grouping is implemented to diminish the multicollinearity between the A C C E P T E D M A N U S C R I P T 44 cohort and recession-born dummies. 25 The inclusion of the aggregated cohort dummies is advantageous because they can be used to control for other cohort-specific trends in the cost and use of credit. 26 Hence, in these cohort models, the recession effects are identified from within the cohort group variation between the firms born in the recession and non-recession years. 27 Recall that industry, region and time fixed effects are included in each model in addition to the firm characteristics, such as the observed creditworthiness. The further specifications also include the age, gender, and birth cohort of the firm CEOs as controls. 28 The further specifications also include an interaction between the recession-born firm dummy and the recession-grown CEO dummy to study the role of firm-and CEO-specific 25 The correlation remains high (that is, close but below 0.80) between the recession-born dummy and the cohort group 1990-1993 containing the firms born during the worst depression years of the 1990s. However, the grouping of cohorts at this interval results in natural and balanced cohort groups, which still avoid the perfect multicollinearity while retaining more accuracy than more coarse groupings, such as the decade fixed effects used by Schoar and Zuo (2016). 26 As a reminder, the age profiles based on the grouped cohorts are less precise than the ones based on the individual birth-year dummies as observed in the earlier analysis. 27 In the case of the Finnish Great Depression of the 1990s, the identification comes from the differences between the cohort born in 1990, a year of modest, close-to-zero growth, and the cohorts of 1991-1993, born after the sudden collapse of the Soviet Union and in the middle of the banking crisis. Indeed, comparisons of these particular years are common (cf. Gorodnichenko et al. 2012). 28 Even though the data consist of corporations rather than proprietorships, it is worth noting the following factors: First, the sample firms are mostly very small (median firm size: three persons), suggesting that the firm behavior and personal traits of the CEOs are likely to be more closely linked than among larger firms. Furthermore, the line between personal and corporate assets is more blurred among micro-sized corporations compared to larger, publicly listed firms.

A C C E P T E D M A N U S C R I P T
45 factors in the recession cohort effects. Furthermore, an interaction between the recession-born dummy and the CEO age analyze whether the effects are different for young and old cohorts.
The previous literature suggests that young individuals could react more strongly to macroeconomic shocks than older ones because the recent experiences constitute a more significant part of their lifetime (see Giuliano and Spilimbergo 2014).  (6)). The dependent variables are defined as follows: Financing costs is financial expenses divided by the average interest-bearing debt between t and t-1. Bank debt is a ratio of outstanding loans from financial institutions divided by total assets at t. The independent variables are defined as follows: Age, modeled using dummies for each age, is the age of the firm defined as the years since the initial incorporation at t. Ln(Sales) is a natural logarithm of net sales at t-1. Tangibility is a ratio of fixed to total assets at t-1.
Profitability is EBITDA divided by total assets at t-1. Credit score measures the observed creditworthiness of the firms (i.e., the probability of default) at the scale 3-100 (scaled by dividing by 100), where higher values mean lower creditworthiness. Recession-born is an indicator equal to one, if the firm was born during the period of negative real GDP growth (years 1991, 1992, 1993, 2009, and 2012), and zero otherwise. Ln(CEO age) is a natural logarithm of the age of the CEO. When the entrepreneur-specific controls (i.e., the age, gender, and birth cohort of the CEO) are included in the model, the recession-born dummy is positive but weakly significant. The coefficient is somewhat more conservative, suggesting 11 basis points higher financing costs in this subsample for which the entrepreneur-specific controls are available.
The recession-grown CEO dummy is insignificant in column (3) Regarding the use of bank loans, the recession-born dummy is negative and highly statistically significant. That is, the recession-born cohorts use lower amounts of bank loans than the non-recession cohorts. The coefficient remains negative and statistically significant, albeit somewhat more conservative, after the entrepreneur-specific controls are included. 30 The recession-grown CEO dummy is insignificant, suggesting that the experience of a recession in the yearly adulthood is not significantly reflected in the use of bank loans for these particular cohorts and time periods. Columns (4) shows that the interaction between the recession-born dummy with the recession-grown CEO dummy is negative albeit insignificant, suggesting that the recession-born effects are not significantly larger for firms with CEOs grown during recessions in the full sample. Columns (5)   The concentration on the older cohorts born in the 1990s should provide further insights on whether the recession effects have a lasting rather than transitory impact on the firms.
The results that focus on the depression cohorts of the 1990s in the pre-2008 period remain very similar to the previous estimates as can be seen from specification (1). The additional specifications with the entrepreneur-specific controls also show similar albeit statistically insignificant estimates. The recession-born effect in the financing costs of these cohorts is 18 basis points, in line with the results observed in the case of the more broadly defined recession cohort.
In the case of bank loans, the interaction between the recession-born dummy and the recession-grown CEO dummy is negative and now weakly significant in column (4), suggesting that the recession-born firm effects are larger in absolute term for the CEO cohorts grown during the Finnish Great Depression. The average marginal effects of the recessionborn dummy for the recession-grown CEOs are -0.014 and significant at the 5% level. The economic magnitude of the effect appears to be significant; the predicted values of the bank debt (0.124) for the recession-grown CEOs in the recession-born firms are more than nine percent lower than for the non-recession-grown CEOs (0.137). The interactions between the recession-born firm dummy and the CEO age remain positive and significant in columns (5)-(6). The average marginal effects for the recession-born dummy in column (6) evaluated at 31 The correlation coefficient between the recession-born indicator and the cohort group dummy containing cohorts 1990-1993 rises from the previous less than 0.80 to 0.84, which might as such call for some caution, while the above estimates seem to provide no obvious reasons for concern.
A C C E P T E D M A N U S C R I P T 52 the (log of) CEO ages 25 and 45 are about -0.019 and -0.006, respectively. They are significant at the 5% level for the younger cohort and insignificant for the older cohort.
Overall, these findings suggest that the depression cohorts of the 1990s are a key group behind the persistent differences between the recession and the non-recession cohorts observed previously in the analysis. However, the further (unreported) analysis suggests that the significance of the younger recession cohorts has risen relative to the older depression cohorts during the more recent periods in the aftermath of the international financial crisis. 32 The magnitude of the recession-born effect on the financing costs (up to 20 basis points) roughly matches and even exceeds the magnitude of a five-point change in the credit score from the lower bound of the credit rating class AA+ (credit score: 20) to the lower bound of the credit rating class AA (credit score: 25). This credit rating change would increase the predicted financing costs approximately 14 basis points (i.e., from 4.97% p.a. to 5.11% p.a).
The persistent recession-born effect of a similar magnitude is intriguing because the observed creditworthiness of the firms is controlled for in the regressions. The recession-born effect on the amount of bank debt used also appears to be significant in economic terms. Because the mean value of bank debt is 0.137, the baseline recession-born estimate of approximately -0.009 suggests that the recession cohorts use an amount of bank debt that is more than six percent lower than the amount used by the non-recession cohorts. The observation of such a 32 The baseline estimates for the Finnish Great Depression cohort remain quite robust up to year 2010.
The statistical significance and size of the estimates for this cohort is weaker and more conservative, respectively, in the sample covering the later years of the post-crisis period. This could indicate that the younger recession cohorts start to dominate in the sample during the more recent years and that ignoring them would underestimate the effect. Indeed, the prolonged Finnish recession that began in 2012 has continued in 2013-2014, already matching the length of the 1990s depression.
A C C E P T E D M A N U S C R I P T 53 lasting impact suggests persistent differences either in the firms' perceived riskiness or in the entrepreneurs' attitudes towards bank finance. This result suggests that severe recessions and periods of financial instability could have scarring effects.
The previous literature suggests that certification provided by banks is particularly important for firms that do not have access to public debt markets (see, e.g., Diamond 1991;Slovin et al. 1993). This suggests that any problems in the lending relationships could have Second, the earlier literature suggests that corporate managers who started out during recessions could have less faith in financial markets and could utilize external finance more conservatively (cf. Graham and Narasimhan 2004;Malmendier and Nagel 2011;Malmendier 33 Alternatively, one could also make an argument that these firms at least survived the depression, unlike other potentially lower quality firms. This potential selection effect is worth entertaining in the interpretation. However, a survivorship bias of this kind could in fact predict an opposite sign for the estimates and suggest even larger absolute effects than observed here.   suggest that recession cohorts, having witnessed a major financial crisis, may be debt averse and lean excessively towards internal finance. Schoar and Zuo (2016) also suggest that recession-born managers make more conservative capital structure choices, including lower use of leverage. Indeed, these differences in the attitude towards bank finance could explain why the recession cohorts use bank loans in smaller amounts than other cohorts.
The findings of this study also indicate that the recession-born effects are largest among the young CEOs, suggesting that the scarring effects are most damaging for them. These findings appear to be in line with the psychological and economic literature that suggests that macroeconomic shocks have a more significant effect on young individuals (see Giuliano and Spilimbergo 2014). However, the recession cohort effect appears to be more related to the firms started during recessions rather than to the earlier experiences of the CEOs grown up in recessions, at least in the sample covering also the recent post-crisis years of the international financial crisis. Meanwhile, the analysis of the Finnish Great Depression cohorts in the pre-2008 period suggests that the long-term impact of the recession-born firm effect on the use of bank loans is larger for the recession-grown (and young) CEOs (i.e., those, who were still young adults when their firm was started). Taken together, the findings of the study suggest that the experience of a recession during the (early years of) entrepreneurship appears to be an important factor driving the findings on the long-lasting recession impact. 34 The previous studies that analyze the cohort-specific effects of the Great Depression and other U.S.
recessions have focused on publicly listed firms. The current study suggests that similar 34 Although the intergenerational transmission of the recession effect may not be ruled out, say, in the case of family firms.
A C C E P T E D M A N U S C R I P T 55 persistent effects are observed among privately held small businesses in a different institutional environment in Finland.

Conclusions
This paper studied the life-cycle profiles of small firms' financing costs and the use of bank financing. The study used an extensive panel of Finnish firms from the 1999-2013 period and paid attention to disentangling age, period, and cohort effects in the empirical models.
This identification problem has been largely ignored in the earlier corporate finance literature.
The findings of the current study suggest that the choice of method affects the conclusions drawn about the relationship between financing costs and firm age. The cross-sectional age profiles of financing costs are hump-shaped and consistent with the hold-up theories. In contrast, the methods that control for cohort or firm fixed effects suggest that the financing costs decrease monotonically as the firms mature, in line with the prediction of Diamond (1989). The findings suggest that these differences in the life-cycle profiles relate to cohort effects. Moreover, the age profiles of the use of credit indicate that firms are more dependent on financial intermediaries in the early periods of their lives.
A few main findings are made about the cohort effects. First, the younger cohorts face lower costs of credit than the older cohorts. While the source of this cohort effect was not formally tested, the longer-term trend of decreasing cohort-specific financing costs would generally appear to be consistent with the hypothesis regarding improvements in the financial system and information environment. Second, the findings suggest that the cohorts born in recessions, particularly the Finnish Great Depression and accompanying banking crisis of the 1990s and the more recent international financial crisis, face higher financing costs and use lower amounts of bank loans in a persistent fashion. This effect is robust to controlling for the A C C E P T E D M A N U S C R I P T 56 creditworthiness of the firms with commercial credit scores. The recession-born firm effect is larger for younger CEOs, in line with the prediction that macroeconomic shocks have a more significant effect on young individuals. However, the recession cohort effect appears to be more related to the experience of starting-up the firm in the recession than to the CEOs growing up in a recession during their early adulthood. Overall, these findings suggest that recessions and periods of financial instability could have a lasting impact on the perceived riskiness of the firms and their use of external finance in the future. Such persistent effects, observed even many years after the depression and banking crisis of the 1990s, are intriguing and might call for additional research to further understand their causes.
The findings could also prove useful in the designing of policies to avoid lasting adverse effects from recessions and periods of financial instability. First, the decrease in the cost of credit has diminished the case for government intervention. Second, the life-cycle profiles of the cost and use of credit indicate that potential policy interventions would likely have best rationalization when targeting younger firms. Finally, the long-lasting recession effects faced by the recession cohorts imply that the periods of financial instability might call for some policy measures targeted to bank-dependent small businesses. However, an effective implementation of the policy measures remains a key challenge.
Taken together, the findings of the paper suggest that the choice of method in disentangling age, period, and cohort effects matters for the conclusions drawn about the lifecycle effects in small business finance. One key implication from the analysis is that the lifecycle profiles estimated from cross-sectional datasets, whose use has been a common practice in the previous corporate finance literature, should be interpreted with caution. Moreover, the existence of cohort effects in the cost and use of credit observed in this study also suggests that the identification problem should not be overlooked either in the repeated cross-section and Bankruptcy. In addition, two-digit-level industry dummies and two-digit zip-code-level regional dummies are included in specifications 1-5. The models with the following additional controls are estimated: 1) time fixed effects 2) cohort fixed effects 3) cohort fixed effects (birth years grouped at the four-year level) and time fixed effects 4) cohort fixed effects and macro controls 5) cohort and time fixed effects 6) firm and time fixed effects.