Three Essays in Corporate Finance: The Evolution of Capital Structure and the Role of Institutional Investors on Cash Holdings and on Firm Value

Transcription

1 Three Essays in Corporate Finance: The Evolution of Capital Structure and the Role of Institutional Investors on Cash Holdings and on Firm Value Yangyang Chen Submitted in total fulfilment of the requirements of the degree of Doctor of Philosophy November 2010 Department of Finance The University of Melbourne

2 Abstract Chapter 1. Capital Structure Convergence: Real or Mechanical Lemmon, Roberts, and Zender (2008) show two features of the data on capital structure: convergence and persistence. I replicate their results and then explore the source of the features. I show that capital structure convergence is likely to be mechanical rather than real. It sources from a statistical fallacy called the regression fallacy. The stationarity of the leverage ratio series and the misclassification problem in portfolio construction give rise to the apparent convergence. Finally, I test two implications of the misclassification argument and propose a method that is able to mitigate the mechanical convergence. Chapter 2. Institutional Ownership and Firm Cash Holdings The precautionary theory of firm cash holdings argues that firms hold cash to protect themselves against adverse cash flow shocks or external financial constraints that might force them to default on payments or forgo valuable investment opportunities. The aggressive trading of a group of institutional investors magnifies the firm s stock price volatility and makes its external financing costly. This may induce the firm s precautionary demand for cash. Consistent with the hypothesis, I find that firms with greater institutional ownership have more cash holdings. The relationship is purely driven by the ownership of short-term institutions (i.e., institutions that trade frequently for short-term trading profits). The ownership of long-term institutions (i.e., institutions that trade infrequently for long-term holding profits) is negatively associated with firm cash holdings. Further tests show that the effect of both types of institutional ownership on firm cash holdings is more significant for growth firms that rely more on external financing and suffer more from cash shortages. Chapter 3. On the Relationship between Institutional Ownership and Firm Value: The Role of Information Prior studies that explore the relationship between institutional ownership and firm value focus exclusively on institutional monitoring. However, the information role of institutional investors may also drive the relationship. Institutional investors improve the firm's information environment through their trading and holding activities. This increases firm value via reducing the estimation risk of the stock and thus the cost of capital. Consistent with the information hypothesis, I find that institutional ownership is associated with a reduction in the cost of capital. Even for the group of institutions that are less likely to monitor the management, their ownership still enhances firm value and the enhancement is exclusively through reducing the cost of capital. Finally, I show that the effect of institutional ownership is more prominent for firms potentially subject to more information asymmetry. 1

3 Declaration This is to certify that i. the thesis comprises only my original work towards the PhD except where indicated in the Preface, ii. due acknowledgement has been made in the text to all other material used, iii. the thesis is less than 100,000 words in length, exclusive of tables, maps, bibliographies and appendices or the thesis is [number of words] as approved by the RHD Committee. Signature 2

4 Acknowledgements I am indebted to my supervisors Professor Christine Brown, Professor Bruce Grundy, and Dr. Chander Shekhar. Without their guidance and encouragement, it is impossible for me to complete the thesis. I also thank my wife and parents. It is their continuous support that enables me to insist on the research. 3

5 Table of Contents Chapter Introduction Data Replication of LRZ Testing Capital Structure Convergence The Misclassification Argument and Its Implications Conclusion References Figures Tables Chapter Introduction Hypotheses Theories of Firm Cash Holdings Hypothesis Development Data Institutional Ownership Firm Cash Holdings and Control Variables Sample and Summary Statistics Empirical Results Univariate Analysis Regression Analysis Robustness Checks Endogeneity Time-series Trend Regression on Changes Alternative Institutional Investor Classification Further Evidence Conclusion Reference

6 Tables Appendix Chapter Introduction Hypothesis Development The Information Role of Institutional Investors Information in Security Pricing Hypothesis Data Institutional Ownership Firm Value and Control Variables Sample and Summary Statistics Empirical Results Endogeneity Omitted Variables Reverse Causality and Simultaneity Further Evidence Isolating the Role of Information Tests on the Information Hypothesis Reconciling Chapter 2 and Chapter Conclusion Reference Tables

7 Chapter 1 Capital Structure Convergence: Real or Mechanical? 6

8 1. Introduction The modern theory of capital structure starts with Modigliani and Miller s (1958, 1963) proof of capital structure irrelevance. Subsequent research relaxes assumptions of the Modigliani-Miller theorem and develops various capital structure theories. Among them, the most noteworthy ones are the trade-off theory and the pecking order theory, which dominated the literature for decades. In recent years, capital structure research has diverged from existing theories in an attempt to explore the evolution and determinants of capital structure in a broader setting (see e.g., Baker and Wurgler (2002), Welch (2004), and Malmendier, Tate, and Yan (2005)). In a recent paper, Lemmon, Roberts, and Zender (LRZ, 2008) show two features of capital structure data that are not identified by previous studies, namely, capital structure convergence and persistence. They sort firms into quartiles in terms of their leverage ratios and examine the evolution of portfolio leverages in subsequent years. The results show that firms with relatively high (low) leverage tend to move toward more moderate levels of leverage; and despite the convergence, firms with initially high (low) leverage tend to maintain relatively high (low) leverage. As supporting evidence of capital structure persistence, they include the firm s initial leverage in a leverage regression and show that initial leverage plays an essential role in the firm s future leverage choices. This study replicates LRZ to examine more closely the convergence feature of capital structure. I follow their sample construction procedures as closely as possible and replicate their main findings first. In particular, I construct the figures that present the evolution of leverage and perform the regression that includes initial leverage as one of the leverage determinants. My results are similar to those of LRZ; both convergence and persistence features are fairly strong and initial leverage is highly significant both economically and statistically. Therefore, I am able to replicate the main results of LRZ in my sample. LRZ propose an explanation of the convergence feature based on the firm's active management of capital structure. However, it is more likely that the feature is mechanical rather than real. It may come from a statistical fallacy called the regression fallacy. To confirm this conjecture, I examine the portfolio leverages before the portfolio formation year. If the convergence feature is real, portfolio leverages must have the same behavior no matter whether I look backward or forward. 7

9 In sharp contrast, when the portfolio leverages before the formation year is examined, a divergence, instead of convergence, pattern is observed. Next, I examine one implication of capital structure convergence: if leverage ratios converge over time, their cross-sectional variance must decrease. I find evidence counter to this implication. Taken together, the two tests provide strong evidence of the mechanical nature of the apparent convergence in my sample. Since my sample is similar but not identical to that of LRZ, I am unable to affirm that the convergence feature in their sample is also mechanical. Nevertheless, the evidence at least suggests that caution should be exercised in investigating the feature. As the observed convergence in my sample is likely to be mechanical, an immediate question arises: where does the mechanical convergence come from? I propose a misclassification argument that answers this question. As demonstrated in LRZ and a number of papers in the target adjustment literature (e.g., Hovakimian, Opler, and Titman (2001), Fama and French (2002), Flannery and Rangan (2006), Alti (2006), Huang and Ritter (2009)), leverage ratios are stationary and revert back to the mean over time. It is precisely the stationarity of leverage ratio series, together with the misclassification problem in portfolio construction, which induces the convergence in capital structure. For example, if firms are sorted according to the portfolio formation year leverage, it is highly probable that some firms are classified into wrong portfolios. When the leverage ratio series revert back to the mean, high and low leverage portfolios are pushed by a single force to the centre, inducing the convergence of these portfolios. The leverage ratios of the moderate leverage portfolios remain stable as the mean reversion generates two forces that cancel out each other. 1 The misclassification argument has two important implications: capital structure convergence will be more severe and initial leverage will be less significant for firms with high leverage ratio variance. I test the two implications by splitting the sample into subsamples with different levels of leverage ratio variance. The results are consistent with the implications. Furthermore, I propose a new sorting method based on the misclassification argument. As the portfolio formation year leverage tends to be extreme, I employ the firm s average past leverage in portfolio construction and find that the mechanical convergence is mitigated. 1 The misclassification argument will be spelt out fully in Section 5. 8

10 The rest of the paper is organized as follows. Section 2 describes the data. Section 3 replicates the key results of LRZ. Section 4 performs tests on capital structure convergence. Section 5 proposes the misclassification argument, and Section 6 concludes. 2. Data This section gives a brief description of the sample construction process as well as the summary statistics of the sample. Following LRZ, my primary sample consists of all nonfinancial firms in the merged Compustat/CRSP industrial annual database between 1965 and I require that all firm-year observations have nonmissing data for book assets and that the leverage ratios (both book and market) lie in the closed unit interval. All the other variables in the multivariate analysis are required to have nonmissing data and are winsorized at both the upper and lower one-percentile. Variable definitions are presented in Table 1. Table 2 presents the summary statistics for the entire sample (All Firms), the subsample of firms with at least 20 years of data (Survivors) and the subsample of the remaining firms (Non-Survivors). The Survivors sample is used by LRZ as a robustness check to address the potential survivorship bias. Panel A presents the mean, median, and standard deviation for each variable. The statistics are close to those of LRZ, indicating the similarity of my sample to theirs. Furthermore, comparison of the three samples reveals that firms in the Survivors sample tend to be large, mature firms with higher leverage ratios, especially market leverage. Firms in the Non-Survivors sample are just the opposite. The Survivors sample also has smaller leverage ratio dispersion, as the standard deviation of leverage ratios for this sample is smaller than that of the Non-Survivors sample. To explore this issue further, I calculate the percentage of observations in each leverage ratio interval for the three samples and present the results in Panel B. The results show that the Non-Survivors sample has a much larger fraction of observations that have either zero leverage or extremely high leverage. For example, only 7.33% (2.39%) of observations in the Survivors sample have zero (above 0.8) book leverage while the percentage in the Non-Survivors sample is 12.34% (5.62%). As most firms in the Non-Survivors sample are small, growth firms, it is not surprising that they tend to have extremely low leverage. Moreover, since firms in this sample have short lives, 9

11 it is also not surprising that a higher fraction of firms in this sample have extremely high leverage ratios just prior to bankruptcy. 3. Replication of LRZ My empirical tests start with replicating the key results of LRZ. Their first remarkable finding comes from diagrams presenting the evolution of leverage portfolios 2. The figures are constructed in the following way: First, each calendar year, firms are sorted into quartiles according to their current year leverage ratios, denoted as: Very High, High, Medium, and Low. The portfolio formation year is denoted event year zero. Second, the average leverage for each portfolio is calculated in each of the subsequent 20 years, holding the portfolio composition constant unless a firm exits the sample. Third, the two steps of sorting and averaging are repeated for every calendar year in the sample period. This process generates 39 sets of event time averages, one for each calendar year in the sample. Fourth, the average leverage of each portfolio across the 39 sets is computed and plotted by event year. I perform the same exercise as LRZ and present the results in Figure 1. The figures for the All Firms sample in Panel A show that there is a large dispersion among leverage portfolios at event year zero and the dispersion decreases sharply over time. However, despite the dispersion decrease, the leverage portfolios still retain their relative position even after 20 years. The figures of the Survivors sample in Panel B exhibit similar patterns. It is clear from the figures that the convergence and persistence features of capital structure shown in LRZ are reproduced in my sample. Further, I replicate graphs for unexpected leverage portfolios as well. Each calendar year, I run a cross-sectional regression of leverage on one-year lagged value of firm size, market-to-book ratio, profitability, tangibility, and industry dummy variables (Fama-French 38 industry classification). Unexpected leverage is defined as the residuals from the cross-sectional regressions. I then sort firms into four portfolios based on unexpected leverage and then track the average actual leverage for each portfolio over the subsequent 20 years. The purpose of using unexpected leverage is to see whether capital structure convergence and persistence are driven by variations in firm characteristics. The results are presented in Figure 2 and are identical to those in LRZ. 2 Figure 1 of LRZ. 10

12 Next, I replicate their regressions that include initial leverage as one of the leverage determinants. The regression specification is as follows: LEV INLEV SIZE MB PROF i, t 1 i 1 i, t 1 2 i, t 1 3 i, t 1 TANG INDML CFV DIVPAY Yr 4 i, t 1 5 i, t 1 6 i, t 1 7 i, t 1 t i, t (1) where i indexes firms, t indexes years, Yr is year fixed effect, and ε is a random error term. The dependent variable is leverage ratio (LEV), either book leverage (BLEV) or market leverage (MLEV). The independent variable of interest is the firm's initial leverage (INILEV), which is proxied by the first nonmissing value of LEV. Control variables include firm size (SIZE), market-to-book (MB), profitability (PROF), tangibility (TANG), industry median leverage (INDML), cash flow volatility (CFV), and dividend payer (DIVPAY). The regression is performed by pooled ordinary least squares (pooled-ols), with the t-statistics computed using standard errors robust to both clustering at the firm level and heteroskedasticity. To avoid an identity at the initial year, the first observation is dropped for each firm. The regression results are presented in Table 3. The results for the All Firms sample show that initial leverage (both book and market) is statistically significant in all three regression specifications. In the regression that includes all the leverage determinants, the coefficient of initial leverage is larger in magnitude than that of all the other determinants except for industry median leverage. Similar results are observed for the Survivors sample. Replicating those in LRZ, the regression results confirm the finding in the graphs with regard to the persistence feature of capital structure. As to the source of the features, LRZ propose that firms active management of capital structure may contribute to the convergence in capital structure. After demonstrating that the key results in LRZ are replicable in my sample, I will conduct a closer examination of the convergence feature in the next section. 4. Testing Capital Structure Convergence Section 3 has reproduced the convergence and persistence features of capital structure. However, it seems from the history of statistics that the convergence feature sources 11

13 from a common fallacy in statistical analysis called the regression fallacy 3 and is thus mechanical rather than real. A classic example of the fallacy is Secrist (1933) 4 who analyzes the profitability of 49 department stores in U.S. between 1920 and He first divides the stores into four equal groups based on the initial year profitability and then calculates average profitability of these groups over time. He finds a remarkable tendency of the group profitability to converge toward a moderate level. The same pattern is observed on a large number of other industrial time series. Nevertheless, little convergence is observed when Secrist extends the analysis to a 10-year series of July temperatures in 191 U.S. cities. He thus concludes that convergence only occurs where competitive forces held sway, and the outcomes were under human control. However, Hotelling (1933) points out in a review of Secrist s work that the apparent convergence of the group averages is a statistical accident, resulting from the method of grouping. Since the groups are formed by possibly extreme values, the average value of each group will converge to a moderate level as the individual values regress to the mean. If the groups are formed by the values in the last year rather than the first year, the figures will exhibit divergence rather than convergence. Hotelling (1933, p.464) writes, these diagrams really prove nothing more than that the ratios in question have a tendency to wander about. As to the lack of convergence among the cities temperature, he figures out that this arises because the temperature series do not move about. Hotelling points out further that the real evidence of a tendency to convergence is a decreasing cross-sectional variance, not among group averages, but among individual values. 5 As the portfolio sorting method in LRZ is similar to that of Secrist (1933), I follow Hotelling s argument and propose two tests to examine the validity of capital structure convergence. The first is the backward-looking test. In the test, I create new figures with the portfolio formation year being the last year of the event period rather than the first year. The figures are constructed in the same way as Figure 1 except for the second step: I examine the average leverage for each portfolio in each of the previous 20 years, instead of the subsequent 20 years, holding the portfolio composition 3 I thank Prof. Stephen Brown for referring me to the statistical literature related to the fallacy. 4 See Stigler (1996) for a detailed description of this example. Other examples of the fallacy include Baumol et al. (1989) and Sharp (1985, p.430). 5 Detailed illustration of the regression fallacy as well as examples of the fallacy can be found in Friedman (1992), Stigler (1997), and Becker and Greene (2001). 12

14 constant unless a firm has not entered the sample. If the convergence feature in capital structure is a statistical fallacy, the new figures will exhibit divergence rather than convergence. The figures are presented in Figure 3. As expected, although the leverage portfolios still exhibit persistence and are much more volatile, they diverge to more extreme values over event time rather than converge to a moderate level. The observed divergence casts doubt on the validity of LRZ conclusion of capital structure convergence. To explore this issue further, I calculate portfolio leverage for both the previous and subsequent 10 years of the portfolio formation year and present the graphs in Figure 4. It seems from the graphs that the greatest dispersion among leverage portfolios is always observed at the portfolio formation year. No matter whether I move forwards or backwards, the dispersion decreases, leading to divergence before the portfolio formation year and convergence after the portfolio formation year. The second test is the variance test. In this test, I investigate the evolution of the cross-sectional variance of leverage ratios among individual firms. To ensure consistency with Figure 1, I construct the figures in the following manner. First, a portfolio is formed each calendar year, including all the firms that exist in that year. The portfolio formation year is denoted event year zero. Second, cross-sectional variance of leverage ratios are calculated in each of the subsequent 20 years, holding the portfolio composition constant unless a firm exits the sample. Third, the calculation is repeated for every year in the sample period. This process generates 39 sets of event time variances, one for each calendar year in the sample. Fourth, the average variance of the portfolio across the 39 sets is computed and plotted in the figure by event year. The generated figures are presented in Figure 5. Panel A shows that the variance of each of book and market leverage of the All Firms sample declines slightly over time. It seems that the variance decrease provides some supporting evidence for capital structure convergence. However, caution must be exercised in interpreting the results. As shown in Section 2, the Non-Survivors sample, which contains firms with short lives, have a larger fraction of observations that have either zero or extremely high leverage ratios. When firms in this sample exit the portfolios over event time, the cross-sectional variance of leverage ratios must decrease. Therefore, it is likely that the observed decline in Panel A is induced by the survivorship bias in the All Firms sample. To address this issue, I perform the same exercise on the Survivors sample, which suffers less from the survivorship bias. The 13

15 figure is presented in Panel B and reveals that the cross-sectional variance of this sample is fairly stable and does not have any decreasing trend. In words, both the backward-looking test and the variance test provide evidence against the validity of the apparent capital structure convergence in my sample. It is likely that the convergence feature shown in LRZ is mechanical as well. 5. The Misclassification Argument and Its Implications The tests in Section 4 demonstrate that the convergence feature in Figure 1 is illusory. I will illustrate briefly why this happens. As demonstrated in LRZ and the target adjustment literature (e.g., Hovakimian, Opler, and Titman (2001), Fama and French (2002), Flannery and Rangan (2006), Alti (2006), and Huang and Ritter (2009)), leverage ratios are stationary and fluctuate around a long-term mean. However, at each portfolio formation year, the leverage ratio is not necessarily equal to the mean. Rather, it may be far larger or smaller. The consequence of sorting firms based on current year leverage gives rise to a misclassification of firms which results in convergence of the portfolio averages. The event year averaging method in LRZ magnifies the convergence further as the misclassification problem occurs each calendar year. Take the Very High portfolio for example. Although my intention is to include all the firms with the highest level of leverage ratio in this portfolio, I may include some firms with high leverage in the portfolio formation year but low mean leverage and miss some firms with low leverage in the portfolio formation year but high mean leverage. Because the Very High portfolio forms the upper boundary, the misclassification only occurs at its lower boundary. When I move away from the portfolio formation year (no matter whether forwards or backwards), the reversion of the leverage ratio towards the mean generates a single force that pushes the portfolio leverage downward. As the partial adjustment literature indicates, the mean reversion in leverage ratio series is a short-term phenomenon. For example, Fama and French (2002) estimate an adjustment period of 5-11 years, while Flannery and Rangan (2006), Huang and Ritter (2007) and LRZ estimate a shorter adjustment period of 3-8 years. This explains why most of the convergence or divergence occurs in the first few years following the portfolio formation year. The same principle applies to the Low portfolio but in an opposite direction. As to the High portfolio and the Medium 14

16 portfolio, since they lie in the center, the misclassification occurs at both their upper and lower boundaries. The mean reversion of the leverage ratio generates two forces, one pushes the portfolio leverage up and the other pushes it down. The two forces cancel out each other, resulting in the relatively stable portfolio leverage of the High portfolio and the Medium portfolio. One immediate implication of the misclassification argument is that capital structure convergence will be more significant for firms with a higher leverage ratio variance, since the more volatile the leverage ratio is, the higher the probability that the portfolio formation year value is extreme and hence more misclassification problems will occur. In an attempt to test this implication, I sort the sample according to the time-series leverage variance of individual firms. I then split the sample into two subsamples with an equal number of firms. I construct figures for the subsamples separately in the same manner as that used in constructing Figure 1. The figures are presented in Figure 6 and confirm the implication. The figures show that for both the All Firms sample and the Survivors sample, a high leverage ratio variance gives rise to strong convergence. For the subsample of firms with high leverage ratio variance, the High and Medium book leverage portfolios even switch their relative position at the end of the event period. In sharp contrast, the leverage ratios of the subsample of firms with low leverage ratio variance are fairly persistent. Only slight convergence is observed among the leverage portfolios. Another implication of the misclassification argument is related to the regression specification. Not only the portfolio formation year leverage, but also the initial leverage, proxied by the first nonmissing value of leverage, is more likely to be extreme for firms with a higher leverage ratio variance. This results in the observed negative relationship between the significance of initial leverage and the leverage ratio variance. To verify the implication, I split the samples into four subsamples with equal number of firms, according to the leverage ratio variance of individual firms and perform the same regression as Equation (1) on the subsamples separately. The regression results are presented in Table 4. The results of the both the All Firms sample and the Survivors sample show that after controlling for firm characteristics, both the coefficient and the statistical significance of initial leverage (both book and market) decrease sharply as the leverage ratio variance increases. The evidence presented above suggests caution in drawing conclusions about the convergence and persistence features in capital structure. In an earlier version of 15

17 Chang and Dasgupta (2009), the authors replicate Figure 1 of LRZ for the actual data as well as simulated samples. They find that capital structure persistence disappears only in the simulated sample in which the financing deficit and change in retained earnings are drawn randomly and firms finance the deficit by a coin toss (i.e., the firm has equal probability of using debt or equity financing ). The leverage portfolios of the random deficit sample converge rapidly and become almost indistinguishable at the end of the event period. They then exercise the same regression specification as Equation (1) on the random deficit sample as a double check and find that the significance of initial leverage is far smaller for this sample than for the actual data as well as other simulated samples generated with the actual deficit. The initial leverage becomes even insignificant in the random deficit sample when the first 20 years are excluded from the regression. The authors conclude that the firm-specific component in the financing deficit and change in retained earnings is most likely to be the source of the capital structure persistence. However, it seems from the evidence presented in Figure 6 and Table 4 that the weakening persistence in the random deficit sample is more likely due to the increasing leverage ratio variance, as the authors show subsequently that the variance of the leverage ratios in the random deficit sample is much higher than that in all the other samples. As the misclassification argument suggests, the apparent capital structure convergence arises because I classify firms by a leverage ratio differing from the firm s long-term mean leverage. Since the mean leverage is not observable ex ante, a better proxy would be the firm s average past leverage. I reconstruct Figure 1 but with a slight difference in the first step: each calendar year, I sort firms into quartiles according to their average past leverage, rather than the current year leverage. The figures are presented in Figure 7. Compared with Figure 1, the new figures exhibit much less convergence. Take the book leverage portfolios of the All Firms sample for example. The difference between the Very High portfolio and Low portfolio at event time 0 is around 33%, much smaller than that of the corresponding figure in Figure 1 (50%), while the difference between the two portfolios at event year 20 is about 17%, similar to that of the corresponding figure (15%). The flatter portfolio leverages across event time indicates a reduced convergence produced by this sorting measure. Although adopting average past leverage does not remove the mechanical convergence completely, it reduces the convergence to some extent. The remaining convergence comes from the poor performance of average past leverage as a proxy for 16

18 the long-term mean leverage in the first several years, as the number of observations of past leverage is not enough to generate a convincing proxy for long-term mean. In unreported results, I repeat the operations on three samples in which the first three, five and ten observations are dropped after calculating average past leverage, respectively. The results show that capital structure convergence decreases as the number of observations dropped increases, confirming the above argument. 6. Conclusion I show in this study that the capital structure convergence documented in LRZ is likely due to the regression fallacy in statistics. I replicate their results and conduct tests on the apparent convergence in my sample. More specifically, if capital structure convergence is real, it will remain whether the pre or post portfolio formation year leverage is examined. Also, the cross-sectional variance of leverage ratios will decrease over time. The test results support none of the above implications, instead supporting the argument that the convergence feature observed in my sample is mechanical in nature. In fact, the mechanical convergence sources from the stationarity of the leverage ratios time series and the misclassification problem in portfolio construction. I test the implications of the misclassification argument and find consistent evidence. I also propose a method based on the argument in which firms are sorted by average past leverage. The resulting reduced convergence provides further support for the mechanical nature of capital structure convergence in my samples. LRZ make a significant contribution to the literature by documenting the persistence and convergence features in capital structure. However, my findings show that the convergence feature in their sample may not be real. It cautions against studies that try to explore the source of this feature. Subsequent studies that explore the persistence feature in leverage ratio series are suggested to find ways to control for the mechanical convergence, as it may distort the conclusion about the persistence feature. As Friedman (1992, p.2131) states I suspect that the regression fallacy is the most common fallacy in the statistical analysis of economic data. It is likely that this fallacy is also the source of other findings in finance and economics studies. A detailed investigation on this issue is necessary and is left for future research. 17

19 References Alti, Aydogan, 2006, How persistent is the impact of market timing on capital structure, Journal of Finance 61, Baker, Malcolm, and Jeffrey Wurgler, 2002, Market timing and capital structure, Journal of Finance 57, Baumol, William J., Sue Anne Batey Blackman, and Edward N. Wolff, 1989, Productivity and American leadership: The long view, (The MIT Press, Cambridge, Massachusetts). Becker, William E., and William H. Greene, 2001, Teaching statistics and econometrics to undergraduates, Journal of Economic Perspectives 15, Chang, Xin, and Sudipto Dasgupta, 2009, Target behavior and financing: How conclusive is the evidence? Journal of Finance 64, Fama, Eugene F., and Kenneth R. French, 2002, Testing trade-off and pecking order predictions about dividends and debt, Review of Financial Studies 15, Flannery, Mark J., and Kasturi P. Rangan, Partial adjustment toward target capital structures, Journal of Financial Economics 79, Friedman, Milton, 1992, Communication: Do old fallacies ever die? Journal of Economic Literature 30, Hovakimian, Armen, Tim Opler, and Sheridan Titman, 2001, The debt-equity choice, Journal of Financial and Quantitative Analysis 36, Hotelling, Harold, 1933, Review of The triumph of mediocrity in business by Horace Secrist, Journal of American Statistics Association 28, Huang, Rongbing, and Jay R. Ritter, 2009, Testing theories of capital structure and estimating the speed of adjustment, Journal of Financial and Quantitative Analysis 44, Lemmon, Michael L., Machael R. Roberts, and Jaime F. Zender, 2008, Back to the beginning: Persistence and the cross-section of corporate capital structure, Journal of Finance 63, Malmendier, Ulrike, Geoffrey A. Tate, and Jun Yan, 2005, Corporate financial policies with overconfident managers, working paper. Modigliani, Franco, and Merton H. Miller, 1958, The cost of capital, corporation finance and the theory of Investment, American Economic Review 48, Modigliani, Franco, and Merton H. Miller, 1963, Corporate income taxes and the cost of capital: A correction, American Economic Review 48,

20 Secrist, Horace, 1933, The triumph of mediocrity in business, (Bureau of Business Research, Northwestern University). Sharpe, William F., 1985, Investments, 3 rd edition, (Prentice Hall, New Jersey). Stigler, Stephen M., 1996, The history of statistics in 1933, Statistical Science 11, Stigler, Stephen M., 1997, Regression towards the mean, historically considered, Statistical Methods in Medical Research 6, Welch, Ivo, 2004, Capital structure and stock returns, Journal of Political Economy 112,

21 0 0 Portfolio Leverage Portfolio Leverage Portfolio Leverage Portfolio Leverage Figures Figure 1 Evolution of Leverage Portfolios in Event Time The sample consists of all nonfinancial firms in the merged Compustat/CRSP Industrial Annual database between 1965 and I require that all firm-year observations have nonmissing data for book assets and that the leverage ratios (both book and market) lie in the closed unit interval. All the other variables in the multivariate analysis are required to have nonmissing data and are winsorized at both the upper and lower one-percentile. Book leverage (BLEV) is defined as the ratio of book debt over book assets. Market leverage (MLEV) is defined as the ratio of book debt over market assets. Each figure presents the average leverage of four portfolios in event time. Each calendar year, firms are sorted into four portfolios based on their leverage ratios. The portfolio formation year is event year zero. The average leverage for each portfolio is calculated in each of the subsequent 20 years, holding the portfolio composition constant unless a firm exits the sample. The two steps of sorting and averaging are repeated for every calendar year in the sample period. This process generates 39 sets of event time averages. The average leverage of each portfolio across the 39 sets is computed and plotted by event year. Panel A presents figures for the entire sample (All Firms); Panel B presents figures for firms with at least 20 years of data (Survivors). Panel A: All Firms Book Leverage Portfolios Market Leverage Portfolios Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Panel B: Survivors Book Leverage Portfolios Market Leverage Portfolios Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage 20

22 Portfolio Leverage Portfolio Leverage Portfolio Leverage Portfolio Leverage Figure 2 Evolution of Unexpected Leverage Portfolios in Event Time The sample consists of all nonfinancial firms in the merged Compustat/CRSP Industrial Annual database between 1965 and I require that all firm-year observations have nonmissing data for book assets and that the leverage ratios (both book and market) lie in the closed unit interval. All the other variables in the multivariate analysis are required to have nonmissing data and are winsorized at both the upper and lower one-percentile. Book leverage (BLEV) is defined as the ratio of book debt over book assets. Market leverage (MLEV) is defined as the ratio of book debt over market assets. Firm size is defined as the natural logarithm of book assets. Market-to-book ratio is defined as the ratio of market assets over book assets. Profitability is defined as the ratio of operating income over book assets. Tangibility is defined as the ratio of property, plant and equipment over book assets. Each figure presents the average unexpected leverage of four portfolios in event time. For each calendar year, we form four portfolios by ranking firms based on their unexpected leverage ratios. Holding the portfolios fixed for the next twenty years, we compute the average leverage for each portfolio. Unexpected leverage is defined as the residuals from a cross-sectional regression of leverage on firm size, market-to-book ratio, profitability, tangibility, and industry dummy variables (Fama-French 38 industry classification), where all independent variables are lagged one year. Panel A presents graphs for the entire sample (All Firms); Panel B presents graphs for firms with at least 20 years of data (Survivors). Panel A: All Firms Unexpected Book Leverage Portfolios Unexpected Market Leverage Portfolios Event Time (Years) Low Unexpeced Lev. High Unexpeced Lev. Medium Unexpeced Lev. Very High Unexpeced Lev Event Time (Years) Low Unexpeced Lev. High Unexpeced Lev. Medium Unexpeced Lev. Very High Unexpeced Lev. Panel B: Survivors Unexpected Book Leverage Portfolios Unexpected Market Leverage Portfolios Event Time (Years) Low Unexpeced Lev. High Unexpeced Lev. Medium Unexpeced Lev. Very High Unexpeced Lev Event Time (Years) Low Unexpeced Lev. High Unexpeced Lev. Medium Unexpeced Lev. Very High Unexpeced Lev. 21

23 0 0 Portfolio Leverage Portfolio Leverage Portfolio Leverage Portfolio Leverage Figure 3 History of Leverage Portfolios in Event Time The sample consists of all nonfinancial firms in the merged Compustat/CRSP Industrial Annual database between 1965 and I require that all firm-year observations have nonmissing data for book assets and that the leverage ratios (both book and market) lie in the closed unit interval. All the other variables in the multivariate analysis are required to have nonmissing data and are winsorized at both the upper and lower one-percentile. Book leverage (BLEV) is defined as the ratio of book debt over book assets. Market leverage (MLEV) is defined as the ratio of book debt over market assets. Each figure presents the average leverage of four portfolios in event time. Each calendar year, firms are sorted into four portfolios based on their leverage ratios. The portfolio formation year is event year zero. The average leverage for each portfolio is calculated in each of the previous 20 years, holding the portfolio composition constant unless a firm has not entered the sample. The two steps of sorting and averaging are repeated for every calendar year in the sample period. This process generates 39 sets of event time averages. The average leverage of each portfolio across the 39 sets is computed and plotted by event year. Panel A presents figures for the entire sample (All Firms); Panel B presents figures for firms with at least 20 years of data (Survivors). Panel A: All Firms Book Leverage Portfolios Market Leverage Portfolios Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Panel B: Survivors Book Leverage Portfolios Market Leverage Portfolios Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage 22

24 0 0 Portfolio Leverage Portfolio Leverage Portfolio Leverage Portfolio Leverage Figure 4 Evolution and History of Leverage Portfolios in Event Time The sample consists of all nonfinancial firms in the merged Compustat/CRSP Industrial Annual database between 1965 and I require that all firm-year observations have nonmissing data for book assets and that the leverage ratios (both book and market) lie in the closed unit interval. All the other variables in the multivariate analysis are required to have nonmissing data and are winsorized at both the upper and lower one-percentile. Book leverage (BLEV) is defined as the ratio of book debt over book assets. Market leverage (MLEV) is defined as the ratio of book debt over market assets. Each figure presents the average leverage of four portfolios in event time. Each calendar year, firms are sorted into four portfolios based on their leverage ratios. The portfolio formation year is event year zero. The average leverage for each portfolio is calculated in each of the previous and subsequent 10 years, holding the portfolio composition constant unless a firm exits the sample. The two steps of sorting and averaging are repeated for every calendar year in the sample period. This process generates 39 sets of event time averages. The average leverage of each portfolio across the 39 sets is computed and plotted by event year. Panel A presents figures for the entire sample (All Firms); Panel B presents figures for firms with at least 20 years of data (Survivors). Panel A: All Firms Book Leverage Portfolios Market Leverage Portfolios Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Panel B: Survivors Book Leverage Portfolios Market Leverage Portfolios Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage 23

25 Portfolio Variance Portfolio Variance Figure 5 Evolution of Cross-sectional Leverage Variance in Event Time The sample consists of all nonfinancial firms in the merged Compustat/CRSP Industrial Annual database between 1965 and I require that all firm-year observations have nonmissing data for book assets and that the leverage ratios (both book and market) lie in the closed unit interval. All the other variables in the multivariate analysis are required to have nonmissing data and are winsorized at both the upper and lower one-percentile. Book leverage (BLEV) is defined as the ratio of book debt over book assets. Market leverage (MLEV) is defined as the ratio of book debt over market assets. Each figure presents the average cross-sectional variance of leverage ratios in event time. Each calendar year, a portfolio is formed, including all the firms that exist in that year. The portfolio formation year is denoted event year zero. Cross-sectional variance of leverage ratios is calculated in each of the subsequent 20 years, holding the portfolio composition constant unless a firm exits the sample. The calculation is repeated for every year in the sample period. This process generates 39 sets of event time variances. The average variance of the portfolio across the 39 sets is computed and plotted in the figure by event year. Panel A presents figures for the entire sample (All Firms); Panel B presents figures for firms with at least 20 years of data (Survivors). Panel A: All Firms Cross-sectional Variance of Leverage Event Time (Years) Book Leverage Market Leverage Panel B: Survivors Cross-sectional Variance of Leverage Event Time (Years) Book Leverage Market Leverage 24

26 0 0 Portfolio Leverage Portfolio Leverage Portfolio Leverage Portfolio Leverage Figure 6 Evolution of Leverage Portfolios in Event Time for Leverage Variance Subsamples The sample consists of all nonfinancial firms in the merged Compustat/CRSP Industrial Annual database between 1965 and I require that all firm-year observations have nonmissing data for book assets and that the leverage ratios (both book and market) lie in the closed unit interval. All the other variables in the multivariate analysis are required to have nonmissing data and are winsorized at both the upper and lower one-percentile. Book leverage (BLEV) is defined as the ratio of book debt over book assets. Market leverage (MLEV) is defined as the ratio of book debt over market assets. The samples are split into two subsamples with equal number of firms in terms of the time-series leverage variance of individual firms. Each figure presents the average leverage of four portfolios in event time. For each subsample, I sort firms into four portfolios based on their leverage ratios each calendar year. The portfolio formation year is event year zero. The average leverage for each portfolio is calculated in each of the subsequent 20 years, holding the portfolio composition constant unless a firm exits the sample. The two steps of sorting and averaging are repeated for every calendar year in the sample period. This process generates 39 sets of event time averages. The average leverage of each portfolio across the 39 sets is computed and plotted by event year. Panel A presents figures for the high variance subsamples of the entire sample (All Firms); Panel B presents figures for the low variance subsamples of the entire sample (All Firms); Panel C presents figures for the high variance subsamples of firms with at least 20 years of data (Survivors); Panel D presents figures for the low variance subsamples of firms with at least 20 years of data (Survivors). Panel A: All Firms - High Variance Subsample Book Leverage Portfolios Market Leverage Portfolios Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Panel B: All Firms - Low Variance Subsample Book Leverage Portfolios Market Leverage Portfolios Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage 25

27 0 0 Portfolio Leverage Portfolio Leverage Portfolio Leverage Portfolio Leverage Panel C: Survivors - High Variance Subsample Book Leverage Portfolios Market Leverage Portfolios Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Panel D: Survivors - Low Variance Subsample Book Leverage Portfolios Market Leverage Portfolios Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage Event Time (Years) Low Leverage High Leverage Medium Leverage Very High Leverage 26