Corporate Diversification and the Cost of Capital

Transcription

1 Corporate Diversification and the Cost of Capital April 2011 Abstract We examine whether organizational form matters for a firm s cost of capital. Contrary to conventional view, we argue that coinsurance among a firm s business units can reduce systematic risk through the avoidance of countercyclical deadweight costs. We find that diversified firms have on average a lower cost of capital than comparable portfolios of standalone firms. In addition, diversified firms with less correlated segment cash flows have a lower cost of capital, consistent with a coinsurance effect. Holding cash flows constant, our estimates imply an average value gain of approximately 5% when moving from the highest to the lowest cash flow correlation quintile.

2 1. Introduction The conventional view among practitioners and researchers is that organizational form does not matter for a firm s cost of capital because, while the imperfect correlation of business unit cash flows may help reduce idiosyncratic risk, this should have no effect on systematic risk. Long a part of mainstream thought, the conventional view is widely disseminated through standard finance textbooks and classroom teaching. The notion that corporate diversification cannot affect systematic risk is usually covered explicitly in the mergers and acquisitions chapter 1 or implicitly through the standalone principle in the capital budgeting chapter. In this paper, we present evidence that is contrary to the conventional view. We find that diversified firms have a lower cost of capital than comparable portfolios of standalone firms. We also find that the reduction in cost of capital is strongly related to the correlation of business unit cash flows, consistent with a coinsurance effect. We argue that organizational form can affect a firm s cost of capital, and in particular, coinsurance the imperfect correlation of cash flows among a firm s business units can reduce systematic risk through the avoidance of countercyclical deadweight costs. Using deadweight costs of financial distress as an illustrative example, if coinsurance reduces default risk (Lewellen (1971)) and enables a diversified firm to avoid countercyclical deadweight costs of financial distress (Elton, Gruber, Agrawal, and Mann (2001), and Almeida and Philippon (2007)) that its business units would have otherwise incurred as standalone firms, then coinsurance should lead to a reduction in the diversified firm s systematic risk and hence its cost of capital. Costly financial distress is of course just one example of deadweight costs faced by firms. Other examples include adverse selection and transaction costs of external finance and resulting 1 Systematic variability cannot be eliminated by diversification, so mergers will not eliminate this risk at all. (Ross, Westerfield, and Jaffe, p. 823) 1

3 investment distortions, forgone business opportunities due to defections by important stakeholders such as suppliers, customers, or employees, and so on. Many of these costs tend to arise following low cash flow realizations making them countercyclical since low cash flow realizations are more likely during bad economic times. Amplification mechanisms such as the credit channel or asset fire sales can also add to the countercyclical nature of these costs. Our general argument is that coinsurance should enable a diversified firm to transfer resources from cash-rich units to cash-poor units in some states of nature and thereby avoid some of the countercyclical deadweight costs that standalone firms cannot avoid on their own. As a result, cash flows of diversified firms should contain less systematic risk than those of comparable portfolios of standalone firms. In addition, the reduction in systematic risk should depend on the extent of coinsurance among diversified firms business units. We test these predictions using a sample of single- and multi-segment firms spanning the period 1988 to Our main cost of capital proxy is the weighted average of cost of equity and cost of debt. We use ex ante measures of expected returns for both components of financing: implied cost of equity constructed from analyst forecasts to proxy for expected equity returns and yields from the Barclays Capital Aggregate Bond Index to proxy for expected debt returns. We estimate implied cost of equity based on the approach of Gebhardt, Lee, and Swaminathan (2001), which has been recently employed in several asset pricing contexts (Lee, Ng, and Swaminathan (2009) and Pastor, Sinha, and Swaminathan (2008)). Our empirical analyses are based on an excess cost of capital measure that benchmarks the cost of capital of a diversified firm against that of a comparable portfolio of standalone firms. We find that diversified firms on average have a significantly lower cost of capital than comparable portfolios of standalone firms, rejecting the conventional view that organizational 2

4 form does not matter for a firm s cost of capital. We consider cash flow and investment correlations among a firm s segments as an inverse measure of coinsurance. Consistent with a coinsurance effect, we find a significant positive relation between excess cost of capital and cross-segment correlations. In addition, we examine the role that financial slack also plays in helping firms avoid deadweight costs. We find that coinsurance effects are stronger for firms with less financial slack (higher net leverage and lower cash holdings). These findings are robust to controlling for potential analyst biases and using alternative measures of (i) implied cost of equity (Claus and Thomas (2001) and Easton (2004)), (ii) cost of equity not reliant on analyst forecasts, (iii) cost of debt inferred from publicly-traded bonds or private loans, and (iv) coinsurance. They are also robust to controlling for selection effects in a Heckman two-stage analysis and using changes in coinsurance over which managers arguably have no control (Lamont and Polk (2002)). Our findings are also economically significant. Our estimates imply an average percentage reduction of approximately 3% in cost of capital and an average value gain of approximately 5% when moving from the highest to the lowest cash flow correlation quintile. The rest of the paper proceeds as follows. Section 2 provides a discussion of the setting and related research. Section 3 outlines the valuation approach that we use in estimating the implied cost of equity along with the construction of excess cost of capital and coinsurance measures. Section 4 describes our sample. Section 5 presents our findings. Section 6 concludes. 3

5 2. The Setting 2.1. Systematic Risk and Cost of Capital in a Model of Coinsurance Our hypotheses about organizational form and cost of capital are based on a model of coinsurance in the spirit of Lewellen (1971). This section summarizes the model s basics and outlines the assumptions under which the imperfect correlation of business unit cash flows lowers a diversified firm s cost of capital relative to a comparable portfolio of standalone firms. 2 To illustrate our main ideas, suppose that firms incur certain deadweight losses when their projects experience low cash flow outcomes. Examples of deadweight losses include forgone business opportunities due to defections by important stakeholders such as suppliers, customers, or employees, financial distress or external finance costs, and so on. A large body of research in finance shows that the expected value of such deadweight losses is higher during worse economic times, possibly due to the higher incidence of low cash flow outcomes, or due to amplification mechanisms such as the credit channel or asset fire sales. As a result, firms face deadweight losses that are partly countercyclical and firms cash flows contain more systematic risk than they otherwise would in a frictionless world. That is, countercyclical deadweight losses add to the systematic risk of firms. In such a setting, it is straightforward to show that a diversified firm s systematic risk would be lower than that of a comparable portfolio of standalone firms. The imperfect correlation of business unit cash flows allows resources to be transferred from cash-rich units to cash-poor units in some states of nature to avoid some of the countercyclical deadweight losses that standalone firms cannot avoid on their own. More generally, a diversified firm with less correlated business unit cash flows and hence greater coinsurance potential would have less 2 In a previous version of the paper, we used the model to derive additional testable predictions, which we later state in this section. The formal analysis can now be found in an Internet Appendix. 4

6 systematic risk. Only in the case of perfectly correlated business unit cash flows would a diversified firm s systematic risk approach that of a comparable portfolio of standalone firms. For these results to hold, two further assumptions are needed. First, it must be costly for standalone firms to enter into state-contingent financing contracts with each other to replicate the extent of deadweight loss avoidance achieved by diversified firms. Second, it must be costly for firms to hold first-best amounts of financial slack to avoid all future deadweight losses. Both assumptions strike us as accurate descriptions of the real world. Verifiability and enforcement frictions likely render state-contingent financing contracts expensive or infeasible. In addition, tax and agency costs likely discourage firms from holding first-best amounts of financial slack. In the Internet Appendix, we consider two extensions of the basic model. First, we allow for the possibility of agency costs of diversification and the possibility of inefficient internal capital markets to address a model prediction that some might see as counterfactual the basic model without any cost of diversification predicts a diversification premium. We show that these costs do not change the qualitative implications of the model about countercyclical coinsurance. So, it is possible to observe both a diversification discount and a coinsurance effect at the same time. Second, we extend the model to include debt alongside equity and show that the coinsurance results apply to both debt and equity financing. To summarize, the model setting outlined above has the following testable predictions. First, diversified firms should have a lower cost of capital than comparable portfolios of standalone firms. Second, the reduction in cost of capital should be related to expected coinsurance opportunities. Diversified firms with less correlated business unit cash flows and thus greater coinsurance potential should have a lower cost of capital. Third, financial slack, which firms can use to avoid deadweight losses, should reduce the importance of coinsurance 5

7 and its effect on cost of capital. Consequently, coinsurance effects should be smaller for diversified firms with greater financial slack Related Literature The notion of coinsurance among a firm s business units goes at least as far back as Lewellen (1971). The ensuing stream of research studies coinsurance in the context of conglomerate mergers (Higgins and Schall (1975) and Scott (1977)) and examines whether such mergers lead to wealth transfers from shareholders to bondholders (Kim and McConnell (1977)). Importantly, this literature does not recognize the possibility that coinsurance can affect a firm s systematic risk. For example, standard textbooks emphasize the irrelevance of corporate diversification and coinsurance when explaining the standalone principle of capital budgeting by either implicitly following or explicitly citing Schall s (1972) analysis. To our knowledge, our study is the first to establish a link between coinsurance and cost of capital. Our study also complements the literature on corporate diversification and firm value (Lang and Stulz (1994), Berger and Ofek (1995), Campa and Kedia (2002), Graham, Lemmon, and Wolf (2002), Mansi and Reeb (2002), and Villalonga (2004)) by exploring an important dimension that thus far has received little attention, namely, cost of capital. The discussion in this literature revolves mostly around future cash flow differences between conglomerates and standalone firms, and confounding selection effects. An exception is Lamont and Polk (2001), who raise the possibility that valuation differences may arise due to differences in expected returns. They find a significant and negative relation between excess values and future returns 3 It is worth noting that a model of contagion would generate the opposite predictions. For instance, if the liquidity concerns of cash-poor units spread to other units of the firm and cause deadweight losses that standalone firms would not incur on their own, then diversified firms would incur greater deadweight losses than comparable portfolios of standalone firms. 6

8 for diversified firms, suggesting that valuation differences are explained in part by differences in expected returns. While their study introduces the important role of expected returns in understanding the valuation of diversified firms, their main focus is to explain the cross-sectional variation in excess value, and not how diversification affects a firm s cost of capital. Our work deepens the foundations of this literature by exploring whether the cross-sectional variation in cost of capital is due to coinsurance. Our work is also related to an extensive literature on the deadweight costs of external finance, and the ability of different organizational forms to avoid them. Livdan, Sapriza, and Zhang (2009) show that more financially constrained firms are riskier and earn higher expected stock returns than less financially constrained firms. Dimitrov and Tice (2006) show that during recessions both sales and inventory growth rates drop more for bank-dependent standalone firms than they do for rival segments of bank-dependent diversified firms. Yan, Yang, and Jiao (2010) show that standalone firms experience investment declines relative to diversified firms during periods of depressed conditions in external capital markets. Related work by Yan (2006) also shows that diversified firms have higher valuations when external capital is more costly. Hovakimian (2011) shows that more financially constrained diversified firms allocate capital more efficiently during recessions. Using the financial crisis as a natural experiment, Kuppuswamy and Villalonga (2010) show that the value of diversified firms increased relative to standalone firms due to financing and investment advantages. Studying deadweight costs of asset fire sales, Pulvino (1998) finds that financially constrained airlines receive lower prices than their unconstrained rivals when selling used narrow-body aircraft. Consistent with deadweight costs of asset fire sales being countercyclical, Ortiz-Molina and Phillips (2009) find that firms with more liquid real assets have a lower cost of capital. Finally, Duchin (2010) studies the 7

9 relation between coinsurance and firms cash retention policies. Our paper combines with Duchin s to form a nascent literature examining the implications of coinsurance for corporate finance in general. 3. Empirical Design The coinsurance hypothesis outlined in Section 2.1 relates a diversified firm s cost of capital to the extent of coinsurance among its business units. In this section, we discuss our main proxies for these constructs Cost of Capital Prior research in finance has generally used ex post realized returns to proxy for expected returns and cost of capital (Fama and French (1997), Lamont and Polk (2001)). However, realized returns are noisy proxies for expected returns due to contamination by information shocks, which can lead to biased inferences in finite samples (Elton (1999)). To address this concern, recent literature in accounting and finance has developed an ex ante approach to measuring expected returns by estimating the implied cost of equity (Claus and Thomas (2001), Gebhardt, Lee, and Swaminathan (2001), Easton (2004)). The implied cost of equity is the internal rate of return that equates the current stock price to the present value of all expected future cash flows to equity. Namely, the value of the firm at time t can be expressed as P t Et[ FCFEt i], i (1 r ) i 1 e where P t is the market value of equity at time t, FCFE t+i is free cash flow to equity at time t+i, and r e is the implied cost of equity. 8

10 In our main analysis, we follow the approach of Gebhardt, Lee, and Swaminathan (2001) (hereafter, GLS) to estimate the implied cost of equity. The GLS measure has been successfully employed in several asset-pricing contexts (Lee, Ng, and Swaminathan (2009), Pastor, Sinha, and Swaminathan (2008), Chava and Purnanandam (2010)). Following GLS and much of the implied cost of equity literature, we use I/B/E/S consensus analyst forecasts to proxy for future earnings (see the Appendix for details). We then follow an approach similar to Lamont and Polk (2001) to estimate total cost of capital (COC). Lamont and Polk define total cost of capital as the weighted average of a firm s realized equity return and the return on an aggregate bond index. To avoid the pitfalls of using ex post realized returns, we use ex ante measures to proxy for expected equity and debt returns. More specifically, the COC for firm i in year t is computed as follows: COC i,t = D i,t-1 Y BC t + (1 - D i,t-1 ) COEC i,t, where Y t BC is the aggregate bond yield from the Barclays Capital Aggregate Bond Index (formerly, the Lehman Brothers Aggregate Bond Index), COEC i,t is the implied cost of equity (GLS), and D i,t-1 is the firm s book value of debt divided by total value (book value of debt plus market value of common equity). 4 To compare a diversified firm s cost of capital to the cost of capital that its business units would have as standalone firms, we compute excess COC as the natural logarithm of the ratio of the firm s COC to its imputed COC. The imputed COC of the firm is a value-weighted average of the imputed COC of its segments: 4 Book value of debt is long-term debt (Compustat Item #9) plus short-term debt (Compustat Item #34); market value of equity is fiscal year-end stock price (Compustat Item #199) multiplied by shares outstanding (Compustat Item #25). 9

11 icoc i n n k 1 k 1 imv ik imv ik icoc ik, where n is the number of the firm s segments, icoc ik is the imputed COC of segment k, which is equal to the median COC of single-segment firms in the segment s industry, and imv ik is the imputed market value of segment k, calculated as in Berger and Ofek (1995). The procedure for estimating segments imputed market values is described in detail in Berger and Ofek (1995). In short, the procedure consists of: (1) estimating the median ratio of enterprise value to sales for all single-segment firms in the industry to which the segment belongs, and (2) multiplying the segment s sales by the median industry ratio. Industry definitions are based on the narrowest SIC grouping that includes at least five single-segment firms with at least $20 million in sales and has a non-missing COC estimate Coinsurance Measuring the level of coinsurance among a diversified firm s business units is empirically challenging because the joint distribution of future business-unit cash flows is not observable. Moreover, using the distribution of historical business-unit cash flows is problematic because firm composition changes over time. Accordingly, we construct coinsurance proxies using correlations of industry-level cash flows based on single-segment firms. While our main proxy is based on business unit membership in 2-digit SIC industries, we also consider several other proxies using different industry definitions (see Section 5.3 for details). To ensure that estimated pairwise industry correlations are not contaminated with systematic risk, we perform the computation in two stages. First, for each industry in a given year, we compute idiosyncratic industry cash flows for the prior ten years as residuals from a 10

12 regression of average industry cash flow on average market-wide cash flow. Next, for each year in our sample, we estimate pairwise industry correlations using prior ten-year idiosyncratic industry cash flows. As coinsurance of investment opportunities can also help firms avoid deadweight costs of external finance (Matsusaka and Nanda (2002)), we similarly estimate pairwise industry correlations using prior ten-year idiosyncratic industry investments. 5 These estimated correlations serve as inputs to our coinsurance measures described below. As an inverse measure of coinsurance, we compute a sales-weighted portfolio correlation measure ρ it(n) for firm i in year t with n business segments as n n p 1 q 1 w w Corr ( j, k), ip( j) iq( k ) [ t 10, t 1] where w ip(j) is the sales share of segment p of firm i operating in industry j (similarly for business segment q of firm i operating in industry k), and Corr (, ) [ t 10, t 1] j k is the estimated correlation of idiosyncratic industry cash flows or investments between industries j and k over the ten-year period before year t. We obtain similar results using an alternative coinsurance measure, which also includes the standard deviation of industry cash flow and investment (Duchin (2010)). Note that a single-segment firm s sales-weighted cash flow or investment correlation measure equals one by definition. This is also true for a multi-segment firm whose segments operate in the same industry. 4. Sample and Data 4.1. Sample Selection 5 As is standard practice, we measure cash flow as operating income before depreciation (Compustat Item #13) scaled by total assets (Compustat Item #6) and investment as capital expenditures (Compustat Item #128) scaled by total assets (Compustat Item #6). 11

13 We obtain our sample from the intersection of the Compustat and I/B/E/S databases for the period 1988 to We construct cost of capital measures by combining firm-level accounting information from the Compustat annual files with analyst forecasts from I/B/E/S. The excess cost of capital measures and the coinsurance measures require availability of segment disclosures from the Compustat segment-level files. Additionally, we impose the following sample restrictions. First, we follow Berger and Ofek (1995) and require that (1) all firm-years have at least $20 million in sales to avoid distorted valuation multiples; (2) the sum of segment sales be within 1% of the total sales of the firm to ensure the integrity of segment data; (3) all of the firm s segments for a given year have at least five firms in the same 2-digit SIC industry with non-missing firm value to sales ratios and GLS COC estimates; and (4) all firms with at least one segment in the financial industry (SIC codes between 6000 and 6999) be excluded from the sample. Second, we require the following data to estimate the GLS COC measure: (1) one- and two-year-ahead earnings forecasts; (2) either a three-year-ahead earnings forecast or the long-term growth earnings forecast and a positive two-year-ahead earnings forecast; and (3) positive book value of equity. The initial sample with available GLS excess cost of capital estimates consists of 38,371 firmyear observations, of which 26,451 (11,920) are single-segment (multi-segment) firms. With additional data requirements for the control variables (discussed in the next section), the final sample consists of 29,080 firm-year observations, of which 19,996 (9,084) observations pertain to single-segment (multi-segment) firms. Some of the sensitivity analyses impose further data restrictions, as discussed in the corresponding sections of the paper. 6 The start of our sample period is driven by our use of pairwise industry correlation estimates based on prior tenyear single-segment data, which start in

14 4.2. Control Variables To ensure that our results on the relation between coinsurance and cost of capital are distinct from the well-documented return patterns (Fama and French (1992) and Jegadeesh and Titman (1993)), we control for size, book-to-market, and momentum as proxied by the log of market capitalization, the book-to-market ratio, and lagged buy-and-hold returns over the past 12 months, respectively. Including a measure of momentum also controls for sluggishness in analyst forecasts. Recent revisions in the stock market s earnings expectations, although immediately reflected in stock prices, may not be incorporated in analyst forecasts on a timely basis, which could induce a negative correlation between past returns and implied COE estimates. 7 Recent research by Hughes, Liu, and Liu (2009) shows that when discount rates are stochastic, implied COE estimates can deviate from expected returns and these deviations can be related to the volatility of, as well as the sample correlation among, expected returns and cash flows, expected growth in cash flows, and leverage. They argue that the resulting measurement error in implied COE estimates may therefore be correlated with variables that are traditionally not associated with systematic risk exposure, explaining the significant correlation between implied COE and leverage, expected earnings growth, and forecast dispersion documented in prior research (Gode and Mohanram (2003)). Therefore, we include these variables as additional controls to avoid spurious results. The timeline of variable measurement is depicted in Figure 1 and the definitions of control variables are summarized below (numbered items refer to the Compustat annual database): 7 It is possible that we are overcontrolling by including size and the book-to-market ratio in our regressions. First, book-to-market may be associated with coinsurance-related forward-looking betas in a conditional asset-pricing model (Petkova and Zhang (2005)). Second, size may serve as an alternative proxy for coinsurance. Larger firms are likely to have a greater number of unrelated projects and thus experience greater coinsurance benefits. 13

15 Log(market capitalization) = Natural logarithm of fiscal year-end stock price times shares outstanding from Compustat (#199*#25); Leverage = Book value of debt divided by the sum of book value of debt and market value of equity from Compustat (#9+#34)/(#9+#34+#199*#25); Book-to-market = Ratio of book value of equity to market value of equity from Compustat (#60/(#199*#25)); Log(forecast dispersion) = Natural logarithm of the standard deviation in analysts one-year-ahead earnings forecasts from I/B/E/S; Long-term growth forecast = Consensus (median) long-term growth forecast from I/B/E/S; Lagged 12-month return = Buy-and-hold stock return from the beginning of June (t-1) until the end of May (t) from CRSP. 5. Empirical Results 5.1. Summary Statistics: Excess Cost of Capital In Table 1, we present summary statistics for excess COC for multi- and single-segment firms separately. For the multi-segment subsample, both mean and median excess COC are negative and significant ( and ). For the single-segment subsample, the median value of excess COC is zero by construction because the imputed values are calculated using the COC of the median single-segment firm in each industry. The mean excess COC is negative ( ) and significant, indicating that the distribution of excess COC is negatively skewed. The difference in means between the multi- and single-segment subsamples is negative (-0.010) and different from zero at better than the 1% level of statistical significance, rejecting the conventional view that organizational form does not matter for a firm s cost of capital. Recall that our excess cost of capital measure is defined as the natural logarithm of the ratio of the firm s COC to its imputed COC based on comparable single-segment firms. Hence, 14

16 when we discuss percentage differences in excess cost of capital, we imply logarithmic percentage differences throughout the paper. Using the estimate above as an example a logarithmic percentage difference of -1% (-0.010) between multi- and single-segment firms the cost of capital of a multi-segment firm would be roughly 9.9% if the cost of capital of a singlesegment firm were 10%. The modest difference between the cost of capital of multi- and single-segment firms is likely due to the pooling of all multi-segment firms, many of which operate within a single industry and enjoy little cross-segment coinsurance. In the next section, we test for differences in cost of capital between multi-segment firms with lower and higher levels of coinsurance Analysis of Excess Cost of Capital and Coinsurance Nonparametric Results In Table 2, we sort our sample of multi-segment firms into quintiles based on crosssegment cash flow and investment correlations (defined in Section 3.2) and report the average excess COC for each quintile in the left and right panels, respectively. Because the results are similar across the two sorts, we focus our discussion on the first sort based on cross-segment cash flow correlations. We also present the results for single-segment firms. Note that singlesegment firms can be viewed as limit observations with respect to the degree of coinsurance for these firms, cash flow and investment correlations equal one by definition. Consistent with the coinsurance hypothesis, we observe a monotonic increase in excess COC from the lowest correlation quintile (Q1) with the most coinsurance to the highest correlation quintile (Q5) with the least coinsurance. The mean difference between Q1 and Q5 is a statistically significant Similarly, the mean difference between the cost of capital of 15

17 multi-segment firms in the lowest correlation quintile (Q1) and single-segment firms is , consistent with a significant coinsurance effect. These results reject the conventional view in favor of the coinsurance hypothesis diversified firms that consist of businesses with less correlated cash flows have a lower cost of capital Main Regression Results Next, we investigate whether the nonparametric evidence in Table 2 is robust to controlling for the set of firm characteristics discussed in Section 4.2. The results of this analysis are presented in Table 3, with robust standard errors (heteroskedasticity consistent and double clustered by firm and year (Petersen (2009)) reported in brackets below corresponding coefficients. In the first set of regressions, Models 1 and 2, we regress excess COC on cross-segment cash flow and investment correlations, respectively, and control for all variables except for the number of segments and the natural logarithm of market capitalization. We exclude these two measures because they likely capture some degree of coinsurance. Larger firms or firms with more segments are more likely to have business units with imperfectly correlated cash flows. Therefore, including them in the regressions could overcontrol for the coinsurance effect. In the second set of regressions, Models 3 and 4, we use the number of segments and the natural logarithm of market capitalization, respectively, as alternative measures of coinsurance. In the last set of specifications, Models 5 and 6, we include all control variables, including the number of segments and the natural logarithm of market capitalization, to disentangle other possible size effects from the coinsurance effect that is captured by cash flow and investment correlations. We therefore view this last set of specifications as the most demanding test of the coinsurance hypothesis. 16

18 Because the results for cash flow and investment correlations are similar in all specifications, we focus our discussion on cash flow correlations. Consistent with the nonparametric results, the coefficient estimate on cross-segment cash flow correlations in Model 1 is positive (0.056) and different from zero at the 1% level of statistical significance. In Models 3 and 4, we find negative and significant coefficients on the number of segments and the natural logarithm of market capitalization. As noted earlier, while these results are consistent with the coinsurance hypothesis, it is difficult to attribute them solely to the coinsurance effect as firm size and the number of segments may also proxy for other factors. Finally in Model 5, the coefficient estimate on cross-segment cash flow correlations is positive (0.053) and different from zero at the 1% level of statistical significance. The coefficient estimate is also similar to that in Model 1, suggesting that cross-segment correlations provide a source of variation in coinsurance that is relatively independent from that provided by firm size and number of segments. Overall, our results strongly reject the conventional view in favor of the coinsurance hypothesis. Firms with lower cross-segment cash flow correlations and hence greater coinsurance potential have a lower cost of capital Financial Slack As discussed in Section 2.1, financial slack can serve as an alternative mechanism for firms to limit or avoid deadweight costs. Thus, one would expect financial slack to reduce the benefit of coinsurance and its effect on cost of capital. We test this prediction in Table 4 using two measures of financial slack, net leverage (an inverse measure of slack) and cash holdings. The main coefficient of interest is the interaction term between cross-segment correlations and financial slack. Models 1 and 2 report the results for net leverage quintile rank, which ranges from 1 for firms in the lowest net leverage quintile (most financial slack) to 17

19 5 for firms in the highest net leverage quintile (least financial slack) in a given year. Models 3 and 4 report the results for cash quintile rank, which ranges from 1 for firms in the lowest cash quintile (least financial slack) to 5 for firms in the highest cash quintile (most financial slack) in a given year. The coefficient estimates on the interaction between cross-segment correlations and net leverage quintile rank are positive (0.024 in both Models 1 and 2) and significant at the 1% level. In addition, the coefficient estimates suggest that the coinsurance effect is approximately zero for firms in the lowest net leverage quintile ( and in Models 1 and 2, respectively), but rather substantial for firms in the highest net leverage quintile (0.091 and in Models 1 and 2, respectively). In Models 3 and 4, the coefficient estimates on the interaction between cross-segment correlations and cash quintile rank are negative ( in both Models 3 and 4) and significant at the 1% level, consistent with a weaker coinsurance effect for firms with higher cash holdings. These results suggest that the coinsurance effect is weaker for firms with more financial slack (lower net leverage or higher cash holdings), consistent with the prediction that coinsurance benefits are smaller for these firms Controlling for Selection Effects Our estimates of the coinsurance effect might be biased due to selection effects arising from firms decisions to diversify, an issue that has been addressed extensively in the diversification discount literature. However, it is unclear how a strong monotonic relation between our continuous coinsurance measures and excess cost of capital would be driven by a dichotomous selection mechanism that pushes some business units to conglomerate. In addition, one might think a priori that high-risk business units, which have the most to gain from coinsurance, are more likely to diversify than low-risk business units, in which case the selection bias would work against us finding a coinsurance effect. 18

20 Nevertheless, we acknowledge that selection is an important concern and we address this issue in two ways. First, we estimate Heckman two-stage regressions to correct for potential selection biases. Second, we follow an approach that is similar in spirit to that of Lamont and Polk (2002) and examine the relation between exogenous changes in coinsurance and changes in excess COC. As we describe below, the results from both analyses suggest that our estimates of the coinsurance effect in Table 3 are unlikely to be contaminated by selection effects or firms decisions to diversify. Heckman s Two-Stage Analysis To control for potential selection biases using Heckman s two-stage procedure, we first estimate a first-stage probit model for firms decisions to diversify. The dependent variable in the probit model is equal to 1 for a multi-segment firm and 0 for a single-segment firm. We estimate two different first-stage models. The first model ( No Instrument ) includes all of the control variables in our main regression model. The second model ( With Instruments ) further includes two instruments used in Campa and Kedia (2002), namely, PNDIV (the fraction of all firms in the industry which are conglomerates) and PSDIV (the fraction of sales accounted for by conglomerates). The second-stage regressions control for the inverse Mills ratio estimated from these two first-stage models. The results of the second-stage regressions are reported in Panel A of Table 5. The first two columns present the results using the inverse Mills ratio from the No Instrument first-stage probit model whereas the last two columns present the results using the inverse Mills ratio from the With Instruments first-stage probit model. In all models, the estimated coefficients on cross-segment correlations are positive and different from zero at the 1% level of statistical significance. Importantly, the magnitudes of the coefficients are similar to those reported in 19

21 Table 3. Further consistent with the absence of significant selection effects, the estimated coefficients on the Mills ratio are small and only marginally significant in the first two columns. Exogenous Changes in Coinsurance and Changes in Excess COC We also follow an approach that is similar to that of Lamont and Polk (2002) to address the issue of selection effects. Specifically, we decompose changes in cross-segment correlations into two components: an exogenous component that reflects changes in pairwise industry correlations that are arguably outside the control of managers, and an endogenous component that reflects changes in firm segment structure that managers can control. Specifically: ( s, c ) ( s, c ) t t t t 1 t 1 ( st, ct) ( st 1, ct) ( st 1, ct) ( st 1, ct 1) endogenous change in exogenous change in where s t and c t represent the firm s segment structure and estimates of pairwise industry correlations in year t, respectively. Next, we regress changes in excess COC on exogenous and endogenous changes in crosssegment correlations as well as changes in the control variables from Table 3. The results are reported in Panel B of Table 5. Models 1 and 3 are analogous to the main regression models in Table 3 (Models 5 and 6), but in a first-differenced form, which effectively controls for firm fixed effects. Models 2 and 4 decompose total changes in cross-segment correlations into exogenous and endogenous changes. Similar to the main regression results in Table 3, the coefficient estimates on crosssegment correlations in Models 1 and 3 are both positive and significant. More importantly, in Models 2 and 4, the coefficient estimates on exogenous changes in cross-segment correlations are also positive and significant, and they are similar in magnitude to those in Table 3. 20

22 It is worth noting that while our main focus is on exogenous changes in cross-segment correlations, endogenous changes are also of interest as a firm s cost of capital should change in response to changes in its organizational structure. Consistent with this prediction, the coefficient estimates on endogenous changes in cross-segment correlations are also positive, although significant only for cross-segment cash flow correlations in Model Economic Significance To evaluate the economic significance of our findings, we estimate the effect of coinsurance-related reduction in cost of capital on firm value. In the simple Gordon growth model, under a zero dividend growth assumption, a 1% decrease in cost of capital from 10% to 9.9% approximately translates into a 1% increase in firm value. However, the relation between cost of capital and firm value is in general nonlinear and depends on other inputs in the valuation formula expected earnings and earnings growth. Our analysis compares actual firm values to as-if firm values calculated using imputed cost of capital (i.e., the cost of capital on a comparable portfolio of single-segment firms) while holding cash flows constant in the GLS valuation model (described in the Appendix). The excess value attributable to differences in cost of capital is calculated as the natural logarithm of the ratio of actual firm value to as-if firm value. Using this approach, we find an economically significant 5.5% (5.4%) average gain in total firm value when moving from the lowest to the highest coinsurance quintile based on crosssegment cash flow (investment) correlations. We note that these estimates might represent a lower bound for the coinsurance effect on firm value because our proxies are limited to segment data and do not capture coinsurance among different product lines or geographic areas. 21

23 5.3. Robustness Tests Excluding Single-segment Firms To investigate the possibility that our results may be spuriously driven by differences between single- and multi-segment firms, we perform our main analysis using multi-segment firms only. The results, reported in Table 6, are qualitatively and statistically similar to those reported in Models 5 and 6 of Table 3. In particular, the coefficient estimates on cross-segment cash flow and investment correlations are both positive and significant at the 1% level. Alternative Measures of Coinsurance Our main coinsurance proxy is constructed using segment membership in 2-digit SIC industries, and we estimate idiosyncratic industry cash flows and investments using a singlefactor market model. We now check whether our results are robust to using: 1) more refined industry definitions, 2) a multi-factor (as opposed to a single-factor market) model to estimate idiosyncratic industry cash flows and investment, and 3) firm-specific segment data (as opposed to industry-level data) to compute cross-segment cash flow and investment correlations. First, we repeat our main analysis using three alternative coinsurance measures based on the following industry definitions: Fama and French (1997) 48 industries ( FF industries ), 3- digit SIC codes ( 3-digit SIC industries ), and the narrowest SIC grouping that includes at least five single-segment firms with at least $20 million in sales over the last ten years ( Narrowest SIC industries ). The results are presented in Panel A of Table 7. The coefficient estimates on cross-segment cash flow and investment correlations are positive and significant in all regressions, with the exception of the 3-digit SIC industries regression where the coefficient estimate on cash flow correlations is positive but insignificant. The weak statistical significance appears to be driven by some 3-digit SIC industries having very few firms in some years, 22

24 resulting in relatively noisy industry cash flows. This problem is alleviated when we adopt the narrowest SIC industry approach, which imposes a minimum of five single-segment firms. Next, we investigate whether our results are robust to a multi-factor model that includes two additional Fama and French (1995) factors size and book-to-market for estimating idiosyncratic industry cash flow and investment series. The results based on this alternative coinsurance measure ( FF multi-factor earnings ), reported in the first two columns of Panel B of Table 7, are qualitatively and statistically similar to those reported in Table 3. Finally, we perform our main analysis using two alternative coinsurance measures based on firm-specific segment cash flow and investment data. In order to provide a reasonable period for estimating cross-segment correlations, the analysis is performed using a subset of firms whose segment structures remain unchanged for 5 or 7 years ( 5-year sample and 7-year sample, respectively). The results are reported in the last four columns of Panel B of Table 7. As expected, a significant drawback of this approach is a small sample size 13,191 (9,847) firm-year observations in the 5-year (7-year) sample compared to 29,080 firm-year observations in our main analysis. Despite the substantial drop in sample size, the estimated coefficients on cross-segment correlations remain positive and statistically significant, with magnitudes similar to those reported in Table 3. Analyst Forecast Errors A potential limitation of implied cost of equity measures is measurement errors arising from biases in analyst forecasts. We use two approaches to address this concern. First, we control for one- and two-year-ahead unexpected and expected forecast errors in our main regression models. In particular, we follow Ogneva, Subramanyam, and Raghunandan (2007) and estimate expected forecast errors using the prediction model in Liu and Su (2005). Our parsimonious 23

25 version of the model includes the following predictors that proxy for systematic biases in analyst forecasts: (1) past stock returns, (2) recent analyst earnings forecast revisions, and variables related to overreaction to past information, namely, (3) forward earnings-to-price ratios, (4) longterm growth forecasts, and (5) investments in property, plant, and equipment. Estimation of the predicted forecast error is performed separately for one- and two-year-ahead forecast errors. Unexpected forecast errors are computed as the difference between realized errors and their predicted component. Because one- and two-year-ahead expected errors are highly collinear, we use the average expected errors over the two years as the control measure. The results, reported in Panel A of Table 8, continue to show a positive and significant coefficient on cross-segment cash flow and investment correlations, suggesting that our main findings are unlikely driven by systematic differences in analyst forecast biases between single- and multi-segment firms. Second, Easton and Monahan (2006) find that the reliability of implied cost of equity estimates increases as analyst forecast accuracy improves. Accordingly, we partition our sample into terciles using absolute forecast errors in one-year-ahead earnings and estimate cost of capital regressions within each subsample. The results are reported in Panel B of Table 8. The coinsurance effect is weakest in the subsample with high absolute forecast errors. These results suggest that our findings are unlikely driven by measurement errors in the implied cost of equity estimates that are induced by biased forecasts. Rather, our results are weakened by them. Alternative Measures of Implied Cost of Equity We also repeat our main analysis using two alternative implied cost of equity measures, CT (Claus and Thomas (2001)) and PEG (Easton (2004)), which differ from the Gebhardt, Lee and Swaminathan (2001) measure (GLS) used in our main analysis in terms of the assumptions regarding terminal value computation and the form of the valuation model. (See the Appendix 24

26 for details on both measures estimation). The results based on these two alternative implied cost of equity measures are reported in Table 9. Consistent with our earlier findings, the coefficient estimates on cross-segment cash flow and investment correlations are positive and significant. Alternative Measures of Cost of Equity In this subsection, we perform our main analysis using three alternative proxies for cost of equity that do not rely on implied cost of equity estimations and hence analyst forecasts: 1) expected returns from the Fama-French three-factor model (hereafter, FF), 2) realized annual buy-and-hold returns, and 3) earnings yield. The first two measures have been widely used as proxies for expected returns in the literature. 8 The last measure, earnings yield (E/P), while admittedly a crude proxy for cost of equity, provides a simple robustness check; it is computed as the ratio of net income to beginning of the year market value of equity. 9 Because earnings yield also contains information about growth opportunities, we include a second measure, E/P growth-adjusted, that incorporates the effect of earnings growth; it is calculated as the sum of earnings yield and growth in net income over the previous year. Excess FF (E/P and E/P growth-adjusted) is the logarithm of the ratio of the firm s FF (E/P and E/P growth-adjusted) to its imputed FF (E/P and E/P growth-adjusted). Excess realized return is a simple difference between the firm s realized return and its imputed realized return. The derivation of the corresponding imputed values is similar to that of our main cost of equity measure (see Section 3.1 for details). 8 To calculate FF expected returns, we estimate factor loadings using 24 months of prior excess returns, multiply the loadings with corresponding historical risk premiums, and add the yield on the 10-year Treasury note. We exclude observations with negative FF cost of equity estimates from the analysis (about 8% of our sample) Realized returns are 12-month buy-and-hold returns starting in June of year t+1 (see Figure 1 for timing convention). 9 Earnings yield is calculated only for observations with positive net income. 25

27 We repeat our main analysis using these three alternative excess cost of equity proxies alongside excess GLS implied cost of equity as a benchmark. The results are reported in Table 10. In the FF and realized return regression models, we include the same set of control variables as in our main regressions, with the exception of forecast dispersion and long-term growth forecast because these two variables are meant to control for measurement error when implied cost of capital measures are used as the dependent variable. For the same reason, we also exclude forecast dispersion in the E/P and E/P growth-adjusted regressions, but include long-term growth forecast to control for expected earnings growth in the E/P model. With the exception of realized returns, all coefficient estimates on cross-segment cash flow and investment correlations are significantly positive, with magnitudes comparable to those in the benchmark specifications using excess GLS implied cost of equity. While these results show that our coinsurance findings are not robust to using realized returns, the other proxies that are less susceptible to cash flow shocks (in particular, FF expected returns) provide further support for our ex ante approach to measuring expected returns. As pointed out by Elton (1999), ex post realized returns can be noisy proxies for ex ante expected returns and may lead to biased coefficient estimates in finite samples due to contamination by cash flow shocks. Several recent papers (Campello, Chen, and Zhang (2008) and Chava and Purnanandam (2010)) show that these biases can be substantial, and our analysis bears out a similar conclusion. Excess Cost of Debt The model outlined in Section 2.1 predicts coinsurance effects for both equity and debt. The results for excess GLS implied cost of equity in Table 10 show that coinsurance benefits are recognized by equityholders. In this subsection, we examine whether they are also recognized by debtholders. 26