Dissecting Conglomerates November 18, 2017 ABSTRACT We develop a new method to estimate Tobin s qs of conglomerate divisions without relying on standalone firms. Divisional qs differ considerably from qs of standalone firms and, consistent with theories of corporate diversification, are less volatile and less sensitive to macroeconomic shocks. In contrast to prior evidence based on standalone qs, divisional investment is highly sensitive to divisional qs. This sensitivity disappears after a division is spun-off into a standalone firm. Moreover, divisional qs predict both acquisition announcement returns and the volume of diversifying acquisitions. Overall, our estimates provide new opportunities to study conglomeration and internal capital allocation. JEL Classification: G32, G34. Keywords: boundaries of the firm, segment valuation, conglomerate investment, internal capital markets, quantile regressions.
Most large US firms operate in multiple divisions. Such conglomerates account for the lion s share of financing, investment, and value of US companies. Consequently, a vast literature studies conglomerates, focusing on conglomerate investment allocation and the value of conglomeration. It is notoriously challenging to investigate these empirically, however, because direct estimates of divisional valuations and investment opportunities are not available. In this paper, we address this challenge by developing a new method to estimate divisional valuation multiples, and use our estimates to study conglomerate investment and diversifying acquisitions. To understand our method, consider the traditional approach to studying conglomerates, pioneered by Lang and Stulz (1994) and Berger and Ofek (1995), which synthetically replicates the overall conglomerate by a portfolio of standalone firms. Specifically, this approach imputes v, the vector of valuations of diversified firms, as ˆv = W q sa, where W is a matrix of firms industry exposures of fundamentals (e.g., sales), and q sa is a vector of industry multiples derived from standalone firms. By using q sa, this approach assumes that industry multiples of standalone firms accurately proxy for those of conglomerate divisions. Thus, it does not account for systematic differences between conglomerates and standalone firms that may arise due to the endogeneity of firms organizational structure. Whited (2001) shows that estimates of Tobin s q of standalone firms are inappropriate for the study of investment by conglomerate divisions. We overcome this limitation by forming portfolios of conglomerates to mimic standalone firms. Specifically, we use data on v and W to directly estimate a vector of conglomerate-implied industry valuation multiples. As a simple example, consider two conglomerates that operate in the same two industries. Assume that the first conglomerate has unit exposure of fundamentals to each industry, and the second conglomerate has exposures of two and one to these industries. It is immediately clear that a portfolio long the second conglomerate and short the first conglomerate has industry exposures of one and zero. Thus, the value of this portfolio equals the conglomerate-implied valuation of divisions operating in the first industry, and provides a proxy of Tobin s q for 1
these divisions. In this example, the solution to the problem is unique and can be obtained by inverting the matrix containing the proportions of conglomerate exposures to each industry. For practical applications, the number of conglomerates exceeds the number of industries, and this matrix is not invertible. We show in Monte-Carlo simulations that median regressions of conglomerate valuation multiples on the matrix of exposure weights provide reliable estimates of the multiples of conglomerate divisions. Overall, this approach generates model-free estimates of conglomerate divisions qs that vary across industries and over time, and unlike traditional measures, are obtained without using industry multiples of standalone firms. The estimates therefore allow us to investigate the systematic differences between conglomerate divisions and standalone firms that operate in similar industries, and to shed new light on theories of corporate diversification and the efficiency of conglomerate investment. In our first set of analyses, we investigate the properties of divisional qs and compare them to traditional measures derived from standalone firms. We find large differences between these multiples. In the cross-section of industries, the average ratio of divisional and standalone qs varies considerably, ranging from -55% to 17%. Relative to standalone firms, divisional qs are lowest in the energy, high-tech, and healthcare industries, whereas divisions in the consumer nondurables and telecommunication industries have higher qs than their standalone counterparts. In the time series, the qs of conglomerate divisions are less volatile and less sensitive to macroeconomic shocks than those of standalone firms. The standard deviation of divisional qs is on average 15% lower than that of standalone qs. The sensitivity of divisional qs to different macroeconomic shocks, such as stock market-wide movements and productivity shocks, is a staggering 48% lower. These results suggest that conglomerate divisions are more insulated from external economic forces than standalone firms operating in the same industry, consistent with theories of corporate diversification (e.g., Lewellen, 1971, Matsusaka and Nanda, 2002). 2
Overall, these findings have broad implications for prior research on conglomerates, as they reveal systematic differences between qs of conglomerate divisions and qs of standalone firms. Consequently, analyses that rely on qs of standalone firms, such as the study of internal capital allocation, may mis-characterize intra-conglomerate processes and their efficacy. Moreover, conclusions from industry-specific analyses need not extrapolate to other industries, where the relation between divisional and standalone qs can differ considerably. 1 Having established that divisional qs differ considerably from traditional measures of q that rely on standalone firms, in the second set of analyses we employ these measures to study conglomerate investment. These tests are motivated by existing theory, which offers diverging views on the efficacy of capital allocation inside conglomerates. On the one hand, internal capital markets in conglomerates may allow raising more external finance (Lewellen, 1971, Hadlock, Ryngaert, and Thomas, 2001) and allocating capital more efficiently (Alchian, 1969, Weston, 1970, Stein, 1997, Matsusaka and Nanda, 2002). On the other hand, conglomerates may suffer from agency problems and, in particular, from the rent-seeking behavior of divisional managers. Consequently, the ceo may distort the conglomerate s internal capital allocation toward weaker divisions to retain divisional managers (Scharfstein and Stein, 2000) and to control their rentseeking behavior (Rajan, Servaes, and Zingales, 2000). Our empirical approach follows the analyses in Shin and Stulz (1998) and Ozbas and Scharfstein (2010). These studies use the neoclassical relation between investment and Tobin s q to investigate the efficacy of conglomerates internal capital allocation. Using estimates of qs derived from standalone firms, they find that investment of conglomerate divisions is less sensitive to investment opportunities than that of standalone firms. Our findings confirm that conglomerate investment is indeed less sensitive to qs estimated from standalone firms. Specifically, conglomerate divisions elasticity of in- 1 Examples of studies that focus on a single industry include Lamont (1997), who analyzes investment decisions of diversified oil companies following the oil price shock of 1986; Khanna and Tice (2001), who investigate the responses of diversified firms to Wal-Mart s entry into their market; Campello (2002), who studies the reactions of financial conglomerates to monetary policy; and Guedj and Scharfstein (2004), who analyze the development strategies and performance of biopharmaceutical firms. 3
vestment to standalone q is insignificant at 0.6%. This estimate is significantly lower than the elasticity of standalone firms (5.5%). However, conglomerate investment is highly sensitive to divisional qs: the elasticity of conglomerate divisions investment to divisional q is significantly positive at 2.6%, whereas the estimate for standalone firms is insignificant at -1.3%. These findings are less consistent with the hypothesis that conglomerates invest inefficiently, and suggest that our method generates clean measures of investment opportunities of conglomerate divisions. To shed light on the operation of internal capital markets, we also investigate the sensitivity of divisional investments to divisional cash flows. When we use qs estimated from standalone firms, our results confirm those of Shin and Stulz (1998). First, a division s investment is more sensitive to its own cash flow than to the other divisions cash flows. Second, the sensitivity of a division s investment to other divisions cash flows is not lower when it has better investment opportunities. These findings were traditionally interpreted as evidence against efficient internal capital markets. However, when we introduce the new measures of divisional qs into the analysis, we find strikingly different results. A division s investment becomes more sensitive to other divisions cash flows than its own cash flow. Further, the sensitivity to other divisions cash flows is lower when investment opportunities are higher. This evidence suggests that capital markets inside conglomerates facilitate the reallocation of resources across divisions towards divisions with high marginal products. A possible concern is that the sensitivity of investment to divisional qs is correlated with potentially unobservable firm-level characteristics. To address this concern, we estimate the elasticity of investment to standalone and divisional qs around the spinoff of conglomerate divisions. This analysis holds constant the firm and traces its investment policy through the organizational changes that it undergoes, thereby minimizing the confounding effects of potentially unobservable differences across firms. The results suggest that prior to the spinoff, divisional investment is highly sensitive to divisional q and insensitive to standalone q. Following the spinoff, however, invest- 4
ment becomes insensitive to divisional q and highly sensitive to standalone q. These results suggest that while the investment opportunities of conglomerates and standalone firms are different, investment efficiency at conglomerates and standalone firms is not. In our final set of analyses, we use divisional qs to study mergers and acquisitions (m&a) activity and the origins of conglomerate value. First, we investigate acquisition announcement returns. We find that the difference between divisional qs and standalone qs, which captures the implied value gains due to the change in the organizational structure, is a strong predictor of acquisition announcement returns. These results hold for target, acquirer, and combined announcement returns, and are economically important. For example, in the full sample of m&a announcements, which includes both focused and diversifying acquisitions, an increase of one standard deviation in the difference between standalone and divisional qs corresponds to an increase of 1.15% in the combined announcement returns. The economic magnitude increases by more than 50% when we restrict the sample to acquisitions that diversify across two of the 10 Fama-French industries. This last result indicates that divisional qs are strong indicators of the value of combining diversified divisions into a conglomerate. Second, we investigate the volume of m&a activity. Our findings indicate that for the full sample of m&a announcements, both divisional and standalone qs predict the volume of activity. In contrast, only divisional qs predict the volume of diversifying acquisitions. This finding provides further evidence that divisional qs capture the value of diversifying. It also suggests that, on average, firms make optimal diversifying acquisitions that exploit high valuations of corporate diversification. A possible concern with these findings is that they seem hard to reconcile with the diversification discount documented by Lang and Stulz (1994), Berger and Ofek (1995), and the literature that followed. However, studies such as Campa and Kedia (2002), Graham, Lemmon, and Wolf (2002), and Villalonga (2004b) show that after accounting for firms endogenous decision to diversify, conglomerates no longer appear discounted. Our findings support this view by providing direct evidence on the value 5
and volume of diversifying acquisitions. Consistent with these findings, Matsusaka (2001) develops a dynamic search model in which diversification is an optimal search process by which firms seek to acquire businesses that are good matches for their capabilities. In his model, diversified firms are discounted because they choose to diversify when their value decreases and not because they make bad diversification decisions. To test these predictions, we compare the divisional qs of conglomerates core and peripheral divisions. We find that the qs of core divisions are deeply discounted compared to the qs of peripheral divisions. These results suggest that lower conglomerate qs do not arise solely from diversifying into low-q peripheral divisions. Instead, discounted firms are those that choose to diversify. Overall, our paper contributes to the literature on corporate diversification and internal capital markets. It makes a step towards a better understanding of conglomerate investment and the efficacy of diversification decisions by providing clean disaggregated estimates of division-level valuation multiples. I. The New Method Our method inverts the traditional approach of Lang and Stulz (1994) and Berger and Ofek (1995). Rather than building up synthetic conglomerates from individual pieces (standalone firms), we break down actual conglomerates into their components. We group these components across conglomerates into classes that share observable characteristics, such as industry affiliation. We then use quantile regressions to obtain median conglomerate-implied estimates for the valuation multiples, or qs, of each class. Comparing these qs with those of standalone firms in each class allows us to analyze the systematic differences between conglomerates and standalone firms at the granular level of a single class. A. Overview of the Estimation Method Let W denote the matrix of I conglomerates by K classes that contains the fundamentals (e.g., sales) of a cross-section of conglomerates. For example, in the analysis of 6
Berger and Ofek (1995), K represents the number of 4-digit sic industries in which the I conglomerates operate. If divisions in all classes function independently, the value of the conglomerate, v, should equal the weighted sum of the valuation ratios of the classes represented by the K 1 vector q c : v = W q c. (1) In the special case where W contains the replacement costs of capital, q c corresponds to Tobin s q associated with each class. Because replacement costs are not observable, the corporate diversification literature relies on asset or sales multiples to proxy for Tobin s q (Lang and Stulz, 1994). The traditional approach of Lang and Stulz (1994) and Berger and Ofek (1995) imputes the value of conglomerates using industry-level qs estimated from standalone firms. In particular, they estimate the qs of each division as the median q of standalone firms operating in the same industry, ˆq c = q sa. A standard result in this literature is that the imputed values of conglomerates, ˆv = W q sa, on average exceed their market capitalization, suggesting that diversified firms are valued at a discount. Our method also builds on Equation (1), but we aim to estimate q c using only conglomerate-level information. In particular, we scale Equation (1) by the total fundamentals of each conglomerate to obtain ṽ = W q c, (2) where ṽ(i) = v(i)/ k W (i, k) and W (i, k) = W (i, k)/ k W (i, k) are valuation multiples and class weights of the conglomerates. Ostensibly, estimating ˆq c from Equation (2) could be achieved via an ordinary least squares (ols) regression of conglomerate multiples on class weights. However, the ols approach is problematic since valuation ratios are positively skewed. The prior literature addresses the skewness in valuation ratios by taking their natural logarithms, an approach not suitable for our purposes since logs are not additive. To resolve the problem of skewed valuation ratios, we base our analysis on medians rather than means, 7
and use quantile regressions. Specifically, we run median regressions of conglomerate multiples ṽ on class weights W cross-sectionally to back out the median class valuation ratios ˆq c. B. Quantile Regressions Before describing our analysis in detail, we provide a short review of quantile regressions. Our goal is to fit the median of the target variable y i conditional on the explanatory variables X i. When estimating Equation (2), y i corresponds to the valuation ratio of conglomerate i, ṽ(i), and X i is the ith row of the weight matrix W. The median, or 50th percentile, of y i is defined from its inverse probability distribution function P 50 (y i ) = inf {y : P rob (y i < y) 0.50}. (3) We can express the median as the solution to an optimization problem P 50 (y i ) = arg inf u E y i u, (4) which is particularly convenient for handling conditioning information sets such as the explanatory variables. 2 We follow the seminal quantile regression specification of Koenker and Bassett (1978), and assume that the median of y i conditional on X i is a linear function of the explanatory variables. This implies P 50 ( y i Xi ) = arg inf u E ( y i u Xi ) = γ0 + γ 1 X i. (5) The assumed linear relation is reminiscent of standard ols specifications. However, median regressions model the conditional median of y i, rather than its mean, as a linear function of X i. C. Monte Carlo Simulations We use Monte Carlo simulations to show that the qs of division classes can be robustly estimated using median regressions rather than ols regressions if conglomerate valuation ratios are positively skewed (e.g., Berger and Ofek, 1995). Specifically, we simulate 2 Equation (4) is a special case of the general quantile regression representation, where the quantile loss function for quantile τ is given by ρ τ (x) = x ( τ I (x<0) ) and the optimization problem is P τ (y i) = arg inf u E [ρ τ (y i u)]. 8
fundamentals (i.e., sales) across K = 5 classes for I = 500 conglomerates. Half of the conglomerates operate in two classes, a third in three classes, and a sixth in four classes, approximately in line with the empirical distribution of industries. 3 Fundamentals W are drawn from a lognormal distribution that is based on a Gaussian distribution with unit mean and a standard deviation of 0.8. 4 The valuation ratios of conglomerates, ṽ, are calculated as in Equation (2), where the class valuations are given by q c = [0.5 1.0 1.5 2.0 2.5]. These class valuation approximately reflect standalone qs of the 10 Fama-French industries, which range from 0.50 to 3.08. Conglomerates are then exposed to multiplicative valuation shocks that have a median of one and are drawn from either a Gaussian distribution that is truncated at zero or a lognormal distribution. These shocks reflect the significant empirical variation in excess values documented, for example, by Lamont and Polk (2001) who report cross-sectional standard deviations of excess values between 0.36 and 0.63. shocks. Correspondingly, we consider 0.3 and 0.6 as standard deviations for our Table I shows averages and standard deviations of the differences between estimated and actual class valuations, ˆq c q c, across 100,000 simulations. In the last column, it also reports excess value measures computed as in Berger and Ofek (1995). Panels A and B correspond to multiplicative shocks with standard deviations of 0.3 and 0.6, respectively. When conglomerate valuation ratios are normally distributed, and therefore nonskewed, both median regressions and OLS regressions yield unbiased estimates of class valuations. Specifically, ˆq c q c = 0 across all columns for both median and OLS regressions. The cross-simulation averages of the median excess value are zero, but those of its mean are negative (-0.05 in Panel A and -0.18 in Panel B). This is because 3 In our sample of XXX conglomerate-years, XXX operate in two industries, XXX operate in three industries, and XXX operate in four industries. 4 The standard deviation of 0.8 implies that, on average, conglomerates are well diversified and not dominated by a single division. The average Herfindahl index across conglomerates is about 0.5. It ranges from 0.35 for four-division firms to 0.6 for two-division firms, closely matching the empirical moments we obtain in untabulated analysis. 9
the use of the logarithm in the excess value calculation directly aims to eliminate the effects of positive skewness. With a normally distributed shock, that is, without positive skewness, Jensen s inequality implies a downward-biased measure of excess value. Several observations about the cross-simulation standard deviations, shown in parentheses, are noteworthy. First, they are increasing across the five classes. This is simply an artifact of having increasing valuation ratios across the classes, and a multiplicative valuation shock. Second, as expected with normally distributed residuals, ols is more efficient than median regressions. Lastly, the standard deviations of class valuations are significantly higher than those of excess values. This is not surprising given that the same data are used to obtain one estimate of excess value but five estimates of class valuation multiples. The finer granularity comes at a cost of reduced efficiency. The ols-based inferences change dramatically when valuation ratios are positively skewed. With lognormal shocks, ols yields strongly upward-biased estimates of class valuations. In Panel A these estimates range from 0.02 to 0.11, representing about 4% of the true valuation multiple. The drawbacks of ols regressions become even more pronounced in Panel B, where we assume a higher cross-sectional variation of 0.6 in excess values. In this case, the bias reaches 20%. Since valuation ratios are known to be positively skewed, our simulation evidence strongly suggests that ols should not be used to obtain class valuation multiples. In contrast, estimates from median regressions are unbiased and, like excess values, remain robust to different distributional assumptions. Overall, the simulations show that valuation ratios of classes can be estimated well using quantile regressions on the median, the approach we follow throughout the paper. II. Data and Empirical Approach A. Data We obtain firm-level accounting variables and sic industry classifications from Compustat, and division-level variables from the Compustat Segment files. Our sample period 10
starts in 1978, when Compustat segment data become available, and ends in 2015. Following the literature, we exclude firms with at least one division in the financial sector, (sic codes 6000-6999), in agriculture (sic codes below 1000), and in government, other non-economic activities, or unclassified services (sic codes 8600, 8800, 8900, and 9000). Since we are interested in studying division investment opportunities, we exclude firms whose sales or assets at the level of business segments are unavailable on Compustat. We also exclude divisions with zero sales, such as corporate accounts. Following the literature, we further require total sales from the Compustat annual files to be greater than $20 million and within one percent of the sum of division sales. We follow Custodio s (2014) suggestion and focus our analysis on sales-based multiples. While earlier literature also considers assets-based multiples, Custodio (2014) offers a caveat by noting that the accumulation of goodwill in merger and acquisition accounting biases the book value of assets of conglomerates upwards, and that conglomerates have more flexibility in allocating assets across divisions. Consequently, we base our analysis on sales multiples. We define a conglomerate as a firm that operates in at least two distinct industries. The literature has frequently classified industries by 4-digit sic codes. A shortcoming of this definition is that the number of conglomerate divisions that operate in each industry in a given year is small, with a median number of divisions per industry-year of just two. Therefore, our subsequent empirical analyses rely on the 10 Fama-French industries, which are sufficiently coarse to ensure that a meaningful number of divisions operate in each industry in a given year, and thus that the matrix of industry weights for conglomerates, W, is well populated and has full column rank. 5 A benefit of this coarser classification is that our analysis is less exposed to the criticism of Villalonga (2004a), who highlights problems in the 4-digit sic assignment by Compustat. While a substantial number of divisions may be misclassified into wrong 4-digit sic codes, they likely remain in the same 10 Fama-French industry. 5 The industries are defined in Appendix A. We also consider the Fama-French five, seventeen, and thirty industries. A finer classification leads to missing observations for some industry-years but produces qualitatively similar results. 11
Table XXX reports summary statistics for the sample. Overall, the sample includes 2,512 conglomerates (12,521 firm-year observations) and 6,279 divisions (28,701 division-year observations). In this sample, an average (median) conglomerate has annual sales of $3.98 ($0.71) billion CPI-adjusted 2004 dollars, owns book assets valued at $4.20 ($0.70) billion, has a Tobin s q of 1.11 (0.93) based on asset multiples and of 1.39 (0.94) based on sales multiples, operates in 2.29 (2) business segments, and has annual capital expenditures of 9.13 (4.31) percent of sales. An average (median) division has annual sales of $1.77 ($0.23) billion and owns book assets valued at $1.88 ($0.24) billion. Requiring that conglomerates operate in two distinct industries based on the 10 Fama-French industry classification reduces the number of conglomerates by about a third relative to the 4-digit sic-based definition. However, the characteristics of this smaller sample are consistent with those reported in the literature (e.g., Custodio, 2014). Moreover, the average excess values are also similar across the two samples (-13.3% vs. -15.5%). B. Empirical Approach In this section, we describe the empirical approach to estimating divisional qs. Every year, we perform cross-sectional quantile regressions, without intercept, of conglomerate qs onto their sales exposure to the 10 Fama-French industries. The resulting coefficients are estimates of the median q of conglomerate divisions in each industry for the year. The number of divisions in each industry range from 27 (Telecommunications) to 176 (Manufacturing) on average, which ensures that each regression coefficient is well identified. Our approach generates direct estimates of conglomerate divisions qs that do not rely on the industry qs of standalone firms. It therefore eliminates measurement errors that may arise due to the systematic differences between conglomerates and standalone firms. In fact, by generating independent conglomerate qs, it allows us to characterize the differences between the qs of conglomerate divisions and those of standalone firms. 12
While our method eliminates biases due to differences in valuations between conglomerates and standalone firms, a remaining identifying assumption is that conglomerate divisions in each of the K classes have common valuation ratios. We test this assumption directly in Figure 1 by comparing the excess values implied by our new method to those from the traditional approach of Berger and Ofek (1995). If our identifying assumption holds, and if our class-by-class estimates of q are indeed better estimates of divisional qs, then they should more accurately approximate the overall value of the conglomerate, resulting in more accurate estimates of conglomerate excess values. Figure 1 investigates whether this is the case by illustrating the distribution of standard deviations of two excess value estimates. For each firm and year, we follow Berger and Ofek (1995) and define excess value as the log-difference between the conglomerates true and imputed values. Imputed values are calculated as sales-weighted sum of either the median qs of standalone firms in the respective 4-digit sic industries, or the new divisional qs obtained from quantile regressions using the 10 Fama-French industries. We then compute the time-series standard deviations of both excess value measures for each firm with at least five consecutive observations, and plot the kernel density of these standard deviations. The evidence in Figure 1 clearly shows that our method results in excess value estimates with a lower standard deviation than the estimates produced by the traditional approach of Berger and Ofek (1995). This evidence suggests that the new estimates of intra-conglomerate qs are less noisy than the standard industry-median qs of standalone firms. This evidence is particularly convincing because the traditional Berger and Ofek (1995) estimates are based on a finer industry classification, and should arguably yield superior excess value estimates. We therefore conclude that our identifying assumption approximately holds, and that our approach provides cleaner estimates of divisional qs compared to the traditional approach based on qs of standalone firms. Next, we investigate the systematic differences between divisional qs and stan- 13
dalone qs, and use the new measures of divisional qs to investigate the efficacy of conglomerate investment and diversifying acquisitions. III. The Properties of Divisional q s A. Cross-Industry Analysis We begin our analysis by presenting evidence on the average valuation multiples, or qs, of divisions relative to standalone firms across industries. Our focus on industry-level valuation estimates is motivated by the standard definition of a diversified firm in corporate finance: a firm that operates in more than one industry. As suggested in Coase (1937) and Maksimovic and Phillips (2007), corporate diversification matters only if the conglomerate s industry composition affects its transaction costs and hence its optimal boundaries. Thus, for corporate diversification to be of interest, it must be that the industry composition of conglomerates is related to firm value. Table II shows the estimated average valuation multiples of conglomerate divisions, standalone firms, and the resulting relative division qs (rdqs) across the 10 Fama- French industries. Similar to Berger and Ofek (1995), we define the rdq of an industry as the log of the ratio of valuation multiples of conglomerate divisions and standalone firms. As such, rdqs can be used to examine the differences between conglomerate qs and standalone qs across industries. To establish the statistical properties of the time-series average of rdqs, we rely on bootstrapping. In particular, each year we resample residuals of the quantile regression 1,000 times and re-estimate the regression. We then compute average rdqs for each bootstrapped sample. The standard errors of rdqs are calculated using the standard deviation of the bootstrapped estimates. To account for a possible asymmetric distribution of the test statistic, we also report bootstrapped p-values for the test that rdqs are less than or equal to zero. There are two main takeaways from Table II. First, the valuation multiples of conglomerate divisions are significantly different from the valuation multiples of standalone 14
firms. The correlation between the valuation of divisions and standalone firms ranges from 0.37 in the health sector to 0.94 in the manufacturing sector. Furthermore, rdqs are statistically significantly different from zero in all sectors except manufacturing. These estimates suggest that industry multiples of standalone firms are noisy proxies for division multiples and can introduce large measurement errors. Second, Table II shows that the valuation multiples of divisions relative to standalone firms vary considerably across industries. Seven industries have a significantly negative rdq, with the largest discounts in the energy (-55%), high-tech (-35%), and healthcare (-28%) industries. In contrast, conglomerate divisions in two industries show substantial premia. Divisions in the nondurable goods sector are valued at a 16% premium and divisions in the telecommunications sector are valued at a 17% premium. Table II also reveals considerable within-industry variation in rdqs over time. The standard deviation in rdqs within industries ranges from 11% in the manufacturing and energy sectors to 38% in the healthcare sector. This variation is also illustrated in Figure 2, which shows the time-series of rdqs for each industry. Overall, these findings suggest that the aggregate estimates of conglomerate value provided by Berger and Ofek (1995) and the numerous studies that follow do not reflect the considerable within-conglomerate variation in valuation. In particular, our disaggregated estimates indicate that the value of conglomeration varies systematically across industries. Our findings have broad implications for prior research on corporate diversification. First, existing theories of the value of diversification cannot explain our findings. Neither the cross-subsidization of weak divisions (e.g., Rajan, Servaes, and Zingales (2000), Scharfstein and Stein (2000)), the endogenous decision to diversify via acquisitions (e.g., Graham, Lemmon, and Wolf (2002), Campa and Kedia (2002), Villalonga (2004b)), or strategic accounting to appear as artificially low performers (Villalonga, 2004a) readily accommodate systematic heterogeneity in conglomerate valuations across industries. Second, the variation in rdqs across industries also suggests that one needs to exer- 15
cise caution in extrapolating results based on industry-specific analyses. For example, a number of studies have used the Longitudinal Research Database to investigate the value of internal capital markets (e.g., Maksimovic and Phillips, 2002) and the productivity of conglomerate divisions (e.g., Schoar, 2002). However, this database tracks only manufacturing plants, a limitation that the authors of these studies acknowledge. While our estimates confirm that there is no significant discount in the manufacturing sector, they at the same time uncover deep discounts in the energy and high-tech sectors and large premiums in the non-durable and telecommunication sectors. One caveat with this analysis is that the value of operating an industry inside the conglomerate can be affected by its other divisions. For example, Hoberg and Phillips (2015) show that conglomerates tend to operate in economically related industries. Thus, a conglomerate s choice of industries is not random, and endogenous matching can create additional value through synergies. To assess this possibility, we study within-conglomerate cross-industry pairs. Each column in Panel A of Table III corresponds to all conglomerates operating a division in the given industry, and each row reports the proportion of conglomerates that also have divisions in the industry indicated by the row. The sum of each column exceeds one because conglomerates can operate in more than two industries. Consistent with nonrandom industry matching, we find that industry-pairs are not equally distributed inside conglomerates. This can be seen from the nonuniform distribution within each row. For example, 51% of conglomerates with a division in utilities also operate one in energy, whereas only 4% of conglomerates with a division in nondurables do. Importantly, however, we find an insignificant relation between divisions qs and their pairing with other industries inside the conglomerate. Panel B of Table III shows average rdqs that are obtained by omitting from the sample all conglomerates that operate in the industry indicated in the row label. For example, when the sample is limited to conglomerates that have a division in the nondurables industry but not in the durables industry, the nondurables divisions have estimated qs that are 16% larger than 16
those of standalone firms. Across all industries, the exclusion of industry pairs does not have a sizeable effect on the average rdq. These exclusions generate a small variation in rdqs within each industry (average standard deviation = 3.8%). Furthermore, the rdqs estimated after excluding industry pairs are statistically different from the fullsample rdqs at the 5% level only in 6 out of 90 cases (6.7%). Overall, this evidence suggests that the cross-industry variation in rdqs continues to hold after accounting for the endogenous within-conglomerate industry matching. Taken together, the results in this section suggest that conglomerate-implied valuations vary systematically across industries and are substantially different from the commonly used industry valuations based on standalone firms. In the next subsection, we compare the time-series properties of divisional and standalone qs. B. Time-Series Analysis Theories of corporate diversification (e.g., Lewellen, 1971, Matsusaka and Nanda, 2002) suggest that conglomerate divisions are more insulated from external economic forces than standalone firms. Consequently, the valuation multiples, or qs, of conglomerate divisions should be less volatile and less affected by macroeconomic shocks than those of standalone firms. Table II shows the time-series standard deviations of our estimates of divisional and standalone qs. While divisional qs are more volatile in half of the industries, on average across all industries the their standard deviation is 15% lower than that of standalone qs. This difference is economically large and statistically significant in a joint test (p = 0.00). [XXX It would be nice to add more details on how we estimate the difference between the standard deviations with/without adjusting for estimation noise XXX] Importantly, the 15% difference in standard deviations between divisional and standalone qs is understated due to estimation noise. In general, the volatility of an estimated series is inflated by the average volatility of the noise in the estimation. While this affects both divisional and standalone qs, medians are generally estimated with 17
less noise than the quantile regression coefficients. When we manually adjust the timeseries volatility estimates by subtracting the average cross-sectional estimation noise from bootstrapped samples, the standard deviation of divisional qs is 20% lower than that of standalone qs. In Table IV, we test the exposure of both divisional and standalone qs to macroeconomic shocks. For each industry, we regress the annual changes in q on contemporaneous changes in macroeconomic conditions, and report the average slope coefficients across the ten Fama-French industries. We consider nine proxies for macroeconomic conditions: market return, industry return, changes in market dividend yield, changes in the VIX index, changes in the default spread, changes in the industrial production, total factor productivity shocks, and the expansion indicator. These data are collectively gathered from XXX. The sensitivity of standalone qs to macroeconomic conditions is higher than that of divisional qs in all cases. For example, a positive shock to industrial production is associated with increases in both standalone and divisional qs. However, the increase in standalone qs is much higher (6.55) than that in divisional qs (3.80), and the difference of 2.75 is statistically significant. On average, across all these macroeconomic shocks, the sensitivity of divisional qs is a staggering 48% lower. Overall, the evidence in this section suggests that divisional qs are significantly different from standalone qs both cross-sectionally and in the time series. In the next section, we investigate the implications of these findings for conglomerate investment and internal capital allocation. IV. Conglomerate Investment Neoclassical theory suggests that, absent financial frictions, investment should depend only on investment opportunities measured by marginal Tobin s qs. Obtaining good estimates of these marginal Tobin s qs has been focus of extensive research, which acknowledges that it is particularly challenging for conglomerate divisions since their valu- 18
ation multiples are not observable. In this Section, we use our estimates of divisional qs to revisit the evidence on the investment-q sensitivity of conglomerate divisions. More concretely, Shin and Stulz (1998) and Ozbas and Scharfstein (2010) use the neoclassical relation between investment and q to study the efficacy of conglomerates internal capital allocation. Both paper proxy for a division s q by the median q of standalone firms in its industry. Ozbas and Scharfstein (2010) compare the sensitivity of investment to Tobin s q in conglomerate divisions and standalone firms. They find that investment is less sensitive to q in conglomerate divisions than in standalone firms. These results are broadly interpreted as evidence that conglomerates overinvest when opportunities are low and underinvest when they are high. Rather than comparing investments of conglomerates and standalone firms, Shin and Stulz (1998) focus solely on divisional investments, estimating their sensitivity to industry Tobin s q and divisional cash flows. They argue that if internal capital markets are working efficiently, (1) divisional investment will depend mostly on the cash flow of the firm as a whole and not on divisional cash flow, and (2) the sensitivity of investment to cash flow will be lower in divisions with a high q. In contrast, they find that divisional investment is more sensitive to its own cash flow than the cash flow of the firm as a whole, and that the sensitivity of a division s investment to cash flow does not depend on the quality of its investment opportunities. They interpret their evidence as consistent with inefficient internal capital markets and socialism divisions are treated alike irrespective of their investment opportunities. One concern with these studies is their use of standalone firms to proxy for the qs of conglomerate divisions. Whited (2001) and Maksimovic and Phillips (2002) demonstrate that estimates of investment opportunities derived from qs of standalone firms are inappropriate for the study of investment by conglomerate divisions. In particular, they show that these estimates suffer from measurement errors that arise due to potentially unobservable differences between conglomerate divisions and standalone firms. In contrast, the estimates of divisional qs generated by our method do not rely 19
on standalone firms. We therefore use these qs to investigate the differences in the q- sensitivity of investment between conglomerates and standalone firms (Ozbas and Scharfstein, 2010), and the sensitivity of conglomerates internal capital allocations to divisional qs and cash flows (Shin and Stulz, 1998). Columns (1)-(4) of Table V compare the sensitivity of investment to q in conglomerate divisions and standalone firms. As in Ozbas and Scharfstein (2010), the dependent variable is capital expenditures over sales, the regressions include year and industry fixed effects, and the standard errors are clustered by industry-year. In columns (1) and (3), the key explanatory variables are the industry median q of standalone firms (Standalone q), and its interaction with SA, an indicator variable equal to one for standalone firms and zero for conglomerate divisions. Hence, the coefficient on Standalone q represents conglomerate divisions sensitivity of investment to q estimated from standalone firms, and the interaction term captures the incremental sensitivity of standalone firms. In column (3), we also include the ratio of cash flows to sales, CF S. Following Ozbas and Scharfstein (2010), we normalize cash flows by sales instead of assets because conglomerates may have more discretion in allocating assets across divisions. The results in columns (1) and (3) are similar to those obtained by Ozbas and Scharfstein (2010). Conglomerate divisions exhibit lower q-sensitivity of investment than do standalone firms, as evidenced by a statistically significant positive coefficient on the interaction term Standalone q SA. Based on column (3), standalone firms sensitivity of investment to Standalone q is 2.8% higher than that of conglomerate divisions. In columns (2) and (4), we augment the regressions with Divisional q, the divisionlevel estimates of Tobin s q generated by our method. Our results are striking. We find that investment of standalone firms depends on Standalone q, but is insensitive to Divisional q. In contrast, investment of conglomerate divisions is highly sensitive to Divisional q, but is unrelated to Standalone q. For example, based on column 20
(4), the sensitivity of a standalone firm s investment to Standalone q is large and significant at 1.5% + 4.8% = 6.3%, while its exposure to Divisional q is negligible (2.9% 3.4% = 0.5%). In contrast, divisional investment is not significantly related to Standalone q, but highly sensitive to our estimate of the divisional q (2.9%, t-stat = 3.31). These results indicate that conglomerate investment is more sensitive to investment opportunities measured using conglomerate firms than to industry multiples estimated from standalone firms. This finding is more consistent with Whited s (2001) critique that investment opportunities are measured with error and less consistent with the hypothesis that conglomerates invest inefficiently. In columns (5) and (6) of Table V, we investigate the sensitivity of division investment to cash flows using the sample of conglomerate divisions as in Shin and Stulz (1998). This analysis allows for a division s investment to depend on its own investment opportunities and cash flows, as well as those of other divisions (indicated by j ). The investment-cash flow sensitivity can further depend on the division s investment opportunities (Standalone q CF S j and Divisional q CF S j ). When we use industry multiples (Standalone q) to proxy for a division s investment opportunities in column (5), our results are similar to those of Shin and Stulz (1998). First, a division s investment is more sensitive to its own cash flow than to the other divisions cash flows. Second, the sensitivity of a division s investment to other divisions cash flows is not lower when it has better investment opportunities. As noted above, these findings are traditionally interpreted as evidence against efficient internal capital markets. However, when we augment the specification with our division-level multiples Divisional q in column (6), we find different results. A division s investment is more sensitive to other divisions cash flows than its own cash flow (coefficients of 0.168 vs. 0.095). Further, the sensitivity to other divisions cash flows is lower when investment opportunities are higher. This can be seen from the negative coefficient (0.333) on 21
the interaction term Divisional q CF S j. This evidence is broadly consistent with efficient internal capital allocation. Overall, we obtain strikingly different results about a division s q-sensitivity of investment when we use our division level q, Divisional q, rather than the industry median q of standalone firms, Standalone q. These results suggest that prior findings should be interpreted with caution because they may arise due to measurement error in Tobin s q rather than inefficient allocation of capital inside conglomerates. Moreover, these findings suggest that our method generates clean measures of qs inside conglomerates. While our regressions control for firm-level characteristics (XXX? do they), the sensitivity of investment to divisional qs could be correlated with potentially unobservable firm-level characteristics. Under this view, the estimates arise due to firm-level attributes that are unrelated to its organizational form. To address this concern, we estimate the elasticity of investment to standalone and divisional qs around the spinoff of conglomerate divisions. This analysis provides clean estimates that minimize confounding effects as it traces the investment policy of the same firm over time as the firm undergoes a change in its organizational structure. Our analysis focuses on spun-off conglomerate divisions that start operating as public standalone firms. We collect these data by augmenting the spinoff data provided by John McConnell with a sample of spinoffs from sdc. 6 We then collect capital expenditures and sales data for six years around the event. We estimate the investment-q sensitivity regressions after introducing a post-spinoff indicator (P OST ) that allows sensitivities to change at the time of the spinoff. Table VI summarizes the results. In the first two columns, we do not impose any further data restrictions, and analyze 1,561 firm-year observations that correspond to 478 spinoffs. They show that if Divisional q is omitted, investment in the spinoff 6 McConnell s sample contains crsp permnos, which facilitates matching to Compustat. The sdc sample is matched to Compustat based on company name and manually verified. Compustat provides the needed information prior to the spinoff for some firms, and we manually collect missing data from 10-k filings from edgar when available. We thank John McConnell for making the data available on http://www.krannert.purdue.edu/faculty/mcconnell/database. 22