Valuation-Driven Profit Transfer among Corporate Segments

Valuation-Driven Profit Transfer among Corporate Segments Presented by Dr Haifeng You Assistant Professor Hong Kong University of Science and Technology #2012/13-11 The views and opinions expressed in this working paper are those of the author(s) and not necessarily those of the School of Accountancy, Singapore Management University.

Valuation-Driven Profit Transfer among Corporate Segments Haifeng You * School of Business and Management Hong Kong University of Science and Technology July 14, 2012 Abstract This paper investigates whether the desire to achieve higher equity valuations induces conglomerates to manipulate their segment earnings. I extend the Stein (1989) model to a multisegment setting and show that conglomerates have incentives to transfer profits from segments operating in industries with lower valuation multiples to those with higher multiples, even if the market is not fooled in equilibrium. If companies engage in such manipulation, segments with relatively high (low) valuations should report abnormally high (low) profits. The empirical tests confirm this prediction and further show that the relation is stronger for firms with more dispersed segment valuations. Finally, this paper also demonstrates that the simple sum-of-theparts valuation with multiples tends to overestimate the enterprise values for conglomerates, and the measurement errors increase with segment valuation dispersions. JEL Classification: M40, M41, G34 Keywords: earnings management, segment reporting, market efficiency, diversification discount. Data Availability: Data are available from the public sources identified in the paper. * School of Business and Management, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong, E-mail: achy@ust.hk I am grateful for comments from Sudipto Basu, Phil Berger, Kevin Chen, Peter Chen, Gerry Garvey, Allen Huang, SP Kothari, Ryan LaFond, Charles Lee, Jim Ohlson, Panos Patatoukas, David Reeb, Zvi Singer, Richard Sloan, Rodrigo Verdi, Frank Yu, Amy Zang, Guochang Zhang, Weining Zhang and seminar participants at the Five Star Finance Forum at the Renmin University of China, Hong Kong University of Science and Technology, Shanghai University of Finance and Economics, UC Berkeley and the University of Hong Kong. All remaining errors are mine.

1. Introduction Valuation with multiples is a popular equity valuation technique. A simple application of this method to value multi-segment firms involves summing up the imputed values of individual segments. The imputed segment value is often estimated by multiplying segment earnings or sales revenues by the corresponding industry valuation ratios. This method, often referred to as sum-of-the-parts (SOTP hereafter) valuation, is widely used by both academics and professionals 1 (see Lang and Stulz 1994 and Berger and Ofek 1995, among others). If investors adopt the SOTP method mechanically, managers should have incentives to transfer earnings from segments operating in industries with lower valuation multiples to those with higher multiples, as doing so will lead to higher equity valuations. However, it is not necessary for investors to be naïve about the incentives of such profit transfer in order for the manipulation to exist. In this paper, I extend Stein (1989) s model to a multi-segment setting and show that conglomerates transfer earnings from segments with lower valuations to those with higher multiples, even if investors are rational and are not fooled by such manipulation in equilibrium. In the model, managers are trapped into a bad equilibrium. The more appealing cooperative outcome of no profit transfer cannot constitute a sustainable equilibrium. In the presence of information asymmetry, investors cannot observe earnings management directly, and managers have no way of making a credible commitment not to engage in profit transfer. Under this circumstance, it would be irrational for investors to expect managers to transfer no earnings 1 In Appendices I and II, I provide the sample spreadsheets from two analyst reports that use such a method to come up with the price target forecasts for News Corporation and General Electronics, respectively (see Pollak and Chin 2011 and Tusa et al. 2011). 1

among segments. If investors otherwise hold such a belief, managers will have incentives to transfer profits to achieve higher equity valuations. In equilibrium, even though investors cannot observe profit transfer directly, they rationally conjecture that certain amounts of earnings will be transferred among segments and take the manipulation into consideration when valuing conglomerate firms. Given such conjectures, it is also optimal for managers to transfer the same amounts of profits among segments. If conglomerates manipulate segment earnings this way, we would expect divisions operating in industries with high (low) valuation multiples to report abnormally high (low) profitability because earnings are transferred to (from) those segments. In order to test this prediction, I develop a measure of segment abnormal profitability (ABROA) by comparing the profitability of a segment to both single-segment firms operating in the same industry and other segments of the same firm. Empirical tests confirm the prediction and show that the mean ABROA is significantly positive for segments with relatively high valuations and is significantly negative for those with relatively low valuations. The differences in mean ABROA between the two groups range from 0.465 to 0.715 percent, and are all statistically significant. The results are robust to several alternative valuation metrics such as the sales multiple (the ratio of the enterprise value to net sales), the EBIT multiple (the ratio of the enterprise value to earnings before interest and tax) or the EBITDA multiple (the ratio of the enterprise value to earnings before interest, tax, depreciation and amortization). Controlling for the inefficient segment investment and resource allocation documented in Rajan et al. (2000) and other determinants of abnormal profitability does not alter this conclusion. 2

The costs and benefits of transferring profits among segments vary across firms. Other things being equal, firms with more dispersed segment valuations should potentially benefit more from the same amount of profit transfer, as the effect of the transfer on firm valuation should be greater. Profit transfer among segments should thus be more pronounced for firms with segments exhibiting more dispersed valuation multiples. I test this prediction and document that the positive relation between abnormal segment profitability and relative valuations is stronger for firms with more dispersed segment valuations. For example, the differences in ABROA between relative high and low valuation segments range from 0.613 to 1.012 percent for firms with high dispersions of sales multiples. In contrast, the differences are only about -0.09 to 0.056 percent for those with low dispersions of sales multiples. The model also predicts that the SOTP valuation tends to overestimate the values of conglomerates if not adjusted for the inter-segment profit transfer. Furthermore, the estimation errors in the imputed values should be greater for firms with more dispersed segment valuations, where managers have greater incentives to manipulate segment profits. If investors anticipate this manipulation and adjust for it, even partially, the market values of conglomerates will be lower than the imputed values, and the discrepancy between the two values should be greater for firms with more dispersed segment valuations. Consistent with this prediction, the empirical tests show that the logarithms of the ratios of market values to the imputed values, i.e., the excess values (EXV hereafter) developed by Berger and Ofek (1995), are negative on average, and are significantly lower for firms with more dispersed segment valuations. The results that the market values of conglomerates are on average, lower than the imputed values, dubbed as the diversification discount, are first documented by Lang and Stulz 3

(1994) and Berger and Ofek (1995). The literature often interprets the discount as evidence that diversification destroys value. Specifically, it is often argued that the conglomerate form of organization exacerbates managerial agency problems and leads to investment inefficiencies such as overinvestment and inefficient cross-subsidization (e.g., Berger and Ofek 1995; Denis, Denis and Sarin 1997; Rajan, Servaes and Zingales 2000; Scharfstein and Stein 2000; Billett and Mauer 2003; Ozbas and Scharfstein 2010). However, additional analysis suggests that investment inefficiencies are unlikely the main drivers of the negative association between EXV and segment valuation dispersion. First, I find a significant association between changes in EXV and changes in segment valuation dispersion for a sample of firm-years without any changes in segment compositions, where the perceived tendency of overinvestment should be largely unchanged. Second, untabulated tests also indicate that that the relationship between EXV and segment valuation dispersion remains intact, even after controlling for the (in)efficiency of crosssubsidization. Agency problem-driven investment inefficiency 2 does not seem to account for the positive association between abnormal profitability and relative segment valuation, either. First, the coefficient on relative segment valuation is statistically significant, even in the presence of a proxy for abnormal investment in the regression. Second, further analysis also reveals that the positive association is stronger for firms with higher managerial equity incentives. If agency problem-driven investment inefficiencies drive the positive association between abnormal 2 Prior research documents that agency problems induce managers to over- (under-) invest in divisions with poor (good) investment opportunities due to inefficient cross-subsidization (e.g., Berger and Ofek 1995; Rajan et al 2000; Billett and Mauer 2003). Under the assumption of diminishing marginal returns to investment, overinvestment (underinvestment) results in an abnormally low (high) average ROA. Given that segments with good (poor) investment opportunities tend to have high (low) valuation, inefficient cross-subsidization and investment may also lead to a positive association between relative segment valuation and abnormal profitability. 4

profitability and relative segment valuation, we would expect the results to be weaker for these firms, as higher equity incentives should mitigate such agency problems. This paper contributes to the extant literature in a number of ways. First, this research enriches the earnings management literature by documenting evidence of valuation-driven earnings management of transferring profits from segments with low industry valuations to those with high industry valuations. The existing literature shows that managers have incentives to manipulate firm-level earnings under various circumstances. Compared to firm-level earnings, segment-level reports often involve more managerial discretions and are subject to less auditor scrutiny (e.g., Givoly, Hayn and D Souza 1999; Berger and Hann 2003). However, little research has been conducted with respect to whether managers take advantage of the discretion and manipulate segment earnings. This paper contributes to the literature by showing that capital market considerations induce managers to report segment earnings opportunistically. This paper also adds to the equity valuation literature by highlighting a potential issue with the simple sum-of-the-parts valuation method. I demonstrate that this method tends to overestimate firm values if not adjusted for the inter-segment profit transfer, and the estimation errors increase with segment valuation dispersions. The findings should be of particular interest to capital market participants. Understanding managers incentives to manipulate segment-level earnings is critical in pricing those companies properly. If investors fail to take such earnings manipulation into consideration, they may make suboptimal investment decisions and may suffer from investment losses. This study also contributes to the literature on the diversification discount. A popular explanation for the diversification discount is that diversification destroys firm values because 5

diversified firms are subject to a more serious agency problem (e.g., Berger and Ofek 1995; Dennis et al. 1997; Rajan et al. 2000; Ozbas and Scharfstein 2010). This paper shows that failing to take valuation-driven profit transfer into consideration leads to overestimated imputed values, which could also lead to the diversification discount, even if diversification is not valuedetrimental. It is also worth emphasizing that the estimation error problem is unlikely to be the complete explanation for the diversification discounts. First, the diversification discount remains noticeable, albeit low, for firms with low (or zero) segment valuation dispersions. Second, the imputed values calculated from segment assets and the corresponding asset multiples are also higher than the market values, on average. Unless we can argue that reported segment assets are subject to similar manipulations, we cannot attribute the discounts to measurement errors in reported segment earnings or sales. The rest of the paper is structured as follows: Section 2 reviews the literature. Section 3 presents the analytical model and develops the hypotheses. Section 4 describes the data and sample selection. Section 5 provides the empirical results. Section 6 examines several additional tests. Section 7 concludes. 2. Literature review Existing accounting research shows that segment-level financial information is useful for investors when it is available. For example, Baldwin (1984) documents that analysts forecast accuracy improved after multi-segment firms started to report segment-level information. Swaminathan (1991) shows that divergence of beliefs was lower in May 1971 after the release of SEC-mandated segment disclosure, compared to May 1970. Furthermore, researchers have also 6

shown that the new segment reporting requirements of SFAS 131 provided more disaggregated information, affected market valuation, improved monitoring (Berger and Hann 2003) and increased the stock price informativeness with respect to future earnings (Ettredge, Kown, Smith and Zarowin 2005). Chen and Zhang (2003) show that segment reporting provides information for valuation purposes incremental to firm-level aggregated accounting information, and the incremental value relevance of segment data relates to the heterogeneity of investment opportunities across segments. Despite its frequent use as valuation input, segment-level financial information tends to involve more managerial discretion and have poorer quality, compared to firm-level financial reporting. For example, Givoly et al. (1999) document that measurement error in segment financial information, particularly earnings, is significantly greater than that in financial information reported by single-segment firms. While there is some evidence that SFAS 131 increased the number of reported segments and provided more disaggregated information (Berger and Hann 2003), it is still unclear to what extent the quality of segment earnings has been improved, as the new standards still give managers considerable discretion in reporting segment earnings 3. Indeed, Chen and Zhang (2007) document some indirect evidence of segment earnings management, both before and after SFAS 131. The potential usefulness of segment reporting, together with the discretion allowed in segment reporting calls for further research to better understand segment reporting incentives, the causes and consequences of segment-level earnings management and the implications for financial information users. Such 3 In fact, the new rule of SFAS 131 does not even define the measure of profit or loss to be disclosed. Instead, it allows any measure used for internal decision making to be reported as segment profits. If managers abuse the flexibility to manipulate segment earnings, the quality of segment earnings could even be worse under SFAS 131. 7

research has so far been scarce. One exception is the work of Hann and Lu (2009), who show that managers exploit their discretion in cost allocation to avoid reporting losses. This paper is also related to the large body of literature showing that capital market considerations induce managers to manipulate earnings. Stein (1989) shows analytically that managers who care about short-term stock prices may engage in myopic earnings management, even if the stock market is efficient. Teoh, Welch and Wong (1998a and 1998b) and Teoh, Wong and Rao (1998) subsequently confirm that firms manipulate their accruals prior to initial public offerings and seasoned equity offerings to obtain higher prices. Erickson and Wang (1999) show that acquiring firms manage their earnings upward in order to increase their stock prices prior to stock for stock mergers. Most of the research, however, focuses on firm-level earnings management. This study extends the literature by examining whether equity valuation consideration induces earnings management at the segment level. Specifically, I investigate whether conglomerate firms shift earnings from segments operating in industries with low valuations to those in industries with high valuations, and how such transfer varies across firms. Although more sophisticated valuation methods such as the discounted cash flow model or the residual income model have long been available, valuation with multiples remains one of the most popular valuation techniques (Liu, Nissim and Thomas 2002). For example, Bradshaw (2004) shows that a special multiple valuation method, the price-earnings-to-growth (PEG ratio) model, explains sell-side analysts recommendations better than present-value approaches, such as residual income models. A simple application of this method to value multi-segment firms involves summing up the values of individual segments, which are estimated from multiplying reported segment accounting variables by the appropriate industry valuation ratios (e.g., Berger 8

and Ofek, 1995). This method, often called the sum-of-the-parts approach, is also widely used by investment professionals (see for example, Pollak and Chin 2011 and Tusa et al. 2011). If the market applies this valuation technique mechanically without adjusting for the potential earnings manipulation among business segments, managers should have an incentive to transfer earnings from segments operating in industries with low valuation multiples to those operating in industries with higher valuation multiples, as doing so will lead to higher equity valuations. Market inefficiency, however, is not a necessary condition for the existence of such earnings manipulation. More sophisticated investors may rationally expect that managers will engage in segment-level earnings manipulation and adjust for it properly when valuing those stocks. Nevertheless, firms may still engage in such manipulation under this circumstance. The intuition is similar to the rationale of corporate earnings management modeled by Stein (1989). Managers are trapped in equilibrium where the market correctly conjectures the degree of earnings manipulation; and earnings manipulation is the managers optimal choice given the market expectations. Following a different analytical framework, Fan (2007) also shows that income-increasing earnings management exists before the IPO as the outcome of a fully separating equilibrium in which investors rationally anticipate the manipulation. In the following section, I extend Stein (1989) s signal jamming model to examine managers inter-segment profit transfer behavior with rational investors. 3. Theoretical model and hypotheses Consider a conglomerate with two segments, labeled as segment 1 and 2. True economic earnings for the two segments that are unobservable to investors are denoted as and, 9

respectively; and are mutually independent, and both follow normal distributions with mean and precision 4. The manager of the firm privately observes signals of actual earnings with noise, denoted as and, respectively; and are independently distributed, and both follow random normal distribution with mean 0 and precision. After observing his private signals, the manager may transfer a certain amount of earnings from one segment to another, and he may then report the manipulated numbers to investors. The amount of earnings transferred is denoted as. The reported earnings, denoted as and, are given by the following equations: e a v b and. We assume that managers bear personal costs for profit transfer, which can be regarded as reputation loss or other costs incurred in manipulating earnings. The cost function is assumed to be an increasing convex function of, with 0 0 and 0 0. Valuation multiples for the two segments are exogenously determined, and are denoted as and, respectively; measures the difference (or dispersion) in valuation multiples between the two segments. Without loss of generality, I assume 0. The task of the market is to determine the market value of the firm, given reported earnings and a conjecture of profit transfer :,, 1, is the conditional expectation of the true profit, given reported earnings and. 4 We assume the economic earnings for two segments have the same mean and precision for the sake of simplicity. Relaxing the assumption, however, does not change the propositions or implications derived from the model. 10

Taking the conjectured profit transfer into consideration, the market adjusts the reported earnings in forming expectations about true earnings, and can be expressed as follows: 2 The manager s objective is to maximize his utility, which, with appropriate normalization, can be defined as the market value of the firm less his personal cost of profit transfer by choosing an optimal level of profit transfer : 3 Utility maximization suggests that the following condition be satisfied: From (2) and (4), we obtain: 0 4 0 5 The derivatives / and / take the values of -1 and 1, respectively, so the firstorder condition can be written as follows: 0 6 In a rational framework, the manager will choose an optimal level of profit transfer, given the market pricing function, and the market will rationally take the incentive of profit transfer into consideration when pricing the firm. In steady-state equilibrium, the market s conjectured profit transfer equals the optimal amount of profit transfer chosen by the manager. Solving for the equilibrium, we have the following proposition: 11

Proposition 1: Under the assumption that is an increasing convex function with 0 0 and 0 0, there exists a unique equilibrium characterized by the following properties: 1) For any 0, the equilibrium amount of profit transfer 0. 2) The equilibrium amount of profit transfer is an increasing function of. 3) On average, the equilibrium market value is lower than the imputed value estimated with valuation multiples, i.e.,. The expected discount equals: 7 Property 1 suggests that in equilibrium, conglomerate firms will always transfer some earnings from one segment to another, as long as the valuation multiples are different. It is worth emphasizing that the market is not fooled by such earnings management in equilibrium, even though investors cannot directly observe the amount of profit transfer. In equilibrium, the market takes the potential benefit of the profit transfer into consideration and correctly conjectures the amount of such earnings manipulation. Even though the manager is unable to fool the market, he still transfers earnings between segments because otherwise, the valuations would have been even lower, given the market expectations. Note also that zero profit transfer cannot constitute a sustainable equilibrium. If the market holds a fixed conjecture of no profit transfer, the manager will have incentives to deviate from this conjecture in order to achieve higher valuations. The empirical prediction of property 1 is that managers will transfer earnings from segments with lower valuations to those with higher valuations, even under the assumption that investors are rational and perfectly conjecture the amount of profit transfer. In reality, investors may not always observe the potential benefits and costs; therefore, they may not always be able 12

to perfectly conjecture the manipulation. As discussed earlier, managers should still have incentives to transfer profits among segments under this scenario. In order to test this prediction, I need to have a measure of earnings management at the segment level. Researchers typically use various discretionary accrual models to identify earnings management (e.g., DeAngelo 1986; Jones 1991; Dechow, Sloan and Sweeney 1995). However, the detailed cash flow or balance sheet data necessary to calculate accruals are rarely available at the segment level. To deal with this problem, I infer earnings management indirectly by comparing segment profitability to the profitability of single-line-of-business firms operating in the same industry and other segments of the same firm. If multi-segment firms systematically transfer earnings from one segment to another, we would expect the subsidizing segment to report abnormally low profitability and the subsidized segment to report abnormally high profitability. Property 1 suggests that segments with high (low) valuation multiples are more likely to be the subsidized (subsidizing) segments. This notion leads to the first testable hypothesis, stated in the alternative form: Hypothesis 1: The abnormal profitability of a segment is positively associated with its relative valuation within the firm. Property 2 predicts that the amount of profit transfer increases with the difference in valuation multiples between the two segments. The intuition is straightforward. Not all conglomerate firms have the same incentives to transfer earnings among their segments. Other things being equal, firms with more dispersed segment valuations potentially benefit more from the same amount of profit transfer, and therefore will have a greater incentive to do so. For example, if investors adopt the SOTP valuation mechanically, a firm can potentially increase its 13

market value by $10 by transferring a $1 profit between two segments with PE multiples of 10 and 20. The valuation impact of the same amount of profit transfer would be only $2 if the PE multiples for the two segments are 14 and 16. Hence, the firm should have greater incentives to transfer profit in the former case. Property 2 shows that similar results can be expected at equilibrium, even if investors are rational. The market rationally expects that firms with more dispersed segment valuations will transfer more profits due to the larger potential benefit, and thus correctly adjusts for it, on average. Given this market expectation, it is also optimal for managers to do so. Consequently, the second hypothesis, stated in the alternative form is: Hypothesis 2: The positive association between abnormal profitability and relative industry valuation is stronger for conglomerate firms with more dispersed valuation multiples. The model also points out a potential problem of valuing multi-segment firms with the simple sum-of-the-parts valuation with multiples. The SOTP valuation generates an imputed value of. However, is on average, higher than the actual value implied from the (unobservable) economic earnings, i.e.,. This happens because the simple valuation multiples approach fails to adjust for the positive amount of profit transferred from the segment with lower valuation to the one with higher valuation. Property 3 further suggests that is higher than the firm value estimated from rational expectations, i.e., the market value. Although the economic earnings and the action of profit transfer are unobservable to investors, they form rational expectations about the amount of profit transfer and properly adjust for the earnings management when valuing the firm. Property 3 also generates a prediction with respect to the cross-sectional variation of the measurement errors in the imputed value estimates. Equation (7) implies that the measurement 14

errors increase with, which captures the dispersion of segment valuation multiples, and therefore, the potential benefits of profit transfer. The measurement errors are larger for firms with more dispersed valuation multiples for two reasons. First, the effect of a unit of profit transfer on valuation estimates is larger for firms with greater. Second, as shown in property 2, firms with greater also tend to transfer more earnings among segments. In the model, investors are rational and perfectly conjecture and adjust for the amount of profit transfer. In reality, as long as the market does not completely ignore such inter-segment earnings management, the market values of conglomerate firms will be lower than the imputed firm values. Additionally, the discount of the market value relative to the imputed value, i.e., the diversification discount measure of Berger and Ofek (1995), should be greater for firms with a higher dispersion of valuation multiples. We therefore have the third hypothesis, stated in the alternative form: Hypothesis 3: Ceteris paribus, the discounts of market values relative to the imputed values are greater for conglomerate firms with more dispersed segment valuations. It worth noting that hypothesis 3 is a joint test of the predictions that 1) higher segment valuation dispersion induces more profit transfer from low to high valuation segments; and 2) the market understands such cross-segment profit transfer and adjusts for it, at least partially. If the market adopts the same naïve multiple valuation technique without adjusting for the crosssegment profit transfer, we will neither observe a (diversification) discount, nor will we expect the predicted association between the diversification discount and segment valuation dispersion, as described in hypothesis 3. 15

4. Data and sample selection I collect financial accounting data from the Standard & Poor s Compustat North America Fundamental database. Business segment data are available on Compustat Segments database. Stock returns and prices are obtained from the University of Chicago s Center for Research in Security Prices (CRSP). I eliminate firms in the financial services (SIC code: 6000-6999) and utilities (SIC code: 4400-5000) industries. Firm-year observations with sales or total assets less than $20 million are also excluded. The initial sample includes 198,855 business segments reported in the Compustat Segments database for the period 1998 to 2007. The sample period starts at year 1998 because the data before that year was backfilled by Compustat to conform with SFAS 131, which took effect on December 15, 1997. After dropping segments with missing SIC codes and corporate, reconciliation, and elimination segments, the sample consists of 158,320 segments for 90,783 unique firm-years. Among them, 35,561 observations are firm-years with multiple segments, and the remaining 55,222 are single-segment firm-years. The single-segment firm-year observations are used to calculate industry-level valuation multiples and profitabilities. After 1) merging with CRSP to obtain market capitalization data; 2) merging with the Compustat North America Fundamental database to obtain additional financial data; 3) eliminating firm-years with total assets or sales revenue less than $20 million; and 4) excluding observations where the total sales revenue from the Computat Segment database differs from that reported by the Compustat annual Fundamental database by more than 5%, I still have 17,822 observations for single-segment firms and 37,184 segments for multisegment firms. I require at least five firms in an industry (defined by a three-digit SIC code) to 16

compute the industry median profitability and valuation multiples. I am able to compute the industry-level profitability and valuation multiples for 2,917 industry-years. The final sample of segments of conglomerate firms includes 15,279 observations for which abnormal profitability and industry-adjusted valuation multiples can be calculated. 5. Research design and empirical results 5.1. Empirical measures 5.1.1. Abnormal profitability (ABROA) The first hypothesis concerns whether or not the abnormal profitability of a segment is positively associated with its relative valuation multiples. I develop a measure of the abnormal profitability of a segment by comparing its profitability to both single-segment firms in the same industry and other segments of the same firm. Specifically, I calculate the abnormal return on assets (ABROA) as follows: ABROA s,t = (ROA s,t IROA s,t ) (Σ s (ROA s,t IROA s,t ) )/S (8) where ROA s,t is the ratio of operating income after depreciation and amortization to identifiable assets, multiplied by 100 for segment s at year t. IROA s,t is the median ROA of all single-segment firms with the same three-digit SIC code as segment s in fiscal year t. S is the total number of segments of the firm. Thus, the first term in equation (8) is the industry-adjusted ROA for segment s, while the second term is the mean industry-adjusted ROA of all segments of the firm. I adjust the industry-adjusted ROA of a segment by subtracting the corresponding firmlevel mean for two reasons. First, multi-segment firms may also manipulate firm-level earnings 17

by inflating profits for all segments. Under this circumstance, industry-adjusted ROA tends to be positive for all segments. We would incorrectly attribute the positive industry-adjusted segment ROA as inter-segment profit transfer without further adjustment. Second, prior research suggests that conglomerate firms have a more serious agency problem than single-segment firms (e.g., Shleifer and Vishny 1989; Denis et al. 2002). If the agency problem adversely affects the performance of multi-segment firms, then the industry-adjusted segment ROA would be negative, even in the absence of inter-segment profit transfer. Thus, subtracting the mean industry-adjusted ROA helps make ABROA a cleaner proxy of inter-segment profit transfer. 5.1.2. Segment relative valuation (RV) The relative valuation multiple of a segment is calculated as the difference between its valuation multiple and the weighted average valuation multiples of the firm, i.e.: RV s,t =VM s,t Σ s (VM s,t * ω s,t ) (9) where VM s,t is the median valuation multiples of all single-segment firms in the same industry as segment s in year t. The industry for each segment is defined as the narrowest SIC grouping that includes at least five single-segment firms with sales of at least $20 million. ω s,t is the weight of segment sales as a fraction of the firm s total sales 5, so ω t =Sales s,t /(Σ s Sales s,t ). I compute RV using three valuation multiples: RV_SALE, calculated from the sales multiple (the ratio of the enterprise value to net sales); RV_EBIT, calculated with the EBIT multiple (the ratio of the enterprise value to earnings before interest and taxes); and RV_EBITDA, calculated from the EBITDA multiple (the ratio of the enterprise value to earnings before interest, taxes, depreciation 5 I also tried to use segment identifiable assets as an alternative-weighting variable. The results were similar, both qualitatively and quantitatively. 18

and amortization). Enterprise value (EV) is defined as the sum of the market cap of equity, longterm debt, short-term debt and preferred stock. 5.1.3. Valuation dispersion (VDISP) Tests of the second and third hypotheses require measures of the dispersion of segment valuations. I construct valuation dispersions (VDISP) as follows: VDISP t =Σ s ( RV s,t * ω s,t )/Σ s (VM s,t *ω s,t ) (10) where RV s,t is as defined in equation (9), which measures how far the valuation multiple of a segment differs from the weighted average valuation of all segments of its parent firm. We calculate the segment valuation dispersion of a firm as the sales-weighted average absolute value of RV s,t, i.e., the weighted average absolute deviation 6 of VM s,t. The weight of a segment ω t =Sales s,t /(Σ s Sales s,t ). Firms with a higher level of valuation tend to have a higher level of valuation dispersion. To control for this effect, we further standardize the valuation dispersion measure by scaling it by the weighted average valuation multiples of all segments of the firm. 5.1.4. The ratio of the market value relative to the imputed value (EXV) Hypothesis 3 predicts that overestimation errors in the imputed value measures increase the dispersion of segment valuations. If the market takes the profit transfer into consideration in pricing the conglomerate, we can use the market value as a benchmark to assess the degree of measurement errors. Following Berger and Ofek (1995) 7, I calculate the excess value of a firm s market value to the imputed value as EXV= log (MV/IV), where MV is the market value of the firm (enterprise value), calculated as the sum of the market cap of equity, long-term debt, short- 6 I use the average absolute deviation instead of the standard deviation because the former measures are less influenced by outliers. 7 The detailed procedures to calculate IV can be found on Page 60 of Berger and Ofek (1995). 19

term debt and preferred stock. IV is the imputed value of a firm, calculated as the sum of the imputed segment values, i.e., Σ s (VM s,t *AI s,t ). VM s,t is the median valuation multiples of all single-segment firms in the same industry, and AI s,t is the accounting variable of interest for segment s. Following Berger and Ofek, I calculate three EXV measures, EXV_S, EXV_A and EXV_E, which are estimated from sales, identifiable assets and EBIT multiples, respectively. 5.2. Empirical results 5.2.1. Tests of Hypothesis 1 Hypothesis 1 predicts that the abnormal profitability of a segment should be positively associated with its relative segment valuations. Table 1 provides the descriptive statistics for ABROA, the relative valuation measures and other control variables. Panel A indicates that the medians of ABROA and the three RVs are all zero, and the means are also close to zero. Consistent with the prediction of hypothesis 1, the correlation matrix in Panel B shows that both the Pearson and Spearman correlations between ABROA and the three RVs are all positive. I first conduct a univariate test of hypothesis 1 by examining mean abnormal profitability for segments with different levels of relative valuations. I place observations into three groups based on each of the relative valuation measures. Segments with RV<0 are placed in the LOW RV group, those with RV>0 in the High RV group, and the rest in the Medium RV group. Panel A of Table 2 reports the mean ABROA for the three groups. When the portfolios are formed based on RV_SALE, the results show that the mean ABROA for segments in the LOW RV_SALE group is -0.172 percent, while the mean for the HIGH RV_SALE group is 0.293 percent. The difference of 0.465 percent is statistically significant at the 5 percent level. The results are stronger for portfolios formed on RV_EBIT and RV_EBITDA in the last two rows, showing that the mean 20

ABROA is significantly negative for segments in the LOW RV group, and significantly positive for those in the HIGH RV group. The differences between the two groups are 0.706 and 0.715 percent for the portfolios formed on RV_EBIT and RV_EBITDA, respectively. In order to control for other determinants of segment profitability, I also test Hypothesis 1 with the following multivariate regression model: ABROA s,t =a 0 +a 1 SRV s,t +a 2 ABROA s,t-1 +a 3 ABINV s,t-1 +a 4 RELSIZE s,t-1 +a 5 MKTSHR s,t +a 6 HINDX s,t +a 7 SIZE t +a 8 BM t +ε t (11) where: ABROA s,t is the same, as defined earlier. SRV s,t is a transformation of the relative valuation RV s,t, which equals 1 if RV s,t,>0, 0 if RV s,t,=0, and -1 if RV s,t,<0. ABINV s,t is the abnormal investment, calculated as (INV s,t IINV s,t ) (Σ s (INV s,t IINV s,t )/S), where INV s,t is the ratio of the segment capital expenditure to identifiable assets, multiplied by 100. IINV s,t is the median ratio of capital expenditure to total assets of all single-segment firms in the same SIC three-digit industry in the fiscal year. I include this variable to control for the effect of inefficient segment investment, as documented in Rajan et al. (2000) on segment profitability. MKTSHR s,t is the market share, calculated as the sales of segment s as the fraction of the total sales of all firms/segments with the same three-digit SIC code. We include the market share as a control variable because prior research shows that entities with high market shares tend to have superior performance (e.g., Buzzell, Gale and Sultan 1975; Prescott, Kohli and Venkatraman 1986). 21

HINDX s,t is the Herfindahl index for the industry to which a segment belongs, where an industry is again defined by a three-digit SIC code. The Herfindahl index is included to control for the effect of competition on segment performance and financial reporting decisions (Harris 1998). Following Berger and Hann (2007), we also control for the segment s size relative to the firm (RELSIZE s,t ), firm size (SIZE t ) and firm growth opportunity (BM t ). RELSIZE s,t is calculated as the ratio of segment assets to firm assets at year t; SIZE t is the logarithm of the firm s market cap at the beginning of fiscal year t; BM t is the firm s book-to-market ratio at the beginning of year t. The regression results are provided in Panel B of Table 2. The table reports the mean coefficients of the annual cross-sectional regressions and the t-statistics of the annual coefficients adjusted for serial correlation of one lag using the Newey and West (1987) procedure. Consistent with the results presented in Panel A, Panel B shows that the coefficients on all three relative valuations are positive and significant, even in the presence of the other control variables. The results are consistent with the prediction of hypothesis 1 that conglomerate firms transfer profits from segments with low valuations to those with high valuations. Rajan et al. (2000) document that diversified firms may transfer resources from divisions with good opportunities to those with poor opportunities. This transfer may lead to overinvestment (underinvestment) in divisions with poor (good) investment opportunities. Under the assumption of diminishing marginal returns to investment, overinvestment (underinvestment) may results in an abnormally low (high) average ROA. Consistent with this notion, the coefficients on lagged abnormal investment are negative. These coefficients are, however, 22

insignificant 8. More importantly, relative valuation multiples have significant explanatory power over abnormal profitability, even in the presence of abnormal investment, suggesting that the results are unlikely driven by inefficient investment or resource allocation. 5.2.2. Tests of Hypothesis 2 The second hypothesis investigates whether the positive association between relative valuation multiples and abnormal profitability is stronger for firms with more dispersed segment valuation multiples. In order to test hypothesis 2, I first split the sample into two equal size groups each year based on VDISP, and then I examine the difference in ABROA between segments with high and low relative valuations for the two groups separately. The results appear in Table 3. In Panel A, I separate the sample into low and high dispersion groups based on the dispersion of the sales multiple (VDISP_SALE). The results show that for the high dispersion group, segments with high relative valuation report positive and significant abnormal profitability, while those with low relative valuation have negative abnormal profitability and are mostly significant, as well. The differences in abnormal profitability between segments with high and low relative valuation range from 0.613 to 1.012 percent. In contrast, for the low dispersion group, neither high nor low relative valuation segments report statistically significant abnormal profitability. The differences in ABROA between segments with high and low relative valuation are insignificant for all three RV measures. Panels B and C show similar results when the sample is split based on VDISP_EBIT and VDISP_EBITDA. I also test hypothesis 2 with the following multivariate regression model: 8 The relatively short sample period may account for the low statistical significance. The sample period covers ten years between 1998 and 2007. The t-statistics with Newey-West correction for serial correlations for a time-series of ten observations may lack the statistical power needed to show high significance. 23

ABROA s,t =a 0 +a 1 SRV s,t + a 2 VDISP t + a 3 SRV s,t * VDISP t +a 4 ABROA s,t-1 +a 5 ABINV s,t-1 +a 6 RELSIZE s,t-1 +a 7 MKTSHR s,t +a 8 HINDX s,t +a 9 SIZE t +a 10 BM t +ε t (12) where all variables are as defined earlier. If the positive association between relative valuation multiples and abnormal profitability is stronger for firms with more dispersed valuation multiples, a 3 should be positive and significant. The regression results are reported in Table 4. Panel A shows the results where the segment valuation dispersion is proxied by the dispersion of the sales multiple (VDISP_SALE). The coefficients on the interaction terms between RV and VDISP_SALE are positive for all three RV measures and are statistically significant for RV calculated from the sales and EBITDA multiples. Panels B and C present stronger results, where I use VDISP_EBIT and VDISP_EBITDA to proxy for segment valuation dispersion. The coefficients on the interaction terms are all positive and statistically significant, at least at the 5 percent level. The overall results in Tables 3 and 4 are consistent with the prediction of hypothesis 2, that is, conglomerates with more dispersed segment valuations transfer more profits from segments with low valuations to those with high valuations, as they enjoy more benefit from doing so. 5.2.3. Tests of Hypothesis 3 The final hypothesis analyzes the measurement errors in the imputed value estimates resulting from failing to adjust for the inter-segment profit transfers. Specifically, I investigate whether the measurement errors are larger for firms with more dispersed segment valuations. As discussed earlier, I use the excess value of Berger and Ofek (1995) as a proxy for the measurement error in the imputed value estimate. If IV overestimates firm value, it is likely to be greater than the market value MV, and EXV would be negative. If hypothesis 3 is true, we 24

would expect EXV to be more negative for firms with more dispersed segment valuations because the measurement errors would be greater. Consistent with the findings of Berger and Ofek (1995), I find that both the mean and median of all three excess value measures, EXV_S, EXV_A and EXV_E, are negative and significant. For this sample, multi-segment firms are on average, traded at about 13.3, 12.8, and 21.2 percent discounts relative to the imputed values estimated from sales, assets and EBIT multiples, respectively (untabulated). Table 5 provides the univariate tests of hypothesis 3, where I report the mean EXV for conglomerates with different levels of segment valuation dispersions. In Panel A, I place all firms into five portfolios based on VDISP_SALE. Portfolio (1) includes those with VDISP_SALE=0; the rest of the stocks with VDISP_SALE>0 are separated into four equal-sized quartiles with portfolio (2) containing those with the lowest and (5) containing those with the highest VDISP_SALE. The results show that all three EXV measures are lower for portfolio (5) than not only portfolio (1), but also portfolio (2). The results are particularly strong for EXV_S. The mean EXV_S is -0.076 for the portfolio with VDISP_SALE=0, and -0.098 for the lowest quartile of a positive VDISP_SALE. In contrast, the mean EXV_S is 0.205 for the highest quartile of a positive VDISP_SALE. Panels B and C present similar results for portfolios sorted on VDISP_EBIT and VDISP_EBITDA. The above results are generally weaker for the diversification discount, calculated from asset multiples (EXV_A). This finding is perhaps not surprising, given that the measurement errors in firm value estimates resulting from not adjusting for cross-segment profit transfer should only affect earnings and sales-based valuation estimates directly. The weak negative 25

association between EXV_A and dispersions of earnings/sales multiples may indicate that conglomerate firms also transfer or manipulate assets among segments, even though segment assets are more difficult to manipulate than earnings and sales. The overall results in Table 5 suggest that the dispersion of valuation multiples has an economically significant impact on the diversification discount. It is also worth noting that failing to adjust for potential profit transfer among segments is not a complete explanation for diversification discounts, as even firms in the lowest dispersion quintiles are still traded at a discount relative to their imputed firm values. These results, therefore, cannot exclude the possibility that certain agency problems also contribute to the diversification discount. The multivariate tests of hypothesis 3 are conducted with the following model 9 : EXV t =a 0 +a 1 VDISP t +a 9 NSEG t + a 2 RELATED t +a 3 OPMG t +a 4 LOGASSET t +a 5 CAPEX t +a 6 LEV t +ε t (13) where EXV is as defined above. I also control for other determinants of the diversification discounts (Berger and Ofek 1995; Bens and Monahan 2004; Hoechle et al. 2012): NSEG t : the logarithm of the number of segments reported by a firm; RELATED t : the relatedness of a firm s operations, which is equal to the difference between the total number of reported segments and the number of segments with different two-digit SIC codes; OPMG t : operating margin, calculated as EBIT divided by net sales; LOGASSET t : the logarithm of the firm s total assets in millions; 9 Following Berger and Ofek (1995), I drop OPMG in the regression with EXV, calculated from earnings multiples to avoid spurious inferences. 26