DOI: 10.1111/j.1475-679X.2010.00397.x Journal of Accounting Research Vol. 00 No. 0 xxxx 2011 Printed in U.S.A.

Earnings Quality Based on Corporate Investment Decisions FENG LI∗ Received 25 July 2007; accepted 20 September 2010

ABSTRACT

In this paper, I examine a new approach for measuring earnings quality, defined as the closeness of reported earnings to “permanent earnings,” based on firm decisions with regard to capital and labor investments. Specifically, I measure earnings quality as the contemporaneous association between changes in the levels of capital and labor investment and the change in reported earnings. This approach follows the reasoning that (1) firms make investment decisions based on the net present value (NPV) of investment projects and (2) reported earnings with higher quality should more closely associate with real investment decisions. I find that measures of earnings quality based on managerial labor and capital decisions correlate positively with earnings persistence and have incremental explanatory power relative to earningsquality measures used in the accounting literature. Furthermore, investmentbased earnings-quality measures are less informative when managers tend to overinvest.

1. Introduction Prior research on earnings quality generally relies on one of two approaches: studying the properties of accounting numbers or extracting

∗ Stephen M. Ross School of Business, University of Michigan. I thank Ray Ball, Phil Berger, Ilia Dichev, Kenneth Merkley, workshop participants at the University of Chicago, and especially an anonymous reviewer and Richard Leftwich (the journal editor) for their comments. 1 C , University of Chicago on behalf of the Accounting Research Center, 2011 Copyright 

2

F. LI

information from stock prices.1 This paper explores a new measure of earnings quality by examining firm investment decisions.2 Managerial investment decisions likely contain information about earnings quality because managers make many decisions based on future profitability, and arguably have more precise and complete information about their firm’s profitability than do other stakeholders. Therefore, to the extent that information asymmetry exists between managers and outsiders, the earnings quality inferred from managerial decisions can provide incremental information to existing empirical measures based on the information set of outside investors or on the properties of the accounting numbers.3 In this study, I examine whether corporate investment decisions contain information about earnings quality. In a simplified setting, managers invest more in projects with a higher net present value (NPV). All else being equal, if a firm’s expected future earnings or permanent earnings increase, then it makes additional investment, because permanent earnings are equivalent to annuitized NPV (Black [1980], Beaver [1998], Ohlson and Zhang [1998]). Hence, if a firm experiences an increase in reported earnings and management views this earnings innovation to be permanent (i.e., the reported earnings have high “quality”), then that firm usually increases its investment level. However, if the innovation in reported earnings is purely transitory, then there should not be a corresponding change in the investment level. This reasoning suggests that earnings surprises that are more associated with changes in corporate investment decisions are more likely to be permanent and of higher quality than are earnings surprises that are less associated with such changes. Inferring earnings quality from corporate investment decisions has limitations. Because of agency problems, managers have incentives to overinvest for empire building and other reasons (Stein [2003]). As a result, project profitability does not solely determine observed investment decisions and this reduces these decisions’ informativeness for assessing earnings quality. Ultimately, whether one can derive useful and reliable measures of earnings quality from management investment decisions is an empirical question. In this paper, I provide empirical evidence to answer this question by first examining whether corporate investment-based earnings-quality measures are informative about future earnings. I measure earnings quality by examining decisions regarding capital and labor investments, two of the 1 See, for instance, Sloan [1996], Dechow and Dichev [2002], Francis LaFond and Olsson [2005], Basu [1997], Collins Maydew and Weiss [1999], Francis and Schipper [1999], and Ecker Francis and Kim [2006]. 2 I define earnings quality as the closeness of reported earnings to the “permanent earnings” following Dechow and Schrand [2004] and use earnings persistence to operationalize this concept. 3 Consistent with this argument, prior papers find that managerial decisions, which include dividend policy (Skinner and Soltes [2009]) and disclosure quality (Li [2008]), contain information about earnings quality.

EARNINGS QUALITY AND CORPORATE INVESTMENT

3

most important investment decisions managers make. I construct earningsquality measures by regressing changes in the number of employees and the amounts of capital and R&D expenditures on the change in reported earnings, using rolling data from the last 10 years for every firm year. The slopes from these regressions capture the sensitivity of investment to the innovation in reported earnings and measure the information content of those earnings for expected future profitability as reflected in corporate investment decisions. Because this regression approach requires a long time series of data, I also construct two other earnings-quality measures calculated as the current changes in capital investment and labor investment (as proxied by the number of employees) divided by the change in reported earnings for every firm year. The empirical findings can be summarized as follows. First, the earningsquality measures based on corporate investment decisions do not correlate highly with other commonly used measures of earnings quality. This finding suggests that the information contained in corporate investment decisions differs somewhat from that reflected by stock prices and the properties of historical accounting numbers. I then show that earnings-quality measures based on corporate investment decisions positively associate with earnings persistence. Furthermore, the predictive power of investment-based earnings quality for earnings persistence still holds and is economically meaningful even after controlling for typical measures of earnings quality such as the absolute amount of accruals (Sloan [1996]), estimation errors in accruals (Dechow and Dichev [2002]), the earnings–returns association, and the volatility of earnings and accruals. I also find that the investmentbased earnings quality contains significant information about future earnings only for firms with a relatively low tendency to overinvest, measured using the amount of free cash flows, the sensitivity of investment to cash flows, and the amount of excess investment based on Richardson [2006]. Finally, I find that the investment decisions are more informative about future earnings for capital-intensive firms, firms from highly unionized industries, and firms with a high frequency of earnings increases. These additional tests further validate the utility of corporate investment decisions for assessing earnings quality. Overall, the evidence indicates that there is substantial information in corporate investment decisions about earnings quality and it is incremental to other commonly used earnings-quality measures. However, researchers or investors need to consider the severity of possible overinvestment due to agency problems when using the investment-based earnings-quality measures. The remainder of this paper is organized as follows. Section 2 discusses the literature on earnings quality and defines earnings-quality measures based on corporate investment decisions. Section 3 presents the data, summary statistics of the measures, and their relation with firm characteristics and other earnings-quality measures used in the literature. Section 4 provides a discussion of the empirical results, and section 5 concludes.

4

F. LI

2. Literature and Hypotheses Development 2.1

LITERATURE ON EARNINGS QUALITY

There is an extensive literature on the earnings quality (see Dechow and Schrand [2004] for a comprehensive review). However, because of its context-dependent nature, there is no consensus on the underlying conceptual construct that “earnings quality” represents. In this paper, I follow Dechow and Schrand [2004] and define earnings quality as the closeness of reported earnings to the “permanent earnings” (Black [1980], Beaver [1998], and Ohlson and Zhang [1998]). I use earnings persistence to operationalize this concept. Earnings quality varies with many factors, including a firm’s business model and economic situation, estimation errors (Dechow and Dichev [2002]), and earnings management (Healy and Wahlen [1999]). To capture earnings quality, prior studies generally follow one of two approaches. The first approach measures earnings quality by using properties of the observed accounting numbers. The measures based on this approach include the level of accruals (Sloan [1996]), the estimation error in accruals (Dechow and Dichev [2002]), and the volatility of earnings (Dichev and Tang [2009]). Because of the historical nature of the current accounting system, the information contained in accounting numbers is unlikely to be complete concerning future profitability. The second approach focuses on the association between earnings and stock returns (e.g., Basu [1997], Collins Maydew and Weiss [1999], Francis and Schipper [1999], and Ecker Francis and Kim [2006]). This approach assumes market efficiency and extracts information about future earnings from stock prices. I take a different approach by emphasizing the management perspective. Managers arguably have more complete information about earnings quality than do outsiders. Therefore, earnings-quality measures based on the information set of managers can provide better proxies for earnings quality than measures that are based on historical accounting numbers or on the information set of outside equity investors.

2.2

EARNINGS QUALITY BASED ON FIRM INVESTMENT DECISIONS

In a simplified setting, making corporate investment decisions is straightforward: a firm invests more if the marginal NPV of the investment project is positive. In accounting terms, the NPV of future investment is a monotonic function of the expected “permanent earnings,” which is essentially the annuitized NPV (Black [1980], Beaver [1998], and Ohlson and Zhang [1998]). Therefore, a firm invests (disinvests) if its permanent earnings increase (decrease). Ignoring potential agency problems, the association between firms’ observed investment decisions and reported earnings captures the closeness of the reported earnings to the permanent earnings. Hence, the investment-earnings association provides information on the quality of the reported earnings. To the extent that managers have private information that investors do not have, corporate investment decisions can

EARNINGS QUALITY AND CORPORATE INVESTMENT

5

provide informative signals about earnings quality relative to the marketbased earnings-quality measures. One point worth discussing is the parallel between a firm’s investment decision and the valuation of its equity by outside investors. In both cases, the involved parties (managers or investors) want to make their decisions by doing valuations based on the expected profitability of the firm. Measuring earnings quality by using stock price information requires a maintained assumption of stock market efficiency, but assessing earnings quality through managerial investment decisions relies on the assumption that managers make optimal investment decisions. I seek to contribute to the literature that explores the implications of managerial decisions for earnings quality. Skinner and Soltes [2009] study the information content of dividend decisions by firms for earnings quality. Investment and dividend policies are both important managerial decisions and are likely to contain information about future earnings. Compared with dividend policy, firm labor and capital-investment decisions are simpler in the sense that they are less likely to be influenced by signaling considerations. A subtle difference between this paper and Skinner and Soltes [2009] is that their emphasis is on testing the dividend information content hypothesis, which has been examined in the dividend literature and has not received much support. The purpose of this paper is to test whether corporate investment decisions, despite potential agency problems, can provide information about earnings quality that is incremental to other typical earnings-quality constructs. Consequently, it is important to control for other measures of earnings quality. Empirically, I examine two types of investment decisions and their associations with reported earnings: labor and capital investment. Labor and capital are the two major factors that determine the output of a firm and they are also the main managerial decision parameters in microeconomics. The labor- and capital-investment decisions can be affected by different economic factors. Therefore, examining both decisions can complement each other.

2.3

OVERINVESTMENT AND INVESTMENT- BASED EARNINGS QUALITY

In this subsection, I explore the implications of firms’ nonoptimal investment decision making for the investment-based earnings quality. My motivation is the existence of agency problems, a central theme in the corporate finance literature with a lineage going back to Berle and Means [1932] and Jensen and Meckling [1976]. Agency problems can lead to overinvestment by managers (Stein [2003] provides a comprehensive review of the literature). One consequence of the agency problem is that managers have an excessive taste for running large firms, as opposed to simply profitable ones. This “empire building” tendency is emphasized by Williamson [1964], Donaldson [1984], and Jensen [1986], among many other studies. Agency problems can also give rise to overinvestment through channels other than “empire building.” Bertrand and Mullainathan [2003] argue that a managerial preference for the “quiet life”—effectively, a resistance to

6

F. LI

change—can lead to excessive continuation of negative-NPV projects. In a somewhat similar vein, Baker [2000] builds a model in which reputational concerns deter managers from discontinuing negative-NPV projects, because this would be an admission of failure. Many empirical studies provide evidence on corporate overinvestment, including evidence from specific industries (e.g., the oil industry overinvestment documented by Jensen [1986]) and evidence of poor acquisitions (Blanchard de Silanes and Shleifer [1994]). More recently, Richardson [2006] finds that investment decisions by firms are excessively sensitive to current cash flows, which is a symptom of overinvestment. Ceteris paribus, if a firm is more likely to overinvest, its labor- and capitalinvestment decisions are less likely to be a useful signal of earnings quality because the investment decisions can be affected by other considerations (e.g., empire building motivations) and are not solely determined by the profitability of the project. I therefore examine whether the information content of the investment-based earnings-quality measures varies with the overinvestment tendency cross-sectionally. I use three empirical constructs to measure the overinvestment tendency. First, Richardson [2006] shows that firms with a large amount of free cash flows tend to overinvest. This finding implies that investmentbased earnings-quality measures are less informative for firms with more free cash flows. Second, a high sensitivity of investment to the free cash flows available for investment can indicate potential agency problems (Stein [2003]).4 Richardson [2006] also finds that firms that tend to overinvest have higher investment–cash flow sensitivity. Hence, I use the investment–cash flow sensitivity as the second measure of the overinvestment tendency. Third, I rely directly on the overinvestment measure constructed by Richardson [2006] for the cross-sectional tests. Because this measure directly relates to capital expenditure, I focus on the capital investment-based earnings-quality measures. To summarize, I expect that for firms that are likely to overinvest (i.e., firms that have more free cash flows, higher sensitivity of investment to cash flows, and more excess investment), the investment-based measures of earnings quality less strongly associate with earnings persistence and are less useful in predicting future earnings.

3. Estimation of Earnings Quality Based on Investment Decisions 3.1

EMPIRICAL ESTIMATION OF EARNINGS QUALITY

I obtain my sample from the Compustat annual industrial and research files between 1952 and 2004. For every firm i in year T , I estimate the 4 The corporate finance literature also argues that firms with higher investment–cash flow sensitivity tend to have more severe financing constraints, in addition to overinvestment problems. Nevertheless, the financial constraint interpretation of the investment–cash flow sensitivity also leads to a prediction of suboptimal investment for firms with higher investment–cash flow sensitivity.

EARNINGS QUALITY AND CORPORATE INVESTMENT

7

following two regressions using the data from year T − 9 to T : (NEMP i,t − NEMP i,t−1 )/TAi,t−1 = αL,iT + βL,iT (E i,t − E i,t−1 )/TAi,t−1 + L,it (1) and (CAPX i,t + RND i,t − CAPX i,t−1 − RND i,t−1 )/TAi,t−1 = αC,iT + βC,iT (E i,t − E i,t−1 )/TAi,t−1 + C,it ,

(2) where T − 9 ≤ t ≤ T , NEMP is the number of employees at the end of a fiscal year (#29 of the Compustat annual files), CAPX is the amount of capital expenditure for the year (#128), RND is the amount of R&D expenditure for the year (#46), E is the operating earnings (#178) with some possible adjustment (details in the next paragraph), and TA is the book value of assets at the end of the fiscal year (#6). Similar to prior studies (e.g., Richardson [2006]), I include R&D expenditure in the capital investment together with capital expenditure. My motivation is the fact that, even though R&D is fully expensed under current U.S. generally accepted accounting principles (U.S. GAAP), prior studies show that the market views it more like an investment (Lev and Sougiannis [1996]). There is also a potential endogeneity problem in the estimation procedure—an increase in capital and R&D expenditures in a given year reduces the reported earnings because of additional expensing. This reduction leads to a mechanical negative relation between investment and earnings. To mitigate this problem, I adjust the reported earnings by adding back the current R&D expense and the depreciation expense due to the new capital expenditure. Specifically, for firm i in year t, E it is calculated as E it = #178 + RND it + CAPX it /(PPE it /DEP it ),

(3)

where PPE is the average of the beginning and end values of the gross amount of property, plant, and equipment (#7) and DEP is the depreciation expense (#14). This adjustment assumes that the new assets, due to current capital expenditure, are depreciated using the same rate as existing assets-in-place with a straight-line depreciation method. Item #29 of Compustat represents the number of company workers (including all employees of consolidated domestic and foreign subsidiaries, all part-time and seasonal employees, full-time equivalent employees, and officers, and excluding consultants, contract workers, directors, and employees of unconsolidated subsidiaries) as reported to shareholders.5 The amount of salary expense is a better variable to capture the amount of investment in labor than the number of employees, especially when the scaling variable is the book value of assets. However, only about 20% of firm years in Compustat report a nonmissing labor expense (#42), but about 95% of the firm 5 This figure is reported by some firms as an average number of employees and by others as the number of employees at year end. If both are given, the year-end figure is used. There is no reason to believe that this difference introduces a systematic bias to our estimates.

F. LI

8

years have the number of employees (#29). Using the number of employees therefore can increase the number of observations dramatically. The number of employees scaled by the book value of assets captures the labor intensity (i.e., the number of employees per dollar of assets). To the extent that the salary per employee remains relatively stable over time for the same firm, scaling the number of employees by total assets is not a problem. In unreported analysis, I also redo all the tests by replacing the dependent variable in equation (1) with log(NEMP t /NEMP t−1 ) and the results are similar. I use a firm-level regression because managerial decisions most naturally apply at the firm level. I expect that a firm-level specification is better than cross-sectional specifications because the regression coefficients are likely to differ across firms because managers have firm-specific information about future profitability. The β L and β C are the “response coefficients” of corporate investment level to current reported earnings. For a firm to have β L and β C estimated in year t, it needs to have nonmissing data in the 10 years from t − 9 to t to estimate equations (1) and (2). Because this regression approach uses 10 years of rolling data and therefore shrinks the sample size greatly, I also construct two other measures of earnings quality based on investment decisions for firm i in year T as follows: γL,iT = (NEMP iT − NEMP i,T −1 )/(E iT − E i,T −1 )

(4)

and γC,iT = (CAPX iT + RND iT − CAPX i,T −1 − RND i,T −1 )/(E iT − E i,T −1 ), (5) where all the definitions of the variables are the same as in equations (1) and (2). Thus, γ L and γ C capture the response of investment to the change in earnings in a given year. Compared with β L and β C , the advantage of these measures is that they require only two years of data, but the cost is that they might not measure the association between earnings and investment precisely and that it is more difficult to interpret the measures when the change in earnings is negative.6 Also, I construct two measures that combine the information content of both the capital and labor investment decisions. To smooth out any possible nonlinear effect of the variables, I average the measures based on capital and labor decisions using their decile ranks EQ 1 = (Decile(βL ) + (Decile(βC ))/2

(6)

EQ 2 = (Decile(γL ) + (Decile(γC ))/2,

(7)

and

where De cile (·) is the decile rank of a variable in a year. 6

This is especially true during recession years.

EARNINGS QUALITY AND CORPORATE INVESTMENT

9

TABLE 1 Summary Statistics Variable βC RSQC βL RSQL γC γL DD ABSACC VOL CFO VOL ACC VOL EARN βR SALGRW MTB OPCYC DIV

N 34,594 34,594 34,594 34,594 83,848 83,848 34,594 83,848 34,594 34,594 34,594 28,948 34,594 74,348 77,100 83,628

Mean 0.187 0.166 0.035 0.223 0.209 0.050 0.024 0.060 0.060 0.052 0.044 3.234 0.088 1.506 0.051 0.702

Standard Deviation 0.561 0.185 0.067 0.223 1.584 0.533 0.016 0.060 0.032 0.030 0.029 5.815 0.064 1.422 0.549 0.458

25th Pctl −0.080 0.022 0.000 0.035 −0.442 −0.009 0.012 0.020 0.036 0.029 0.023 0.328 0.048 0.960 0.014 0

50th Pctl 0.142 0.096 0.013 0.147 0.162 0.009 0.020 0.043 0.053 0.045 0.038 1.678 0.085 1.183 0.020 1

75th Pctl 0.426 0.252 0.048 0.355 0.982 0.078 0.031 0.080 0.077 0.067 0.058 4.391 0.125 1.640 0.036 1

p-Value 0.00 – 0.00 – 0.00 0.00 – – – – – 0.00 – – – –

This table presents the summary statistics of the investment-based earnings-quality measures and other variables. The p-value is for the test that examines whether the variable is significantly different from zero. The β C is estimated for every firm year as the slope coefficient from the regression of change in capital and R&D expenditures (scaled by lagged book value of assets) on the change in reported earnings (adjusted for the impact of current capital and R&D expenditures and scaled by lagged book value of assets) using data from the last 10 years. The RSQ C is the adjusted R-squared from the regression. The β L is estimated for every firm year as the slope coefficient from the regression of change in the number of employees (scaled by lagged book value of assets) on the change in reported earnings (scaled by lagged book value of assets) using data from the last 10 years. The RSQL is the adjusted R-squared from the regression. For a firm year to have β C and β L , it must have nonmissing data for the last 10 years. The γ C is the change in capital and R&D expenditures divided by the change in reported earnings. The γ L is the change in the number of employees divided by the change in reported earnings. The DD is estimated for every firm year as the standard deviation of the Dechow and Dichev [2002] residuals from the regression of accruals on lagged CFO (operating cash flows), current CFO, and next year’s CFO using the last 10 years of data. The ABSACC is the absolute amount of accruals scaled by lagged book value of assets. The VOL CFO, VOL ACC, and VOL EARN are the volatility of operating cash flows, accruals, and earnings (all scaled by lagged book value of assets) calculated using data from the last 10 years. The β R is estimated for every firm year as the slope coefficients from the regression of stock returns on the change in reported earnings (scaled by lagged book value of assets) using data from the last 10 years. The SALGRW is the average sales growth in the last 10 years. The MTB is the market value of a firm’s asset divided by the book value of the assets. The OPCYC is the operating cycle of a firm, calculated as 360/(Sales Average AR) +360/((Cost of Goods Sold)/(Average Inventory)). The DIV is a dummy variable that equals 1 if a firm pays dividend and 0 otherwise.

To summarize, β L , β C , and E Q 1 (i.e., the decile rank average of β L and β C ) capture the earnings quality inferred from labor- and capitalinvestment decisions by using a regression approach; γ L , γ C , and E Q 2 (i.e., the decile rank average of γ L and γ C ) represent the response of corporate investments to earnings innovation in the current year.

3.2

SUMMARY STATISTICS

Table 1 presents the summary statistics of the investment-based earningsquality measures and other variables needed in later analysis. Every year, if a firm has nonmissing data for the past 10 years, its β L is the estimation of the slope coefficient in the regression of the change in employment on the change in earnings following equation (1), and β C is the estimation of

10

F. LI

the slope coefficient in the regression of the change in capital investment on the change in earnings using equation (2). Because of the 10-year data requirement, the sample size is relatively small (34,594 firm years or about 800 firms per year).7 As predicted, current changes in reported earnings positively relate to the contemporaneous changes in the level of labor and capital. As indicated in table 1, the mean coefficient on earnings change in the labor regression (β L ) is 0.035 and the median is 0.013. The mean and median of β C are 0.187 and 0.142, respectively. The RSQL and RSQC are the adjusted R-squared from the regressions that have a mean of 0.223 and 0.166, respectively, and indicate that earnings changes can explain 17–22% of the variations in the changes of labor and capital investment. Based on the cross-sectional distribution of the coefficients, β L and β C are both statistically significant with p-values of 0.00. The variations in the two measures are substantial: the standard deviations of β L and β C are 0.067 and 0.561 and their inter-quartile ranges are 0.048 and 0.506, respectively. Table 1 also presents the summary statistics of γ L and γ C , the estimates of earnings quality defined in equations (4) and (5). Because the estimates only require two years of data, the sample size is much bigger—83,848 firm years (or about 2,000 firms per year). The γ L and γ C have a mean of 0.050 and 0.209, respectively and both have substantial variations with standard deviations of 0.533 and 1.584. The table also presents the summary statistics of other commonly used measures of earnings quality. The DD is the standard deviation of accruals that cannot be explained by cash flows as defined in Dechow and Dichev [2002], and is calculated using data from the last 10 years. The β R is the earnings–returns association calculated as the slope coefficient in the regression of stock returns on the change in reported earnings using data from the last 10 years. The ABSACC is the absolute amount of accruals scaled by lagged book value of assets. The V OL CF O, V OL ACC, and V OL E ARN are the volatilities of cash flows, accruals, and earnings (all scaled by lagged book value of assets) calculated using data from the last 10 years.

3.3

INVESTMENT- BASED EARNINGS QUALITY AND OTHER MEASURES OF EARNINGS QUALITY

In this subsection, I examine the associations between the earningsquality measures developed in this paper (i.e., β L , β C , γ L , and γ C ) and firm characteristics and other earnings-quality measures, such as the Dechow and Dichev [2002] measure, absolute value of accruals, volatility of earnings, earnings–returns association, growth, market-to-book ratio, and firm operating cycle. Table 2 presents the Pearson correlation coefficients between the investment-based earnings-quality measures and these variables.

7 To remove the influence of extreme observations, all variables are winsorized at the 1% and 99% levels.

βC

βL 0.23