Journal of Financial Economics 54 (1999) 423}460 The impact of cash #ow volatility on discretionary investment and the costs of debt and equity "nancingq Bernadette A. Minton!,*, Catherine Schrand" !Fisher College of Business, The Ohio State University, Columbus, OH 43210-1144, USA "The Wharton School, University of Pennsylvania, Philadelphia, PA 19104, USA Received 30 April 1998; received in revised form 20 November 1998; accepted 12 August 1999 Abstract We show that higher cash #ow volatility is associated with lower average levels of investment in capital expenditures, R&D, and advertising. This association suggests that "rms do not use external capital markets to fully cover cash #ow shortfalls but rather permanently forgo investment. Cash #ow volatility also is associated with higher costs of accessing external capital. Moreover, these higher costs, as measured by some proxies, imply a greater sensitivity of investment to cash #ow volatility. Thus, cash #ow volatility not only increases the likelihood that a "rm will need to access capital markets, it also increases the costs of doing so. ( 1999 Elsevier Science S.A. All rights reserved. JEL classixcation: G31 Keywords: Cash #ow volatility; Investment; Cost of equity "nancing; Cost of debt "nancing q Previous versions of this paper were titled &Costs of Accounting Income versus Cash Flow Volatility'. We thank Gordon Bodnar, John Core, Peter Easton, Chris GeH czy, Paul Gompers, Jarrad Harford (the referee), Bob Holthausen, Steve Kaplan, Andrew Karolyi, Sara Moeller, Tim Opler, Andre Perold, Tony Sanders, Bill Schwert (the editor), ReneH Stulz, Ralph Walkling, Franco Wong, and workshop participants at Dartmouth, the Federal Reserve Bank of New York, Harvard, Michigan, Minnesota, Ohio State, Purdue, Rochester, and Wharton for valuable comments, and Howard Yeh for research assistance. Minton thanks the Dice Center for Financial Economics for "nancial support. * Corresponding author. Tel.: #1-614-688-3125; fax: #1-614-292-2418. E-mail address: minton.15@osu.edu (B.A. Minton) 0304-405X/99/$ - see front matter ( 1999 Elsevier Science S.A. All rights reserved. PII: S 0 3 0 4 - 4 0 5 X ( 9 9 ) 0 0 0 4 2 - 2 424 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 1. Introduction &As risk managers, we spend much of our time examining the factors that cause cash #ows to #uctuate. This is important work, since low cash #ows may throw budgets into disarray, distract managers from productive work, defer capital expenditure or delay debt repayments. By avoiding these deadweight losses, risk managers can rightly claim they add to shareholder value'. (See Shimko, 1997.) Consistent with this claim that cash #ow volatility is costly, we document that cash #ow volatility is associated both with lower investment and with higher costs of accessing external capital. Higher cash #ow volatility implies that a "rm is more likely to have periods of internal cash #ow shortfalls. Our analysis indicates that "rms do not simply react to these shortfalls by changing the timing of discretionary investment to match cash #ow realizations. Rather, "rms forgo investment. Firms could smooth internal cash #ow #uctuations using external capital markets. However, Myers and Majluf (1984) show that external capital is more costly than internal capital. Consequently, "rms that require more external capital relative to internal capital will have lower investment, all else equal, assuming "rms follow the basic net present value (NPV) decision rule for capital budgeting. A higher frequency of cash #ow shortfalls, however, is not the only reason that volatility a!ects investment decisions. Cash #ow volatility also is positively related to a "rm's cost of accessing external capital. Volatility can a!ect capital costs because of capital market imperfections including information asymmetry and contracting (e.g., debt covenants). For example, consider that analysts are less likely to follow "rms with volatile cash #ows. Assuming that lower analyst following implies greater information asymmetry and a higher cost of accessing equity capital, "rms with higher cash #ow volatility will have higher equity capital costs. Together, the two e!ects of cash #ow volatility imply that reductions in cash #ow volatility through risk management activities can reduce a "rm's expected &underinvestment' costs (Froot et al., 1993; Myers, 1977). The basic "nding of the analysis is that cash #ow volatility is associated with lower investment in average annual capital expenditures, research and development costs, and advertising expenses, even after industry-adjusting and controlling for the level of a "rm's average cash #ows and its growth opportunities. In addition, "rms experiencing cash #ow shortfalls in a given year relative to their peers or relative to their own historical experience have signi"cantly lower discretionary investment in that year than "rms that are not experiencing shortfalls. Sensitivity analyses indicate that the results are not driven by "rms in "nancial distress or cross-sectional di!erences in investment opportunities. Fazzari, Hubbard, and Petersen (FHP, 1988,1998), Hoshi et al. (1991), Kaplan and Zingales (KZ, 1997), and Lamont (1997) "nd a negative contemporaneous relation between annual investment levels and liquidity. These studies cannot B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 425 distinguish whether "rms with volatile cash #ows time their investment decisions to match internal cash #ow realizations or actually decrease their overall level of investment. Our "ndings reveal a negative relation between volatility, measured over a period, and the average level of investment measured over the same period, suggesting that "rms that experience shortfalls ultimately forgo investment. The magnitude of the forgone investment is large. Capital expenditures by "rms with high cash #ow volatility (in the highest quartile relative to "rms in the same industry) are 19% below the mean level of capital expenditures for the sample while capital expenditures by "rms with low cash #ow volatility are 11% above the mean. Three pieces of related evidence emerge from tests designed to further explain our basic "nding. First, the negative relation between volatility remains after controlling for a "rm's cost of accessing external capital. Second, there is a direct relation between capital costs and investment levels. Moreover, "rms that we claim have higher costs of accessing external capital (e.g., small "rms) have a higher sensitivity of investment to volatility. Third, cash #ow volatility is positively related to the costs of accessing external capital. Speci"cally, higher cash #ow volatility is associated with worse S&P bond ratings, higher yields-tomaturity, lower analyst following, lower dividend payout ratios, higher bid}ask spreads, and higher weighted average costs of capital. Taken together, the evidence suggests that the basic "nding of an association between investment and cash #ow volatility is not just a relation between investment and project risk in disguise. The results provide a benchmark for assessing the value of risk management activities. However, the sensitivity of investment to volatility does not suggest that "rms should necessarily reduce or eliminate cash #ow volatility. We recognize that volatility is a choice variable and assume that managers make rational decisions based on all available information. Our results provide an additional source of information that managers can use to assess the bene"ts of reducing cash #ow volatility. Firms must weigh these bene"ts against the costs, which can vary across "rms and industries. Risk management costs are likely to be low, for example, for "rms in the oil and gas, mining, and agriculture industries where liquid, well-developed derivatives markets exist for a risk that represents a signi"cant source of a "rm's cash #ow volatility. In contrast, hedging costs are likely to be higher for "rms in which signi"cant cash #ow volatility results from factors that are relatively uncorrelated with interest rates, foreign exchange prices, or commodity prices. The cross-sectional variation in these costs, relative to the potential bene"ts of reduced volatility, leads to interesting cross-sectional implications about risk management decisions. The positive association between a "rm's current cost of external capital and its historical cash #ow volatility is a subtle but important distinction for risk managers. One interpretation of this result is that debt and equityholders use historical volatility to predict future cash #ow volatility when they set prices. 426 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 This interpretation implies that a "rm's cost of accessing capital will depend on the expected persistence of cash #ow volatility into future periods. Hence, cross-sectional di!erences in the persistence of the e!ects of risk management activities will be associated with cross-sectional di!erences in the association between volatility and the "rm's cost of accessing capital. In the extreme case, risk management activities that have successfully reduced volatility, but which are not expected to have a persistent e!ect on volatility in future periods, will not necessarily reduce a "rm's current cost of accessing external markets. One conjecture is that debt and equityholders do not view the use of short-term "nancial derivatives to reduce volatility in the same way as the use of longerterm risk reduction activities, such as moving a plant overseas to reduce foreign exchange price risk. Understanding how di!erent types of risk management activities a!ect the costs that we document is an interesting avenue for future research. Although this paper provides the "rst direct evidence that cash #ow volatility is related to lower investment, we are not the "rst to make this claim. Shapiro and Titman (1986), Lessard (1990), Stulz (1990), and Froot et al. (1993) propose a link between volatility and investment in the context of explaining hedging activities that reduce cash #ow volatility. Consistent with these theories, Dolde (1995), GeH czy et al. (1997), Mian (1996), Nance et al. (1993), and Tufano (1996) "nd that "rms that have the greatest expected bene"ts from reducing volatility are more active in risk management activities. These papers jointly test two hypotheses: (1) volatility is costly for the reasons predicted by a particular theory (or theories), and (2) "rms engage in a speci"c risk-management activity (such as using derivatives) to reduce the volatility that creates the cost. Our direct evidence of an association between volatility and discretionary investment complements the "ndings of these indirect tests. The paper proceeds as follows. Section 2 provides an outline of the various predictions and tests. Section 3 describes the measure of cash #ow volatility and the methodology for the analysis of the association between volatility and investment. Section 4 reports the results of these tests. In Section 5, we examine the relation between costs of accessing capital markets and investment. Section 6 presents the analyses of the relations between cash #ow volatility and these costs. Section 7 provides concluding remarks. 2. Overview of the paper This paper analyzes a large and representative sample of "rms over a sevenyear period. The primary advantage of this sample is that the evidence can be generalized to a broad class of "rms and investment decisions. A disadvantage is that the results are particularly susceptible to criticisms related to endogeneity issues and omitted correlated variables, despite the use of industry-adjusted data B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 427 in the analysis. To mitigate these concerns, we perform three separate sets of tests of the e!ects of cash #ow volatility on investment and the costs of accessing capital, and we support these tests with numerous sensitivity analyses. While there may be questions about the interpretation of the results from any of the three individual tests, the results taken together support our conclusions. This section provides a detailed outline of the approach. The "rst analysis examines the direct association between investment and cash #ow volatility. We predict that a "rm's cash #ow volatility during a period will be negatively associated with its average discretionary investment measured over the same period. We test for this negative relation using annual crosssectional regressions of industry-adjusted capital expenditures, research and development costs, and advertising expenses on industry-adjusted cash #ow volatility. The methodology is described in Section 3 and the results are presented in Section 4.1. One interpretation of the negative relation is that volatility captures the likelihood that a "rm experiences a cash #ow shortfall. Further evidence to support this interpretation is based on an examination of "rm-level investment during periods of cash #ow shortfalls (Section 4.2). Section 4 also includes robustness checks of the regression results to assess whether the negative association between volatility and investment merely represents the relation between investment and "rm characteristics that are correlated with cash #ow volatility but omitted from the analysis. In particular, prior research on "rms' investment decisions has found that investment is positively related with cash #ow levels and suggests that the investment-cash #ow sensitivities could di!er for "nancially constrained and healthy "rms.1 Section 4.1 includes an assessment of cross-sectional di!erences in investmentvoltatility sensitivities across cash #ow levels. Section 4.3 examines alternative explanations for the results. Speci"cally, we address concerns about the causality of the relation between investment and volatility, the impact of "nancially distressed "rms on the results, and variable speci"cation issues. The remainder of the paper examines whether the source of the negative relation between investment and volatility is a positive relation between volatility and the costs of external "nancing. In Section 5, we re-estimate the sensitivity of investment to cash #ow volatility as in the "rst analysis and include a proxy for a "rm's cost of accessing external capital and an interaction variable that is the product of this proxy and cash #ow volatility. We predict that "rms with higher costs of accessing capital will have lower investment, all else equal. The coe$cient on the proxy measures the direct association between capital costs 1 FHP (1988) show that investment-cash #ow level sensitivities are greater for "rms with low dividend payout ratios. The perspective of the FHP paper is that investment-cash #ow sensitivities proxy for a "rm's degree of "nancing constraint. However, there is some debate about the interpretation of the FHP results, with the debate focusing on the de"nition of "nancing constraints (i.e., KZ, 1997; FHP, 1988,1998) 428 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 and investment, while the coe$cient on the interaction term indicates whether the negative association between volatility and investment is a!ected by crosssectional di!erences in the costs of accessing external capital. We use nine proxies for these costs: S&P bond ratings, yields-to-maturity on debt, stock market betas, total equity price risk, weighted average costs of capital, analyst following, dividend payout ratios, "rm size, and bid}ask spreads on common stock. In Section 6, we examine the associations between the proxies for the costs of accessing external capital (except "rm size) and cash #ow volatility. We estimate eight separate sets of annual cross-sectional regressions that measure the associations between cash #ow volatility and each of the proxies for capital costs. Based on existing theory and empirical research, we predict that cash #ow volatility is positively associated (in most cases) with costs of accessing external capital. The regressions include controls for a "rm's level of cash #ows as well as variables that have been identi"ed in prior research as determinants of the proxies. 3. Methodology Section 3.1 de"nes cash #ow and the methodology for measuring cash #ow volatility. Section 3.2 de"nes the proxies for investment and the regression equations used to estimate the association between investment and cash #ow volatility. 3.1. Measures of cash yow and cash yow volatility Operating cash #ow is computed quarterly for all non-"nancial "rms on Compustat as sales (Compustat data item 2) less cost of goods sold (item 30) less selling, general and administrative expenses (item 1) less the change in working capital for the period. Working capital is current assets other than cash and short-term investments less current liabilities and is calculated as the sum of the non-missing amounts for accounts receivable (item 37), inventory (item 38), and other current assets (item 39) less the sum of the non-missing amounts for accounts payable (item 46), income taxes payable (item 47), and other current liabilities (item 48). Quarterly selling, general and administrative expenses exclude one-quarter of annual research and development costs (item 46) and advertising expenses (item 45) when those data items are available. Thus, operating cash #ow represents the cash #ow available for discretionary investment. Cash #ow volatility is de"ned as the coe$cient of variation in a "rm's quarterly operating cash #ow over the six-year period preceding each of the seven sample years from 1989 through 1995. Thus, for the sample year 1995, the B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 429 Table 1 Summary of "rms and industries in the sample The sample includes all "rms in two-digit SIC industries in which there are at least ten "rms with operating cash #ow data available. The number of sample "rms reported is the number with data available for 1995. The "rms are on the Compustat quarterly data tapes from 1989 to 1994 Industry name Two-digit SIC code Number of sample "rms Metal mining Oil and gas extraction Building construction * general contractors, operative builders Non-building construction Food and kindred products Textile mill products Apparel and other "nished products Lumber and wood products, except furniture Furniture and "xtures Paper and allied products Printing, publishing and allied Chemicals and allied products Petroleum re"ning and related industries Rubber and miscellaneous plastic products Stone, clay, glass, and concrete products Primary metal industries Fabricated metal, except machinery, transportation equipment Machinery, except electrical Electrical, electrical machinery, equipment, supplies Transportation equipment Measuring instruments; photographic goods; watches Miscellaneous manufacturing industries Water transportation Communication Electric, gas, sanitary services Durable goods * wholesale Non-durable goods * wholesale General merchandise stores Food stores Apparel and accessory stores Furniture, home furnishings stores Eating and drinking places Miscellaneous retail Business services Amusement, except motion pictures Health services Environmental services 10 13 15 16 20 22 23 24 25 26 27 28 29 30 32 33 34 35 36 37 38 39 44 48 49 50 51 53 54 56 57 58 59 73 79 80 87 20 64 14 11 47 24 22 14 15 30 36 105 33 30 16 48 49 125 106 43 63 22 10 19 23 41 33 22 16 16 11 24 32 57 13 19 14 1,287 430 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 coe$cient of variation is calculated using 24 quarters of data from the "rst "scal quarter of 1989 to the fourth "scal quarter of 1994. A "rm is included in the sample for a given year if it has at least 15 non-missing observations during the 24 quarters. The coe$cient of variation is the standard deviation of operating cash #ow scaled by the absolute value of the mean over the same period. The resulting metric is a unitless measure of variation that has been used by Albrecht and Richardson (1990) and Michelson et al. (1995). The coe$cient of variation of each "rm-year observation is adjusted relative to the median for all sample "rms in the same two-digit SIC code for the same sample year. This continuous measure of a "rm's industry-adjusted coe$cient of variation in operating cash #ows is denoted CVCF. Industry-adjusted coe$cients of variation control for di!erences across industries in the quarterly seasonality of cash #ows and in the nature of "rms' operations. Because of the industry adjusting, we eliminate "rms in industries with less than ten "rms with available data. We also delete seventy "rm-year observations representing twenty-six "rms that are classi"ed as being in reorganization or liquidation based on their Standard & Poor's (S&P) stock ratings. The "nal annual samples consist of between 897 "rms (1989) and 1287 "rms (1995) with available operating cash #ow data. Table 1 summarizes the number of "rms by industry for the 1995 sample. The distributions of "rms in other sample years are similar. The sample represents "rms in 37 separate two-digit SIC codes and is consistent with the distribution of "rms on Compustat except that our sample excludes "rms in the "nancial services industry. 3.2. Volatility and discretionary investment The following model examines the relation between investment and volatility: INVESTMENT"a #a CVCF# + a CONTROL #e . (1) 0 1 i i 1 i/2,3 INVESTMENT is one of three proxies for discretionary investment: capital expenditures, R&D costs, or advertising expenses. Capital expenditures (Compustat data item 90), R&D costs (item 46), and advertising expenses (item 45) are all scaled by the "rm's total assets at the beginning of the year. The extant literature on the sensitivity of investment to cash #ow levels, including (FHP, 1988,1998) and KZ (1997), scales the only proxy for investment, capital expenditures, by beginning of period total "xed assets. In this paper, scaling by total assets provides a consistent scaling variable across all three proxies for investment.2 2 The results are qualitatively similar if the investment variables are scaled by sales revenue. B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 431 We compute average capital expenditures, R&D costs, and advertising expenses for the same rolling six-year periods over which we measure volatility. Because the average investment variables are measured contemporaneously with volatility, the results of the regression analyses indicate whether "rms with higher volatility during a given period make lower average investments during that same period. The three proxies for discretionary investment are adjusted relative to the median for all sample "rms in the same two-digit SIC code for the same sample year. Industry-adjusting the proxy variables for investment controls for variation across industries in capital intensity and growth during the sample period. The model includes two control variables (CONTROL) that measure growth. FHP (1988) identify sales growth as a signi"cant determinant of capital expenditures. Sales growth is the average annual change in sales, scaled by beginning of period sales, for the same rolling six-year periods as volatility. The second proxy for growth opportunities is the average annual book-to-market ratio, measured for the same rolling six-year periods as volatility.3 FHP and KZ also include additional "rm and industry characteristics in the regression equations that estimate the determinants of investment. However, it is the growth variables that are consistently signi"cant across various studies. Like the dependent variable and the coe$cient of variation, the control variables are industryadjusted. Eq. (1) is estimated annually using ordinary least squares regressions for each of the seven samples from 1989 to 1995. We present the means of the annual coe$cient estimates. To test the hypothesis that the mean coe$cient estimate is statistically di!erent from zero, we calculate and report a Z-statistic (Z"tM /[p(t)/J(N!1)]) where tM and p(t) are the average and standard deviation of the annual t-statistics, respectively, and N is the number of annual observations.4 We estimate two equations that control for the potential relation between investment and cash #ow levels. Eq. (2) partitions the sample "rms based on the level of industry-adjusted cash #ows: INVESTMENT"b #b LO#b HI#b CVCF#b CVCF]LO 0 1 2 3 4 #b CVCF]HI# + b CONTROL #e . 5 i i 2 i/6,7 (2) 3 Both FHP (1988) and KZ use variants of Tobin's Q as a proxy for growth opportunities. KZ measure Tobin's Q as the ratio of the market value of assets to the book value of assets. FHP measure Tobin's Q using replacement costs. 4 An alternative test statistic is Z@"1/J¹ +N t /Jk /(k !2) where t is the t-statistic for year t/1 i i i i i and k is the degrees of freedom. Z@ assumes the annual parameter estimates are independent and is i likely overstated; Z corrects for the potential lack of independence. !0.609 !0.616 0.073 136 0.016 0.012 0.030 129 Industry-adjusted CV of operating cash #ow Mean !1.713 !0.905 Median !1.504 !0.901 Std Deviation 0.565 0.096 N 101 145 Industry-adjusted operating cash #ow Mean 0.020 Median 0.013 Std Deviation 0.032 N 89 0.015 0.011 0.033 129 Decile 3 Decile 2 Decile 1 (LOW) !0.000 !0.000 0.037 455 0.104 0.000 0.430 533 Deciles 4}7 Panel A: Firms ranked based on industry-adjusted coezcient of variation of operating cash yow !0.018 !0.015 0.050 90 1.949 1.937 0.497 119 Decile 8 !0.037 !0.026 0.070 91 4.775 4.403 1.498 120 Decile 9 !0.053 !0.037 0.067 86 26.382 17.747 21.250 120 Decile 10 (HIGH) Means, medians, and standard deviations of industry-adjusted coe$cients of variation (CV) and levels of operating cash #ow scaled by beginning of period total assets. The coe$cient of variation equals the ratio of the standard deviation to the absolute value of the mean of operating cash #ow, calculated using quarterly data from 1989 to "scal year end 1994. In Panel A, "rms are ranked on the basis of industry-adjusted coezcients of variation of operating cash #ow within each two-digit SIC code. Decile 1, for example, contains all "rms with coe$cients of variation that are in the lowest decile in their respective two-digit SIC code. Because each SIC code does not have a multiple of ten "rms, the number of observations varies across deciles. In Panel B, "rms are ranked on the basis of industry-adjusted levels of operating cash #ow scaled by beginning of period total assets. A "rm is classi"ed as LO, MED, or HI if it is in the lowest three, middle four, or highest three decile rankings in its industry, respectively. The results are reported for 1995. Results for other sample years are similar Table 2 Descriptive statistics of industry-adjusted coe$cients of variation and levels of operating cash #ow 432 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 !0.000 0.000 0.008 440 0.859 !0.300 5.483 431 Industry-adjusted CV of operating cash #ow Mean 6.574 Median 1.375 Std Deviation 14.103 N 319 MED (Deciles 4}7) !0.073 !0.037 0.321 329 Industry-adjusted operating cash #ow Mean Median Std Deviation N LO (Deciles 1}3) Panel B: Firms ranked based on industry-adjusted level of operating cash yow 0.368 !0.425 6.030 320 0.042 0.037 0.024 329 HI Deciles (8}10) B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 433 434 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 For each annual estimation, LO (HI) is an indicator variable equal to one if the "rm is in the lowest (highest) three deciles of the sample "rms when they are ranked based on the industry-adjusted average annual level of operating cash #ows scaled by beginning of period total assets. The inclusion of the cash #ow level variables controls for the observed sensitivity of investments to cash #ow levels documented by FHP (1988,1998), Cleary (1998), Hoshi et al. (1991), KZ (1997), and Lamont (1997).5 Alternatively, Eq. (3) is an augmented version of Eq. (1) that includes a continuous measure of industry-adjusted annual operating cash #ows (in levels) scaled by beginning of year total assets averaged over the same six-year period as volatility (OPCF): INVESTMENT"c #c OPCF#c OPCF2#c CVCF 0 1 2 3 #c CVCF]OPCF# + c CONTROL #e . (3) 4 i i 3 i/5,6 OPCF2 controls for potential nonlinearities in the relation between investment and industry-adjusted average annual operating cash #ow. The interaction variable (CVCF]OPCF) measures the impact of a "rm's cash #ow level on the estimated sensitivity of investment to cash #ow volatility. Table 2 provides descriptive evidence about the cash #ow volatility variable (CVCF) used in Eqs. (1)}(3) and average annual cash #ow levels. In Panel A, "rms are ranked into deciles (annually) based on industry-adjusted coe$cients of variation in operating cash #ows. Each "rm in decile one, for example, has a coe$cient of variation of operating cash #ow that is in the lowest ten percent relative to other "rms in its industry.6 Statistics are reported for decile one (the lowest volatility measure), decile two, decile three, the middle four deciles as a group (deciles four through seven), decile eight, decile nine, and decile ten (the highest volatility measure) for the sample year 1995. Results for other sample years (not reported) are similar.7 Panel A of Table 2 illustrates that the increases in the coe$cients of variation are non-linear across the deciles. This pattern emerges even though we remove the top ten percent of decile 10 (top one percent of the sample "rms). The mean 5 We also estimate speci"cations that include a third control variable which is the industryadjusted log of "rm size where size is de"ned as the market value of a "rm's equity plus the book value of its debt. The results are qualitatively similar to those reported. 6 Because the number of "rms in the two-digit SIC codes is not always a multiple of ten, each decile contains a di!erent number of observations. As an example, if there are 32 "rms in an industry for a sample year, SAS allocates the two extra observations to deciles three and six. If there are 34 "rms, SAS allocates the four extra observations to deciles two, four, six, and eight. 7 The number of observations per decile in Panels A and B are di!erent because cash #ow volatility data can be missing for "rms that have annual cash #ow level data. We require that a "rm has at least 15 quarters of non-missing data to calculate the volatility measure. B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 435 results are driven by some extreme observations. The medians follow a similar, although less dramatic, pattern. In the regression analyses, we set all coe$cients of variation that are greater than 100 equal to 100, which approximately represents the 99th percentile. In addition, in#uential observations in the annual regressions are downweighted by the method of Welsch (1980). Panel A also indicates that there is a negative association between CVCF and mean levels of average annual operating cash #ow. Industry-adjusted cash #ow levels range from 0.020 for "rms in the lowest decile ranked on CVCF to !0.053 for "rms in the highest decile. This pattern emerges despite the use of the coe$cient of variation to measure volatility, which reduces the likelihood of a mechanical relation between volatility and levels. This negative relation justi"es the use of levels as control variables in Eqs. (2) and (3). The negative association between volatility and levels is con"rmed in Panel B in which "rms are ranked into deciles (annually) based on industry-adjusted cash #ow levels. For "rms that have cash #ows that are in the lowest three deciles when compared to "rms in their respective two-digit SIC code (LO), the average CVCF is 6.574, compared to a CVCF of 0.368 for "rms in the highest three deciles (HI). Panel B also indicates that the standard deviations of the coe$cients of variation vary across "rms with di!erent cash #ow levels, which suggests that Eq. (2) may be mis-speci"ed. The speci"cation of Eq. (2) as a pooled regression with separate parameter estimates across groups is the most e$cient speci"cation only if the standard deviations of the independent variables are similar across the groups (Greene, 1993, p. 236). In the case of dissimilar variances, the appropriate technique is to estimate Eq. (1) separately for each group. The results from the separately speci"ed equations are qualitatively similar to those obtained from the pooled regression. 4. Results Sections 4.1 and 4.2 present empirical results on the relation between cash #ow volatility and investment. Section 4.3 provides robustness checks of the analyses. 4.1. Regression analysis of investment and volatility Table 3 reports the means of the annual coe$cient estimates from regression Eqs. (1)}(3) using industry-adjusted average capital expenditures, R&D costs, and advertising expenses as proxies for discretionary investment. We do not present the coe$cient estimates on the control variables.8 8 In all regressions in Table 3, the estimated coe$cients on the control variables for growth are signi"cant and of the predicted signs. 0.0025 (18.738) 0.0028 (9.791) 0.0065 (10.388) 0.0040 (8.597) 0.0106 (25.205) 0.0031 (2.708) R&D costs Advertising expenses !0.0013 (!0.459) !0.0071 (!21.002) !0.0032 (!4.971) 0.0170 (8.582) 0.0091 (6.861) 0.0010 (2.810) !0.0008 (!8.957) 0.0003 (0.501) !0.0006 !(5.574) !0.0008 (!3.121) !0.0002 (!20.212) !0.0003 (!4.843) CVCF !0.0007 (!0.831) 0.0008 (2.795) 0.0002 (2.361) LO HI Intercept LO CVCF* Intercept* Capital expenditures Dependent variable Panel A: Regressions using indicator variables to control for operating cash yow levels !0.0032 (!5.782) !0.0041 (!4.481) !0.0008 (!6.203) HI 3.27}4.83 0.13}1.18% 3.03}10.22 0.19}6.70% 4.59}7.93 2.60}4.26% Range of Adj. R2 Means of annual regressions of industry-adjusted capital expenditures, research and development (R&D) costs, and advertising expense (proxy variables for discretionary investment) on industry-adjusted operating cash #ow volatility (CVCF) and industry-adjusted sales growth and book-to-market ratios. Operating cash #ow volatility is measured as the coe$cient of variation of a "rm's quarterly operating cash #ow over the six-year period preceding each of the seven sample years from 1989 through 1995. The dependent variables and the control variables represent averages of annual amounts measured over the same six-year period. LO and HI are indicator variables equal to one if a "rm is in the lowest or highest three decile rankings, respectively, based on the level of its six-year average annual industry-adjusted operating cash #ows. Operating cash #ow is sales!cost of goods sold!selling, general and administrative expenses (excluding R&D and advertising)!the change in working capital. For each equation, means of the seven annual least squares values for each coe$cient (a6 ) are presented. Z-statistics to test the hypothesis that E(a6)"0 are shown in parentheses. In#uential observations in the it annual estimations are downweighted by the method of Welsch (1980). Coe$cient estimates on control variables included in the regressions (industry-adjusted book-to-market ratio and industry-adjusted sales growth) are not presented Table 3 Means of annual regressions of proxies for discretionary investment on cash #ow volatility 436 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 Intercept 0.0025 (20.547) 0.0042 (5.560) 0.0056 (11.468) Dependent variable Capital expenditures R&D costs Advertising expenses 0.2676 (25.329) 0.2050 (17.429) 0.0361 (5.955) OPCF 1.3667 (5.657) 0.4727 (3.967) !0.0780 (!2.088) (OPCF)2 !0.0006 (!4.184) !0.0007 (!4.484) !0.0003 (!12.413) CVCF Panel B: Regressions using continuous variables to control for operating cash yow levels !0.0059 (!2.749) !0.0118 (!2.867) !0.0033 (!2.235) CVCF]OPCF 5.24}7.26% 4.15}11.68% 3.71}5.93% Range of Adj. R2 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 437 438 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 Taken together, the results in Table 3 indicate that discretionary investment levels are sensitive to cash #ow volatility, and that the degree of the sensitivity is a function of cash #ow levels. In Panel A, in the regressions that include an intercept, the coe$cient of variation, and the control variables for growth, higher industry-adjusted operating cash #ow volatility is associated with signi"cantly lower industry-adjusted capital expenditures, research and development costs, and advertising expenses. However, the results from Eq. (2) indicate that volatility is not an equally signi"cant determinant of investment across all levels of cash #ows. For low cash #ow "rms, the negative relations between volatility and capital expenditures (!0.0003) and R&D costs (!0.0008) are mitigated (CVCF]LO"0.0002) or eliminated (CVCF]LO"0.0008), respectively. Advertising expenses are negatively associated with cash #ow volatility, but only for "rms with high levels of cash #ows (CVCF]HI"!0.0032). Although investment is not related to cash #ow volatility for low cash #ow "rms, these "rms exhibit a lower absolute level of average industry-adjusted capital expenditures and R&D costs than "rms with moderate or high cash #ow levels. The sum of the intercept and the coe$cient estimate on the LO dummy variable indicates that average annual capital expenditures as a percentage of total assets for this group are 0.0004 below the industry median. Similarly, R&D costs are 0.0031 below the industry median for low cash #ow "rms. Thus, the average level of a "rm's investment over time is lower for low cash #ow "rms, but cash #ow volatility does not have a signi"cant marginal e!ect on investment.9 The regression results presented in Panel B include the continuous measure of cash #ow levels (Eq. (3)). These results also indicate that volatility is negatively associated with investment, and that this relation varies across "rms as a function of cash #ow levels. As in Panel A, "rms with higher levels of cash #ows have higher levels of investments, ceteris paribus. The interaction variable between CVCF and OPCF has a negative and signi"cant association with each of the three proxies for discretionary investment. Thus, the sensitivity of investment to volatility is stronger as cash #ows increase. This result is consistent with the evidence in Panel A that the impact of volatility is second order relative to the e!ect of cash #ow levels for "rms with low cash #ows. By adding cash #ow volatility, OPCF2, and CVOPCF]OPCF to the regression, we enhance the explanatory power of the basic investment-liquidity model 9 We also estimate Eqs. (1)}(3) using the sum of non-missing capital expenditures, R&D costs, and advertising expenses as a measure of total discretionary investment, because "rms can potentially substitute across these investments. The results (not reported) show a negative and signi"cant relation between volatility and total investment in the analysis that excludes controls for a "rm's operating cash #ow level. In the analysis that controls for cash #ow levels, there is a signi"cant, negative relation between volatility and investment for moderate and high cash #ow "rms, consistent with the results for capital expenditures reported in Panel A of Table 3. B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 439 in the existing investment literature that includes only the level of operating cash #ow (OPCF) and the control variables for growth. The average incremental R2 from adding these variables to the basic speci"cation (averaged across the individual annual regressions) is 1.79, 2.48, and 2.52 for the regressions measuring the association between volatility and capital expenditures, R&D costs, and advertising expenses, respectively. 4.2. Cash yow shortfalls and investment As more direct evidence that shortfalls in cash #ows are associated with lower investment, we examine the capital expenditures, R&D costs, and advertising expenses of "rms that experience shortfalls. A "rm is de"ned to be in a shortfall position for a sample year if it is in the lower quartile of the sample based on one of four separate metrics: 1. average annual industry-adjusted operating cash #ows, or 2. average annual operating cash #ows that are not industry-adjusted, or 3. the di!erence between its annual operating cash #ows and its own historical average annual operating cash #ows measured over the prior six-year period, or 4. the di!erence between its annual industry-adjusted operating cash #ows and its own historical average industry-adjusted annual operating cash #ows measured over the prior six-year period.10 The "rst measure de"nes a shortfall relative to the "rm's industry peers and the second measure de"nes a shortfall relative to all "rms in the sample. The last two measures de"ne a "rm to be in a shortfall position relative to its own historical cash #ow levels. As a benchmark against which to evaluate the investments of these groups, we examine the investments of "rms in excess cash #ow positions (the upper quartile in each analysis).11 The results in Table 4 indicate that "rms that experience shortfalls relative to their industry peers or to the sample "rms have lower industry-adjusted levels of discretionary investment than "rms with excess cash #ow levels, consistent with 10 As in Table 2, we delete all "rm-year observations with a CVCF above the 99th percentile. The results are qualitatively the same if we include these observations. 11 The upper quartile is not an appropriate benchmark if "rms with &excess' cash #ow overinvest (e.g., Lang et al., 1991). We also compare the shortfall "rms to "rms in the third quartile. The di!erence between industry-adjusted capital expenditures of the two groups remains signi"cant when "rms are sorted based on industry-adjusted cash #ows, but the signi"cance of the di!erence between industry-adjusted R&D and advertising declines (p-values between 0.10 and 0.15). Comparing the shortfall "rms to the upper half of the sample, the di!erences between all three proxies for investment are signi"cant. 440 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 the regression results in the previous section.12 The di!erences are signi"cant at less than the 1% level in tests including observations from the full sample period. In annual tests, these di!erences are signi"cant at conventional levels in all sample years for all three proxies for investment. As Table 4 reports, industryadjusted capital expenditures are negative for "rms that experience cash #ow shortfalls. Thus, these "rms spend less on capital expenditures than the median "rm in their respective industries. However, for R&D and advertising, the results indicate only that the shortfall "rms spend less than "rms in excess cash #ow positions (i.e., the amounts are statistically lower but positive). Finally, when we de"ne a shortfall relative to a "rm's own historical cash #ows, only capital expenditures are signi"cantly di!erent from those of "rms with excess cash #ows over the full sample period. In annual tests, this di!erence is signi"cant in four of the seven years. 4.3. Sensitivity analysis The "rst sensitivity analysis examines whether "nancially distressed "rms drive the results in Table 3. Because cash #ow levels and cash #ow volatility are potentially correlated with a "rm's probability of "nancial distress, and "nancial distress is potentially correlated with investment decisions, we perform two analyses to examine the e!ects of "nancially distressed "rms on the results. First, we include in regression Eq. (3) an industry-adjusted measure of leverage as a proxy for "nancial distress. The leverage proxy equals the average annual debt-to-equity ratio de"ned as the book value of long-term debt scaled by the sum of the book values of long-term debt, common equity, and preferred stock. The coe$cient on this variable is negative and signi"cant, consistent with the prediction that more levered "rms invest less on average (Lang et al., 1996). However, the signi"cance of the association between volatility and investment holds.13 12 This evidence is not a necessary condition for a relation between cash #ow volatility and average investment. Consider investment projects that require staged "nancing. A "rm with volatile cash #ows may not invest in such projects even if the "rm has su$cient cash #ows to fund the "rst stage if the "rm anticipates that it will be in a shortfall position when funding is required at later stages. This scenario implies a negative relation between cash #ow volatility and average investment in a cross-sectional analysis. However, this scenario will not imply a di!erence between investment levels for "rms in shortfall and excess positions. 13 Annual correlation coe$cients between the industry-adjusted average leverage ratios and industry-adjusted CVs are positive and less than 0.15. Annual correlation coe$cients between industry-adjusted average leverage ratios and industry-adjusted sales growth are either insigni"cant or less than 0.10 in absolute value. Annual correlation coe$cients between industry-adjusted average leverage ratios and industry-adjusted average book-to-market ratios are negative and less than 0.15 in absolute value. B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 441 Table 4 Discretionary investment for "rms in a cash shortfall versus cash excess position Capital expenditures, R&D costs, and advertising expense for "rms in a cash shortfall versus cash excess position. A "rm is in a shortfall (excess) position and classi"ed as Short (Excess) if it is in the lowest (highest) quartile based on its average annual industry-adjusted ratio of operating cash #ow to beginning of period total assets (OPCF/TA), its average annual unadjusted OPCF/TA, the di!erence between current year unadjusted OPCF/TA and the six-year average unadjusted OPCF/TA, and the di!erence between current year industry-adjusted OPCF/TA and the six-year average industry-adjusted OPCF/TA. Results for the discretionary investment variables represent averages of the annual variable across the seven sample years from 1989 to 1995. The next to last column reports t-statistics for the di!erences in the means of discretionary investment for the short and excess quartiles. The last column reports the number of sample years in the annual analyses for which the di!erence in the means is statistically signi"cant at better than the 10% signi"cance level (d sig years) Firm's cash position measured Firm's cash position Short t-stat d sig years Excess Relative to industry peers (ranked on average annual industry-adjusted operating cash yow) Industry-adjusted capital expenditures !0.0004 0.0045 8.419 7 Industry-adjusted R&D costs 0.0018 0.0173 7.864 7 Industry-adjusted advertising expense 0.0053 0.0293 9.466 7 Relative to sample xrms (ranked on average annual operating Capital expenditures 0.0159 R&D costs 0.0379 Advertising expense 0.0376 cash yow) 0.0217 0.0529 0.0649 9.135 6.461 9.692 7 7 7 Relative to its own historical average (ranked on annual operating cash yow less six-year average annual operating cash yow) Capital expenditures 0.0162 0.0192 3.729 4 R&D costs 0.0469 0.0465 !0.119 1 Advertising expense 0.0435 0.0448 0.377 1 Relative to its own historical industry-adjusted average (ranked on annual industry-adjusted operating cash yow less six-year average annual industry-adjusted operating cash yow) Industry-adjusted capital expenditures 0.0007 0.0038 4.358 4 Industry-adjusted R&D costs 0.0104 0.0079 !1.108 1 Industry-adjusted advertising expense 0.0046 0.0040 !0.244 0 Second, we identify and eliminate "nancially distressed "rms from the sample and re-estimate the relation between volatility and investment (Eqs. (1) and (2)). Since there is no consensus on a measure of "nancial distress, we identify distressed "rms using seven di!erent metrics that are proposed by existing studies. A "rm-year observation is considered distressed if it has: (1) speculative grade debt (S&P bond ratings of BB and worse); (2) a negative earnings-price ratio; (3) negative average annual asset growth calculated over the six-year 442 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 period preceding the sample year; (4) average total assets in the lowest quartile of total assets; (5) a debt-equity ratio in the sample year in the highest quartile; (6) an average dividend payout ratio less than 10%, or (7) an average interest coverage ratio in the lowest quartile. These metrics are based on results in Barth, Beaver, and Landsman (BBL, 1997), FHP (1988), and KZ (1997). The results for each of the seven reduced samples (not presented) are qualitatively similar to those reported in Table 3. Thus, "nancially distressed "rms do not appear to drive the results. The second sensitivity analysis examines the causality of the relation between cash #ow volatility and discretionary investment. Our interpretation of the results in Table 3 is that cash #ow volatility, on average, leads to lower investment. However, an alternative explanation is that di!erent levels of investment (the dependent variable) produce di!erent volatilities due to the nature of the investments. To some degree, and with a lag, expenditure choices and patterns may determine the nature of the cash #ow stream, both in levels and volatilities. This concern is partially mitigated by the analysis in Section 5 that shows that investment-volatility sensitivities are related to the costs of accessing capital. Such a relation would not be expected if investment levels determine volatility. While we cannot provide conclusive evidence that higher volatility leads to lower investment rather than lower investment leading to higher volatility, we present additional results that are more consistent with our interpretation of the results. First, cash #ow volatility is not highly correlated with our proxies for growth. We would expect a signi"cant and positive correlation if investment determines cash #ow volatility. Over the seven-year sample period, the correlation coe$cients between industry-adjusted operating cash #ow volatility and average annual sales growth and book-to-market ratios (industry-adjusted), respectively, are 0.05 and 0.02. Second, industry-adjusted cash #ow volatility is signi"cantly and positively related to the industry-adjusted volatilities of the three proxy variables for discretionary investment across all levels of cash #ows. We would expect this positive relation if cash #ow volatility leads to lower investment. However, if di!erent levels of investment (the dependent variable) produce di!erent volatilities, we expect no association between the volatility of investment and cash #ow volatility.14 Third, we "nd that earnings volatility is not related to average investment levels and that the inclusion of earnings volatility in Eqs. (2) and (3) does not change the negative relation between cash #ow volatility and investment. This result is consistent with our interpretation that greater cash #ow volatility 14 There are production functions under which more volatile expenditures might produce more volatile cash #ows. In this case, investment volatility would lag cash #ow volatility. Because we observe a contemporaneous relation, this concern is somewhat mitigated although not eliminated. We thank the referee for o!ering this alternative interpretation. B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 443 implies more frequent periods of cash #ow shortfalls, and that these shortfalls are related to lower investment. If the causality works in the other direction and investment decisions a!ect volatility, it is unclear why investment decisions would a!ect only cash #ow volatility and not earnings volatility. In sum, these three pieces of evidence are consistent with our interpretation that cash #ow volatility is associated with lower average investment because it measures the incidence of cash #ow shortfalls. The third sensitivity analysis shows that the results are insensitive to our method of industry adjusting. An alternative to industry adjusting the CV is to "rst industry-adjust the quarterly cash #ows and then compute the CV of this measure. The results using this alternative measure of volatility to control for quarterly seasonality in operating cash #ows are qualitatively similar to those presented in Table 3. The fourth sensitivity analysis indicates that the results are not in#uenced by cross-sectional variation in growth opportunities. This evidence supplements the controls for growth provided by sales growth and book-to-market ratios and industry-adjusting. Volatility remains a signi"cant negative determinant of investment (in Eq. (3)) for both the upper and lower deciles of "rms partitioned based on book-to-market ratios as an indicator of growth. The "nal sensitivity analysis indicates that the results are robust to alternative de"nitions for &volatility' in cash #ows. The coe$cient of variation attempts to control for a mechanical relation between volatility and levels by scaling the standard deviation of cash #ows by the absolute value of the mean. An alternative methodology is to scale cash #ows, e.g., by total assets, and to compute the standard deviation of the ratio. We present the results using the coe$cient of variation because the scaling choice (assets in this example) will induce the results if asset levels are correlated with investment decisions. However, we also perform the analyses in Table 3 using the standard deviation of cash-return on assets, cash-return on the book value of equity, and cash-return on the market value of equity as the measure of cash #ow volatility. The results are qualitatively similar. 5. Investment, volatility, and the costs of accessing external capital This section investigates whether "rms' investment decisions are directly related to the costs of accessing capital markets and whether these costs a!ect the sensitivity of investment to cash #ow volatility. This analysis also demonstrates whether cash #ow volatility remains a signi"cant determinant of investment after controlling for a "rm's cost of accessing capital. This cost, in part, captures a "rm's average project risk. Therefore, the analysis provides evidence about whether project risk is a correlated omitted variable that explains our basic "nding of a negative relation between investment and volatility. 444 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 5.1. Methodology The regression model is an augmented version of Eq. (2) that includes a proxy for the industry-adjusted cost of accessing capital (CAPCOST), an interaction variable equal to the product of CAPCOST and the industry-adjusted coe$cient of variation of operating cash #ow (CVCF), and controls for cash #ow levels: INVESTMENT"d #d CVCF#d CAPCOST 1 2 3 #d (CVCF]CAPCOST) 4 # + MD [d #d CVCF#d CAPCOST i 1i 2i 3i i/LO,HI #d (CVCF]CAPCOST)]N#e . (4) 4i 4 D (D ) is an indicator variable equal to one if the observation has cash #ows LO HI in the lowest (highest) three deciles based on its industry-adjusted average annual level of operating cash #ows scaled by beginning of period total assets. CAPCOST is included in the regression to directly measure the relation between the costs of external capital and investment, but it also serves as a proxy for project risk. The interaction variable (CVCF]CAPCOST) measures whether cross-sectional variation in the costs of accessing capital markets mitigates (or exacerbates) the impact of volatility on investment levels. Nine separate variables are used as proxies for a "rm's costs of accessing debt and equity markets. Table 5 outlines the calculation of each variable. Five of the nine variables are related to a "rm's risk-adjusted cost of capital: (1) S&P bond rating (SPBOND), (2) yield-to-maturity (YTM), (3) systematic risk (BETA), (4) total equity price risk (pRET), and (5) weighted average cost of capital (WACC). We predict that "rms with a higher risk-adjusted cost of capital on an industry-adjusted basis will have lower industry-adjusted levels of investment, ceteris paribus. SPBOND and YTM are proxies for a "rm's cost of accessing debt capital. Calomiris et al. (1995) and Ogden (1987) suggest that "rms with worse (higher) S&P bond ratings have higher debt "nancing costs.15 WACC is a combination of a "rm's YTM and the annual average of its daily equity return (RET) from 15 Using S&P bond ratings as a proxy for the cost of accessing debt capital in an OLS regression assumes that the yield spreads between ratings categories are equivalent. However, Simkins (1998) reports that the average spreads for 1991}1995 vary. For example, the average spreads between A and BBB "rms is 42 basis points, between BBB and BB "rms is 110 basis points, and between BB and B "rms is 183 basis points. Section 5.2 provides a sensitivity analysis of this variable as a proxy for the cost of debt. B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 445 Table 5 De"nitions of variables that proxy for the cost of accessing external capital markets Variable Variable name De"nition Cost of debt capital S&P bond rating SPBOND The average S&P rating from Compustat (data item 280) and S&P Bond Guides. Yield-to-maturity YTM Weighted-average YTM on long-term debt (excluding convertible debt), calculated using data from S&P Bond Guides. Cost of equity capital Systematic risk BETA The annual beta of common stock obtained from CRSP stock "les. Total equity price risk pRET The annual standard deviation of daily market returns obtained from CRSP stock "les. Other costs of accessing capital markets Weighted average cost WACC of capital The after-tax YTM times the book value of long-term debt scaled by SIZE plus the return on equity from CRSP times the market value of equity scaled by SIZE. Analyst following ANALYST The maximum number of analysts making a forecast of earnings during the sample calendar year from I/B/E/S. Dividend payout ratio DIV The ratio of total cash dividends paid during the "scal year (Compustat data item 21) to the sum of net cash #ows and total cash dividends paid during the year. Bid}ask spread BASPRD Annual average of the daily di!erence between the bid and ask prices scaled by the daily closing price. Total "rm capitalization SIZE The market value of equity plus the book value of debt plus preferred stock (Compustat data item 130). The market value of equity is share price times the number of common shares outstanding (Compustat data item 199]data item 25). The book value of debt equals long-term debt plus the current portion of long-term debt (Compustat data item 9#data item 34). CRSP. The after-tax YTM is weighted by the debt-to-equity ratio (where the denominator is the market value of total common equity plus the book value of debt) and RET is weighted by one minus the debt-to-equity ratio. BETA and pRET measure a "rm's cost of accessing equity capital. In a Sharpe-Lintner world, cross-sectional variation in "rms' costs of equity is the direct result of cross-sectional variation in "rms' betas. Thus, if the assumptions 446 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 underlying the Sharpe-Lintner CAPM hold, the higher a "rm's stock beta, the higher is its cost of raising equity capital. If these assumptions do not hold, then other systematic factors not captured by beta or unsystematic risk can a!ect the cost of equity. In this case, the higher the total risk of a "rm's stock (systematic risk plus unsystematic risk), the higher is its risk-adjusted cost of raising equity capital. We predict that four additional proxies are associated with a "rm's cost of accessing capital because of capital market imperfections: (1) analyst following (ANALYST), (2) bid}ask spreads (BASPRD), (3) "rm size, and (4) dividend payout ratios (DIV). These proxies are related to the costs of accessing capital because they measure information asymmetry or the demand for liquidity, both of which can lead to cross-sectional variation in the costs of accessing equity. Lang and Lundholm (1996) and Brennan and Hughes (1991) show that analyst following is negatively associated with information asymmetry. Amihud and Mendelson (1988) "nd that bid}ask spreads are positively related to information asymmetry. As summarized in Botosan (1997), information asymmetry can have a positive association with a "rm's cost of equity for two reasons. First, greater information reduces transaction costs which creates greater demand for a "rm's securities. The greater demand increases liquidity and &liquidity-enhancing policies can increase the value of the "rm by reducing its cost of capital'. (See Amihud and Mendelson, 1988.) Second, greater information reduces estimation risk about the value of a "rm's equity. Lower estimation risk will reduce the cost of equity if estimation risk is non-diversi"able. In summary, greater analyst following and lower bid}ask spreads represent a lower cost of accessing external capital. Firm size (the natural logarithm of SIZE) is also a proxy for information asymmetry. Atiase (1985), Brennan and Hughes (1991), and Collins et al. (1987) suggest that large "rms have less information asymmetry than small "rms. Consistent with this lower information asymmetry, Ritter (1987) "nds that large "rms have lower costs of issuing securities. Thus, we predict that large "rms have lower costs of accessing capital markets. Dividend payout ratios (DIV) measure the liquidity of an investment in a "rm's stock. Asquith and Mullins (1983), Aharony and Swary (1980), Lang and Litzenberger (1989), and Hepworth (1953) indicate that capital markets value dividends because of liquidity constraints when equityholders are unable to borrow and lend freely, or because dividends provide a credible signal of management's private information. Because liquidity is associated with a lower cost of accessing capital markets, we predict that high dividend payout "rms have lower costs of accessing capital.16 The dividend payout ratio is de"ned as 16 Firms can also provide liquidity with stock repurchases or special dividends. These strategies bias against observing a relation between dividend payout ratios as a measure of the cost of accessing capital and investment. B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 447 the ratio of total dividends paid during the year to net cash #ows (before dividends) during the year because we focus on cash #ow volatility. The results are qualitatively unchanged using the traditional de"nition of a dividend payout ratio, dividends per share divided by earnings per share. 5.2. Results Results are reported in Table 6. Coe$cient estimates on the control variables and intercepts are not reported. Three main results emerge from this analysis. First, even after controlling for the relation between CAPCOST and investment, a negative and signi"cant association between volatility and investment remains across all regression equations (except when yield-to-maturity is used as a proxy for the cost of accessing external capital). Therefore, the documented association between investment and volatility does not simply re#ect an omitted correlated variable related to a "rm's cost of external "nancing such as project risk. Rather, the results in Table 6 are consistent with the notion that volatility measures the incidence of internal cash #ow shortfalls. Second, higher costs of accessing external capital are associated with lower investment, on average. As Table 6 reports, lower capital expenditures are signi"cantly associated with worse S&P bond ratings (higher numerical code),17 higher total equity price risk, higher bid}ask spreads, higher weighted average costs of capital (marginally signi"cant), and lower information asymmetry as measured by analyst following and "rm size. As before, cash #ow levels appear to have a "rst-order e!ect on investment. The bene"ts of "rm size in terms of higher investment are increasing in the level of a "rm's cash #ows. Similarly, the negative relation between investment and bid}ask spreads is eliminated for "rms with high cash #ows. Two results are contrary to our predictions. Lower investment is associated with lower systematic risk and higher dividend payout ratios. We expect the opposite relations assuming that beta is positively associated with the cost of accessing external capital and that the dividend payout ratio is negatively associated with these costs. One possible explanation for the negative association between dividends and investment is that "rms consider dividend payments to be non-discretionary. Alternatively, Smith and Watts (1992) suggest that dividend paying "rms are more mature and have fewer investment 17 The analysis includes only "rms with S&P bond ratings. Gilchrist and Himmelberg (1998) and Calomiris et al. (1995) contend that rated "rms have a lower cost of accessing external capital relative to unrated "rms. To control for this potential selection bias, following Gilchrist and Himmelberg (1998) we estimate Eq. (4) using as the proxy for CAPCOST an indicator variable that is equal to one if the "rm has an S&P bond rating and equal to zero otherwise. The results are similar except that the coe$cient on the interaction term between CAPCOST and CVCF is insigni"cant. Predicted Sign ! ! ! ! ! # # ! # Proxy for CAPCOST SPBOND YTM WACC BETA p(RET) ANALYST DIV BASPRD SIZE !0.0003 (!3.294) 0.0001 (0.062) !0.0006 (!1.680) 0.0024 (5.278) !0.1030 (!5.823) 0.0001 (2.751) !0.0009 (!2.677) !0.0366 (!2.363) 0.0004 (3.186) CAPCOST !0.0001 (!0.639) !0.0001 (!0.867) !0.0003 (!0.456) !0.0022 (!1.646) !0.0285 (!1.254) 0.0001 (1.160) 0.0006 (1.742) !0.0280 (!0.211) 0.0003 (1.567) !0.0002 (!0.920) !0.0007 (!0.826) !0.0011 (!1.191) 0.0041 (3.639) 0.0388 (1.693) 0.0001 (1.333) !0.0015 (!2.985) 0.0341 (2.470) 0.0006 (2.696) CAPCOST] LO HI !0.0003 (!2.517) !0.0002 (!0.445) !0.0004 (!1.840) !0.0005 (!8.040) !0.0006 (!4.198) !0.0004 (!6.967) !0.0003 (!4.164) !0.0007 (!6.497) !0.0002 (!2.934) CVCF 0.0002 (1.104) 0.0001 (0.122) 0.0001 (0.796) 0.0005 (4.835) 0.0005 (2.522) 0.0002 (3.279) 0.0002 (2.319) 0.0006 (4.350) 0.0001 (1.103) !0.0010 (!4.303) !0.0012 (!1.808) !0.0018 (!2.226) !0.0006 (!4.117) !0.0013 (!4.651) !0.0008 (!7.444) !0.0008 (!6.108) !0.0003 (!1.599) !0.0009 (!6.316) CVCF] LO HI 0.0001 (1.822) 0.0004 (0.472) 0.0007 (0.418) 0.0002 (0.422) !0.0120 (!1.248) 0.00001 (1.255) 0.0000 (0.382) !0.0085 (!1.175) 0.0001 (2.746) CVCF] CAPCOST !0.0001 (!2.150) !0.0007 (!1.248) !0.0006 (!0.428) !0.0004 (!0.708) 0.0062 (0.507) !0.0000 (!0.305) 0.0000 (0.090) 0.0121 (1.069) !0.0001 (!3.588) 0.0000 (0.117) !0.0005 (0.469) !0.0002 (0.936) !0.0018 (!2.246) !0.0254 (!1.580) !0.0000 (!1.250) 0.0001 (!0.037) 0.0201 (0.284) !0.0003 (!4.565) CVCF]CAPCOST] LO HI 5.85}8.83 !1.21}7.89 4.57}8.42 5.88}10.53 6.04}14.39 6.21}15.41 !1.00}22.10 3.42}10.46 4.49}11.07% Range of Adj. R2 Means of annual regressions of average annual industry-adjusted capital expenditures on industry-adjusted operating cash #ow volatility (CVCF), proxies for the average annual costs of accessing external capital (CAPCOST), and an interaction variable equal to the product of cash #ow volatility and a proxy for the average annual cost of accessing external capital (CVCF]CAPCOST). The proxies for the cost of accessing external capital are S&P bond ratings (SPBOND), yield-to-maturity (YTM), weighted average cost of capital (WACC), equity beta (BETA), standard deviation of returns (pRET), analyst following (ANALYST), dividend payout ratios (DIV), bid}ask spread (BASPRD), and the natural logarithm of "rm size (SIZE). All of the proxies are industry-adjusted. Detailed de"nitions are in Table 5. The regressions also include controls for industry-adjusted growth as measured by industry-adjusted average annual sales growth and book-to-market ratios. LO and HI are indicator variables equal to one if a "rm is in the lowest or highest three decile rankings, respectively, based on the level of its industry-adjusted operating cash #ows. Operating cash #ow equals sales!cost of goods sold!selling, general and administrative expenses (excluding R&D and advertising)!the change in working capital. Operating cash #ow volatility is measured as the coe$cient of variation of a "rm's quarterly operating cash #ow over the six-year period preceding each of the seven sample years from 1989 through 1995. Averages of all other variables (including the dependent variable) are calculated over the same period for which cash #ow volatility is measured. For each equation, the mean of the seven annual least squares values of the coe$cient on the interaction variable (a6 is presented. Z-statistics to test the hypothesis that E(a6)"0 are shown in parentheses. In#uential observations in it the annual estimations are downweighted by the method of Welsch (1980). Coe$cient estimates on intercepts and control variables included in the regressions (industryadjusted book-to-market ratio and industry-adjusted sales growth) are not presented. Table 6 Means of annual regressions of capital expenditures on cash #ow volatility, proxies for the costs of accessing external capital, and interaction variables between volatilities and proxies for the costs of accessing external capital 448 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 449 opportunities. This explanation is also consistent with the positive association between systematic risk and investment if beta re#ects information about growth that is not captured by book-to-market ratios or the other controls for growth. Third, signi"cant coe$cients on some of the interaction variables (CVCF]CAPCOST) indicate that the negative association between operating cash #ow volatility and capital expenditures is mitigated for "rms with lower costs of accessing external capital markets. Large capitalization "rms which we claim have a lower cost of accessing external capital have a less negative and marginally signi"cant sensitivity of investment to cash #ow volatility, on average. However, large "rms with either extremely low or extremely high cash #ows do not bene"t from their size. Similarly, lower costs of accessing equity capital, as measured by either beta or total equity price risk, mitigate the negative impact of cash #ow volatility on investment for "rms with high cash #ows. Conditional on the negative relation between investment and the costs of accessing debt capital as measured by S&P bond ratings, lower costs (lower numerical ratings) do not mitigate the impact of volatility on investment. Assuming that "rms with higher cash #ow volatility are more likely to have insu$cient internal capital in some periods and require external capital to fund investment, the negative association between volatility and investment is consistent with the Myers and Majluf (1984) pecking order hypothesis. As further evidence on this issue, we also estimate the relation between investment and volatility separately for "rms with low or high average annual cash balances. This speci"cation is based on Myers' (1984) evidence that internal cash can act as a substitute for external "nancing. A "rm is a low-cash "rm (high-cash "rm) if it is in the lowest three (highest three) deciles of "rms ranked on the basis of its industry-adjusted annual cash balances (Compustat data item 1) averaged over the same six-year period over which volatility is measured. The results show that the association between volatility and investment is more negative for the low cash "rms. These "rms are more likely to require external capital to fund investment because they lack su$cient internal cash bu!ers. 6. Volatility and the cost of accessing capital markets This section presents evidence that the negative association between investment and volatility is consistent with the basic NPV investment rule by showing that volatility is directly related to the costs of accessing external capital. Unlike the tests of the association between volatility and investment, however, the dependent variables in each of the separate regressions of CAPCOST on volatility are measured at the end of the period over which volatility is measured. For example, volatility measured over the six-year period 1988}1994 is matched with the "rm's S&P bond rating for 1995. In contrast, in the tests of 450 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 discretionary investment, average investment and volatility are measured over contemporaneous six-year periods. The dependent variable is measured di!erently because our predictions about why volatility a!ects investment di!er from our predictions about why it a!ects the proxies for the cost of accessing capital. The contemporaneous measurement of volatility and investment in the discretionary investment tests re#ects the prediction that higher cash #ow volatility over a period, and consequently more likely realizations of cash shortfalls, is associated with lower investment during that same period. In the tests of the association between volatility and the costs of "nancing, the prediction is that historical volatility is relevant because debt and equityholders use historical volatility to predict future volatility. In this case, a 1995 bond rating, for example, re#ects the creditor's assessment of future volatility as of 1995, and historical data is one factor that creditors can use to make this assessment. This di!erence in the analysis calls into question whether cash #ow volatility is the measure that investors use to assess the risk of future cash #ows. Debt and equityholders alternatively could use earnings volatility to assess future cash #ow volatility.18 Consequently, one could argue that earnings volatility is an omitted variable in the analyses in this section. As a robustness check of the results, we estimate all of the regressions outlined in this section including not only cash #ow volatility and cash #ow levels but also earnings volatility and earnings levels. When the earnings variables a!ect the results, we discuss the e!ects. 6.1. Volatility and the costs of accessing debt and equity We predict positive associations between cash #ow volatility and the two proxies for a "rm's cost of accessing debt capital, which are S&P bond ratings and yields-to-maturity. With interim payments, volatility increases a "rm's probability of default, other things equal. For a "rm to avoid technical default, cash #ows in every period must be su$cient to cover the "rm's debt service requirements. Higher cash #ow volatility increases the probability that the "rm's cash #ow realization in any given payment period will not cover its debt service requirements.19 18 Sloan (1996) "nds that earnings levels are a better predictor of future cash #ow levels than are historical cash #ow levels. The conditions under which earnings volatility is a better predictor of future cash #ow volatility is an open question. 19 Trueman and Titman (1988) make a similar prediction. They demonstrate that the incentives to smooth income and the costs of volatility are related to industry classi"cation because the probability and costs of bankruptcy vary across industries. B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 451 Because debtholders have a claim only on cash #ows after the results of all "rm decisions including investment decisions, the cash #ows that are relevant in debt valuation are cash #ows after investment (CFAI).20 Quarterly CFAI equals operating cash #ow, as de"ned in Section 3.1, less net capital expenditures, research and development costs (Compustat item 46 divided by four), and advertising expenses (item 45 divided by four). Net capital expenditures equal gross capital expenditures (item 90) less capitalized interest (item 147 divided by four) less the after-tax proceeds from sales of PPE (item 83 times one minus the tax rate). In all calculations, the tax rate (TR) is equal to 46% before 1987, 38% in 1987, and 34% after 1987. The correlation between industry-adjusted average operating cash #ow and cash #ow after investment is 97.5%. Although we are not aware of any direct empirical evidence on the association between cash #ow volatility and the cost of debt, indirect evidence is consistent with a positive association between earnings volatility and the cost of debt. Collins et al. (1981) and Lys (1984), for example, "nd negative returns at announcements of accounting rule changes that are predicted to increase earnings volatility and indicate that the magnitude of the reaction is positively related to a "rm's debt constraints. In cross-sectional studies, Bartov (1993) and Imho! and Thomas (1988) show that "rms adjust their real activities to avoid volatility, and that the extent of these adjustments varies with "rms' debt constraints. These studies suggest that managers have incentives to smooth earnings because smoother earnings reduce debt-related costs. We predict positive associations between volatility and systematic risk (BETA) and total equity price risk (pRET), which represent two of our proxies for the costs of accessing equity capital. As discussed by Beaver et al. (1970), estimations of these associations test the joint hypothesis that cash #ow volatility is correlated with a price-relevant risk and that the market impounds this information in security prices. We do not make a directional prediction about the association between volatility and analyst following as a proxy for a "rm's cost of accessing equity. The ultimate product of an analyst is a report that makes a stock buy or sell recommendation, but one element of the report is the "rm's earnings forecast. If analysts value forecast accuracy and it is more di$cult to predict earnings for high-volatility "rms, then volatility can negatively a!ect the analyst following decision. Beidleman (1973), Brennan and Hughes (1991), and articles in the popular press suggest that analysts are less likely to follow stocks of "rms with more volatile earnings because it makes their job of estimating &normal' earnings more di$cult. In addition, Schipper (1991) notes that &readers of analyst reports may use forecast accuracy as a quantitative measure of the quality of the overall 20 When the measure of cash #ows is cash #ows after investment, the measure of earnings is operating income, which is after depreciation. 452 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 report; this e!ect will create a preference for accuracy2'. These arguments suggest a negative relation between analyst following and volatility assuming a positive correlation between cash #ow volatility and earnings volatility. However, Barth et al. (1998) argue that analysts add the greatest value, and thus potentially reap the highest compensation, when information asymmetry is greatest. In this case, assuming a positive association between cash #ow volatility and information asymmetry, analysts would prefer to follow high-volatility "rms. We predict that dividend payout ratios are negatively associated with cash #ow volatility. Aharony and Swary (1980) show that negative stock price reactions to dividend decreases are larger in magnitude than positive reactions to dividend increases. This evidence suggests that equityholders value stable dividends. If dividend stability is a priority, "rms with higher cash #ow volatility are forced to maintain lower dividends to avoid the costs associated with cutting a dividend. We predict a positive association between volatility and bid}ask spreads. This prediction is based on an assumption that historical cash #ow volatility is associated with greater uncertainty about future cash #ows. Amihud and Mendelson (1988) show that greater uncertainty is associated with higher bid}ask spreads. Because equityholders have a claim only on residual cash #ows after debtholders are paid, we examine whether systematic risk, total equity price risk, analyst following, bid}ask spreads, and dividend payout ratios are associated with net cash #ows, after both investment and interest charges.21 Quarterly net cash #ows (NETCF) are measured as cash #ow after investment (CFAI) less after-tax interest expense (Compustat item 22 times one minus the tax rate) plus after-tax capitalized interest (item 147 divided by four times one minus the tax rate). The correlation between industry-adjusted average operating cash #ow and net cash #ow is 88.0%. In summary, we predict a positive association between volatility and S&P bond ratings, yields-to-maturity, stock market betas, total equity price risk, and bid}ask spreads. The "rst four predictions also imply a positive association between volatility and a "rm's weighted average cost of capital (WACC). We predict a negative association between volatility and dividend payout ratios. No prediction is made about the association between volatility and analyst following. 6.2. Results We estimate variations of Eqs. (1) and (2) to examine whether volatility is associated with the proxies for the cost of accessing external capital. Each 21 When the measure of cash #ows is net cash #ows, the measure of earnings is net income. B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 453 regression equation includes control variables that prior research has identi"ed as determinants of the dependent variable. The control variables are di!erent in each equation. Because the control variables are not the focus of our analysis, they are not described in detail. Table 7 summarizes the control variables, the predicted signs, and the source that justi"es the use of the variable as a control.22 As in the investment regressions, we compute the means of the seven annual ordinary least squares regression coe$cients for 1989}1995 for each dependent variable. Coe$cient estimates for the control variables are not presented. The results for these variables are consistent with the predictions from the literature cited in Table 7. Table 8 reports the results. In the regressions that exclude the controls for a "rm's cash #ow level, the mean coe$cient estimates on the volatility of cash #ow are statistically signi"cant and of the predicted sign in all regressions except when BETA is the dependent variable. However, once we control for the level of a "rm's cash #ows, the impact of volatility changes. The discussion focuses on the regressions that control for the level of a "rm's cash #ows. As Table 8 reports, the volatility of cash #ow after investment is associated with worse S&P bond ratings (higher numerical codes) and higher yields-tomaturity.23 These results are consistent with the prediction that higher volatility increases the likelihood that a "rm will not be able to meet its debt payments, all else equal.24 Similarly, the volatility of cash #ow after investment is positively related to a "rm's WACC. Net cash #ow volatility is not signi"cantly associated with stock market betas or total equity price risk once we control for the level of a "rm's net cash #ows.25 However, net cash #ow volatility is signi"cantly related to the proxies for the costs of accessing equity capital that result because of market imperfections. Speci"cally, volatility has a signi"cant positive association with bid}ask spreads.26 This positive association is consistent with the joint claim that 22 The control variables in the WACC regression are the same as the control variables in the YTM regression. 23 As a robustness check of this result, we exclude "rms with S&P bond ratings below investment grade and re-estimate the relation between bond ratings and volatility. The results are qualitatively similar to those presented in Table 8. 24 Because of the high correlation between bond ratings and volatility, we orthogonalize the S&P bond rating variable with respect to the volatility of cash #ow after investment before including it as a control variable in the yield-to-maturity and WACC regressions. 25 Total equity price risk is statistically and positively related to earnings volatility for highearnings "rms when earnings volatility is included along with cash #ow volatility in the regression. 26 We orthogonalize (pRET) with respect to net cash #ow volatility before using it as a control variable in the bid}ask spread regression. When bid}ask spread is regressed only on volatility, we observe a positive relation. However, when pRET (unorthogonalized) is added to the regression along with trading volume and "rm size, there is a negative and signi"cant relation between BASPRD and volatility. KU, Ogden (#) KU (#) Ogden (!) Ogden (!) EYR (#) EYR (#) BKS (#) BKS (!) Hamada, BKS (#) Banz (!) CN (!) CN (!) CN (#) BKS (!) AB (!) Standard deviation of returns SW (!) SW (#) SW (#) Dividend policy Beta S&P bond rating Yield-tomaturity Proxies for costs of accessing equity markets Proxies for costs of accessing debt markets OB, BH (#) OB (#) (#)! OB, BH, Bhushan (#) Analyst following HL, MP (!) HL, MP (!) HL (!) HL, MP, AM (#) Bid}ask spread !Many papers propose that analyst following is related to information asymmetry. We use bid}ask spread as a measure of information asymmetry. "For betas and dividend payout ratios, we use market-to-book ratios and separately sales growth as proxies for growth. For analyst following, consistent with the methodology of O'Brien and Bhushan (1990), growth is the net entry of "rms into the sample "rm's industry over the "ve-year period prior to the sample year. Abnormal returns Growth" Share Price Beta Bid}ask spread Trading volume Dividend payout ratio Total equity risk Firm size S&P bond rating Leverage Control variables Summary of control variables that are used in the regressions that estimate the relation between volatility and the costs of accessing external capital markets. The proxies are S&P bond ratings, yield-to-maturity, equity beta, standard deviation of returns, dividend payout ratios, analyst following, and bid}ask spreads. The table lists the citation(s) for the predicted coe$cient estimate as well as the predicted sign of the coe$cient estimate (in parentheses) for the control variable. The sources are: AB: Alford and Boatsman (1995); AM: Amihud and Mendelson (1989); Banz: Banz (1981); BKS: Beaver, Kettler, and Scholes (1970); Bhushan: Bhushan (1989); BH: Brennan and Hughes (1991); CN: Cheung and Ng (1992); EYR: Ederington, Yawitz, and Roberts (1987); Hamada: Hamada (1972); HL: Howe and Lin (1992); KU: Kaplan and Urwitz (1979); MP: Menyah and Paudyal (1996); OB: O'Brien and Bhushan (1990); Ogden: Ogden (1987), and SW: Smith and Watts (1992) Table 7 Summary of control variables 454 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 455 volatility is related to information asymmetry and that information asymmetry is related to higher spreads. In addition, volatility has a marginally signi"cant negative association with analyst following as a proxy for information asymmetry. These negative associations are consistent with the joint claim that analysts are more likely to make erroneous stock buy/sell recommendations when volatility is high and that analysts attempt to reduce this likelihood by not following "rms with volatile cash #ows. Including earnings volatility in the regression with controls for cash #ow and earnings levels, we "nd that only cash #ow volatility a!ects analyst following even though earnings are the more frequently forecasted number. Net cash #ow volatility also is negatively related to dividend payout ratios. However, including controls for earnings volatility and the levels of earnings, both cash #ow volatility and earnings volatility are negatively associated with dividend payout ratios. The observation that dividend payout ratios are negatively associated with earnings volatility is consistent with Smith and Warner's (1979) observation that dividend restrictions in bond covenants are frequently based on accounting earnings realizations. Finally, as in Table 3, the level of a "rm's net cash #ows has a "rst-order e!ect on the costs of accessing external capital as measured by some proxies. In particular, the intercepts indicate that low cash #ow "rms have worse S&P bond ratings, higher equity betas, and higher equity price risk than "rms with median cash #ows. High cash #ow "rms have better S&P bond ratings and lower dividend payout ratios. 7. Summary and conclusions This paper provides direct evidence that cash #ow volatility is associated with lower average levels of investment in capital expenditures, research and development costs, and advertising expenses. Cash #ow volatility remains a signi"cant negative determinant of investment even after controlling for the costs of accessing external capital. Moreover, cash #ow volatility increases these costs. In particular, cash #ow volatility is related to worse S&P bond ratings, higher yields-to-maturity, higher weighted average costs of capital, higher bid}ask spreads, lower analyst following, and lower dividend payout ratios. The results related to the role of capital costs in the investment decision and the importance of cash #ow volatility in the presence of these costs imply that the sensitivity of investment to volatility does not result because volatility is a proxy for project risk. Rather, cash #ow volatility is related to investment because it increases the likelihood that a "rm will need to access capital markets and it also increases the costs of doing so. Taken together, the results suggest that "rms do not completely smooth cash #ow volatility through time to maintain investment levels, but rather forgo some Intercept Intercept] LO Intercept] HI CV 0.0206 (2.313) !0.0424 (!0.766) 0.4801 (3.676) 0.1555 (1.007) 0.0192 (4.390) 0.4675 (4.605) 0.0293 (2.805) 0.2494 (1.915) 0.0104 (0.017) 0.2429 (2.045) 0.0303 (6.272) 0.1507 (0.997) WACC !0.0094 (!0.271) !0.0046 (!0.187) 0.0390 (0.404) 0.0178 (0.342) 0.0093 (1.840) 0.0055 (2.517) Costs of accessing debt and equity markets: Cash yow is dexned as cash yow after investment (CFAI) Yield-to-maturity S&P bond rating Costs of accessing debt markets: Cash yow is dexned as cash yow after investment (CFAI) Dependent variable 0.0062 (0.083) !0.0125 (!0.224) 0.0098 (0.536) CV] LO !0.0078 (!1.384) !0.0031 (!0.140) 0.0173 (0.354) CV] HI 62.11}85.78 60.59}85.63% 28.89}51.28 29.62}51.53% 57.02}61.87 57.31}62.01% Range of Adj. R2 Means of annual regressions of proxies for the costs of accessing capital markets on cash #ow volatility. Proxies for these costs are de"ned in Table 5. LO and HI are indicator variables equal to one if a "rm is in the lowest or highest three decile rankings, respectively, based on the level of its industry-adjusted cash #ow. Cash #ow after investment (CFAI) is operating cash #ow!net capital expenditures including R&D and advertising. Net cash #ow (NETCF) is operating cash #ow!net capital expenditures (including R&D and advertising)!aftertax interest charges. Cash #ow volatilities represent the coe$cient of variation (CV) in cash #ow estimated over the 24 quarters in the six years prior to the year of the calculation of the dependent variable. Cash #ows and other control variables are measured in the same year as the dependent variable. All variables are industry-adjusted. For each equation, means of the seven annual least squares values for each coe$cient (a6 ) are presented. Z-statistics to test the hypothesis that E(a6)"0 are shown in parentheses. In#uential it observations in the annual estimations are downweighted by the method of Welsch (1980). Coe$cient estimates on control variables included in the regressions (summarized in Table 7) are not presented. Table 8 Means of annual regressions of proxies for the cost of accessing capital markets on cash #ow volatility 456 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 Bid}ask spread Dividend policy Analyst following Standard deviation of stock returns Stock market beta 0.0000 (0.663) !0.0001 (!1.253) 77.58}86.73 !0.0008 (!1.345) 0.0002 (3.584) !0.0008 (!1.217) !0.0048 (!6.433) 21.64}28.32 15.22}23.46% 66.46}70.64 66.98}70.68% 78.09}88.39 77.63}88.54% 3.95}20.08 76.53}86.90% 0.0006 (0.999) 0.0047 (0.361) 0.0001 (1.558) 0.0042 (0.787) 0.0002 (3.001) !0.0014 (!0.639) !0.0242 (!1.084) !0.00002 (!0.587) !0.0009 (!0.644) 4.78}16.01% !0.0052 (!5.632) !0.0014 (!2.439) !0.1449 (!14.411) 0.2282 (17.154) !0.1455 (!8.129) !0.0012 (!2.963) 0.1525 (15.724) !0.0234 (!1.621) !0.3925 (!1.518) 1.4104 (4.606) !0.2750 (!1.035) !0.0231 (!4.491) 1.2347 (3.914) !0.0000 (!0.065) !0.0002 (!0.934) 0.0026 (2.685) 0.0012 (2.109) 0.00002 (2.050) 0.0029 (2.576) 0.0008 (0.284) !0.0912 (!5.661) 0.2743 (12.754) 0.1241 (3.358) 0.0019 (1.251) 0.2532 (14.024) Costs of accessing equity markets: Cash yow is dexned as net cash yow (NETCF) B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 457 458 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 investment. We do not claim that these results imply that "rms should reduce or eliminate volatility. Rather, the e!ects of volatility represent one factor that a "rm should consider in its risk management decisions. Firms must decide how to trade-o! the expected negative impact of volatility on investment levels against the other e!ects of managing volatility. References Aharony, J., Swary, I., 1980. Quarterly dividend and earnings announcements and stockholders' returns: an empirical analysis. Journal of Finance 35, 1}11. Albrecht, D., Richardson, F., 1990. Income smoothing by economy sector. Journal of Business, Finance and Accounting 17, 713}730. Alford, A., Boatsman, J., 1995. Predicting long-term stock return volatility: implications for accounting and valuation of equity derivatives. Accounting Review 70, 599}618. Amihud, Y., Mendelson, H., 1988. Liquidity and asset prices: "nancial management implications. Financial Management 7, 5}15. Amihud, Y., Mendelson, H., 1989. The e!ects of bid-ask spread, residual risk, and size. Journal of Finance 44, 479}486. Asquith, P., Mullins, D., 1983. The impact of initiating dividend payments on shareholders' wealth. Journal of Business 56, 77}95. Atiase, R., 1985. Predisclosure information, "rm capitalization, and security price behavior around earnings announcements. Journal of Accounting Research 23, 21}36. Banz, R., 1981. The relationship between return and market value of common stocks. Journal of Financial Economics 9, 3}18. Barth, M., Beaver, W., Landsman, W., 1997. Relative valuation roles of equity book value and net income as a function of "nancial health. Unpublished Working Paper. Stanford University. Barth, M., Kasznik, R., McNichols, M., 1998. Analyst coverage and intangible assets. Unpublished Working Paper, Stanford University. Bartov, E., 1993. The timing of asset sales and earnings manipulation. Accounting Review 68, 840}855. Beaver, W., Kettler, P., Scholes, M., 1970. The association between market determined and accounting determined risk measures. Accounting Review 35, 654}682. Beidleman, C., 1973. Income smoothing: the role of management. Accounting Review 38, 653}667. Bhushan, R., 1989. Collection of information about publicly traded "rms: theory and evidence. Journal of Accounting and Economics 11, 183}206. Botosan, C., 1997. Disclosure level and the cost of equity capital. Accounting Review 72, 323}349. Brennan, M., Hughes, P., 1991. Stock prices and the supply of information. Journal of Finance 46, 1665}1691. Calomiris, C., Himmelberg, C., Wachtel, P., 1995. Commercial paper, corporate "nance, and the business cycle: a microeconomic perspective. Carnegie-Rochester Series on Public Policy 42, 203}250. Cheung, Y., Ng, L., 1992. Stock price dynamics and "rm size: an empirical investigation. Journal of Finance 47, 1985}1997. Cleary, S., 1998. The relationship between "rm investment and "nancial status. Journal of Finance 54, 673}692. Collins, D., Kothari, S.P., Rayburn, J., 1987. Firm size and the information content of prices with respect to earnings. Journal of Accounting and Economics 9, 111}138. Collins, D., Roze!, M., Dhaliwal, D., 1981. The economic determinants of the market reaction to proposed mandatory accounting changes in the oil and gas industry. Journal of Accounting and Economics 3, 37}71. B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 459 Dolde, W., 1995. Hedging, leverage, and primitive risk. Journal of Financial Engineering 4, 187}216. Ederington, L., Yawitz, J., Roberts, B., 1987. The information content of bond ratings. Journal of Financial Research 10, 211}226. Fazzari, S., Hubbard, R., Petersen, B., 1988. Financing constraints and corporate investment. Brookings Papers on Economic Activity 1, 141}195. Fazzari, S., Hubbard, R., Petersen, B., 1918. Investment * cash #ow sensitivities are useful: a comment on Kaplan and Zingales. Unpublished Working Paper. Columbia University. Froot, K., Scharfstein, D., Stein, J., 1993. Risk management: coordinating investment and "nancing policies. Journal of Finance 48, 1629}1658. GeH czy, C., Minton, B., Schrand, C., 1997. Why "rms use currency derivatives. Journal of Finance 52, 1323}1354. Gilchrist, S., Himmelberg, C., 1998. Investment, fundamentals, and "nance. Unpublished Working Paper. Columbia University. Greene, W., 1993. Econometric Analysis 2nd edition. MacMillan Publishing Co., New York. Hamada, R., 1972. The e!ect of the "rm's capital structure on the systematic risk of common stocks. Journal of Finance 27, 435}452. Hepworth, S., 1953. Smoothing periodic income. Accounting Review 28, 32}39. Hoshi, T., Kashyap, A., Scharfstein, D., 1991. Corporate structure, liquidity, and investment: evidence from Japanese industrial groupings. Quarterly Journal of Economics 56, 33}60. Howe, J., Lin, J., 1992. Dividend policy and the bid}ask spread: an empirical analysis. Journal of Financial Research 15, 1}10. Imho!, E., Thomas, J., 1988. Economic consequences of accounting standards: the lease disclosure rule change. Journal of Accounting and Economics 10, 277}310. Kaplan, R., Urwitz, G., 1979. Statistical model of bond ratings: a methodological inquiry. Journal of Business 52, 231}261. Kaplan, S., Zingales, L., 1997. Do investment-cash #ow sensitivities provide useful measures of "nancing constraints?. Quarterly Journal of Economics 112, 169}215. Lamont, O., 1997. Cash #ow and investment: evidence from internal capital markets. Journal of Finance 52, 83}111. Lang, L., Litzenberger, R., 1989. Dividend announcements: cash #ow signalling vs. free cash #ow hypothesis. Journal of Financial Economics 24, 181}191. Lang, L., Ofek, E., Stulz, R., 1996. Leverage, investment, and "rm growth. Journal of Financial Economics 40, 3}30. Lang, L., Stulz, R., Walkling, R., 1991. A test of the free cash #ow hypothesis. Journal of Financial Economics 29, 315}335. Lang, M., Lundholm, R., 1996. Corporate disclosure policy and analyst behavior. Accounting Review 71, 467}492. Lessard, D., 1990. Global competition and corporate "nance in the 1990s. Journal of Applied Corporate Finance 3, 59}72. Lys, T., 1984. Mandated accounting changes and debt covenants: the case of oil and gas accounting. Journal of Accounting and Economics 6, 39}65. Menyah, K., Paudyal, K., 1996. The determinants and dynamics of bid}ask spreads on the London Stock Exchange. Journal of Financial Research 19, 377}394. Mian, S., 1996. Evidence on corporate hedging policy. Journal of Financial and Quantitative Analysis 31, 419}439. Michelson, S., Jordan-Wagner, J., Wootton, C., 1995. A market based analysis of income smoothing. Journal of Business Finance and Accounting 22, 1179}1193. Myers, S., 1977. Determinants of corporate borrowing. Journal of Financial Economics 5, 147}175. Myers, S., 1984. The capital structure puzzle. Journal of Finance 39, 575}592. Myers, S., Majluf, N., 1984. Corporate "nancing and investment decisions when "rms have information that investors do not have. Journal of Financial Economics 13, 187}221. 460 B.A. Minton, C. Schrand / Journal of Financial Economics 54 (1999) 423}460 Nance, D., Smith, C., Smithson, C., 1993. On the determinants of corporate hedging. Journal of Finance 48, 267}284. O'Brien, P., Bhushan, R., 1990. Analyst following and institutional ownership. Journal of Accounting Research 28, 55}76. Ogden, J., 1987. Determinants of the ratings and yields on corporate bonds: tests of the contingent claims model. Journal of Financial Research 10, 329}339. Ritter, J., 1987. The cost of going public. Journal of Financial Economics 19, 269}281. Schipper, K., 1991. Commentary on analysts' forecasts. Accounting Horizons 5, 105}121. Shapiro, A., Titman, S., 1986. An integrated approach to corporate risk management. In: Stern, J., Chew, D. (Eds.), The Revolution in Corporate Finance. Basil Blackwell, New York, pp. 331}354. Shimko, D., 1997. Yearnings per share. Risk 10, 37. Simkins, B., 1998. Asymmetric information, credit quality and the use of interest rate derivatives. Unpublished Working Paper. Oklahoma State University. Sloan, R., 1996. Using earnings and free cash #ow to evaluate corporate performance. Journal of Applied Corporate Finance 9, 70}78. Smith, C., Warner, J., 1979. On "nancial contracting: an analysis of bond covenants. Journal of Financial Economics 7, 117}161. Smith, C., Watts, R., 1992. The investment opportunity set and corporate "nancing, dividend, and compensation policies. Journal of Financial Economics 32, 263}292. Stulz, R., 1990. Managerial discretion and optimal "nancing policies. Journal of Financial Economics 26, 3}28. Tufano, P., 1996. Who manages risk? An empirical examination of risk management practices in the gold mining industry. Journal of Finance 51, 1097}1137. Trueman, B., Titman, S., 1988. An explanation for accounting income smoothing. Journal of Accounting Research 26, 127}139. Welsch, R., 1980. Regression sensitivity analysis and bounded-in#uence estimation. In: Kmenta, J., Ramsay, J. (Eds.), Evaluation of Econometric Models. Academic Press, New York, pp. 153}167.