UFDC Home  Search all Groups  UF Institutional Repository  UF Institutional Repository  UF Theses & Dissertations  Vendor Digitized Files   Help 
Material Information
Subjects
Notes
Record Information

Full Text 
ECONOMIC DETERMINANTS OF EARNINGS PERSISTENCE By SUNDARARAMAN THIAGARAJAN A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 1989 Copyright 1989 by Sundararaman Thiagaraj an ACKNOWLEDGEMENTS It is hard to list the names of the numerous friends of mine who have helped me in this marathon effort. I would like to thank Professor Robert Freeman, my chairman, and Professor Prakash Loungani for getting me interested in the area of persistence and for their constant guidance especially in the formative stages of the project. My cochairman, Professor Bipin Ajinkya, provided detailed comments on my work and also provided constant moral encouragement for which I am ever grateful. Professor Senyo Tse asked penetrating questions which made me return to the drawing board several times. I ground my teeth at that time but am extremely thankful now for his insight and incisive comments. Professor Linda Bamber went through my dissertation with a finetoothed comb. She made me infuse common intuition into all the concepts and made me rewrite large portions of the paper. It was painful to do so at that time, but the dissertation improved so much that I am eternally grateful that she agreed to be on my committee. Professor Miles Livingston made me think and work efficiently. He encouraged me to work hard and finish the dissertation with his sound advice on several aspects. Professor Steve Cosslett provided a lot of help in the formative stages of the project which helped write efficient computer programs. All of my committee members put up with my tardiness but were not tardy in their responses. The most difficult stage of the dissertation is the formative stage and the help obtained at this stage is the most valuable. Sarmila Banerjee, Charlie Chi, Steve Kachelmeier, John Neill, and K. Ramesh served as bouncing boards for several of my ideas, and once the idea was formalized, provided plenty of help in the tests. I am extremely grateful to them. Perhaps the most important was the moral support and research help provided by my wife, Bhama, without whose patience, constant encouragement, and faith in my abilities this dissertation would not have seen the light of day. Finally my family back home provided a lot of encouragement to finish the dissertation and the faith in the Lord provided me with the spiritual energy to carry me through all the difficult stages of the project. TABLE OF CONTENTS page ACKNOWLEDGEMENTS.............................................. iii ABSTRACT. ..................................................... vii CHAPTERS 1 INTRODUCTION........................................... 1 2 MEANING OF PERSISTENCE... ....................... ... .. 5 Early Studies...................................... 5 Miller And Rock Framework........................... 6 Kormendi And Lipe Framework........................ 8 A Comment Regarding KL Specification ............... 9 Cochrane's Specification............................. 12 Comparison With The KL Specification ................ 14 Estimation Of Cochrane's Measure Of Persistence................................... 16 Estimates Of Persistence Based On Analysts' Forecasts ........................... 17 3 MOTIVATION AND POSSIBLE ECONOMIC DETERMINANTS.................................. 19 Introduction........................................ 19 Motivation.......................................... 19 Economic Determiants Of PersistenceThe Framework....................... 22 Persistence Of Industry Earnings................... 24 Firm Specific Factors.............................. 26 Effects Of Financing Decisions..................... 29 Effects Of Accounting Techniques................... 30 Conclusion......................................... 34 4 MODELS, MEASUREMENT OF VARIABLES AND SAMPLE CHARACTERISTICS........................... 35 Basic Model....................................... 35 Measurement Of Variables... ....................... 37 Kormendi And Lipe Model............................ 40 Tests For The Effects Of Accounting Methods................................ 42 CHAPTERS page 5 SAMPLE SELECTION, DESCRIPTIVE STATISTICS AND RESULTS OF MODEL ESTIMATION................. 44 Sample Selection.................................. 44 Matters Concerning Empirical Specification.................................... 47 Descriptive Statistics............................. 51 Results Of Estimation Of Basic Model............... 56 Robustness Checks For The Basic Model.............. 57 Kormendi And Lipe Model........................... 62 Effects Of Accounting Techniques .................. 65 Effects Of Differential Accounting For R&D Variable................................. 69 Generalization Of The Results..................... 70 6 SUMMARY AND CONCLUSIONS................... ........... 77 APPENDICES 1 RELATIONSHIP BETWEEN KL AND COCHRANE SPECIFICATIONS OF PERSISTENCE.................. 79 2 ASSOCIATION OF ISP WITH ECONOMIC VARIABLES............ 87 REFERENCES................................................... 92 BIOGRAPHICAL SKETCH.......................................... 96 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy ECONOMIC DETERMINANTS OF EARNINGS PERSISTENCE By Sundararaman Thiagarajan August 1989 Chairman: Robert N. Freeman Major Department: Fisher School of Accounting This dissertation examines the association of earnings persistence with investment factors internal to the firm. The purpose of this examination is to better understand an important attribute of accounting earnings numbers. This attribute has been shown to have valuation implications, but it is not clear what factors internal to the firm contribute to this attribute. The results of the investigation suggest that earnings persistence is positively associated with the investment decisions of the management. Specifically, earnings persistence is found to be associated with the investments in capital expenditure, R&D and advertising. In addition, it is shown that the positive association between earnings persistence and the firmspecific investment factors is robust with respect to the measurement of earnings persistence and the investment factors chosen as determinants. Apart from enabling us to understand some of the components which comprise vii earnings persistence, the results confirm that persistence is a meaningful construct. viii CHAPTER 1 INTRODUCTION The relationship between earnings and security prices has been the focus of intensive research effort since the publication of the well known studies of Ball and Brown (1968) and Beaver (1968). The general conclusion of these studies and followup studies [e.g., Brown and Kennelly (1972) and May (1971)] is that earnings announcements are informative for pricing of stocks. This conclusion is "robust across statistical methodologies, time periods and stock exchanges" [Lev and Ohlson (1982, p. 261)]. Furthermore, any given firm's accounting numbers depend on factors such as the firm's production, investment, and financing decisions, the general economic environment, and the firm's accounting techniques [Foster (1986, p. 134)]. Since these firm specific factors can vary crosssectionally, stock prices may react to earnings announcements in a manner that reflects these differences. Crosssectional differences in the information content of earnings have been a matter of theoretical, econometric and empirical investigations for a long time [Choi (1984), Burgstahler (1981), Beaver, Clark and Wright (1979)]. Recently, Miller and Rock (1985) [MR] used a simple twoperiod model of firm valuation to suggest that the magnitude of return reaction to an earnings innovation, is related to the "persistence of earnings" (hereafter, "persistence"). They define persistence as the impact of an earnings innovation (or earnings surprise) on the expectations of future earnings.1 Thus, one possible reason for crosssectional differences in the information content of earnings, as measured by the stock reaction to an earnings innovation, is differential persistence. Kormendi and Lipe (1987) [KL] tested MR's proposition in a crosssectional setting and found that the magnitude of the return reaction is positively associated with persistence. This dissertation focuses on the association of earnings persistence with firm specific variables. The variables of interest are the firmspecific investment variables. Earnings persistence, which is a measure of the longterm behavior of the earnings series, is shown to have a valuation impact both theoretically (Miller and Rock) and empirically (Kormendi and Lipe). However, it is not clear what firmspecific factors contribute to the difference in such longterm behavior of earnings. The motivation for this examination arises from the desire to understand the economic reasons underlying the longterm behavior of accounting earnings, as measured by persistence. Since the longterm behavior of earnings, as captured by persistence, is seen as a valuation construct by the market, it is important to understand the firmspecific factors which influence this longterm behavior of earnings. This dissertation explores the relationship between earnings persistence and certain economic 1The terms "earnings surprise," "earnings innovation" and "shock in earnings" are all used synonymously in this dissertation. Such earnings innovation is observed at the time of the earnings announcements. characteristics of the firm. Specifically, it examines the association of persistence with the investment decisions of the firm. The research question is whether firmspecific investment decisions explain cross sectional differences in persistence. Documenting an association between persistence and the selected economic factors confirms that persistence is a meaningful construct. This, in turn, should enable us to examine whether earnings persistence is a reduced form characterization of certain economic factors internal to the firm. This examination should further our general understanding of the determinants of earnings behavior of firms with differences in estimated persistence. The association between persistence and the selected economic factors is examined using two different measures of persistence. The two different measures are a parametric measure of persistence suggested by KL and a nonparametric measure suggested by Cochrane (1988).2 An appendix to the dissertation (appendix 1) examines the relationship between Cochrane's measure and KL's measure. The structure of the dissertation is as follows. The next chapter briefly explains the meaning of persistence as used in the different studies and compares the two measures of persistence employed in this dissertation. Chapter 3 motivates the research question and discusses possible economic determinants of persistence. Chapter 4 specifies the 2 The term "nonparametric" is not used in the sense of statistical analysis based on ranks. As explained later, this measure of earnings persistence is nonparametric in the sense that the specification of the measure does not depend on the underlying process that describes earnings; it is flexible enough to allow for a variety of earnings processes. In other words, no assumptions need to be made about the process that describes earnings. 4 models to be estimated and the basis for the measurement of the variables. Chapter 5 describes the sample characteristics and discusses the results of the models estimated. The sixth and final chapter concludes summarizes the results and provides some suggestions for future research. CHAPTER 2 MEANING OF PERSISTENCE This chapter briefly explains the different measures of persistence and compares the two measures of persistence used in this study. Though the notion of persistence has attracted a lot of research attention in recent years, the concept has been used in industrial organization studies for more than two decades. 2.1 Early Studies Persistence refers to the quality of being persistent which, in turn, is understood as "continuing for a long time into the future."1 This general meaning of persistence was used in economic studies by Stigler (1963), Brozen (1970) and others in examining departures from perfectly competitive conditions [see Scherer (1980) and Clarkson and Miller (1982) for a review of the relevant literature]. These studies hypothesized that noncompetitive conditions (i.e., monopolies) are characterized by persistence of high rates of return (i.e., continuance of high rates of return).2 'Webster's dictionary defines persistence as "existing for a long time or longer than usual time or continuously" [Webster's New Collegiate Dictionary, 1977, p.855]. 2Even though the preponderance of evidence suggests that non competitive conditions are characterized by persistence of high rates, there is controversy regarding the measurement of the relevant variables [see Scherer(1980) for details]. Since the notion of persistence is the focus here, the controversy is not discussed. 5 Persistence was defined in these studies as the correlation of rates of return at two distinct points in time such as t and t+l see Stigler (1963, table 18, p. 71)]. High correlation implied high persistence. Stigler showed that high rates of return were persistent for noncompetitive industries. 2.2 Miller And Rock Framework The recent MR study has sparked considerable interest in the notion of persistence. In the MR model, the firm exists for only two periods (Periods 1 and 2) and is liquidated at the end of the second period. At the beginning of period 1 and at the beginning of period 2, investments I1 and 12 are made respectively.3 The earnings from investments I1, made at the beginning of period 1, are denoted as xl and the earnings from investments 12, made at the beginning of period 2, are denoted as x2 (a tilde over xt indicates a random variable; xt without a tilde indicates the realized value of xt). Expectations are formed at the beginning of period 1 for the earnings of periods 1 and 2. These expectations are denoted as E1(xi) and El(i2) respectively. At the end of period 1, the actual earnings for period 1, xl, are revealed. As soon as the actual earnings are revealed, two events occur: (1) The surprise in earnings for period 1 is calculated. This surprise in earnings, denoted el, is the difference between the actual earnings for period 1 (xl) and the expectation of such earnings at the beginning of period 1 [El(x1)]. In other words, el = x, El(xi). (2) The expectation for the earnings of time period 2 is revised. 3 The notation has been slightly altered from the original MR model to help exposition. In the absence of any information, at the beginning of period 1, the expected values of the earnings surprises for the two periods (el, e2) are zeroi.e., El(ei) = El(e2) = 0. On the release of information x, and the calculation of el, at the end of period 1, the expected earnings for the next period are revised. MR denote this revision as Tre, where r is called the coefficient of persistencei.e., E2(e2el) = re1. This coefficient measures the extent to which the current earnings innovation (el) is incorporated in the expectations of earnings for next period. If r=0, then the expected earnings for the next period are not revised and the earnings surprise for the current time period is not expected to continue in the futurei.e., it is not persistent. In this case, the earnings surprise, el, is "transitory." If, on the other hand, r=l then the expected earnings for time period 2 are revised by the entire amount of the current period's earnings surprise. Thus, the entire current period shock is expected to continue in the futurei.e., the current shock is persistent. In this case, the earnings surprise is "permanent." Typically, r takes the value between 0 and 1.4 In summary, the MR model defines persistence as the extent to which the current period earnings surprise is expected to continue in 4 Conceivably, the r coefficient can take a value greater than 1 which would imply that a $1 innovation in the current period results in more than $1 revision in the expected value of earnings in the future. the future.5 The future in the MR model is restricted to one period (period 2). The analytical extension of the MR model to several periods is likely to require certain restrictive assumptions. 2.3 Kormendi And Lipe Framework As indicated earlier, KL tested the MR proposition that the cross sectional difference in the information content of earnings announcements (captured by the earnings response coefficient) can be explained by differential persistence. Kormendi and Lipe needed a measure of the revision in the expectation of future earnings induced by the current period earnings innovation. To derive this, KL assumed that the earnings expectations of each firm in their sample could be approximated by the following second order autoregressive process: (1) AX, = b + caAXt1 + C2AXt2 + et, where AXt is the difference earnings at time ti.e., AXt = XtXt. 6 b is the intercept, and e, is the earnings surprise. 5 The discussion above suggests that each earnings shock of a firm could have different persistence. In terms of the MR model, the r could be different for each shock. Hence, in a multi period framework,r would carry a time subscript. Operationally, however, it is not possible to calculate the persistence of individual earnings shocks when the time series properties of earnings are used to estimate persistence. Consequently, persistence estimators typically measure the average persistence of all the earnings surprises during the period of estimation. Such average persistence of all the earnings shocks is called the persistence of the firm's earnings. 6 The earnings series were difference by KL to obtain a stationary series. Conditional on the assumption that the expectations of future earnings can be approximated by equation 1, simple techniques in time series analysis enable the calculation of the effect of a current earnings innovation (et) on the revisions in the expectations of future earnings (see appendix 1 for details). The present value of these revisions (denoted as PVR), calculated using an assumed discount rate, is the KL measure of persistence. Firms with high PVRs are interpreted as having high persistence, and vice versa. Kormendi and Lipe's notion of persistence, therefore, empirically extends the MR notion to a multi period framework, and estimates the revisions in expectations for all the future periods.7 2.4 A Comment Regarding KL Specification The proper estimation of persistence under the KL specification requires the identification and estimation of the underlying process for each sample firm.8 A popular way of estimating persistence is to avoid the problem of identification of the underlying process for each sample firm by assuming that the earnings process of all the firms can 7 Present value of revisions (PVR) is not the same as the measure of persistence employed in the earlier studies. To illustrate this, consider two firms, both of which have an identical $10 innovation in earnings at time t. For firm 1, this innovation results in a $8 revision in expectation in time t+l and no revisions in any subsequent period. For firm 2, this innovation results in $.25 revisions for each of the future time periods up to the sixth period. According to the KL measure firm 1 would have a higher persistence, whereas according to the definition used in earlier studies, firm two, which experiences the effect of the first period shock for as long as the sixth period, has higher persistence. 8 The process of identification and estimation is likely to involve errors in judgement which are directly reflected in the estimates of persistence. Such estimation errors are independent of the errors in the specification. 10 be described by a certain process such as an autoregressive process of a specified order [See Lipe (1986) and Kormendi and Lipe (1987)]. This creates a problem since the underlying process may be quite different from the one assumed. If the underlying model is some other ARMA process, such as a moving average process instead of the autoregressive process assumed by KL, the magnitude of estimated persistence would be altered [see Collins and Kothari (1987, figure 1)]. To illustrate this, figure 1 shows the effect of calculating the KL persistence measure for different orders of the autoregressive process for six firms based on timeseries data from 19501985. The Xaxis indicates the assumed order of the autoregressive process and the Yaxis the magnitude of persistence of net income (before extraordinary items and discontinued operations). Both the magnitude of the estimated persistence and the rankordering of firms with respect to persistence change when the underlying assumption about the expectation model is altered.9 Both of these errors arise out of the need to specify an expectation model so that the innovation can be estimated. While it is not clear that specification errors resulting from these sources systematically biased KL's results, it is quite likely that attempts to reduce these errors could result in better estimates of persistence. 9In this example, the assumption is that all the underlying firms can be described by an autoregressive process. Based on this assumption, the effect of altering the order of the process was revealed in figure 1. In reality, each firm could be described by a different process. Firm 1 may follow an Arma (1,3) process while firm 2 may follow an arma (3,2) process. In such a case, the assumption of an autoregressive process, in addition to the assumption of the order of the process would be in error. 11 28 26 24 22 4) U 20 14 8 10 a 2 4 6 7 Order of the autoregressive process Kormendi And Lipe Measure Of Persistence For Different Lags Figure 1 2.5 Cochrane's Specification This specification is based on the premise that any series can be modeled as the sum of two components, a permanent component and a temporary component. An estimate of the permanent component is the suggested measure of persistence. An intuition behind this measure can be obtained by examining two extreme processes illustrated in the following two cases. Case 1 Consider the following model: (2) xt = xt_ + ut. In this case, the earnings at time t (xt) depend only on the past earnings (xtl) and a white noise error term (ut). In this case, all the innovations are permanent and are incorporated in the expectations of future earnings. To see this, we can write out the equation for xt1i.e., (3) Xt1 = Xt2 + Ut1l Substituting for xt in equation 2 we have xt = ,t2 + ut1 + ut. By recursive substitution it can be shown that (4) ix = Utk + utk1 + ... + Uti + Ut. Thus, all the past innovations are imbedded in the expectations of earnings at time t. All the innovations are permanent and the series itself has only a permanent component since no part of the earnings is transitory. This is the familiar random walk model. Case 2 If all the innovations are transitory, then the current earnings innovation does not cause a revision in the expected future earnings. This can be illustrated with the following model: (5) xt ut In this case, earnings are modeled as white noise. Consequently, the current earnings innovation, ut, does not affect the expectations of earnings in the future. The expectation of earnings at any point is measured as E(ut) which is zero. In case 2 all the innovations are transitory and the earnings do not have a permanent component.10 These two cases illustrate that if the current earnings innovation does not result in any revisions in the expectations of future earnings, then such innovation is transitory (case 2). If it results in some revision of expected future earnings, it has a degree of permanence (case 1). Persistence measures the extent to which the innovations are permanent. This was identified as the r coefficient in the MR model. These two cases further illustrate that if all the innovations are permanent, then the series itself would have only a permanent component and no transitory component (Case 1). On the other hand, if all the innovations are transitory, then the series would have only a transitory component and no permanent component (Case 2). Most earnings series 10 Case 2 is clearly a hypothetical process and is used for illustration only. It is difficult to conceive of any earnings process which is pure white noise. To illustrate the idea that current earnings innovation does not affect future expectations, we can construct a similar case where earnings are described by the following process xt = a + ut. In this case, the current innovation, ut, does not affect the expectation of xt+k which is a. The permanent component is an intertemporal constant, a. Innovations in earnings do not affect the permanent component (a) unlike the random walk model where all the innovations are incorporated in the permanent component. The extreme white noise process was chosen to highlight the effect of the temporary component. fall between these two extremes.11 Thus, earnings surprises typically have both a temporary component and a permanent component. By inference, therefore, the earnings series also has both a permanent component and a temporary component. The Cochrane specification directly estimates the permanent component of the series. The estimated magnitude of the permanent component is Cochrane's suggested measure of persistence. 2.6 Comparison With The KL Specification In the KL framework, persistence is defined as the present value of future revisions in expected earnings. The revisions themselves are estimated from a time series model such as an autoregressive model which is assumed to be intertemporally constant. Thus, the KL specification measures the cumulative effect of a shock on the level of a series. On the other hand, the Cochrane specification adopts the notion that if the permanent component of a series is high, then the shocks or innovations are highly persistent and vice versa. This is clear from the two cases illustrated above. In case 1, all the shocks are persistent and the series has only a permanent component. In case 2, however, none of the shocks is persistent and this series has no permanent component. Thus, measuring the magnitude of the permanent component of the series should 11 This assertion is not without basis. If most of the earnings series fell into either of the two extremes, then there would be little variability in any estimate of persistence. The results of this research (reported in chapter 5) and those in KL (Table 1 of KL) suggest significant variability in persistence. In addition, the results in a companion paper with Loungani, applying recent tests, (Thiagarajan and Loungani [1989]) suggest that the random walk hypothesis can be rejected for annual earnings. This suggests that earnings are a mixture of these two extreme processes. 15 reveal the persistence of the shocks in the series without the need for modeling the series and estimating the unobservable innovations. This is the intuition behind the Cochrane specification wherein the magnitude of the permanent component of the observed series is measured directly. The Cochrane specification focuses on measuring the permanent component of the observed series. As explained in the appendix 1, the permanent component can be estimated from the autocorrelations of the observed series and does not require the estimation of the unobserved innovations or the unobserved revisions arising from such unobserved innovations. Since the innovations need not be estimated, there is no need for an expectations model in the Cochrane specification. Thus, the problems resulting from estimation error arising from the need to parameterize an expectations model are eliminated. Some of these problems are associated with the KL specification and were discussed in section 2.4. Furthermore, the Cochrane specification provides more information about the revisions than the KL specification. In particular, the Cochrane specification measures both the sum of the future revisions and the variance associated with such future revisions whereas the KL specification measures only the sum of future revisions, appropriately discounted. This relationship is explained further in appendix 1. The Kormendi and Lipe specification, however, has the advantage of conforming to the intuitive notion of persistence as the present value of the future revisions in the series, for a given innovation. In summary, the Cochrane measure captures the long term effect of an innovation as does the KL specification but avoids the need to have an expectation model and thereby avoids some of the problems of 16 estimation inherent in the KL specification. In addition, the Cochrane specification provides additional information about the nature of revisions in the series. Both the Cochrane and the KL specifications are used for estimating persistence in the first essay, which examines the economic determinants of earnings persistence. They are also used in the second essay wherein the association between persistence and the earnings response coefficient is examined. 2.7 Estimation Of Cochrane's Measure Of Persistence To provide an intuitive description of the estimation procedure of the Cochrane's measure of persistence, we can express the earnings at time t, Xt, as the sum of two components: (1) a permanent component Zt and (2) a transitory component Ct. Thus, (6) Xt = Zt + Ct. We can rewrite this equation in first differences as (7) AX, = AZt + ACt.12 Cochrane suggests measuring persistence with the ratio Var(AZt)/Var(AXt). This ratio measures the fraction of the variance of the difference earnings series that is accounted for by the variance of the permanent component. This ratio is directly related to the size of the permanent component of the series. This definition of persistence provides an appealing measure of the relative strength of the permanent component. 12 The series is difference for two reasons: (1) to make it stationary and (2) to get finite bounds on the measure of persistence. Since first differencing may not make all the series stationary, in empirical tests all the series that were not stationary on first differencing were eliminated. The details are in Chapter 5. Cochrane shows that in finite samples this ratio can be consistently estimated by taking a linear combination of the sample autocorrelations of the difference series. This can be expressed in the following equation: Var(AZt) k kj (8) = 1 + 2 Z r Var(AXt) j=1 k where rj is the jth autocorrelation of the difference series AXt and k is the number of lags up to which the autocorrelations are estimated. This is an appealing measure of persistence, because the correlation between AXt and AX,+j, captured by Fr, measures the extent to which the innovation at time t is felt at time t+j. By including the appropriate number of lags one can obtain the relative strength of the permanent component.13 2.8 Estimates Of Persistence Based On Analysts' Forecasts14 Kormendi and Lipe define persistence as the discounted sum of the revisions made in the expectations of future earnings arising from the 13 Monte Carlo simulations by Lo and Mackinlay (1987) and by Campbell and Mankiw (1987) indicate inclusion of too many lags may result in unreliable estimates of persistence. It is difficult to define "too many" since it depends on the series. As a rule of thumb, one could adopt the suggestion of Lo and Mackinlay, in their specification test for random walk, to restrict the lag length to one eighth of the sample size. 14 Lipe (1986) proposes another measure of persistence which is very similar to the KL measure, except that the expectation model used to estimate the innovations is a firstorder autoregressive model instead of the secondorder specification used in KL. This is not discussed because it is a simplified version of the KL model. 18 current earnings innovation. Under this definition, analysts' forecast revisions can be used to measure persistence directly, as long as their revisions are available for several periods in the future. However, such revisions are available only for very short periods into the future.15 This requires an assumption that the short term revision is sufficient for capturing the effects on all the future periods. This assumption might be problematic especially in light of the Brown et. al., (1985) finding that the capital market appears to employ a multi year forecast horizon. If the market adopts a multiyear perspective, then the market agents may use the information in the current earnings innovation to revise expectations for a long period into the future. Consequently, calculating persistence based on revisions over the short horizon might result in estimates of persistence very different from those estimated by market agents which is implicit in security prices.16 Analysts' forecasts have been employed in recent studies to estimate persistence [e.g., Easton and Zmijewski (1986) and Regier (1986)]. This approach is not used here because of potential differences in the magnitude of revisions of expectations for short versus long horizons. 15The revisions are available for about four quarters ahead from the Value Line Investment Survey. 16 It is tempting to conclude that the use of the shortterm revisions of analysts' forecasts results in systematically downward biased estimates of persistence. This may not be true in all the cases as trend reversions can cause an innovation to have alternating effects on expectations over the long horizon. CHAPTER 3 MOTIVATION AND POSSIBLE ECONOMIC DETERMINANTS This chapter provides the motivation for the examination of the economic determinants of earnings persistence and outlines the determinants chosen for analysis in the first essay. 3.1 Introduction This essay examines some potential determinants of earnings persistence. In their overview of the research issues in timeseries analysis, Ball and Foster (1982) indicate "at present our knowledge as to why certain statistical properties are found for the annual earnings series of firms is very meager indeed" (page 212). The proposed research investigates the reasons for crosssectional differences in one statistical property of annual earningspersistence. Specifically, the objective is to address the following research question: "What is the role of industry factors and more importantly firmspecific factors in explaining crosssectional differences in persistence?" 3.2 Motivation Prior research has examined the relationship between persistence and security prices in general and earnings response coefficients in particular [KL and Easton and Zmijewski (1986)]. Market efficiency suggests that investors use the information about the resource stocks and flows of the firm to form their expectations about the future. By positing a positive association between persistence and earnings response coefficients, KL assume that the estimated measure of persistence captures, in a meaningful way, the economic factors that influence the market's assessments of future earnings/cash flows. The observed positive association between earnings response coefficients and persistence suggests that persistence has valuation implications.1 This, in turn, implies that persistence should be related to the underlying firmspecific variables which generate firm value. This link is illustrated in figure 2. To be meaningful for valuation, therefore, persistence must reflect information about the real resource stocks and flows induced by the firm's decisions and other factors, such as industry factors, which describe the firm's economic environment. It is, therefore, important to examine the extent to which persistence reflects the economic factors which influence the resource flows of the firm. Explicit documentation of the association between persistence and the firmspecific variables would confirm that persistence is a meaningful construct. In addition, documentation of the association between persistence and firmspecific variables could provide evidence of the information content of extant financial disclosures. Such explicit documentation is one of the goals of the research in this essay. A question of significant interest to accounting research is whether different accounting techniques contribute to crosssectional 1 The tests of KL are joint tests of the validity of the underlying model and the robustness of their measures. This research investigates whether estimates of persistence are related to decisions within the firm which affect the value of the firm. EFFICIENT MARKET HYPOTHESIS EXISTING RESEARCH Kormendi and Lipe (1987) Easton and Zmijewski (1986) PROPOSED RESEARCH ECONOMIC STUDIES Scherer (1980) Framework For The Research Question Figure 2 differences in persistence of accounting income. This is of interest because persistence has been documented as having valuation implication. If accounting techniques are found to be meaningfully associated with persistence, then it could be inferred that accounting techniques have economic consequences.2 Examination of the association between firm specific factors and persistence can help address this question. If the determinants of earnings persistence are not affected by the choice of different accounting techniques, then we may be able to infer that accounting techniques have either no effect or only a second order effect on persistence. Identifying the firmspecific factors is the first step in this process. 3.3 Economic Determinants Of PersistenceThe Framework Income arises from past firmspecific decisions (e.g., investment and operating decisions) and is influenced by a variety of factors at the industry level such as labor conditions and barriers to entry.3 From an economic perspective, income has a fundamental rolethat of a "driving force for change" [Lev (1983), p. 33]. This perspective implies that income levels and income trends signal the course of future 2The inference that accounting techniques may have economic consequences arises from the valuation implication of persistence. If the component of persistence attributable to accounting techniques can be isolated, a direct test of the economic consequences can be performed. Isolation of such a component, a formidable task, is outside the scope of this dissertation. 3 A barrier to entry is a condition which prevents potential rivals from entering the market and competing away the advantage enjoyed by the incumbent firms. This can take several forms such as absolute cost advantage enjoyed by the incumbent firms in the form of scale economies, or economies in purchasing, capital raising or promotion not attainable by the rivals. [See Scherer(1980), Chapter 8 for a discussion of the forms of barriers to entry.] 23 action both at the industry level (such as entry into and exit from the industry) and at the firm level (such as change in the level of research and development expenditures, investment in new products).4 Since earnings result from past firmspecific decisions and also provide the guidelines for future decisions, such as investment decisions, an association can be expected between these firmspecific decisions and the stochastic properties of the earnings series.5 In this essay, an association is posited between some of the these firmspecific decisions and one property of the earnings seriespersistence. Earnings persistence is a characteristic of the earnings process which reflects the extent to which an innovation (or earnings surprise) results in revisions of expected levels of future earnings. It can also be viewed as a measure of the magnitude of the permanent component of the earnings series (see appendix 1). In an efficient market economy, where the flow of resources to activities earning excess profits is relatively quick, profits above the equilibrium level should quickly 4Both historical income and expected future income provide guidelines for future action. Expectations for future income are formed from trends of past historical income. 5This relationship is essentially dynamic. Past decisions affect current income and current income and income trends affect future decisions and so on. Thus, over time, the stochastic properties of accounting income should be related to the general pattern of firm specific decisions. 6 It is natural to think that since the effects of all these firm specific decisions eventually flow through the income statement, there should be relationship by construction. It needs to be noted that persistence is a well defined specific property of the series. It is not at all clear there would be a systematic relationship between this property of earnings and an accounting policy such as expensing of the research and development expenses. disappear. This implies that persistence of profits in any industry will attract new capital to the industry and persistence of losses will detract capital away from the industry. If this process is quick and efficient, the firms in any industry would earn only the longrun or equilibrium rate of return and crosssectional differences in persistence across industries would be minimal. Consequently, existence of different levels of persistence across industries should reflect certain characteristics at the industry level (such as entry barriers) which may contribute to this disequilibrium. Similarly, existence of different levels of persistence across firms within an industry should reflect certain firm specific actions (e.g., product differentiation). Persistence can, therefore, be characterized as a function of both industry factors and firmspecific factors as illustrated in figure 3. The proposed examination of the determinants of persistence is restricted to the chosen firmspecific factors. Consequently, the proposed examination recognizes the industrylevel characteristics and examines certain firmspecific actions which are hypothesized to result in differential persistence. 3.4 Persistence Of Industry Earnings Several industryspecific factors influence a firm's earnings seriese.g., barriers to entry, demand variability, product type [see Lev(1983)]. For example, an industry may have several barriers to entry, such as the scale of operations required to enter the industry and legal restrictions in the form of licenses. FACTORS INFLUENCING PERSISTENCE Industry Factors Firm Specific Factors Nature of the product Demand variability Labor conditions Barriers to entry Investment Decisions Other Decisions Capital expenditures R&D Advertising Classification Of Factors Influencing Persistence Figure 3 The persistence of profits in such an industry (and hence the persistence of profits of the firms in that industry) is likely to be higher than the persistence in an industry which does not have such barriers to entry. Since the primary objective of the first essay is to examine the relationship between earnings persistence and selected firm specific variables, the industryspecific determinants of persistence 26 must be controlled.7 One way of controlling for industry factors is to develop appropriate empirical surrogates for all the variables affecting industry earnings. An alternative is to use the persistence of industry earnings to capture the effects of all such industry factors on individual firms' earnings persistence. The second approach is adopted here. It is expected that the higher the persistence of the industry's earnings (for whatever reason), the higher the persistence of the earnings of a firm in that industry. Thus, industry persistence should be positively related with firmspecific persistence. 3.5 Firm Specific Factors The firmspecific decisions which could influence earnings include investment decisions, labor decisions, production decisions and financing decisions. All these decisions could potentially influence the firm's earnings persistence. However, for parsimony and given the difficulty in obtaining data on other decisions (such production decisions in the form of product quality changes), the proposed empirical examination is restricted to the role of the firm's investment decisions in determining persistence. Investment decisions are defined broadly to include research and development (hereafter, R&D) and advertising which have potential longterm effects and arise from the discretionary investment spending of the management. Since management's investment decisions have a major impact on the stream of future 7The specific industry level determinants of persistence are interesting in themselves (see appendix 2 for some results in this regard). However, a majority of these determinants (such as entry barriers) are outside the firm's control. The scope of this essay are restricted to controllable firmspecific factors. earnings, these decisions are expected to be a major determinant of persistence. In summary, out of the several factors which influence earnings persistence at the firm level, this research examines the role of investment decisions as potential explanators of crosssectional differences in earnings persistence. The framework for the choice of factors and the hypothesized direction of their effects on earnings persistence are based on the industrial organization literature in economics [Scherer (1980)]. 3.5.1 Investments In Property. Plant And Equipment. Investment provides evidence that management has identified positive net present value projects and an earnings innovation confirms these expectations. A high rate of investment indicates that management expects sustained positive earnings innovations from their investments. In addition, increased investment in property, plant and equipment enables the firm to increase the scale of operations and potentially reduce the longterm average cost of production. Sustained investment in profitable projects coupled with the reduction in the average cost imply that the persistence of earnings for such a firm should be high. In other words, high rates of investment are expected to result in high persistence of earnings. This conjecture is supported (at the industry level) by Stigler's (1963) finding of a positive correlation between expected profits and investment growth in the nonrecession years and by Wenders (1971) finding of a positive association between profitability and asset growth. Thus, the expected relationship between rate of investments and persistence is positive.8 3.5.2 R&D And Advertising Industrial organization literature suggests that both R&D and advertising create barriers to entry that enable incumbent firms to sustain high profits. This is done by discouraging firms seeking to enter the industry through increased production costs, R&D and advertising costs that have to be borne by the potential entrant [see Clarkson and Miller (1982, chapter 16)]. Barriers to entry enable the firm to sustain high profits by discouraging competition resulting in persistence of such high profits. Thus, R&D and advertising are potential determinants of persistence. Research and development. Investments in research and development might result in new products and/or cost saving technological change, resulting in higher profits over a long period of time.9 The higher the R&D expenditure, the better the position of the firm to discover and develop new products or lower costs and sustain the profits arising out of such products/technological changes. Empirical support for this expectation can be found in Branch (1974) whose results support the causal links from profitability to R&D efforts and from R&D efforts to increased profitability. Thus, incremental spending on R&D is expected 8If there is a decline in earnings which is perceived to be permanent, there is likely to be disinvestment in order to reduce the persistence of losses. In a growing industry, investment is positively associated with persistence of profits and in a declining industry, investment is positively associated with persistence of losses. 9 Registered patents arising out of R&D efforts can be retained for 17 years. to result in persistence of profits and persistence of high profits is likely to encourage R&D spending."1 Therefore, the expected relationship between persistence and the magnitude of R&D is positive. Advertising. Economic studies [Comanor and Wilson (1977), Weiss (1969)] indicate that high advertising expenses create a barrier to entry and result in high persistence of profits. A firm which advertises heavily, relative to the industry, may be able to create "brand loyalty" and thus reap sustained high profits. Thus, the more the firm spends on advertising, the greater their chance to preserve an earnings increase arising out of the demand so captured. Thus, the magnitude of advertising is expected to be positively associated with persistence. 3.6 Effects Of Financing Decisions Earnings are affected not only by the management's decisions concerning investments, advertising and R&D, but also by financing decisions. Different capital structures result in different finance charges, and hence differences in earnings behavior." Since the framework for empirical investigation is provided by the investment decisions of the management, the effects of different financing decisions are outside the scope of this study. An additional reason for excluding the effects of different financing decisions is the absence of 0Just as increased profits arising out of R&D is likely to spur increased R&D spending, increased profits from capital expenditure and advertising are also likely to encourage further spending on such investment. 1 Differences in capital structure result in different earnings levels which could result in different time series properties. a clear direction of the effects of different choice on persistence. For example, it is not clear that the choice of debt over equity would have a systematic effect on persistence. Therefore, this study abstracts from the effects of financing decisions on persistence by defining persistence in terms of operating income, rather than net income. Operating income excludes interest income and interest expenditure and hence abstracts from the effect of financing decisions. 3.7 Effects Of Accounting Techniques A question of considerable interest is whether crosssectional differences in the choice of accounting methods can explain differential persistence of reported accounting income. As indicated earlier (in chapter 2) this is of interest because of the observed valuation implication of persistence (Kormendi and Lipe). If choice of accounting methods is a determinant of persistence, then, by inference, accounting methods have economic consequence. This question arises because accounting income is not only affected by the firmspecific decisions on investments, R&D etc., but also by the accounting techniques used to measure them. Consequently, the estimated persistence of accounting income is likely to be influenced by the choice of accounting techniques. The effect on persistence arises due to two sources. The first source is the independent effect of different choices due timing differences. To provide an example, the income numbers of a firm using straight line depreciation will be different from those of a firm using declining balance methods. The timing difference from this source gets reflected in the measured persistence. The second source is the interactive effect. To continue the prior example, the firm using straight line depreciation could be making systematically different investment decisions compared to a firm using declining balance methods. The economic consequences literature posits interdependence between the choice of accounting methods and the substantive attributes of the firm's decisions. The choice of accounting methods may affect management's decisions about production, investment, marketing and financing (PIMF) because of links arising out of management compensation and contracting costs [see Gonedes and Dopuch (1979) and Watts and Zimmerman (1986)]. These effects may be explained as follows: Accounting Income = f(economic environment; firm's production, marketing and finance [PIMF]; and accounting techniques). Firmspecific Persistence (FSP) = F(Accounting Income). Therefore, FSP = G(economic environment, PIMF and accounting techniques). Economic consequences theory suggests that PIMF = h(accounting techniques). This study captures some of the aspects of the economic environment through the industry persistence measure. Production, investment and marketing decisions are captured by the investment, advertising and R&D variables and abstracts from the effects of financing decisions.12 12 It is expected that most (although not all) aspects of production, investment and marketing decisions may not be captured by the investment, advertising and R&D variables. In summary, the choice of accounting methods may have both an independent and an interactive effect on the measured FSP. The research interest arises primarily because of the interactive effect of accounting techniques on FSP through management decisions on PIMF. The independent effect of choice of accounting methods on persistence could pose a validity threat to the study and confound the effects of the economic determinants.13 That is, any difference in persistence across firms could be attributed to the economic determinants when it may have arisen from different accounting methods. To provide some insight into the effect of accounting techniques on persistence, this study employs two procedures: First, the analysis is replicated using two different dependent variables which are expected to be less affected by choice of accounting techniques than is operating income.14 These variables are (1) persistence of sales and (2) persistence of cash flow from operations (CFO). Differential choice of accounting techniques affects sales much less than operating income [see Lev (1983) for a similar treatment]. However, sales may not capture all the effects of the independent variables. For example, cost savings arising out of R&D will not be reflected in sales. This implies that there may be some loss of information in the persistence of sales. 13 The interactive effect poses less of a validity threat. If accounting techniques have an interactive effect, then it should be considered another or part of the other economic determinants. 14 An alternative way of examining whether the estimated measure of earnings persistence, based on accounting income, reflects real resource flows is to examine the association between firm specific persistence and nonaccounting variables such as production, employment etc. This is investigated in a related paper. Consequently, the analysis is also replicated based on the persistence of CFO as a second proxy for persistence of earnings. The advantage of this proxy is that it is likely to reflect more of the effects of the economic determinants; the disadvantage is that it may not be completely independent of the effects of accounting techniques.15 Replication of the analysis using these two different dependent variables should enable us to infer if the choice of accounting techniques affects the relationship between the hypothesized determinants and firmspecific persistence. If this procedure indicates that choice of accounting techniques has an effect on the estimated persistence, then the inference is that such an effect could be due to either the independent effect due to the timing differences inherent in the different methods of accounting or due to economic consequences or a combination of both. It would be possible to isolate these two effects if we could specify the direction and/or magnitude of either the timing difference effect or the economic effect. Unfortunately this is not an easy task for either the independent effect or the interactive effect.16 Hence, the most 15The reason why this may not be completely independent of the effects of accounting techniques is because the cash flow variable is generally derived from the income variable after removing the effects of all noncash items such as depreciation and amortization. Consequently, some of the effects of items such as choice of inventory accounting method is likely to be reflected in the derived cashflow variable. 16 Specification of the independent effect involves the delineation of all combinations of the different accounting choices and the examination of the effect of each of the combination on persistence. Since the nature of relationship between accounting choices and persistence is not well specified, it is not possible to specify the nature and magnitude of the interactive effect. For example, the effect of a portfolio of accounting choices such as an incomedecreasing strategy on persistence is not clear. 34 that could be done is to perform overall tests to check if differential choice of accounting methods makes any overall differences in the inferences to be drawn. Analysis of the interactive effects of accounting techniques on persistence is interesting due to the potential insights on the economic consequences of the choice of accounting methods. However, a rigorous analysis is hampered due to two reasons: (a) limited data availability of detailed information on firm's accounting techniques across time and (b) inability to postulate an unambiguous direction of the effects of any particular choice of accounting technique on persistence. The scope of this study is limited to examining if crosssectional differences in choice of accounting techniques have an effect on persistence. Specification of the source of such an effect, if it exists, is outside the scope of this dissertation. 3.8 Conclusion The variables chosen as possible determinants of persistence are based on economic reasoning supported by empirical findings found in the industrial organization literature. This study does not claim that these variables constitute a complete structural model of the relationship between a systematic property of accounting numbers, namely persistence, and the attributes of firms' decisions. It is hoped that the empirical relationships observed in this study will pave the way for a more rigorous model of the determinants of earnings persistence. CHAPTER 4 MODELS, MEASUREMENT OF VARIABLES AND SAMPLE CHARACTERISTICS 4.1 Basic Model The analysis involves a crosssectional regression with firm specific persistence (FSP) as the dependent variable. Since FSP is a function of both industry factors and firm specific factors, as explained in chapter 3, the independent variables include industry variables and firm specific variables. The industry factors are controlled through the measure of industry specific persistence (ISP). ISP is used as a covariate in the proposed cross sectional regression. The other independent variables are the firm specific variables which include investment, R&D and advertising variables. FSP and ISP are primarily operationalized through Cochrane's persistence measure which is calculated as a linear combination of autocorrelation coefficients.1 Since autocorrelation coefficients are scale free, the estimates of FSP and ISP are also scale free. The other independent variables, the firm specific economic factors, are normalized by sales to make them 1 To test for robustness of the association between the independent and the dependent variables, the basic model is replicated using the KL measure of persistence. This is explained further in the chapter. 35 independent of the units of measurement.2 This process makes the definition of all the variables consistent. All the firm specific variables are defined as averages calculated over the sample period. This definition makes the independent variables compatible with the FSP and ISP which are based on the average magnitude of the permanent component of all the shocks in the earnings series of the firm and the industry, respectively. Based on this reasoning the following cross sectional model will be estimated: (9) FSPij = a + Po ISPj + fi Average investment(n)ij + 02 Average R&D(n)ij + P3 Average advertising(n)ij + eij, where FSP is firmspecific persistence ISP is industryspecific persistence n denotes the respective variables being normalized by sales i = 1,...,p is the firm subscript j = 1,...,q is the industry subscript. Since each of the economic determinants identified in sections 3.4 and 3.5 are expected to be positively associated with FSP, the expected sign on all the P's is positive. Conceptually, the beta coefficients in this model measure the extent to which a dollar increase in normalized investment impacts on the permanent component of operating earnings. 2 Normalizing by sales also helps control for size effects in the independent variable. For example, the magnitude of advertising, one of the selected independent variables, may be influenced by size but persistence would not be so influenced, because autocorrelations are impervious to scale. 4.2 Measurement Of Variables FSP. Firmspecific persistence (FSP) is calculated for each firm using Cochrane's specification. As indicated earlier, Cochrane suggests the limit of the ratio of variances as a measure of persistence. This ratio is defined in the following way: 1 Var (xt+k xt) k Var (xt+l xt) where x, denotes the value of the difference operating income at time t and k denotes the lag length. It is important to choose the appropriate lag length relative to the sample size. Campbell and Mankiw (1987) caution that including too few lags may obscure the trend reversion manifested in higher autocorrelations. On the other hand, including too many lags, relative to the sample size, results in a downward bias in the estimates of persistence.3 Moreover, Lo and Mackinlay (1987) [LM] find that lag lengths in excess of oneeighth of the sample size, result in unreliable inferences.4 Using this rule of thumb to estimate persistence, this study limits the lag length to oneeighth of the available timeseries observations for each firm in the sample. Since the minimum number of 3 Campbell and Mankiw (1987) find it difficult to distinguish between a random walk and an AR(2) process in their Monte Carlo simulations for large lags (relative to the sample size) due to the downward bias in the persistence measures calculated using large lags [See Campbell and Mankiw, p. 113, Table 1]. It is important to note that sample size here indicates the number of time series observations. 4 The suggestion of LM was made in the context of a specification test for random walk. However, since the specification test is based on variance ratios, this benchmark may be applied here. 38 timeseries observations (on earnings) available for each sample firm is 35 (the maximum is 36), FSP is calculated for each firm using a lag length (k) of five.5 To improve the smallsample behavior of the estimates of persistence, this study adopts both the suggestions of Cochrane(1988). In the first place, unbiased estimates of variances are used by calculating s2 instead of a2. Secondly, the sample mean of the first differences is used as an estimate of the drift term in the process that describes the permanent component instead of estimating a different drift term for each k.6 Without these corrections, there is a downward bias in the estimates of persistence for finite sample as the lag length is increased (relative to the sample size). To control for nonstationarities arising out of price level changes, the operating income is adjusted for price level changes using the GNP deflator. In summary, FSP is measured by the ratio of the unbiased variance of the fifth lag of the difference real operating income series to the 5 One fifth of 35 is 4.38. Since the lag length cannot be in fractions it is rounded to the next highest number. Preliminary investigation of my sample indicates that the rank correlation between estimates of persistence using lag length of four and lag length of five is .953 (product moment correlation is .98). This suggests that inferences based on persistence estimates based on lag length of five would not be very different from those based on lag length of four. 6 While the intuition behind the use of s2 is clear, the intuition behind the estimation of the drift term is not that obvious. This correction is required to prevent the variance of k differences from declining toward zero as k, the lag length, tends toward t, the sample size (see Cochrane[1988] for details). 39 unbiased variance of the first lag.7 The operating income series is obtained from Compustat.8 ISP. Industry specific persistence (ISP) is calculated in the same manner as FSP using the aggregation of sample firms' earnings in the different industries.9 Average investments(n). The normalized investment for each firm i, for each year t is defined as Investment in property, plant and equipment Salesit. The average of such ratios calculated over the sample period forms the average investment(n) for each firm in the sample. Investment in property, plant and equipment of firm i at time t is obtained from Compustat item thirty. Additions to property, plant and equipment due 7 Cochrane's measure of persistence can be calculated from the estimated sample autocorrelations. In this study, FSP is estimated directly as a ratio of variances. There is no difference in these two methods of estimating persistence. 8 The operating income is obtained from Compustat items thirteen (operating income before depreciation) and fourteen (depreciation). 9 Industrywide aggregate earnings are not available on the Compustat tape for the early years of the sample. In order to have a consistent definition of industry earnings, over time, an equally weighted index of industry earnings was created from the sample firms. This aggregation of earnings was made out of the firms in the sample at the twodigit SIC industry code level. To provide an example, if there were two firms one with the SIC industry code of 2800 and the second with a code of 2834, both of these firms had the same industry persistence as a control for industry factors. Such industry persistence was measured on a series of industry earnings formed from the aggregation of the earnings of all the firms with the 2 digit SIC code of 28. The limitations of this definition are discussed in chapter to merger activity are included in the definition of investment. The reason for this is that operating income, which provides the basis for the measurement of the dependent variable, is likely to be influenced by such additions. The sales for firm i at time t is obtained from Compustat item twelve. Average R&D(n). The normalized value of R&D is calculated for each firm 1, for each year t, as the ratio of R&D to sales. The average of such ratio calculated over the sample period forms average R&D(n) for each firm. The R&D for each year is obtained from Compustat item forty six and the sales from item twelve. Average advertising(n). The average magnitude of advertising is calculated in a manner analogous to the calculation of the average R&D(n). Advertising for each firm in each year is obtained from Compustat item fortyfive and the sales from item twelve.10 4.3 Kormendi And Lipe Model This study compares the results of the basic model (equation 1) across both the definitions of persistence discussed in Chapter 3, namely the nonparametric measure suggested by Cochrane and the parametric measure employed by Kormendi and Lipe. If the nature and the strength of the relationship is similar across the two definitions, then Cochrane's nonparametric approach provides only a computational 10 It will be noticed that the dependent variable, FSP and one of the independent variables ISP are adjusted for pricelevel changes using the GNP deflator, whereas none of the other independent variables are so adjusted. The reason for this is that all the other independent variables are normalized by sales and any adjustments for pricelevel changes has to be done both to the numerator and the denominator. Such adjustments both to the numerator and the denominator will be canceled out making them equivalent to the unadjusted ratios. advantage with respect to this sample. If the relationship is different, then the nature and the magnitude of the relationship specified in the basic model (equation 1) is sensitive the measurement of persistence.1 In order to examine this, the basic model is reestimated using the KL measure of persistence. The Kormendi and Lipe specification is independent of the units of measurement analogous to the Cochrane specification. Consequently, the definition of the variables in this model are analogous to the basic model (equation 1). To keep the measurement of FSP and ISP consistent, both of them are measured using the KL specification and denoted FSP(KL) and ISP(KL) respectively. The KL model may be stated as follows: (10) FSP(KL), = a** + **oISP(KL)j + 8**i Average investment(n)ij + P**2 Average R&D(n)ij + P** Average advertising(n)ij + eij As in the basic model, all the P** 's are expected to be positive in the KL model. In summary, the KL model is estimated to assess the sensitivity of the relationship postulated in equation 1 to the measurement of persistence. Conceptually, the beta coefficients in the KL model 11 Clearly, the relevant question here is which specification is better. The appendix explains that the Cochrane specification provides more information about the series than the KL specification. If the results of the two specifications are similar, then such additional information does not affect the determinants specified in this study. If the results are different, it can be argued that the such additional information does affect the determinants. measure the extent to which a dollar increase in the normalized investment impacts on the present value of the revisions of future operating earnings. As suggested earlier, the KL model is conceptually more appealing than the Cochrane model, but the purpose of the estimation here is to assess the sensitivity of the relationship to the nature of measurement of persistence. 4.4 Tests For The Effects Of Accounting Methods Overall Tests. To provide some insight into the effects of cross sectional differences in the choice of accounting methods on persistence, equation (1) is reestimated using persistence of sales [FSP(S)] and persistence of cash flow from operations [FSP(CFO)] as the dependent variables. The models are: (11) FSP(S)i = a' + P'o ISP(S)j + P3 Average investment(n)ij + 80'2 Average R&D(n)ij + 3 Average advertising(n)ij + e'ij (12) FSP(CFO),j = a + "o ISP(CFO)j + + /3" Average investment(n)ij + /"2 Average R&D(n)ij + pf Average advertising(n)ij + e ij. Both sales and CFO are obtained from Compustat.12 A priori, cross 12In order that CFO reflect the effects of total accruals, both current and noncurrent, but not the effects of financing decisions it is defined as follows [adapted from Wilson(1987)]: CFOt = OIBDt + (CL, DCLt) (CLt DCIt1) (CAt CASH,) + (CAtg_ CASHt,1) where sectional differences in accounting techniques are not expected to significantly affect the substantive relationship between the economic determinants and FSP. Consequently, the same relationships should hold for the persistence of sales and the persistence of CFO as for the persistence of operating income. Thus, P coefficients in equations 11 and 12 are expected to be positive. OIBD is the operating income before depreciation and amortization at the end of year t (13) CL is the current liabilities at the end of year t (5) DCL, is the current portion of long term debt included in current liabilities at the end of year t (34) CAt is the current assets at the end of year t (4) CASHt is the cash and short term investments at the end of year t (1). Numbers in parenthesis refer to Compustat data items. OIBD adjusts the operating income for noncurrent accruals such as depreciation, amortization, minority interest, deferred taxes etc., and the other items represent the adjustment for current accruals (i.e., the net changes in all the working capital accounts except for changes in cash, marketable securities, and short term debt). CHAPTER 5 SAMPLE SELECTION,DESCRIPTIVE STATISTICS AND RESULTS OF MODEL ESTIMATION This chapter outlines the basis for the choice of the sample, discusses the issues concerning the operational definition of the variables and provides the relevant descriptive statistics. This chapter also discusses the results of the estimation of the different models, summarizes the inferences to be drawn from the results. 5.1 Sample Selection The sample was selected by combining data from two Compustat tapes, the first carrying data from 1950 through 1969 and the second carrying data from 1966 through 1985. There were several restrictions for a firm to be included in the sample. These restrictions can be classified as follows: (a) restrictions concerning the measurement of persistence, (b) restrictions concerning the formation of the industry index to calculate industry persistence and, (c) restrictions concerning the measurement of the firm specific variables. Each of these restrictions is discussed in the following paragraphs. Restrictions concerning measurement of persistence. The calculation of persistence requires the availability of observations on a continuous basis. The primary restriction was the availability of data on a continuous basis on operating and net income.1 This step yielded 403 firms. The calculation of both the Cochrane measure of persistence and the KL measure of persistence requires that the difference earnings be stationary.2 Stationarity was checked by examining the autocorrelation series of the difference real operating income to ensure that they dampen out at higher lags. The autocorrelation coefficients were examined upto 24 lags to ensure that there were no nonstationary series. This procedure resulted in the elimination of 51 firms. Restriction required to calculate industry persistence. Persistence of industry earnings was calculated from an index of industry earnings formed from the earnings of sample firms.3 Hence, it was important to ensure that the selected firms were defined to be in the same major industry group during the sample period. The definition of industry, in this study, is based on the 2 digit SIC code.4 The industry affiliation of the sample firms throughout the sample period 1 The need to ensure that data on net income were available on a continuous basis was required to enable the calculation industry persistence on net income. 2 Stationary is required in the Cochrane specification to get bounds on the persistence measure. In the KL specification, if the series is nonstationary, the series would not be invertible. 3 This problem arises because aggregate data on industry earnings were not available on the Compustat tape for the early years. 4 As explained in footnote 9 of Chapter 4, the earnings of the sample firms were accumulated to form an index of industry earnings. This accumulation was done at the 2 digit industry code level and not at the 4 digit code level because there was not a sufficient number of observations at the 4 digit level to form any reasonable estimate of industry earnings. was checked with the CRSP tape and the firms which had changed affiliation at the two digit industry level during the sample period were eliminated. This data restriction resulted in the elimination of 26 firms. Since the earnings of the sample firms were accumulated to form an index of industry earnings, it was important to ensure that there were adequate number of firms in each industry to get a representative index of industry earnings. To illustrate, in the automotive service and repair industry (SIC code 7500) there is only one firm in the sample. To use the earnings of a single firm to form an index of industry earnings, would not be representative. Firms were included in the sample only if they represented at least 20% of all firms in the industry.5 The index of industry earnings formed using this truncation rule represented, on an average, 36% of the earnings of each industry at the level of 2 digit SIC code.6 This step resulted in the elimination of 29 firms. Restrictions concerning measurement of firm specific variables. As indicated in Chapter 4, the independent variables are capital expenditures, R&D and advertising. While information on most firms' capital expenditures is available on the Compustat tape, information is not always available on the other firm specific variables, R&D and advertising. In the utility industry, for example, there are 81 sample 5 Industry affiliation was determined by checking the Compustat files. 6 Preliminary analysis showed that a higher truncation rule (30%) made little difference in the estimates of industry specific persistence. Since a higher truncation rule results in loss of firms, all the analyses were based on industry earnings formed using the 20% truncation rule. 47 firms with no information on R&D and advertising. Only firms which had information available on advertising and R&D for at least 5 years during the sample period were included.7 This procedure resulted in the elimination of 111 firms, including 81 firms in the utility industry. The average number of years of data availability was 12 years for advertising and 17 years for R&D. All the above steps resulted in a final sample of 186 firms as summarized in table 1. Table 1 Sample Selection Total number of firms from two Compustat files 403 Removed because of nonstationarity 51 Removed due to change of industry 26 Removed to enable the calculation of industry specific persistence 29 Removed due to non availability of information on R&D and advertising 111 Final Sample 186 The next section discusses the issues concerning the empirical specification of the variables in the different models. 5.2 Matters Concerning Empirical Specification The basic model to be estimated,as specified in chapter 4, is as follows: (1) FSPij = a + po ISPj + pl Average investment(n)ij + P2 Average R&D(n),i + P3 Average advertising(n)ij + eij 7 Using a higher truncation rule, while restricting the number of firms did not change any of the results. The estimation plan for each of the variables was discussed in chapter 4. However, the operational definition of some of the variables gives rise to certain measurement issues. These are discussed in the following paragraphs. Specification of ISP. There are three issues concerning the specification of ISP. The first relates to the appropriateness of the choice of industry persistence to surrogate for the industry factors. As indicated earlier, industry specific persistence is measured by using the Cochrane specification of persistence on an index of industry earnings. One way of controlling for industrylevel determinants is to develop appropriate surrogates for each determinant. An alternative adopted in the dissertation is to use industry specific persistence. In order to ensure that industry specific persistence does indeed reflect some of these industry level determinants of persistence, the association between ISP and some of these determinants was examined. The details of this analysis is provided in appendix 2. The results suggest that the estimated ISP reflects the major determinants of industry persistence. Thus, the choice of ISP to surrogate for the industry level determinants seems appropriate. The second and the third issues relate to the actual measurement of ISP. The second issue is the definition of income for estimating ISP and the third issue is the definition of the sample to construct the industry earnings index. Since ISP is used to surrogate industry level determinants of earnings persistence, it is important that the choice of income for aggregation to measure ISP be such that the industry level determinants are well represented. Income could be defined as operating 49 income or net income (before extraordinary items). Net income includes items such as cost of idle capacity, foreign exchange adjustments and finance charges. Industrial organization literature suggests that excess capacity could be important deterrent to entry by potential rivals [see Wenders (1971) and Spence (1977)]. Foreign exchange adjustments are likely to reveal characteristics about the industry such as the nature of markets in which they operate. Finance charges are likely to reveal information about the absolute cost advantages enjoyed by the firms in the industry.8 Thus, in terms of being able to reflect industry characteristics, net income seems more appropriate than operating income.9 The analyses reported in this chapter are all based on the definition of ISP based on net income. ISP was measured on an index of earnings constructed from the sample firms. The index was created by aggregating sample firms' income. Even though the final sample is 186 firms, the aggregation of industry earnings was based on all of the 403 firms excluding the 29 firms which were removed because of they did not satisfy the 20% cut off 8It is reasonable to argue that use of net income while reflecting certain characteristics of the industry might introduce some noise due to differential choice of accounting methods for items such as pension costs or investment tax credit. This is less of a concern for industry earnings than the earnings of the firm, due to the aggregation process. 9ISP was calculated using both these definitions of income. The product moment correlation between these two measures is .88 and the Spearman correlation is .68. This suggests that the choice of the definition of income is less of a concern with respect to the estimated measure of industry persistence. All the analyses were repeated using the alternative definition of ISP and produced no change in the results. rule (see table 1). This provides a large number of firms and hence a more representative index of industry earnings.10 Specification of firm specific variables. The firm specific variables, outlined in chapter 4, require the calculation of the average over the sample period. In addition, all the three chosen variables need to be standardized by a size variable. This gives rise to certain measurement issues. The size variable chosen for standardization could either be sales or the book value of assets.1 The average calculated over time could be either the median or the mean. Each of the independent variables, R&D, advertising and capital expenditures were standardized by both sales and the book value of assets. The correlation between the two versions of the operational definition is very high. The Spearman correlation between versions ranges from .78 (for capital expenditure) to .98 (for both R&D and 10 The rationale for adding back the firms is that the reasons for excluding the firms from the final sample such as (a) nonstationarities of firms' earnings or (b) nonavailability of information on investment, R&D etc and (c) change of industry should not affect the aggregation of industry earnings. Nonstationarities of firm's earnings does not imply nonstationarity of the industry earnings. Since ISP is measured on the aggregated earnings, the first difference of the index so formed should be stationary. This was checked and the indices for all the industries were stationary. Nonavailability of information on the independent variables should not affect the aggregation of industry earnings. Finally, if a firm changes industry at the two digit level during the sample period, the income of such firm should be included in the appropriate industry to get a representative index. 11 A third alternative for a size variable would be the market value of the firm at the end of the fiscal year. This alternative was not considered because the fiscal year endings of the firms in the sample were not uniform either across the firms in the sample or across time making it more difficult to obtain the market value of each firm for each year. 51 advertising).12 Since the correlations are very high (and significant) the inferences should not be affected by the choice of size variable for deflation. The independent variables standardized by sales is the operational definition used for estimation of the different models.13 The rank correlation between the mean and the median of the different variables is also very high ranging from .93 to .98 (Pearson correlation ranges from .93 to .99). This suggests that the definition of the average makes no difference in the inference to be drawn. The reported results are based on the mean of each variable.14 5.3 Descriptive Statistics In this section the mean and the correlations of the relevant variables are discussed. Table 2 shows the mean, median and the standard deviation of the variables used in the different models. An examination of the descriptive statistics suggests that the mean of FSP is fairly high when compared to the theoretical maximum of 1 (the random walk case) suggesting a high permanent component. This is true with respect to the KL measure of persistence as well where the maximum is 10 (assuming an interest rate of 10%). However, given the mean, the standard deviation is large enough to provide a reasonable degree of variability in the FSP measure. 12The Pearson correlation ranges from .68 (for capital expenditure) to .96 (for R&D). 13 In the preliminary stages of the analyses, all the key models were estimated using both the versions of the variables. There was no differences in any of the inferences. 14 As in the case of choice of size variable, all the key model were estimated using the median of the respective variables, and none of the inferences changed. Table 2 Descriptive Statistics Median Standard Deviation FSP(C) FSP(KL) ISPNI INDNI(KL) RDSMN ADVSMN CEMSMN MNREALRD* MNREALADV* MNREALCEM* MNNOMRD* MNNOMADV* MNNOMCEM* MNREALSAL* MNREALASST* MNNOMSAL* MNNOMASST* * indicates that the unit .81 .69 9.03 8.19 .60 .58 9.87 8.14 .02 .013 .014 .004 .07 .05 .99 .846 .54 .47 2.34 2.03 73.21 12.26 43.61 4.3 146.30 19.04 29.87 26.98 23.7 22.77 1835.2 607.7 1421.98 485.8 of measurement is a million Legend: FSP(C) Cochrane's specification of firm specific persistence FSP(KL) KL specification of firm specific persistence ISPNI(C)  ISPNI(KL) RDSMN ADVSMN CEMSMN MNREALRD MNREALADV MNREALCEM MNNOMRD MNNOMADV MNNOMCEM MNREALSAL MNREALASST MNNOMSAL MNNOMASST Cochrane's specification of industry specific persistence KL specification of industry specific persistence. Mean Mean Mean Mean Mean Mean Mean Mean Mean Mean Mean Mean Mean of real R&D expenditures standardized by sales of real advertising expenditures standardized by sales of real capital expenditures standardized by sales of real R&D expenditures of real advertising expenditures of real capital expenditure of nominal R&D of nominal advertising expense of nominal capital expenditures of real sales of real assets of nominal sales of nominal value of assets. * Mean refers to the crosssectional mean of the specified variable, say RDSMN, where the variable itself is defined as the timeseries average of R&D standardized by sales for each firm. Variable Mean .45 3.27 .31 4.27 .02 .022 .07 2.7 1.18 6.8 230.96 106.02 544.64 75.72 55.62 4068.3 2973.6 dollars * It is interesting to note that the average of all the measures of persistence, (FSP, ISPOI and ISPNI) are well within the theoretical limit of 1 (the random walk case) in the case of the Cochrane specification and within the limit of 10 in case of the KL specification. This is encouraging since the number of time series observations used to calculate the persistence measures is not very large compared to the use of this measure in other applications such as GNP or stock prices.15 The averages of the independent variables, RDSMN, ADVSMN AND CEMSMN, are quite small in magnitude."1 It is important to note that standardization by sales makes these averages very small. This is clear from the examination of the unstandardized averages of real R&D, advertising and capital expenditures, denoted as MNREALRD, MNREALADV and MNREALCEM respectively and the examination of their nominal counterparts MNNOMRD, MNNOMADV and MNNOMCEM respectively. The average of real R&D expenditures across the sample, for example, is close to a million dollars per year whereas the median nominal R&D expenditure, unadjusted for price level changes, is about 12.5 million dollars. The general inference is that even though the absolute magnitude of the independent variables appear small, the 15 Cochrane (1988) uses the measure for GNP and Fama and French(1987) use the measure in examining the permanent component of stock prices. 16 RDSMN ranged from 0 to .10 million dollars per year, ADVSMN ranged from 0 to .14 million dollars per year and CEMSMN ranged from .009 to .764 million dollars per year. underlying dollar amounts of the variables are not trivial.17 The median sales of firms in the sample, adjusted for price level changes is approximately 24 million dollars and the average value of assets, so adjusted, is about 27 million dollars. The average size of the firms in the sample is better revealed by examining the median nominal sales which is about 608 million dollars, whereas the median nominal value of the assets is close 490 million dollars. This suggests that the average size of the firms in the sample is quite large. This is not surprising since the sample firms includes only those which have been in existence for at least 35 years. In summary, the descriptive statistics suggest that both the size of the firms in the sample and the magnitude of the relevant variables are large. The need to standardize the chosen independent variables diminish the underlying magnitude. It is difficult to make inferences based on the statistics in table 2, because the numbers represent averages over a long period of time (19501985) and across the sample of 186 firms. The next section discusses the correlations among the different variables used in the basic model and the KL model. Correlations among the variables used in the different models. Panel A of table 3 discusses the relationship between the dependent and the independent variables used in the basic model and panel B discusses the same relationship for the KL model. The correlations in panel A of table 3 suggests that firm specific persistence (FSP) is positively and 17 Approximately 3% (.99/29.87) of the priceadjusted revenue was being spent on R&D, on an average and approximately 1.8% of the price adjusted revenue was being spent on advertising. significantly associated with each of the independent variables in a significant manner. The correlation between FSP and ISPNI (Industry specific persistence based on net income) is the highest. Table 3 PANEL A Pearson Correlations Among The Dependant And Independent Variables For The Basic Model ISPNI RDSMN ADVSMN CEMSMN FSP .27 .18 .13 .16 (.0002) (.014) (.07) (.03) ISPNI .17 .10 .07 (.02) (.16) (.30) RDSMN .001 .009 (.9037) (.9000) ADVSMN .186 (.011) PANEL B Pearson Correlations Amongst The Dependant And Independent Variables For The KL Model ISPNI(KL) RDSMN ADVSMN CEMSMN FSP(KL) .22 .123 .14 .117 (.003) (.099) (.06) (.115) ISPNI(KL) .12 .28 100 (.109) (.0001) (.18) RDSMN .015 .007 (.8413) (.921) RADVMEAN 194 (.008) The numbers in parenthesis are the significance probability of correlation and denote PROB > IRI where R is the probability. The examination of the correlations between the independent variables does not suggest a multicollinearity problem. Even though the correlation between ISPNI and RDSMN and the correlation between ADVSMN and CEMSMN appear significant, examination of the tolerance factor (which measures the strength of the interrelationship between the regressor variables) does not suggest multicollinearity.18 Panel B of table 3 yields similar inferences with respect to the KL model. FSP(KL) is positively and significantly associated with all the independent variables. Even though the Pearson correlation between ISPNI(KL) and ADVSMN is high, the rank correlation is .13 suggesting the existence of some extreme observations. Further tests (reported later) suggest that there is no problem of multicollinearity. The next section reports the results of estimation of the different models. 5.4 Results Of Estimation Of Basic Model The results of estimating the basic model are provided in table 4. As expected, all the variables are positive and significant. Table 4 Results Of Estimation Of Basic Model Basic Model: FSP = a + p1ISP + p2RDSMN + f3ADVSMN + 64CEMSMN Parameter Standard Error T INTERCEPT .41 .08 5.075** ISPNI .35 .10 3.5* RDSMN 2.97 1.51 1.96* ADVSMN 2.93 1.42 2.06* CEMSMN 1.38 .47 2.9 ** significant at .01 level * significant at .05 level Test of the Goodness of Fit for the Model F Value: 7.531 P > jIF .0001 Adjusted R Square of the Model .1237 n = 186. 18 The details of the different tests to examine possible violations of the OLS model is provided in a later section. 57 The model is significant and the adjusted r square (R2) for the model is .1237 which seems reasonable considering that the model is cross sectional. The coefficients suggest that the higher the expenditure on R&D, advertising and capital items (property, plant and equipment), for a given level of sales, the higher is the firm specific persistence. 5.5 Robustness Checks For The Basic Model Since the statistical procedure used is ordinary least squares, conformity to assumptions of the linear regression procedure such as homoscedasticity and orthogonolity of the independent variables needs to be examined. According to White's (1980) test, there was no evidence of heteroscedascticity. Table 5 shows the relevant statistics for the heteroscedasticity and multicollinearity tests. Table 5 Robustness Tests For The Basic Model White's test to examine the presence of heteroscedasticity Chi Square Value 9.95 P Value .77 Null hypothesis of homoscedasticity not rejected. Test to detect the presence of multicollinearity Variable Tolerance ISP .96 RDSMN .97 ADVSMN .96 CEMSMN .96 The tolerance parameter is calculated as 1R2 that results from the regression of the other variables in the model on that regressor. If the variables are orthogonal to each other, Tolerance is 1 for all the variables. The inference is that there is no evidence of multicollinearity among the variables.19 "1As a further step to examine the problem of multicollinearity, the procedure suggested in Besley,Kuh and Welsh (1980) was employed. This procedure effectively examines if any particular variable contributes to the variance of two or more estimates. This procedure The collinearity diagnostics revealed no evidence of multicollinearity amongst the independent variables. The results in table 5 suggest that the assumptions of ordinary least squares are satisfied in the case of the basic model. The next step in checking the robustness of the results was to examine the presence of outliers which may be distorting the inference. This was done by examining the studentized residual, Cook's D, DFFITS an DFBETAS [See Besley, Kuh and Welsh (1980) for a detailed explanation of these statistics]. Each of these statistics examines the effect of each observation on the estimated parameters and the predicted value. Based on the examination of these parameters 13 outliers were removed and the basic model was reestimated.20 The results of the estimation are provided in Table 6. TABLE 6 Basic Model After Removing Outliers Model: FSP = a + fiISP + 02RDSMN + p3ADVSMN + 14CEMSMN Parameter Standard Error T INTERCEPT .429 .077 5.55** ISP .260 .082 3.16** RDSMN 4.522 1.35 3.34** ADVSMN 2.965 1.39 2.13** CEMSMN 1.03 0.64 1.61s ** significant at .01 level Test of Goodness of Fit: F value 8.039 $ significant at .052 level P value 0.0001 Adjusted R Square of the model .141 did not suggest the presence of multicollinearity. 20 The determination of outliers is necessarily somewhat arbitrary. When outliers are removed based on criteria such as the standardized residual and Cook's D and the model reestimated, the reestimated model would still reveal new outliers. The basic model was estimated iteratively and conservative cut off rules were used determine which observations should be discarded. 59 The results from the estimation of the basic model after removing the outliers are essentially the same as those reported in table 4. The R&D variable becomes more significant and the significance of capital expenditure declines. While the reason for this change is not entirely clear, all the variables continue to be significant at conventional levels.21 The removal of the outliers improves the fit for the model as measured by the adjusted R2 but not overwhelmingly. In summary, the results in tables 5 and 6 suggest that the assumptions of ordinary least squares seem to satisfied and outliers do not pose a problem for the results of the basic model. Two additional procedures were employed to test the robustness of the basic model's results. The first procedure was to test the stability of the model over a period shorter than the 36 year period used for the estimation of the basic model reported in table 4. To check for robustness, the basic model was estimated over a shorter period, after dropping the first 11 time series observations. Thus, all the estimates of persistence and the other firm specific variables were 21 The significance of the capital expenditure drops and is significant only .052 level. The sample selection procedure suggests one possible reason for this loss of significance. Table 1 suggests that more than 100 firms had to be dropped from the sample in order to ensure that information was available for the R&D and advertising variables. With respect to the capital expenditure variable, firms had been dropped for no systematic reason. Adding back the firms and reestimating the model with only capital expenditure suggests that capital expenditure is very significant. This is not reported here since the results in table 5 suggests that capital expenditure is quite close to significance at conventional levels. measured over 25 observations.22 The results of the estimation of this restricted model are provided in Table 7. TABLE 7 Results Of Estimation Of Restricted Model Restricted Model: FSP = a + f1ISP + z2RDSMN + 3sADVSMN + 04CEMSMN Parameter Standard Error T INTERCEPT .53 .096 5.54** ISP .37 .089 4.12** RDSMN 2.81 1.29 2.17** ADVSMN 1.98 1.3 1.52s CEMSMN 1.11 0.424 2.61** ** significant at .01 level $ significant at .06 level Test of Goodness of Fit for the Model F Value 7.542 P Value 0.0001 Adjusted R Square of the Model .1239 n = 186 The results of the restricted model are generally consistent with the results of the basic model.23 The advertising variable is not significant at the conventional level of 5%, but is significance at 6% level. It is important to realize, however, that estimating persistence with just 25 observations may not be entirely reliable. The mean of estimated persistence based on 36 observations was .81 whereas the mean 22 In the estimation of both firm specific persistence and industry specific persistence, the lag length used was 1/8th of the number of observations. This rule was preserved in this restricted model and persistence was measured using a lag length of 3 instead of 5 used in the estimation of the basic model. 23 All the diagnostic checks such as the conformity with assumptions of regression, effects of outliers etc were checked for this and the other models. The reported results give effect to the required correctional procedures. For example, if outliers are detected on examining the studentized residual, it is removed and results are reported after such removal. If homoscedasciticity is violated, a heteroscedastic consistent covariance matrix is estimated and test statistics are reestimated based on such matrix. based on 25 observations was 1.008 and a paired t test to compare the means of the two measures of persistence had a value of 9.69 (p value .0001). This test strongly rejects the null of equality of the two means. This is not very surprising since the negative autocorrelations at the higher lags have been ignored by restricting the measure of persistence to just 3 lags when it was based on only 25 observations. Thus, even though it is encouraging to note that the results of the basic model are robust when estimated over a shorter horizon, the results need to interpreted with caution. The second procedure to check the robustness of the results of the estimation of the basic model was to perform a regression on ranks. The diagnostic checks applied to the OLS regressions reported in table 5 suggested that a linear regression was appropriate. However, if the relationship between the independent variables and persistence was nonlinear in nature, then fitting a parametric linear regression may not be most appropriate. Puri and Sen (1985) suggest that a rank regression may be appropriate in such cases. In addition performing a regression on ranks requires no distributional assumptions with respect to the residuals or the independent variables. Table 8 reports the results of a regression on the ranks of the dependent and the independent variables. The results of the rank regression mirror the results of the parametric linear regression. Capital expenditure is significant at .08 level, but is not significant at the conventional levels. Table 8 Rank Regression Of The Basic Model Model: FSPRANK = a + 81ISPRANK + f2RDRANK + p3ADVRANK + f4CEMRANK Parameter Standard Error INTERCEPT 20.44 ISPRANK 0.367 RDRANK 0.1464 ADVRANK 0.148 CEMRANK .103 ** significant at .01 level * significant at .05 level $ significant at .054 level # significant at .08 level Test of goodness of fit for the model: Adjusted R square of the Model .183 12.68 1 0.07 5 0.069 2 0.0746 1 0.072 1 F Value P Value n = 173 10.641 .0001 In summary, both parametric and nonparametric tests of the basic model suggest that firm specific investment variables are significant in explaining the crosssectional variation of persistence. 5.6 Kormendi And Lipe Model As indicated in Chapter 4, KL suggest an alternative measure of persistence which measures the present value of revisions in expectations of the levels of future earnings to a dollar innovation in current earnings. Appendix 1 explains the technical details of this measure. The Kormendi and Lipe model uses this alternative measure of persistence as the dependent variable. The rate of discount used to discount the revisions in the expectations of future earnings is 10%.24 24The use of different discount rates alters the magnitude of persistence, but does not alter the empirical rank ordering of the firms with respect to persistence. The choice of 10% is essentially arbitrary and is chosen to make it consistent with the choice of KL. The rate of 10% was applied uniformly to all the firms. It is very likely that the appropriate rates of discount would be different for different firms and T . 612s .21 ** .092* .988* ..42 # Analogous to the Cochrane specification of persistence, the KL specification is devoid of the units of measurement. Consequently, the independent variables are the same ones used in the basic model and measure the standardized averages of real R&D, real advertising and real capital expenditure. Table 9 presents the results of the regression of the KL model. Table 9 Results Of KL Model KL Model: FSP(KL) = a + i1ISPNI(KL) + P2RDSMN + 93ADVSMN + f4CEMSMN Parameter Standard Error T INTERCEPT 6.58 .468 14.054 ISPNI(KL) 0.096 .0375 2.55** RDSMN 24.69 8.22 3.003** ADVSMN 11.67 8.62 1.35$ CEMSMN 7.008 2.54 2.763** ** implies significant at .01 level $ implies significant at .09 level Test of Goodness of Fit for the Model F Value 6.43 P Value 0.0001 Adjusted R square of the model .112 n=173 The results of this model are generally consistent with the results of basic model. The advertising variable is significant at .09 level, whereas all the other variables are significant at .01 level. The collinearity and heteroscedasticity diagnostics were examined and the residuals were plotted to check for normality. None of the assumptions of the classical linear regression model were violated in may undergo changes for the same firm over time. Since the choice of different rates for different firms would be arbitrary, this improvement is not explored at this stage. the estimation of the KL model. As indicated in footnote 25, the magnitude of the KL specification is dependant on the assumption of the interest rate. To enhance comparability with the Cochrane specification requires an estimation strategy of the KL model which is less influenced by the assumption about the interest rate. A rank regression of the KL model would accomplish this purpose. Table 10 presents the results of a rank regression of the KL model. Table 10 Rank Regression Of The KL Model Model: FSP(KLRANK) = a + PISP(KLRANK) + 02RDRANK + f3ADVRANK + 4CEMRANK Parameter Standard Error T INTERCEPT 33.782 14.245 2.371 ISP(KLRANK) 0.1926 0.074 2.603** RDRANK 0.1232 0.0748 1.647* ADVRANK 0.1186 0.077 1.539$ CEMRANK .1936 0.076 2.526** ** significant at .01 level * significant at .05 level $ significant at .07 level Test of Goodness of Fit for the Model F Value 3.949 P Value 0.0095 Adjusted R square of the model .08 n = 173 The results of the rank regression are consistent with the parametric estimation which suggests that the relationship posited in the KL model is not dependent on the assumption about the interest rate. Thus, the results reported in tables 9 and 10 together suggest that the results of estimating the basic model do not change when the basis of measurement of persistence is changed. Both the Cochrane and KL specifications provide information about the duration of the memory of a shock in a series. Tables 9 and 10 suggest that the operating decisions of the firm which determine such duration is not sensitive to how such duration is measured.25 This is encouraging as it conforms to the intuition that the economic decisions of a firm that determine the persistence of the resultant earnings series should not be sensitive to the technique employed to measure such persistence. 5.7 Effects Of Accounting Techniques In order to assess the effect of accounting techniques on estimated persistence, firm specific persistence was estimated on (1) sales (2) cash flows after adjusting for noncurrent accruals26 (3) cash flows after adjusting for current and noncurrent accruals.27 If the results from the estimation of these models were similar to those of the basic model (table 4), then it could be inferred that choice of accounting techniques does not have either an independent effect or an interactive effect. If the results are different, then additional structure (and/or assumptions) is required in the model is required to disentangle the two effects. Sales. Table 11 reports the results of estimation of the basic model using persistence of sales as the dependent variable. 25 This also suggests that the additional information provided in the Cochrane specification does not change the rank ordering of the firms on the persistence dimension as determined by the operating decisions of the firm. 26As explained later, two specifications of cash flow were used to assess that the results of estimation are robust to the measurement of cash flow 27 To estimate persistence of sales, 4 firms were omitted because of non availability of sales data. In the estimation of cash flow after adjustment of noncurrent and current accruals, 39 firms were omitted because of non availability of data about items such as current portion of long term debt, cash, current asset or current liability. Table 11 Model Estimation Based On Persistence Of Sales28 Model: FSP(S) = a + 01ISP + 02RDSMN + f3ADVSMN + 04CEMSMN Parameter Standard Error T INTERCEPT .84 .114 7.35** ISP .32 .142 2.24** RDSMN 4.697 2.14 2.459** ADVSMN 4.92 2.00 2.459** CEMSMN 0.66 0.67 0.983 ** significant at .01 level Test of Goodness of Fit for the Model F Value 4.795 P Value 0.0011 Adjusted R Square for the Model .08 N=182 The results reported in Table 11 are similar to the results reported for the basic model in table 9. All the variables are significant except capital expenditure. As suggested in footnote 21, one possible reason for this is the loss of firms due to missing data on R&D and advertising variables. This reduces the power of the test with respect to the capital expenditure variable. Although this test is not a conclusive test of the effects of the choice of accounting methods, it provides limited evidence that accounting methods are not important in explaining the variation in persistence. Cash Flows Adjusted For Current And Non Current Accruals fCF1. The procedure for estimating cash flows after adjustments for current and noncurrent accruals was rendered difficult as a result missing data for 28 This model was reestimated using ISP based on Sales instead of net income and the results were identical. Hence the results of this model are not reported here. 67 some items such as longterm debt included in current liabilities. This resulted in a loss of 39 firms and the estimation was based on only 147 firms. The results of the estimation of this model are provided in table 12.29 Table 12 Model Estimation Based On Cash Flow (Definition 1) Model: FSP(CF) = a + fIISP + f2RDSMN + f3ADVSMN + f4CEMSMN Parameter Standard Error T INTERCEPT .286 .046 6.22** ISP .07 .05 1.28s RDSMN 1.04 .878 1.19 ADVSMN 1.10 .78 1.4s CEMSMN 1.52 .378 4.009** ** significant at .01 level * significant at .10 level Test of goodness of fit for the model: F Value 6.411 P Value 0.0001 n = 147 Adjusted R Square for the model .13 The results of the estimation of this model are not consistent with the results of the basic model estimated earlier. Only the capital expenditure variable is significant at conventional levels. It is interesting to note that even the industry persistence variable is not significant at conventional levels, while advertising approaches significance in a direction opposite to the one expected. One possible reason is the reduction in the power of the test due to the reduction in the sample size. A more important reason could be that the cash flow 29 There were not enough observations to enable the calculation of industry persistence based on cash flows. Thus, industry persistence is based on net income as in the case of the basic model. 68 variable, after possibly removing the effect of differential accounting methods might introduce noise as a result of differential cash management techniques across firms. Thus, differences in credit and cash management policies could be reflected in this specification of cash flow variable. Such differences in cash management policies are not purported to be explained by the chosen firm specific variables. In order to investigate this possibility the analysis is repeated using a definition of cash flow based non current accruals. Cash Flow Variable Adjusted For Non Current Accruals [CFNC1. This variable is measured by adding back depreciation and amortization to the operating income variable. Since the desire is to examine the effects of differential accounting techniques in the operating income variable, the amount of depreciation and amortization are added back and the cash flow variable is defined as operating income before depreciation and amortization.30 Based on this measure of cash flow, the basic model was reestimated wherein the dependent variable is persistence based on this measure of cash flow. Table 13 provides the results of the estimation of this model.31 30 A direct way to get this variable to get the working capital from cash flow (Compustat item # 110). Information is not available for this variable for the early years. 31 The model estimated here uses a measure of industry persistence based on net income. This model was estimated again using a measure of industry persistence based on cash flows (after adjustment for non current accruals). The results were very similar. Table 13 Model Estimation Based On Cash Flow (Definition 2) Model: FSP(CFNC) = a + f1ISP + #2RDSMN + J3ADVSMN + f4CEMSMN Parameter Standard Error T INTERCEPT .41 .081 5.08** ISP .35 .1007 3.50** RDSMN 2.971 1.513 1.96* ADVSMN 2.94 1.422 2.065* CEMSMN 1.383 .476 2.91** *significant at .01 level significant at .05 level Test of Goodness of Fit for the Model F Value 7.531 P Value 0.0001 Adjusted R Square for the Model .1237 n = 186 The results of this model are very similar to the basic model reported in table 7. All the firm specific variables are significant and the adjusted R square is comparable to that of the basic model. In summary, the combination of evidence provided in tables 11 (sales) and 13 (cash flow adjusted for noncurrent accruals) suggests that differential choice of accounting techniques is not likely to have either an independent or interactive effect on estimated persistence for this sample. 5.8 Effects Of Differential Accounting For R&D Variable Firms could differ with respect to the treatment of items such as depreciation, inventory and R&D. The analysis presented in tables 11 through 13 examine the overall effects of cross sectional difference in the choice of accounting methods. One of the variables chosen for our analysis is R&D. Before 1974 firms had the option to either capitalize or expense R&D. The sample period spans a period of 36 years from 1950 through 1985. More than half of this period is before 1974. The decision to capitalize or expense has an effect on operating income and the persistence of operating income. Hence, the choice of accounting method for R&D (before 1974) could have an effect on estimated persistence. This poses an internal validity threat with respect to the inferences for firms which had capitalized R&D before 1974. In order to check this, data on the method of accounting were obtained for about 324 firms from the data base discussed in Shehata (1988).32 However, there is an overlap of only 40 firms between the sample of Shehata (1988) and the sample used in this study. Out of this small sample of firms there are only 2 firms which capitalized R&D expenditures. Thus, these 2 firms were the only ones which had different accounting methods for R&D before and after 1974. This sample is too small to detect any effects of the choice of accounting method for R&D.33 5.9 Generalization Of The Results Persistence, however measured, is independent of any sign or units of measurement. Since the firms in the sample could have either positive or negative income, the invariance of persistence to the sign 32 Shehata (1988) included all the NYSE and AMEX firms for which data were available during the sample period discussed in that study. 33 Nevertheless, the basic model was run without these two firms and produced no change. This is not surprising since two firms is too small a sample size for any type of statistical procedure to reveal any systematic pattern even if it exists. As an additional step the basic model was estimated with just these 40 firms and in this regression, choice of accounting method with respect to treatment of R&D expenditure did not seem to have an effect. of income gives rise to the following issues:34 (a) is the model generalizable across firms with both positive and negative innovations? (b) If the model is not generalizable, does the presence of negative observations in the sample affect the results in this study? The Issue. Persistence describes the duration of memory of the innovations in the series. In the extreme case of a random walk, the memory of the innovation is the highest and in the case of a stationary series, it is the lowest. Since the phenomenon described is the duration of the memory, it is clearly devoid of any sign. In other words, persistence describes the duration of the memory of an innovation irrespective of whether such innovation itself is positive or negative. In fact, in the extreme case where there are two series of realizations with the only difference between them being the sign on the realizations, the estimated persistence would be identical. Thus, for a series with the innovations 100,250,445... and another with the same numbers but opposite sign 100,250,445... persistence would be the same under both the KL and Cochrane specifications of persistence. In terms of the research question in this essay, this gives rise to an interesting issue as to investment behavior of firms wherein the realizations from investment (and consequently the innovations) are 34The dependent variable in the estimated models is the permanent component of the income series. The permanent component of the income series is related to the permanent component of the innovations as described in chapter 2. Irrespective of the nature of the underlying series, income or income innovations, persistence is independent of the sign on the observations of the series. primarily negative as opposed to a firm wherein the such realizations are primarily positive. Genaralizability of the model. The model proposed in this study suggests that persistence is positively associated with investment. In other words, firms which have high investment are likely to have high persistence, and vice versa. The model does not, however, suggest anything about the sign of the resulting realizations from such investment. It is possible that firms with either income consistently declining over time and/or primarily negative realizations would make adjustments in either the nature and/or magnitude of investments in order to reduce the persistence of losses.35 The model proposed in this essay is not a dynamic model which can characterize the changes in the investment program of a firm as a result of negative realizations from investments. The basic premise of the model is that investment is positively associated with persistence. That is, by committing to a high investment program, ceteris paribus, the management is committing itself into a high persistence profile in the income series. The dynamics of the adjustment in the investment profile arising from negative realizations is outside the scope of this model. The sample selection procedure used in this study diminishes the magnitude (and possibly the existence) of this potential concern. A firm was included in the sample only if it survived in the same industry for a minimum of 35 Such adjustments are likely to happen only if the losses from investments are considered permanent due to reasons such as permanent changes in demand. Temporary shifts in demand are likely to be handled by altering the variable factors of production such as labor or raw material. 73 35 years. Survivorship arguments suggest that a firm would not survive unless the realizations from the investments are nonnegative on average. Possible effects of negative realizations. Firms that have primarily negative income and/or declining income, could be making more frequent adjustments in its investment program in order to reduce the persistence of losses. Consequently, the variance of the investment, over time, of such firms is likely to be higher than firms which do not have negative or declining income.36 If data were available for a reasonable number of firms with such declining income and/or negative income such a hypothesis could be tested. However, as suggested by the data in table 14, there are very few firms which have either negative or declining income. This is not surprising given the sample selection criterion which required that the firm be in the same industry over the sample period of 35 years. Table 14 suggests that more than 90% of the firms in the sample had positive income for about 90% of sample period (32 years or more). A comparison of the real operating income and the absolute value of real operating income suggests that there is little difference between them, confirming the relatively small number of negative observations in the sample. A regression of real operating income on time suggests that only 20 firms had income declining over time (table 14). An limited examination of the behavior of variance of investments for such firms in this sample was conducted. As table 14 36 It is not clear that the mean of investment of such firms would be lower because the firms shift the production to other related products or shift investment to other related areas within the industry. suggests there are 20 firms with income declining over time and there are 26 firms with negative income during at least 10% of the sample period ( 4 years or more). Table 14 Descriptive Statistics On Firms With Negative Numbers Number of years of Number of firms Cumulative negative 01 Percentage 0 114 62 1 23 75 2 12 80 3 11 86 4 10 91 5 5 94 6 2 95 7 2 96 8 4 98 10 2 99.5 15 1 100 186 Number of firms for which average real operating income was negative 0 Mean Median (1) Real operating income (in million $) 2.94 .83 (2) Absolute value of real operating income (in million $) 3.06 .85 Pearson Spearman Correlation between (1) and (2) .98 .97 Model: Real operating income = a + fTime. Coefficient on 3 Number of firms Average T value Positive 132 8.83 Negative 20 3.59 Not Significant 34 There is an overlap of 14 firms between these two sets of firms resulting in 32 firms with negative/declining income. The variances of real capital expenditure (RCE), real advertising (RADV) and real R&D (RRD) of these 32 firms were compared with their respective counterparts of the rest of sample (154 firms).37 Table 15 reports the results of this analysis. A folded form of F statistic, F', was used to test the equality of the variances of these two groups.38 Table 15 Test Of Hypothesized Effects On Variance Of Firms With Negative/Declining Income RCE RRD RADV Variance for 32 firms 142.08 25.32 2.27 Variance for 154 firms 25.53 3.63 1.22 F' Statistic 5.57 6.98 1.86 The results in table 15 suggest that the variance of investment decisions of firms with negative and/or declining income was higher than the other firms and they were all statistically significant at .01 level. The inference to be drawn from this result is that there is larger variation in the amounts of investment in the firms which had not been as successful in obtaining positive realizations from investments 37 The decision to include only firms which had negative income during at least 10% of the sample period is essentially arbitrary. It is implicitly assumed that negative income during a shorter period, 3 years or less, may not trigger changes in investment behavior. Exclusion of other firms biases the test against finding an effect and is hence conservative. Since the choice of this 10% rule is arbitrary, the analysis in table 14 was repeated with just the 20 firms which had income declining over time and the results were unchanged. 38 Refer to Steel and Torie (1980) for details. A folded form of F statistic is expressed as F' = (larger of s12,s22)/(smaller of s12,s22) This is a two tailed F distribution since we do not specify which s2 is larger. The P value gives the probability of a greater F under the null of a12 = '22 76 compared to other firms which had been more successful.39 The necessary caveat, however, is that the sample size for this analysis is very small and inferences, therefore, may not be unambiguous. In summary, persistence, however measured, provides information about the duration of the memory of a shock without regard to whether such shock is positive or negative. Thus a firm with nothing but negative income would still have positive persistence. In terms of the research question in this essay, it gives rise to an interesting issue as to the possible differences in the investment behavior of firms which have negative and/or declining income. Such investigation is outside the scope of the model proposed in this essay. This is because the model is not dynamic and does not address the response to negative realizations from investment. However, it is conjectured that the variance of the investments would be higher in firms with such declining/negative income due to the adjustments that need to be made to reduce the persistence of losses. A subsample of 32 firms with such declining/negative income exhibits significantly higher variance when compared with the rest of the sample. 39 Comparison of the variances of the firms with negative numbers might be better when interindustry differences in variances are controlled. Thus it might be better to compare the variances of such firms with those of other firms in the same industry. This was difficult because the 32 firms were spread over 16 industries at the two digit SIC code level. Thus there were very few industries where such firms could be considered a distinct group within an industry. Only in two industries, [blast furnaces and steel (SIC code 33) and motor vehicles, parts and accessories (SIC code 37)] was this possible. In both these industries, the variances of the chosen firms were higher than the industry average. CHAPTER 6 SUMMARY AND CONCLUSIONS The evidence presented in this dissertation strongly suggests that firm specific investment variables, namely capital expenditure, R&D and advertising explain a significant amount of the cross sectional differences in estimated persistence after controlling for industry factors.1 This result is robust across different statistical procedures and different persistence measures. Choice of accounting techniques did not seem to play a major role in explaining cross sectional variation in estimated persistence. This inference arises because the significance of the chosen variables was unchanged when the analysis was repeated using two different instrumental variables. The contribution of the research in this dissertation is the explicit documentation of the association between economic attributes of the firm and persistence. This confirms that persistence is a meaningful construct. This in turn provides one underlying reason for the positive association between persistence and earnings response coefficient observed in prior studies (KL). 'Capital expenditure is not significant at conventional levels in certain analyses. As suggested in footnote 21 of chapter 5, one possible reason is the reduction in the number of firms due to non availability of observations on the other firm specific variables. The major limitation of the analysis presented here is that lack of a rigorous theory to guide the choice of the variables. This study's choice of variables and the reasoning behind the expected direction of the variables' effects on persistence is somewhat heuristic and arises from economic arguments rather than from a rigorous economic model. In the absence of a rigorous theory to guide the choice of variables and the expected direction of effects, the tests conducted herein are joint tests of the validity of the hypothesized effects and the validity of the estimates of variables used. Said differently, specification errors and estimation errors are indistinguishable in the absence of a theory giving rise to the joint nature of the tests. Future research should focus on trying to build a rigorous model to guide the choice of variables which determine the magnitude of the permanent component of earnings. While such a task may be quite formidable, it is important to consider other variables associated with the financial and operating decisions of the management which are likely to affect the magnitude of persistence. Other possible avenues of research involve the examination of the association of estimated earnings persistence with other characteristics of the firm such as production, employment and management compensation. The primary interest in the construct of persistence arises because of the observed valuation implication of persistence. The association between earnings persistence and firm specific returns could be reexamined to evaluate the sensitivity of the association to measurement errors in the estimation of both earnings response coefficients and persistence. APPENDIX 1 RELATIONSHIP BETWEEN KL AND COCHRANE SPECIFICATIONS OF PERSISTENCE This appendix explains the relationship between KL and Cochrane measures of persistence. Before explaining the relationship between the two measures, it is important to understand how these measures are calculated. Sections 1 and 2 of this appendix explain the KL specification and the Cochrane specification respectively. Section 3 examines the relationship between the two and section 4 provides some concluding remarks. Section 1: Kormendi and Lipe specification Kormendi and Lipe assume that earnings, Xt, can be reasonably well approximated by an AR(2) process in the first differences. Specifically, (13) AX, = b + cAX t1 + c2AX2 + et, where AXt is the first difference earnings at time t and u, is a white noise error term. Suppose a forecaster or the market agent knows all the parameters and knows all the u's and the X's through time tl, then his optimal forecast of X, at time t1 would be (14) Et_(AXt) b + clAX t1 + c AXt2. (15) AXt Et_1(AXt) = ut. Persistence can be measured by finding out by how much the forecaster would revise his forecast of earnings for a future period t+n, i.e., Xt+n, given the surprise information at time t, i.e., ut. To see this, it is convenient to rewrite equation 15 in a moving average form in the levels using standard causality conditions. Specifically, (16) Xt = 6(L)ut where 8(L)ut = ut + elUt_1 + 82t2 + 3Ut3 + ... c and ej is the jth moving average coefficient. To obtain the revisions in forecasts for all the future periods we would need (17) [Et(Xt+1) Et_(Xt+i)] + [Et(Xt+2) Et(X+2)] + ** It can be shown that (18) Et(Xt+) Et1(Xt+) = 8nut. Using equation 18, we can get an expression for equation 17. This is shown hereunder. [Et(Xt+1) Et_1(Xt+1)] + [Et(Xt+2) Et1(Xt+2)] + * = 81Ut + e2Ut + 83Ut + 84Ut = u, [8e + E2 + 83 + ...] (19) = u, [A(1)] where A(1) denotes the sum of the moving average coefficients e8 +82 + ... Equation 19 can be rewritten as follows: (20) A(l) =Zkek k = 1 to . Equation 20 provides the basis for KL's measure of persistence. They needed a measure of the present value of the revisions in the expected future earnings. Hence, their measure was the sum of the moving average coefficients appropriately discounted. Their measure was (21) Zk6 (k) k = 1 to c where S = l/(l+i) and i is the discount rate.1 In summary, the KL measure of persistence is the discounted sum of moving average coefficients arising out of an estimated timeseries expectation model. This implies that the KL specification compares the persistence of two series based on the discounted sum of the moving average coefficients and a series with a higher discounted sum is ranked higher in persistence. Assuming the same discount rate, however, it should be possible to compare two different series on the basis of the sum of the undiscounted moving average coefficients, i.e., A(1). The undiscounted sum, A(1), will be used compare the KL and Cochrane specifications in Section3. Section 2: Cochrane Specification Cochrane (1988) suggests measuring persistence via an estimate of the relative magnitude of the permanent component of a series. To understand this measure, we can write the earnings series, xt, in terms 1 KL's reported results are based on a 10% discount rate. Altering the discount rate affected the magnitude of persistence but did not alter the rankordering of the firms with respect to persistence both in their analysis and in the analysis done in this study. of permanent and stationary components. Specifically, (22) Xt = Zt + ct , where z, is the permanent (i.e., random walk or trend) component and ct is the temporary or stationary component.2 As equation 22 is nonstationary in the levels, assuming that first differencing makes it stationary it can be written in the form of a first differences as follows: (23) Axt = Azt + Act Cochrane suggests the calculation of the ratio a2Az/o2AX as a measure of persistence. This measures the ratio of variance of the permanent component to the variance of the first differences, and answers the question "what portion of the variance of the change in income is accounted for by the variance of the permanent component?" This gives a convenient measure of the relative strength of the permanent component of a series. If this ratio is high, the variance of the permanent component accounts for a high proportion of the variance of the difference series. This suggests that the series has a high permanent component, which implies that a change in income will have a large longrun or permanent effect. Estimation of Cochrane ratio. Since the random walk component is unobserved, estimation of this component becomes important. Cochrane 2 Writing the earnings equation in this form does not imply any restrictions on the series other than the series having to be stationary in the first differences. This is a standard assumption. Cochrane (1988) shows that all first difference stationary processes can be represented as a sum of stationary and random walk components. proves that the variance of the random walk component can be estimated for any series by calculating the limit of the variance of the kth differences. Cochrane shows that 1 (24) Lim Var(xt+k xt) = a2(AZ). k c k From equation 24, we can get an expression for the suggested variance ratio a2AZ/a2Ax. Rewriting a2(Ax) as Var(xt_ xt) we have a2Az 1 Var(xt+k xt) (25) = Lim Sk k Var(xt+i xt). Cochrane also shows that 1 (26) Lim Var(xt+k xt) = (1 + 2 Zj [(kj)/k j]} a2Ax , k k k where rj is the jth autocorrelation coefficient of the difference series xt. Equation 26 suggests that the variance of the permanent component of a series, a2(Az), is a function of a linear combination of the autocorrelation coefficients of the difference series. From Equations 25 and 26 we have a2AZ 1 Var(xt+k xt) (27) = Lim Sk k Var(xt+i xt) (28) = 1 + 2 Ej [(kj)/k Fj]. Cochrane suggests equation 27 as a measure of persistence. This ratio can be interpreted as an estimate of the magnitude of the permanent component of the series. Equation 28 tells us that the magnitude of the permanent component of the series can be estimated as a linear combination of the sample autocorrelation coefficients of the difference series. Since the variance ratio measures the magnitude of the permanent component, it is a natural measure of persistence as it estimates the long run effect of the shocks in the series (This was illustrated in chapter 2 of the dissertation). Section 3: Relationship Between The Two Measures: Recall that the KL measure is the discounted sum of the moving average coefficients, which, for convenience, can be referenced by A(1)3. As indicated earlier, (29) A(1) = Ekek k = 1 to Cochrane expresses the relationship between the variance ratio and the A(1) representation in the following equation: a2Az (8ek)2 (30) = k = 1 to . a2Ax (e82') Using rules of algebra, we can rewrite 8e 2k as follows: (31) ke82k = nVar(ek) + n(m )2 where 8m = mean of k4 Also (32) k, Ok = nem Substituting equations 31 and 32 in equation 30, we have 3 For the purposes of exposition, the KL measure is referenced by the A(1) notation which is the sum of the undiscounted moving average coefficients. Discounting does not change the inferences in any substantial manner. 4 The use of "n" arises from the assumption that the moving average coefficients are finite. This is a reasonable assumption, since for most stationary series the moving average coefficients tend to zero after a reasonable number. (zek )2 (ne)2 (33) = (X82k,) nVar(8k) + n(B.)2 Equation 32 is the undiscounted sum of moving average (MA) coefficients, A(1), and forms the basis for the KL measure of persistence which is the discounted sum of MA coefficients. Equation 33 is the variance ratio. The difference between the two is that the variance ratio standardizes the KL type measure by the variance of the MA coefficients 8k. Thus, the KL specification captures only one of the moments of the MA coefficients, the mean, whereas the Cochrane specification incorporates the second moment as well, namely the variance of the MA coefficients.5 Section 4: Conclusion In summary, both the specifications provide measures of the long run effect of a shock to a series. The difference between the measures lies in the fact that the Cochrane specification incorporates the second moment of the distribution of MA coefficients, namely the variance whereas the KL specification is restricted to the first moment the mean. The variance ratio can be estimated as a linear combination of the sample autocorrelations of the difference values of the observed series; it does not require estimation of the unobserved innovation. This reduces possible estimation error arising out of the assumption that a particular time series model describes annual earnings. As 5 Under conditions of uncertainty and heterogeneity of liquidity preferences and endowments on the part of market agents, it may be argued that more information about the distribution of the earnings series, i.e., more information about the moving average coefficients is desirable. 86 indicated in chapter 2 of the dissertation, these errors relate to (1) the use of a parsimonious model to capture expectations, and (2) the use of univariate models to measure innovation giving rise to the possibility of correlated omitted variables. APPENDIX 2 ASSOCIATION OF ISP WITH ECONOMIC VARIABLES This appendix explains the association between ISP, the variable chosen to surrogate the industry level determinants of earnings persistence, and certain economic factors which have been identified as determinants of industrylevel persistence in the industrial organization literature. Section 1: Introduction As suggested in chapter 4, industry level persistence (ISP) was chosen to surrogate industry level determinants because of the difficulty in obtaining data for all the industrylevel determinants over the entire sample period. The purpose of the analysis presented in the appendix is to validate that the chosen surrogate is appropriate by examining the association between ISP and certain major industrylevel determinants. The determinants of industry persistence chosen for examination are the same as chosen by Lev (1983), namely barriers to entry and durability of the product. Lev examined the association between the first and the second order autocorrelation of earnings and factors such as barriers to entry and durability of the product, firm size etc. Factors such as barriers to entry and durability of the product influence the correlation structure of the earnings (and hence, the persistence) through industry level persistence. The reason for this is that such industry characteristics are generally out of the control of the firm. The next section provides the conceptual link between the chosen factors and industry persistence. Section 3 provides the details regarding the data and also provides the results of estimation of the different models. Section 3 also provides some concluding comments. Section 2: Conceptual Link The chosen factors are barriers to entry and durability of the product. The association between these factors and industry level persistence is explained in this section. Barriers to entry. The higher the barriers to entry in an industry, the easier it is for the incumbent firms in the industry to preserve an earnings increase for a longer period of time. Hence, the higher the barriers to entry, higher the persistence of industry earnings. This suggests that the expected sign for the coefficient on barriers to entry is positive. Durability of the product. Friedman's 'permanent income theory' implies that the consumption of nondurables and services is a function of permanent income, whereas the consumption of durables is a function of transitory income. This suggests that the consumption series for nondurables and services would be more durable over time when compared to the consumption series of durables.1 This, in turn, suggests that the income of industries producing durable commodities would be more 1 This is confirmed by the fact that the prediction errors for the consumption of durable commodities are much larger for durables. [See Zarnowitz (1972)]. volatile when compared to the income of industries engaged in nondurables and services. Thus, a random walk series is likely to describe the earnings process of a durable industry better than any other model. It has been shown in Chapter 2 that a random walk model is the one with the highest magnitude of persistence. Persistence of earnings is expected to increase with durability. Thus the expected association between persistence and durability of the product is positive. Based on the arguments stated above, industry level persistence (ISP) can be viewed as a function of two factors: barriers to entry and durability of the product. The empirical model to be estimated may be states as follows: ISP, = a + fi Barriers to entry + P2 Durability of product + ei Section 3: Data. Measurement Of Variables And Results. Data and measurement of variables: Data on durability of products of different industries were obtained from the Survey Of Current Business published by the US Department Of Commerce.2 Industries with durable products were coded 1 and the other industries coded 0. Industry is defined at the 4digit SIC code level for the purpose of analysis in the Lev model. Data on the barriers to entry were obtained for the industries in the sample from a study by Palmer (1973). Palmer's classification is based on the four digit industry 2The Survey of current business is published monthly by the Department of Commerce. The monthly issue of January 1985 was used to measure of the durability of the product of the industry. Even though ISP is measured over 36 years, the independent variable is chosen at a point in time. This is a less of a concern because it is unlikely that the industries would be classified differently over time. code and is compiled from four sources: Bain (1965), Mann(1970), Shepherd (1970) and a 1967 Federal Trade Commission study.3 Firms with high barriers to entry were coded as one and those with low barriers to entry were coded 0.4 Persistence was measured using the Cochrane specification. The results of estimation are provided in table 16. Table 16 Panel A Association Of ISP (net income based) With Economic Variables MODEL: ISPNI = a + PfBarriers to Entry + )2Durability of Product Parameter Standard Error T INTERCEPT .47 .055 8.48*** HBE .094 .059 1.60s DUR .104 .046 2.26** Adjusted R Square of the Model .04 Panel B Association of ISP (operating income based) With Economic Variables MODEL: ISPOI = a + fiBarriers to Entry + i2Durability of Product Parameter Standard Error T INTERCEPT .55 .05 9.48** HBE .102 .059 1.739s DUR .166 .046 3.59** *** significant at .01 level Legend: HBE High barriers to entry ** significant at .05 level DUR Durability of product. $ significant at .06 level. Adjusted R Square of model .08. 3 While barriers to entry may change over time, Palmer's classification was used for three reasons: (1) it is a combination of the information from four different sources (2) the choice is consistent with a previous study and hence useful to make comparisons and (3) the period of Palmer's study coincides approximately with the middle of the sample period in this study (19501985). 4 Palmer's study has three levels of barriers to entry high, substantial and low. For the purpose of this analysis, industries classified as having substantial barriers to entry were classified as high. 91 Panel A of table 16 provides the results of the estimation of the model using the Cochrane specification on aggregate industry income based on net income. In Panel B, persistence is based on operating income. The tables in both panel A and B suggest that the chosen independent variables are significantly and positively associated with persistence as expected. In both the models, the adjusted R square is small suggesting that there is substantial amount of unexplained variation in estimated persistence measure. However, it is encouraging to note that the effects of two major economic determinants of industry persistence are reflected in the estimated measure of ISP. In summary, industry persistence is a function of several factors, chief among which are barriers to entry and durability. The variable chosen to surrogate all the industry factors in this dissertation is industry specific persistence. The results presented in this appendix validate the choice of this surrogate. It is shown that industry specific persistence is associated with these key factors in the expected direction in a significant way. All the variability in industry specific persistence is not explained by these two factors (the adjusted R square is low). This could be due to reasons: (1) noise in the measurement of ISP due to the aggregation process and more likely (2) omitted variables in the model. REFERENCES Bain, J., 1965, Barriers to the New Competition, Harvard University Press, Cambridge. Ball, R. and P. Brown, 1968, An empirical evaluation of accounting income numbers, Journal of Accounting Research 6, 159178. Ball, R. and G. Foster, 1982, Corporate Financial Reporting: A methodological review of empirical research, Journal of Accounting Research 20 (Supplement), 161234. Beaver, W.H., 1968, The information content of annual earnings announcements, Journal of Accounting Research 6 (Supplement), 6792. Beaver, W, R. Clarke and W. Wright, 1979, The association between unsystematic security returns and the magnitude of earnings forecast errors, Journal of Accounting Research 17, 316340. Besley, D.A., Kuh,E. and Welsh,R.E, 1980, Regression Diagnostics, John Wiley & Sons, New York. Branch, S. Ben, 1974, Research and Development Activity and Profitability: A Distributed Lag Analysis, Journal of Political Economy 82, 9991011. Brown, P., G. Foster and E. Noreen, 1985, Security Analyst Multiyear Earnings Forecasts and the Capital Market, Accounting Research Study Series #21, American Accounting Association, Sarasota, Florida. Brown, P. and J.W. Kennelly, 1972, The information content of quarterly earnings: An extension and some further evidence, Journal of Business 45, July 1972, 40315. Brozen, Yale, 1971, Bain's concentration and rates of return revisited, Journal of Law and Economics 14, 351369. Burgstahler, D., 1981, Crosssectional and crosstemporal differences in the relationship between annual earnings and security returns, Unpublished dissertation, University of Iowa, Iowa City, Iowa. Campbell, J. and G. Mankiw, 1987, Permanent and transitory components in macroeconomic fluctuations, American Economic Review 77, 111117. 