Econometrics of the Basu Asymmetric Timeliness Coefficient and Accounting Conservatism

Authors


  • Accepted by Philip Berger. We acknowledge the helpful comments of Sudipta Basu, John Core, Andrei Kovrijnykh, Jerry Zimmerman, two anonymous referees, and workshop participants at the University of Chicago.

ABSTRACT

A substantial literature investigates conditional conservatism, defined as asymmetric accounting recognition of economic shocks (“news”), and how it depends on various market, political, and institutional variables. Studies typically assume the Basu [1997] asymmetric timeliness coefficient (the incremental slope on negative returns in a piecewise-linear regression of accounting income on stock returns) is a valid conditional conservatism measure. We analyze the measure's validity, in the context of a model with accounting income incorporating different types of information with different lags, and with noise. We demonstrate that the asymmetric timeliness coefficient varies with firm characteristics affecting their information environments, such as the length of the firm's operating and investment cycles, and its degree of diversification. We particularly examine one characteristic, the extent to which “unbooked” information (such as revised expectations about rents and growth options) is independent of other information, and discuss the conditions under which a proxy for this characteristic is the market-to-book ratio. We also conclude that much criticism of the Basu regression misconstrues researchers’ objectives.

1. Introduction

Conservatism has been a central accounting principle for centuries (Watts and Zimmerman [1986], Basu [1997], Watts [2003a]). Basu ([1997, p. 7] defines conservatism as “the accountant's tendency to require a higher degree of verification to recognize good news as gains than to recognize bad news as losses,” a definition that is consistent with the adage “anticipate no profits but anticipate all losses.” In an efficient market, stock return reflects all new public information, and thus is a valid proxy for economic shocks to value. Then, in a piecewise-linear regression of accounting income on fiscal-period stock return, the incremental coefficient on negative return (the proxy for negative shocks, or “bad news”) is assumed to be a valid measure of asymmetrically timely loss recognition.1 Basu predicts and finds that the incremental coefficient indeed is positive, indicating timelier incorporation of negative economic shocks than positive shocks.2

Under this definition of conservatism, how accounting income incorporates shocks to firm value depends on their sign, so Ball and Shivakumar [2005] and Beaver and Ryan [2005] term it conditional conservatism. This contrasts with defining conservatism as reporting unconditionally low earnings or book value of equity. The key difference between these concepts is that conditional conservatism carries new information.3 Basu's formulation of conservatism in this fashion was an important breakthrough in our understanding of financial reporting rules and practices.

In a comparatively short period of time, the Basu [1997] piecewise-linear regression of accounting income on stock return has become one of the principal models of the financial accounting literature. The range and importance of these applications is testimony to the pervasiveness of conservatism as a property of financial reporting, and also to researchers’ confidence in the validity of their estimates of it.4 Furthermore, these studies regularly report differences in conditional conservatism that are consistent with plausible hypotheses, which provides added confidence in the validity of the estimates.

Nevertheless, researchers have assumed that the Basu regression produces valid measures of conditional conservatism due largely to its intuitive appeal, without rigorous analysis.5 The need for formal econometric analysis is heightened by claims that the Basu asymmetric timeliness coefficient is not a valid measure of conservatism (Dietrich, Muller, and Riedl [2007]); that it is unduly affected by variables, such as the book-to-market ratio (Pae, Thornton, and Welker [2005], Givoly, Hayn, and Natarajan [2007], Roychowdhury and Watts, [2007]); and that the Basu coefficients should be avoided by researchers (Patatoukas and Thomas [2011]).

We build our analysis on a model of the relation between accounting income and economic income that captures the salient properties of income recognition as it is practiced. We then use the model to derive and analyze the Basu earnings asymmetric timeliness coefficient. The model addresses shocks to firm value that arise from different types of information, and, consequently, our analysis is conducted entirely in the context of “news” and its incorporation in earnings. In particular, we ignore variation in expected returns across both time and firms. This framework allows a formal derivation of the Basu measure of conservatism, and also improves our understanding of how that measure interacts with characteristics of a firm's information environment.

The model distinguishes four components of information that financial reporting rules and practices cause to be incorporated in accounting income at different points in time. One information component is incorporated contemporaneously. The second information component is incorporated contemporaneously or with a lag, depending on its sign or magnitude. This earnings component is the source of conditional conservatism, which arises because negative news about future cash flows is subject to a lower accounting verification threshold than positive news. The third information component in our model always is incorporated with a lag, such as news about “unbooked” rents or growth options. The first three information components are assumed to be reflected in security prices, even when they are not recognized in current earnings. The model also has accounting income incorporating “noise,” for example, due to random errors in counting inventory or in valuing accounts receivable, that reverses over time. We believe this model incorporates the salient properties of accounting recognition rules and practices.

The primary result in this paper is that the Basu regression provides econometrically valid estimates of conditional conservatism. In particular, we show that, holding other things constant, the Basu regression identifies conditional conservatism only when it exists. The model also allows us to pursue the secondary goal of this paper, which is to show how conditional conservatism is a function of the relative importance of the various information components. Consequently, we are able to formalize and extend the results in Roychowdhury and Watts [2007] on the relation between conditional conservatism and market-to-book ratios. We also address the Dietrich, Muller, and Riedl [2007] claim that return endogeneity and sample truncation lead to biased Basu regression estimates.

Section 'A Framework for Interpreting Regressions of Earnings on Returns' outlines the model of the income–return relation that incorporates salient properties of accounting income. Section 'Econometrics of the Basu Regression' then proceeds with a formal analysis of the Basu regression in the context of that model. Section 'Conditional Conservatism, Firm Characteristics, and Empirical Implications' analyzes the relation between timeliness and firm characteristics, the role of the market-to-book ratio, and other empirical implications. Section 'Return Endogeneity and Sample Truncation' addresses return endogeneity and sample truncation. A short summary and our conclusions appear in section 'Summary and Conclusions'.

2. A Framework for Interpreting Regressions of Earnings on Returns

We develop a model of the relation between accounting and economic incomes that incorporates realistic and intuitive assumptions about how accounting rules and practices incorporate different types of information about economic value into accounting income, depending on properties of the information. In subsequent sections, we use this model as a framework to analyze the validity of Basu regression measures of asymmetrically timely recognition of economic gains and losses.

2.1 A Model of Accounting Income Recognition

The origins of the following model can be traced back to Ball and Brown [1968], Beaver, Lambert, and Morse [1980], Fama [1990], Kothari and Sloan [1992], Basu [1997], Kothari [2001], Beaver and Ryan [2005], Roychowdhury and Watts [2007], and others. We begin by assuming that capital markets are informationally efficient.6 We focus on shocks to firm value, ignoring variation in expected returns, with the implicit assumption being that variation in expected returns either does not occur or is controlled for as discussed in Ball, Kothari, and Nikolaev [2013].

While informationally efficient returns reflect all publicly available information in a timely fashion, we assume that accounting rules and practices emphasize verifiability, objectivity, and conservatism, as reflected, for example, in the historical cost principle and revenue recognition rules, and hence accounting income in any period incorporates some but not all the information that becomes publicly available during the period. The consequence is that accounting income incorporates some information with a lag: that is, “prices lead earnings.”7

Information in our model can be made public via earnings or any other information channel. We do not implicitly or explicitly assume that accounting systems and earnings announcements have no role in price formation. We assume only that, to the extent the market responds to reported earnings, it understands its components. In particular, the market distinguishes the components that reflect information made public in prior periods from those that reflect current-period information.

In line with the above assumptions, the total revision in security price (i.e., stock return) comprises three components, which are incorporated in accounting income differently:

display math(1)
display math(2)

where the subscripts t and t – 1 refer to time periods, and

  • Rt = total unexpected security return;8
  • xt = portion of the total unexpected return Rt that invariably is contemporaneously captured in accounting income, It;
  • yt = portion of the total unexpected return Rt that is not contemporaneously captured in It unless required by conservative accounting;
  • gt = portion of the total unexpected return Rt that never is contemporaneously captured in It, but always is incorporated with a lag;
  • It = accounting income;
  • wt = an indicator variable that takes the value of one when conservative accounting rules and practices lead to recognition of y in the current period; and
  • ɛt = “noise” in accounting earnings that reverses in the next period.

The unexpected return components xt, yt, and gt are stationary and time-independent random variables (Bachelier [1900], Samuelson [1965], Fama [1970], Campbell, Lo, and MacKinlay [1997]) whose variances are denoted by σx2, σy2, and σg2, respectively, and:

display math(3)

Accounting-induced noise ɛt has variance σɛ2 and is assumed independent of return components.

To keep the analysis tractable, we make the following linearity assumption. We assume that the information components x, y, and g can be expressed as linear functions of each other (e.g., math formula where f(., .) is linear and e is an uncorrelated residual), which we operationalize by assuming that the variance–covariance matrix of x, y, and g is independent of the sign of news. This implies that math formula; or more generally, math formula and math formula, so that math formula.9 This assumption allows us to focus on the effect of conservatism and ignore other potential sources of nonlinearity in the returns–earnings relation, such as an arbitrary nonlinear relation between the information components x and y or between x and g.10 To the extent that such nonlinearities are present, they comprise a confounding factor (akin to an omitted variables problem) that needs to be understood and controlled for when estimating conditional conservatism.

In practice, the distribution of returns is skewed, which often indicates nonlinearity in the data. Conditional conservatism implies a particular source of nonlinearity and hence calls for a particular functional form, which the Basu [1997] regression accommodates. If other sources of nonlinearity are present, the specification can be augmented accordingly to measure the separate effect of conservatism. As in all empirical work, it is important to use a specification that controls for the confounding effects of any economic forces other than conservatism when parsing out its effect on the returns–earnings relation. This issue is addressed in more detail in Ball, Kothari, and Nikolaev [2013]. In this paper, we ignore skew in returns in order to focus on the effects of accounting rules and practices per se.

2.2 An Interpretation of The Model in Terms of Accounting Practices

In this model, financial reporting rules and practices lead to accounting income always contemporaneously incorporating the information component xt, regardless of its sign or magnitude (i.e., without conditional conservatism). The intuition is that this component represents the least costly source of information to verify, and thus it invariably is recognized in the same period as it affects returns. This type of information could include current-period news about current-period cash flow, such as learning the actual cash realizations of current-period revenues and expenses in comparison with their expectations. It also could include news about future cash flows that is verifiable at low cost, and that is incorporated symmetrically in accounting income via working capital accruals. Symmetrically, low-cost verification typically applies to current operating-cycle information, such as cash receipts and payments, and the quantities of accounts receivable, accounts payable, and inventories. For example, accruals are used to adjust current-period cash flow for both increases and decreases in the quantity of closing inventory relative to opening inventory, because they are approximately equally low in cost to verify. Similarly, cash collections from customers are adjusted for both increases and decreases in accounts receivable. These working capital accruals incorporate into accounting income current information about future cash flows. For example, other things equal, an increase in closing inventory is information that less cash will be spent on purchasing inventory in future periods, and it is verifiable at low cost regardless of its sign. In some limited circumstances, symmetrically low-cost verification can apply to long-cycle information as well, an example being index funds, in which gains and losses on long-term investments in traded stocks are symmetrically low cost to verify and, in practice, are both accounted on a daily basis.

The second component of stock return, yt, is incorporated in accounting income either contemporaneously or with a lag, depending on the accounting operator wt. The intuition here is that yt represents information that is costly to verify, and more cost is incurred in verifying negative than positive news. Consequently, the verification threshold is lower for negative news (Basu [1997], p. 4), and this information component is incorporated asymmetrically. Such information could include the current-period revision in the expectation of unrealized future-period cash flows from “booked” assets, including long-term assets in place and purchased intangible assets, such as patents and goodwill. It could include information about the marketability of current assets, such as inventory. This component also could include news about the present value of future-period cash outflows, such as lawsuit settlements. Using accrual accounting to bring forward shocks to expected future cash flows is known as timely gain and loss recognition. Conditional conservatism implies that wt is more likely to be triggered by bad news (adverse shocks to expected future cash flows) than good news, and hence that loss recognition generally is timelier than gain recognition.

In the event that timely recognition is not triggered and thus the return component yt is not contemporaneously incorporated in income, it is incorporated with a lag, the intuition being that some revisions in expectations of future cash flows are not reflected in accounting income until the actual cash flow realizations occur. Conditional conservatism implies that incorporation with a lag is more likely for good news than bad. The effect of asymmetrically delayed incorporation of information is that, in a two-period model, the total effect on current-period income is wt yt + (1 – wt–1)yt–1.

In practice, the extent of the timely recognition asymmetry is determined by a number of factors. An extensive literature studies determinants that include economic incentives, debt and compensation contracting, governance, GAAP, regulation, and taxes. Since our objective is not to provide an equilibrium model of the extent of conditional conservatism, but rather to investigate the validity of the Basu [1997] measure used in this literature, our model takes the extent of the asymmetry as an exogenous given, and studies the properties of its measurement. In our setting, timely recognition is triggered when yt is below an exogenous threshold c.11

The third component of stock return, gt–1, invariably is incorporated in accounting income with a lag. One source of this return component is revisions in the value of growth options or unbooked intangibles (Beaver and Ryan [2005]). This information component is similar in nature to “rents” in Roychowdhury and Watts [2007]. Because growth options by definition are not “booked” as assets on balance sheets, these shocks to firm value are symmetrically incorporated in earnings with a lag, when the associated cash flows are (or are not) realized.

The model also has accounting income incorporating uncorrelated “noise” that reverses over time. An important source of this earnings component would be imperfect accounting accruals. For example, miscounting inventory affects current and future earnings with opposing signs. Other examples are errors in estimating uncollectible accounts receivable, errors in forecasting deferred tax liabilities, the effects of using historical cost interest rates on debt, and errors in estimating assets’ useful lives. Because accounting errors reverse over time, in our two-period model the error term is reversed out in the following period. For tractability we assume accounting error is uncorrelated with return components, and thus we ignore nonrandom “earnings management,” including “smoothing,” that plausibly is negatively correlated with returns.12

In this formulation, yt–1 and gt–1 are the two sources of delayed recognition. They generate the anticipated or stale component of earnings whose value consequences were previously reflected in returns. They therefore are uncorrelated with current-period return Rt, which is influenced only by information arriving contemporaneously. We view this lagged recognition of some components of stock return as a natural feature of the income recognition process in accounting, which can be measured and investigated (e.g., as a function of firm characteristics, managers’ incentives, or countries’ economic and political institutions).

While yt–1 and gt–1 are assumed to be uncorrelated with xt, yt, and gt, and, therefore, with Rt, there are economic reasons to expect a positive correlation among xt, yt, and gt. Recall that all three components are a consequence of economic news affecting investors’ cash flow expectations. However, only the news generating the stock price growth rate component xt, and possibly also yt (i.e., when wt = 1), is incorporated in accounting income contemporaneously, whereas news generating the return component gt finds its way into accounting income in the following period.

The earnings process modeled in equation (2) is intended to capture the important properties of asymmetric accounting recognition rules and practices, and their effect on how new information is incorporated in reported earnings. We next use this framework to investigate the properties of the Basu measure and develop empirical implications.

3. Econometrics of the Basu Regression

3.1 Representing Timeliness as The Conditional Expectation of Earnings

The Basu [1997] model is used to measure the (asymmetric) timeliness of earnings with respect to economic news, where timeliness is the extent to which a dollar of unexpected economic income is reflected, on average, in accounting income over a certain period. Over a firm's life (or a sufficiently long period), and assuming “clean surplus” accounting, each dollar of stock return translates into a dollar of earnings. Over shorter time periods (e.g., years), however, there is no one-to-one mapping between earnings and returns and thus they can be thought of (and modeled) as random variables sharing a joint probability distribution math formula. Statistically, earnings timeliness is thus the expectation of earnings conditional on return news math formula, or math formula. Note that this conditional expectation is not a point estimate but is a function math formulaof math formula. The conservative nature of accounting suggests the shape of this function is nonlinear, which the Basu [1997] model accommodates by specifying a pricewise linear regression of earnings math formula on returns math formula.

To see if the least-squares regression of earning on returns achieves the objective of measuring earnings timeliness, consider the following (general) model specification:

display math(4)

where μ(.) is an arbitrary function, and where math formula and math formula are observed contemporaneously. A well-known result is that:

display math(5)

It follows directly from result (5) that, for a correctly specified functional form, the least-squares estimation of (4) yields an unbiased conditional expectation. This is true irrespective of the distributional assumptions made with respect to math formula and math formula. Thus, a regression of earnings on returns fulfills its research objective of representing timeliness.

3.2 Simple Case With no Accounting Asymmetry

We now use the framework from section 'A Model of Accounting Income Recognition' to examine a baseline case with no accounting asymmetry (i.e., wt = 0), and hence where financial reporting incorporates all news of type yt with a lag regardless of its sign, in the same fashion as gt. Consider the econometric properties of regressing accounting income on security return:

display math(6)

In this case, the slope coefficient reflects the extent to which accounting income contemporaneously captures “good” and “bad” news equally. Let math formula denote the OLS estimate, such that:

display math(7)

The lagged income components, yt–1 and gt–1, do not correlate with xt, yt, and gt, so in the symmetric accounting case:

display math(8)

In equation (8), the estimated slope coefficient from a regression of earnings on contemporaneous returns increases in the timeliness of earnings, var(xt)/var(yt + gt), which is the ratio of the information in returns that accounting income incorporates in a timely fashion to the information it incorporates with a lag. Equivalently, as timeliness increases the contribution to current accounting income It of stale news, yt–1 and gt–1, is relatively lower.

In one extreme, accounting income is 100% timely with respect to economic income if var(yt) and var(gt) both are zero, in which case math formula converges to one and accounting income perfectly correlates with stock returns. This extreme is approximated by unlevered firms investing in actively traded securities that are marked to market, such as mutual funds. In the other extreme, accounting income is completely untimely for firms consisting entirely of growth options, as approximated by start-up firms investing almost entirely in intellectual property. In this case, var(xt) and var(yt) both are zero, math formula converges to zero, and accounting income is uncorrelated with the stock return.

3.3 Asymmetric Timeliness

We now consider the more general case with asymmetric timeliness, where the function math formula is expected to be nonlinear due to accounting conservatism. We demonstrate that conservatism implies that a piecewise linear regression model following Basu [1997] will correctly identify the asymmetry in earnings timeliness. In terms of our model (2), we now consider a case where financial reporting incorporates the information component y without a lag whenever it is below a threshold (i.e., wt > 0), but otherwise with a lag.

The Basu [1997] asymmetric timeliness coefficient is estimated from the regression model:

display math(9)

where Dt = 0 if Rt ≥ 0, Dt = 1 if Rt < 0, and β2 is the asymmetric timeliness coefficient. β2 then is the incremental coefficient on negative return (the proxy for negative economic income), and is predicted to be positive because conditionally conservative accounting incorporates negative economic income into accounting income sooner than it incorporates positive economic income.

Let math formula denote the OLS estimate of β2 from equation (9). Then we have:

display math(10)

where math formula and math formula

Timely loss recognition in accounting is not triggered by all types of economic loss. Our model (2) represents this by assuming symmetrically timely recognition for the verifiable information component xt, asymmetrically timely recognition of the information component yt that is triggered only when it falls below a threshold value such as zero, and symmetrically untimely recognition of the information component gt. These individual components are unobservable to the researcher, so the Basu regression model (9) conditions the regression slope on total return Rt, which is observable. Because total return incorporates information components that are not recognized asymmetrically, there is a question as to whether the Basu regression provides a valid measure of conditional conservatism. We now show that the incremental regression coefficient β2 = 0 if earnings are not conditionally conservative, and β2> 0 if earnings are conditionally conservative.

Our analysis incorporates what we believe to be a realistic scenario, that xt, yt, and gt are positively correlated. The algebra under that assumption is tedious, so we relegate derivations to the appendix.

3.3.1. Basu Regression Coefficients with Conditionally Conservative Accounting

We next derive the estimate for the β2 coefficient under conditionally conservative accounting, and then show that it behaves in a predictable fashion as a function of the firm's information environment.

We begin by considering a simple case in which the information component g is zero (this is relaxed in section 'Conditional Conservatism, Firm Characteristics, and Empirical Implications'). Then:

display math(11)
display math(12)

Equation (12) follows from the assumption that all shocks to stock returns (and thus total returns) are independent over time. Similarly:

display math(13)

The assumption that the variance–covariance matrix of the three information components in returns is independent of the sign of economic news implies that var(Rt | Rt ≥ 0) = var(Rt | Rt < 0), and we have:

display math(14)

Henceforth, we drop the subscript t, and all variables occur at time t unless explicitly subscripted t – 1.

We model conditional conservatism as recognition of y in the current period only when it is sufficiently “bad news.” We assume math formula if math formula and zero otherwise, where c is a threshold below which current-period recognition occurs. In a limiting case where math formula, timely recognition of y always occurs, there is symmetric recognition of good and bad news, and, thus, math formula = 0. Symmetry also exists in the limiting case of math formula, where timely recognition of y never occurs. The setup then degenerates to w = 0, as analyzed in section 'Simple Case With no Accounting Asymmetry'.

We set the threshold c to zero in line with the adage “anticipate no profits, but anticipate all losses” and to reconcile with the empirical formulation of the Basu [1997] regression. In practice, c could be expected to be negative and close to zero. To the best of our knowledge, this assumption is not restrictive and does not bias the results.

To show that math formula is positive under these assumptions, it suffices to show that:

display math(15)

Equation (15) can be expressed as:

math image(16)

where φ(., .) is a bivariate probability density function. This equation results from expanding equation (15) to separate its x and y components and expressing it in the form of expectations. The last term follows conveniently from the linearity assumption that the variance–covariance matrix of x and y is independent of the news sign, which implies that math formula = math formula and that math formula.

To explicitly evaluate the integrals, we further introduce the assumption that φ(., .) follows normality. Under this assumption we have (see appendix for derivation13):

display math(17)

where math formula

It further can be shown that for any positive α, the expression for Δ is positive (one way to see this is to note that math formula and that the first derivative of Δ is positive over the interval (0, +∞)).14 This result implies that the Basu incremental coefficient on negative returns reflects the existence of conditional conservatism in accounting income.

3.3.2. Basu Regression Coefficients in the Absence of Conditional Conservatism

As a validity check, consider the asymmetric timeliness coefficient in the absence of conditional conservatism. In this case, accountants symmetrically do not recognize any of the information in yt in the current period t, regardless of how negative, that is, wt = 0 for all yt. Under the linearity assumption, equation (14) implies math formula Thus, under the null hypothesis of no conditional conservatism, the Basu coefficient is zero. Indeed, this should be the case if the model is well specified and the coefficient is a valid measure of conditional conservatism.

3.3.3. Basu Regression Coefficient Validity

We conclude that the Basu asymmetric timeliness coefficient math formula is positive in the presence of conditional conservatism, and zero in the absence of conditional conservatism, consistent with it being a valid estimator. This result contradicts the claims of Dietrich, Muller, and Riedl [2007], to which we return below. Further, the coefficient is not driven by a single structural parameter, but rather is a function of conditional accounting conservatism and other attributes of the information environment, including information about growth options (Roychowdhury and Watts [2007]), which we discuss in more detail in the following section.

4. Conditional Conservatism, Firm Characteristics, and Empirical Implications

In this section, we demonstrate that the asymmetric timeliness coefficient varies with firm characteristics affecting their information environments, such as the length of the operating and investment cycle, and the degree of diversification. We particularly examine one characteristic, the relative amount of symmetrically “unbooked” information (notably, revised expectations about rents or growth options) and the extent to which it is correlated with other information components. We then discuss the conditions under which a proxy for this characteristic is the market-to-book ratio, formalizing and extending the analysis in Roychowdhury and Watts [2007], among others.

4.1 The Role of Information About Growth Options and Rents

Consider the case where the variance of the information component g in economic income is nonzero. Current-period accounting income then includes a component that is a function of past shocks to the firm's growth options. In our two-period model, equation (2) has shocks to growth options showing up in accounting income with a simple one-period lag.

Define

display math(18)

The Basu asymmetric timeliness coefficient then is given by (see the appendix for the derivation):

math image(19)

where math formula

Several limiting cases are of interest. To avoid cumbersome notation, here we loosely refer to plim math formula as β2. As the correlation coefficient between z and y approaches 1, we have:

display math(20)

This result is intuitive and suggests that the asymmetric timeliness coefficient depends on the variance of y relative to the total variance of returns, as one would expect when the components are perfectly correlated.15 As one also would expect, math formula, and math formula. In the first case, the variation in y subsumes all other components, while in the second there is no role of conditionally conservative accounting. Finally, math formula and math formula, which mirror the previous results and also are intuitive. The limiting cases suggest that the Basu specification captures the extent of accounting conservatism in a meaningful and intuitive way.

We now show that the Basu asymmetric timeliness coefficient declines in the variance of growth options. The derivative of β2 with respect to math formula is:

math image(21)

where math formula. This expression is negative for economically meaningful values of the parameters and the expression for β2 is symmetric in math formula and math formula. Since math formula is increasing in both math formula and math formula, it follows that, holding the correlation coefficient ρ constant, β2 is decreasing in math formula and math formula. This result is intuitive, in that the x and g components of revision in price are treated symmetrically in financial reporting (fully incorporated and fully ignored, respectively), whereas the y component is treated asymmetrically (incorporated only if below a threshold).

Figure 1 plots the asymmetric timeliness coefficient as a function of math formula and math formula. Other parameters are fixed at the following levels: math formula and math formula.16 As expected, the coefficient increases in the variance of y (reaching 1 in the limit), as timely loss recognition becomes more important. Figure 1 also illustrates that asymmetrically timely loss recognition decreases in the variance of g (information about growth options), reaching zero in the limit. This result also is intuitive. Shocks to growth expectations are not captured contemporaneously in accounting income, so in a regression of accounting income on stock returns their variability math formula dampens the coefficient on returns. These results confirm the original conjecture of Ball, Kothari, and Robin [2000, p. 48] that asymmetric earnings timeliness is a function of the amount of news about “unbooked” growth options relative to news from other sources.

Figure 1.

β2 coefficient as a function of math formula and math formula.

4.2 Effect of Correlation Among Components of Earnings Innovations

The literature is silent on how asymmetric timeliness is affected by complementarities or correlations among the information components that drive returns but that are reflected in earnings in different periods. We use equation (19) to plot β2 as a function of the correlations math formula and math formula (we do not separately consider math formula because it is easy to see that its effect is the same as that ofmath formula).

Figure 2 presents results when other parameters are fixed at the levels: math formula and math formula.17 As the figure shows, asymmetric timeliness increases in math formula, the correlation between y and g, and it decreases in math formula, the correlation between x and g. Intuitively, the first result obtains because in our model y exhibits timely recognition, but the researcher observes total return Rt, which also contains other information components. As the correlation between the return components y and g strengthens, the Basu model's asymmetric timeliness coefficient increases. Similarly, the second result obtains because variation in the sum of the symmetrically recognized x and g components becomes more pronounced due to their covariability, and the conditional conservatism asymmetry becomes less salient. Thus, the correlations and interactions between different information components are expected to have an effect on timely loss recognition empirically.

Figure 2.

β2 coefficient as a function of math formula and math formula

4.3 Earnings Timeliness and Market-to-Book

In this subsection, we discuss the relation between the (beginning-of-period) market-to-book ratio (math formula) and asymmetric timeliness, which has been the subject of some controversy in the literature. A number of studies (e.g., Pae, Thornton, and Welker [2005], Givoly, Hayn, and Natarajan [2007]) show that this relation is negative and, to the extent that math formula is also a measure of conservatism, these studies suggest that the Basu measure is biased. Roychowdhury and Watts [2007] provide the most recent and comprehensive treatment on this subject by proposing a theoretical framework that explains the inverse relation of asymmetric timeliness and math formula (see also Beaver and Ryan [2005]). We start by summarizing their arguments and then explain how we complement and extend their analysis.

Roychowdhury and Watts [2007] decompose equity value into separable assets and rents. Separable assets include unverifiable assets in excess of cost, verifiable assets in excess of cost, and assets at cost (net of depreciation and impairments). Rents stand for above-competitive future returns and may (but need not) include growth options, synergies, monopoly power, etc. They argue that accounting excludes the measurement of rents or their changes in the financial statements in part because they are unverifiable. Further, they maintain that conservatism is characterized by the degree of understatement of net separable assets on the balance sheet, not due to the omission of rents from the financial statements. Consequently, in the context of the Basu regression model, Roychowdhury and Watts [2007] view rents and changes in rents as measurement error arising from using stock prices and stock returns, respectively, to infer properties of balance sheet and income statement information. Because rents are reflected in the numerator of m/b, econometrically math formula serves under this interpretation as a proxy for the measurement error in returns, the independent variable in the Basu regression model, which in turn attenuates the Basu measure of conservatism.18

We start by noting an important conceptual point of distinction of our model. Viewed from the perspective of timeliness, news about rents (in our notation, g) is not measurement error because rents eventually are converted into net assets (ultimately, cash) and go through the income statement. By the time the conversion takes place, the stock price already has reflected g. Therefore, the matching between stock returns and earnings is a matter of time, or leads and lags. Given that the objective of the Basu model is to measure timeliness, g is a natural information component that affects overall and asymmetric timeliness (as figure 1 indicates). Since math formula is likely correlated with the relative variance of g, the negative relation between math formula and conditional conservatism is a natural property of accounting earnings in our model.19

Our model advances the analysis in Roychowdhury and Watts [2007] in three ways. First, our model shows that the effect of math formula on asymmetric timeliness is more complex than it first appears. The analyses in sections 'The Role of Information About Growth Options and Rents' and 'Effect of Correlation Among Components of Earnings Innovations' demonstrate that the Basu coefficient is not determined by the amount of rents (or growth options) relative to other assets per se, but by the relative amounts of new information about them and how this information is correlated with other information: that is, it is a second moment effect. For example, consider a firm, such as a regulated utility where shocks to the value of asymmetrically booked assets (y in our notation) are negligible, that is, math formula. At this extreme, all information about firm value is incorporated symmetrically in earnings, so here too the incremental Basu slope is zero, independent of math formula. More generally, a higher math formula could imply a simultaneous increase in the variance of g and a decrease in the variance of x. As equation (21) and figure 1 indicate, the effect on asymmetric timeliness is ambiguous in such a case. Further, while Roychowdhury and Watts [2007] assume that rents are independent of other information types (p. 11), we show that the correlation between shocks to assets in place and growth options can affect the relation between earnings timeliness and market-to-book ratios. In particular, figure 2 indicates that, if higher math formula also leads to a higher correlation between y and g, the link of math formula and asymmetric timeliness is not necessarily negative.

Second, Roychowdhury and Watts [2007] argue that asymmetric timeliness should become more pronounced over longer earnings horizons because rents convert to separable assets and thus their negative effect on the Basu measure diminishes. Our analysis generalizes this argument, but also suggests that the effect of horizon is not always clear. Intuitively, as the return period is extended back, current-period earnings reflect more of the growth options at the beginning of the period. Further, shocks to growth options are more likely to flow through the financials over longer horizons. One can think of this as increasing the variance of x or y relative to the variance of g. On the one hand, to the extent that a longer horizon means that g can be treated as y, that is, the relative variance of y widens, asymmetric timeliness increases. On the other hand, over an even longer horizon x can replace g, in which case the relative variance of x widens and asymmetric timeliness decreases (intuitively, over sufficiently long horizons earnings approximate returns and hence the asymmetry should disappear).

Finally, our model implies that there are important nonlinearities and interactions in the relation between asymmetric timeliness and the variances of x, y, and g (and hence math formula), as illustrated in figure 1.

An important related question is whether the validity of the Basu asymmetric timeliness coefficient as a proxy for conditional conservatism is threatened because, in fact, Basu coefficients and market-to-book ratios are negatively correlated.20 Indeed, the negative correlation has lent credence to criticisms that the Basu asymmetric timeliness coefficient is flawed (Dietrich, Muller, and Riedl [2007], Patatoukas and Thomas [2011]). Our analysis suggests otherwise, as it demonstrates that Basu coefficients derive from an interaction between asymmetrically conservative accounting rules and practices and the underlying economic characteristics of the firm. Depending on one's research objectives, controlling for firm characteristics may or may not be appropriate. As an example, consider Nikolaev's [2010] hypothesis that financial covenants are effective in distressed situations only if there is timely recognition of losses in earnings, and evidence that financial covenants in public debt contracts indeed are correlated with firms’ Basu coefficients. For this research question, it is not essential whether the size of the Basu coefficient derives from firm characteristics, from accounting, or from the way they intersect.21

4.4 Other Determinants of Conditional Conservatism

The result that asymmetric timeliness is decreasing in math formula is intuitive in our setup and presents several additional empirical predictions. For example, companies with short operating cycles, short investment cycles, or short asset maturity, or companies with a substantial realized component of economic income, are expected to exhibit lower levels of timely loss recognition. This reconciles with the evidence of Khan and Watts [2009], and also with their prediction that conditional conservatism increases in environmental uncertainty, which we expect to be inversely related to the variance of x and positively related to the variance of y (due to different verification requirements for earnings components that are more uncertain and more costly to verify).

Finally, we note that asymmetric timeliness is not affected by math formula, the amount of accounting-induced noise. This is a direct consequence of accounting income being specified as the dependent variable (as well as noise being assumed independent of all return components). The implication is that, while asymmetric timeliness is a valid property of accounting income, it is by no means a complete measure of accounting quality. In particular, it does not reflect random accruals errors, such as errors in recording inventory receivables and payables.

5. Return Endogeneity and Sample Truncation

Dietrich, Muller, and Riedl [2007], henceforth DMR, argue that the Basu regression specification is misspecified for two reasons: earnings cause returns; and conditioning on the sign of the returns induces sample truncation bias. Both claims are based on a misconception that the objective of a Basu specification is to estimate structural parameters in the model (DMR, p. 100):

display math(22)

where the error term math formula is independent of It, and γis known as the “earnings response coefficient.”22

DMR show that the OLS slope coefficient from a “reverse” regression of earnings I on returns R is a biased estimator of the inverse of the earnings response coefficient, 1/γ, because the error term in the reverse model is not independent of Rt. This well-known result is irrelevant if the research objective is to measure how accounting income incorporates the information in stock return, that is, to estimate the functional shape of the conditional expectation E(I | R), in which case (22) is the incorrect specification. As shown in section 'Representing Timeliness as The Conditional Expectation of Earnings', least squares regression achieves that research objective without bias.

Second, DMR (p. 102) argue that conditioning the regression of earnings on returns upon the sign of returns induces “sample truncation bias.” Specifically, DMR argue that the conditions math formula and math formula imply a truncation bias in a Basu regression. However, these conditions only are relevant in a regression of returns on earnings, as in their (22).23 The relevant condition for a Basu association study regression is math formula, where math formula. Indeed, math formula The effect of sample truncation on the least squares outcome for equation (4) can be expressed as:

display math(23)

This implies that truncation bias arises when math formula. However, truncating the sample on R (or a function of R) does not lead to a bias as, using the law of iterated expectations, math formula. Consequently, when the research objective is to estimate the functional shape of the conditional expectation E(I | R), return is the correct independent variable, and conditioning on it does not induce bias.

6. Summary and Conclusions

We present a model of accounting income that is based on salient properties of income recognition in accounting, and use it to analyze the econometrics of the Basu asymmetric timeliness coefficient. The analysis addresses conceptual and econometric challenges to the coefficient's validity as a measure of conditional conservatism. We show that the Basu measure is unbiased under the null hypothesis of zero asymmetry, and that under the alternative hypothesis it captures conditional conservatism as formulated in our model.

We also demonstrate that conditional conservatism is a function of a variety of firm characteristics. Notably, we formalize the intuition and extend the analysis in Roychowdhury and Watts [2007], demonstrating econometrically that a negative relation between market-to-book ratios and asymmetric timeliness coefficients is expected, but is more complex than previously thought. The market-to-book ratio does not per se determine asymmetric timeliness; its effect arises because it is correlated with the relative amount of total price revision associated with revisions in “unbooked” components, such as growth option expectations. We also discuss the effect of firm characteristics, such as operating and investment cycle length, asset maturity, the correlation between the components of returns (hence the degree of diversification), and the length of the fiscal period studied. An important issue that context-specific research design must consider is whether it is appropriate to control for such firm characteristics.

Appendix

Analytical Derivation of Asymmetric Timeliness: The Case of Correlated x, y, and g

We derive the asymmetric timeliness coefficient for a general case where all three components of return, x, y, and g, are considered. The cases when some of these components are omitted follow immediately. We omit subscripts t when all random variables are contemporaneous.

We start by noting that:

math image(A1)

and

display math(A2)

Exploiting symmetry we have:

math image(A3)

where math formula is an indicator variable taking a value of one in case y < 0 and zero otherwise.

Define math formula, math formula and

display math

To show that β2 is positive it suffices to show that math formula, where:

display math

Note further that:

display math(A4)

where math formula is bivariate normal density and math formula.

The integrals can be evaluated explicitly under the normality assumption, first noting that symmetry implies math formula

display math(A5)

Further, making the same transformation, and integrating by parts it can be shown that:

display math(A6)

where math formula, and math formula

By analogy (making the same substitution as above) we have:

display math(A7)

Using integration by parts, it can be further shown that:

display math(A8)

and, by analogy,

display math(A9)

Now we can use these results to evaluate the expression for Δ:

display math(A10)

Rearranging this expression yields:

math image(A11)

Now the asymmetric timeliness coefficient can be computed as follows:

display math(A12)

where we use the result that math formula.

  1. 1

    We use the term “validity” as shorthand for “construct validity,” which Peter [1981, p. 134] notes “generally is used to refer to the vertical correspondence between a construct which is at an unobservable, conceptual level and a purported measure of it which is at an operational level.” See the classic Cook and Campbell [1979] text and its following editions.

  2. 2

    The model and its origins are described in Basu [2009].

  3. 3

    Basu [2005, section 'A Framework for Interpreting Regressions of Earnings on Returns'], Ball, Kothari, and Robin [2000, fn. 15], and Ball and Shivakumar [2005, 2008].

  4. 4

    As of July 19, 2013, Basu [1997] has 2,116 citations in Google Scholar and 355 citations in the Social Sciences Citation Index, making it one of the more highly referenced papers in the modern accounting literature.

  5. 5

    Basu [1997] offers no formal analysis. Increasingly formal but nevertheless limited analyses are offered in Ball, Kothari, and Robin [2000], Pope and Walker [1999], Ryan and Zarowin [2003], and Beaver and Ryan [2005].

  6. 6

    Studies using the Basu model typically use annual returns to coincide with the periodicity of reported earnings and the settlement period in many contracts based on earnings (notably, debt, supply, and compensation agreements). As argued in Ball, Kothari, and Robin [2000], annual horizons substantially mitigate concerns about the informational efficiency of security prices.

  7. 7

    That prices lead earnings is clear from the graphs in Ball and Brown [1968], Foster, Olsen, and Shevlin [1984], and Bernard and Thomas [1990], and from the longer horizon studies of Beaver, Lambert, and Morse [1980] and Kothari and Sloan [1992].

  8. 8

    We ignore dividends, the implicit assumption being Miller–Modigliani dividend irrelevance.

  9. 9

    In contrast, the distribution of earnings generally will be left skewed, not symmetric, due to conditional conservatism.

  10. 10

    Equations (2) and (3) show that these would induce nonlinearity in the relation between I and R. As discussed in section 'Representing Timeliness as The Conditional Expectation of Earnings', the shape of the return distribution per se does not invalidate the Basu [1997] measure.

  11. 11

    Alternatively, timely loss recognition can occur with probability p, which can be a function of y, that is, Pr(w = 1| y) = p(y).

  12. 12

    Knowledge of ɛt–1 would help predict earnings, but not returns. We assume the market unravels the current accounting error ɛt and does not react to it. This assumption can be relaxed without altering our conclusions.

  13. 13

    The appendix provides derivation for a more general case where the information component g also is present. The result below is thus a special case that obtains when setting g = 0 or, in other words, setting z = x.

  14. 14

    Taking the second derivative indicates that the first derivative is increasing over the interval (0,math formula] and is decreasing over the interval [math formula, +∞), but asymptotes to a positive constant.

  15. 15

    Equation (19) seems to suggest that β2 approaches zero as ρ approaches 1. However, α1 then tends to plus infinity.

  16. 16

    Analogous results obtain when other parameters are fixed at different levels, including correlation coefficients of zero. For illustration, we report the results only under one set of parameters.

  17. 17

    Our model accommodates any correlation between shocks to components of firm value. Positive correlation is assumed only for realism, but can be relaxed without altering our main results and specifically without changing the sign of the expression given by equations (17), (19), and hence (14). In other words, the incremental Basu coefficient (β2) remains positive. To see this, first consider a case with only two news components, x and y, that share a negative correlation. While, for some parameter values, the argument α can become negative if math formula, the expressions (17), and hence (14), remain positive. We previously have shown that expression (17) is positive for any positive alpha, and it can be shown by analogy that it also is positive for any negative alpha. To see this, note that the function given by math formula is symmetric about the vertical axis math formula. In a more general case, where the component g is also present, by analogy the assumption that math formula is inconsequential as g behaves in a similar way as x. Finally, a negative correlation between x and g will not affect the expression (4.2), and hence (14), because this equation is only affected by the correlation of z = x + g and y, that is, only math formula and math formula discussed above matter here.

  18. 18

    A similar conclusion follows from Beaver and Ryan's [2005] simulation that alters intangibles intensity, changes in which are viewed as noise.

  19. 19

    Whether one calls this a “measurement error” problem or a natural property of accounting earnings is at one level a semantic issue. Nevertheless, this issue is important as the appropriate econometric solutions differ, depending on the context (as discussed further below).

  20. 20

    Gigler and Hemmer [2001], Richardson and Tinaikar [2004], Beaver and Ryan [2005], Price [2005], Pae, Thornton, and Welker [2005], Givoly, Hayn, and Natarajan [2007], and Roychowdhury and Watts [2007].

  21. 21

    Ball, Robin, and Sadka [2008] raise the related issue of whether firm characteristics truly are exogenous in cross-country research on the effects of accounting rules and practices, or whether they are determined jointly.

  22. 22

    Because this model suppresses conditional conservatism, the variable I in this section is different from the variable I in our analyses above. In particular, DMR's (9.3) and derivative equations cannot be substituted into our earlier equations.

  23. 23

    Truncation is figuratively depicted incorrectly in DMR (figure 2), where the regression of earnings on returns minimizes the horizontal (squared) deviations from the regression line, not the vertical deviations. After one notes this, it immediately becomes clear from the graph that no truncation bias is present in a Basu regression. DMR also offer a “simulation” that they claim should not exhibit nonzero Basu coefficients. Ball, Kothari, and Nikolaev [2013, fn. 17] disprove this claim.

Ancillary