Conditional conservatism and disaggregated bad news indicators in accrual models

Abstract

Conditional conservatism is an integral but often unmodeled part of the normal accrual process. The standard economic determinants of accruals contain information about unrealized losses. We argue that accountants recognize these unrealized losses as disaggregated write-downs for small asset pools. Modeling disaggregated impairments yields new economic insights about accruals and improved accrual models. We predict that accrual conservatism manifests as a sum of asymmetries for a vector of news indicators, rather than as an asymmetry for a scalar aggregate news proxy. We argue that more detailed segment-level and quarterly indicators have an incremental effect on annual firm-level accruals. We also predict a dynamic effect of successive loss indicators because accountants look for consistent patterns in these variables. Empirical results for U.S. firms support our predictions. The asymmetries in accruals are consistent with conservatism in validation tests. We also document improved statistical power and type I error in earnings management tests.

1 Introduction

Research on abnormal accruals requires a good understanding of the normal accrual process (e.g., Dechow et al. 1995). We argue that this normal process should incorporate conditional conservatism—more timely recognition of bad news than good news (Basu 1997). We examine how accrual conservatism arises from asset impairment tests and predict that these impairments are triggered by a vector of loss indicators that is mismeasured by aggregated scalar proxies for bad news. Using the Allen et al. (2013) model of working capital accruals as a benchmark, we find large asymmetric effects of individual bad news indicators, consistent with conservatism at a low aggregation level. Our analysis offers new insights about accruals and improved benchmark models for earnings management tests.

Using stock return as an aggregate proxy for news, Basu (1997) shows that net income and accruals incorporate bad news more quickly than good news and infers conditional conservatism. Numerous studies verify this asymmetric effect and show that it varies in agreement with conservatism theory (e.g., Ball et al. 2000; Watts 2003b; Ball et al. 2013; Collins et al. 2014a).

Unlike conservatism research, the accruals literature typically uses a linear specification with multiple economic determinants of accruals, such as sales growth and gross property, plant, and equipment in the Jones (1991) model and multi-period cash flows in the Dechow and Dichev (2002) model. Bushman et al. (2011b) and Allen et al. (2013) combine and refine these standard models, using changes in sales and total employees as measures of firm growth and past, current, and future cash flows as proxies for the matching role of working capital accruals.

These economic variables likely contain information about unrealized losses. For example, negative cash flow can indicate a rise in uncollectible accounts or a decline in selling prices (Ball and Shivakumar 2005, 2006). A sales decrease likely predicts lower demand for the inventory on hand. Layoffs likely reflect managers’ expectation of an enduring demand decrease. These unfavorable indicators can trigger asset write-downs (i.e., negative accruals). In contrast, favorable indicators rarely lead to asset write-ups under U.S. GAAP.

We build our predictions from the micro-foundations of conservatism—asset impairment tests. Different asset classes, such as inventory and receivables, are tested for impairment separately, and accountants can divide each asset class into smaller pools (e.g., they can test total inventory as a single pool or test distinct inventory items in separate pools). The composition of these pools affects impairment amounts (e.g., Kieso et al. 2013, p. 477). Suppose that a firm has an unrealized loss of $10 on asset X and an unrealized gain of $10 on asset Y. If these assets are combined into a single pool, then the impairment is zero based on total value change for the pool (ΔX + ΔY = −$10 + $10 = $0). If the assets are tested separately, then the impairment is $10 based on their individual value changes (ΔX = −$10, which triggers a write-down, and ΔY = $10, which does not lead to a write-up). Thus the firm can conceal the unrealized loss on asset X by pooling it with the unrealized gain on asset Y or the firm can disclose this loss by using disaggregated asset pools.

Disaggregated losses convey useful information. For example, when a firm has large unrealized losses on inventory of an unsuccessful product, the firm’s future performance can be improved by eliminating, redesigning, or repositioning this product. This information is likely viewed as material by stakeholders (Statement of Financial Accounting Concepts 2, FASB 1980). Therefore impairment tests are likely conducted separately for small pools of assets (e.g., narrow inventory categories such as subcompact economy cars, mid-size economy cars, or mid-size luxury cars) to identify disaggregated losses. The write-down determinants likely differ across these pools. Therefore we predict that working capital accruals exhibit asymmetries for a vector of bad news indicators because each indicator triggers impairment for some of the pools.

Second, we argue that segment and quarterly indicators influence total write-downs. Even when firm sales are increasing, a segment with declining sales likely has impaired assets. Even when annual sales increase, a fourth quarter sales decrease can indicate impairment of assets at fiscal year-end. Therefore, conditional on the asymmetric effects of annual firm-level indicators, these more detailed indicators will have an asymmetric effect on annual firm-level accruals.

Third, we argue that, because repeated cash losses are unlikely to be caused by normal variation in working capital (Dechow 1994; Ball and Shivakumar 2006), they signal permanent impairment. Therefore asymmetric timeliness is likely greater for successive cash losses.

We use U.S. Compustat/CRSP data from 1962–2007. The results support our predictions and are robust. Vuong (1989) tests show that conservatism in firm-year data should be modeled using asymmetries for individual news indicators rather than a single asymmetry for a summary measure of news (that aggregates the same underlying indicators).¹ These asymmetries are consistent with conservatism in additional tests, and they improve statistical power (by up to 15 %) and type I error (by a factor of 10 in some cases) in earnings management tests.

Our main contribution is to develop new insights about accruals. Conservatism pervades accounting practice (e.g., Sterling 1967; Basu 2009) and standards (e.g., Lawrence et al. 2013), such as Accounting Standards Codification (ASC) topics 310 and 330 for receivables and inventory, respectively (FASB 2016c, d).² Although asymmetric loss recognition via accruals is widely documented (e.g., Basu 1997; Ball and Shivakumar 2005, 2006; Hsu et al. 2011, 2012), these insights have not yet influenced the standard accrual models. With few exceptions, this research does not examine the role of detailed accounting data. Furthermore, while prior research often assumes that accrual asymmetry reflects conservatism, we test it against alternative explanations, such as curtailment (Lawrence et al. 2016) and cost stickiness (Banker et al. 2016b).

We develop the hypotheses in Sect. 2, describe the data and the estimation models in Sect. 3, present the empirical results in Sect. 4, and conclude in Sect. 5.

2 Hypotheses development

Prior research has identified two primary drivers of working capital accruals. First, many accruals relate to firm growth (e.g., Jones 1991; Dechow et al. 1998). For example, to support higher sales, a firm will likely increase its inventory and accounts receivable. Second, accruals play a matching role by aligning related cash inflows and outflows in the same period (e.g., Dechow 1994; Dechow and Dichev 2002). For example, if a customer pays for merchandise later than usual, this will manifest as a temporary decrease in cash flow and a temporary increase in accounts receivable, both of which will reverse when the payment is received. Bushman et al. (2011b) and Allen et al. (2013) integrate the growth component, as in the Jones (1991) model, and the matching component, as in the Dechow and Dichev (2002) model, and show that the combined model yields new insights.³ They model the growth component as a linear function of percentage changes in sales and number of employees and model the matching component as a linear combination of lagged, current, and future cash flows (Dechow and Dichev 2002). We use the Allen et al. (2013) model as the benchmark model for our analysis.

While accrual models are typically linear, conservatism research identifies an important asymmetry in both earnings and accruals. Conditional conservatism is interpreted as the higher verification threshold to recognize good news as gains than to recognize bad news as losses (Basu 1997; Watts 2003a). Basu shows that conservatism implies asymmetric timeliness of earnings and accruals with respect to good versus bad news. Using stock return as a summary measure of news about future cash flows, he finds that net income and accruals respond more to negative returns (bad news) than to positive returns (good news), consistent with his predictions. Ball and Shivakumar (2005) introduce cash flow as a conservatism indicator for private firms that lack stock prices. Ball and Shivakumar (2006) document asymmetric timeliness for both stock return and cash flow in multi-indicator accrual models, which increase explanatory power over linear models.⁴

Conservatism research shows that asymmetric timeliness is pervasive (see reviews by Watts 2003b; Mora and Walker 2015; Ruch and Taylor 2015) and that it varies predictably with various drivers of conservatism (Watts 2003a), such as litigation exposure (Basu 1997; Holthausen and Watts 2001; Qiang 2007; Chung and Wynn 2008), country and industry characteristics (Pope and Walker 1999; Ball et al. 2000, 2008; Dhaliwal et al. 2014), corporate governance (Garcia Lara et al. 2009), managerial stock ownership (LaFond and Roychowdhury 2008), debt contracting (Zhang 2008; Wittenberg-Moerman 2008; Nikolaev 2010; Jayaraman and Shivakumar 2013), and information asymmetry (LaFond and Watts 2008).

For consistency with Allen et al. (2013), we focus on working capital accruals, but the predictions generalize to broader accrual measures.⁵ Conditional conservatism can have an asymmetric effect on many working capital accounts. For example, bad news about the value of inventory on hand is likely to trigger an inventory write-down, i.e., a negative accrual that reflects these expected future losses per the lower-of-cost-or-market rule (ASC 330-10-35-1).⁶ Similarly, unrealized losses on receivables are likely to be recognized early (ASC 310-10-35-41). In contrast, unrealized gains on inventory and receivables cannot be recognized early as an asset write-up (ASC 310-10-35-41 and ASC 330-10-35-16, respectively).⁷ Therefore working capital accruals incorporate bad news more quickly and fully than good news, resulting in an asymmetric relation between accruals and various indicators of future gains and losses.

Decreases in sales and total employees can signal unrealized losses. Because sales changes are persistent (Fairfield et al. 2009), a current sales decrease predicts further deterioration in demand. Future selling prices for inventory on hand will likely be lower than originally expected, which can lead to an inventory write-down.⁸ Receivables are appraised “in the light of the current economic environment” (ASC 310-10-35-10), which includes expected demand trends. Therefore a sales decrease can also signal impairment of receivables. Pessimistic managers are more willing to lay off workers (e.g., Banker et al. 2014). Therefore a reduction in the number of employees signals additional bad news, which can trigger further write-downs.

Lagged, current, and future cash flows likely influence impairments. For example, negative current period cash flow can indicate that the proportion of uncollectible receivables is higher than expected or that the selling prices are lower than expected, which can cause impairments of receivables and inventory, respectively (Ball and Shivakumar 2005, 2006).

How accountants combine these indicators depends on the composition of asset pools in impairment tests. For example, suppose that a firm has two assets X and Y with current book values of $100 each. The fair value estimates for these assets are $x and $y, respectively, which we assume are observable for simplicity. If the assets are tested for impairment as a single pool, then the fair value of the pool as a whole ($x + $y) is compared to its book value ($200). When $x + $y < $200, the pool is written down to its fair value $x + $y. The resulting asset write-down (coded as a negative number) is min{0, $x + $y − $200}, i.e., an asymmetric function of the linear combination $x + $y. In contrast, if the assets are tested as two separate pools, then the fair value of each asset ($x or $y) is compared to its book value ($100). When $x < $100, asset X is written down to $x (regardless of $y), and when $y < $100, Y is written down to $y. The total firm-level write-down is min{0, $x − $100} + min{0, $y − $100}, i.e., the sum of asymmetric functions of the individual indicators $x and $y. Thus, depending on the composition of asset pools, conservatism in firm-level data can manifest either as an asymmetry for a scalar summary measure ($x + $y) or as a sum of asymmetries for the components of the vector ($x, $y).

These scenarios have different implications for the firm’s stakeholders. Suppose that the fair values of assets X and Y are $90 and $110, respectively, i.e., the firm has an unrealized loss of $10 on X and an unrealized gain of $10 on Y. If X and Y are combined into a single pool, the total write-down is $0 [= min{0, $90 + $110 − $200}] because the loss on X is offset by the gain on Y. Thus managers can conceal bad news through aggregation of loss and gain assets (Basu 2005).⁹ In contrast, if X and Y are separate pools, then the total write-down is $10 [= min{0, $90 − $100} + min{0, $110 − $100}], i.e., bad news about asset X is revealed quickly. The write-down reduces reported earnings by $10, which can decrease managers’ performance-based compensation and trigger accounting-based debt covenants. Notably, this disaggregated scenario gives managers an incentive to quickly terminate, restructure, or adapt unprofitable projects (Burgstahler and Dichev 1997), improving the firm’s economic performance.

We argue that impairment tests are conducted for multiple asset pools and that the fair value determinants (i.e., write-down triggers) differ across these pools. First, asset classes such as inventory and receivables are tested for impairment separately (ASC topics 330 and 310, respectively). The fair value of receivables reflects the last stage of the operating cycle, i.e., conversion of sales into future cash. The fair value of inventory reflects an earlier stage, i.e., conversion of inventory into future sales, and is likely based on different indicators. For example, because firms commit labor resources in advance to produce goods, change in the number of employees is a forward-looking indicator of managers’ future sales plans. Therefore it likely has a stronger association with the fair value of inventory on hand (which is the source of future sales) than with the fair value of receivables (which is expected cash from past sales). Because inventory is converted into cash with a longer lag than receivables, additional forward-looking indicators may also have a greater impact. Thus the relevant indicators for inventory and receivables likely differ and cannot be combined into a summary news measure that would accurately capture value changes for both asset classes.

Second, each asset class comprises asset subsets with distinct fair value determinants. For example, consider a firm that sells economy cars to low-income customers and luxury cars to high-income customers. The fair value of inventory is based on the expected selling prices. The prices likely vary with the demand shocks for each customer type and competition in each product category. These price determinants likely differ between economy and luxury cars. For example, income shocks likely have a greater impact on purchases for low-income customers, and price competition is likely more intense in the economy category. The fair value of receivables from sale of these cars reflects expected collections from low-income and high-income customers, respectively, which likely have different determinants (e.g., Bhat et al. 2014). The fair value determinants could also vary with the product’s expected life. For example, a short-term demand decrease likely has a much greater impact on prices of perishable inventory, which must be sold quickly, than on prices of long-lived inventory, which can be sold after the demand recovers. After controlling for short-term demand indicators, a long-term demand decrease likely has a much smaller effect on perishable inventory than on long-lived inventory because perishable inventory on hand will be sold or discarded quickly.

Disaggregated impairments for these dissimilar items can convey important information to stakeholders. For example, if a firm reports impairment for an unsuccessful product, then this suggests that future profitability can be improved by eliminating, refining or repositioning the product. In contrast, if a firm conceals the impairment through aggregation with successful products, then it communicates mediocre profitability for the pool without a clear path to improvement. Therefore, if accountants aim to report informative impairments, they will form separate pools for distinct subsets of inventory. Similarly, they will likely form distinct pools of receivables based on economic factors such as customer types or time outstanding (e.g., ASC 310-10-50-7A).

Although accountants exercise limited discretion in forming pools within each asset class, they are unlikely to try to cover up disaggregated losses by pooling economically dissimilar assets. Statement of Financial Accounting Concepts 2 (FASB 1980) explains that even small items can be material “if they arise in abnormal circumstances” (paragraph 123). Economic impairment constitutes “abnormal circumstances” and is likely to be interpreted as material information that must be disclosed. Staff Accounting Bulletin (SAB) 99 of the Securities and Exchange Commission (SEC 1999) clarifies that an economic fact is material if it significantly alters the total mix of available information that affects various stakeholders. For example, even when an unrealized loss on a particular subset of inventory is offset by unrealized gains for other inventory items, this loss can be material because its disclosure (1) conveys essential information to investors about unsuccessful economic activities of the firm, (2) reduces reported earnings, (3) can reduce managers’ performance compensation (e.g., Potepa 2014) or the market value of their equity holdings, and (4) can trigger debt covenants and other contractual restrictions that are based on reported earnings.¹⁰ Therefore accountants likely recognize unrealized losses at a low aggregation level, treating asset subsets with distinct fair value determinants as separate pools.

In annual firm-level data for accruals, this disaggregated process is best modeled as a sum of asymmetric effects of individual explanatory variables, rather than an asymmetric effect of a single summary measure (that aggregates these variables in a way that best fits the data). The firm-level explanatory variables are likely associated with the detailed internal indicators used by accountants (e.g., firm-level sales growth reflects the total of product-level sales changes). For some of the disaggregated asset pools, the fair value will relate most closely to sales growth (e.g., perishable inventory and short-term receivables, which are most affected by trends in the immediate future). Write-downs for these pools will be an asymmetric function of firm-level sales growth (as an empirical proxy for detailed internal sales growth indicators). For some of the other pools, employee growth will be more relevant (e.g., long-lived inventory that is sensitive to indicators of sustained demand decreases such as layoffs). The associated write-downs will be an asymmetric function of employee growth.

Ball and Shivakumar (2005, 2006) predict and find that current period cash flow has an asymmetric effect on accruals. We extend their analysis to future and lagged cash flows. Future cash flow is an ex post proxy for forward-looking information that managers have in the current period (Dechow and Dichev 2002). When future cash flow is negative (as observed ex post by a researcher), this suggests that the forward-looking information in the current period is likely unfavorable, which can trigger write-downs for some of the pools. Conservatism for expected future cash flow also leads to an asymmetry of the opposite sign in deferred recognition of lagged cash flow, because lagged cash losses were likely recognized previously.¹¹ The firm-level accrual includes total write-downs for all pools, i.e., the sum of all of these asymmetric effects.

H1

Conservatism for firm-level accruals is better approximated by a sum of asymmetric effects of individual news indicators than by an asymmetric effect of an aggregate summary measure of news that combines all of these indicators.

H2a

Accruals exhibit asymmetric timeliness with respect to concurrent sales growth and growth in the number of employees.

H2b

Accruals exhibit asymmetric timeliness with respect to future cash flow and an asymmetry in the opposite direction with respect to lagged cash flow.

While accrual models typically use annual firm-level indicators, they can be enriched with more detailed indicators. To maintain focus, we only decompose variables from our main model in ways that are relevant to impairment practice. First, in multi-segment firms, most asset impairment tests are conducted within individual segments. Even when aggregate firm sales are increasing, segments with decreasing sales will likely have asset write-downs, whereas segments with increasing sales are unlikely to have asset write-ups.¹² Therefore, conditional on firm-level information, segment-level sales changes will have an asymmetric effect on firm-level accruals.

H3

After controlling for the asymmetric effects of the firm-level news indicators, firm-level accruals exhibit asymmetric timeliness with respect to segment-level sales growth.

Assets are tested for impairment at the date of the financial statements (e.g., ASC 330-10-35-2 and ASC 310-10-35-8 for inventory and receivables, respectively). If sales increase for the full year but deteriorate in the fourth quarter, signaling a decrease in future revenue from the available inventory at fiscal year-end, the firm will likely write down inventory. Sales changes for earlier fiscal quarters are less likely to have an asymmetric effect on annual accruals. First, many unrealized gains and losses from the interim quarters are fully realized by the fiscal year end; therefore they are recognized symmetrically in accruals. Second, each interim period is primarily viewed as an integral part of the fiscal year (ASC 270-10-45-1, FASB 2016a; previously mandated in APB Opinion 28, APB 1973), and interim financial statements need not be audited before issuance (ASC 270-10-S99-1). Firms allocate many annual estimates to interim quarters and typically adjust for estimation errors at the fiscal year-end (Rangan and Sloan 1998). Because auditors tend to prefer more conservatism than managers, and auditor adjustments typically flow through fourth quarter earnings, conservatism increases in the fourth quarter (Elliott and Hanna 1996; Basu et al. 2002). Therefore sales change in the fourth quarter (but not necessarily in other quarters) will have an incremental asymmetric effect on annual accruals.¹³

H4

After controlling for the asymmetric effects of annual firm-level variables, annual accruals exhibit asymmetric timeliness with respect to sales growth in the fourth quarter.

Temporary fluctuations in cash flow subsequently reverse (e.g., Dechow 1994; Ball and Shivakumar 2006). For example, if a firm had unusually large sales at the end of year t − 1 and paid the supplier at the beginning of year t, then its cash flow is higher than usual in year t − 1 and lower than usual in year t. Thus, if cash flow is negative in the current period but positive in adjacent periods, then this likely reflects normal variation in the timing of cash flows rather than an economic loss. This is unlikely to be sufficient evidence for a write-down (ASC 310-10-35-4 and 330-10-35-4). However, if cash flow is negative 2 years in a row (i.e., there is no reversal or the loss is persistent), then it provides stronger evidence of permanent impairment. Therefore we predict that asymmetric timeliness for cash flow in year t is greater when it is accompanied by a negative cash flow in year t + 1, which indicates that accountants do not expect a reversal in year t + 1 based on their forward-looking information.¹⁴ Similarly, gain recognition for a positive cash flow in year t is likely to be weaker when it is followed by a negative cash flow in year t + 1.

H5a

Gain recognition for current period cash flow is smaller when future cash flow is negative.

H5b

Asymmetric loss recognition for current period cash flow is greater when future cash flow is negative.

Although Hypotheses 1–5 arise from some major provisions of U.S. GAAP, they require more than just formal compliance with the accounting standards. Lawrence et al. (2013) distinguish between nondiscretionary conservatism, which arises from unbiased application of U.S. GAAP, and voluntary conservatism, which reflects accountants’ discretion in implementing the accounting guidance. For example, while firms must follow the lower-of-cost-or-market rule, they have leeway in the implementation details. If their objective is to report accurate estimates, then they will form disaggregated asset pools for distinct inventory items (Hypothesis 1), use all relevant predictors (Hypotheses 2–4), and examine cash flows for multiple periods (Hypothesis 5). However, they can also disregard much of this information without violating the standards.¹⁵ Similarly, the guidance for receivables allows for substantial discretion (ASC 310-10-35-4). Thus our predictions hinge on accountants’ trying to faithfully represent the economic drivers of impairment rather than just the formal structure of accounting standards.

3 Data and empirical models

3.1 Sample selection and descriptive statistics

We use annual Compustat/CRSP data from 1962 to 2007. We end the sample in 2007 to exclude the financial crisis, which triggered massive write-downs for many firms, but the results are robust to extending the sample to 2014. Following Allen et al. (2013), we exclude financial firms (SIC codes 6000–6999) and restrict the sample to domestic firms (Compustat items POPSRC = D and FIC = USA) traded on a major U.S. exchange (CRSP exchange codes 1–3).

All variables used in the paper are defined in Table 1. We use Allen et al.’s definitions of accruals and cash flows, which they measure using the balance sheet approach. Working capital accruals are defined as the change in noncash current assets less the change in current operating liabilities (Compustat items ΔACT − ΔCHE − [ΔLCT − ΔDLC − ΔTXP]). Cash flows are computed as operating income (Compustat item OIBDP) less accruals, using an income measure that incorporates working capital accruals and excludes special items, depreciation, extraordinary items and discontinued operations.¹⁶ Accruals and cash flows are scaled by average total assets (Compustat item AT).

Table 1 Variable definitions

We discard firm-year observations with changes in fiscal year-end, missing or invalid data for the main regression variables, and two-digit SIC codes with insufficient data for industry-specific estimation.¹⁷ Following Allen et al., we winsorize scaled accruals and cash flows at ±1 (the results are robust to winsorizing at the 1 and 99 percentile levels). All other variables are winsorized at the bottom and top 1 %. The final sample comprises 109,735 observations for 10,962 firms.

The univariate descriptive statistics are presented in Panel A of Table 2. On average, working capital accrual (ACC) equals 1.7 % of total assets, and the median is 1.1 %. Average (median) annual sales growth (SGR) is 15.6 % (9.9 %). Sales decreases (DS = 1) are 25.2 % of the sample. Average (median) growth rate for total employees (EGR) is 7.4 % (2.6 %). Decreases in total employees (DE = 1) account for 36.7 % of the sample. On average, cash flow (CF) equals 9.6 % of total assets, and the median is 11.8 %. 16.8 % of observations have negative cash flow (DC = 1).

Table 2 Descriptive statistics

Panel B of Table 2 presents the correlation matrix. Working capital accrual is positively associated with changes in both sales and total employees (cor(ACC, SGR) = 0.280 and cor(ACC, EGR) = 0.291, respectively), indicating that many accruals relate to firm growth (Jones 1991). Working capital accrual is negatively correlated with concurrent cash flow (cor(ACC, CF _t ) = −0.323) and is positively correlated with lagged and future cash flow (cor(ACC, CF _t−1) = 0.100 and cor(ACC, CF _t+1) = 0.110, respectively), consistent with the matching role of accruals (Dechow 1994; Dechow and Dichev 2002). The correlations between each of the growth variables and each of the matching variables are all less than 0.1, which suggests that these two groups of variables capture different economic factors.

In Panel C of Table 2, we test for asymmetric association of accruals with each of the independent variables. We partition the sample based on the sign of the respective variable and compute its correlation with accruals within each subsample. Accrual is more correlated with decreases in sales and employees than with increases in these variables (cor(ACC, SGR) = 0.217 for negative SGR versus 0.188 for positive SGR, and cor(ACC, EGR) = 0.232 versus 0.201, respectively; both differences are significant). These results are consistent with conservatism for the growth variables (Hypothesis 2a). For both concurrent and future cash flow, we find a significantly higher correlation of accrual with cash losses than with cash profits, consistent with asymmetric recognition of contemporaneous and future cash losses (Ball and Shivakumar 2005 and Hypothesis 2b, respectively). As expected, the correlation pattern for lagged cash flow is reversed (i.e., accrual is more correlated with cash profits than with cash losses), indicating less deferred recognition of past cash losses relative to past cash profits (Hypothesis 2b).

3.2 Empirical models

Allen et al. (2013) and Bushman et al. (2011b) use the following model for scaled accruals:

$$ACC_{t} = \alpha_{0} + \alpha_{1} SGR_{t} + \alpha_{2} EGR_{t} + \alpha_{3} CF_{t - 1} + \alpha_{4} CF_{t} + \alpha_{5} CF_{t + 1} + \varepsilon_{t} ,$$

(1)

where SGR _t is percentage change in sales; EGR _t is percentage change in total employees; and CF _t−1…CF _t+1 are lagged, current, and future cash flows, respectively, scaled by average total assets for the corresponding year. The firm index is omitted for brevity. Because working capital accruals exclude depreciation, Allen et al. (2013) and Bushman et al. (2011b) do not control for gross property, plant and equipment, as in the Jones (1991) model. We expect a negative coefficient on CF _t and positive coefficients on CF _t−1 and CF _t+1 due to the matching role of accruals (Dechow 1994; Dechow and Dichev 2002) and positive coefficients on the growth variables SGR and EGR.

We examine two asymmetric models for firm-year variables. First, we build on the measurement error argument of Ball and Shivakumar (2006) and Roychowdhury and Watts (2007), in which accountants infer the total bookable value change from multiple noisy indicators and then use this total value change to assess impairment. This argument is best modeled as asymmetric loss recognition with respect to a summary measure of news:

$$\begin{aligned} ACC_{t} &= \alpha_{0} + \alpha_{1} SGR_{t} + \alpha_{2} EGR_{t} + \alpha_{3} CF_{t - 1} + \alpha_{4} CF_{t} + \alpha_{5} CF_{t + 1} + \beta_{1} DX_{t} + \beta_{2} DX_{t} \times X_{t} + \eta_{t} \\ X_{t} & = \alpha_{1} SGR_{t} + \alpha_{2} EGR_{t} + \alpha_{3} CF_{t - 1} + \alpha_{4} CF_{t} + \alpha_{5} CF_{t + 1} , \\ \end{aligned}$$

(2)

where DX _t is a dummy variable that equals 1 if the linear combination of indicators X _t is negative and zero otherwise. DX _t  = 1 indicates that this linear combination as a whole conveys bad news. We estimate this model using nonlinear least squares.¹⁸ Conditional conservatism implies that the coefficient on DX×X is positive, i.e., the linear combination X _t (our empirical summary measure of news) has a greater impact on accruals when it indicates bad news.

In our main asymmetric specification, we model conservatism for firm-level accruals as a sum of asymmetric effects of the individual indicators:

$$\begin{aligned} ACC_{t} & = \alpha_{0} + \alpha_{1} SGR_{t} + \alpha_{2} EGR_{t} + \alpha_{3} CF_{t - 1} + \alpha_{4} CF_{t} + \alpha_{5} CF_{t + 1} \\ &\quad + \beta_{1} DS_{t} + \beta_{2} DS_{t} \times SGR_{t} + \beta_{3} DE_{t} + \beta_{4} DE_{t} \times EGR_{t} \\ &\quad + \beta_{5} DC_{t - 1} + \beta_{6} DC_{t - 1} \times CF_{t - 1} + \beta_{7} DC_{t} + \beta_{8} DC_{t} \times CF_{t} + \beta_{9} DC_{t + 1} + \beta_{10} DC_{t + 1} \times CF_{t + 1} + \omega_{t} , \\ \end{aligned}$$

(3)

where DS _t , DE _t , and DC _t−1 … DC _t+1 are dummy variables for bad news, which equal 1 if SGR _t , EGR _t , and CF _t−1 … CF _t+1, respectively, are negative and zero otherwise. This model extends Ball and Shivakumar’s (2006) multi-indicator accrual models by incorporating asymmetries for all of the included indicators, which is a key implication of our disaggregated-impairment theory.

Hypothesis 1 predicts that model (3) better fits the data than the aggregate model (2) because it captures asymmetries at a low aggregation level. Hypothesis 2a predicts that the coefficients on DS×SGR and DE×EGR are positive, i.e., accruals are more sensitive to decreases than to increases in sales and employees, reflecting asymmetric loss recognition for the growth variables. The coefficient on DC _t ×CF _t is expected to be positive (Ball and Shivakumar 2005), representing conservatism with respect to concurrent cash flow. Hypothesis 2b implies that the coefficient on DC _t−1×CF _t−1 is negative and the coefficient on DC _t+1×CF _t+1 is positive, i.e., accruals exhibit lower deferred recognition and higher early recognition of cash losses relative to cash profits.

To test Hypothesis 3, we add a measure of segment-level sales decreases

$$segSGR_{t}^{ - } \equiv {{\sum\nolimits_{s} {\Delta segSALES_{s,t}^{ - } } } \mathord{\left/ {\vphantom {{\sum\nolimits_{s} {\Delta segSALES_{s,t}^{ - } } } {SALES_{t - 1} }}} \right. \kern-0pt} {SALES_{t - 1} }},$$

(4)

where ∑ _s ΔsegSALES ⁻ _s,t is the sum of sales changes for all segments s that have decreasing sales in year t, and SALES _t−1 is lagged firm-level sales.¹⁹ If all segments have sales decreases, then segSGR ⁻ _t equals the firm-level SGR _t . However, if some segments have a sales increase, then segSGR ⁻ _t adds information. We estimate the following model for multi-segment firms:

$$\begin{aligned} ACC_{t} & = \alpha_{0} + \alpha_{1} SGR_{t} + \alpha_{2} EGR_{t} + \alpha_{3} CF_{t - 1} + \alpha_{4} CF_{t} + \alpha_{5} CF_{t + 1} \\ &\quad + \beta_{1} DS_{t} + \beta_{2} DS_{t} \times SGR_{t} + \beta_{3} DE_{t} + \beta_{4} DE_{t} \times EGR_{t} \\ &\quad + \beta_{5} DC_{t - 1} + \beta_{6} DC_{t - 1} \times CF_{t - 1} + \beta_{7} DC_{t} + \beta_{8} DC_{t} \times CF_{t} + \beta_{9} DC_{t + 1} + \beta_{10} DC_{t + 1} \times CF_{t + 1} \\ &\quad + \delta_{1} segDS_{t} + \delta_{2} segDS_{t} \times segSGR_{t}^{ - } + \mu_{t} , \\ \end{aligned}$$

(5)

where segDS _t is a dummy variable equal to 1 if any of the firm’s segments has a sales decrease in year t and zero otherwise, and the remaining variables were defined previously. Hypothesis 3 predicts that the coefficient on segDS×segSGR is positive, i.e., even after incorporating the asymmetric effects of all firm-level variables from model (3), including sales growth SGR _t , segment-level sales decrease plays an incremental role in asymmetric loss recognition.

To test Hypothesis 4, we include fourth-quarter sales growth in the model for annual accruals, adapting the approach of Stubben (2010):

$$\begin{aligned} ACC_{t} & = \alpha_{0} + \alpha_{1} SGR_{t} + \alpha_{2} EGR_{t} + \alpha_{3} CF_{t - 1} + \alpha_{4} CF_{t} + \alpha_{5} CF_{t + 1} \\ &\quad + \beta_{1} DS_{t} + \beta_{2} DS_{t} \times SGR_{t} + \beta_{3} DE_{t} + \beta_{4} DE_{t} \times EGR_{t} \\ &\quad + \beta_{5} DC_{t - 1} + \beta_{6} DC_{t - 1} \times CF_{t - 1} + \beta_{7} DC_{t} + \beta_{8} DC_{t} \times CF_{t} + \beta_{9} DC_{t + 1} + \beta_{10} DC_{t + 1} \times CF_{t + 1} \\ &\quad + \delta_{1} DS4_{t} + \delta_{2} SGR4_{t} + \delta_{3} DS4_{t} \times SGR4_{t} + \nu_{t} , \\ \end{aligned}$$

(6)

where SGR4 _t is percentage change in sales in the fourth quarter of year t, computed relative to the same quarter of year t − 1 to remove seasonality; DS4 _t is a dummy variable equal to 1 if SGR4 _t is negative and zero otherwise; and the remaining variables are defined previously and refer to annual data. We omit interim fiscal quarters for brevity but include them in robustness checks. Hypothesis 4 implies that the coefficient on DS4×SGR4 is positive, i.e., annual accruals exhibit an incremental asymmetry with respect to fourth quarter sales growth.

We examine the effect of successive cash losses (Hypothesis 5) using the following model:

$$\begin{aligned} ACC_{t} & = \alpha_{0} + \alpha_{1} SGR_{t} + \alpha_{2} EGR_{t} + \alpha_{3} CF_{t - 1} + \alpha_{4} CF_{t} + \alpha_{5} CF_{t + 1} \\ &\quad + \beta_{1} DS_{t} + \beta_{2} DS_{t} \times SGR_{t} + \beta_{3} DE_{t} + \beta_{4} DE_{t} \times EGR_{t} \\ &\quad + \beta_{5} DC_{t - 1} + \beta_{6} DC_{t - 1} \times CF_{t - 1} + \beta_{7} DC_{t} + \beta_{8} DC_{t} \times CF_{t} + \beta_{9} DC_{t + 1} + \beta_{10} DC_{t + 1} \times CF_{t + 1} \\ &\quad + \delta_{1} DC_{t - 1} \times DC_{t} + \delta_{2} DC_{t} \times DC_{t + 1} + DC_{t} \times (\delta_{3} CF_{t - 1} + \delta_{4} DC_{t - 1} \times CF_{t - 1} ) \\ &\quad + DC_{t - 1} \times (\delta_{5} CF_{t} + \delta_{6} DC_{t} \times CF_{t} ) + DC_{t + 1} \times (\delta_{7} CF_{t} + \delta_{8} DC_{t} \times CF_{t} ) \\ &\quad + DC_{t} \times (\delta_{9} CF_{t + 1} + \delta_{10} DC_{t + 1} \times CF_{t + 1} ) + \varsigma_{t} , \\ \end{aligned}$$

(7)

where all variables are as defined earlier. Hypothesis 5a predicts that the interaction coefficient on DC _t+1×CF _t is negative, i.e., gain recognition for current period cash flow (the full coefficient on CF _t ) is smaller when future cash flow is expected to be negative (DC _t+1 = 1). Hypothesis 5b implies that the interaction coefficient on DC _t+1×DC _t ×CF _t is positive, i.e., asymmetric loss recognition for current cash flow (the full coefficient on DC _t ×CF _t ) is greater when the current period cash loss is not expected to reverse (DC _t+1 = 1). For completeness, we also control for the parallel interactions of current period cash losses (DC _t  = 1) with lagged and future cash flows.

4 Empirical results

The estimates of models (1)–(3) are presented in Table 3. Consistent with Allen et al. (2013) and Bushman et al. (2011b), working capital accruals in the symmetric Allen et al. model (1) incorporate both a significant growth component (i.e., positive coefficients on SGR and EGR) and a significant matching component (a negative coefficient on CF _t and positive coefficients on CF _t−1 and CF _t+1).

Table 3 Estimates of the Allen et al. (2013) model and asymmetric models for annual firm-level working capital accruals

Asymmetric models have incremental explanatory power. First, we add an asymmetry for concurrent cash flow (DC _t ×CF _t in the second column), following Ball and Shivakumar (2006). The adjusted R ² increases by 1.0 percentage points (=53.8 − 52.8). We then replace it with an aggregate news measure X ≡ α ₁ SGR + α ₂ EGR + α ₃ CF _t-1 + α ₄ CF _t  + α ₅ CF _t+1, which combines all five indicators with weights α _i that best fit the dependent variable in model (2). This change increases the adjusted R ² by 0.2 percentage points (=54.0 – 53.8), which is statistically significant (Vuong Z = 1.78). Thus aggregation of multiple indicators results in a slightly more informative measure of bad news than current period cash loss alone.

Our main model (3) incorporates asymmetric effects of individual bad news indicators (DS×SGR, DE×EGR, DC _t−1×CF _t−1, DC _t ×CF _t , and DC _t+1×CF _t+1). Consistent with Hypothesis 1, this decomposition of conservatism into multiple asymmetries improves explanatory power considerably relative to model (2). The incremental adjusted R ² is 3.4 percentage points (=57.4 − 54.0), significant at the 1 % level (Vuong Z = 24.52).²⁰ Model (3) also significantly outperforms the Ball and Shivakumar model (the incremental adjusted R ² is 3.6 percentage points, Vuong Z = 30.02). Thus the accrual process for firm-level data is best approximated by asymmetric effects of the individual explanatory variables in model (3) due to the disaggregated nature of asset impairment tests. To examine whether these asymmetries reflect disaggregation across or within asset classes, we estimate our models for changes in inventory and receivables. As expected, we find significant differences between the estimates for inventory and receivables (F = 113.84 and 154.80 in models (2) and (3), respectively; untabulated). For both variables, model (3) significantly outperforms model (2) (Vuong Z = 16.66 and 16.83, respectively; untabulated). This suggests that the asset pools in impairment tests are disaggregated not only across asset classes but also within each asset class.

As expected (Hypothesis 2a), the coefficients on DS×SGR and DE×EGR in model (3) are positive and significant, indicating that accruals are more sensitive to decreases in sales and total employees than to increases. The slope coefficient on sales change SGR is 0.054 for sales increases versus 0.149 (=0.054 + 0.095) for sales decreases, i.e., the sensitivity to bad news is 176 % larger (=[0.149/0.054] − 1). The coefficient on change in total employees EGR is 0.015 for increases versus 0.053 (=0.015 + 0.038) for decreases, a difference of 253 % (=[0.053/0.015] − 1). These results are consistent with conservatism for the growth variables. Furthermore, these variables primarily capture decline accruals, such as write-downs of inventory or receivables, rather than growth accruals related to firm expansion.

The coefficient on DC _t ×CF _t indicates conservatism for concurrent cash flow (Ball and Shivakumar 2005). As predicted (Hypothesis 2b), the coefficient on DC _t+1×CF _t+1 is positive and significant, indicating quicker recognition of expected future cash losses, while the coefficient on DC _t−1×CF _t−1 is negative and significant, indicating smaller deferrals of past cash losses. Current period accrual incorporates 23.5 % (=0.181 + 0.054) of next period’s cash losses versus just 18.1 % of next period’s cash profits, i.e., early recognition is 30 % (=[23.5/18.1] − 1) greater for losses. It incorporates 27.4 % of lagged cash profits versus 18.9 % (=0.274 − 0.085) of lagged cash losses, i.e., deferred recognition is 45 % (=[27.4/18.9] − 1) greater for profits.

The relative asymmetry in the growth component (176–253 % for SGR and EGR) is much larger than that in the matching component (less than 50 % for CF _t−1, CF _t , and CF _t+1). The growth component asymmetries also have greater explanatory power than the matching component asymmetries. The incremental adjusted R ²s are 3.0 and 1.5 percentage points, respectively. These results suggest that asymmetric timeliness primarily flows through accruals related to firm growth and decline, a new effect that we predict in Hypothesis 2a, rather than through the matching component, which was the focus of prior asymmetric accrual models.

In Table 4, we examine the effect of segment information in model (5), after controlling for the asymmetric effects of all firm-level variables from our main model (3). The incremental adjusted R ² of segment sales data is 0.4 percentage points (=64.0 − 63.6). Consistent with Hypothesis 3, we find a significant positive coefficient on segDS×segSGR ⁻, which indicates asymmetric timeliness with respect to segment-level sales decreases. The size of this effect (0.069) is comparable to that for annual sales (0.071 for DS×SGR in the same model), i.e., the segment data plays an economically significant asymmetric role in firm-level accruals. In the last column, we interact the segment variables segDS and segDS×segSGR ⁻ with dummies for firm-level sales decreases and increases (DS and IS, respectively). The asymmetry for segment sales is smaller during firm-level sales decreases (DS×segDS×segSGR ⁻) than during firm-level sales increases (IS×segDS×segSGR ⁻), but both are significant. Thus, even when firm sales are decreasing, segment sales decreases add information (e.g., a large negative segSGR ⁻ can signal very poor performance for some of the segments). When firm sales are increasing, segment decreases are even more useful because they convey material contrarian information.

Table 4 Effect of segment-level sales decreases on annual firm-level working capital accruals

Table 5 presents the effect of fourth-quarter sales change SGR4 in model (6), after controlling for the piecewise-linear effects of all annual variables from our main model. Consistent with Hypothesis 4, we find a significant asymmetry for the quarterly sales change (DS4×SGR4). The incremental adjusted R ² of this variable is 0.4 percentage points (=57.6 − 57.2). The results are robust to controlling for sales changes in the interim quarters (untabulated). In the last column, we estimate the effect of SGR4 separately for annual sales decreases and increases (DS and IS, respectively). The asymmetry for SGR4 is smaller during annual sales decreases (DS×DS4×SGR4) than during annual sales increases (IS×DS4×SGR4), but both are positive and significant. Thus, even when a fourth quarter sales decrease is a part of an annual decrease, it is incrementally informative (e.g., a rapid demand decrease in the fourth quarter can signal large unrealized losses at fiscal year-end). It has an even greater impact when it provides a contrarian indicator of a more recent demand decrease, which is not visible in annual data.

Table 5 Effect of fourth-quarter sales change on annual firm-level working capital accruals

In Table 6, we examine the effect of successive cash losses in model (7). The adjusted R ² improves by 1.6 percentage points (=59.0 – 57.4), relative to our main model (3). Consistent with Hypothesis 5a, the interaction coefficient on DC _t+1×CF _t is negative and significant, i.e., gain recognition for current period cash flow (CF _t ) is smaller when future cash flow is negative (DC _t+1 = 1). When both CF _t−1 and CF _t+1 are positive, asymmetric timeliness for current period cash flow (the coefficient on DC _t ×CF _t ) is −0.085, which indicates that a one-time cash loss in period t does not trigger conservatism. Consistent with Hypothesis 5b, the interaction coefficient on DC _t+1×DC _t ×CF _t is positive and significant, i.e., asymmetric timeliness for cash flow in period t (DC _t ×CF _t ) is greater when a firm has a cash loss in period t + 1 (DC _t+1 = 1). The asymmetric timeliness for CF _t increases to 0.115 (=−0.085 + 0.200, t = 2.38), which is consistent with conservatism. These results suggest that accountants distinguish between negative timing shocks, which reverse quickly, and adverse economic shocks, which cause persistent cash losses. In other words, they respond to consistent economic patterns rather than viewing each indicator in isolation.

Table 6 Dynamic effects of successive cash losses on annual firm-level working capital accruals

The results in Tables 3, 4, 5 and 6 are robust to alternative definitions of cash flows and accruals, including measures derived from the SFAS 95 (FASB 1987) statement of cash flows, following Collins and Hribar (2002); measures that combine SFAS 95 data with earlier data from the funds flow statement, following Xie (2001); and broader accrual measures based on net income. The estimates are robust to controlling for market-to-book quintiles as a proxy for expected long-term growth (Collins et al. 2014b), realized changes in sales and employees in year t + 1 as a short-term expected growth proxy, asset market-to-book ratio above one as a control for nondiscretionary conservatism (Lawrence et al. 2013), and fiscal-year stock return as a standard news indicator (Basu 1997; Ball and Shivakumar 2006). The results hold when we discard mergers (using Compustat footnote codes) and divestitures (using discontinued operations in excess of $10,000 as a proxy), where both criteria follow Collins and Hribar (2002).²¹ The results also hold when we screen for mergers using the SDC data. The asymmetries in accruals increase significantly with the length of the operating cycle (Dechow 1994), consistent with the forward-looking nature of conservatism. We also examine alternative bad news thresholds (Basu 2005) for the growth and cash flow variables. We set the bad news indicators DS, DE, and DC to 1 if the corresponding continuous variable (SGR, EGR, and CF, respectively) is either below average or is in the bottom 20, 30, 40 or 50 % of its distribution, where the average and the percentiles are computed either for the pooled sample or separately for each two-digit industry. In all cases, the results continue to hold and are significant statistically and economically.

4.1 Tests of alternative explanations for accrual asymmetry

While accrual asymmetry is often interpreted as conditional conservatism (e.g., Basu 1997; Ball and Shivakumar 2006; Collins et al. 2014a), it could reflect asymmetric operational effects, such as cost stickiness (Banker et al. 2016b) or curtailment (Lawrence et al. 2016). In Table 7, we examine some of these alternative explanations.

Table 7 Validation tests for accrual asymmetry

Conservatism flows primarily through assets (Ijiri and Nakano 1989). We estimate model (3) for the main asset-related working capital accruals, i.e., changes in accounts receivable and inventory. We find significant asymmetries for both of these variables, consistent with conservatism (columns 1 and 2 in Panel A of Table 7).²² Asset write-downs directly capture conservatism and are not confounded by other potential asymmetries in accruals (e.g., Lawrence et al. 2013). Compustat does not separately report current asset write-downs, which flow through working capital accruals, but has data on write-downs of long-lived tangible assets and goodwill, which flow through broader operating accruals. For both of these write-down categories, we find significant asymmetric effects of the growth variables and future cash flow (columns 3 and 4 in Panel A of Table 7), which indicates that these variables trigger conservatism.

Banker et al. (2016b) show that the piecewise-linear effect of sales changes on operating accruals is partly attributable to sticky costs. Cost stickiness arises from asymmetric adjustment of physical resources such as employees and equipment (e.g., Anderson et al. 2003). It manifests as an asymmetry in operating costs, earnings, and major operating accrual components such as depreciation (Banker et al. 2016b). However, because cost stickiness affects costs rather than revenue, it cannot explain the asymmetric effect of sales changes (DS×SGR) on receivables in Panel A of Table 7. For inventory, cost stickiness predicts an asymmetry of the opposite sign from conservatism. When production costs are sticky, they fall less for sales decreases than they rise for sales increases (Anderson et al. 2003). Because the carrying value of inventory embeds these costs, it will be less sensitive to sales decreases than to sales increases (i.e., the coefficient on DS×SGR will be negative). This is contrary to our estimates for inventory (column 2 in Panel A of Table 7). Thus, while cost stickiness affects operating accruals as shown by Banker et al. (2016b), it does not explain our results for working capital accruals.

Lawrence et al. (2016) argue that the asymmetries in accruals are partly due to curtailment of underperforming operations. For example, if a firm discontinued an unsuccessful product line during the fiscal year, it likely liquidated the associated inventory and receivables. This could explain the asymmetries for the growth variables. Lawrence et al. (2016) argue that a sales decrease or an employee decrease by itself does not indicate curtailment and use simultaneous decreases in sales and employees as a proxy for curtailment. Thus, if the asymmetric effects of the growth variables on accruals primarily reflect curtailment (using Lawrence et al.’s proxy), then these asymmetries should be larger when both sales and employees are decreasing.

To test this prediction, we add higher-order interaction effects of simultaneous decreases in sales and employees (DE×DS×SGR and DS×DE×EGR) in our main model (3). The estimates are presented in Panel B of Table 7. The coefficients on both DE×DS×SGR and DS×DE×EGR are insignificant, i.e., the asymmetric effects of sales and employees (DS×SGR and DE×EGR, respectively) are not significantly associated with the curtailment proxy (DS = DE = 1).

While the results are consistent with conservatism, they do not rule out asymmetry in operations. For example, nonlinear changes in credit policy during demand decreases could generate an asymmetric effect of the growth variables on receivables even if conservatism does not play a major role. Physical inventory levels might be adjusted nonlinearly to smooth production or avoid capacity constraints. These operational decisions likely respond to forward-looking information that is also relevant for asset impairment. Furthermore, stakeholders demand conservatism because it gives managers an incentive to quickly adapt or terminate underperforming operations (e.g., Watts 2003a). Therefore conservatism could be a fundamental cause of many operational asymmetries. For example, if managers aggressively cut inventory and limit credit sales during demand decreases to avoid the risk of future write-downs, this incentive effect of conservatism can cause an asymmetry in inventory and receivables.²³ Managers might also incur additional costs to quickly adapt unprofitable projects (e.g., Collins et al. 2014a; Schrand 2014), causing further asymmetries. Thus conservatism and operational asymmetries likely co-exist and are intertwined both conceptually and empirically.

4.2 Implications for earnings management tests

Following prior studies (Dechow et al. 1995, 2012; Kothari et al. 2005; Collins et al. 2014b), we simulate earnings management tests for different accrual models. For brevity, we focus on four models: the Allen et al. model (1), the Ball and Shivakumar model, our main asymmetric model (3), and the extended asymmetric model (7) with dynamic effects of successive cash losses.²⁴

First, we examine these models’ power to detect earnings management. We randomly select 100 earnings management observations and add a discretionary accrual equal to 1 or 2 % of total assets following Dechow et al. (2012). We estimate each model and test whether the abnormal accrual (i.e., regression residual) in the earnings management years differs from zero.²⁵ We repeat all simulations 1000 times. Our asymmetric models incorporate more parameters than the Allen et al. and Ball and Shivakumar models, which could reduce test power (despite the increase in adjusted R ²) due to estimation noise. Thus, if test power improves, this would suggest that the added asymmetries in our models are sufficiently informative to outweigh this noise.

The results are presented in Panel A of Table 8. We use a significance level of 5 % in a one-tailed test. When accruals are managed upwards by 1 % of total assets for 100 firm-years, earnings management is detected in 43.9 % of simulations for the Allen et al. model and 47.1 % of simulations for the Ball and Shivakumar model. Test power improves to 48.0 % in our main asymmetric model (3) and 50.5 % in the extended asymmetric model with dynamic effects. Thus our extended model enhances the researcher’s ability to detect moderate earnings management by 15 % (=[50.5/43.9] − 1), relative to the Allen et al. model, and by 7 % (=[50.5/47.1] − 1), relative to the Ball and Shivakumar model. When accruals are manipulated by 2 % of total assets, the rejection rate is 88.0 % in the Allen et al. model, 89.8 % in the Ball and Shivakumar model, 90.4 % in our main asymmetric model, and 91.5 % in our extended asymmetric model. In other words, the proportion of false inferences (nonrejection of a false null hypothesis) is reduced by almost one-third, from 12.0 % (=100 − 88) in the Allen et al. model to just 8.5 % (=100 − 91.5) in our extended asymmetric model. We find a comparable improvement in test power when we simulate a discretionary accrual of 0.25 or 0.5 % of total assets for 2000 observations (Collins et al. 2014b).

Table 8 Simulation results for earnings management tests

We next examine type I error. We randomly select “suspected earnings management” observations from either the full sample or subsamples with extreme economic performance. We estimate the models using the original accruals data for the full sample and test for earnings management in the suspect firm-years. By construction, the earnings management dummy has no causal effect on accruals, i.e., the null hypothesis of no earnings management is true. Therefore findings of a significant abnormal accrual constitute type I error or rejection of a true null hypothesis. When the suspect firm-years are drawn from a particular subsample, they might be correlated with omitted determinants of accruals, which can result in mis-specified tests.

First, following Dechow et al. (1995, 2012), we draw 100 suspected earnings management observations from extreme performance deciles. Dechow et al. (1995) report that all of their accrual models over-reject the null hypothesis of no earnings management for firm-years in the extreme deciles of earnings. Notably, because earnings incorporates concurrent accruals, the expected abnormal accrual in extreme earnings deciles almost surely differs from zero even without any earnings management.²⁶ In other words, even if a model correctly captures all the determinants of normal accruals, it will have excessive rejection rates for firms with extreme earnings performance due to selection on the dependent variable. To assess this selection problem, we use two earnings metrics: ROA (Compustat item IB, scaled by average total assets) as in prior studies, and adjusted ROA, which is based on earnings net of working capital accruals. We also examine the rejection rates for extreme deciles of our main news indicators to determine whether our models adequately control for these observable determinants of accruals.

The type I errors are presented in Panel B of Table 8. We combine significant positive and significant negative test results because both constitute false rejection of the null hypothesis and use a two-tailed test with a 10 % significance level. When the earnings management years are selected from the full sample, the rejection rates in all models are 9.9–11.4 %, consistent with the nominal significance level. All models have rejection rates above 99 % in extreme ROA deciles, consistent with our selection bias argument. As expected, the rejection rates improve when we form the performance deciles based on adjusted ROA. The rejection rates in the bottom decile decrease to 14.4 % for the Allen et al. model, 12.7 % for the Ball and Shivakumar model, 10.5 % for the main asymmetric model, and 8.3 % for the extended asymmetric model. In the top decile of adjusted ROA, all models over-reject the null hypothesis, indicating a confounding effect of correlated omitted variables. However, our asymmetric models partly mitigate the over-rejection, yielding type I error of 50.2–58.1 %, versus 97.5 % for the Allen et al. model and 64.6 % for the Ball and Shivakumar model.

Our piecewise-linear models are likely better specified in extreme deciles of the growth and cash flow variables than the linear Allen et al. model and the partly linear Ball and Shivakumar model. As expected, the rejection rates for our asymmetric models (3) and (7) are generally consistent with the nominal significance level, and even the largest rejection rate is just 17.2 %. In contrast, the symmetric model has a rejection rate of 94.5 % in the bottom sales growth decile, 74.8 % in the bottom employee growth decile, and 49.1 % in the top cash flow decile. The Ball and Shivakumar model incorporates asymmetry only for cash flow. As expected, it has valid rejection rates for the cash flow deciles but over-rejects considerably in the extreme growth deciles (e.g., the rejection rate is 93.1 % in the bottom sales growth decile and 75.1 % in the bottom employee growth decile). Thus a researcher should incorporate asymmetries for all indicators to avoid type I error due to mis-specified linear functional form.²⁷

We next examine type I errors in the simulation protocol suggested by Collins et al. (2014b), who argue that the treatment sample should be larger and more heterogeneous. They simulate a suspect subsample of 2000 observations, half of which are selected from a given extreme performance partition and the other half from the remainder of the sample. The results are presented in Panel C of Table 8. Consistent with our selection bias argument, all models have a rejection rate of 100 % for the extreme ROA partitions. For partitions based on the growth variables, the rejection rates in our asymmetric models are 9.2–15.4 %. In contrast, the parallel rejection rates in both the Allen et al. model and the Ball and Shivakumar model are 76.3–100 %, indicating over-rejection due to functional form mis-specification.

5 Conclusion

We examine in depth the implications of conditional conservatism for accrual research. Study of abnormal accruals requires an accurate benchmark model of the normal accrual process (Ball 2013), which should incorporate the main features of accounting practice. Conservatism has been described as “the most ancient and probably the most pervasive principle” in accounting practice (Sterling 1967, p. 110). We examine the accounting guidance for working capital accounts (ASC topics 310 and 330 for receivables and inventory, respectively; previously based on ARB 29, 30, and 43) and show that these standards incorporate asymmetric treatment of unrealized losses versus unrealized gains for small asset pools. Many unrealized losses are recognized early as asset write-downs (i.e., negative accruals), whereas unrealized gains are not recognized as asset write-ups. We argue that the standard explanatory variables in accrual models signal future gains and losses for disaggregated asset pools and predict that conservatism in firm-year data is best approximated by a sum of asymmetric effects of individual news indicators (rather than an asymmetry for an aggregate news measure). We argue that segment-level and quarterly indicators have incremental explanatory power for annual firm-level accruals. Because accountants distinguish temporary and permanent cash losses, we predict a dynamic effect of successive negative cash flows.

Estimates for U.S. Compustat/CRSP data are consistent with our predictions. While Ball and Shivakumar (2006) document asymmetric timeliness of accruals with respect to concurrent cash flow (and additional indicators in some of the tests), we focus on how different firm-level and disaggregated indicators should be incorporated in accrual models. The results support our disaggregated-information argument for both firm-level indicators and more detailed segment-level and quarterly indicators. Our improved accrual models also have greater statistical power and lower type I error in earnings management tests.

Recent advances in empirical accrual research, such as Allen et al. (2013) and Bushman et al. (2011b), examine economic drivers of accruals. In contrast, we focus on disaggregated asymmetries in accruals to develop new insights and improved empirical tests for a variety of research settings. While we attribute the results to conservatism (and rule out some alternative explanations in validation tests), they could partly reflect asymmetries in firms’ operations. Further, conservatism can cause operational asymmetries by giving managers an incentive to quickly adapt or terminate unsuccessful projects, while operational decisions can affect future cash flows that are relevant for conservatism. Thus it is conceptually difficult to fully disentangle conservatism from operational effects. While accrual asymmetry can have alternative interpretations, we show that the default linear specification of accrual models is unjustified both theoretically and empirically.

Because our asymmetric models have more parameters that the standard accrual models, researchers should exercise judgment to avoid unfocused variable proliferation (Roychowdhury and Martin 2013). For example, if a researcher seeks to identify new determinants of normal accruals, then a more parsimonious linear model might be preferred for expositional convenience (if the results are robust). If a researcher aims to provide credible evidence of earnings management (or its absence), then high statistical power and low type I error likely matter more than model parsimony. The evidence in Sect. 4.2 suggests that our asymmetric models perform better than the standard models in earnings management tests, but a researcher could (and probably should) conduct simulations for the specific empirical context to identify the most appropriate model. Similarly, if different metrics such as absolute versus signed discretionary accruals produce conflicting results (e.g., Hribar and Nichols 2007), a researcher could examine whether a more extensive asymmetric model resolves the conflict.²⁸ Future research should consider the asymmetric nature of accruals.