Management of the loss reserve accrual and the distribution of earnings in the property-casualty insurance industry $

Transcription

1 Journal of Accounting and Economics 35 (2003) Management of the loss reserve accrual and the distribution of earnings in the property-casualty insurance industry $ William H. Beaver, Maureen F. McNichols, Karen K. Nelson* Department of Accounting, Graduate School of Business, Stanford University, Stanford, CA 94305, USA Received 2 March 2002; received in revised form 14 January 2003; accepted 27 January 2003 Abstract We document that property-casualty insurers with small positive earnings understate loss reserves relative to insurers with small negative earnings. Furthermore, loss reserves are managed across the entire distribution of earnings, with the most income-increasing reserve accruals reported by small profit firms, and the most income-decreasing reserve accruals reported by firms with the highest earnings. We analyze this pattern separately for public, private, and mutual companies, and find that public companies and mutuals manage loss reserves to avoid losses, but that private companies do not. We also find evidence of reserve management to avoid losses by financially healthy and distressed firms. r 2003 Elsevier B.V. All rights reserved. JEL classification: M41; G22; G28 Keywords: Earnings management; Earnings distribution; Discretionary accruals; Property-casualty insurance $ We appreciate the helpful comments and suggestions of Jerry Zimmerman (the editor), an anonymous reviewer, and workshop participants at the American Accounting Association Annual Meeting, the 11th Annual Conference on Financial Economics and Accounting, Athens University of Economics and Business, Columbia University Burton Workshop, Harvard University Kennedy School of Government, Indiana University, MIT, University of Oregon, Stanford University, and Syracuse University. Research assistance was provided by Qintao Fan and Yvonne Lu. We gratefully acknowledge the support of the Stanford Graduate School of Business Financial Research Initiative. *Corresponding author. Present address: Department of Accounting, Jones Graduate School of Management, Rice University, Houston, TX 77005, USA; Tel.: ; fax: address: nelsonk@rice.edu (K. Nelson) /03/$ - see front matter r 2003 Elsevier B.V. All rights reserved. doi: /s (03)

2 348 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) Introduction This study examines the relation between discretionary loss reserve accruals and the distribution of reported earnings for a sample of property-casualty (P&C) insurers. Two distinctive features of the P&C insurance setting motivate studying earnings management in this sector. First, required disclosures for a material accrual, the claim loss reserve, allow us to develop a relatively reliable measure of management s exercise of discretion over earnings. Second, a variety of ownership structures exist within the industry, including public, private, and mutual companies. This variation allows us to examine more directly than in most prior research how incentives inherent in different ownership structures affect earnings management behavior. Our first research objective is to examine the relation between management of the loss reserve accrual and the distribution of reported earnings. For a sample of non-insurance firms, Burgstahler and Dichev (1997), hereafter BD, and Degeorge et al. (1999) find a discontinuity in the distribution of earnings in the region immediately around zero, which they interpret as evidence that firms manage earnings to avoid reporting losses. However, lacking a model of accruals or an estimate of discretionary accruals, there is ambiguity in interpreting these results. We document that P&C insurers report small positive earnings with greater frequency than expected given the relative smoothness of the remainder of the earnings distribution, and that these firms significantly understate the loss reserve accrual relative to firms with small negative earnings. Thus, our evidence is consistent with P&C firms managing accruals to avoid losses. We also provide evidence on management s exercise of discretion over the entire earnings distribution rather than only in the area immediately around zero. The magnitude and frequency of earnings management is of significant interest to standard setters and regulators, but existing research provides little relevant evidence (Healy and Wahlen, 1999; Dechow and Skinner, 2000). Our analysis indicates that earnings management occurs across the entire distribution of reported earnings, with earnings management by small profit firms accounting for only a fraction of total earnings management activity. Specifically, we find that the least profitable firms understate reserves relative to the most profitable firms. This evidence is consistent with P&C firms managing loss reserves to smooth earnings rather than to take an earnings bath. Our second research objective is to examine the effect of ownership form and financial condition on management of the loss reserve accrual and the distribution of earnings. Because the extent of owner manager and owner policyholder conflicts varies across public, private, and mutual insurers (Mayers and Smith, 1988, 1994), incentives to manage earnings are also likely to vary. 1 For public and mutual insurers, we find that small profit firms significantly understate loss reserves relative 1 Concurrent research by Beatty et al. (2002) examines whether there is a discontinuity in the distribution of earnings changes for public and private banks.

3 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) to small loss firms. In contrast, there is no evidence that private insurers manage reserves to avoid reporting losses. For all three ownership forms, there is evidence of income smoothing, as loss reserves are significantly overstated by the most profitable firms. Taken together, our results show that managing reserves to avoid losses and smooth earnings is not unique to public companies. Finally, we examine whether insurers financial condition affects reserve management across the distribution of reported earnings. Following prior research (e.g., Petroni, 1992), we partition our sample according to whether the firm is financially healthy or financially distressed. Holding financial condition constant, we find that management of reserves to avoid losses is more pronounced in the sample of healthy insurers. However, both the sample of financially distressed insurers and the sample of financially healthy insurers understate loss reserves in the left tail of the earnings distribution relative to the right tail of the distribution. However, because distressed insurers occur more frequently in the left tail of the distribution, they tend to report loss reserves that are understated, on average, relative to healthy insurers. Our study makes several contributions to the literature on earnings management in general, and on insurers management of the loss reserve accrual in particular. First, we provide direct evidence that P&C firms who avoid reporting small losses do so by managing loss reserves. This finding also documents a form of discretionary behavior management of the reserve accrual to avoid losses not previously known to exist for P&C firms. Second, our results indicate that earnings management is not concentrated in these firms, but instead is pervasive across all levels of earnings. Our evidence thus extends Beaver and McNichols (1998) who find wide-spread evidence of positive serial correlation in discretionary loss reserve accruals. Finally, our results complement the literature focusing on earnings management by public firms by documenting that similar patterns of earnings management are observed for firms that are not public. The layout of the paper is as follows. Section 2 provides an overview of the accounting for claim loss reserves which underlies our estimate of management s discretionary accrual behavior. Section 3 develops our hypotheses. Section 4 describes the sample and data. Section 5 presents our empirical findings. Section 6 concludes. 2. Estimating discretionary loss reserve accruals Claim loss reserves are generally the largest liability on a P&C insurer s balance sheet, and the income effect of the related provision is substantial. In our sample, the loss reserve liability is 56 percent of total liabilities, and the provision for losses is 71 percent of premium revenue. The matching principle requires insurers to charge claim losses to operations in the period they are incurred and the related premium revenue is recognized. Although some claims are settled in the year incurred, the majority will remain outstanding for several years. The uncertainty surrounding the estimation of the cost to settle incurred but unpaid claims provides an opportunity

4 350 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) for substantial earnings management because understating (overstating) the reserve accrual increases (decreases) reported earnings. 2 As information becomes available regarding prior period claims, insurers revise their original estimate of loss reserves with a charge to current period operations. After all claims for a period have been settled, policy losses for that period are known with certainty. Insurers are required to disclose the year-by-year revisions, called loss reserve development, for each of the past 10 years. Using the revised estimate as a proxy for the unbiased expectation of policy losses, we calculate the discretionary loss reserve accrual as follows: 3 DEV tþj;t ¼ RESERVE tþj;t RESERVE t;t =ASSETSt ð1þ where RESERVE t,t is the loss reserve for year t reported in year t, and RESERVE tþj;t is the revised estimate of the year t loss reserve reported in year t þ j: As in Petroni (1992), we scale loss reserve development by total (admitted) assets in year t ðassets t Þ: 4 We calculate DEV tþj;t for a 5 year development period (i.e., j=5), if available; otherwise, we calculate DEV t+j,t using the most recent revised estimate (i.e., j=1 to 4). 5 In the absence of discretion, the expected value of DEV conditional on earnings in year t is zero. This is because loss reserve development is analogous to a forecast error. If a forecast is unbiased, then the subsequent forecast error has an expected value of zero. Thus, DEV is a function of ex post surprises (which, by definition, have an expected value of zero) and the effects of discretion. Essentially, discretion causes a non-zero expected value. Moreover, loss reserve development does not possess any tendency for reversal, as do accruals such as accounts receivable. Rather, loss reserve development measures the reversal that will occur in the reserve accrual due to misstatements in the originally reported reserve. Specifically, DEV is positive (negative) if originally reported reserves are understated (overstated). 2 Although we measure discretion using the loss reserve accrual (i.e., the liability) rather than the provision for claim losses (i.e., the expense), errors in the reserve accrual directly affect the computation of earnings. However, if an insurer understates (or overstates) the reserve accrual by a constant amount each year, then earnings after the first year will be the same as they would have been if no error had occurred, although they will still be misstated relative to what they would have been if the insurer were to report the correct reserve accrual in subsequent years. Nevertheless, in our sample the error in reserves varies across years for all but a few insurers. 3 We suppress firm subscripts to simplify the notation. See the appendix for an illustration of the required claim loss disclosures and a numerical example of the calculation of loss reserve development. 4 Certain economic resources recognized as assets under Generally Accepted Accounting Principles (GAAP) are required by Statutory Accounting Practices (SAP) to be excluded from the balance sheet, and are thus referred to as non-admitted assets. Non-admitted assets are generally non-liquid items such as property, plant, and equipment, furnishings and supplies, and leasehold improvements. Henceforth, references to total assets should be interpreted as total admitted assets. Our results are robust to scaling by the developed loss reserve, RESERVE tþj;t. 5 We also calculated DEV t+j,t using the longest development period available (maximum of 10 years). The untabulated findings using this measure of discretionary loss reserves are consistent with the results reported below.

5 3. Hypotheses W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) Based on a higher than expected frequency of slightly positive earnings, BD and Degeorge et al. (1999) conclude that firms manage earnings to avoid losses. However, neither paper directly relates the clustering of observations immediately above zero to a measure of discretionary behavior. We hypothesize that P&C insurers manage the loss reserve accrual to avoid reporting losses. Although there are other discretionary choice variables available to insurers attempting to avoid a small loss (e.g., security gains, management of other accruals), the magnitude of the loss reserve accrual and the considerable subjectivity inherent in its estimation suggest that it is the primary means of exercising discretion in this industry. Furthermore, neither auditors (Petroni and Beasley, 1996) nor regulators (Gaver and Patterson, 2002) appear to be effective in detecting and/or mitigating this form of managerial discretion. Thus, we test the following hypothesis, stated in null form: H1: P&C insurers do not manage the loss reserve accrual to avoid losses. Stated in terms of our measure of discretionary loss reserve accruals, the null hypothesis is that there is no difference in the loss reserve development of small profit and small loss firms. The alternative hypothesis is that loss reserve development is significantly higher for small profit firms relative to small loss firms. The first hypothesis focuses on a narrow range of the earnings distribution immediately around zero. However, there is relatively little prior evidence about the magnitude and pervasiveness of earnings management across the entire earnings distribution. Healy (1985) and McNichols and Wilson (1988) provide evidence consistent with firms recognizing income-decreasing discretionary accruals when performance is either weak or strong, and income-increasing accruals at intermediate levels of performance. However, the magnitude of managers discretion over earnings is not easily identified from these studies. The proxy for discretionary accruals used in the first of these studies, the change in accruals, likely includes a significant non-discretionary component, while the proxy in the second study, the residual provision for bad debts, likely reflects only a small portion of managements discretion over earnings. We test for evidence of pervasive earnings management using a discretionary accrual measure that we expect is less likely to contain a nondiscretionary element and more likely to be comprehensive than discretionary accrual measures used outside the P&C insurance industry. Our second hypothesis, stated in null form, is: H2: P&C insurers do not manage the loss reserve accrual outside the region immediately above zero. Stated in terms of our measure of discretionary loss reserve accruals, the null hypothesis is that loss reserve development is zero in all earnings intervals. The alternative hypothesis is that loss reserve development is non-zero in at least one earnings interval outside the interval immediately above zero. The comprehensive alternative hypothesis captures two principal earnings management incentives,

6 352 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) income smoothing and taking a bath. Income smoothing implies that loss reserve development is significantly higher for the least profitable firms relative to the most profitable firms, while bath-taking implies that loss reserve development is significantly lower for the least profitable firms. Incentives to manage earnings likely vary with the ownership structure of the firm. A distinctive feature of our sample is that it includes public, private, and mutual companies. Mayers and Smith (1988, 1994) argue that incentive and agency issues differ across these ownership structures because of differences in how each structure combines the management, ownership, and policyholder (in effect, debt holder) functions. According to Mayers and Smith, the potentially complete separation of these functions in public companies raises issues of incentive alignment between owners and managers as well as between owners and policyholders. In private companies, the owner manager conflict is mitigated by the ability of a concentrated set of owners to monitor management actions at a relatively low cost; however, the owner policyholder conflict is still present. Finally, in a mutual company, the owner and policyholders functions are merged, thus mitigating this type of conflict. However, Mayers and Smith argue that owner manager conflicts could be aggravated in mutual companies because of the lack of takeover threat. 6 Because each ownership form faces diverse incentive issues, and because earnings management can arise in response to either owner manager or owner policyholder conflicts, the incentives for earnings management may vary across ownership forms. P&C insurers have two additional incentives to manage earnings, taxation and regulation. Because the claim loss provision is tax deductible, insurers, particularly profitable ones, have an incentive to overstate the loss reserve accrual. However, the Tax Reform Act of 1986 largely eliminated differences in the taxation of mutuals and stock companies (Mayers and Smith, 1994). Industry regulation focuses on the financial solvency of insurers and on the rates they are permitted to charge. Although each state has an insurance commission, regulatory efforts are coordinated through the National Association of Insurance Commissioners (NAIC). Furthermore, solvency and rate regulation apply to all insurers, and we are aware of no evidence to suggest that this varies with organizational form. Thus, although we expect tax status and regulation to influence earnings management behavior of P&C insurers, we do not expect these factors to induce differences between insurers with different organizational forms. This discussion motivates our third set of hypotheses, also stated in null form: H3a: Public insurers do not manage the loss reserve accrual to avoid losses or to manage earnings outside the region immediately around zero. H3b: Private insurers do not manage the loss reserve accrual to avoid losses or to manage earnings outside the region immediately around zero. H3c: Mutual insurers do not manage the loss reserve accrual to avoid losses or to manage earnings outside the region immediately around zero. 6 Although proxy fights to remove existing management are a potential control mechanism, they are expensive and ineffective, and thus are rarely used in practice (Mayers and Smith, 1994).

7 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) Because of the diverse incentive issues discussed above, we do not offer specific alternative hypotheses regarding the pattern of reserve development for each form of ownership. Finally, we examine whether financial condition, determined using the NAIC system for monitoring financial health, is related to incentives to manage earnings, both within a narrow interval around zero and over the entire earnings distribution. Prior evidence (e.g., Petroni, 1992) suggests that financially distressed insurers significantly understate loss reserve accruals relative to financially healthy insurers. However, there is no evidence to indicate whether this behavior is uniform over the entire earnings distribution. Thus, we test the following set of hypotheses, stated in null form: H4a: Financially healthy insurers do not manage the loss reserve accrual to avoid losses or to manage earnings outside the region immediately around zero. H4b: Financially distressed insurers do not manage the loss reserve accrual to avoid losses or to manage earnings outside the region immediately around zero. If financial distress motivates earnings management, then the pattern of reserve development discussed for H1 and H2 will be evident for distressed insurers but not for healthy insurers. 4. Sample and data The primary data source is the NAIC Property-Casualty Annual Statement Database. 7 The sample is selected using the following criteria: (i) the firm is a stock or mutual company domiciled in the United States, (ii) the firm is not primarily a reinsurer (i.e., direct premiums written are greater than premiums assumed), 8 (iii) net income and total assets are available, (iv) the firm reports positive loss reserves, and (v) at least one calendar year of loss reserve development is available. This selection process yields 13,807 firm-year observations for To control for extreme errors in the loss reserve accrual, we exclude observations with an original loss reserve estimate that differs from the revised estimated by greater than 50 percent in absolute value. We also exclude observations in the upper or lower one percent of the DEV distribution. The final sample consists of 11,460 firm-year observations. We control for size differences across observations by scaling net 7 Data Source: National Association of Insurance Commissioners, by permission. The NAIC does not endorse any analysis or conclusions based upon the use of its data. 8 Loss reserves are generally determined by the insurer that originally writes the business. Thus, firms that are primarily reinsurers have less discretion over the reported loss reserve.

8 354 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) income by total assets. However, we also calculated our primary results scaling by policyholders surplus and earned premiums, with similar results. Table 1, Panel A presents descriptive statistics for scaled values of reported earnings. The distribution of scaled earnings is relatively stable during the sample period, with a sample-wide mean (median) of 2.5 (2.9) percent of total assets. Our sample consists of relatively fewer loss firms than BD who report negative earnings at the first quartile in most of their sample years. 9 This finding is likely due to solvency regulation in the P&C industry which reduces the likelihood that a firm will report a loss. Table 1, Panel B presents descriptive statistics for loss reserve development (DEV). The results indicate that loss reserves are overstated a mean (median) of 0.2 (0.7) percent of total assets during the sample period. However, there is substantial crosssectional variation; loss reserves are overstated 3.9 percent at the 25th percentile but understated 2.2 percent at the 75th percentile. Loss reserve development also varies across the sample period, with reserves in the early years generally more understated than in the later years. This calendar year pattern is consistent with evidence documented in Beaver and McNichols (1998). Table 1, Panel C presents descriptive statistics for other firm characteristics. Loss firms (LOSS) comprise 19.7 percent of the sample, while financially distressed firms (DISTRESS) account for 35.3 percent. As in prior research (e.g., Petroni, 1992), we identify insurer-years as financially distressed if more than one non-reserve ratio used by the NAIC in the Insurance Regulatory Information System (IRIS) is outside the range considered acceptable. Mean (median) total assets (ASSETS) are approximately $288 ($44) million, indicating that the size distribution of the sample is skewed by some particularly large insurers. Approximately two-thirds of the sample write lines of business with a settlement period, or tail (TAIL), of at least 5 years. Relative to total premiums, 11.1 percent of premium revenue, on average, is attributable to workers compensation (WC), 13.0 percent to liability lines (product liability, other liability, and medical malpractice) (LIABILITY), and 35.8 percent to auto lines (AUTO). 5. Results 5.1. Discretionary loss reserve accruals and the distribution of reported earnings Fig. 1 shows the distribution of scaled earnings for all firms in the sample using interval widths of Assuming a smooth probability distribution, the expected number of observations in an interval is the average of the two adjacent intervals. The difference between the actual and expected number of observations, divided by 9 Only 1 year in our sample, 1996, has negative earnings at the first quartile. Despite the slight clustering of loss observations, inferences are unchanged when we exclude 1996 observations from the analyses. 10 Consistent with Degeorge et al. (1999), the interval width is approximately twice the interquartile range of scaled earnings times the negative cube root of sample size.

9 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) Table 1 Descriptive statistics for sample of 11,460 observations for the years Year N Mean Std. Dev. 25% 50% 75% Panel A: Scaled values of reported earnings Total 11, Panel B: Loss reserve development (DEV) Total 11, Panel C: Other firm characteristics LOSS 11, DISTRESS 11, ASSETS 11, TAIL 11, WC 11, LIABILITY 11, AUTO 11, Reported earnings is net income, scaled by total assets; DEV is the difference between the developed and originally reported loss reserve, scaled by total assets; LOSS is an indicator variable equal to one if net income less than zero, and zero otherwise; DISTRESS is an indicator variable equal to one if the firm has more than one IRIS ratio outside the range considered acceptable by the NAIC, and zero otherwise; ASSETS is the dollar amount of total assets, in millions; TAIL is an indicator variable equal to one if the firm requires 5 or more years for 75 percent of claims to be paid, and zero otherwise; LIABILITY is net premiums earned for product liability, other liability, and medical malpractice as a percentage of total net premiums earned; WC is net premiums earned for workers compensation as a percentage of total net premiums earned; and AUTO is net premiums earned for private and commercial auto as a percentage of total premiums earned.

10 356 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) Frequency Earnings Interval Fig. 1. The distribution of reported net income, scaled by total assets. The distribution interval widths are 0.006, and the location of zero on the horizontal axis is marked by the vertical line. The first interval to the right of zero contains observations in the interval [0.000, 0.006), the second interval contains [0.006, 0.012), and so forth. the estimated standard deviation of the difference, indicates whether there is a significant discontinuity in a particular interval. Under the null hypothesis of a smooth distribution, these standardized differences are distributed approximately Normal with mean 0 and standard deviation 1. As expected, we find a significant discontinuity in the earnings distribution around zero. Earnings slightly greater than zero occur more frequently than expected given the relative smoothness of the remainder of the distribution (the standardized difference for the first interval above zero is 4.33). 11 To provide evidence on whether insurers manage the loss reserve accrual to avoid reporting losses, we examine the distribution of DEV conditional on the level of scaled earnings. We sort the observations on scaled earnings, and form equal-sized portfolios of 600 observations. The size of the portfolios approximates the number of observations in the first interval above zero in Fig The portfolio boundaries are defined relative to zero. Under the null hypothesis of no earnings management, DEV is expected to be zero in each of the earnings portfolios. If insurers understate loss reserves to avoid reporting losses, DEV will be significantly higher in the portfolio immediately above zero compared to the portfolio immediately below zero. Fig. 2 shows the conditional distribution of median DEV. Consistent with predictions, there is an upward shift in DEV in the portfolio immediately to the right of zero. Untabulated statistics indicate that the median difference in DEV between 11 In untabulated analyses, we also find evidence of a discontinuity at zero in the distribution of earnings changes (the standardized difference is 2.91), although, consistent with BD, this result is not as significant as the discontinuity in the distribution of earnings levels. 12 We form equal-sized portfolios rather than use the equally spaced earnings intervals in Fig. 1 to reduce the impact of outliers on intervals in the tails of the distribution that have a small number of observations.

11 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) Median DEV Earnings Portfolio Fig. 2. The distribution of the median percentage loss reserve development, conditional on the magnitude of scaled earnings. DEV is the difference between the developed and originally reported loss reserve, divided by total assets. Each portfolio contains 600 observations based on the magnitude of scaled earnings relative to zero. the small loss and small profit portfolios is significant at the 0.01 level. Although not predicted, we also find that median DEV is significantly higher at the 0.01 level in the second portfolio to the right of zero. Over the entire earnings distribution, these are the only two portfolios in which the median loss reserve is understated. There is also evidence in Fig. 2 of an overall negative association between earnings and DEV. The correlation between median portfolio earnings and DEV is negative at the 0.01 level. This finding is consistent with income smoothing; loss reserve accruals are more overstated at higher levels of earnings. Untabulated findings reveal a similar pattern in the mean loss reserve development, except that mean DEV is positive in the left tail of the earnings distribution. However, inferences remain the same; mean DEV is significantly higher in the two portfolios to the right of zero than in the portfolio immediately to the left of zero, and there is a significant negative correlation between earnings and mean DEV. To further test the relation between management of the loss reserve accrual and the distribution of earnings, we estimate the following model: DEV tþj;t ¼ a t þ b 1 NEGATIVE t þ b 2 BELOW t þ b 3 ABOVE t þ b 4 POSITIVE t þ b 5 TAIL t þ b 6 TAILSQ t þ b 7 WC t þ b 8 LIABILITY t þ b 9 AUTO t þ e t : ð2þ The intercept varies across years because of calendar year differences in loss reserve development (see Table 1), which may reflect common non-discretionary events that ex post cause the average development to be non-zero in a given year. However, the year-specific intercepts may extract discretionary elements of loss reserve development as well. Therefore, we also estimate Eq. (2) without the year effects to examine the sensitivity of the results. The first four regressors in Eq. (2) are indicator variables

12 358 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) identifying the position of an observation in the earnings distribution, i.e., to the left of the first portfolio below zero (NEGATIVE), in the portfolio immediately below zero (BELOW), in the portfolio immediately above zero (ABOVE), and to the right of the first portfolio above zero (POSITIVE). The remaining variables in Eq. (2) are firm-specific determinants of loss reserve development. Insurers that primarily write long-tailed lines of business have greater flexibility to manage loss reserves than do insurers that primarily write short-tailed lines of business. TAIL is an indicator variable equal to one if the firm requires 5 or more years for 75 percent of claims to be paid. Nelson (2000) reports that loss reserves contain an implicit discount to present value, implying that development is increasing in the length of the settlement period, but at a decreasing rate. Thus, we also include the square of the settlement period (TAILSQ). Reserve development also likely varies across lines of business. Following Petroni et al. (2000), we control for three types of business, each measured as the proportion of premiums earned for that business relative to total premiums earned. Loss reserves for workers compensation (WC) policies are typically discounted to present value, and thus we expect positive reserve development for this line. LIABILITY captures lines of business prone to exogenous ex post shocks, i.e., product liability, other liability, and medical malpractice. AUTO captures lines of business less subject to exogenous ex post shocks, i.e., private and commercial auto. Petroni et al. (2000) find little evidence of a relation between reserve development and LIABILITY, but a generally negative relation between reserve development and AUTO. Table 2 reports regression summary statistics (Panel A) and tests of coefficient restrictions (Panel B). We report results for three specifications of Eq. (2): (i) including only the earnings distribution indicator variables; (ii) including the year intercepts and the earnings distribution indicator variables; and (iii) the full model in Eq. (2). 13 The first estimation parallels the analysis in Fig. 2 and allows us to assess the unconditional magnitude of discretionary reserve accruals in each of the four regions of the earnings distribution. Under the null hypothesis of no discretion in any of the earnings portfolios, the coefficient estimates on all four indicator variables are expected to be zero. Including the year intercepts in the second estimation allows us to assess the magnitude of reserve development in each region of the earnings distribution conditional on the average yearly reserve development. For parsimony, we do not tabulate the coefficient estimates on the year intercepts. The final specification provides additional evidence on the robustness of the earnings distribution results to other determinants of loss reserve development. The results from the estimation of the first specification indicate that loss reserves are significantly understated in the left tail of the earnings distribution (NEGATIVE coefficient estimate=0.01, p-value=5.02) and in the region immediately above zero (ABOVE coefficient estimate=0.02, p-value=5.20), and are significantly overstated in the right tail of the distribution (POSITIVE coefficient 13 We examined all estimations for statistical outliers but none were detected. For comparability, we present results for the sample of observations with all data available to estimate the full model stated in Eq. (2).

13 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) Table 2 Regression of loss reserve development on earnings portfolios DEV tþj;t ¼ a t þ b 1 NEGATIVE t þ b 2 BELOW t þ b 3 ABOVE t þ b 4 POSITIVE t þ b 5 TAIL t þ b 6 TAILSQ t þ b 7 WC t þ b 8 LIABILITY t þ b 9 AUTO t þ e t Panel A: Regression summary statistics Excl. year intercepts Including year intercepts Variable Coefficient t-statistic Coefficient t-statistic Coefficient t-statistic NEGATIVE BELOW ABOVE POSITIVE TAIL TAILSQ WC LIABILITY AUTO Adjusted R N 11,310 11,310 11,310 Panel B: Tests of coefficient restrictions a Restriction F-statistic p-value F-statistic p-value F-statistic p-value All equal o o o0.01 ABOVE=BELOW 8.57 o o o0.01 NEGATIVE=POSITIVE o o o0.01 DEV is the difference between the developed and originally reported loss reserve, divided by total assets. NEGATIVE is equal to one if the observation is to the left of the first portfolio below zero; BELOW is equal to one if the observation is in the first portfolio below zero; ABOVE is equal to one if the observation is in the first portfolio above zero; and POSITIVE is equal to one if the observation is to the right of the first portfolio above zero. Each portfolio contains 600 observations, based on the magnitude of scaled earnings relative to zero. The control variables are defined as follows: TAIL is an indicator variable equal to one if the firm requires 5 or more years for 75 percent of claims to be paid, and zero otherwise; TAILSQ is the square of the number of years required for 75 percent of estimated claims to be paid; LIABILITY is net premiums earned for product liability, other liability, and medical malpractice as a percentage of total net premiums earned; WC is net premiums earned for workers compensation as a percentage of total net premiums earned; and AUTO is net premiums earned for private and commercial auto as a percentage of total net premiums earned. a p-values for the test of the restriction ABOVE=BELOW are one-sided; all others are two-sided. estimate= 0.01, p-value= 7.44). Moreover, small profit firms significantly understate loss reserves relative to small loss firms. The tests reported in Panel B formally reject the null hypothesis that ABOVE=BELOW at less than the 0.01 level. Conditioning on the average yearly reserve development increases the coefficient estimates on the earnings distribution indicator variables, consistent with the samplewide negative reserve development reported in Table 1. However, we continue to find

14 360 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) incremental reserve management associated with the distribution of reported earnings. Development is positive and significant in all regions of the earnings distribution, but the magnitude of the reserve understatement is greatest for small profit firms and for firms in the left tail of the earnings distribution. Moreover, the difference in coefficient estimates between the ABOVE and BELOW portfolios remains significant at less than the 0.01 level. In addition, the least profitable firms understate reserves relative to the most profitable firms, allowing us to reject the null hypothesis that NEGATIVE=POSITIVE at less than the 0.01 level. This evidence is consistent with firms exercising discretion over the loss reserve accrual to smooth earnings. The last two columns in Table 2 report the estimation of the full model in Eq. (2). The results for the earnings distribution indicator variables are virtually unchanged from the previous model, indicating that the relation between discretionary loss reserve accruals and the distribution of earnings is robust to other significant determinants of loss reserve development. Although not the focus of our analysis, the coefficient estimates on the control variables are generally consistent with expectations. To summarize, in all three specifications reported in Table 2 we reject the null hypothesis that ABOVE=BELOW at less than the 0.01 level, consistent with firms managing reserves to avoid reporting a loss. In all three specifications we also reject the null hypothesis that NEGATIVE=POSITIVE at less than the 0.01 level, consistent with income smoothing. To further assess the relation between loss reserve development and the distribution of reported earnings, we estimate a variant of Eq. (2) that allows for separate coefficient estimates on each of the earnings portfolios. We estimate this model for the three specifications in Table 2, and plot the earnings portfolio coefficient estimates in Fig. 3. The pattern of loss reserve development across the earnings distribution is strikingly similar in all three specifications. All three models reveal an upward shift in DEV in the two portfolios to the right of zero, and an overall negative association between DEV and reported earnings. The coefficient estimates for the model including only the year intercepts as controls and the model including all controls are nearly identical in all portfolios, and differ from the coefficient estimates in the no controls specification by an amount that is relatively constant across all portfolios. Thus, the results reveal a systematic pattern of loss reserve development across the earnings distribution that is robust to several control variables. 14 If the discontinuity in the earnings distribution is induced by discretion, then the distribution of pre-managed earnings should be smooth around zero. We calculate pre-managed earnings by subtracting the discretionary loss reserve 14 To examine the effect of tax status on loss reserve development, we define a dummy variable equal to one for firms with high tax rates, identified, as in Petroni (1992), as insurers that are currently paying taxes. We find that although insurers with high tax rates understate reserves less than insurers with low tax rates, inclusion of the tax control variable in Eq. (2) does not alter our findings concerning the pattern of loss reserve development across the earnings distribution (results not tabulated). However, estimating Eq. (2) separately for the high and low tax subsamples reveals that overstatement of loss reserves in the right tail of the earnings distribution is primarily due to insurers with high average tax rates.

15 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) Coefficient Estimates Earnings Portfolio all controls year intercepts no controls Fig. 3. Coefficient estimates from a regression of DEV on earnings portfolio indicator variables (PORTFOLIO), year intercepts, and control variables for firm-specific determinants of loss reserve development, as follows: X DEV tþj;t ¼ a t þ g i PORTFOLIO i;t þ b 1 TAIL t þ b 2 TAILSQ t þ b 3 WC t i þ b 4 LIABILITY t þ b 5 AUTO t þ e t The variables are defined in Table 2. Results are plotted for three specifications including: (i) all controls, (ii) only the year intercepts, and (iii) no controls. accrual for year t (i.e., RESERVE tþj;t RESERVE t;t ) from reported earnings in year t. 15 Fig. 4 compares the distribution of pre-managed earnings to that of reported earnings previously shown in Fig. 1. Consistent with expectations, there is no evidence of a discontinuity around zero in the distribution of pre-managed earnings. 16 Fig. 4 also indicates that the distribution of pre-managed earnings is more dispersed than that of reported earnings, as evidenced by the higher density of observations in both tails of the distribution. This finding is consistent with insurers exercising discretion over the loss reserve accrual to smooth reported earnings. 15 There are at least two concepts of pre-managed earnings in a world where discretion occurs in multiple years. The first concept assumes that managers exercise discretion not only over loss reserves incurred in year t but also over development of reserves incurred in prior years. This is the measure of loss reserve development used to adjust reported earnings in Fig. 3. The second concept assumes that managers do not exercise discretion over losses incurred prior to year t. Thus, reported earnings for year t are adjusted to eliminate discretion only on losses incurred in year t, i.e., there is no catch-up adjustment for prior years incurred losses. Because there are plausible arguments for this alternative definition of pre-managed earnings, we also examine its distribution. Consistent with our primary results, untabulated findings reveal no discontinuity around zero. 16 Degeorge et al. (1999) suggest that deflation can lead to a spurious buildup in the density at zero. However, this does not appear to explain our results because we deflate both reported earnings and premanaged earnings, yet only find a discontinuity at zero in the distribution of reported earnings.

16 362 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) Frequency Earnings Interval Fig. 4. The distribution of reported net income (shaded) and pre-managed net income (bars), scaled by total assets. Pre-managed net income is equal to reported net income less the discretionary loss reserve accrual (i.e., the difference between the developed and originally reported loss reserve). The distribution interval widths are The first interval to the right of zero contains observations in the interval [0.000, 0.006), the second interval contains [0.006, 0.012), and so forth. However, as discussed above, reserve development naturally reflects ex post surprises. Thus, even if management provided unbiased estimates of policy losses, there would be greater dispersion in pre-managed earnings due to the development surprises. To summarize the findings reported in this section, we report evidence suggesting that firms understate the loss reserve accrual to avoid reporting small losses. More generally, we find that the loss reserve accrual is managed over the entire distribution of reported earnings, rather than exclusively or primarily in the region around zero. Firms in the left tail of the earnings distribution understate reserves relative to those in the right tail of the earnings distribution, consistent with income smoothing but inconsistent with P&C firms taking an earnings bath Ownership structure and the distribution of reported earnings In this section, we examine the effect of ownership structure on management of the loss reserve accrual and the distribution of reported earnings. We determine ownership structure from information reported in the annual editions of Best s Insurance Reports. Consistent with Mayers and Smith (1994), we classify firms according to the ownership structure of the ultimate owner(s). If the ultimate owners are the policyholders, we classify the firm as a mutual. We also classify the firm as a mutual if it is organized as a stock company but the majority of its shares are owned by a mutual. Mayers and Smith (1994) report that the activities of stock companies owned by mutuals are more like those of mutuals than stock companies. If the ultimate owner is included in the CRSP database or the EDGAR system of Securities and Exchange Commission filings, we classify the firm as public. Finally, if Best s Insurance Reports indicates that ownership rests with an individual, family,

17 W.H. Beaver et al. / Journal of Accounting and Economics 35 (2003) private consortium, etc., or we cannot locate the firm or its ultimate owner on CRSP or EDGAR, we classify the firm as private. Although the classification of insurers organizational structure is consistent with prior work, there are limitations to this approach in our research context. First, we cannot rule out the possibility that some of the firms we classify as private because of a lack of evidence to the contrary are actually public or owned by a mutual. Second, because of data constraints, we conduct our tests at the level of the individual insurance company rather than at the level of the ultimate owner(s). Although this approach has the advantage of categorizing firms according to their actual business activities rather than their nominal ownership structure (Mayers and Smith, 1994), it may not capture the entity at which the incentives to manage earnings operate. This limitation affects all three categories of ownership structures, although possibly to different degrees. Finally, our analysis of public insurers is based on earnings and loss reserves reported according to SAP rather than GAAP. Although SAP and GAAP measures of earnings and loss reserves are highly correlated, our tests of earnings management by public companies may be less powerful if their focus on managing GAAP numbers is not well captured by SAP. Table 3 reports descriptive statistics for public, private, and mutual insurers. The sample observations are divided fairly evenly between the three ownership types, although private and mutual insurers are somewhat more prevalent than public insurers. Whether measured by the level of scaled earnings or the incidence of losses, public insurers are more profitable than private insurers, which are in turn more profitable than mutual insurers. However, the greater degree of reserve overstatement exhibited by mutuals, median DEV for mutuals is 0.012, compared to for public firms and for private firms, suggests that pre-managed earnings are more similar across the organizational types. 17 Private insurers are classified as financially distressed more frequently than either public or mutual firms. Private insurers are also smaller than either public or mutual insurers. Consistent with prior research (e.g., Mayers and Smith, 1988; Lamm-Tennant and Starks, 1993; Harrington and Danzon 1994), mutual insurers, on average, write less long-tailed and LIABILITY lines of business that are prone to exogenous ex post shocks. This finding suggests that the greater magnitude of reserve development for mutuals is not due to their writing these less predictable lines of insurance. Fig. 5 shows the scaled earnings distributions of public (Panel A), private (Panel B), and mutual (Panel C) insurers. For all three ownership structures, there is evidence of a discontinuity at zero. The standardized difference is 1.96 for public firms, 3.79 for private firms, and 2.15 for mutual firms. 18 Surprisingly, the 17 For example, median net income before loss reserve development would be (0.036 plus 0.005) for public insurers compared to (0.023 plus 0.12) for mutual insurers. 18 We also examined the earnings distribution of publicly-traded P&C insurers (SIC code 6331) using data obtained from Compustat on earnings determined using GAAP. Consistent with the findings reported in Fig. 5, we find evidence of a significant discontinuity in the distribution of earnings immediately around zero.