University Park, PA 16802, USA. East Lansing, MI 48824, USA. Received 23 November 1998; received in revised form 08 November 1999

Transcription

1 Ownership concentration and sensitivity of executive pay to accounting performance measures: Evidence from publicly and privately-held insurance companies Bin Ke a, Kathy Petroni b*, Assem Safieddine b a The Smeal College of Business Administration, Pennsylvania State University, University Park, PA 16802, USA b Eli Broad Graduate School of Management, Michigan State University East Lansing, MI 48824, USA Abstract Received 23 November 1998; received in revised form 08 November 1999 We investigate the relation between CEO compensation and accounting performance measures as a function of ownership structure. We use publicly-held property-liability insurers to consider the relation for firms with diffusely-held ownership and use privately-held property-liability insurers to consider the relation for firms with closely-held ownership. We find a significant positive association between return on assets and the level of compensation for publicly-held insurers. Consistent with optimal contracting theory, we find no such relationship for privately-held insurers. Results suggest that within closely-held firms CEO compensation is less based on objective measures like accounting information and more on subjective measures. JEL classification: C23; J33; M41 Key words: Management compensation; Ownership structure; Insurance * Corresponding author. Tel: (517) ; fax (517) ; petroni@msu.edu We would like to thank Charles Hadlock, Hamid Mehran, Kevin Murphy, Melinda Newman, Sheridan Titman, an anonymous reviewer, and accounting seminar participants at the University of Michigan for insightful comments and Rebecca Shortridge and Ryan Thies for research assistance.

2 Ownership Concentration and Sensitivity of Executive Pay to Accounting Performance Measures: Evidence from Publicly and Privately-held Insurance Companies 1. Introduction A long literature starting with Berle and Means (1932) recognizes that the diffuse ownership structure may diminish shareholder incentives to monitor managers and to limit managerial perquisite-consumption. Contracting theory predicts that by tying pay to performance these agency costs can be partially mitigated. Consistent with this theory, empirical research has documented a positive relation between Chief Executive Officer (CEO) compensation of publicly-held companies and accounting performance measures. 1 In firms with less diffuse ownership, owners have more incentives to directly monitor management and therefore explicit incentive pay contracts may become less important. Shleifer and Vishny (1986) demonstrate that large non-management shareholders of publicly-held firms find positive benefits from directly monitoring management. Mehran (1995) finds that publiclytraded firms with a larger percentage of their shares held by large non-management shareholders use less equity-based compensation than those with fewer outside blockholders. He doesn't consider the use of explicit accounting-based contracts. The purpose of this study is to examine whether accounting-based incentive pay contracts are less prevalent in companies without diffuse ownership. Accordingly, we compare the use of accounting-based incentive pay contracts across firms with diffuse ownership, i.e., those that are publicly-held, and firms with the most concentrated ownership, i.e., those that are privately-held. 1 For example, see Murphy (1985), Barro and Barro (1990), Gibbons and Murphy (1990), Lambert and Larcker (1987), and Sloan (1993).

3 Published studies on compensation in privately-held companies are essentially nonexistent because the data generally has not been accessible. However, data for the compensation of CEOs of privately-held insurance companies are available from state regulatory agencies. In this paper, we use this unique data set to consider CEO compensation in the property-liability insurance industry across different ownership concentration levels. One could consider whether there are differences in the use of accounting-based contracts across different levels of ownership concentration within publicly-held firms. We primarily rely on the private versus public distinction because this dichotomy allows a more powerful test and it doesn't require the researcher to make an arbitrary distinction between diffuse and concentrated ownership. 2 Mayers and Smith (1992) use similar data to consider differences in compensation for stock versus mutual life insurance companies. Based on differences in corporate control across the two organizational forms, they argue that compensation for mutuals should be less sensitive to firm performance than compensation for stock insurers. Consistent with their prediction they do find a pay-to-performance link for stocks but not mutuals. Their sample of stock insurers 2 In addition, we don't focus on ownership concentration within publicly-traded firms because it is not clear to us that there are many public companies with sufficient ownership concentration to warrant that the firm forgo the use of explicit accounting-based incentive contracts. As discussed in Shleifer and Vishny (1997) large share holdings, especially majority ownership, are relatively uncommon for public firms in the United States. More specifically, Mikhail (1999) reports that 96% of his sample of publicly-traded life insurers use bonus programs based on operating performance measures. For completeness, however, in a later part of this paper, we discuss the results of a supplemental test that we performed to examine differential pay-to-accounting performance sensitivities within the publicly-held insurers. 2

4 includes both public and private companies. They do not separately analyze the pay-toperformance link for the public and private stock companies. We argue that significant differences in ownership concentration should be associated with different pay- to- performance sensitivities. For closely-held insurers there should be more direct monitoring of management by owners and less reliance on contracts that link management s compensation to explicit performance measures. Therefore we hypothesize that objective accounting measures will be less associated with compensation of CEOs of closelyheld insurers than with compensation of CEOs of widely-held insurers. The research hypothesis is tested using a sample of 45 CEOs of privately-held propertyliability insurers and 18 CEOs of publicly-held property-liability insurers during Consistent with optimal contracting theory, we find a significant positive association between return on assets and the level of compensation for publicly-held insurers while we find no such relationship for privately-held insurers. We also find that the change in compensation is significantly more sensitive to the change in return on assets for publicly-held insurers than privately-held insurers. Results suggest that within privately-held firms, the reliance on accounting information for performance measurement is reduced due to relatively less separation between ownership and control. Due to the separation of ownership and control in many corporations, executive compensation design is one of the most important issues in corporate governance. Many previous studies examine the pay-to-performance sensitivity in isolation (see Murphy 1998 for a review). This study contributes to this literature by demonstrating the importance of one alternative managerial control mechanism, share ownership concentration, as a determinant of the optimal structure of executive compensation. Recently there is a surge in interest in 3

5 comparing executive compensation structures among different countries. Because share ownership differs substantially among many countries (Shleifer and Vishny 1997), our empirical findings suggest that any differences in pay-to-performance sensitivity found in these studies should be interpreted with caution. Our results also contribute to the literature on earnings management within public versus private firms. Research suggests that private companies are more aggressive in managing taxes (Beatty and Harris 1999 and Mikhail 1999). The primary argument for this finding is that managers of publicly-held companies are more likely to have accounting-based incentive compensation contracts and are therefore more concerned about the negative impact of tax planning on their compensation. Importantly our paper is the first to empirically address this argument by testing for differences in pay-to-accounting performance sensitivities across a sample of public and private firms. 3 Readers should interpret our results with caution because our sample is very small and examines only one specific industry. It would be inappropriate to draw conclusions from this one study about the general nature of management compensation as a function of ownership structure. The remainder of this paper is organized as follows. In the next section, we provide the motivation for our hypothesis in greater detail. In section 3, we describe our sample and report descriptive statistics. Section 4 describes our empirical methods and results, and Section 5 concludes the paper. 3 Cloyd, Pratt, and Stock (1996), based on a survey of manufacturing managers, report that private firms are less likely to have incentive-based compensation with incentive-based compensation loosely defined. 4

6 2. Hypothesis development 2.1. Theories of management compensation in widely-held companies As discussed in Fama and Jensen (1983a and 1983b), in large publicly-held companies, agency problems result from the separation of residual risk bearing and decision making. Within these organizations, agency problems are controlled by decision systems that separate the initiation and implementation of decisions (i.e., decision management) from the ratification and monitoring of decisions (i.e., decision control). Monitoring of decisions involves measuring the performance of decision agents and implementing rewards. As discussed by Holmstrom (1979), if the agent's actions are not directly observed (because it is impossible or prohibitively costly), imperfect information may be used for performance measurement. There are costs of relying on performance measures that are based on imperfect information in compensating agents. First, introducing a link between pay and the performance measures requires the agent to bear risk and since the agent is assumed to be risk averse, the agent must be compensated for bearing this risk (Harris and Raviv 1979). Consistent with this theory, empirical evidence in Aggarwal and Samwick (1998) demonstrates that the degree to which CEO pay is sensitive to performance is decreasing in the variance of the firm's stock returns. Similarly, managers may engage in activities that reduce the firm's risk even if it is not in the shareholders' best interest (Jensen and Meckling 1976). For example, empirical evidence in Amihud and Lev (1981) suggests that managers engage in conglomerate mergers to decrease their own risk, at the shareholders' expense. Second, as argued by Baker, Jensen, and Murphy (1988), given an explicit incentive contract, managers may choose non-optimal decisions in order to maximize their welfare, referred to as "gaming the system". Holmstrom and Milgrom (1991) support this notion by 5

7 analytically demonstrating that linking pay to objective output measures in multidimensional tasks is costly because an increase in an agent's compensation in any one task will cause some reallocation of attention away from other tasks. Empirically, Dechow and Sloan (1991) find that managers reduce research and development spending as they near expected retirement, apparently to maximize bonus payments tied to accounting earnings. Apparently, for most publicly-traded firms, the benefits of relying on imperfect performance measures outweigh the costs. Accounting measures are widely used in compensation contracts of publicly-traded firms Theories on management compensation in closely-held companies In their seminal paper, Berle and Means (1932) argue that in publicly traded firms, diffuse ownership structures reduce shareholder incentives to monitor managers' wasteful expenditures. As Black (1992) argues, diffuse shareholders face the problem of collective action since they receive a very small fraction of the benefits from monitoring, but must bear the full cost of their monitoring efforts. The literature has recognized, however, that large non-management shareholders have different incentives and technologies for monitoring management relative to diffuse small shareholders. For example, Shleifer and Vishny (1986) demonstrate in a theoretical model that large non-management shareholders of publicly-held firms serve as effective monitors of management because they have a lower marginal cost of acquiring and disseminating information, and receive a bigger share of the benefits due to their large shareholdings. 4 For example,see Healy (1985) and Holthausen, Larcker, and Sloan (1995). 6

8 Moreover, these large non-management shareholders have both the incentive and ability to initiate proxy contests and takeovers. To date, however, little attention has been given to the monitoring mechanisms employed in companies that are essentially owned by only a few large shareholders (i.e., privately-held companies). Fama and Jensen (1983a and 1983b) argue that the residual claims of privately-held firms are largely restricted to important decision agents or to agents who have a special relationship with the decision agent, such as a family member or a business associate. In these cases, agency problems between residual claimants and decision agents are avoided, thus costly mechanisms for monitoring decisions are not needed. In some privately-held firms, however, the CEO may not own 100% of the firm or may not have a special relationship with any or all of the residual claimants. In these cases, agency problems will still exist and there is a need for owners to monitor management. We argue that the monitoring mechanism employed is less likely to be an explicit management compensation contract with objective performance-based measurements. Rather, it is based on more direct monitoring combined with subjective performance measurement. Assuming complete observation is possible, Holmstrom (1979) demonstrates the costs of relying on imperfect performance measures can be avoided and the first-best solution to the agency problem can be achieved by employing a forcing contract that penalizes dysfunctional behavior. In privately-held firms, the owners can not completely observe management. However, relative to diffuse shareholders of publicly-held firms, we can assume they have a greater incentive to observe management, as well as better access to management. As a result, owners of privately-held firms should be more likely to employ a forcing contract with penalty of firing. 7

9 Several studies on publicly-held firms report evidence that is consistent with our hypothesis. McConnell and Servaes (1990) demonstrate that shareholders who own large stakes in a firm are more effective monitors of a firm. They also find that these firms have a higher Tobin's Q, supporting the notion that large non-management shareholders of publicly-held firms play a value maximizing role in the firm. Mehran (1995) finds that publicly-traded firms with a larger percentage of their shares held by outside blockholders (i.e., large non-management shareholders) use less equity-based compensation than those with fewer outside blockholders. This suggests that direct monitoring by outside blockholders substitutes for equity incentive pay contracts. In addition, Brickley, Lease and Smith (1988) find that large blockholders vote more actively on antitakeover amendments than nonblockholders. Apparently, the blockholders have more incentive to invest in firm-specific information and to participate in decision making. If a forcing contract is used effectively, dysfunctional behavior must be penalized. Penalizing dysfunctional behavior in the form of firing a manager is less costly in a privatelyheld firm than in a publicly-held firm. Baker, Jensen, and Murphy (1988) argue that board members of publicly-held firms, who often own only a trivial fraction of their firm's common stock, are in no sense perfect agents for the shareholders who elected them. Board members are reluctant to terminate or financially punish poor-performing CEOs because they bear a disproportionately large share of the non-pecuniary costs (e.g., personal discomfort with the task, loss of important friendships, and the possibility of being sued for illegal discharge), but receive essentially none of the pecuniary benefits. Owners in privately-held firms will more directly 8

10 receive the pecuniary benefits and therefore will be more likely to fire an underperforming manager. 5 There is an alternative to our hypothesis. One could argue that CEO compensation in privately-held firms should be more sensitive to accounting measures than that in publicly-held firms because privately-held firms are more limited in the types of objective performance measures and technologies available to align managers' incentives with shareholders. For example, privately-held firms do not have market determined performance measures such as stock returns or the choice to grant their employees stock options (see Hall and Liebman 1998 for evidence on the dominance of these factors in the determination of compensation for CEOs of publicly-held firms). 6 Our empirical tests are designed to distinguish between the two hypotheses. 3. Sample selection and description 3.1. Sample selection for privately-held insurers To compile the sample of privately-held, stock property-liability insurers with CEO compensation data, we began by contacting the insurance departments of fourteen states with the 5 Challenging management in publicly-traded firms via proxy contents (see Pound 1988) or via takeover bids (see Grossman and Hart 1980) are also costly mechanisms. 6 Some privately-held firms reward managers with stock. In these cases the value of the stock may be based on year-end book value and may be sold back to the firm on retirement. We are not aware of any empirical studies that quantify the prevalence of these types of incentive arrangements. Although they are on the rise, they appear to be uncommon (Geer 1997). 9

11 largest number of property-liability insurers domiciled in the state. 7 We requested from each of these states' insurance departments the Supplemental Compensation Exhibit for 1996 for each insurer domiciled in the state. This exhibit is part of the annual report that insurers are required to file with state insurance regulators in its state of domicile. The data on this exhibit include the salary, bonus, and all other compensation for the CEO for The insurance departments of seven states (FL, TX, OH, MI, MO, and AZ) sent us copies of the requested exhibit; so our sample consists of privately-held insurers domiciled in these states. The remaining insurance departments either did not keep the exhibits on file or were not willing to copy and mail us the exhibits at a reasonable cost. 9 We received compensation information for the CEOs of 209 insurers. We then traced each insurer to A.M. Best's Insurance Reports: Property and Casualty, 1997 (referred to as Best's Reports) and determined that 72 of the insurers were privately-held stock insurers. Of these 72 insurers, 20 were omitted because the CEOs of these insurers were also CEOs of more than one insurance company. 10 An additional seven insurers were omitted because they were not active insurers (i.e., they did not have positive net premiums written), they were under regulatory control as a result of financial difficulties, or because accounting data for these insurers could not 7 The states with the most insurers were identified using the National Association of Insurance Commissioners (NAIC) Property Annual Statement Database for State insurance department contacts were identified using the NAIC Insurance Department Directory. 8 This exhibit also includes the salary, bonus, and other compensation of the four highest paid employees and five additional directors, officers, or employees if their compensation exceeds $100,000 a year. 9 One exception is Wisconsin. We received compensation data from Wisconsin's insurance department but in a different format than that used by the other states. As a result we did not include them in the sample. 10 Many insurance companies are members of insurance groups with common ownership and management. 10

12 be located. The final sample consists of 45 privately-held stock insurers, with CEO compensation data on 114 insurer-years. Financial information for each company was retrieved from the NAIC Property Annual Statement Databases Sample selection for publicly-held insurers To compile the sample of publicly-held property-liability insurers, we started with the 100 companies included in the 1996 COMPUSTAT annual industrial data base with an SIC code 6331 (fire, marine, and casualty insurance). From this initial sample, we omitted 13 foreign insurers, 29 firms that are not primarily property-liability insurers, 12 four insurers that went public after 1993, one firm that is on the ITC Bulletin Board, one firm that is majority owned by another firm in the sample, two firms that could not be located in the SEC's EDGAR database, and five firms for which we could not locate CEO compensation data. This resulted in a potential sample of 45 publicly-held insurers. In order to minimize the size difference between the public and private insurers, we further reduced the sample of public insurers to include only those that have mean assets during the period that are smaller than the largest private insurer's mean assets. This resulted in a final sample of 18 public insurers with CEO compensation data on 53 insurer-years. Data for each company was obtained from proxy statements, COMPUSTAT, the annual reports, and the NAIC Property Annual Statement Databases. 11 Data source: National Association of Insurance Commissioners (NAIC), used by permission. The NAIC does not endorse any analysis or conclusions based on the use of these data. 12 Firms that reported on at least one income statement during that less than 90% of total revenues were generated by property-liability operations were not considered to be primarily property-liability insurers. We applied this rule to our public companies because all of our private companies are 100% property-liability insurers. 11

13 3.3. Descriptive statistics Table 1 provides descriptive data about our privately-held and publicly-held sample insurers and two-sample tests for differences across the two organizational types. All monetary variables are expressed in constant (1994) dollars (in millions). The mean (median) total compensation for the privately-held insurers is $174,000 ($134,000) while for publicly-held insurers it is $482,000 ($394,000). The difference is not surprising given the differences in insurer size. The mean public insurer is 7 times larger than the mean private insurer both in terms of assets and net premiums written. In addition, public insurers have significantly more geographically dispersed operations. On average, public insurers write business in 31 states while on average the private insurers write business in only 12 states. Both appear to be equally profitable, since the mean and median return on assets are between 3-4% for both groups. [INSERT TABLE 1 APPROXIMATELY HERE] Based on the composition of total compensation, it appears that CEOs of private insurers are less likely to have a direct pay- to- performance link relative to CEOs of public insurers. The mean bonus for CEOs of private insurers is less than 10 percent of their total compensation. While less than one-half of the CEO-years for private insurers received a bonus, bonuses are much more common for CEOs of public insurers. For example, the bonus component makes up just less than one-quarter of the total compensation for both the mean and median CEO of public insurers. Table 2 provides descriptive statistics on the ownership structure of our sample firms in For public firms we report the percentage of the insurer's stock (including exercisable stock options) owned by the CEO, the percentage of stock owned by the largest outside 12

14 blockholder that is an individual or trust, private corporation, mutual fund, brokerage house, or investment advisor, 13 and the total of the CEO and largest blockholder percentages. The data for public firms should be complete since it is obtained from proxy statements. The data for the private insurers is obtained from A.M. Best disclosures and A.M. Best does not have a consistent disclosure policy or standard reporting format. As a result, the value of this data is limited. If A.M. Best doesn't disclose CEO ownership data we recorded the CEO as having no ownership in the firm. If no information on ownership by a blockholder is given we also recorded the insurer as having no large blockholder. This table, therefore, serves only as a lower bound on the extent of CEO and large blockholder ownership for the private insurers. For purposes of this table, CEO ownership percentages include ownership by members of the CEOs' families. 14 [INSERT TABLE 2 APPROXIMATELY HERE] Even with the underreporting of CEO and block ownership for the private insurers, private insurers exhibit more concentrated ownership. CEOs on average own 31% of the private insurers while for public insurers they own 15%. On average, the largest blockholder owns 28% of the private insurers while for public insurers the largest blockholder owns 10%. Approximately 87% (25%) of the median private (public) insurer is owned by the CEO and/ or the largest blockholder. The ownership concentration for the private insurers is significantly greater than that of the public insurers at the.02 level based on the Wilcoxon rank-sum test. 4. Empirical tests and results 13 We exclude blockholders that are publicly-traded corporations and mutual insurance companies because we only want to consider blockholders that represent concentrated ownership. 14 Family members are identified as individuals with the same last names. This likely causes us to understate the level of CEO ownership. 13

15 4.1. Regression model We test for a differential relationship between the use of accounting performance measures and CEO compensation by examining the relation between the compensation and return on assets for our sample insurers. For this purpose, we use the following regression model which pools the observations both cross-sectionally and temporally: COMP it = α t + β 1 PUBLIC i + β 2 ASSETS it + β 3 STATES i + β 4 ROA it + β 5 ROA it *PUBLIC i + e it, where i = insurer index ; t = year index for ; COMP = the natural log of total compensation (not including stock options and other long-term incentive compensations for public insurers) 15 ; PUBLIC = dummy variable, taking the value one if the insurer is public, otherwise zero; ASSETS = the natural log of total assets; STATES = the natural log of the number of states in which the insurer does business in 1996; ROA = net income over beginning-of- period total assets; e = error term. An intercept that varies across time is included to control for the effects of exogenous economic factors on compensation over the time period. PUBLIC is included to control for differences in the level of CEO compensation in public versus private insurers. STATES is 15 COMP omits stock options and other long-term incentive compensations for public insurers because the focus of our study is on accounting performance measures. 14

16 included to control for differences in the geographic dispersion of the insurers' operations. The coefficient on PUBLIC and STATES should be positive if CEOs of publicly-held and more geographically disperse insurers make more complex decisions and therefore require a higher quality of managerial talent. The managerial talent hypothesis predicts that higher quality management will have higher levels of compensation (see Rosen 1982). ASSETS is included to control for differences in the size of the insurer that the CEO manages. We expect this coefficient to be positive because a positive relation between firm size and the level of management compensation has been well documented by many previous researchers (see Kostiuk 1989), consistent with the managerial talent hypothesis. ROA is the accounting performance measure we use to test our hypothesis. 16 A similar measure was used by Blackwell, Brickley and Weisbach (1994) in an examination of the reliance on accounting-based performance measures in the internal performance evaluation of bank managers. 17 ROA*PUBLIC is included as a means to test whether the relation between compensation and ROA is different for public insurers. Our hypothesis predicts a positive coefficient on ROA*PUBLIC (i.e., the relation between CEO compensation and ROA should be stronger for public insurers). We do not make a prediction for the coefficient on ROA, whether it is positive or zero is an empirical question and depends on whether accounting measures are used at all in private insurers. However, if our empirical model yields results consistent with past 16 Net income and total assets for the public insurers are based on Generally Accepted Accounting Principles (GAAP) while for the private insurers they are based on Statutory Accounting Principles (SAP). In section 4.5 we discuss the differences between SAP and GAAP and the tests we performed to assess the impact of these differences on our analysis. 17 We also estimated our model using return on equity and the results are qualitatively identical. 15

17 research on pay-to-performance sensitivity, we expect (β 4 + β 5 ) to be positive, showing that there is a positive sensitivity of pay to performance in publicly held companies. We test the model both in levels and first differences. The levels model estimates the relation between the level of compensation and return on assets, while the change model more directly analyzes the sensitivity of pay to performance. 18 The levels model requires less data, however, and therefore maximizes the sample size. Previous studies on CEO compensation include CEO age and tenure as control variables (see, e.g., Kostiuk 1989, Barro and Barro 1990, and Mayers and Smith 1992). We searched several sources to determine the age of the CEOs in our sample, and the number of years they held the office of Chief Executive. This effort included searching the CEOs' biography in Who's Who in Insurance and conducting telephone inquiries with the insurance companies. Since we could only obtain the age and job years of the CEO for less than half of our insurer-years we do not include them in our primary model. 19,20 18 In the change model the dependent variables is the change in the logs of total compensation. We also estimated the change model using the change in the levels of total compensation and obtained similar results. 19 Mayers and Smith (1992) also found weak evidence of an association between CEO pay and the percentage change in net premiums written. We tried including the change in net premiums written in our model but it was not explanatory. 20 Since, in addition to CEO age and tenure, there are likely to be other omitted CEO- and insurer-specific variables that are important in determining compensation, we also estimate our levels regression using a fixed effects model. Kostiuk (1989) similarly applies a fixed effects model. Results using the fixed effects model suggest that we do not have omitted CEO- or insurer-specific variables that are confounding our analysis because results on our coefficients of interest are similar to those reported for the OLS model. In addition, applying a Hausmen test suggests that a random effects model could also be used. Results using the random effects model are also similar to those of OLS. Kennedy (1993, pg ) includes a general discussion of the fixed and random effects models. 16

18 4.2. Regression Results on the levels of compensation In Panel A of Table 3 we present the OLS results of the pooled cross-sectional regression using the log of CEO compensation as the dependent variable. The error terms in the OLS model fail White's (1980) specification test and are likely serially correlated. As a result we report t-values that account for heteroskedasticity and serial correlation (Rogers 1993). There are 159 CEO-year observations used to estimate the model. 21 The regression is explanatory with an adjusted R 2 of 61 percent. As expected, compensation is increasing in the size of the insurer as well as the level of geographic dispersion, evidenced by significant coefficients on ASSETS and STATES. The coefficient on PUBLIC is negative but not significant. Apparently, ASSETS and STATES better capture complexity of the firm than the distinction between being public or private. The coefficient on ROA is essentially zero, while the coefficient on ROA*PUBLIC is significantly positive at less than the.01 level (t-statistic=3.87). 22 [INSERT TABLE 3 APPROXIMATELY HERE] Panel B presents the results of the regression on the 53 public CEO-years. Consistent with the results in panel A, ROA is associated with compensation for public insurers at less than the.01 level. Panel C reports the results of the regression on the private insurers. The regression is estimated across all 106 private CEO-years as well as for the 45 CEO-years in which the CEO 21 We omit 8 observations using the Cook's distance criteria (Cook 1977). OLS results on the full sample of 167 observations are similar except the significance on our variable of interest (ROA*PUBLIC) is reduced to the 5 percent level. 22 All reported probabilities on t-statistics are one-tailed tests if the sign of the association is predicted, otherwise two-tailed probabilities are reported. 17

19 owns greater than 5% of the insurer and the 61 CEO-years in which the CEO owns less than 5% of the insurer. We separate the private insurers based on CEO ownership because in cases where the CEO owns a significant portion (all) of the insurer, there should be little (no) need to monitor the CEO. In cases where the CEO owns less than 5% of the firm, monitoring, either objective or subjective, becomes essential. The coefficient on ROA is insignificant across all three estimations, even for the subsample with the highest agency costs. This suggests that even CEOs with little ownership interest are not compensated on objective earnings-based measures. Because the public insurers are still on average larger than the private insurers we performed two additional tests on the pooled regression. Both tests suggest the difference in firm size across the public versus private insurers is not driving the noted difference in the sensitivity of pay to performance. First we added the interaction of ROA and ASSETS to the model. The coefficient on this variable is insignificant (t=-0.71). Second, we created a subsample that includes the 28 private CEO-years that are bigger, in terms of assets, than the smallest public CEO-year and the 28 smallest public CEO-years. The mean assets for the private insurers is 99 million and for public insurers 130 million. Results on this smaller sample are less significant but consistent with those for the full sample. The coefficient on ROA is negative and insignificant, while the coefficient on ROA*PUBLIC is positive and significant at the 8% level. Another difference between the private and public insurers is the distribution of ROA. The variance of ROA for the private firms is twice as large as the variance of ROA for the public firms. The lowest ROA for the private (public) observations is -37% (-5%) and the largest ROA for the private (public) observations is 27% (11%). Since explicit bonus plans often have lower and upper bounds (see Healy 1985) this may be biasing our tests towards finding no association between compensation and ROA for the private insurers. To address this issue we created a 18

20 subsample that includes the 48 public observations with positive ROA and the 79 private observations with positive ROA and ROA less than the largest ROA of the public insurer (i.e., less than 11%). The variances of the ROAs across the private and public observations in this subsample are 2.4% and 2.7% respectively. We tested our model on this subsample and found that the coefficient on ROA is insignificant (t=0.49) and the coefficient on ROA*PUBLIC is positive and significant at the 8% level Regression Results on changes in compensation Table 4 presents the results of our OLS regression model on 101 CEO-years after replacing COMP with COMP, ASSETS with ASSETS, ROA with ROA and finally ROA*PUBLIC with ROA*PUBLIC. STATES is not included in this regression because this variable is zero for our sample. 23 The model has an adjusted R 2 of 25 percent and appears well specified. White's (1980) specification test does not reject the null hypothesis of homoskedastic errors or independence of the error terms and explanatory variables. [INSERT TABLE 4 APPROXIMATELY HERE] In panel A, the results on the entire sample are reported. The positive, significant coefficient on PUBLIC demonstrates that CEO compensation increased more for public insurers over the period than for private insurers. Consistent with the result in the levels regression, CEO compensation of the public insurers is significantly more sensitive to performance than that of private insurers (the t-statistic on ROA*PUBLIC is 2.17). Interesting, the coefficient on ROA is positive and marginally significantly different from zero (the t- 23 We collected STATES using 1996 disclosures. Based on reading the 1996 annual reports of our public insurers it appears like this variable is very stable over the sample period. 19

21 statistic is 1.75). Consistent with our hypothesis, however, the coefficient for the public insurers on ROA is more than four times as big as that for private insurers. Panel B and C report the results of estimating the change model separately on the private and public insurers. These separate regressions are consistent with the pooled regression. In Panel C we also report results of the change model for the 30 (37) private observations with the CEO owning greater (less) than 5%. The coefficient on ROA for the insurers with little CEO ownership is only 0.61 with a t-statistic of 0.86 while for insurers with significant CEO ownership it is 1.84 with a t-statistic of Apparently insurers with significant CEO ownership drives the significance of the positive relation between ROA and COMP. Again, this suggests that compensation of CEOs is not determined by explicit earnings-based contracts for private insurers Additional control variables To rule out alternative explanations for our findings we re-estimated our primary regression model including several other control variables as well as ROA*ASSETS. The results of this regression are reported in Table 5. Importantly, the addition of these additional controls did not change the conclusions regarding RAO*PUBLIC (the t-statistic is 2.10). Below we discuss each of these additional controls. [INSERT TABLE 5 APPROXIMATELY HERE] We collected data on the net premiums written by line of business for each of our sample insurers to control for any systematic differences in the operations of the public and private insurers. The public insurers, on average, write business in ten different lines, while the private insurers write business in only two lines. This is not surprising given that our public insurers are 20

22 on average larger than our private insurers. It is not clear what impact writing more lines of business will have on CEO compensation. It may be the case that managing fewer lines of business requires less managerial talent and therefore insurers writing fewer lines of business will have lower paid CEOs. Alternatively, insurers writing business in only a few lines may be more likely to write specialized policies and therefore have CEOs with more specialized expertise that demands higher compensation. To test whether the difference in the number of lines of business is important in the pay-to-performance sensitivity, we include the number of lines of business (LINES) as well as an interactive term LINES*ROA. We find that the coefficients on LINES and LINES*ROA are not significantly different from zero. We also consider, on average, how much insurance public and private insurers write in three major lines of business (i.e., multiple peril, liability, and automobile). Public and private insurers both write 25% of their business in multiple peril lines. Public (private) insurers write 7% (9%) of their business in medical malpractice, product and other liability. The average private insurer writes 36% of its business in automobile insurance while the average public insurer writes 47%. The difference in automobile lines is statistically significant at the 10% level. Since public insurers write marginally more automobile insurance and it is possible that writing auto insurance requires more complex decision making, we estimated our regression model including the percentage of total net premiums written in automobile insurance (AUTO) as well as an interactive term AUTO*ROA. The coefficient on AUTO is positive with a t- statistic of The coefficient on AUTO*ROA is insignificant. We also consider the regulatory environment because previous researchers have argued and demonstrated that the regulatory environment impacts the pay-to-performance relationship. The managerial talent hypothesis suggests that if regulation decreases the complexity of the 21

23 decisions made by management, then there should be a weaker link between pay and performance for more highly regulated firms. Consistent with this hypothesis, Smith and Watts (1992) find that the use of incentive contracts is less likely for regulated utilities than for manufacturing concerns. Hubbard and Palia (1995) and Ezzell and Miles (1995) find that within the commercial banking industry the sensitivity of CEO pay to differences in performance increased significantly, as bank activities became less regulated in the early 1980s. Our sample insurance companies face different regulatory environments because insurance regulations vary by state and each insurer essentially follows the regulations of its state of domicile. 24 In order to determine if differences in the regulatory environment are confounding our results, we identified the state of domicile for each insurer as reported in Best's Reports. 25 A proxy for the regulatory environment was then created based on a Conning & Company's survey of insurance managers in The survey assesses the relative freedom among states to manage personal and commercial lines of business and each state is ranked from one to 51 for each line. The higher the ranking, the more restrictions the state places on insurers' ability to design, price, and market their products. The mean of the personal and commercial rankings is significantly lower for our privately-held insurers than for our publicly-held insurers. 26 This suggests that the private insurers included in our sample operate in less restrictive environments than the public insurers. The presence, therefore, of a regulatory impact on pay-to-performance sensitivity would have biased us against finding our empirical result that there is a weaker relation between COMP and ROA for private insurers relative to public insurers. 24 See Petroni and Shackelford (1995) for a discussion of the variations in regulatory environments across states. 25 For the publicly-traded insurers, we identified the state of domicile of the subsidiary with the largest assets. 26 Recall that our privately-held insurers are from six states that were not randomly selected. 22

24 For completeness, however, we estimated our model including the mean rank (STRICT) as a proxy for the regulatory environment in which each insurer operates as well as an interaction between the mean rank and ROA. The coefficients on the regulatory climate variables are not significantly different from zero. We therefore conclude that differences in regulatory environments are not confounding our results Controlling for differences in GAAP versus SAP earnings Lastly we consider whether our tests are confounded by the fact that ROA for the public insurers is based on Generally Accepted Accounting Principles (GAAP) while for private insurers it is based on Statutory Accounting Principles (SAP). There are essentially five items that cause GAAP and SAP to differ. First, under GAAP, deferred policy acquisition costs are capitalized, while they are immediately expensed for SAP. Second, under GAAP, salvage and subrogation recoverable are reductions to the liability for outstanding claims, while under SAP they are not recognized until received. Third, under GAAP, deferred taxes are recognized, while under SAP taxes are accounted for on the cash basis. Fourth, some GAAP assets are considered non-admitted and are recognized as direct reductions to stockholders' equity (i.e., policyholders' surplus) for SAP. These non-admitted assets include furniture and fixtures, automobiles, and accounts receivable and investments not meeting certain requirements. Fifth, SAP doesn't recognize the fair values of fixed maturity securities, while GAAP recognizes the fair values of those fixed maturity securities that are not considered 'held-to-maturity'. Private insurers generally do not prepare GAAP financial statements and public insurance companies, that are essentially holding companies of individual insurance companies, do not prepare consolidated SAP financial statements. We therefore argue that it is appropriate that we 23

25 measure performance using SAP for the private insurers and GAAP for the public insurers. One could possibly argue, however, that we find a weaker pay-to-performance relation for our private insurers because our performance measure for private insurers is based on SAP and CEOs of private insurers are compensated based on GAAP. We can not directly address this issue because we do not have GAAP accounting numbers for our private insurers. Fortunately, however, the public insurers disclose in the notes of their financial statements the sums of the insurers' subsidiaries' statutory incomes and statutory policyholders' surpluses. This allows us to consider whether a pay to performance relation can be found using SAP measures for public insurers. We collected the disclosed statutory incomes for and statutory surpluses for from the annual reports to the extent available for our public insurers. 27 Based on these disclosures, we calculated a statutory return on equity (ROE) for 122 public insurer-years. 28 The correlation between the SAP ROE and GAAP ROE is highly significant (the Pearson correlation coefficient is.86). We then tested the relation between pay and performance on the public insurers (using a regression model similar to that estimated in table 2, panel C) using SAP ROE rather than GAAP ROA. Similar to the results on GAAP ROA, the coefficient on SAP ROE is positive and significant (at less than the 1 percent level) for public insurers. Apparently the SAP and GAAP performance measures capture similar constructs. So, therefore, even if our private insurers are being compensated as an explicit function of GAAP 27 Seventeen of the public insurers also report a reconciliation between GAAP income and equity and SAP income and equity. For these insurers, we calculated a consolidated SAP income and equity by adjusting the sums of the SAP incomes and equities for intracompany transactions and transactions from non-insurance operations. For insurers without a reconciliation we simply used the reported sums of the SAP incomes and equities. 28 SAP ROA would be optimal but public insurers usually do not report statutory assets. 24

26 income, a strong relation between pay and SAP income for the private insurers should still be present. This suggests that our documented differential relation between pay and performance is not being driven by differences in accounting for public versus private insurers but is being driven by differences in ownership structure Supplemental analysis of publicly-held insurers One limitation of our findings is that the difference in pay-to-performance sensitivities noted between the public and private insurers may be caused by some unspecified difference between these two organizational forms other than the degree of diffuse ownership. One way to address this issue is to determine if the degree of diffuse ownership within the sample of publicly-held insurers is associated with different pay to performance sensitivities. 29 This may not be an appropriate test, however, if the degree of ownership concentration for all public insurers is too small for direct subjective monitoring or if the level of variation in ownership concentration is not sufficient for our tests to detect differences in the reliance on explicit accounting-based incentive contracts. In spite of these limitations, we estimated the primary regression model on only the publicly-held insurers substituting CONCEN for PUBLIC. CONCEN is a variable that is coded as 1 if the insurer's CEO and largest blockholder combined own greater than 40% of the insurer (i.e., a public insurer with relatively concentrated ownership). There are 4 insurers (12 insurer- 29 We also ran a test on the full sample of public and private insurers using an alternative, albeit arbitrary, dichotomous measure of ownership concentration. We coded all public insurers with ownership less than 40% by the largest blockholder and/or CEO as DIFFUSE. We substituted DIFFUSE for PUBLIC in our primary regression. Consistent with expectations, we find that the coefficient on ROA is insignificant (t-statistic is 0.61) and the coefficient on ROA*DIFFUSE is positive and highly significant (t-statistic = 4.31). 25