Medical Loss Ratio Malpractice?

Transcription

1 Medical Loss Ratio Malpractice? Evan M. Eastman a Florida State University David L. Eckles b University of Georgia June 5, 2017 Abstract Under the Patient Protection and Affordable Care Act (PPACA), health insurers are required to spend a certain portion of premium revenue on consumers. This spending requirement is measured by the medical loss ratio. However, rather than a simple ratio of outflows to inflows, the medical loss ratio allows health insurers to include loss reserves. Reserves include payments that are expected to be made for losses that have not been finalized, have not been reported, or have not even occurred. By definition, these components of reserves are estimated by insurers and are subject to managerial discretion. Using two natural experiments, we exploit variations in state level regulation (pre-ppaca) as well as changes in insurer behavior post-ppaca to study the health insurer loss estimation practices. We find that regulation around spending requirements incentivizes insurers to report higher losses and thus may not necessarily change actual spending practices. We also find that the use of a Big 4 auditor and actuarial firm can help mitigate the management of reserves. Keywords: Health Insurers; Affordable Care Act; Medical Loss Ratio; Accounting Discretion; Reserve Management; Accruals; Earnings Management; Audit Quality JEL classification: G22, G24, I13, M41 The authors would like to thank Jim Carson, Cameron Ellis, Anne Ehinger, Steve Fier, Josh Frederick, Chuck Nyce, Brad Karl, Mary Kelly, Jonathan Ketcham, and Marc Ragin for helpful comments. a College of Business, William T. Hold/The National Alliance Program in Risk Management and Insurance, Florida State University, 519 RBA, Tallahassee, FL, 32306, Tel.: , Fax: , eeastman@business.fsu.edu. b Terry College of Business, Department of Insurance, Legal Studies, and Real Estate, University of Georgia, 206 Brooks Hall, Athens, GA 30602, Tel.: , Fax: , deckles@uga.edu.

2 Medical Loss Ratio Malpractice? Abstract Under the Patient Protection and Affordable Care Act (PPACA), health insurers are required to spend a certain portion of premium revenue on consumers. This spending requirement is measured by the medical loss ratio. However, rather than a simple ratio of outflows to inflows, the medical loss ratio allows health insurers to include loss reserves. Reserves include payments that are expected to be made for losses that have not been finalized, have not been reported, or have not even occurred. By definition, these components of reserves are estimated by insurers and are subject to managerial discretion. Using two natural experiments, we exploit variations in state level regulation (pre-ppaca) as well as changes in insurer behavior post-ppaca to study the health insurer loss estimation practices. We find that regulation around spending requirements incentivizes insurers to report higher losses and thus may not necessarily change actual spending practices. We also find that the use of a Big 4 auditor and actuarial firm can help mitigate the management of reserves. Keywords: Health Insurers; Affordable Care Act; Medical Loss Ratio; Accounting Discretion; Reserve Management; Accruals; Earnings Management; Audit Quality JEL classification: G22, G24, I13, M41

3 1. Introduction In attempt to ensure insurers were returning a significant portion of collected premiums to policyholders (in the form of medical care or medical improvements), the Patient Protection and Affordable Care Act (PPACA), Section 2718 of the Public Health Service Act was modified to require most health insurers spend at least 80% of their premiums on medical claims (including non-claims costs ) and improvements to health quality. 1 These spending requirements, commonly referred to as the medical loss ratio (MLR), are designed to promote spending on behalf of insureds. To further connect the requirement to policyholder expenses, if an insurer s MLR is below the appropriate standard, the insurer is required to issue premium rebates to insureds. 2 Table 1 shows the rebates for the first four years of MLR regulation. Insurers were required to pay substantial rebates overall, and specifically in individual, small group, and large group lines of business. The use of these MLRs is not a concept unique to the PPACA. Prior to PPACA implementation, a collection of states utilized similar regulatory requirements (often informed by model guidelines from the National Association of Insurance Commissioners (NAIC)) requiring health insurers report their MLR when making insurance rate filings. 3 Indeed, 34 states had some form of MLR reporting requirement prior to the implementation of the PPACA. Generally, these states had minimum MLRs for which the rate must support. That is, if a requested (or previously filed) rate would result in a MLR below the threshold set in the state, the rate would not be approved. Further, six states offered more stringent regulation and required insurers to offer policyholders rebates if the MLR threshold was not met. Much discussion has surrounded the implementation of the PPACA as a whole, though 1 The 80% threshold applies to policies written in the individual market and small group market. For those issued in the large group market, a higher threshold, 85% applies. Additionally, these thresholds can be altered by regulation at the state level. 2 The premium rebate is set at a level such that the insurer will meet the MLR standard. 3 For many personal lines of insurance, including health insurers, state-level regulators require insurers regularly report the prices (i.e. rates) charged to consumers. 1

4 very little attention has been paid to the implementation and oversight of the MLR. In particular, no academic studies (of which we are aware), and only a few policy-based reports have mentioned the potential for insurers to manipulate the MLR. Further, no studies have investigated the potential manipulation of the MLRs surrounding the state-level regulation pre-ppaca. Whether at the state or national level, manipulation of the MLR could allow insurers to avoid issuing rebates and effectively mitigate one of the more discussed purported advantages of the PPACA. The purpose of this study is not to question the motives, appropriateness, or implementation of the PPACA. On the whole, we will remain agnostic as to the PPACA. Rather, we seek to show that health insurers not only can manipulate accounting results (such as the MLR), but that health insurers appear to be doing so, prior and subsequent to the implementation of the PPACA. We also show, however, auditor and actuarial external oversight as well as market discipline can mitigate the manipulation. We utilize two natural experiments to show apparent manipulation of MLRs by insurers. First, we exploit the variation in state-level regulation of health insurers before the implementation of PPACA. Here we show robust evidence that those health insurers in states with MLR regulation systematically over-stated their losses relative to those health insurers in states without MLR regulation. 4 Second, we exploit the implementation of the PPACA as an exogenous shock to health insurers. Since some insurers (pre-ppaca) did not face MLR regulation, using a difference-in-differences approach, we can alleviate macroeconomic concerns and interpret changes in loss estimation behavior to that of the regulation itself. Here, we again find significant changes in the behavior of health insurers post-implementation of PPACA. More specifically, we find that firms operating in states without pre-existing MLR regulation significantly increased their loss estimates post-ppaca, relative to those firms operating in states with pre-existing MLR regulation. We further show that this result is 4 Institutional details, discussed below, allow for a rather clean test. 2

5 more pronounced for health insurers that are subject to less market discipline. The remainder of this paper proceeds as follows. In Section 2 we provide background on health insurer loss reporting, including a broad review of the previous literature on insurer loss manipulation. In Section 3 we examine our testable hypotheses and Section 4 describes our research design. Section 5, then, describes our data and provides our empirical results and Section 6 concludes. 2. Background 2.1. Health Insurer Regulation In 2010, PPACA was signed into law as a remedy to what some perceived as a systematic failure of the U.S. health care system to provide health care to the population at a reasonable cost. A significant component to the legislation involved an overhaul of the private provision of health insurance. Amongst other goals, the legislation aimed to both reduce the cost of health insurance and increase the availability (and take-up rates) of health insurance. One mechanism put in to place to minimize costs was a policyholder rebate if an insurer did not spend a certain fraction of premiums on health care expenditures (Kirchoff, 2014). This mechanism is designed to ensure that the health insurers are not charging excessive premiums. As a consequence of PPACA, health insurers are now required to report, on an annual basis, their medical loss ratio (MLR). If the insurer is not spending enough as measured by the MLR, the health insurer must issue refunds to policyholders such that their MLR meets the minimum set by the PPACA regulation. The minimum spending requirement varies for the type of insurance sold. For insurance plans covering individuals and small groups, health insurers must spend 80% of premiums on medical expenses. For large group plans, PPACA requires health insurers to spend 85% of premiums on medical expenses. Figure 3

6 1 shows the distributions of medical loss ratios as reported by firms for The red vertical lines represent the minimum requirement. We note that for all three types of policies (individual, small group, and large group) most firms are clustered immediately above the minimum requirement. Additionally, there are a substantial quantity of firms that do not meet this requirement, mandating rebate issuance. A subset of health insurers were subject to a similar form of MLR regulation prior to PPACA. Specifically, thirty four states had previously enacted state-level regulation requiring insurers report their MLR. Here, state regulators were requiring justification that insurer rates were not unduly high. These states (generally) were using the MLR as a rate regulation tool. 5 That is, most states were not requiring that insurers return premium dollars to policyholders ex-post, but rather would use the MLR as an indication that a rate needed to be lowered or could be approved for an increase. 6 Figure 2 shows those states (shaded in blue) with some form of MLR regulation. Figure 3 provides historical industry MLRs by year. The MLR is defined in this figure as total hospital and medical claims less reinsurance recoveries as a percentage of net premium income. While this definition of MLR is slightly different from the one used in the PPACA, this figure can still provide some context for current PPACA minimum MLR requirements. Notably the 25 th percentile of MLR is less than 80 percent in every year going back to 2001, which provides some indication that the current regulation is binding to some firms. Additionally, in many years the median is lower than 85 percent, which is the minimum requirement for large group MLRs under the PPACA Loss Reporting In health insurance (and, indeed all lines of insurance) insurers report losses from three sources. First, insurers report losses from actual claim payments made. Though some of 5 The use of the MLR varied somewhat across states. 6 Six states did have strict rules around the MLR reporting and required policyholder rebates. 4

7 these paid claims may be re-opened or be recouped by the insurer (through a process called subrogation), these paid claims are generally considered to be known. Second, even claims that are reported (and partially paid) can have aspects that cause the claim to extend into the future. 7 These unpaid, reported claims are generally referred to as incurred losses and are estimated by insurers. Finally, insurers also estimate losses from what are called incurred but not reported (IBNR) claims. With IBNR claims, insurers estimate expenses associated with claims which the insurer expects to be made but is not yet aware. The possibility for estimation error (and, therefore, manipulation) varies with these claim types and is greatest (lowest) with IBNR (paid) claims. The sum of insurer estimates for losses is referred to as the loss reserve. Loss reserves are not reserves in the context of a rainy day fund, but are treated as a liability on the balance sheet and an expense on the income statement. 8 Indeed, as loss reserves are increased, the financial position (from an accounting standpoint) of the insurer deteriorates. Both on a relative and absolute basis, loss reserves are economically significant and represent the single largest liability on an insurer s balance sheet. For health insurers, state-level regulation plays two explicitly important roles with regard to loss reserves. First, as discussed above, health insurers were required in some states (pre- PPACA) to report their MLR as a component of their requests for rate approval. The total losses incurred (including paid losses and loss reserves) are included as expected loses in the MLR. To that end, increasing loss reserves allows the health insurer to meet the MLR required. Second, and just as important for our purposes, if not a bit mundane, as part of general regulatory oversight, state-level regulation requires that insurers report the evolution of the loss estimates (referred to as loss development). In particular, insurers are required to report how loss reserves change over time for each reporting year. This reporting allows regulators (and academics) to observe the development of the claims and to observe the 7 Claims can be re-opened or simply have a long payout duration (tail). 8 As insurers estimate losses on a yearly basis, the estimate flows as an expense on the income statement. These loss estimates accumulate over time as a liability on the balance sheet. 5

8 error in the reserves initially set Prior Literature The idea that insurers manipulate reserves is not new. Indeed, a rather robust literature exists in the accounting and insurance literature using loss reserve errors as a measure of managerial discretion (e.g., Petroni, 1992; Grace and Leverty, 2010). In the property-casualty insurance industry, loss reserve errors have been linked to various earnings-related incentives. Early studies utilizing insurer reserve error (e.g., Weiss, 1985; Grace, 1990; Beaver, McNichols, and Nelson, 2003) focused on earnings smoothing. In these studies insurers are found to manage earnings in a way that minimizes the variability of income, thereby minimizing tax liability. Later studies (e.g., Petroni, 1992; Gaver and Paterson, 2004) show insurers to use reserving practices to avoid regulatory scrutiny. In particular, the Insurance Regulatory Information System (IRIS) ratios, used as a regulatory tool by the National Association of Insurance Commissioners, can be manipulated with reserving practices. Petroni (1992) and Gaver and Paterson (1999, 2004) show that insurers do appear to manage reserves to avoid violating enough ratios (four) to trigger regulatory intervention. More recent studies on reserving practices have focused on reserve management around executive compensation issues and ratings attainment. (Eckles and Halek, 2010; Eckles et al., 2011; Eastman et al., 2015) all show reserve management to be associated with managerial compensation, in particular, executive bonuses. Eckles et al. (2011) show that these results are affected by corporate governance and Eastman et al. (2015) show the effect to be most prominent in stock firms (relative to mutual firms). With respect to firm ratings, Eastman, Eckles, and Halek (2016) show insurers with a relatively low rating (relative to their target ) manage earnings upwards via reserves in an attempt to earn a higher rating. In addition to the many studies showing various incentives to manipulate reserves, there also exists studies that suggest potential methods to mitigate the earnings management. 6

9 Eckles et al. (2011) and Eastman et al. (2015) show that corporate governance and organizational form many also help mitigate earnings management, though organizational form and corporate governance changes are arguably difficult to accomplish in a short period of time. Other studies (e.g., Petroni and Beasley, 1996; Gaver and Paterson, 2001, 2007; Grace and Leverty, 2013), however, show that other forms of external monitoring, particularly auditors and actuarial firms, may help mitigate earnings management through loss reserve manipulation. The common theme in the entirety of the extant literature on reserve errors is that the samples examined are always firms in the broad property-casualty insurance industry, none of which are primary health insurers. 9 As a consequence of the dearth of studies on health insurer reserve errors, the specific issue of insurers managing reserves to meet a minimum MLR is unstudied. 3. Hypothesis Development We test three sets of hypotheses. First, we examine the degree to which health insurers appear to (or not) manipulate loss reserves in the presence of MLR regulations. Specifically, we study incentives created by state-level regulation before the implementation of PPACA. We also examine the change in behavior after the implementation of PPACA. Second we examine the degree to which this manipulation is mitigated by market discipline. Finally, we examine the effectiveness of external auditors to mitigate the manipulation of reserves. Our first hypothesis surrounds the incentive for insurers to manipulate their loss reserves to meet a minimum MLR imposed by either state-level regulation or PPACA. Recall that the MLR is essentially a ratio of medical expenses (including loss reserves) to health insurance premiums. Therefore, an increase in the loss reserve increases the numerator without affecting the denominator, thus increasing the MLR. Insurers operating under minimum MLR 9 Some insurers in the samples of these papers may have a small portion of business written in accident and health lines, but none are considered primarily health insurers. 7

10 regimes, therefore, have an incentive to increase loss reserves. This effect is particularly pronounced if the insurer is close to the minimum MLR. An insurer who is close can avoid regulatory scrutiny (or paying rebates) by increasing their MLR to meet the minimum threshold. Therefore, we put forth two related hypotheses: H1a: Health insurers operating with minimum MLR requirements will over-estimate losses relative to those insurers not subject to MLR regulation. H1b: Health insurers close to meeting the minimum MLR requirement will over-estimate losses. Next, we examine the degree to which market discipline minimizes health insurers ability to over-state reserves. Here we take advantage of an institutional component of health insurers and MLR regulation (the PPACA regulation). The minimum MLR requirement varies for health insurance business written for individuals (lowest minimum requirement), small group health plans, and large group health plans (highest minimum requirement). Presumably the difference in regulation is meant to reflect the risk related to each class. That is, individual health plans are generally considered to be high cost (hence the high premiums) for a variety of reasons including adverse selection concerns. Large group plans, however, are thought to spread risk over a larger, more diverse (with respect to health status) pool, therefore reducing the risk-related costs. Small group plans are, then, somewhere in the middle of the spectrum. In addition to risk-related differences between these three groups, there is also a fairly obvious continuum of market power held by these groups. Individual policyholders represent only themselves, and therefore can be considered to be price takers. Large groups, on the other hand, represent (by definition) a large group of insureds and will therefore generally have more power, exert more control, have more oversight, etc. of the insurer (again, the small groups will fall somewhere in between). If the insurance purchasers are exerting this market power and oversight, we would expect health insurers 8

11 that sell primarily to large groups to overstate losses (to meet MLR requirements) less than those focusing on individual health insurance plans (and small group plans). Thus our second set of hypotheses formally stated is written as: H2a: Health insurers focused on individual plans will overstate losses more than those focused on small group plans. H2b: Health insurers focused on individual plans will overstate losses more than those focused on large group plans. H2c: Health insurers focused on small group plans will overstate losses more than those focused on large group plans. In addition to market discipline, we examine the ability of external monitors to mitigate earnings management by the insurers. In particular, we hypothesize that higher quality external monitors (e.g., Big 4 auditors and affiliated actuaries) will reduce the earnings management seen in firms. Formally stated, we propose the following hypothesis: H3: Insurers utilizing high quality external monitors (Big 4 audit firms and Big 4 actuaries) will be associated with less earnings management. We will test these hypotheses by taking advantage of state-level variation in MLR regulation (prior to the PPACA regulation). We will also use the implementation of nation-wide regulation (PPACA) as a natural experiment. 4. Research Design 4.1. Data Our data are from annual statutory filings made by health insurers with the National Association of Insurance Commissioners between 2003 and We exclude firms from 10 The NAIC has several types of statutory statements depending on the operations of insurers. Here, we use the Health statement. There are also statements for Property/Casualty, Life/Accident & Health, 9

12 our sample who have non-positive direct premiums written or surplus. Also, consistent with prior studies examining reserve bias (e.g., Petroni, 1992; Beaver, McNichols, and Nelson, 2003; Grace and Leverty, 2010) we exclude firms with extreme errors in their loss reserves. Specifically, we exclude firms where the revised reserve estimate differs from the initial estimate by more than 50 percent in absolute value. Our sample focuses on affiliated and unaffiliated individual firms. We also exclusively examine firms that operate entirely in states with or without some form of MLR regulation. In other words, we exclude firms that operate across states. Since the majority of health insures operate in a single state, this restriction is not particularly stringent. This focus allows us to perform a clean test on the impact of MLR regulation on reserving. Our pre-ppaca sample consists of 3,455 firm-year observations consisting of 679 unique firms from 2004 to The nature of regulation in the PPACA period allows us to examine a firm s reserve management surrounding their MLR threshold. In order to perform this analysis, we obtain data from Centers for Medicare & Medicaid Services website ( which provides data on insurer MLR filings. Specifically, these data contain an insurer s reported MLR as well as their MLR target. 11 One issue in our empirical design is that the PPACA MLR regulation is assessed at the firm-state-line level. So a firm must meet the MLR regulation in each state it writes business in for each line (individual, small group, and large group). In order to provide a more direct test, we only consider firms in this analysis that operate in a single state. 12 This allows us to construct a clean sample to perform our empirical analysis. We merge these data with insurer loss reserve errors and other financial characteristics from statutory filings from 2010 to Data from 2010 are used to construct variables including lags. Data from 2014 are used to construct 2013 reserve errors. Our PPACA analysis consists Title, and Fraternal. 11 Note that these reported targets reflect any state MLR waivers that have been granted. 12 While we would ideally examine firms that write a single line in a single state to obtain the cleanest possible test, there are insufficient observations to make meaningful inference using this sample. 10

13 of data from 2011, 2012, and These are the first three years in which national MLR regulation was in place Empirical Strategy To test our hypotheses we utilize two general strategies. First, we measure the degree to which loss estimates vary among those health insurers operating in states with MLR regulation (pre-ppaca). Next we use the implementation of national MLR regulation as a natural experiment. Using a difference-in-differences approach, we estimate the degree to which health insurers change their estimation behavior in response to the newly imposed regulation. Finally, we examine reserve management to avoid MLR regulation in the post- PPACA period by specifically looking for evidence of reserve management among firms who just beat the minimum MLR requirement. We further describe our various empirical models more specifically below State-Level Regulation Model Using state-level variation in MLR regulation, we first estimate a model designed to capture differences in loss estimation behaviors by health insurers operating in states with differing MLR regulation. We next augment this model to include a variable measuring the quality of external monitoring. Ultimately, this model provides tests of H1a/b and H3. These models also control for standard firm-level characteristics hypothesized to determine loss estimation. Specifically, we estimate the following model: Error i,t = βmlr Regulation i,t + γx i,t + λi t + ɛ i,t (1) where Error i,t is firm i s one-year loss reserve error scaled by total assets in year t. MLR Regulation i,t is an indicator variable equal to one if health insurer i writes business in a state 11

14 with some form of MLR regulation in year t. 13 We measure the loss reserve error using data from the Underwriting and Investment Exhibit Part 2C Development of Paid and Incurred Health Claims, Section B Incurred Health Claims from the annual statutory statements. An example of this section is presented in Table 2. The boxed values in column (4) minus the boxed values in column (5) provide our measure of loss reserve error. We then scale this by total assets. A positive value for the loss reserve error indicates that the initial estimate was higher than the eventual development (overreserving). A negative value indicates that the initial loss reserve estimate was lower than the eventual development (underreserving). Literature examining property-casualty insurer reserve errors generally use five years of development to measure reserve error (e.g., Petroni, 1992; Beaver, McNichols, and Nelson, 2003; Grace and Leverty, 2010). For a number of reasons, we use only a single year of development in our present study. The first reason is that health claims tend to have a substantially shorter tail compared to liability losses. Table 3 provides data on the percentage of ultimate losses paid after each year. 14 On average, health insurers pay 88 percent of claims in the first year (t 0 ) and 98 percent of losses after the second year (t 1 ). For property-casualty insurers, approximately 88 percent of claims are paid after the fifth year and 98 percent of claims are not paid out until after 10 years (e.g., Grace and Leverty, 2012; Barth, Eastman, and Eckles, 2015). This suggests that in using a one-year error for health insurers, we are capturing a similar percentage of paid claims compared to the majority of studies examining 13 According to America s Health Insurance Plans (AHIP) April 2010 report Thirty-four states...establish MLR guidelines, require the filing or reporting of loss ratio information with state regulators, or impose limitations on administrative expenses for comprehensive, major medical insurance. We use these thirty-four states in our calculation. These states are: Arizona, Arkansas, California, Colorado, Connecticut, Delaware, Florida, Georgia, Iowa, Kansas, Kentucky, Maine, Maryland, Massachusetts, Missouri, Minnesota, New Hampshire, New Jersey, New Mexico, New York, North Carolina, North Dakota, Ohio, Oklahoma, Oregon, Pennsylvania, South Carolina, South Dakota, Tennessee, Utah, Vermont, Virginia, Washington, and West Virginia. 14 We note that our ultimate loss is defined by the last available reporting year (e.g. four years from the initial loss). If losses continue to develop past four years, the statistics reported below would change slightly. 12

15 property-casualty insurer reserve errors using five-year errors. For a second, and more practical, motivation for using one-year errors, we note potential data limitations in constructing loss reserve errors. The strength of loss reserve errors as a measure of managerial discretion comes from the ability to observe revised estimates of initial loss reserves. However, this requires lead years of data in order to observe these revisions. Accordingly, if we use the maximum amount of data available in the Underwriting and Investment Exhibit Part 2C, we could construct a four-year reserve error. However, this requires four lead years of data. For our pre-ppaca analysis, this would not be problematic. 15 However, as a primary motivation of our study is to examine the impact of PPACA MLR regulation, we would not be able to do so using four-year errors. Accordingly, we construct one-year errors which allows us to examine the first three years of PPACA MLR regulation (2011, 2012, and 2013) as we require only one lead year of data (i.e., construct the 2013 error using 2014 statutory filings). In examining loss reserve errors it is important to control for both discretionary and nondiscretionary determinants to isolate the marginal impact of each variable on the error (Grace and Leverty, 2012). We include a control for firm size to account for larger firms having the resources to employ more actuaries (Aiuppa and Trieschmann, 1987). We measure Size i,t as the natural log of firm i s total assets in year t. Harrington and Danzon (1994) suggest that firms attempting to grow can understate reserves in an attempt to improve firm growth (while also increasing insolvency risk). We control for this incentive by including Growth i,t which is the percentage change in firm i s net premium income from year t 1 to year t. 16 We also account for differing levels of product diversification. Providing a more diverse array of product offerings can make it more difficult for firm s to reserve accurately. 17 We 15 In our pre-ppaca results our main results are consistent for three- and four-year reserve errors as our reported results. 16 Data on net premium income is from the Statement of Revenue and Expenses page of the annual statutory health statement. 17 Studies examining the determinants of loss reserve errors commonly control for geographic diversification 13

16 measure Product Herf i,t as a Herfindahl Index for firm i based on net premium income across eight lines of business in year t. 18 Harrington and Danzon (1994) find that firms looking to grow will understate their reserves. These firms can hide this under-reserving using reinsurance. Accordingly, we control for reinsurance usage by including Reinsurance i,t, which is firm i s reinsurance ceded, divided by the sum of direct premiums written and reinsurance assumed. 19 We control for two aspects of a firm s ownership structure. First, we control for whether a firm is organized as a mutual or a stock firm. Mutual ownership structures can create different incentives that could impact reserving decisions (e.g., Mayers, Shivdasani, and Smith, 1997; Cummins, Weiss, and Zi, 1999). We include a binary variable, Mutual i,t, that is equal to one if firm i is organized as a mutual in year t and zero otherwise. Public i,t also controls for ownership structure. Public i,t is an indicator variable equal to one if firm i s ultimate owner is included in the CRSP database in year t and zero otherwise. 20 The second aspect relating to ownership structure is whether an insurer is under common ownership with other insurers (Powell, Sommer, and Eckles, 2008). 21 We include a binary variable, in their models (e.g., Grace and Leverty, 2010). In our setting using, for example, a geographic Herfindahl index is not appropriate due to our sample construction. Since we limit our sample to only firms that operate fully within our outside of states with MLR regulation the vast majority of firms in our sample operate in only one state. Therefore, the only geographic variation that is captured by a geographic Herfindahl index is if firms operate across two or more states with or without MLR regulation. We, therefore, do not include a geographic Herfindahl measure in any of our models. However, we note that our results do not materially change when a geographic Herfindahl index is included in our models. 18 The data we use to construct Product Herf are from the Analysis of Operations by Lines of Business page of the annual statutory health statements. The eight lines of business are Comprehensive (Hospital & Medical), Medicare Supplement, Dental Only, Vision Only, Federal Employees Health Benefit Plan, Title XVIII Medicare, Title XIX Medicaid, and Other. 19 Data on reinsurance premiums are from the Underwriting and Investment Exhibit Part 1 Premiums page of the annual statutory filings. 20 CRSP includes stocks traded on the American Stock Exchange (AMEX), the NASDAQ, and the New York Stock Exchange (NYSE). 21 Group membership is common in the health insurance industry. 85 percent of our sample firms are affiliated with other firms. One example is Cigna Healthcare Group. In 2011, Cigna Healthcare Group was comprised of numerous subsidiaries, such as Allegiance Life & Health Insurance Company, Cigna Healthcare of Georgia, and Cigna Healthcare MidAtlantic. Annual statutory statements for health insurers are reported at the individual company level. 14

17 Group i,t, which is equal to one if firm i is a group member in year t and zero otherwise. Finally, we include a set of variables related to income smoothing. The income smoothing hypothesis suggests that firms will manage reserves to minimize earnings and tax volatility (Weiss, 1985; Grace, 1990; Beaver, McNichols, and Nelson, 2003). Insurers may have an incentive to reduce the variability of their income to make the firm appear less risky to potential shareholders (Froot, Scharfstein, and Stein, 1993). Alternatively, insurers could be incentivized to smooth income (and reduce apparent risk) to appeal to regulators (Grace, 1990). Beaver, McNichols, and Nelson (2003) find empirical evidence suggesting that firms manage their reserves to avoid losses and across the entire earnings distribution. Notably, they find evidence that firms with small positive profits tend to under-reserve, suggesting that these firms only managed to avoid a loss due to under-stating losses. We, therefore, include a set of three binary variables to control for a firm s earnings in a given year (Beaver, McNichols, and Nelson, 2003; Grace and Leverty, 2012). Small Loss i,t is a binary variable equal to one if firm i s earnings fall in the top five percent of the negative earnings distribution in year t and zero otherwise. Small Profit i,t is a binary variable equal to one if firm i s earnings fall in the bottom five percent of the positive earnings distribution in year t and zero otherwise. Profit i,t is a binary variable equal to one if firm i s earnings fall in the top 85 percent of the positive earnings distribution in year t and zero otherwise. Breusch-Pagan Lagrangean multiplier tests indicate that pooled cross-sectional models are not appropriate for our sample (p-value < 0.001). We, therefore, perform a Hausman specification test which suggests that random effects are appropriate for our data (p-value > 0.10). Prior studies find evidence that loss reserve errors for property-casualty insurers are positively serially correlated (Beaver and McNichols, 1998; Grace and Leverty, 2012). As in Grace and Leverty (2012) we perform a test for autocorrelation in panel data from Wooldridge (2002). This test suggests the presence of first-order serial correlation within panels (p-value < 0.10). Additionally, Wald tests indicate that the residuals do not have a 15

18 common variance across panels (p-value < 0.001). This indicates that heteroskedasticity is present in our data. To account for the autocorrelation and heteroskedasticity in our data we use feasible generalized least squares. As noted by Grace and Leverty (2012), this allows us to correct standard errors for the panel-specific heteroskedasticity and serial correlation present in our data. This requires us to drop observations where there is only a single firm-year observations. All models include year fixed effects. We then augment model (1) to include variables related to external monitoring: Error i,t = βmlr Regulation i,t + ψmlr Regulation i,t Big4 i,t (2) +φbig4 i,t + γx i,t + λi t + ɛ i,t Prior studies have examined whether high quality external monitoring influences reserving practices (e.g., Petroni and Beasley, 1996; Gaver and Paterson, 2001, 2007; Gaver, Paterson, and Pacini, 2012). Gaver and Paterson (2001) note that in addition to considering external monitoring by auditors, insurers are subject to external monitoring by actuaries. Since actuaries possess an expertise in regard to establishing the loss reserve, it is important to consider high quality external monitoring by actuarial firms. Insurers are also required to to obtain an opinion from an actuary regarding the adequacy of a firm s loss reserve. 22 To examine whether external monitoring influences a firm s reserving practices, we include two variables to control for high quality external monitoring (Gaver and Paterson, 2001). First, we include Big 4 i,t, which is a binary variable equal to one if firm i had both a Big 4 auditor or a Big 4 actuarial firm in year t and zero otherwise. 23,24 22 Gaver and Paterson (2001) note that although firms are not required to obtain this from an independent actuarial firm, most do so in practice. 23 Consistent with Gaver and Paterson (2001), we define a Big 4 actuarial firm as one that is affiliated with a Big 4 auditor. 24 Data on a firm s auditor and actuarial firm come from the General Interrogatories page of the health annual statutory statements. The auditor is identified from 9 What is the name and address of the inde- 16

19 We also include an interaction term (MLR Regulation i,t Big 4 i,t ) to examine the differential effect of external monitoring in the states with MLR regulation Natural Experiment We next turn to an examination of the natural experiment associated with the implementation of the PPACA. Here, we estimate two models. First, we estimate the model presented above (equation 1 but with indicator variables for: 1) those insurers operating in states without MLR regulation (No MLR i,t ); 2) the years in which the PPACA was implemented (PPACA i,t ); and 3) the interaction between the two variables (No MLR i,t PPACA i,t ). This interaction term is essentially the difference-in-differences estimator. Figure 4 provides a theoretical illustration of our hypothesized relationship. We expect to observe that those not operating in states with MLR regulation will a) have lower loss estimates (analogous to firms overreserving in the presence of MLR regulation) and b) increase those estimates after the implementation of the PPACA. This specification allows us to take advantage of the cross-sectional variation in state adoption of MLR regulation prior to the PPACA, as well as the simultaneous MLR regulation adopted by all U.S. health insurers as mandated by the PPACA in Next, we exploit variation in the PPACA regulation to examine the degree to which those health insurers close to triggering a rebate, as a result of an MLR that is below the required minimum, over-state losses. Specifically we estimate: Error i,t = βbeat MLR Ind i,t + ψbeat MLR SmGrp i,t (3) +φbeat MLR LgGrp i,t + γx i,t + λi t + ɛ i,t pendent certified public accountant or accounting firm retained to conduct the annual audit? The actuary is identified from 10 What is the name, address and affiliation (officer/employee of the reporting entity or actuary/consultant associated with an actuarial consulting firm) of the individual providing the statement of actuarial opinion/certification? 17

20 Here, Beat MLR X is an indicator variable equal to one if the health insurer was above the minimum MLR requirement by ,26 We utilize three variations of this variable (Ind, SmGrp, LgGrp) for health insurers that focus on individual health plans, small group health plans, and large group health plans. 27 Overreserving by firms that just beat their minimum MLR requirement would be consistent with our hypothesis that firms will attempt to avoid regulation through reserve management. These firms, hypothetically, would have not met the minimum MLR requirement in the absence of reserve management, but used discretion over reserves to avoid having to issue rebates. 5. Results 5.1. Summary Statistics Summary statistics for our sample firms are presented in Table 4. Statistics for reserve errors specifically are reported in Table 5. The average reserve error for firms in our sample is , indicating overreserving. Approximately 75 percent of firms in the sample over-stated reserves (positive values of Error i,t ). This is consistent with studies examining reserving practices of property-casualty insurers, where the majority of firms over-reserve (e.g., Grace and Leverty, 2010; Eastman, Eckles, and Halek, 2016). The average firm overreserved in 25 The actual PPACA MLR rules are not uniformly applied. Some states were granted waivers so that health insurers in the state had a MLR lower than the stated MLR from PPACA. Our variable incorporates this variation. 26 While we focus here on the incentive firms have to manage reserves to avoid paying a rebate, there is also an incentive for firms to over-state losses even when they cannot meet the minimum MLR requirements. Since the PPACA MLR regulation requires firms to issue rebates until they meet the minimum requirement, any degree of reserve over-statement would reduce a firm s rebate liability. In untabulated empirical tests, we also include binary variables to control for firms missing their minimum MLR requirement. While our main result holds, the variable for missing the minimum MLR requirement is not significant. We suggest that this is due to increased fixed costs firms incur by going from paying any rebate to paying no rebate. This fixed costs include tangible factors, such as writing checks, as well as intangible factors, such as reputational penalties. 27 We utilize the descriptions of small and large group plans from the PPACA regulation. 18

21 every year aside from 2011 and The median reserve error is positive in every year. 69 percent of firm-years in our sample operated in states with some form of MLR regulation in the pre-ppaca period. Table 6 provides summary statistics in the pre- and post-ppaca periods and also provides univariate differences between the two periods. Sample firms tended to underreserve following the enactment of the PPACA. Firms also tended to be larger following the PPACA, experience less growth, and to report more small losses. Table 7 provides univariate differences and tests of loss reserve errors by whether or not a state was subject to state MLR regulation in the pre-ppaca period ( ). The average firm-year reserve error for firms operating in states with no MLR regulation in while the average firm-year reserve error for firms in states with MLR regulation is While firms subject to MLR regulation tended to overreserve relative to firms not subject to MLR regulation, we see that this difference is not statistically significantly different from zero. The median reserve error for non-mlr firm-years is , while the median for MLR firm-years is Firms operating in MLR states again tend to overreserve relative to firms in non-mlr states, as with the means. The difference for medians, however, is statistically significantly different from zero. This provides some preliminary evidence in support of our hypotheses related to overreserving and MLR regulation Table 8 provides univariate tests of loss reserve errors by whether or not a firm had to pay a rebate in the PPACA period. Panel A is for individual policies, Panel B is for small group policies, and Panel C is for large group policies. Firms that beat their minimum MLR requirement did not have to pay rebates and firms that are below their MLR had to issue rebates until they met their minimum MLR requirements. Firms in individual and small group (Panels A and B of Table 8) tended to overreserve if they beat their MLR requirements relative to firms that had to pay rebates. These differences, however, are not statistically significant for the means or the medians. This result also holds for means in large group 19

22 policies (Panel C of Table 8) in that the mean reserve error for non-rebate-paying firms is higher than rebate-paying firms, but it is not significant. The opposite result holds for large group policies (firms that had to pay rebates overreserved to a greater extent) at a statistically significant level in this case. Overall, these results are generally consistent with firms overreserving to circumvent MLR regulation. While the results in Table 7 and Table 8 provide some preliminary evidence for our hypotheses that MLR regulation induces overreserving, we note that it is important to control for discretionary and non-discretionary determinants of insurer loss reserve errors in order to isolate the impact of MLR regulation. In the next section we, therefore, provide results from our multivariate analyses Results: State-Level Regulation Results from our estimation of equation (1) are presented in Table 9. The dependent variable in both specifications is Error i,t, which is firm i s one-year loss reserve error scaled by total assets in year t. Results in column (1) exclude our variable of interest, MLR Regulation i,t. Positive coefficient estimates indicate overreserving while negative coefficients indicate underreserving. Standard errors are presented beneath each coefficient estimate and account for autocorrelation and heteroskedasticity. The estimated coefficient of MLR Regulation i,t is positive and significant in column (2). This provides evidence that firms operating in states with some form of MLR regulation in the pre-ppaca period tended to overreserve relative to firms operating in states without any MLR regulation. This result provides evidence in support of our first hypothesis. This significant and positive coefficient estimate is evidence consistent with firms overstating initial loss estimates in an effort to circumvent minimum MLR regulation in the pre-ppaca period. The results for the control variables in columns (1) and (2) of Table 9 provide additional 20

23 evidence of reserve management related to incentives other than MLR regulation. Specifically, the estimated coefficient of Size i,t is positive and significant in both specifications, indicating that larger firms, as measured by the natural log of assets, tend to overreserve. More conservative reserving by larger firms is not consistent with smaller insurers maintaining a safety loading (i.e., a buffer against larger than expected losses). However, this is consistent with prior empirical results in studies examining property-casualty insurer reserve errors. The estimated coefficient of Product Herf i,t is negative and significant in both specifications in Table 9. This negative coefficient suggests that insurers that concentrate their operations into relatively few lines of business tend to underreserve. The coefficient estimates for Reinsurance i,t suggests that insurers underreserve as they use more reinsurance. This is consistent with Harrington and Danzon (1994) and provides evidence that firms underreserve in an attempt to hide fast growth. The coefficient estimate on Public i,t is positive and significant in both specifications. This result suggests that publicly-traded firms tend to reserve more conservatively relative to their privately-held or mutual peers. Prior empirical studies document a similar result (e.g, Grace and Leverty, 2010). Affiliated health insurers tend to underreserve relative to unaffiliated firms, as evidenced by the negative coefficient estimates on our indicator variable, Group i,t. Finally, we find evidence of earnings smoothing behavior for health insurers. The negative and statistically significant coefficient estimate on Small Profit i,t provides evidence that firms underreserve in order to avoid losses, consistent with prior empirical studies (e.g., Beaver, McNichols, and Nelson, 2003). Overall, the results in Table 9 provide empirical support for our hypothesis that firms managed reserves in relation to MLR regulation in the pre-ppaca period. These results also support reserve management in response to other incentives, such as earnings smoothing. While the current regulatory environment differs from the one examined in the analysis in Table 9, the results presented here speak to the impact of the PPACA. The regulatory 21