Market Discipline in Life Insurance: Insureds Reaction to Rating Downgrades in the Context of Enterprise Risks

Transcription

1 Market Discipline in Life Insurance: Insureds Reaction to Rating Downgrades in the Context of Enterprise Risks Etti G. Baranoff Associate Professor of Insurance and Finance Virginia Commonwealth University School of Business 1015 Floyd Avenue Richmond, Virginia Thomas W. Sager Professor of Statistics Department of Information, Risk, and Operations Management The University of Texas at Austin, CBA Austin, Texas (512)

2 Market Discipline in Life Insurance: Insureds Reaction to Rating Downgrades in the Context of Enterprise Risks Abstract There is serious debate about the effectiveness of the market to discipline regulated industries like banks and insurance. In this paper, we examine empirically the reaction of consumers to changes in the financial ratings of life insurers. We find that downgraded life insurers in the period experience a decline in demand for life insurance policies in the year after a downgrade. This occurs in spite of mitigation strategies like adjusting premium rates. It occurs in the context of a spectrum of enterprise risk measures and other controls such as financial risk (leverage), product risk (premiums written in various lines), asset risk (a proxy called opportunity asset risk), operational risks (use of derivatives, organizational structure and distribution system), regulatory risk (risk-based capital ratio), and size (total assets). Our findings comport with similar research in banking on the reactions of depositors to changes in the risk taking of banks. Moreover, our model is able to estimate the magnitude of the decline in demand for individual insurers. Using the notion of Granger causality, we are also able to demonstrate that the direction of the relationship flows from ratings downgrade to decline in demand, rather than the reverse. These findings have potentially strong impacts on the issue of appropriateness of risk models that regulators and the industry are developing in parallel: The parallel tracks may need to align with a third rail consumers discipline. 1

3 Market Discipline in Life Insurance: Insureds Reaction to Rating Downgrades in the Context of Enterprise Risks I. INTRODUCTION The literature entertains serious debate about the effectiveness of the market to discipline regulated financial institutions. In the banking industry, studies of market-based discipline focus on the reactions of investors and depositors to increased institutional risk taking as summarized by Flannery (1998 and 2001). Park and Peristiani (2007) note that the new capital adequacy proposals of the Basel Committee for Banking Supervision (Basel II) consider market discipline, along with capital requirements and supervision, as one of the three pillars of support for the banking system. Such a statement is supported by many market discipline studies (some described later) that find stock prices reacting negatively to increased risk and depositors shifting their money elsewhere. Insurance is another financial sector industry that is highly regulated. But we found no studies of the direct exercise of market discipline by consumers for the life insurance industry. The closest approximation is a study by Epermanis and Harrington (2006), who found that premiums (as a proxy for consumer demand) were sensitive to ratings downgrades in the U.S. property/casualty insurance industry. However, Froot (2007) finds that various studies rely on customer preferences to drive insurance pricing. He concludes that consumer s demand falls when financial position of insurers decline and there is evidence in the literature to support this view of excess sensitivity. 1 Two studies that focus on the ability of capital markets to discipline insurers are Brewer and Jackson (2002) and Fenn and Cole (1994). However, these studies explore investors reactions rather than insureds reactions. Moreover, few life insurers are publicly traded. It appears that so far the role of consumers in disciplining life insurers has not been explored. 1 See Phillips, Cummins, and Allen (1998), who estimate the price discounting in relationship to probability of insurer default. 2

4 This paper is the first study known to us to examine empirically the question of market discipline for essentially the entire life insurance industry in the U.S. Our focus is consumer demand for life insurance policies. We model the effect of rating downgrades on the number of in-force life insurance policies in the context of a number of important enterprise risks and other control variables. The A. M. Best ratings of insurers are used, as they are the oldest and most often used in this market, regardless of organizational structure. 2 The enterprise risks include appropriate measures of asset risk, product risk, operational risk, regulatory risk and market risk. Not only do we test whether consumers wield market discipline, but also we can estimate the magnitude of the effect. We find significant and substantial effects. Therefore, this study contributes importantly to the debate on the proper role of free market forces vis-a-vis regulation in life insurance, paralleling the aforementioned three pillars of the banking industry. The study suggests that regulatory risk models, as well as parallel industry and rating agencies models, should not downplay the role of market forces. The importance of the life insurance industry in the U.S. is demonstrated in Table 1, which summarizes aggregate demand and price data for Table 1 shows data for only life insurance products (whole life, term life, group life, etc.). We emphasize that life products represent only a portion of the business of the industry. Life insurers also market health and accident insurance, annuities, and reinsurance. 3 Over this period, the demand for policies increased overall by 24.8 percent and average face amount of coverage per policy increased by 58.4 percent, while the average premium rate for all types of policies decreased from $7 to $4.9 per $1,000 of coverage and mortality rates also decreased. 4 Thus, the growth in premium volume was not generated by price increases. On the contrary, the growth of life insurance premiums was generated by selling more insurance policies and larger death benefits (face amount) per policy. Although the persistence throughout this period of a regime of industry growth might suggest some difficulty in exerting market discipline, our results show clearly that consumers retained that power. 2 Both stock and mutual insurers are rated. 3 For example, in 2003 the life insurance industry collected annuities premiums totaling $279 billion, health and accident premiums of $120 billion, and reinsurance premiums of $65.5 billion. 4 See data at: Table 12. Estimated life expectancy at birth in years, by race and sex: Death-registration States, , and United States, In National Vital Statistics Reports, Vol. 53, No. 6. pp Published November 10,

5 Insert Table 1 In order to measure market discipline for life insurers, it is especially appropriate to focus upon consumers reactions since a life policy is a sold product rather than a bought product. Consumers can be choosy both because there are many choices and also because information about the financial condition of insurers is readily available. 5 The annual statements that insurers submit to regulators include data on the number of life insurance policies and certificates sold, surrendered and paid, as well as coverage totals, and premiums. So these rich data enable the calculation not only of demand for life products, but also extrapolating the average price of life policies (of all types) per $1,000 of coverage. Price is a critical control variable in any analysis of demand. Thus, this study is able to analyze essentially the entire life insurance industry in the U.S. by virtue of its consumer focus. Many studies of market discipline for banks focus upon investors, as noted in the survey provided by Flannery (1998). Such a focus requires availability of stock prices to assess impact on firm value. Unfortunately, very few life insurers have publicly traded stock. Those that do may not be representative of the rest of the industry. We use various types of regression methodologies in this study. The primary response variable is the year-to-year change in the logarithm of the number of life insurance policies and certificates in effect. The predictor of our prime interest is the change in the A. M. Best rating from the preceding year. Control variables include the change in price of coverage, company size, and a number of enterprise risk measures to be listed later. All models lead to consistent results. The following is a brief synopsis of the results of the points investigated: 1. To establish a baseline, we first use OLS regression to contrast the effect of a rating downgrade with the effect of a rating upgrade. We find that a downgrade has a significant and substantial negative impact on demand. We also find a significant interaction between change in price and change in rating that further decreases demand for life policies for the downgraded insurers relative to those upgraded. 2. In order to compare downgraded with no-change insurers, we must correct some technical regression issues that loom larger for the downgrade vs. no-change comparison than for downgrades vs. upgrades. We use GEE (generalized estimating equations) 5 See quotes on the Internet in cites such as or In these quotes the rating of the insurer is always included along with the price. 4

6 regression models for assessing the effect of a rating downgrade vis-à-vis unchanged insurers. Using the same models, we also assess the effect of a rating upgrade. These models can correct within-insurer correlation that results from including the same insurer multiple times in the data (multiple downgrades or multiple no-changes, for example). Again, we find that a downgrade has a significant and substantial negative impact on demand. However, an upgrade has no statistically significant effect on demand. 3. To investigate whether the direction of the relationship runs from ratings change to demand, or whether it really runs in the opposite direction, we use Granger causality tests. We confirm that a rating downgrade Granger-causes a reduction in demand, but that a reduction in demand does not Granger-cause a rating downgrade. In order to illuminate the importance of our findings for the life insurance industry, the next section provides a literature review. Section 3 describes the data. Section 4 discusses the methodology and results. The paper concludes with summary. II. LITERATURE REVIEW Flannery (2001) defines market discipline as the power of market forces to assess and control financial institutions risky behavior. Market forces include investors and consumers. Most of the market discipline research focuses on banks and investigates the reactions of investors and depositors to increased risk taking by these financial institutions. The research on banks appraises the ability of market forces to discipline and supervise banks as a supplement to, or partial replacement for, regulatory supervision. Flannery (1998) provides a survey of the market discipline literature. He notes that the empirical research for banks suggests that regulators could expand their reliance on market discipline. Flannery concludes that even in the presence of bank deposit insurance and other safety nets, the market s assessment of and reaction to bank risks are rational and compare well with regulatory supervisory actions. A number of papers investigate the role of deposit insurance (or its absence) in disciplining thrifts. Park and Perisitiani (1998) created a probability-based index of the risk profile for thrifts. Using this index, they find that riskier thrifts pay higher interest rates and have lower demand for 5

7 their uninsured deposits than less risky thrifts. Goldberg and Hudgins (2002) investigated failed thrifts and found that they exhibited declining demand for uninsured deposits before failure. Their results support the market discipline hypothesis that lowering deposit insurance has the potential to increase market power to punish risk-seeking thrifts. Peria and Schmukler (2001) focused on the interest rates and flows of insured deposits in Argentina, Chile, and Mexico during the 1980s and 1990s. Although all deposits in their study are insured, depositors demand is lower and interest rates are higher for riskier banks. Consumers retain the power to discipline banks in spite of deposit insurance. Using yields on subordinated debentures for , Flannery and Sorescu (1996) conclude that their analysis soundly rejects the hypothesis that investors cannot rationally differentiate among the risks undertaken by the major U.S. banking firms. Sironi (2003) empirically evaluated the risk of the subordinated debt of European banks from , based on Moody s Banks Financial Strength rating, spreads, and accounting ratios. Her empirical results support the hypothesis that investors in subordinated notes and debentures are sensitive to bank risk. A bank rating like Moody s is often taken as a measure of the risk of a bank. But a bank rating is also regarded as a measure of the franchise value of the bank. Several studies investigate the power of franchise value to discipline banks. Herring and VanKudre (1987), Keeley, (1990) and Demsetz, Saidenberg, and Strahan (1997) find that franchise value has disciplinary value, even when deposits are insured. On the other hand, Billet, Garfinkel, and O Neal (1998) find no risk-sensitivity. They found that banks downgraded by Moody s issued more insured deposits after being downgraded. However, their finding did not carry over to the uninsured deposits. More recent study is that of Jorion, Liu, and Shi (2005) who note that because of some exclusions in the Regulation Fair Disclosure, implemented on October 23, 2000, Credit analysts at rating agencies have access to confidential information that is no longer made available to equity analysts, potentially increasing the information content of credit ratings. They did find such advantage in the post-fd period regarding investors reaction to downgrades. Research on market discipline in insurance is much less extensive than in banking. One impediment is that relatively few insurers have publicly traded stock. So the approach to 6

8 assessing market discipline through impacts on stock prices that has been effective in banking is much less effective in insurance. Froot (2007) sees evidence from pricing studies that consumers exert market discipline. The most direct study is that of Epermanis and Harrington (2006), who investigated the relationship between ratings changes and demand for insurance in the U.S. property/casualty insurance industry. Their results are consistent with the research on banks. Using premiums as a proxy for consumer demand (a measure that combines both the quantity and rate), Epermanis and Harrington demonstrate demand sensitivity to ratings downgrades, especially for commercial products, which have less of a safety net than the personal property/casualty products. In insurance the safety net is provided via guaranty funds operating in each state. These funds are analogous to the deposit insurance for banks. Epermanis and Harrington (2006) also provide a detailed summary of the insurance literature dealing with the exercise of market discipline through stock prices for the small minority of insurers that are publicly traded. Other property/casualty studies include Cummins and Lewis (2003) and Doherty, Lamm-Tennant and Starks (2003), who find that property/casualty insurers stock prices rebounded more quickly for higher A.M. Best rated insurers after the terror attacks of September 11, Our paper examines the impact of market discipline for a large panel of life insurers for the years We focus on consumer behavior as a source of market discipline, rather than on investor behavior. Unlike Epermanis and Harrington (2006), we use as demand data the actual number of life policies sold, rather than a proxy like premiums. We analyze the impact of changes in A.M. Best ratings on the demand for life insurance. To our knowledge, this is the first study of market discipline for essentially the entire life insurance industry in the U.S. As noted above, the only other studies of market discipline for life insurers that are known to us are Fenn and Cole (1994) and Brewer and Jackson (2002). Using the small population of publicly traded life insurers, they approach market discipline through investor behavior and find that life insurers with problem assets see greater stock price declines during slumps in commercial real estate and bond markets of III. DATA 7

9 Ratings. The A. M. Best rating of a life insurer is an important element in a consumer s decision to buy a life insurance product. When consumers search the Internet for coverage quotes, the first available data provided are the premiums/rates and the financial rating of the insurer. Our working hypothesis is that a consumer prefers to buy a policy from a more highly rated insurer if rates, benefits, and other conditions are similar. Likewise, when policies are bought from nonexclusive agents, the quality of the carrier will be an item of consideration by the agents. We define two ratings variables, based on the A. M. Best rating of a life insurer: We code UPGRADE = 1 if the insurer s rating improved in a given year compared with its rating in the previous year, and UPGRADE = 0 otherwise. Similarly, we code DOWNGRADE = 1 if the insurer s rating declined from the previous year, and DOWNGRADE = 0 otherwise. Our primary focus is the impact of these indicators on consumer demand for life insurance products of all types combined. Table 2 shows that only about 4.5% of the ratings actions covered by this study were downgrades; about 5.5% were upgrades. 2,053 life insurers were included in one or more years of this study. Of the 2,053 insurers, 1,371 had no ratings change; 280 had at least one upgrade and no downgrades; 221 had at least one downgrade and no upgrades; and 181 experienced at least one upgrade and at least one downgrade. Insert Table 2 Demand. Consumers can exercise market discipline by taking their business elsewhere. But there are alternative means. Life insurance policies can expire or terminate for any of the following reasons: death, maturity of term policies, disability, expiry (the end of the policy term), surrender, and lapses (when premiums are not paid). 6 Our measure of change in demand is the number of life insurance policies and certificates 7 at the end of the year, less the number of 6 Cf. the Exhibit of Life insurance in a life insurer s annual statement (e.g., page 31 of the 2003 annual statement). The list also includes conversions, decreased (net) reinsurance and aggregate write-ins for decreases (which are less consumer actions than insurer-initiated adjustments). A conversion is the change into whole life policy when a holder of a group life certificate terminates participation in the group and wishes to have a whole life policy because he or she is no longer insurable in the open market. 7 Certificate is a term used instead of policy in group life insurance. 8

10 policies and certificates at the beginning of the year. If consumers do not buy enough new policies to overcome reductions, the change is negative. In most of our analyses, we transform the year-end and year-beginning policy counts by taking logarithms and then differencing the logs. The reason for logging is to correct the skew of the count distributions. In addition, the difference between log(year-end policy count) and log (year-beginning policy count) has a convenient interpretation as (approximately) the percentage change in policy count during the year. We emphasize that our demand variables measure only the demand for life insurance policies of all types combined. Life insurers also sell health and accident insurance, annuities, and reinsurance, in addition to life insurance. But we count only life insurance policies and certificates. 8 Interestingly, the median number of policies declined for all groups, but much more so for the downgraded insurers. Median prices also declined slightly. The standard deviations for the unlogged variables bespeak heavy-tailed distributions. Therefore, we recommend the logged variables or the median to assess centrality. Price. The price of life insurance is an obviously necessary control. A downgraded life insurer may try to defend market share by lowering price. Thus the effect of a ratings change may be obscured unless price can be controlled for. We calculate an average price per $1,000 of life insurance coverage (for all types of life policies) as total annual premiums collected for in-force life insurance products divided by the total amount of life insurance in force in $1,000s. Most of our analyses use the logarithm of price. Insert Tables 3A and 3B The whole industry is working with similar mortality tables and prices are reflected in the results shown in Table 3A for all insurers. Risks/Controls. 8 The count data for life insurance policies is cleaner than it is for health and accident. In addition, it is more difficult to calculate a price proxy for the other types. Certificates are used for group life insurance to cover each group member. 9

11 An important feature of our analysis is that we imbed ratings risk in the context of other enterprise risks. We do this by adding selected proxies as controls. Among them are the following: financial leverage, a measure of asset risk (opportunity asset risk OAR based on Baranoff, Papadopoulos and Sager, 2007), the pressure of regulation (RBCRATIO), use of derivatives (IND_DERIV = 1 if yes, = 0 if no) as indication of operational risk per Cummins, Lewis and Wei (2006), income to total capital (RETONCAP), proxies for operational risks via distribution system (BROKER = 1 if products marketed primarily through brokers, = 0 if through agents), organizational structure (NTYPE = 1 if stock, = 0 if other [mutual]), and membership in an affiliated group of companies (NGROUP = 1 if yes, = 0 if no). In addition to the change in rating, we include a numerically equivalent value of the rating, converted from character value as in Colquitt, Sommer and Goodwin (1999). These proxies represent a variety of financial, asset, and operational risks. We proxy product risk by the proportions of premiums collected from life insurance (PLIFE), health and accident insurance (PHEALTH), and annuities (PANNUITY). In addition to these risk variables, we also include price (noted above) and insurer size in the form of total assets as controls. We are cognizant that upgraded, downgraded and no-change insurers may differ in terms of these enterprise risks and other variables that may be related both to the occurrence of an upgrade or downgrade and to the demand for life insurance. Thus it is important to adjust the effects of ratings changes for any pre-existing differences. This is done by including these variables as predictors, along with ratings changes, in our regression model for demand. Insert Table 4 IV. METHODOLOGY AND RESULTS We use regression to estimate the effect of rating changes on the demand for life insurance in the context of our control variables. Although there are some technical issues with OLS, we first present OLS results as a baseline and to introduce ideas. Table 5 shows the OLS regression results for the subset of Table 2 consisting of upgraded or downgraded life insurers, but excluding no-change insurers. Control variables that were statistically insignificant have been deleted from the analysis. The difference between downgraded and upgraded insurers is statistically significant (p = ). Moreover, the effect on sales can be estimated. Compared with an upgraded insurer, a downgraded insurer with otherwise the same characteristics is 10

12 estimated to experience a decline in sales of 1 - exp( ) = 10.09% in the year following the downgrade. In this analysis, a downgraded insurer has the value 1 on the DOWNGRADE variable and an upgraded insurer has the value 0. If the downgraded insurer reacts by lowering prices by 10% (and the upgraded insurer does not), the price reduction should mitigate the loss in sales resulting from the downgrade. According to the model, there are two sources for the mitigation: the variables CHANGE IN LOG PRICE and INTERACTION (the product LAG(DOWNGRADE)*CHANGE IN LOG PRICE.) By itself, the mitigation effect of CHANGE IN LOG PRICE is estimated to be 1 exp( *log(0.90)) = 2.05%; and the separate mitigation effect of INTERACTION is 1 - exp( *log(0.90)) = 5.41%. Their combined mitigation effect is estimated to be 1 exp( *log(0.90) *log(0.90)) = 7.57%. Thus the net demand reduction is estimated to be 1 exp( *log(0.90) *log(0.90)) = 3.28%, ceteris paribus. At any given set of predictor values, an otherwise identical downgraded insurer is estimated to suffer a loss of sales relative to an upgraded insurer. A few comments can be made about the remaining predictors: The annual change in demand is higher for more highly rated insurers, for more leveraged insurers, and for insurers that assume more asset risk in their investment portfolios. And if an insurer is downgraded, having a high (good) Risk-Based Capital ratio will help retain demand a little. Insert Table 5 A technical issue with this analysis arises from having a partial panel of insurers. OLS regression on panel data is generally incorrect because of autocorrelation of the year-to-year values within the same firms. However, using differenced data as the response (year-to-year changes in log policies) greatly mitigates the autocorrelation issue and would remove it altogether for random walk structures. Still, many insurers are in the data multiple times: The subset of Table 2 that is used in the Table 5 analysis numbers 682 insurers, of which 366 had either one upgrade or one downgrade only, and 316 had multiple upgrades, downgrades, or both. Re-running the Table 5 OLS on the further subset of insurers having only one occurrence in the data (and therefore presumably resolving the correlation issue) yielded results consistent with Table 5. 11

13 A still more careful analysis would adjust explicitly for the issue of within-insurer correlation. We elected not to pursue that option with the preceding comparison of downgraded to upgraded insurers. We think that it is more meaningful to compare downgraded and upgraded insurers to no-change insurers, especially since the majority of insurers (1,371 out of 2,053) experience no rating change. Applying the model to the no-change insurers makes the within-insurer correlation issue more acute. Most insurers occur multiple times in the no-change group. Thus it is imperative to adjust for within-insurer correlation when comparing insurers with ratings changes to the no-change group. To do this, we ran a GEE model (generalized estimating equations) with a couple of different within-insurer correlation structures. The results for autoregressive order 1 and unstructured within-insurer correlations were fairly similar, so we present the analysis for the AR1 case in Table 6. Again, predictors that proved statistically insignificant have been deleted from the analysis. Insert Table 6 Table 6 results are similar to those of Table 5 in broad outline. Now, a downgraded insurer, in comparison with a no-change insurer having otherwise the same characteristics, is estimated to experience a decline in sales of 1 - exp( ) = 6.20% in the following year. If the downgraded insurer reacts by lowering prices by 10% (and the no-change insurer does not), the price reduction should mitigate the loss in sales resulting from the downgrade. According to the model, there are two sources for the mitigation: the variables CHANGE IN LOG PRICE and INTERACTION (the product LAG(DOWNGRADE)*CHANGE IN LOG PRICE.) By itself, the mitigation effect of CHANGE IN LOG PRICE is estimated to be 1 exp( *log(0.90)) = 1.87%; and the separate mitigation effect of INTERACTION is 1 - exp( *log(0.90)) = 5.95%. Their combined mitigation effect is estimated to be 1 exp( *log(0.90) *log(0.90)) = 7.92%. Thus the net effect on sales is estimated to be a gain of 1 exp( *log(0.90) *log(0.90)) = +1.23%, ceteris paribus. In this case, the price reduction more than compensates for the downgrade in terms of demand. The annual change in demand is lower for insurers with higher assets, but higher for insurers with more total premiums, more leverage, more portfolio asset risk, and higher Best rating. The demand change was also elevated in

14 Insert Table 7 Table 7 shows the same GEE model as Table 6, but for upgraded insurers in relation to nochange insurers. This model shows only statistically insignificant effects for an upgrade. Both the upgrade indicator itself and its interaction with price change fail to come close to significance. Nor can the lack of significance be attributed to collinearity between the lag and its interaction, for the GEE Wald test for their combined effects lacks significance (p-value = ). Finally, we raise a fundamental question about the direction of the relationship. It could be argued that a decline in demand sets into motion a cascade of related negative financial events that later result in a ratings downgrade for an insurer. Thus, it may be argued, changes in demand lead to changes in ratings, rather than the reverse. On its face, this argument lacks plausibility, for we analyze changes in demand in the year following the rating event. In response, however, it could be argued that nothing in our data or model restricts an insurer from having experienced repeated declines in demand prior to the year in question. To address this issue, we invoke the notion of Granger causality and test for the direction of the relationship. The test involves estimating the following models: log N log 1 t π t π t = = m i= 1 m i= 1 α log N i α log N i t i t i + + m j= 1 m j= 1 β R j β R j t j t j + + r k= 1 r k= 1 λ X k λ X k k, t k, t + ε 1t ( Eq 1) ( Eq 2) In these models, the term r λ k = 1 k X k, t represents the suite of predictor variables (covariates) from Table 6. The term m i= 1 α log is the set of changes in LOG POLICY COUNT for the preceding i Nt i m years, and m β j j= 1 Rt j is the set of indicators for ratings changes for the preceding m years. These two sets of m lags appear in both equations. The idea of Granger causality is to examine the contribution to explanatory power added by the past history of ratings changes ( β j m j= 1 Rt j ) 13

15 in a model for the change in LOG POLICY COUNT that already includes the past history of changes in LOG POLICY COUNT. 9 If the past history of ratings changes is statistically significant in such a model, then the ratings change ( ( log Nt ); otherwise, Rt does not Granger-cause log Nt Rt ) is said to Granger-cause the change in LOG POLICY COUNT. So (Eq 1) permits us to test whether rating changes Granger-cause changes in demand for life insurance. Similarly, by testing the significance of m i= 1 α log in (Eq 2), we can determine whether changes in the demand i Nt i for life insurance Granger cause changes in ratings. 10 We tested for Granger causality in the data for downgraded and no-change life insurers. Results are shown in Table 8. Since we previously had found no significant impact of rating upgrades on demand for life insurance (see Table 7), we did not test upgraded and no-change insurers. In our Granger tests we used lags for the two past years (m = 2). 11 In Table 8A we find that rating downgrades Granger-cause changes in demand (p = without covariates, p = with covariates). 12 In Table 8B we find that changes in demand do not Granger-cause changes in ratings (p = without covariates, p = with covariates). 13 We interpret this result as further evidence for the direction of the relationship between ratings downgrades and reductions in demand for life insurance. 9 In a pure form of the Granger test, the covariates in Eq 1 and Eq 2 would be omitted. We present results for both forms of the test with and without covariates. The fundamental conclusion is the same in each case. 10 In (Eq 2) the response variable is the 0-1 indicator for a ratings downgrade (or upgrade). So an appropriate model is a logistic regression, which is expressed in (Eq 2) in its GLIM form. This form shows the logit link between the mean π t of the response and the set of predictors. π t is also the probability of a ratings change. 11 Lags of more than two were generally not significant; their use also reduced the sample size because fewer insurers have complete data for longer periods. 12 The analysis for Table 8A uses the same GEE methodology as in Table 6, with autoregressive error structure for each insurer over time. The suite of predictors used for the analysis shown in Table 8 necessarily differs somewhat from the predictor set for Table 6: Since the issue investigated by Table 8 is the effect of ratings change lags, the ratings change itself and its interaction with price change used in Table 6 have been removed in Table The within-insurer error structure for Table 8B was modeled as independent because analysis of the logistic regression residuals indicated an absence of statistically significant insurer-specific serial correlation for the pure Granger model. For the Granger model with covariates, a small positive within-insurer autocorrelation of about 0.05 remained and was statistically significant. Such a small autocorrelation is not likely to be of material consequence. Moreover, the effect of positive autocorrelation is to diminish further the reported statistical significance of coefficients (i.e., to increase p-value) because positive autocorrelation acts as though to reduce the effective sample size. Thus the significance of the Granger test reported in Table 8B (with covariates) is conservative. 14

16 Insert Table 8 V. Summary In this paper we have examined evidence that consumers/insureds discipline downgraded life insurers by reducing demand for their life policies. We estimate that a downgraded life insurer experiences a decline of about 6% in the number of in-force life policies in the year following the downgrade, compared with a no-change life insurer with the same risk profile, premium rates, and other study characteristics. By use of Granger causality, we have verified that the direction of the relationship flows from downgrade to demand reduction, rather than the reverse. On the other hand, there is no statistically significant impact on demand resulting from a rating upgrade. The data for the study include essentially the entire life insurance industry from in the U.S., except for smaller companies not followed by the A. M. Best Company, which supplied the rating data. This is the first study of market discipline for substantially all of the life insurance industry. Nearly universal coverage is possible because of our focus on consumer demand rather than investors reaction to downgrades. Life insurer annual statement data are sufficiently rich in information about life policies and certificates to measure the change in consumer demand from year to year, as well as to construct an aggregate price per $1,000 of life coverage. By imbedding a spectrum of enterprise risks into our model, we insure that the effects of rating changes are not confounded with those of other major risks. We expect that our findings will be of interest in the debate over the role of regulation in providing protections. Regulators, rating agencies and insurers have developed parallel models to assess insurer risk profiles in order to protect consumers/insureds from potentially insolvent insurers. Our study suggests that consumers of life insurance are not defenseless and do exert a market discipline upon insurers by reducing demand for life policies from those marked as financially weakened. The self-regulating role of market forces deserves due consideration in regulatory and industry risk models. 15

17 References A.M. Best Company, Special rating events , Oldwick, NJ. Baranoff, Etti G., Savas Papadopoulos and Thomas W. Sager (2007). Capital and Risk Revisited: A Structural Equation Model Approach for Life Insurers, The Journal of Risk and Insurance, 2007, Vol. 74, No. 3, Billet, M., Garfinkel, J., and O Neal, E. (1998). The Cost of Market Versus Regulatory Discipline in Banking, The Journal of Financial Economics 48, Brewer, Elijia III, and Jackson, William E. III (2002). Intra-industry Contagion and the Competitive Effects of Financial Distress Announcements: Evidence from Commercial Banks and Insurance Companies. Federal Reserve Bank of Chicago Working Paper No Colquitt L. Lee, David W. Sommer, Norman H. Goodwin (1999). Determinant of Cash Holdings by Property-Liability Insurers, The Journal of Risk and Insurance, Vol. 66, No 3, Phillips, R. D., J. D. Cummins, and F. Allen, (1998), Financial Pricing of Insurance in Multiple Line Insurance Companies, Journal of Risk and Insurance, 65: Cummins, J. David, and Christopher M. Lewis (2003). Catastrophic Events, Parameter Uncertainty and the Breakdown of Implicit Long-term Contracting: The Case of Terrorism Insurance, Journal of Risk and Uncertainty 26, Cummins, J. D, Christopher M. Lewis, Ran Wei (2006). The Market Value Impact of Operational Loss Events for US Banks and Insurers, Journal of Banking & Finance Volume 30, Issue 10, October 2006, Pages Demsetz, Rebecca S., Marc R. Saidenberg, and Philip E. Strahan (1996). Banks with Something to Lose: The Disciplinary Role of Franchise Value, Economic Policy Review:Federal Reserve Bank of New York 2, Doherty, Neil A., Joan Lamm-Tennant, and Laura T. Starks (2003). Insuring September 11 th: Market Recovery and Transparency, Journal of Risk and Uncertainty 26, Epermanis, Karen and Scott Harrington (2006). Market Discipline in Property/Casualty Insurance: Evidence from Premium Growth Surrounding Changes in Financial Strength Ratings, Journal of Money, Credit, and Banking - Volume 38, Number 6, pp

18 Fenn, George W., and Rebel A. Cole (1994). Announcements of Asset-quality Problems and Contagion Effects in the Life Insurance Industry, Journal of Financial Economics 35, Flannery, Mark J. (1998). Using Market Information in Prudential Bank Supervision: a Review of the U.S. Empirical Evidence, Journal of Money, Credit and Banking 30, Flannery, Mark J. (2001). The Faces of Market Discipline, Journal of Financial Services Research 20, Flannery Mark J., and Sorin M. Sorescu (1996). Evidence of Bank Market Discipline in Subordinated Debenture Yields: , The Journal of Finance 51, Froot, Kenneth A (2007) Risk Management, Capital Budgeting, and Capital Structure Policy for Insurers and Reinsurers The Journal Of Risk And Insurance, 2007, Vol. 74, No. 2, Goldberg, L.G. S.C. Hudgins (2002). Depositor Discipline and Changing Strategies for Regulating Thrift Institutions, Journal of Financial Economics 63 (2002) Herring, Richard J., and Prashant VanKudre (1987). Growth Opportunities and Risk-taking by Financial Intermediaries, Journal of Finance 42, Jorion Philippe, Zhu Liu and Charles Shi (2005) Informational Effects of Regulation FD: Evidence From Rating Agencies Journal of Financial Economics 76 (2005) Park, S. and S. Peristiani (2007) Are Bank Shareholders Enemies of Regulators or A Potential Source of Market Discipline?, Journal of Banking & Finance (2007), doi: /j.jbankfin Park, S. and S. Peristiani (1998). Market Discipline by Thrift Depositors, Journal of Money, Credit, and Banking 30, Peria, Maria S.M., and Sergio L. Schmukler (2001). Do Depositors Punish Banks for Bad Behavior? Market Discipline, Deposit Insurance, and Banking Crises. Journal of Finance 56, Sironi Andera (2003). Testing for Market Discipline in the European Banking Industry: Evidence from Subordinated Debt Issues, Journal of Money, Credit, and Banking, Vol. 35, No

19 Table 1 Statistics for Life Insurance by Year (Data from Annual Statement Exhibits (numbers of insurers vary from 1564 in 1994 to 1009 in 2003) Number of life policies and certificates Life mean premium per policy or certificate Mean face amount per policy Total premiums for life insurance Total life face amount ($1000s) Premiums/ face amount ,840,243, ,550, ,558,076,780 30, ,300,998, ,784, ,364,951,848 31, ,496,209, ,923, ,293,855,939 34, ,794,045, ,737, ,111,596,320 36, ,151,119, ,083, ,086,030,892 38, ,349,201, ,365, ,336,060,004 38, ,507,290, ,230, ,276,017,405 39, ,815,465, ,706, ,181,573,877 41, ,049,200, ,143, ,671,503,834 44, ,766,466, ,923, ,788,757,276 48, Growth factor for life mean premium per policy Growth factor for face amount per policy Growth factor for mean face amount per policy Growth factor for premium/face amount Year-change Life premium growth factor Life policies growth factor Mean growth rate Total growth for the period Source: the Exhibit of Life insurance in a life insurer s annual statement (e.g., page 31 of the 2003 annual statement). 18

20 Table 2. Distribution of ratings changes by year, number of life insurers YEAR TOTAL Downgrade No change Upgrade TOTAL

21 Table 3A. Summary statistics on changes in the number of policies and their prices for downgraded, upgraded and no-change life insurers in the year after a rating change, VARIABLE N MEAN MEDIAN Std Dev Downgrade Change in policy count ,553-1, ,612 Change in log policy count Change in price Change in log price No change Change in policy count , ,111 Change in log policy count Change in price Change in log price Upgrade Change in policy count , ,533 Change in log policy count Change in price Change in log price

22 Table 3B. Nonparametric tests of the equality of downgraded and upgraded insurers in the year after a rating change, VARIABLE Wilcoxon p-value 15 Median p-value 16 Change in policy count Change in log policy count Change in price Change in log price Since some insurers experienced more than one rating change, and some experienced both upgrades and downgrades, the independence assumption that is part of the specification of these and other two-sample tests is technically violated. A more careful analysis was also performed that takes into account the correlation between multiple occurrences of the same insurer. The result shown in Table 3B is preserved: The differences between upgraded and downgraded insurers remain significant for counts and log counts of policies; the differences remain insignificant for price and log price. 15 The two-sample Wilcoxon rank test for equality of two distributions. The two-tailed p-value is shown. 16 The two-sample median test for equality of two distributions. The two-tailed p-value is shown. 21

23 Table 4. Summary statistics on characteristics of upgraded, downgraded, and no-change life insurers, combined data VARIABLE N MEAN MEDIAN Std Dev Downgrade Best rating (numerical) Broker log(total assets) Total assets 485 3,276,646, ,417,959 15,031,686,625 States of licensure Leverage Opportunity asset risk Return on capital RBCratio , Member of group? (1/0) Stock company? (1/0) Deriv_TotalBookValue 503 2,354, ,805,076 Use derivatives? (1/0) Fraction premiums from life Fraction premiums from health Fraction premiums from annuities Life insurance in force, year end ,254,748 1,299, ,911,884 Life insurance in force, prior year end ,211,330 1,549,834 98,925,705 Policy count, year end ,778 62,251 2,868,295 Policy count, prior year end ,276 69,656 2,902,963 Change in policy count ,553-1, ,612 Change in log policy count Price, year end Price, prior year end Change in price Change in log price No change Best rating (numerical) Broker log(total assets) Total assets ,310,220,042 47,017,741 11,060,393,379 States of licensure Leverage Opportunity asset risk Return on capital RBCratio , , Member of group? (1/0) Stock company? (1/0) Deriv_TotalBookValue ,941, ,451,406 Use derivatives? (1/0) Fraction premiums from life Fraction premiums from health Fraction premiums from annuities Life insurance in force, year end ,011, ,268 75,145,381 Life insurance in force, prior year end ,819, ,677 70,369,860 Policy count, year end ,480 15,896 1,774,957 22

24 Policy count, prior year end ,408 16,479 1,780,188 Change in policy count , ,111 Change in log policy count Price, year end , Price, prior year end , Change in price , Change in log price Upgrade Best rating (numerical) Broker log(total assets) Total assets 686 1,719,199,190 96,714,144 6,964,238,493 States of licensure Leverage Opportunity asset risk Return on capital RBCratio , Member of group? (1/0) Stock company? (1/0) Deriv_TotalBookValue 693 1,361, ,409,834 Use derivatives? (1/0) Fraction premiums from life Fraction premiums from health Fraction premiums from annuities Life insurance in force, year end ,978, ,123 65,325,526 Life insurance in force, prior year end ,026, ,701 53,269,514 Policy count, year end ,900 39,864 1,595,717 Policy count, prior year end ,203 41,384 1,211,855 Change in policy count , ,533 Change in log policy count Price, year end Price, prior year end , Change in price , Change in log price

25 Table 5. OLS regression model for downgraded vs. upgraded life insurers, for year after rating change, combined data. Response variable = change in log policy count. VARIABLE Parameter Pr > t estimate Intercept <.0001 Downgrade last year? (1/0) Change in log price <.0001 Best rating (numerical) Leverage <.0001 Opportunity asset risk Interaction: lag(downgrade)*change in log price <.0001 Interaction: lag(downgrade)*rbcratio R-square

26 Table 6. GEE regression model for downgraded vs. no-change life insurers, for year after rating change, combined data, response variable = change in log policy count, autoregressive error structure within insurers. VARIABLE Parameter estimate Standard error Z Pr > Z Intercept Downgrade last year? (1/0) Change in log price <.0001 Interaction: lag(downgrade)*change in log price <.0001 log (Total assets) <.0001 log (Total premiums) <.0001 Leverage Opportunity asset risk Best rating (numerical) Year is 1996? (1/0)