The Geneva Papers, 2017, 42, (193–225) 2016 The International Association for the Study of Insurance Economics 1018-5895/16 www.genevaassociation.org
Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters Vijay Aseervathama, Patricia Bornb, Dominik Lohmaiera,* and Andreas Richtera a
Munich Risk and Insurance Center, Munich School of Management, Ludwig-Maximilians-Universita¨t Munich, Schackstraße 4/III, 80539 Munich, Germany. E-mails:
[email protected];
[email protected];
[email protected] b Department of Risk Management/Insurance, Real Estate and Legal Studies, College of Business, Florida State University, 821 Academic Way, Tallahassee, FL 32306-1110, USA. E-mail:
[email protected]
Prior studies on the effects of catastrophes on insurance markets have either focused on one specific type of hazard or pooled several natural disasters. We argue that insurers evaluate disaster risk with respect to not only the frequency and severity of disasters but also the disaster type. We analyse U.S. property insurers’ supply decisions between 1992 and 2012 and find that insurers’ responses with respect to the reduction of business volume and exit decisions differ across hazards, even after controlling for damage size. The negative effects of catastrophes on supply decisions are more pronounced after extreme hurricane years compared with tornado years. We argue that supply distortions in the aftermath of unprecedented catastrophes are driven primarily by correlated losses besides the damage size of the event. Our results show that the predictability of catastrophe losses poses less-severe threats to insurers. Thus, we propose that insurers and regulators should focus primarily on measures that encourage diversification. The Geneva Papers (2016) 42, 193–225. doi:10.1057/s41288-016-0004-5 Keywords: catastrophic risks; insurance supply; property/casualty insurance JEL Classification: D8; G22 Article submitted 28 September 2015; accepted 14 June 2016; published online 6 December 2016
Electronic supplementary material The online version of this article (doi:10.1057/ s41288-016-0004-5) contains supplementary material, which is available to authorised users.
Introduction The occurrence of natural disasters is particularly challenging for insurance companies because of the enormous accumulated damage size and the high correlation between individual losses in an exposed region. Thus, unprecedented disaster events can cause significant distortions in insurance markets. Importantly, insurers are affected by disasters in various ways. Insurers with exposure in the affected area have to reimburse their customers’ losses and suffer from a reduction of their risk-taking capital. In addition, all insurers independent of their actual risk exposure receive new information about the natural
The Geneva Papers on Risk and Insurance—Issues and Practice
194
hazard’s potential for damage.1 Hence, insurers may reassess the value of their insurance portfolio and the appropriateness of the charged insurance premiums. Consequently, insurers may decide to reduce their exposure to a certain hazard and exit markets after a major disaster event.2 For instance, Allstate Corporation significantly reduced its business in coastal states, and one of the largest personal insurers in Florida, Poe Financial, became insolvent after Hurricane Katrina. In response to insurers’ supply reactions, regulators often intervene in catastrophe insurance markets to ensure the affordability and availability of insurance coverage.3 Various empirical studies have examined insurance market reactions after natural catastrophes. However, prior empirical research is silent on hazard-specific effects on insurance supply or implicitly assumes the non-existence of such effects by limiting the analysis to the size and frequency of disaster events. Studies have either focused on one type of natural hazard,4 or have not differentiated between different types of catastrophes at all.5 Event studies on capital market reactions to catastrophic events find opposing effects on insurers’/insurance brokers’ stock returns depending on the type of hazard.6 Hence, investors’ expectations about insurers’ future profitability seem to depend on the type of event. Using results of pooled-hazard studies or single-hazard studies to derive general recommendations for appropriate regulatory interventions after natural disasters may result in unintended or unnecessary market distortions if hazard-specific supply reactions exist. In contrast to the approach of single-hazard studies, standard homeowners insurance provides coverage against multiple perils. For instance, Grace et al.7 show that ‘‘all perils’’ policies are most common in Florida (approximately 99 per cent of policies) and in New York (approximately 80 per cent of policies). These policies cover almost all major natural hazards except floods and earthquakes. Floods and earthquakes are covered mostly by governmentally sponsored programmes such as the National Flood Insurance Program (NFIP). Thus, a pooled-hazard study may also be inappropriate for examining private insurers’ supply reactions after natural catastrophes because without differentiating by hazard-type analyses might lead to misleading results. Our paper focuses on hurricanes and tornados to examine supply-side effects and, particularly, insurers’ supply reactions on private insurance markets. Both hazards are among the most severe natural hazards in terms of loss size in the U.S.8 and differ with respect to several disaster characteristics. The goal of this paper is to identify differences in insurance market reactions to natural catastrophes by comparing insurers’ supply decisions after major tornado and hurricane years. If insurers evaluate disaster risk not solely with respect to frequency and severity, differences between hazards concerning predictability and correlation of single losses could also be important determinants for insurers’ supply reactions. Determining the most challenging disaster characteristics for insurers that provide coverage against natural catastrophes is important for deriving recommendations 1 2 3 4 5 6 7 8
Froot and O’Connell (1999). Born and Viscusi (2006); Grace and Klein (2006, 2007); Born and Klimaszewski-Blettner (2013). Harrington and Niehaus (2001); Klein (2013); Medders et al. (2014). See, for example, floods or hurricanes; see Browne and Hoyt (2000) or Grace and Klein (2009). See, for example, Aseervatham et al. (2013); Born and Klimaszewski-Blettner (2009, 2013). Cummins and Lewis (2003); Hagendorff et al. (2015); Ragin and Halek (2015). Grace et al. (2004). Swiss Re (2015).
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
195
for insurers and regulators to ensure and improve the availability of insurance coverage in exposed areas. The importance of disaster characteristics for insurers’ supply decisions has been emphasised in prior studies. In particular, unexpected large catastrophes are a main driver of insurers’ market exits,9 and regionally concentrated risks tend to have a more pronounced effect on insurance supply because insurers often refrain from regional diversification to strive for economies of scale.10 In addition, the predictability of an event may also matter because both underwriters and actuaries are reluctant to insure correlated and ambiguous (with respect to loss size and the probability of loss) risks and charge significant markups.11 Despite the empirical evidence for the importance of the abovementioned disaster characteristics, prior research often neglects the differences between hazards. Our empirical setup allows us to control for these differences and to identify the most severe challenges for the insurability of natural catastrophes by comparing insurers’ supply reactions after different disaster events. We examine severe hurricane and tornado seasons in the U.S. in the period between 1992 and 2012. We retrieved data on property damages caused by extreme tornado and hurricane events for each state in the U.S. from SheldusTM. A state is classified as experiencing a catastrophic ‘‘hazard year’’ if annual, inflation-adjusted property damages caused by a certain hazard exceed the prior maximum for this hazard not only in the affected state but in all states in the U.S. To analyse insurers’ supply reactions after these extreme events, we compiled underwriting and financial statement information for the homeowners lines of all U.S. property insurers within this period at the state level from the State Pages of insurers’ annual filings with the National Association of Insurance Commissioners (NAIC). In total, our data sets contain approximately 130,000 firm-stateyear observations and we observe supply decisions of 1,576 unique firms. To compare insurers’ supply reactions (change in business volume and probability to exit a state) after different types of disaster, we run linear/logistic fixed-effects panel data regressions. In addition, this paper also includes a dynamic panel data model to examine effects on insurers’ premium income after both types of disasters utilising a two-step system general method of moments (GMM) estimator12 as the estimator is considered superior when independent variables are potentially endogenous. We find that even after controlling for the size of natural hazards, hurricanes seem to cause more severe challenges for insurers than tornados. An insurer’s probability of exiting a state or reducing business significantly increases after a severe hurricane season. We conclude that the size of a natural catastrophe seems to matter, but correlated losses are also a main driver of supply distortions. Evidence supporting market disruptions in the aftermath of extreme tornado years is rather weak. Thus, we suppose that insurers are more comfortable with addressing ambiguous risks rather than large and correlated losses. Regulatory interventions such as cancellation bans, premium subsidies or the implementation of residual markets, seem to be unnecessary after extreme tornado seasons. In contrast, hurricanes cause significant reductions in insurance supply, and the availability of affordable insurance coverage may be threatened. 9
Born and Klimaszewski-Blettner (2013). Klein and Kleindorfer (1999). 11 Kunreuther et al. (1995); Cabantous (2007); Cabantous et al. (2011). 12 See Arellano and Bond (1991); Arellano and Bover (1995); Blundell and Bond (1998). 10
The Geneva Papers on Risk and Insurance—Issues and Practice
196
The contribution of our paper is threefold: first, we provide guidelines for the evaluation of the results of prior catastrophe research. Because insurance supply reactions differ among hazards, we can conclude that generalising results of single-hazard studies may be misleading and that the reliability of pooled-hazard studies is questionable. Second, we show which disaster characteristics cause the most severe availability constraints. Correlated losses seem to be the major challenge for insurers who write business in catastrophe insurance. Third, we are able to derive recommendations for insurers and for the appropriate design of regulatory frameworks that aim to ensure the availability of sufficient insurance coverage in disaster-prone areas (e.g. encouraging diversification by aligning state regulations). The remainder of this paper is structured as follows. The section ‘‘Supply-side effects in the aftermath of catastrophes’’ provides a brief overview of existing single-hazard and pooled-hazard studies about direct supply effects on insurance markets and indirect effects on capital markets after catastrophic events. In the ‘‘Hazard-specific disaster characteristics’’ section, we emphasise differences in the disaster characteristics (size, predictability and correlation) of hazards in our analysis and develop our main research hypotheses. In the ‘‘Data and variables’’ and ‘‘Empirical model’’ sections, we present our data and empirical methodology. We summarise our results in ‘‘Results’’ section and provide robustness checks in the ‘‘Robustness checks’’ section. We conclude with a discussion of the implications of our findings in the ‘‘Discussion and Policy Implications’’ section.
Supply-side effects in the aftermath of catastrophes Effects of catastrophic events on insurance companies can be measured in two ways. First, effects on insurers’ pricing and insurers’ decisions to provide coverage can be detected by examining the market for catastrophe insurance directly. Second, to determine disasters’ effect on insurers’ future profitability, empirical studies also analyse insurers’ stock price variations immediately after the occurrence of disasters. This section summarises the literature on catastrophes’ effects on insurance markets and capital markets and highlights differences between studies’ disaster definition and analysis of hazard-specific effects. Insurance market effects In general, insurers’ supply decisions are determined by various internal and external factors. For instance, the insurable risk, company-specific characteristics and regulation affect insurers’ price and quantity decisions. Due to solvency regulations (e.g. Solvency II in Europe and Risk-Based Capital (RBC) in the U.S.), the availability of capital limits insurers’ ability to underwrite catastrophic risks.13 Because major disaster events can result in a significant reduction of insurers’ risk-taking capacity, sufficient capital buffers are crucial. According to the ‘‘capacity constraint’’ hypothesis, exposed insurers are forced to reduce supply after a catastrophic event because they are not able to replace lost capital completely, due to capital market imperfections. The reduction of available coverage and an increase in insurance demand after natural disasters should result in significant price 13
Cummins and Sommer (1996).
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
197 14
15
increases. Cagle and Harrington argue that insurers’ ability to recover from a catastrophic loss by increasing insurance prices is limited when policyholders respond to price increases and insurers’ default risk. Thus, damage size and policyholders’ price elasticity are important determinants of insurers’ supply decisions in response to natural disasters. Besides a reduction of risk-taking capital, new information about the potential for damage of a certain hazard after a disaster event can cause premium increases. Froot and O’Connell1 disentangle both effects by analysing reinsurance prices in the aftermath of different types of natural catastrophes and show that capital market imperfections (a shortage of capital) are the main reason for price increases. Insurers’ cost of bearing disaster risks is determined by their leverage ratio, size and degree of diversification. For instance, the determinants of writing earthquake risks are very similar to the costs of writing terrorism risk.16 However, Born and KlimaszewskiBlettner17 emphasise the importance of the line of business for insurers’ supply reactions, showing that homeowners lines perform worse than the commercial lines after a catastrophe. In addition, insurers struggle more to recover from disaster events in states with stricter regulation. Regulatory interventions such as rate regulation, cancellation bans or the provision of a residual market can result in severe market failures.18 Born and Viscusi19 examine the effect of ‘‘blockbuster events’’ on the U.S. insurance market and find a negative supply effect of both large and unexpected catastrophes. Despite the decrease in insurance supply, insurers seem to benefit from major disaster events because losses are accompanied by an increase in insurance demand and insurance prices. In line with Born and Viscusi,19 Zhang and Nielson20 also confirm this ‘‘demand–revenue’’ hypothesis. However, both studies do not differentiate between hazards and focus on single devastating events. For instance, the majority of Born and Viscusi’s19 ‘‘blockbuster’’ events are hurricanes (12 out of 20 events). Thus, it is unclear whether this is a hazard-specific effect or a general observation after disaster events. Klein and Kleindorfer10 and Grace and Klein,21 who investigate the effect of hurricanes on the Florida insurance market, obtain similar results. This may indicate a hurricane-specific supply effect. In particular, Klein and Kleindorfer10 find an increase in prices and a decrease in availability of insurance coverage because of increased hurricane risk. In addition, they mention institutional reasons for different market effects after hurricanes in comparison to earthquakes. Grace et al.22 argue that major disaster events, such as the hurricane season of 2004 and 2005, cause a disruption of insurers’ confidence in their own risk assessment/risk models in addition to the reduction of their risk capital. Consequently, insurers try to mitigate hurricane risks through diversification and risk avoidance when exiting states like Florida which are heavily exposed to disaster risk. However, it is an open empirical question whether insurance markets respond in a similar way to other types of natural hazards. 14 15 16 17 18 19 20 21 22
Gron (1990, 1994); Winter (1994); Cagle and Harrington (1995). Cagle and Harrington (1995). Kleffner and Doherty (1996); Thomann and von der Schulenburg (2006). Born and Klimaszewski-Blettner (2009). Born and Klimaszewski-Blettner (2013); Medders et al. (2014); Zhang and Nielson (2015). Born and Viscusi (2006). Zhang and Nielson (2015). Grace and Klein (2009). Grace et al. (2005).
The Geneva Papers on Risk and Insurance—Issues and Practice
198
Capital market effects Variation in insurers’ stock returns is often used as a proxy for investors’ expectations about insurers’ future profitability. Interestingly, the evidence on catastrophes’ effect on insurers’ stock prices is mixed and seems to depend to some extent on the type of disaster. Earthquakes tend to have a positive effect on U.S. insurers’ stock prices independently of the actual loss exposure.23 Despite the indemnity payments, investors seem to expect positive business prospects due to an increase in insurance demand. In contrast, Lamb24 and Angbazo and Narayanan25 show that hurricane events had a negative effect on insurers’ stock returns. However, only insurers that wrote business in Florida or Louisiana were affected. Unexposed insurers did not experience any stock price effect. Lamb argues that the opposing findings after hurricanes in comparison to earthquakes can be explained by hazard-specific effects. Ragin and Halek,26 who examine insurance broker returns, add that differences in the predictability of earthquakes and hurricanes might be relevant. Although the landfall of a hurricane can be predicted a few days in advance, the occurrence of an earthquake is essentially unpredictable. Cummins and Lewis27 compare the effect of the World Trade Center (‘‘9/11’’) terrorist attack on insurers’ stock prices with the effects of Hurricane Andrew and the Northridge earthquake. In general, 9/11 had a more sustained effect than the other two events. The authors argue that the size of the event, as well as uncertainty about pricing parameters after the event, could cause the effect. Interestingly, conflicting results on the effect of natural disasters can also be found in the Japanese stock market. Takao et al.28 show that stock prices of insurance companies decreased after the Great East Japan Earthquake in 2011. In contrast, stock prices tend to increase after typhoons, which have the same intensity as U.S. hurricanes, make landfall in Japan.29 Hagendorff et al.30 compare the effect of hurricanes on U.S. insurers’ stock prices directly with several other natural hazards. In their analysis, stock prices decrease after ‘‘mega-catastrophes’’ independent of the type of hazard. However, the effect of hurricanes is weaker and, in addition, Hurricane Katrina dampened stock market reactions of subsequent disaster events. The authors argue that hurricanes are the most predictable events because of the most advanced risk models and the insurance industry updated their expectations about the potential maximum loss size of natural disasters after Hurricane Katrina. Summarising the literature on capital market effects after natural disasters, we conclude that effects differ among hazard types and between insurance markets for the same hazard.
23 24 25 26 27 28 29 30
Shelor et al. (1992); Aiuppa et al. (1993). Lamb (1995). Angbazo and Narayanan (1996). Ragin and Halek (2015). Cummins and Lewis (2003). Takao et al. (2013). Yamasaki (2015). Hagendorff et al. (2015).
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
199
Table 1 Overview of catastrophic event definitions Author
Type of study
Definition of catastrophic event
Hazard differentiation
Shelor et al. (1992) Aiuppa et al. (1993) Lamb (1995) Angbazo and Narayanan (1996) Froot and O’Connell (1999) Cummins and Lewis (2003) Born and Viscusi (2006) Grace et al. (2005) Grace and Klein (2009) Born and KlimaszewskiBlettner (2013) Hagendorff et al. (2015)
Capital market
1989 California earthquake
Single hazard
Capital market
Hurricane Andrew
Single hazard
Insurance market
An event that gives rise to USD 15 million or more in insured losses
Different hazards for each region
Capital market Insurance market Insurance market
WTC attack, Hurricane Andrew, Northridge earthquake 20 most costly natural catastrophes
Comparison of three blockbuster events Pooled hazards
Hurricane season 2004 and 2005
Single hazard
Insurance market
Proxy for major disaster events: Natural disasters damages per state [1 Premiums written per state
Pooled hazards
Capital market
19 ‘‘mega-catastrophes’’ with first insured loss estimates exceeding USD 1 billion
Ragin and Halek (2015)
Capital market
43 largest insured-loss catastrophes since 1970
Zhang and Nielson (2015)
Insurance market
An event that generates claims of at least USD 25 million and affect a significant number of policyholders and insurers
Differentiation between hurricanes and other natural disasters Differentiation between earthquakes, WTC attack and other disasters Pooled hazards
Disaster definition and differentiation by disaster type Table 1 provides an overview of definitions of catastrophes used in prior studies. While there is no general definition of catastrophe, two major approaches are predominantly used. Studies either focus on single blockbuster events such as the WTC terrorist attacks or use an arbitrary size criterion, e.g. an event that generates claims at least of USD 25 million. Papers focusing on direct insurance market reactions predominantly pool over all hazards and do not disentangle hazard-specific effects. The majority of studies on capital market effects examine single blockbuster events. Our cross-study comparison provides evidence for differences in market reactions among natural hazards. Some capital market papers differentiate between hazards in their analysis and find significant differences depending on the type of disaster.31 However, capital market studies 31
Cummins and Lewis (2003); Hagendorff et al. (2015); Ragin and Halek (2015).
The Geneva Papers on Risk and Insurance—Issues and Practice
200
have only limited explanatory power for effects on the insurance market. The applied event study methodology of these papers focuses on single devastating events and their short-term effects on the stock market. In addition, changes in stock prices only reflect investors’ expectations, which do not necessarily coincide with the actual developments in the insurance market. Thus, a comprehensive empirical analysis of hazard-specific effects on the insurance market seems worthwhile because actual changes in the insurance market do not occur a few days after a disaster event but may take some time to develop. In addition to the empirical evidence for hazard-specific effects from the capital market, we argue in the following section that differences in disaster characteristics may explain why insurers’ supply reactions after tornados and hurricanes may differ. Based on this theoretical reasoning, we derive several hypotheses which we aim to test in our subsequent empirical study.
Hazard-specific disaster characteristics Besides testing for differences in insurers’ supply reactions depending on the hazard type, a second goal of this paper is to test which disaster characteristics may cause the opposing results. Hence, this section outlines differences between disaster events in characteristics that may affect insurers’ supply reactions after major catastrophic losses. Our analysis focuses on two hazards: hurricanes and tornados. Both hazards belong to the most severe natural hazards in terms of loss size in the U.S.8 Although earthquakes and floods are also large-scale natural hazards in the U.S., we neglect them in our analysis because they are covered, at least in some states, by government-sponsored insurance programmes. These events should cause no reaction or distorted supply reactions in the private insurance market. Importantly, only wind damage caused by hurricanes is normally covered by private insurance policies.32 We argue that hurricanes and tornados differ with respect to size, correlation and predictability. Based on these criteria, we derive differing predictions with respect to supply effects in the aftermath of these catastrophes. We assume that unprecedented disaster seasons cause a re-evaluation of an insurer’s risk exposure and might result in an adjustment of insurance supply. Size/severity Hurricane and tornado events differ with respect to their severity. An analysis of the SheldusTM data, a comprehensive natural hazard database, shows that hurricane damages are, on average, approximately four times greater than tornados in terms of annual property damage. According to the ‘‘capacity constraint’’ hypothesis, negative capital shocks can cause short-term supply distortions with higher prices. Thus, one should expect, ceteris paribus, stronger supply reductions after hurricanes than after tornados. In addition, we observe differences with respect to standard deviation, maximum loss size and skewness. Table 2 presents the descriptive statistics of our hazard-specific annual property damage data at the state-level, which will be discussed further below. 32
A controversy arose after Hurricane Katrina concerning whether damages had been caused by wind or water because the differentiation was difficult in cases where nothing but the building’s foundations remained (e.g. Kunreuther and Michel-Kerjan 2009). However, detailed information about the cause of damage in our loss database allows a differentiated analysis.
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
201
Table 2 Hazard-specific annual property damages per state (time period: 1985–2012) Mean (USD millions)
Std. (USD millions)
Max. (USD millions)
Skewness
Events
Events/ observations
80.3 21.7
1,001 147
23,300 4,220
17.60 21.47
168 1,068
0.12 0.73
Hurricane Tornado
Correlation High correlation of individual claims is a decisive characteristic of catastrophic risks. Raykov33 shows theoretically that, due to the correlation of single losses, market failure can be a stable equilibrium in catastrophe insurance markets. Thus, even after controlling for differences in loss size between hazards, we still expect different supply reactions of insurers after tornados and hurricanes because of differences in correlation. Whereas tornados appear all over the U.S. with a relatively high frequency (in approximately 70 per cent of our state-year observations in the 1985–2012 period), hurricanes occur in approximately 12 per cent of the state-year observations and only in coastal states. Deryugina34 shows that hurricanes primarily affect nine states and that such storms rarely occur outside this region. Figure 1 highlights the differences between the damage distributions of hurricanes and tornados in the U.S. Whereas the lion’s share of hurricane damages occurs in the Gulf coast area, all U.S. states are exposed to tornado risk. In addition, the two maps in Figure 1 indicate that most coastal states with hurricane exposure between 1985 and 2012 face average annual property damages above USD 100 million. In contrast, only two states (Alabama and Missouri) are exposed to similarly sized tornado risk. In line with the correlation of single losses, the scope of the affected area also matters for diversification. The ‘‘footprint’’ of a hurricane—i.e. the scope of the area affected—differs from that of tornados. Whereas single hurricanes usually affect large areas, tornados are rather local phenomena. Hence, insurers are forced to write business in a geographically larger area to generate a diversified hurricane portfolio than the area required for a diversified tornado risk portfolio. However, holding the number of policies constant, insurers with more regionally concentrated portfolios benefit from economies of scale in marketing and sales distribution. Insurers often have to trade-off between regional concentration to receive efficiency gains and diversification benefits.10 We conclude that diversification of tornado risks is easier than diversification of hurricane risks in terms of both temporal and regional diversification. Because insufficiently diversified risks can result in large claims’ payments at one point in time, insurers may struggle more after an unprecedented hurricane event. Thus, insurance supply should be reduced more heavily after hurricanes than after tornados. Predictability The ease of predicting disaster events concerns two dimensions: the event date and the affected area. The predictability of hurricane events differs from tornado events in both 33 34
Raykov (2015). Deryugina (2011).
The Geneva Papers on Risk and Insurance—Issues and Practice
202
Figure 1. Average annual property damages by state and by hazard (time period: 1984–2012). Source: SheldusTM.
dimensions. As mentioned previously, the occurrence of hurricanes is usually limited to the Gulf coast area. In contrast, tornados can basically occur in all U.S. states. Differences in the predictability of both hazards are highlighted by the following statements: ‘‘The Hurricane season in the Atlantic begins June 1st and ends November 30th’’,35 in contrast ‘‘tornados have been known to occur in every state in the United States, on any day of the year, and at any hour’’.36 Hence, there is more ambiguity in assessing tornado risks than in assessing hurricane risks. This evaluation is also backed by the University Corporation for Atmospheric Research; Henson37 states that the Storm Prediction Center of the National Oceanic Atmospheric Administration (NOAA) publishes tornado forecasts eight days in advance at maximum, whereas hurricane seasons can be predicted up to one year in advance. Henson also refers to a member of the NOAA who believes that the tornado outlooks should ‘‘[…] move toward something probabilistic, much like the hurricane outlooks’’. Several studies have shown that insurers are reluctant to underwrite ambiguous risks and, thus, they set a price markup on coverage for these risks. Price markups are particularly large if ambiguity comes from conflicting information instead of imprecise knowledge about probabilities.38 Swiss Re8 highlights the challenges for accurate convective storm 35 36 37 38
National Hurricane Center (2014). National Severe Storms Laboratory (2014). Henson (2013). Kunreuther et al. (1995); Cabantous (2007); Cabantous et al. (2011).
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
203
(tornado) risk modelling and emphasises the shortcomings of existing models. Most models rely on historical data of questionable quality and struggle with the localised nature of convective storms when assessing the risk exposure of a location. In general, model uncertainty and insurers’ ambiguity aversion can cause disruptions in the catastrophe insurance market.39 Hagendorff et al.30 argue that weaker capital market reactions after hurricanes compared with other hazards are due to more advanced hurricane risk models which provide more accurate estimations of potential losses. Grace et al.22 also argue that lost confidence in their risk assessment reduces insurers’ willingness to provide coverage against catastrophe risks after an unprecedented event. If predictability is crucial for underwriting decisions, we should see stronger insurance market reactions in the aftermath of unprecedented devastating tornado years, challenging the predictions of existing risk models. Our analysis can be summarised as shown in Table 3; although hurricanes cause more problems with respect to size and correlation, tornados cause more challenges with respect to predictability. Table 3 depicts a comparison of the insurability of these hazards. A positive sign represents a more demanding challenge for the insurer, whereas a negative sign represents a less important problem. Hence, we are not able to derive clear theoretical predictions for differences between insurers’ supply reactions after hurricanes and tornados. If one of the examined disaster characteristics is more challenging to address for insurers, we should observe differences in our empirical analysis. If the size of a disaster event affects insurers’ supply reactions, we should capture the effect by controlling for disaster losses. Thus, the first statistical hypothesis is as follows: Hypothesis H1: Size Supply reactions after catastrophes depend on the size of disaster losses. In our analysis, we account for the size of a disaster by including property damages caused by tornados and property damages caused by hurricanes separately. If damage size is the only determinant for insurers’ supply reactions after a disaster event, we should not detect any additional hazard-specific effect after controlling for property damages caused by both perils. The analysis of further disaster characteristics will help us to identify disaster-specific effects and to derive implications for regulators and insurers. If hurricanes cause more severe supply distortions after controlling for the severity of a disaster season compared with tornados, we conclude that the correlation of single losses is a more severe challenge for insurers. Hypothesis H2: Correlation Correlation of single losses is the most challenging disaster characteristic after controlling for damage size. If extreme tornado seasons result in more extreme supply effects compared with hurricane seasons after controlling for damage size, we argue that the predictability of a disaster event is a more crucial determinant for insurers supply reactions. Hypothesis H3: Predictability Predictability of a disaster event is the most challenging disaster characteristic after controlling for damage size. 39
Chen and Sun (2012).
The Geneva Papers on Risk and Insurance—Issues and Practice
204
Table 3 Summary of disaster characteristics
Size Correlation Predictability
Hurricane
Tornado
+ + -
+
Data and variables The data set for our empirical analysis is a blend of two data sources. We received financial and underwriting information40 for all U.S. property insurers at the state level from the state pages of insurers’ annual filings with the National Association of Insurance Commissioners (NAIC). We obtained direct premiums earned and losses incurred for all homeowners lines from 1992 to 2012. In total, our data set contains approximately 130,000 firm-state-year observations. Our sample consists of 1,576 unique firms that were active in at least one state.41 We supplement our data with information on the catastrophe exposure of each state. Therefore, we compile natural disasters from SheldusTM, which provides hazard data for 18 different natural hazard event types at the county level. To analyse the hazard-specific supply reactions of private insurance companies, we focus on two hazard types, hurricanes and tornados, as discussed previously. Given that we are interested primarily in large catastrophe damages that threaten the financial viability of private insurers operating in an affected state, we compile the annual property damages42 by hazard type and state. Event definition We classify a state as experiencing a catastrophic ‘‘hazard year’’ if the annual, inflationadjusted property damages for a certain hazard exceed the prior maximum for this hazard in any state. We introduce indicator (dummy) variables for each hazard type, which equal 1 if our event definition is met.43 We use this event definition instead of simply accounting for the occurrence of blockbuster events for several reasons. First, we assume that policyholders and the media focus less on the annual loss exposure of a state, but rather, are more focused on single devastating events.44 Therefore, we are able to reduce distorting effects caused by changes in insurance demand and, as a result, we are more likely to observe a less biased supply reaction. Second, an event definition based on the size of single events would result in an over-representation of hurricanes, which cause large-scale damages at one point in time. Third, many small or medium-sized hazards in a year could also cause solvency problems for insurance companies; this would be ignored if only a single blockbuster event definition were used. Insurers do not evaluate their underwriting performance based on a single event but compare the incurred losses in a year with their 40 41 42 43
44
All financial data were inflation adjusted and expressed in 2012 USD. To exclude inactive insurers, we dropped firms with less than USD 1,000 in premiums earned per year. All damages were expressed in 2012 USD to account for inflated damages in more recent years. Consequently, the reference category is a combination of ‘‘minor hurricane or tornado events’’, ‘‘floods’’, ‘‘earthquakes’’, ‘‘other hazards’’ and ‘‘no events’’. Gallagher (2014).
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
205
premium income in the respective year. Fourth, insurers must also rely on seasonal predictions rather than predictions on an event basis, if at all.37 This also backs the aggregation of hazard damages on an annual basis. To account for the fact that insurers elicit informational value from past events, we extend our natural catastrophe database to years before 1992. Therefore, we apply our event definition already beginning in 1985 and classify a state as experiencing a major ‘‘hazard year’’ if the annual hazard-specific property damages exceed the prior maxima beginning in 1985. According to Swiss Re,45 damages caused by natural catastrophes have increased dramatically since 1985. Thus, we are confident in capturing most major disaster years in insurers’ loss history. Table 4 provides a summary of events after applying our event definition for the entire period. The italicised year-state observations represent the event history, whereas the light-grey-shaded observations are events included in our analysis.46 To control for differences in damage size between hazards, we include hazard-specific property damages for each state in each year. Remaining differences in the effect of catastrophes on insurance supply cannot be explained by the severity of events but might be due to the correlation of single losses or the ease of predicting an event. Dependent variables Because we aim to examine differences in insurers’ supply reactions in the aftermath of natural disasters depending on the type of hazard, we focus our analysis on two criteria: business volume and the probability of exiting a state. We use two approaches to examine the effect of catastrophic events on the business volume of insurers operating in the affected state. First, we analyse the direct effect on premiums earned. Second, we calculate the percentage change in premiums earned between the event year and the following year47 and identify insurers that reduce their business by more than 60 and 80 per cent. Business volume in a state is affected by both price and quantity effects. According to the literature, premiums tend to increase after a disaster event, whereas the quantity of coverage decreases in disaster states.48 Thus, it might be difficult to infer supply reactions from changes in premium volume. Overall premiums earned could increase, although insurance supply decreases. In contrast, if an insurer decides to leave an affected state, quantity will decrease by definition. For the sake of simplicity, in line with Born and KlimaszewskiBlettner,9 we classify an insurer as exiting a state if its premium earnings are zero in the following two years. This definition is rather restrictive because only complete exits are captured. It does not differentiate between insolvencies, voluntary exits, M&A transactions and other events that may trigger an exit. Thus, we measure the overall supply-side effect of catastrophic events. Nevertheless, we also provide estimation results focusing on active management decisions. In contrast to insolvencies and mergers, an insurer who stops writing business in one state but still provides coverage in other states (=voluntary exits) makes a predominantly active strategic management decision.
45 46 47 48
Swiss Re (2014). Without introducing an event history, the first year of our analysis (1992) would be an event year by definition. Specifically, ½Premiums Earnedðt þ 1Þ Premiums EarnedðtÞ=Premiums EarnedðtÞ. See, for example, Grace et al. (2005); Born and Viscusi (2006).
The Geneva Papers on Risk and Insurance—Issues and Practice
206
Table 4 Summary of catastrophic events (time period: 1985–2012) Year
State
Annual property damages in USD millions
1985 1985 1989 1991 1996 1998 1999 2001 2004 2005 2011
MS OH SC KS AR FL OK TX FL LA AL
1,030 127 4,940 185 441 510 1,460 6,300 19,500 23,300 4,220
Hazard Hurricane Tornado Hurricane Tornado Tornado Tornado Tornado Hurricane Hurricane Hurricane Tornado
Control variables Insurers’ business mix and regional diversification determine the cost of risk bearing.49 Insurers’ degree of diversification is captured by two variables. We account for insurers’ regional diversification by using the ‘‘number of states’’ in which an insurer operates as a proxy. The second dimension of diversification, product diversification, is captured by the number of lines of an insurer within each state. Another important determinant of insurers’ supply reactions after disaster events is the access to financial resources within and outside the company. Thus, we add the insurers’ reinsurance ratio50 of the respective line of business. An insurer that transfers the lion’s share of its risk exposure should respond less to extreme catastrophe years. If an insurer belongs to a larger insurance group, capital can be transferred from other parts of the group to single insurers in the case of a major loss event. To account for inter-group relationships, we compute the share of each insurer’s homeowner business in the group’s homeowners business. Insurers’ surplus is used as a proxy for the size and financial viability of an insurer. We also control for the age of an insurer because it is considered a measure of experience in catastrophe insurance. After catastrophic events, new insurance companies enter the market to benefit from rising premiums but often exit the market or go bankrupt after suffering losses in subsequent years.10 Not all insurers in a state suffer to the same extent from a major disaster event, and the underwriting results vary significantly among insurers. The loss ratio indicates whether an insurer received sufficient premium income to cover its loss expenses. To detect to what extent an insurer’s underwriting result is affected by a disaster year, we compute the deviation of the insurer’s loss ratio from the average loss ratio in the last three years. The importance of the business operation in a state compared with the nationwide business affects insurers’ likelihood of exiting a state after a disaster event. Hence, we calculate the share of homeowners business in one state in an insurer’s nationwide homeowners business. In addition, we control for an insurer’s market share in the
49 50
Kleffner and Doherty (1996). In line with Cole and McCullough (2006), we calculate the reinsurance ratio as premiums ceded to non-affiliates divided by the sum of direct premiums written and reinsurance business assumed from non-affiliates.
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
207
homeowners market in each state to account for the competitiveness of an insurer in the respective state. The market environment in each state might also affect insurers’ supply decisions. Regulatory scrutiny is known to influence insurers’ willingness to write business in a state.51 In particular, regulating the rate-setting process can reduce insurers’ ability to recover from a disaster event. We use a dummy variable to indicate the existence of a prior approval rate-regulation system in the homeowners market in the respective state to account for regulation. The wealth level in a state can affect the demand for insurance and loss exposure. Hence, we use the ‘‘per capita income’’ in each state to control for differences in wealth between states. Finally, the degree of competition in a state influences insurers’ ability to raise premiums after a major disaster year and might drive inefficient companies out of the market. We compute a Herfindahl–Hirschman index for the homeowners business in each state to account for the degree of competition among the insurers in a state. A summary of all variable descriptions is provided in Table A1 in the online Appendix. Descriptive statistics The descriptive statistics of our data set are summarised in Table 5. We report our dependent variables, the number of lines within a state, an insurer’s market share in a state, the share of an insurer’s state business in the nationwide business and the deviation of the insurer’s loss ratio from the three-year average for each of our panel observations. Hazard data, regulatory information, ‘‘per capita income’’ and the ‘‘state HHI’’ are reported at the state-year level, and all company-specific control variables are reported at the firm-year level. Insurer-state observations differ strongly in terms of premium income which is not surprising given the structure of the homeowners insurance market in the U.S.52 Based on our data set and our definition of an exit, insurers exited states in approximately 7 per cent of our observations. The average insurer operates in multiple lines (approximately 11) within a state and has business operations in approximately seven states. However, the degree of diversification varies in our sample. Some insurers limit their operations to a single state, whereas others write business across the entire U.S. In contrast to Table 2, we report in Table 5 only the hazard data that we use in our analysis. The mean of the property damage variables is greater than that of the full sample. Competitiveness of the homeowners insurance market varies between states. The Herfindahl–Hirschman index indicates that market concentration in homeowners insurance is differs significantly between states. Approximately a third of the states in our sample have a prior approval rate-regulation system. One concern of our identification strategy may be that our definition of a major disaster event cannot be differentiated from large loss events. Hence, the driver of potential supply effects could be just the size of the event. If our event indicators were just equivalent to large loss events, they should be highly correlated with the annual hazardspecific property damages. The inclusion of both types of variables (event dummies and property damages) in our estimation model would also result in multicollinearity 51 52
Klein (2008); Born and Klimaszewski-Blettner (2013). Insurance Information Institute (2016).
The Geneva Papers on Risk and Insurance—Issues and Practice
208
Table 5 Summary statistics of insurance companies and catastrophe damages in the U.S. (time period: 1992–2012) Variable Year-state-firm observations Premiums earned (in USD thousands) Exits Voluntary insurer exits Business reduction ([80 per cent) Business reduction ([60 per cent) Number of lines in state s Market share in state s Share in nationwide homeowners business Deviation from three-year average loss ratio State-year observations Hurricane event Tornado event Property damages hurricane (in USD thousands) Property damages tornado (in USD thousands) Rate regulation Homeowners HHI of state s Per capita income Firm-year observations Number of states insurer operates Share in group homeowners business Age Surplus (in USD thousands) Reinsurance ratio
Obs.
Mean
SD
Min
Max
8,563.1780 0.0688 0.0540 0.1072 0.1215 10.9145 0.0081 0.1506 0.0336
43,600 0.2531 0.2260 0.3093 0.3267 5.4007 0.0264 0.2900 4.6857
1.008 0 0 0 0 0 0 0 -449.9478
1,740,000 1 1 1 1 33 0.4684 1 1,181.1030
1,050 1,050 1,050
0.0029 0.0038 104,000
0.0534 0.0616 1,170,000
0 0 0
1 1 23,300,000
1,050
27,300
172,000
0
4,220,000
1,050 1,050 1,050
0.2981 0.0941 30,752.98
0.4576 0.0351 8,327.804
0 0.0249 14,651
1 0.2455 58,908
6.6415 0.4796 59.0109 260,000 0.0912
10.7763 0.4324 46.4121 1,070,000 19.8931
1 0 1 34,600 -2389.3340
50 1 218 21,800,000 399.7850
129,390 123,216 123,216 123,216 123,216 129,390 129,390 129,390 100,042
19,482 19,482 19,390 19,361 19,359
problems. By reporting the correlation coefficients between our event indicators and the annual hazard-specific property damages in Table 6, we show that there is only a very moderate to low correlation. The correlation coefficient of the hurricane event dummies and annual hurricane-specific property damages is 0.196. Correlation between the tornado event indicator and annual property damages of tornados is even lower (0.089). To make sure our estimations are not confounded by multicollinearity of our independent variables, we conduct several multicollinearity tests for all independent variables (event dummies, hazard-specific property damages and control variables). None of our tests indicates any multicollinearity issues. For example, all variance inflation indicators range between 1 and 2. We conclude that our estimations are unlikely to suffer from multicollinearity problems and our event definition is not similar to large loss events but contains additional information.
Empirical model To examine whether insurers’ supply reactions after disaster events differ between unprecedented devastating tornado and hurricane seasons, we regress our insurance supply
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
209
Table 6 Correlation tables of main variables
Hurricane event Tornado event ln(property damages hurricane) ln(property damages tornado)
Hurricane event
Tornado event
ln(property damages hurricane)
1.000 -0.003 (0.261) 0.196*** (0.000)
1.000 0.067*** (0.000)
1.000
0.038*** (0.000)
0.089*** (0.000)
0.187*** (0.000)
ln(property damages tornado)
1.000
Note: *, ** and *** denote statistical significance at the 10, 5 and 1 per cent levels, respectively.
measures on the disaster event dummies and on a set of control variables. Therefore, we perform a series of panel data regressions that analyse the effect of both hazard types on premiums earned, the probability of exiting a state, and the probability of reducing business by more than 60/80 per cent. Our data are reported by firms on a state-level basis. Hence, each observation in our data set is a unique state-firm entity. According to Greene (2008), estimating panel data with a pooled ordinary least square regression can result in biased estimates due to unobserved heterogeneity, i.e. unobserved variations between and within cross sections. Thus, we first run fixed-effects panel data regressions to account for this potential bias. By using both firm-state fixed effects and year fixed effects, we are able to control for omitted variable biases. Our empirical model is supported by the Breusch–Pagan Lagrangian multiplier test that rejects the null hypothesis of no entity-specific and period-specific effects (p value \0.01). In addition, the Hausman test supports the choice of a fixed-effects model, because based on the v2 test statistics, we can reject the hypothesis that both random and fixed effects are consistent estimators (p value \0.01). Besides controlling for period-specific effects by including year dummies, we also add lagged event dummies (one and two years after an unprecedented disaster season). We argue that insurers are not able to change all of their business operations in a state immediately and it may take some time until managerial decisions are observable in the market. For instance, cancelling policies or shutting down entire business units within a state is often not possible immediately after a disaster has occurred. Therefore, we believe that analysing the first and second years after a major catastrophic event is particularly insightful. Because we aim to identify other hazard-specific characteristics with our event dummies, we control for the damage size by using the natural logarithm of hurricane and tornado property damages. We rely on the natural logarithm in our empirical model because damage data are truncated at zero and are additionally positively skewed as shown in Table 2. Firm- and state-specific control variables, as outlined in the data section, are summarised by Xist in the regression equation. Hence, we estimate our dependent variables (ln(premiums earned), Pr(Exit), Pr(Business Reduction)) against our set of explanatory variables:
The Geneva Papers on Risk and Insurance—Issues and Practice
210
Table 7 Panel data regression results—linear fixed-effects and system GMM: dependent variable = ln(premiums earned) Variables
Linear fixed effects ln(premiums earned) Coefficient [SE]
System GMM ln(premiums earned) Coefficient [SE]
Hurricane event
-0.254*** [0.082] -0.010 [0.044] -0.048 [0.068] -0.028 [0.035] -0.105 [0.080] -0.065* [0.034] -0.001** [0.001] 0.000 [0.001] -0.001 [0.001] 0.000 [0.001] -0.001 [0.001] 0.000 [0.001] 27.503*** [5.654] 1.958*** [0.150] 0.077*** [0.010] 0.047*** [0.004] 0.408*** [0.111] 0.004 [0.003] 0.000 [0.000] 0.002*** [0.001] -0.005*** [0.002] -0.054* [0.033] -1.984*** [0.763]
-0.276** [0.128] -0.011 [0.105] -0.092 [0.150] 0.064 [0.118] -0.218 [0.156] -0.182** [0.083] 0.004 [0.003] 0.004 [0.004] 0.003 [0.003] 0.004 [0.004] 0.006* [0.003] 0.004 [0.004] 5.954* [3.186] 0.211 [0.322] 0.025** [0.012] 0.015*** [0.004] -0.479*** [0.118] -0.021*** [0.003] 0.000*** [0.000] 0.002 [0.012] 0.011 [0.016] -0.290 [0.256] -2.150 [4.783]
Tornado event Hurricane event(t-1) Tornado event(t-1) Hurricane event(t-2) Tornado event(t-2) ln(property damages hurricane) ln(property damages tornado) ln(property damages hurricane)(t-1) ln(property damages tornado)(t-1) ln(property damages hurricane)(t-2) ln(property damages tornado)(t-2) Market share in state s Share in nationwide homeowners business Number of lines in state s Number of states insurer operates Share in group homeowners business Age Surplus in USD millions Reinsurance ratio Deviation from three-year avg. loss ratio Rate regulation Homeowners HHI of state s
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
211 Table 7 (continued) Variables
Linear fixed effects ln(premiums earned) Coefficient [SE]
System GMM ln(premiums earned) Coefficient [SE]
Per capita income in USD thousands
0.006 [0.005]
0.082** [0.038] 0.739*** [0.046] 2.772*** [0.977] 94,176
ln(premiums earned)(t-1) Constant
10.886*** [0.220] 99,777 0.259 10,865 YES YES 0.258 20.81 43
Observations R2 Number of panelvar Insurer FE Year FE Adj. R2 F value df AR(1) AR(2) AR(3) AR(4) Hansen test Diff-in-Hansen test
10,518 YES
44 -3.36*** -5.98*** -3.09*** 1.37 243.51*** 19.91* (12)
Notes: Standard errors are robust to heteroscedasticity and clustered at the firm-state level in the linear fixed effects model. The two-step GMM estimates are reported with corrected standard errors (Windmeijer 2005). AR(1)– AR(4) are tests for serial correlation in the first-differenced residuals (Arellano and Bond 1991). (Difference) Hansen statistics are v2 distributed; number in brackets behind difference Hansen test provides the number of restrictions/degrees of freedom. *, ** and *** denote statistical significance at the 10, 5 and 1 per cent levels, respectively. Dollar values are inflation-adjusted and expressed 2012 USD and premiums earned were winsorised at the 1/99 per cent level.
Dependent variableist ¼ h0 þ
2 X
hjþ1 HurricaneEvents;tj þ
j¼0
þ
2 X
hjþ4 TornadoEvents;tj
j¼0
hjþ7 lnðHurricaneDmgÞs;tj þ
0
þ cXist þ
2 X
2 X
hjþ10 lnðTornadoDmgÞs;tj
j¼0 21 X j¼1
dj Y j þ
n X
li F i þ eist
i¼1
All control and event dummies introduced in the ‘‘Data and variables’’ section are used in each regression model. Depending on our measure for insurance supply, we conduct a linear panel data regression (premiums earned) or a logistic panel data regression (probability to exit,53 probability to reduce business by more than 60/80 per cent). To 53
Due to our definition of an exit, we lose another year of observations (2012).
The Geneva Papers on Risk and Insurance—Issues and Practice
212
Table 8 Wald test results for premiums earned regression models Dependent variable: ln(premiums earned) Linear fixed effects regression Year
t t-1 t-2
Event
Damages
F-statistic
p value
F-statistic
p value
6.77*** 0.07 0.20
0.0094 0.7915 0.6524
3.08* 1.36 2.63
0.0795 0.2438 0.1054
System GMM model Year
Event 2
t t-1 t-2
Damages 2
v -statistic
p value
v -statistic
p value
3.00* 0.73 0.06
0.0835 0.3932 0.8111
0.01 0.03 0.10
0.9373 0.8544 0.7526
account for a potential within-firm correlation over time, we adjust standard errors in the linear panel data regression for clustering at the insurer level. In addition to the standard fixed effects model, we conduct a general method of moments (GMM) estimation to examine the effect of both hazard types on ‘‘premiums earned’’. This allows the inclusion of lagged dependent variables (ln(premiums earned)t-1)) and relaxes the assumption of strict exogeneity meaning that independent variables may be correlated with past and/or current realisations of the error term. Arellano–Bond54 and Arellano–Bover/Blundell–Bond55 developed two types of dynamic panel data estimator: difference GMM and system GMM.56 Fixed effects are removed by first taking differences of the regressors. However, this imposes correlation between error terms and differenced lagged dependent variables. Both approaches overcome this problem by using instruments. One advantage of both approaches is that they do not rely on external instruments to overcome endogeneity problems, as they use the lags of potentially endogenous variables for instrumenting them. By imposing an additional assumption that first differences are not correlated with fixed effects, the system GMM estimator can increase efficiency dramatically. Hence, we decided to implement a two-step system GMM estimator. By using the system GMM estimator, the model is estimated simultaneously in differences and levels. Lagged levels/lagged differences of the independent variables are used as instruments in the difference/levels equation, respectively. To obtain valid standard errors, we apply a methodology suggested by Windmeijer57 to receive a corrected variance estimator. As the consistency of the estimators relies on the assumption of no serial correlation in the error terms eist and on the validity of the instruments, we conduct two tests after running 54 55 56 57
Arellano and Bond (1991). Arellano and Bover (1995); Blundell and Bond (1998). We use the command xtabond2 developed by David Roodman to implement the GMM estimator in Stata. Windmeijer (2005).
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
213
Table 9 Logit results: dependent variable = response (insurer exit/voluntary insurer exit/business reduction) Explanatory variable
Exits coefficient [SE]
Business reduction [ 80 per cent Coefficient [SE]
Business reduction [ 60 per cent Coefficient [SE]
Hurricane event
0.069 [0.392] -0.184 [0.375] 1.031** [0.406] -0.491 [0.427] 1.202*** [0.427] -0.217 [0.348] 0.011 [0.007] 0.010 [0.008] 0.012* [0.007] 0.008 [0.008] 0.004 [0.007] -0.000 [0.008] -339.248*** [25.868] -0.822*** [0.316] -0.459*** [0.018] -0.153*** [0.008] 0.318 [0.211] -0.024** [0.010] -0.000*** [0.000] -0.005 [0.004] 0.003 [0.002] -0.176 [0.279] 4.496 [4.028]
0.281 [0.271] -0.121 [0.246] 0.877*** [0.284] -0.126 [0.251] 1.153*** [0.321] -0.239 [0.234] 0.008* [0.004] 0.004 [0.005] 0.016*** [0.004] 0.003 [0.005] 0.004 [0.005] 0.003 [0.005] -215.112*** [13.334] -0.968*** [0.226] -0.307*** [0.011] -0.107*** [0.005] -0.109 [0.139] -0.014*** [0.005] -0.000*** [0.000] -0.001 [0.001] 0.005 [0.004] -0.919*** [0.172] 6.083*** [2.303]
0.270 [0.254] -0.139 [0.221] 0.860*** [0.269] -0.212 [0.229] 0.967*** [0.307] -0.265 [0.214] 0.004 [0.004] 0.003 [0.005] 0.016*** [0.004] 0.005 [0.005] 0.005 [0.004] 0.003 [0.005] -168.627*** [10.663] -0.608*** [0.209] -0.261*** [0.010] -0.096*** [0.004] -0.334** [0.130] -0.009** [0.005] -0.000*** [0.000] -0.001 [0.001] 0.006 [0.004] -0.674*** [0.160] 5.266** [2.117]
Tornado event Hurricane event(t-1) Tornado event(t-1) Hurricane event(t-2) Tornado event(t-2) ln(property damages hurricane) ln(property damages tornado) ln(property damages hurricane)(t-1) ln(property damages tornado)(t-1) ln(property damages hurricane)(t-2) ln(property damages tornado)(t-2) Market share in state s Share in nationwide homeowners business Number of lines in state s Number of states insurer operates Share in group homeowners business Age Surplus in USD millions Reinsurance ratio Deviation from three-year avg. loss ratio Rate regulation Homeowners HHI of state s
The Geneva Papers on Risk and Insurance—Issues and Practice
214 Table 9 (continued) Explanatory variable
Exits coefficient [SE]
Business reduction [ 80 per cent Coefficient [SE]
Business reduction [ 60 per cent Coefficient [SE]
Per capita income in USD thousands
-0.058 [0.039] 31,028 YES YES -3168 9,429 0 0.598
-0.072*** [0.021] 42,782 YES YES -7761 8,339 0 0.349
-0.072*** [0.019] 44,144 YES YES -9092 8,341 0 0.314
Observations Insurer FE Year FE Log-likelihood v2 p value Pseudo-R2
Notes: Standard errors (SE) are robust to heteroscedasticity. *, ** and *** denote statistical significance at the 10, 5 and 1 per cent levels, respectively. Dollar values are inflation-adjusted and expressed 2012 USD.
Table 10 Wald test results for Pr(Exit) and Pr(Business Reduction) regression models Dependent variable: Pr(Exit) Year
t t-1 t-2
Event
Damages
v2-statistic
p value
v2-statistic
p value
0.23 6.91*** 6.92***
0.6323 0.0086 0.0085
0.01 0.11 0.16
0.9123 0.7361 0.6879
Dependent variable: Pr(Business Reduction [ 80 per cent) Year
Event 2
t t-1 t-2
Damages 2
v -statistic
p value
v -statistic
p value
1.26 7.34*** 12.70***
0.2609 0.0067 0.0004
0.51 3.83* 0.01
0.473 0.0503 0.9124
Dependent variable: Pr(Business Reduction [ 60 per cent) Year
t t-1 t-2
Event
Damages
V2-statistic
p value
v2-statistic
p value
1.53 9.62*** 11.14***
0.2163 0.0019 0.0008
0.07 2.88* 0.1
0.7945 0.0895 0.7539
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
215
the regression. Besides the results of the Arellano–Bond serial correlation test, we report the Hansen test results of over-identifying restrictions to ensure the validity of our model specification (see footnote 57).58
Results The results of our first set of regressions on the natural logarithm of premiums earned are summarised in Table 7. To ensure our results are not driven by outliers, we winsorised premiums earned at the 1/99 percentile. Because our main research question is concerned about different supply reactions depending on the type of natural disaster, we focus our analysis on a comparison of the major hazard event dummies and the hazard damage variables. Since we control for differences in the magnitude of the hazards by including the annual property damage exposure for each state and hazard, the remaining differences can be explained by other factors beyond damage size, such as the correlation of single damage events or predictability problems. In general, the effect of major disaster years on the business volume in an affected state seems to be negative, as almost all event coefficients in both estimations (linear fixed effects and system GMM) have negative signs. However, we observe large differences between the two hazards. Extreme hurricane years reduce business volume by a significant amount59 in the event year, whereas the effect of extreme tornado years is small and insignificant. This fairly large decrease can be partially explained by insurer defaults after extreme disaster seasons. In addition, insurers which suffered from disaster losses may decide to reduce their exposure in the affected area by not renewing insurance contracts or even cancelling policies in some cases. In addition, one should keep in mind that our analysis is based on our very restrictive event definition. Hence, we cover only very extreme disaster events, which may contribute to the huge observed effect size. To test whether differences between the effects of both hazards on insurance supply are statistically significant, we run Wald tests for each pair of hazard and damage coefficients. The results are summarised in Table 8. The difference between both event dummies is significant at the 1 per cent level (p value = 0.007) in the linear fixed-effects model and at the 10 per cent level (p value = 0.08) in the system GMM model in event year t. The effect size of extreme hurricane years is larger compared with extreme tornado years based on the lagged event dummies, but the effect and the differences between the effects are not significant. Interestingly, two years after an extreme tornado season, we detect a minor negative and significant effect on the premium income. However, the difference between both event dummies is also not significant [p value (fixed effects = 0.65/ p value (system GMM) = 0.81]. Based on the results of the fixed-effects regression, the size of hurricane damages in the event year seems to have a weak negative effect. The effect of hurricane damages is also different from that of tornado damages in the event year at the 10 per cent level (p value = 0.084). However, these results are not supported by the system GMM estimation and the effect size seems to be rather low. 58 59
Hansen (1982) Based on the linear fixed effects model results, business volume is reduced by approximately 22.4 per cent. The marginal effect is calculated by per cent-increase = 100 9 (exp(-0.254) - 1) = -22.43.
The Geneva Papers on Risk and Insurance—Issues and Practice
216
Hence, we obtain only very limited support for our first hypothesis on the size effect of a catastrophe. The differences between tornados and hurricanes after controlling for damage size provide preliminary support for our second hypothesis. Correlation of single losses seems to be a more severe challenge compared with the predictability of disaster events. Consequently, we cannot reject the null hypothesis of hypothesis H3 based on our analysis of premiums earned. Drawing conclusions about insurance supply decisions based on the size of premium income has an important disadvantage. Business volume is determined by the quantity of coverage provided and the price per coverage. Hence, the effects of price increases and quantity reductions might cancel one another out. In addition, both panel data estimation models with premiums earned as the dependent variable have some limitations. The linear fixed-effects panel data model is based on the assumption of strict exogeneity. Hence, we cannot include lagged dependent variables in this model, as estimates would be biased and inconsistent.60 Without including lagged dependent variables, we are not able to account for dynamic relations in our data. In addition, the strict exogeneity assumption may also be violated if some of our control variables are endogenous. Thus, we also conducted the system GMM estimation which overcomes both problems. However, the system GMM model requires no serial correlation of the error and validity of the instruments. The test results reported in Table 7 raise some concerns. We had to restrict the instrument set to lags 4 and longer because of the results of the Arrelano–Bond test for serial correlation which indicates serial correlation of the error terms up to order 3. In addition, the null hypotheses of the Hansen test for joint validity of the instruments is rejected which questions the validity of the model specification of the system GMM estimation. Due to potentially biasing price effects in the premiums earned regressions and the methodological concerns with both models, we conduct further analysis to validate our preliminary results on hazard-specific supply reactions. To analyse changes in insurers’ willingness to provide coverage in the aftermath of catastrophic events more thoroughly, we focus our second set of regressions on the probability of exiting a state and the probability of reducing business operations drastically (by more than 60/80 per cent). In particular, the probability of exiting a state is a pure quantity reduction and is not biased by price increases. In addition, prior empirical evidence shows that major disaster events are usually accompanied by price increases and by positive demand effects.61 Thus, a significant reduction of business volume is unlikely to be driven by a decrease in price per coverage or changes in insurance demand. Our results are summarised in Table 9. We obtain a positive effect on the probability of exiting a state in a hurricane event year and in the two subsequent years after a hurricane event. In contrast, the sign of the tornado event year dummies is always negative. The probability of exiting a state increases significantly one and two years after an extreme hurricane year. This delayed effect is not surprising because an insurer is usually not able to change its strategy immediately. An insurer can stop writing new policies and refrain from renewing expired policies, but an immediate cancellation of in-force business is not possible in general. In addition, shutting down entire business operations in a state is even more time consuming. 60 61
See, e.g. Nickell (1981). See, for example, Froot and O’Connell (1999); Browne and Hoyt (2000); Gallagher (2014).
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
217
We compute the odds ratios to obtain a more intuitive interpretation of the marginal effects of the non-linear model. The odds of exiting an affected state are multiplied by 2.80 one year after an extreme hurricane event and by 3.32 two years after an extreme hurricane event. In contrast, insurers’ exit probability does not increase after extreme tornado years because the event coefficients are negative and insignificant. We again conduct Wald tests to ensure that observed differences are statistically significant. A summary of the results of the comparison of all pairs of event and hazard coefficients is provided in Table 10. The differences between the effects of tornados and hurricanes on insurance supply are significant one and two years after the respective events (p value event dummies (t - 1) = 0.009; p value event dummies (t - 2) = 0.009). In addition, hurricane damages have a significant effect on the probability of exiting a state in the year after the event. However, according to the Wald test results, the difference to tornado damages is not significant. As mentioned in the data section, our definition of an exit is rigorous because insurers are required stop writing business completely. However, due to long-term contracts or regulatory requirements, insurers may be forced to keep some share of their business in subsequent years. Thus, the insurer might have stopped operating and offering new policies within a state but still have positive premium income. Therefore, we examine large-scale reductions of business volume in addition to analysing complete exits. Our results on the probability of reducing business volume by more than 60 or 80 per cent, which are shown in the second and third column of Table 9, confirm prior findings on the probability of exiting a state. Extreme tornado years have a negative and insignificant effect on the probability of reducing business volume. However, insurers tend to decrease business volume after extreme hurricane seasons. The effects one and two years after a hurricane event are significant at the 1 per cent level, and the differences between the hurricane and tornado dummies are also significant (p value event dummies (t - 1) = 0.007; p value event dummies (t - 2) = 0.000), as shown in Table 10. With respect to the damage size, we observe only a significant effect of the size of hurricane damages one year after an event for both ‘‘business reduction’’ models. This difference between the effects of both hazards is also statistically significant at the 10 per cent level. Our findings suggest that extreme hurricane years cause insurers’ exits and result in a significant downsizing of available coverage. In line with our prior findings, extreme tornado years seem to have no direct effects on insurers’ supply decisions. We find significant differences between both hazards one and two years after an extreme disaster season. These differences can be only partially explained by differences in the severity of the disaster years and might be caused by other hazard-specific effects. Similar to our findings on the effect of disaster events on premiums earned, we find limited support for hypothesis H1 which is concerned about the size effect of an event in our ‘‘exit’’ and ‘‘business reduction’’ regression models because at least the size of hurricane damages has a significant effect on our insurance supply measures. In addition, our results on the difference between the effect of unprecedented hurricane and tornado years provide support for hypothesis H2 and do not confirm hypothesis H3. Hence, we argue that the correlation between single losses seems to be a more challenging disaster characteristic for insurers in comparison to more ambiguous risks. Our other control variables are in line with our expectations. More diversified insurers (number of lines per state and number of states) have a lower probability of exiting a state. In addition, the importance of a state for the insurer and the competitiveness of an insurer in a state reduce the exit probability.
The Geneva Papers on Risk and Insurance—Issues and Practice
218
Robustness checks Florida and regulation Due to its high natural catastrophe exposure and specific regulatory environment, Florida is often suspected of driving the results of empirical analyses in catastrophe research.62 Therefore, we exclude all Florida insurer observations from our data set and rerun all regression models. The complete results are summarised in Tables A2 and A3 in the online Appendix. Because our results are largely the same after excluding the Florida observations from our sample, we can rule out the possibility that differences between the supply effects of hurricanes and tornados are driven by the specific characteristics of the state of Florida. Interestingly, after conducting Wald tests for the differences between hurricane and tornado coefficients, we find that differences are even more pronounced in some of our regression models when we exclude Florida, as shown in Table A4 in the online Appendix. For example, the negative effect of extreme hurricane seasons on premiums earned is also statistically significant two years after the event in both model specifications. This may indicate that regulatory interventions after catastrophic events in Florida may have limited insurers’ ability to alter their supply decisions (see, e.g. Klein and Kleindorfer10). Exits as strategic management decision As mentioned in the data section, we conduct a more refined analysis of insurers’ exits. We thereby examine the hazard-specific effect on voluntary exits, i.e. insurers’ decision to cease writing business in the respective state without shutting down their homeowners business nationwide.63 We argue that those decisions are in general not driven by external effects but represent strategic management decisions that contrast with mergers and acquisitions or insolvencies. The occurrence of a catastrophe causes a revaluation of the company’s risk exposure by management, and the voluntary exit is the outcome of this process. Table A5 shows the results of our general ‘‘exit’’ regression compared with the ‘‘voluntary exits’’ regression. The signs of all coefficients do not change. Hurricanes increase the probability of exiting a state, in general, and extreme tornado years tend to decrease the exit probability. Despite the slightly lower significance of our findings, we still find clear differences between both hazards in our analysis, and our main conclusions do not change after applying this more refined definition of an exit. Supply reactions in states with joint exposure States’ risk exposure to both hazards varies strongly, as indicated by Figure 1, which shows the average annual loss exposure of each state for tornados and hurricanes. First, the size of the exposure per hazard exhibits large variation between states. Second, some states are exposed to both risks, whereas other states only face one of these risks. Restricting the analysis to those states with exposure to both hurricanes and tornados allows a cleaner comparison of insurers’ supply reactions within this subsample. We limit the analysis to 62 63
Grace and Klein (2006); Medders et al. (2014). Our ‘‘voluntary exit’’ indicator equals one if an insurer’s premium income is zero in the following two years in one state, but the insurer still has positive premium income in other states.
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
219 64
states that experienced both hurricane and tornado damages between 1992 and 2012 according to the SheldusTM data base. In line with our prior analysis, we examine the effect on premiums earned, the probability of exiting a state, and the probability of reducing business by more than 60/80 per cent. The results are shown in the online Appendix in Tables A6 and A7. The findings for the subsample do not differ from the results in the main part of the paper. Extreme hurricane seasons still cause a significant reduction of premiums earned in the affected states in both (linear fixed effects and system GMM) models, and the probabilities of exiting a state or of reducing business volume by more than 80/60 per cent increase significantly in the two subsequent years after a hurricane event. Our system GMM estimation also shows a smaller negative and statistically significant effect of tornados in the event year and two years after the event. We conduct Wald tests to ensure that the observed differences between hurricanes and tornados are statistically significant. As shown in Table A8, the Wald test results are predominantly in line with our prior results and indicate that the differences are statistically significant in the event for the linear fixed effects and system GMM estimation models and in the two subsequent years after an event for our different logit estimations. Exclude small insurers with limited risk taking capacity As our data set covers the entire U.S. homeowners insurance market, and the distribution of premiums earned is heavily skewed (skewness = 18.70),65 many small insurers are included in our analysis. Since they may be more heavily affected by natural disasters than their larger peers because of their limited risk-taking capacity, we exclude insurers with nationwide premium income66 in the homeowners line below USD 10 million. As a result, approximately 25 per cent of our observations are excluded from the sample. We examine the effect of both hazards on supply decisions of insurers in this restricted sample by conducting our different panel data regressions. For most model specifications, our results of the restricted sample are similar to the full-sample analysis. The system GMM estimation does not confirm the hurricane-specific effect on insurance supply in the event year. In general, the statistical significance of our findings is decreased but the coefficients signs do not change and we still find statistically significant differences between the effects of both hazards. We conclude that the effects of large natural disasters on insurance supply are less pronounced for larger insurance companies. In general, larger insurers tend to have more diversified operations and better access to the capital market which may explain the results. Effect on business volume at the state level We test the robustness of our findings of hazard-specific disaster effects by examining the effect of extreme hurricane and tornado years on the aggregated insurance supply of each state. We rerun both premiums earned panel data regression models with all U.S. states as 64
65
66
The following states belong to our subsample with joint hazard exposure: AL, AR, CT, DE, FL, GA, LA, MA, MD, ME, MO, MS, NC, NH, NJ, NY, OK, PA, RI, SC, TN, TX, and VA. The U.S. homeowners insurance market is rather concentrated. The five largest insurance companies account for almost 50 per cent of the market (Insurance Information Institute, 2016). We focus on nationwide premium income to determine the threshold, as an insurer who operates in multiple states can transfer funds to cover losses in other states. Hence, premium income in one state can be misleading for determining an insurer’s risk-taking capacity.
The Geneva Papers on Risk and Insurance—Issues and Practice
220
the observation unit. Results are summarised in Table A12 in the online Appendix. Hurricane events have still a more negative effect on aggregated insurance supply than tornados in the event year. However, the effect is not statistically significant in both model specifications. Interestingly, aggregated premium income seems to increase two years after an extreme hurricane year. We find a statistically significant positive effect of extreme hurricanes which is also significantly different from the effect of extreme tornado years according to the Wald test results of the linear fixed effects panel data regression which are shown in Table A13. This may indicate that the reduction of insurance supply in the event year is at least partially compensated by price increases and additional coverage supplied by new entrants or insurers that remained in the market in the medium term. Hence, the long-term effect of extreme disaster events may also vary between different insurers. Importantly, all model specification tests support the validity of the system GMM model in contrast to the disaggregated supply estimations. We used lags 2 and more as instruments as the auto correlation tests only indicated serial correlation of the error term up to order 1. The effect of a hurricane’s damage size on insurance supply is negative and significant in the event year and in the subsequent year. The effect of the size of property damages does also differ between both hazards according to our linear fixed effects estimation results. Analysis of two disaster seasons with similar damage size Based on our analysis, we argue that disaster-specific characteristics cause the observed differences between both disaster types in their effect on insurance supply decisions. One limitations of our approach could be that we are not able to control completely for the damage size of disaster years by including the annual hazard-specific property damages. Thus, differences between the event dummies of both hazards could be partially caused by differences in the damage size of tornado and hurricane seasons. To limit difference in damages size in advance to the largest extent, we choose two disaster years with similar property damages in a state, the hurricane season in Texas in 2001 and the tornado season in Alabama in 2011, and compare insurers’ supply reactions after both events. We examine the effect of both events on premiums earned67 and our results are shown in Table A14. As the tornado event occurred in 2011 and our data set covers the period until 2012, we could only compare the lagged effect one year after both disaster seasons. Although, the Texas hurricane season was similar to the Alabama tornado season in terms of damage size, both lead to different insurer supply reactions. In the event year, insurance supply in Texas is significantly reduced, whereas we do not observe a similar effect in Alabama after the 2011 tornado season. We are not able to identify a statistically significant difference between both events in our linear fixed-effects regression, but the results of the system GMM estimation support our previous findings. Interestingly, premiums earned seem to increase after the 2011 tornado season. Although, both extreme disaster years do not differ strongly in terms of damage size, we find differences in their effect on insurers’ supply decisions. Under the caveat that this estimation is based on only two disaster events and, hence, only few insurers were affected by the events, we may conclude that hurricanes seem to have a more significant negative effect on insurance supply, which could be caused by high correlation of single damage events. 67
We were not able to analyse the effect on the probability of exiting a state and the probability of reducing business by more than 60/80 per cent as we were not able to calculate multiple of the 2011 tornado season due to the forward-looking definition of our dependent variables.
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
221
Discussion and policy implications In this paper, we were able to show that the type of a natural disaster matters for insurers’ supply decisions after a major catastrophe year and damage size is not the only determinant. We investigated how an insurer’s decision to write business in a state is affected after large and unprecedented hazard seasons. In general, major hurricane seasons have a more pronounced negative effect on insurers’ premium income compared with tornado seasons. Even after controlling for the size of annual property damages and several insurer- and state-specific characteristics, we observe significant differences in the probability of exiting a state or the probability of reducing business volume. In the following two years after an extreme hurricane season, the odds of exiting the affected state are multiplied by 2.80 and 3.32 after extreme hurricane seasons, respectively. The exit probability is not similarly affected after extreme tornado years. Based on these findings, we conclude that it is easier for insurers to deal with extreme tornado years compared with extreme hurricane seasons. As indicated in Figure 1, hurricanes affect a fairly small area in the U.S. but cause very large damages within this area. Therefore, insurers’ benefits from regional diversification are very limited. Because we account for the size of the hazard damages, the remaining difference in supply reactions after hurricanes and tornados may be caused by differences in the degree of correlation between single-loss events. Thus, we recommend that insurers and regulators focus on encouraging diversification. The results for our control variables show that diversification can reduce the likelihood of exiting a state. Both regional diversification (the number of states in which an insurer operates) and diversification by line of insurance (number of lines in a state) significantly reduce the probability of exiting a state. Currently, writing business in different states is hindered due to state-specific regulations. Insurers might refrain from achieving diversification by operating in multiple states because adapting products and processes to the local regulatory framework might be too costly. According to Ibragimov et al.68 diversification benefits for heavy-tailed risks might exhibit a U-shaped relationship. Thus, an increase in the number of risks might even have a negative effect on the individual insurer’s performance up to certain level of diversification. At the industry level, diversification benefits remain positive. The authors recommend a central agency to achieve coordinated diversification equilibria and unleash diversification benefits. Considering hazard-specific effects on insurance supply is crucial for designing a nationwide disaster pool. Another obvious means to increase diversification is incentivising the purchase of reinsurance coverage, lowering reinsurance requirements, or implementing tax benefits for reinsurance. In particular, international reinsurers could enhance the worldwide diversification of catastrophic risks. Another approach to reducing supply distortions due to poor diversification is implementing all-hazard insurance policies that cover floods and earthquakes in addition to hurricanes and tornados, as suggested by Kunreuther and Michel-Kerjan69 in a different context. Insurers could benefit from bundling more- and less-frequent risks all over the U.S. However, for the design of multi-hazard policies or risk-sharing pools, a better 68 69
Ibragimov et al. (2009). Kunreuther and Michel-Kerjan (2009).
The Geneva Papers on Risk and Insurance—Issues and Practice
222
understanding of disaster characteristics and how they influence insurers’ supply reactions is important. The predictability of natural catastrophes is also mentioned as a challenge when insuring catastrophic risks. We highlight that estimating tornado risk properly is very demanding; however, our results indicate that insurers do not struggle significantly with insuring tornado risks. We suppose that large investments of the insurance industry in catastrophe modelling might have helped to mitigate predictability issues. In the light of increasing losses due to convective storms in the last decade,8 improving catastrophe modelling might be a worthwhile venture but should not have the highest priority at the moment. We conclude that insurance contracts are an appropriate tool for providing coverage against ambiguous catastrophic losses. Risk-sharing pools, which are particularly useful when probabilities of loss are unknown,70 might be a valuable complement to insurance contracts in some cases. However, the insurance industry seems to be comfortable with addressing tornado risk, which is one of the least predictable hazards. The size of an event is often named as a major challenge for insurability. Not surprisingly, we also find that larger losses cause more severe supply distortions. However, damage size is not the only determinant of insurers’ supply reactions. The insurance industry is able to cover large but uncorrelated annual damages more easily. However, if single losses are highly correlated and concentrated in a specific area, we observe more extreme supply reactions. Enhancing the risk-taking capacity of insurers can be beneficial for maintaining sufficient insurance coverage in disaster-prone areas; however, our analysis indicates that other measures may have a stronger influence. Considering the results of our empirical analysis, we find that insurers’ supply reactions differ significantly with respect to the type of hazard. We conclude that results based on one-hazard studies should not be generalised to other hazards because insurers are very likely to react differently after catastrophic events with the same damage size but caused by a different trigger. In addition, prior studies that examined the influence of different natural hazards on insurance markets within one analysis without differentiating between the types of hazards should be extended. Differentiation by hazard type can provide additional insights into insurance market reactions after catastrophic events. Generalising results or pooling hazards within one analysis might result in misleading conclusions. Importantly, our findings on hazard-specific effects are limited to the effect of two significant natural hazards on the U.S. insurance market. In addition, our findings rely on our specific definition of unprecedented natural catastrophes. We opted for this definition for various reasons as outlined in the ‘‘Data and variables’’ section. Nevertheless, our definition is arbitrary like all other definitions introduced in the literature review. Future studies should extend the analysis to a broader scope of hazards and to other insurance markets to validate the existence of hazard-specific supply effects.
Acknowledgements The authors thank the participants at the Annual Meeting of the German Insurance Science Association, 2014; the Annual Meeting of the American Risk and Insurance Association, 2014; the CEAR/ MRIC Behavioral Insurance Workshop, 2014; the 16th Joint Seminar of the European Association of 70
Skogh (1999); Skogh and Wu (2005).
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
223
Law and Economics (EALE) and The Geneva Association, 2015; and in particular, James Carson, Randy Dumm, Michael Hanselmann, Robert Hoyt, Johannes Jaspersen, Stefan Neuß and Martin Spindler for valuable comments. Any remaining errors are our own.
References Aiuppa, T.A., Carney, R.J. and Krueger, T. M. (1993) ‘An examination of insurance stock prices following the 1989 Loma Prieta Earthquake’, Journal of Insurance Issues 16(1): 1–14. Angbazo, L.A. and Narayanan, R. (1996) ‘Catastrophic shocks in the property-liability insurance industry: Evidence on regulatory and contagion effects’, The Journal of Risk and Insurance 63(4): 619–637. Arellano, M. and Bond, S. (1991) ‘Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations’, The Review of Economic Studies 58(2): 277–297. Arellano, M. and Bover, O. (1995) ‘Another look at the instrumental variable estimation of error-components models’, Journal of Econometrics 68(1): 29–51. Aseervatham, V., Born, P. and Richter, A. (2013) Demand reactions in the aftermath of catastrophes and the need for behavioral approaches, Munich Risk and Insurance Center Working Paper No. 13. Blundell, R. and Bond, S. (1998) ‘Initial conditions and moment restrictions in dynamic panel data models’, Journal of Econometrics 87(1): 115–143. Born, P. and Klimaszewski-Blettner, B. (2009) Catastrophes and performance in property insurance: A comparison of personal and commercial lines, Independent Policy Report, Oakland, CA: Independent Institute. Born, P. and Klimaszewski-Blettner, B. (2013) ‘Should I stay or should I go? The impact of natural disasters and regulation on US property insurers’ supply decisions’, The Journal of Risk and Insurance 80(1): 1–36. Born, P. and Viscusi, W.K. (2006) ‘The catastrophic effects of natural disasters on insurance markets’, Journal of Risk and Uncertainty 33(1–2): 55–72. Browne, M.J. and Hoyt, R.E. (2000) ‘The demand for flood insurance: Empirical evidence’, The Journal of Risk and Insurance 20(3): 291–306. Cabantous, L. (2007) ‘Ambiguity aversion in the field of insurance: Insurers’ attitude to imprecise and conflicting probability estimates’, Theory and Decision 62(3): 219–240. Cabantous, L., Hilton, D., Kunreuther, H. and Michel-Kerjan, E. (2011) ‘Is imprecise knowledge better than conflicting expertise? Evidence from insurers’ decisions in the United States’, Journal of Risk and Uncertainty 42(3): 211–232. Cagle, J.B. and Harrington, S. (1995) ‘Insurance supply with capacity constraints and endogenous insolvency risk’, Journal of Risk and Uncertainty 11(3): 219–232. Chen, H. and Sun, T. (2012) Model uncertainty, ambiguity aversion and implications for catastrophe insurance market, working paper, Temple University. Cole, C.R. and McCullough, K.A. (2006) ‘A reexamination of the corporate demand for reinsurance’, The Journal of Risk and Insurance 73(1): 169–192. Cummins, D. and Lewis, C.M. (2003) ‘Catastrophic events, parameter uncertainty and the breakdown of implicit long-term contracting: The case of terrorism insurance’, Journal of Risk and Uncertainty 26(2–3): 153–178. Cummins, D. and Sommer, D.W. (1996) ‘Capital and risk in property-liability insurance markets’, Journal of Banking & Finance 20(6): 1069–1092. Deryugina, T. (2011) The dynamic effects of hurricanes in the US: The role of non-disaster transfer payments, working paper, Massachusetts Institute of Technology Center for Environmental Policy Research. Froot, K. A. and O’Connell, P. G. (1999) ‘The pricing of US catastrophe reinsurance’, in The Financing of Catastrophe Risk. Chicago: University of Chicago Press, pp. 195–232. Gallagher, J. (2014) ‘Learning about an infrequent event: Evidence from flood insurance take-up in the United States’, American Economic Journal: Applied Economics 6(3): 206–233. Grace, M. and Klein, R. W. (2006) After the storms: Property insurance markets in Florida, working paper, Georgia State University. Grace, M. and Klein, R.W. (2007) ‘Hurricane risk and property insurance markets’, working paper, Georgia State University. Grace, M. and Klein, R.W. (2009) ‘The perfect storm: Hurricanes, insurance, and regulation’, Risk Management and Insurance Review 12(1): 81–124.
The Geneva Papers on Risk and Insurance—Issues and Practice
224 Grace, M., Klein, R.W. and Kleindorfer, P.R. (2004) ‘Homeowners insurance with bundled catastrophe coverage’, The Journal of Risk and Insurance 71(3): 351–379. Grace, M., Klein, R.W. and Liu, Z. (2005) ‘Increased hurricane risk and insurance market responses’, Journal of Insurance Regulation 24(2): 3–32. Gron, A. (1990) Property-Casualty Insurance Cycles, Capacity Constraints, and Empirical Results, Cambridge, MA: Massachusetts Institute of Technology. Gron, A. (1994) ‘Capacity constraints and cycles in property-casualty insurance markets’, The RAND Journal of Economics 25(1): 110–127. Hagendorff, B., Hagendorff, J. and Keasey, K. (2015) ‘The impact of mega-catastrophes on insurers: An exposurebased analysis of the U.S. homeowners’ insurance market’, Risk Analysis 35(1): 157–173. Hansen, L.P. (1982) ‘Large sample properties of generalized method of moments estimators’, Econometrica: Journal of the Econometric Society 50(4): 1029–1054. Harrington, S.E. and Niehaus, G. (2001) ‘Government insurance, tax policy, and the affordability and availability of catastrophe insurance’, Journal of Insurance Regulation 19(4): 591–612. Henson, B. (2013) ‘Long-range tornado prediction: Is it feasible?’, from http://www2.ucar.edu/atmosnews/ opinion/9996/long-range-tornado-prediction-it-feasible, retrieved 16 March 2015. Ibragimov, R., Jaffee, D. and Walden, J. (2009) ‘Non-diversification traps in catastrophe insurance markets’, Review of Financial Studies 22(3): 959–993. Insurance Information Institute (2016) ‘Homeowners and Renters Insurance’, from http://www.iii.org/factstatistic/homeowners-and-renters-insurance, retrieved 10 April 2016. Kleffner, A. E. and Doherty, N.A. (1996) ‘Costly risk bearing and the supply of catastrophic insurance’, The Journal of Risk and Insurance 63(4): 657–671. Klein, R. (2008) Catastrophe risk and the regulation of property insurance, working paper, Georgia State University. Klein, R. and Kleindorfer, P. (1999) The supply of catastrophe insurance under regulatory constraints, Wharton Center for Financial Institutions Working Paper No. 99-25. Klein, R.W. (2013) ‘Insurance market regulation: Catastrophe risk, competition, and systemic risk’, in G. Dionne (ed) Handbook of Insurance. New York: Springer, pp. 909–939. Kunreuther, H., Meszaros, J., Hogarth, R.M. and Spranca, M. (1995) ‘Ambiguity and underwriter decision processes’, Journal of Economic Behavior & Organization 26(3): 337–352. Kunreuther, H. and Michel-Kerjan, E.O. (2009) At War with the Weather: Managing Large-Scale Risks in a New Era of Catastrophes. Cambridge, MA: MIT Press. Lamb, R.P. (1995) ‘An exposure-based analysis of property-liability insurer stock values around Hurricane Andrew’, The Journal of Risk and Insurance 62(1): 111–123. Medders, L.A., Nyce, C.M. and Karl, J.B. (2014) ‘Market implications of public policy interventions: The case of Florida’s property insurance market’, Risk Management and Insurance Review 17(2): 183–214. National Hurricane Center (2014) ‘Hurricane season dates’, from http://www.nhc.noaa.gov/, retrieved 16 March 2015 National Severe Storms Laboratory (2014) Severe weather 101’, from https://www.nssl.noaa.gov/education/ svrwx101/tornadoes/faq/, retrieved 16 March 2015. Nickell, S. (1981) ‘Biases in dynamic models with fixed effects’, Econometrica: Journal of the Econometric Society 6: 1417–1426. Ragin, M.A. and Halek, M. (2015) ‘Market expectations following catastrophes: An examination of insurance broker returns’, The Journal of Risk and Insurance. doi:10.1111/jori.12069. Raykov, R.S. (2015) ‘Catastrophe insurance equilibrium with correlated claims’, Theory and Decision 78(1): 89–115. Shelor, R.M., Anderson, D.C. and Cross, M.L. (1992) ‘Gaining from loss: Property-liability insurer stock values in the aftermath of the 1989 California earthquake’, The Journal of Risk and Insurance 59(3): 476–488. Skogh, G. (1999) ‘Risk-sharing institutions for unpredictable losses’, Journal of Institutional and Theoretical Economics 155(3): 505–515. Skogh, G. and Wu, H. (2005) ‘The diversification theorem restated: Risk-pooling without assignment of probabilities’, Journal of Risk and Uncertainty 31(1): 35–51. Swiss Re. (2014) ‘Natural catastrophes and man-made disasters in 2013’, SIGMA 1/2014. Swiss Re. (2015) ‘Natural catastrophes and man-made disasters in 2014: Convective and winter storms generate most losses’, SIGMA 2/2015: 14–20.
Vijay Aseervatham et al Hazard-Specific Supply Reactions in the Aftermath of Natural Disasters
225 Takao, A., Yoshizawa, T., Hsu, S. and Yamasaki, T. (2013) ‘The effect of the Great East Japan earthquake on the stock prices of non-life insurance companies’, The Geneva Papers on Risk and Insurance—Issues and Practice 38(3): 449–468. Thomann, C. and von der Schulenburg, J.-M. (2006) Supply and demand for terrorism insurance: Lessons from Germany, Hannover Economic Papers dp-340, Universita¨t Hannover. Windmeijer, F. (2005) ‘A finite sample correction for the variance of linear efficient two-step GMM estimators’, Journal of Econometrics 126(1): 25–51. Winter, R.A. (1994) ‘The dynamics of competitive insurance markets’, Journal of Financial Intermediation 3(4): 379–415. Yamasaki, T. (2015) ‘Do typhoons cause turbulence in property-liability insurers’ stock prices?’, The Geneva Papers on Risk and Insurance—Issues and Practice 41(3): 432–454. Zhang, L. and Nielson, N. (2015) ‘Solvency analysis and prediction in property-casualty insurance: Incorporating economic and market predictors’, The Journal of Risk and Insurance 82(1): 97–124.
About the Authors Vijay Aseervatham received his Ph.D. in Business Administration from LMU Munich. He is a member of the Munich Risk and Insurance Center at LMU Munich. His research interests include the management of catastrophic risks, insurance regulation and behavioural economics. Patricia Born received her Ph.D. in Economics from Duke University and is currently the Midyette Eminent Scholar of Insurance in the Department of Risk Management/Insurance, Real Estate and Legal Studies at Florida State University. She is a research associate in the Florida Catastrophic Storm Risk Management Center. Her research interests include insurance market structure and performance, professional liability, health insurance and the management of catastrophic risks. Dominik Lohmaier received his Ph.D. in Business Administration from LMU Munich. He is a member of the Munich Risk and Insurance Center at LMU Munich. His research interests include experiential learning and risk management, insurance accounting and the management of catastrophic risks. Andreas Richter received his Ph.D. and his Habilitation from the University of Hamburg and currently holds the Chair in Risk and Insurance at LMU Munich. He is also the Director of the Munich Risk and Insurance Center at LMU Munich. His research interests include the microeconomic theory of insurance, the management of catastrophic risks, behavioural insurance and law and economics.