Professional Documents
Culture Documents
7625 PDF
7625 PDF
Abstract
Adverse selection and moral hazard are commonly expected to cause market failures in natural
disaster insurance markets. However, such problems may not occur if individuals mainly buy
insurance based on risk preferences. Advantageous selection can occur if individuals with
insurance are highly risk averse and seek to reduce risk. This is the first empirical study of
adverse selection and moral hazard effects in natural disaster insurance markets. Statistical
analyses are based on survey data of individual purchases of disaster insurance and risk
mitigation activities in Germany and the United States. Consistent results are obtained in both
Keywords: Adverse selection, Heckman sample selection model, Propensity Score Matching,
1
1. Introduction
Over the last decades economic damages from natural disasters have been increasing and this
trend is likely to continue (IPCC, 2012; Munich Re, 2013). Insurance could play an important
role in managing natural disaster risks and promoting recovery from disasters, because it reduces
financial risks for individuals by spreading risks over many policyholders, helping people to “get
back on their feet” after a disaster occurs (Botzen, 2013). Moreover, insurance can provide
incentives for risk reduction by charging premiums that act as a price signal of risk, or by
providing premium discounts to policyholders who protect their property against disaster damage
(Kunreuther, 1996). On the other hand, insurance coverage may result in a moral hazard effect in
that individuals take fewer measures to limit risk if they expect that insurers will compensate
their damage irrespective of their mitigation efforts (Arnott and Stiglitz, 1988). This poses
problems if the behavior leading to moral hazard cannot be observed by the insurer, meaning that
actual risks are not reflected in the insurance premium (Chiappori and Salanie, 2000). Moreover,
adverse selection may obstruct the adequate functioning of natural disaster insurance marketsi if
mainly individuals who face a high risk demand insurance, while insurers do not adequately
factor such risks into higher prices because of information asymmetries between the insurer and
It is, however, not evident that adverse selection and moral hazard problems arise in
insurance markets. Adverse selection may not occur if individuals are characterized by “bounded
rationality” and misperceive the risk that they face (Kunreuther and Pauly, 2004). In other words,
it is not necessarily the case that individuals facing a high objective risk have a high risk
perception (Botzen et al., 2009) or a high demand for natural disaster insurance coverage (Botzen
and van den Bergh 2012a,b). In general, a lot of evidence shows that individuals have difficulties
with assessing low-probability/high-impact risks (Kunreuther et al., 2001) which may translate
2
into poor decision making with respect to natural disaster insurance purchases (Kunreuther et al.,
2013). Moreover, adverse selection and moral hazard may not be an issue if insurance purchase
decisions are mostly driven by risk aversion, and if highly risk averse people who purchase
insurance also take other precautionary measures which limit risk (Finkelstein and McGarry,
2006; Cohen and Einav, 2007). Such an advantageous selection effect is theoretically supported
by de Meza and Webb (2001). This theory may hold in practice since, for instance, Thieken et al.
(2006) showed that insured households during a 2002 flood in Germany had better risk
question whether moral hazard and adverse selection effects dominate advantageous selection
effects based on preferences, like risk aversion, and potential risk-reducing incentives provided
by insurance.
Several empirical studies conducted on the health insurance markets show that adverse
selection is present (Sloan and Norton, 1997; Finkelstein et al. 2005; Finkelstein and McGarry,
2006; Courbage and Roudaut, 2008), although there are studies that argue the opposite. One such
study is Cardon and Hendel (2001) who provide evidence that information asymmetries are not
significantly present, by showing that observable traits are important drivers for health
expenditure. However, Finkelstein and McGarry (2006) arrive at an opposite finding to that of a
moral hazard effect since individuals with health insurance in the U.S. take more measures to
reduce health risks than uninsured individuals, which may be explained by risk aversion. Einav
et al. (2013) find that greater U.S. health insurance coverage is driven by increasing expected
health risks, risk aversion and moral hazard. Their results show that altering the level of risk
aversion only has small effects on the size of the chosen deductible, while moral hazard and
expected health risks have roughly equal driving effects on the size of the deductible. Moreover,
3
they find that moral hazard is driven by the expected incremental increase of health expenditure,
which implies that individuals who take measures to reduce their marginal health care costs can
display advantageous selection. Likewise, in their investigation of life, acute health, annuities,
long-term care, and Medicare supplemental insurance markets, Cutler et al. (2008) find that those
individuals who engage in less risk reducing behavior (i.e., moral hazard) are less likely to have
each of these insurance types. To the best of our knowledge, a systematic empirical analysis of
the presence of moral hazard and adverse selection in natural disaster insurance markets is
lacking.
This paper aims to fill this gap by presenting an analysis of how flood insurance
purchases relate with experiences of flood damage and flood risk mitigation behavior in
Germany and in the U. S. In particular, for Germany, this paper examines the individual
behavioral component of adverse selection, namely whether people who take out flood insurance
face a higher flood risk. Moreover, it is estimated whether after correcting for adverse selection,
individuals who have flood insurance experience more flood damage than uninsured individuals,
which can be related to moral hazard when insured individuals take fewer flood damage
mitigation measures. The absence or presence of moral hazard in Germany may be related to risk
preferences of individuals, or the deductible on the flood insurance policy for which we cannot
control directly. However, we expect the influence of the deductible to be small. This is
confirmed by our complementary analysis for the U.S. which shows that the deductibleiii has a
negligible impact on policyholders risk mitigation activities. In particular, for the U.S. it is
examined whether people who have taken out separate homeowners insurance (which covers
wind damage only) and flood insurance policies are more, or less likely to take other measures
4
that limit hurricane damage. This analysis provides insights into moral hazard, and controls for
Understanding the potential moral hazard and adverse selection effects in the market for
natural disaster insurance has important public policy relevance in these two countries. In
Germany it has been argued that adverse selection is one of the reasons for the observed low (5-
10%) market penetration of natural hazard insurance, which has resulted in calls for introducing
mandatory disaster insurance (Schwarze and Wagner, 2007; Seifert et al., 2013). In the U.S.,
where flood insurance is primarily a publically underwritten insurance vehicle through the
National Flood Insurance Program (NFIP), there have been recent calls for reform including
more private market involvement (Michel-Kerjan and Kunreuther, 2011). Adverse selection
would be a deterrent in this regard. Moreover, the movement toward risk-based premiums as a
part of the recent flood insurance reform actsiv is aimed at providing incentives for mitigation for
which it is relevant to know how far insurance acts as a mitigation disincentive (moral hazard).
In the U.S. wind insurance market significant wind and hurricane deductibles are a part of a
homeowner’s insurance contract to help avoid potential moral hazard. However, such
deductibles may substantially lower the attractiveness of the insurance for consumers, and
whether these can be applied in recent major events, such has Hurricane Irene and Sandy, has
been a contestable legal issue (Pomerantz and Suglia, 2013). It is, therefore, of interest to
examine whether moral hazard is a major issue in the U.S. natural disaster insurance market, and
whether high deductibles are effective in stimulating policyholders to mitigate risks, as will be
studied here.
5
The remainder of this article is structured as follows. Section 2 describes the methods and
data used. Section 3 presents the results about adverse selection and moral hazard in Germany
Statistical methods
Two methods are applied to examine adverse selection and moral hazard effects in flood
insurance markets in Germany. First, an estimation of how average flood damage differs
between individuals with, and without, flood insurance provides an indication of both adverse
selection and moral hazard effects on flood damage. For example, experienced flood damage
may be higher for individuals with flood insurance because they face a higher flood risk (adverse
selection) or because such individuals take less damage mitigation measures (moral hazard). In
contrast, flood damage of insured individuals could be lower if they have taken more damage
mitigation measures, for example, because individuals with flood insurance are very risk averse.
Alternatively they may have received incentives from their insurer to reduce flood risks,
although the latter is uncommon in Germany (Thieken et al., 2006). Second, this study aims to
estimate the independent effect of moral hazard on experienced flood damage by correcting for
the effects of adverse selection on flood damage by insured individuals using Propensity Score
Matching (PSM).
The reason for using PSM is that taking only a simple comparison of average damage
between the insured group and non-insured group of survey respondents for estimating moral
hazard would likely suffer from selection bias (SB). This is formalized in Equation 1 which
shows that the difference between the damage outcome of people with flood insurance (the
6
treatment group) 𝑦1 and people without insurance (non-treatment group) 𝑦0 is a combination of
the average treatment effect on the treated (ATT) and SB. T is a binary variable for participation
in the treatment group. The ATT is the independent influence of having an insurance policy on
flood damage, which can be influenced by a moral hazard effect. SB arises as a result of the
influence of factors that jointly affect flood damage outcomes and having purchased flood
insurance. Selection bias can be viewed as an adverse selection effect because it represents the
households. The effects of adverse selection are due to the individuals with high(er) risk traits
having a higher tendency to buy insurance, which will mean higher expected damage to be
suffered. The higher expected damage will remain even if the individuals with higher risk traits
PSM is able to remove the SB from a comparison of average damage, and estimate the ATT,
which provides an indication of the presence or absence of moral hazard. PSM is most
commonly used in cases where non-random entry into the control and treatment groups means
that traits that affect both outcomes and treatment participation (confounders) can introduce bias
into evaluation attempts. Important confounders in this application are the characteristics of the
flood hazard faced by individuals and characteristics of their assets exposed to floods which have
made them select into buying flood insurance (adverse selection). Both characteristics include
factors that significantly influence the damage that individuals suffer when a flood occurs. Our
7
approach is in line with other studies that use matching methods to investigate moral hazard in
observational data, such as survey data that will be used here. Only 𝑦1 or 𝑦0 can be observed for
each agent, which implies that we do not observe the level of flood damage insured households
would have suffered in case they would not have purchased flood insurance coverage. Members
of the non-treatment group can be used as the required counterfactual observation for treatment
group members, if SB shown in Equation 1 can be removed. Removing this bias requires the
following conditions to hold, where represents independence, and 𝑝(𝑋) is the estimated
propensity score (PS) as a function of the confounders X (Rosenbaum and Rubin, 1983; Hudson
et al., 2014):
Condition 3: Overlap – The PS distributions for the control and treatment groups share a
purchases in the U.S. which uses binary discrete choice (probit) models to examine whether
insured individuals are more or less likely to implement any measures that reduced hurricane
8
losses. The initial approach employed with the U.S. data is similar to the analyses of moral
hazard in health insurance conducted by Cutler et al. (2008). Specifically, the likelihood of an
individual having either a homeowners or flood insurance policy is estimated as a function of the
activities that were undertaken prior to the arrival of an impending hurricane. Undertaking these
behavioral measures are hurricane risk reducing in that they could reduce damage to one’s
property (putting up storm shutters, taking in furniture, permanent modifications to one’s home,
etc.) or oneself (purchase of food and water supplies, made reservations in case evacuation is
needed, plan to evacuate, etc.). We additionally control for an individual’s risk perception of the
storm through a variable that indicates how safe one feels in staying in their home throughout the
hurricane event. Moreover, additional analyses examine the influence of the insurance deductible
on mitigation activities.
Data collection
The German data are obtained from surveys carried out in the Elbe and Danube river catchment
areas in response to flood events occurring in 2002, 2005 and 2006. The sample population was
collated by using official data to collect all of the streets that suffered from a flood. The sample
population was refined into the experimental sample by drawing a random sample of households
from the identified addresses. The survey was conducted as a 30 minute telephone interview
directed to the head of the household. The surveys provide approximately 2,000 respondents in
total (Kreibich et al., 2005; 2011). Of the usable 640 observation for PSM about 42% have flood
insurance. The surveys were intended to explain both damage outcomes from the flood and if a
respondent had undertaken precautionary flood risk mitigation measures; the overlapping
9
variables are the confounders required for PSM to be successful. These variables which are
included in estimating the propensity score are related to the flood hazard, exposure and
vulnerability, which are of influence on the actual flood risk faced by the respondent and may
indicate the presence of adverse selection. The variables are described in Appendix A. Several
different methods of matching have been employed in order to assess the appropriateness of the
supposed confounders. More detailed information on the applied methodology can be found in
Hudson et al. (2014) who apply PSM to evaluate damage savings from flood preparedness
The U.S. data are obtained from field surveys that measured the evolution of coastal
residents’ risk perceptions and preparation plans as three hurricanes ─ Irene (2011), Isaac (2012),
and Sandy (2012) ─ approached the U.S. during the 2011 and 2012 hurricane seasons. In these
studies, perceptions and preparation decisions were notably measured in real time as they were
being made by residents threatened by the storms. The surveys for these three storms provide
1,698 respondents in totalvi and include information on whether respondents had a homeowner’s
insurance policy that would pay for damages to one’s home resulting from the storm, if they had
a separate flood insurance policy, and whether they knew the amount of their insurance policy
deductible or would have to look it up. In the U.S., homeowner’s insurance policies cover only
wind-related hurricane damages, not any damages due to flooding from the storm. While 85
percent of total respondents indicated having a homeowner’s insurance policy, only 32 percent
answered having a separate flood insurance policy. Answers to these two questions serve as our
indicator variables for whether a respondent has homeowner’s insurance or flood insurance.
We utilize four categorical variables for the behavioral moral hazard measures: no
preparation = 1 if have not undertaken any of the presented short-term preparation activities, 0
10
otherwise; no window protection = 1 if answer “no” to whether their home has any sort of
window protection, 0 otherwise; no mitigation = 1 if answer “no” to whether ever modified their
home to reduce the amount of hurricane wind damage other than having window protection, 0
otherwise; and no evacuation plans = 1 if answer “no” to whether they plan to evacuate to
purchased supplies for the home such as food, water and batteries; filled car with gas; filled
generator with gas (or readied generator); put up storm shutters; took in furniture or other outside
precautions; and made reservations or plans in case evacuation is needed. While only eight
percent of total respondents indicated not doing any short-term preparation activities, 67, 78, and
71 percent did not undertake any window protection, long-term mitigation, or evacuation plans
respectivelyvii. Lastly, in order to account for individual risk perception of the event in relation to
undertaking any risk reducing activities we include a measure of safety perception. Responses to
the following question were given on a 0 to 100 scale: “How safe did one feel about staying in
your home through the storm, considering both wind and water?” “0” indicated “certain that it
will not be safe” and “100” indicated “certain that it will be safe”viii. The mean perception of
safety values for any one storm were all above 75 indicating that survey respondents felt
relatively safe concerning the impending hurricanes. More detailed information on the real-time
hurricane survey methodology, data, and specific questions can be found in Meyer et al. (2014).
3. Results
Flood risk is a function of exposure (the value of what can be damaged), hazard (the probability
and intensity of a flood), and vulnerability (susceptibility of the building or contents to damage)
11
(Kron, 2005), while flood damage is a single realization of the risk faced. The presence of
adverse selection or moral hazard would cause systematic differences in these variables between
the insured and the non-insured samples. A mean comparison of experienced flood damage
between groups of individuals with, and without, flood insurance (Equation 1) reveals that
insured individuals in Germany suffered significantly higher flood damage to contents and
buildings (Table 1). It can be argued that because a mean comparison contains the ATT and a
selection bias (Angrist and Pischke, 2008) it estimates the combined effect of adverse selection
and moral hazard. Selection bias would contain a type of adverse selection effect as high risk
individuals have a greater incentive to buy insurance. Moral hazard is captured by the ATT
(presented in Table 1) which represents the influence of having insurance on the outcome of
flood damage, while correcting for effects on this damage caused by selecting to buy insurance
(adverse selection). Adverse selection would be expected to increase damages, while the moral
hazard effect is more ambiguous as insurance could cause individuals to become more lax or
insured individuals may take more mitigation measures because they are generally very careful
(risk averse). It is thus an empirical issue to estimate whether individuals with flood insurance
Table 2 panel A presents summary statistics of the hazard experience. It shows that an
element of adverse selection may be present because the treatment group scores higher on
various hazard indicators than the control group. In other words, respondents with flood
insurance suffered from a worse flood event, as the difference in water levels shows. This
suggests that households with flood insurance face a higher flood risk as the overall shape of the
water-level distribution can be argued to be the same across different flood magnitudes, although
12
most important variable of influence on flood damage (DEFRA, 2006; Merz et al. 2010). In
conclusion, adverse selection may be present in the German flood insurance market since the
insured households faced higher risk than non-insured households; indicating that higher risk
households have a greater incentive to purchase flood insurance coverage. Such problems with
adverse selection can, in practice, be limited by reflecting in insurance premiums the higher
flood risks of individuals in floodplains who demand flood insurance than individuals in
The PSM estimates can be regarded as providing an indication of the presence of a moral
hazard effect, because the confounders remove selection effects on flood damage. Table 1 shows
that the expected difference in flood damages between the groups with, and without, flood
insurance is lower once adverse selection has been controlled for using PSM. The latter removed
effects on flood damage resulting from risk-related factors (hazard and exposure) that determined
whether people purchased flood insurance. The ATT estimates suggest that moral hazard is not a
problem, because the difference in damage between the insured and non-insured is statistically
insignificant. For this to be the case the behavior of the two groups must be rather similar with
mitigation measures. For moral hazard to be present the insured group would have to undertake
fewer protective measures. The summary statistics displayed in panel B of Table 2, imply that
both groups have an equal probability of employing at least one damage mitigation measure.
Moreover, the insured group also seems to be more informed about the risk they face as well as
being more likely to be a part of a flood support network. It is also arguable that the insured
group is more risk averse than the non-insured group, as every member of the insured group
employed at least one of the flood coping measures indicated in Table 2, while only 56% of the
13
non-insured group did soix. Therefore, it is possible that the higher level of risk aversion has
reduced any negative moral hazard effect. It may be the case that insurance has directed
individuals to undertake actions that improve their ability to cope with floods, such as joining a
flood support group, although this is unlikely based on the results of Thieken et al. (2006).
The results in Tables 1 and 2 are based on data from both the Elbe and Danube River
catchment areas; however, there may be differences between these two catchments. Because of
historical reasons the flood insurance cultures in the two catchment areas have developed
differently. The Elbe catchment is mainly located in the former German Democratic Republic,
where flood insurance was a part of the compulsory insurance policies household must have.
Even after the reunification of Germany, a large number of households in that area still have an
equivalent set of contracts, while insurance penetration in former West Germany (including the
German part of the Danube catchment) is much lower (Thieken et al., 2006). Therefore, although
the factor is controlled for when estimating the propensity score, there is the potential that
households located in the Elbe catchment are being compared with households in the Danube
catchment area. In order to investigate if the results of Table 1 are being driven by regional
effects, the model was estimated using only the sample of households located in the Elbe
catchment area, and again, but restricted to, the Danube catchment. The results of these spilt
Several new insights emerge from the split sample approach. The overall results of the
estimates for contents damage by the Elbe catchment area and building damage by the Danube
catchment area. Consistent with Table 1 is that within both catchment areas there is no
conclusive evidence for the presence of moral hazard, because the ATT estimates are
14
insignificant. There remains evidence for the presence of adverse selection from the results of the
mean comparison: namely, for contents damage in the Elbe catchment and property damage in
the Danube catchment. However, although building damage in the Elbe and content damage in
the Danube appear to be higher for households with flood insurance, this difference compared
with uninsured households is insignificant. Where the results are statistically significant we see a
The difference in insurance culture between the two regions in Germany provides an
opportunity for examining how this translates in different flood protection behavior. In
particular, in the Elbe area insurance is acquired as a matter of habit, while in the Danube area it
is more of a conscious decision to buy insurance. This allows investigating if the risk averse
population has a higher tendency to buy insurance, which could be reflected by the more choice
based insurance culture of the Danube catchment area displaying a higher portion of the insured
population taking measures to protect themselves. While focusing on the Elbe catchment area on
the other hand provides an opportunity to investigate if whether a more social consensus based
reasoning behind insurance purchase encourages less personal risk mitigation. Table 4 provides
an indication of the difference in damage mitigation attempts between the insured and non-
insured population. On the whole, it appears that the insured group has a greater proportion of its
population employing various damage mitigation measures. In the Elbe catchment area this is a
modest increase across all the measures investigated, while for the Danube catchment area
especially large differences can be found in the use of water proofing and water barriers. This
finding indicates that in general those who purchase insurance have also carried out more
damage mitigation actions, and that this effect is greater when the decision to buy insurance is
15
Results for the United States
Table 5 presents for each hurricane the relationship between the lack of undertaking any
hurricane risk reducing behavior and having homeowners and flood insurance coverage in the
U.S. Negative coefficients signs across all three storms indicate that those survey respondents
that engage in no short or long-term ex-ante property risk reducing behavior are less likely to
have homeowner’s or flood insurance. That is, those without homeowners or flood insurance are
more vulnerable due to a lack of risk mitigation measures, and thus those that have insurance do
not exhibit evidence of moral hazard. However, the statistical significance of this relationship
differs between the hurricanes. For example, those survey respondents that conduct no
preparation activities for Hurricane Irene (the largest proportion across all three storms at 11
percent of respondents) are 55 percent less likely to have homeowners insurance and 38 percent
less likely to have a separate flood insurance policy.xi These effects are statistically significant at
the 1 percent and 5 percent level respectively. Respondents from Hurricane Sandy produce a
similar result in that those that conduct no preparation activities are 64 percent less likely to have
homeowners insurance, which is statistically significant at the 1 percent level xii. For all three
hurricanes those that felt safer are more likely to have homeowners insurancexiii. Furthermore,
those respondents that engage in no ex-ante personal risk reducing behavior (have no plans to
evacuate) are more likely to have homeowner’s insurance. These results suggest a trade-off in
risk aversion to property losses vs. risk aversion to personal harm among our respondents.
The results in Table 5 do not indicate the presence of moral hazard in regard to natural
disaster coverage. However, if insurers are concerned that moral hazard may be occurring one
way to offset this is through the use of a deductible. The deductible forces the insured to have
16
“skin in the game” by making them at least partially responsible for any losses incurred.
Moreover, in the U.S. for wind events, like hurricanes, this amount is likely substantial as
separate wind and hurricane deductibles are typical for homeowner’s policies and range from 1
to 5 percent of the insured value of the home. Unfortunately for insurers relying on a deductible
to offset moral hazard, our survey data suggest that homeowners are not aware of their
deductible amount, or if they are aware believe it to be relatively low. For example, from our
1,442 respondents who indicate that they have homeowners insurance, 62 percent do not know
what their deductible isxiv. Furthermore, only 12 percent believe it to be greater than $1000.
Table 6 presents the coefficient results from a series of estimations aimed at primarily
investigating the deductible’s role in the likelihood of undertaking any short and long-term
preparation activities in addition to controlling for having flood insurance in-place, previous
hurricane damage experience, and the perceived level of safety. Specifically, for each storm and
for only those respondents indicating having homeowners insurance we undertake three separate
statistical estimations: 1) Heckman sample selection model where the selection stage is a probit
model of the likelihood of undertaking any short-term preparation and the outcome component
of the model estimates effects of the explanatory variables on the actual number of preparation
activities undertakenxv; 2) a probit model of the likelihood of having window protection in-place;
and 3) a probit model of the likelihood of having done any other home mitigationxvi.
Similar to the Table 5 results we see little evidence of moral hazard for the insured, as the
likelihood of undertaking any short or long-term preparation activities as well as the number of
preparation activities undertaken has a positive and mostly statistically significant relationship
with having flood insurance in place. As would be expected, having experienced hurricane
damage in the past and feeling less safe are generally positively related to undertaking more
17
preparation activities, although not always statistically significant. In terms of the deductible
coverage, model (1) selection stage coefficient values of having a known low deductible suggest
this is a deterrent to the likelihood of undertaking any preparation activities for Hurricanes Isaac
and Sandy. From the model (1) second stage results in comparison with those having an
unknown deductible amount (the omitted category), coefficient signs generally indicate that
knowing one’s deductible increases the number of preparation activities undertaken, but not
always as is apparent from the Sandy results. Similarly, from models (2) and (3) having higher
levels of known deductible coverage does not always translate into an increased likelihood of
undertaking window protection or other home mitigation as would be expected. As our surveys
were not conducted on a panel of respondents over time and were drawn from different
geographic areas for each storm, these types of differences across storms are expected. For
example, while Irene made landfall in New York in 2011, our Sandy respondents were all from
states outside of NY – VA, MD, DE, and NJ – whereas Irene respondents were drawn from NC
and NY. Despite these inherent differences in respondents however, most notable is the general
lack of statistical significance on any of the deductible variables in the three models, indicating
4. Conclusion
It is often suggested that adverse selection and moral hazard create problems in establishing
well-functioning markets for natural disaster insurance. According to our knowledge, this is the
first systematic empirical study that estimates both of these effects for flood insurance. In
particular, Propensity Score Matching has been applied to estimate the influence of adverse
selection and moral hazard on experienced flood damage by households living along the Elbe or
18
Danube Rivers in Germany. The results show that adverse selection can occur since households
with flood insurance experienced a worse hazard during past flood events in the Elbe or Danube
catchments. However, flood damages do not significantly differ after controlling for this adverse
selection effect, meaning that moral hazard has not heightened the vulnerability of insured
households to floods. Actually the opposite of a moral hazard effect may be present since
individuals with flood insurance in Germany are more likely to have undertaken one of the
These findings are overall robust to a split sample analysis of respondents located in the
former German Democratic Republic and former West-Germany. This split sample approach is
of interest since flood insurance purchases in the latter may be more of a conscious choice and
less driven by habits than in the former German Democratic Republic, where flood coverage was
compulsory. An insight obtained from the split sample analysis is that more conscious choices
for flood coverage are related with more other activities to mitigate flood damage. This supports
other studies conducted on health insurance that higher levels of risk aversion among insured
The evidence from Germany is complemented by a study of moral hazard in the markets
for flood insurance and homeowner’s policies that cover wind damage in the U.S. That analysis
uses real time data on hurricane preparedness and shows that those households that engage in no
short or long-term ex-ante property risk reducing behavior are less likely to have homeowner’s
or flood insurance. This also points towards the opposite of a moral hazard effect. Moreover,
respondents have little specific knowledge of their deductible amount and the complementary
statistical analyses show that deductible levels have no significant influence on undertaking short
19
Results of this study have implications for ongoing discussions about reforming natural
disaster insurance markets in both countries. The results for Germany confirm concerns that have
been raised about the presence of adverse selection. Insurers should reflect this higher risk
profile of individuals who demand flood insurance in a risk-based flood insurance premium.
broader insurance coverage would result in fewer mitigation activities by policyholders are not
supported since we find no evidence of moral hazard. Similar findings are obtained for the U.S.
This questions the use of high deductibles by insurance companies in the U.S. for preventing
moral hazard, and supports ongoing reforms to use insurance for providing financial incentives
1. Household contents damage: damage to household contents, where contents are all
2. Household building damage: Damage to the building – repair costs. Measured in euros.
3. Household contents value: The value of all moveable items within the home. Measured in
Euros.
4. Flood duration: The length of time the building was flooded in hours. Measured in hours.
5. Flow speed one: low water speed (stationary water is the base group). From a 0-4 scale
based on the scale developed by the Bureau of Reclamation (Thieken, 2005). This is a
dummy variable taking the value of 1 if the respondent provided a value of 1, and 0
otherwise.
20
6. Flow speed two: medium water speed (stationary water is the base group). From a 0-4
scale based on the scale developed by the Bureau of Reclamation (Thieken, 2005). This is
a dummy variable taking the value of 1 if the respondent provided a value of 1, and 0
otherwise.
7. Elbe: A dummy variable taking the value of 1 if the respondent lived along the Elbe
8. Urban area: A dummy variable taking the value of 1 if the respondent lived in an urban
9. House age (1948): A dummy variable taking the value of 1 if the respondent’s building
10. House age (1964): A dummy variable taking the value of 1 if the respondent’s building
11. House age (1990): A dummy variable taking the value of 1 if the respondent’s building
12. House age (2000): A dummy variable taking the value of 1 if the respondent’s building
13. House quality 2: A dummy variable taking the value of 1 if the respondent said that the
14. House quality 3: A dummy variable taking the value of 1 if the respondent said that the
15. House quality 3 plus: A dummy variable taking the value of 1 if the respondent said that
21
16. Flood risk 1: A dummy variable taking the value of 1 if the respondent said that a flood
17. Flood risk 2: A dummy variable taking the value of 1 if the respondent said that they
18. Flood risk 3: A dummy variable taking the value of 1 if the respondent said that they
19. Flood risk 4: A dummy variable taking the value of 1 if the respondent said that they
20. Flood risk 5: A dummy variable taking the value of 1 if the respondent said that they
21. Water height: The height of floodwaters entering the house in meters.
22. Contaminated water: A dummy variable taking the value of 1 if the respondent’s house
23. Warning duration: The length of time before a flood that a warning was issued in hours.
24. Return 1: A dummy variable taking the value of 1 if the flood recorded at the nearest
25. Return 2: A dummy variable taking the value of 1 if the flood recorded at the nearest
26. Return 3: A dummy variable taking the value of 1 if the flood recorded at the nearest
27. Cellar: A dummy variable taking the value of 1 if the building has a cellar, and 0
otherwise.
22
28. Floor size: The total floor space of the home, including the size of the cellar if present.
Measured in m2.
29. House price: An estimate of the house price based on the M1914 criteria. Measured in
euros.
30. Warning quality 1: A dummy taking on the value of 1 if the perceived quality of the flood
31. Warning quality 2: A dummy taking on the value of 1 if the perceived quality of the flood
32. Warning quality 3: A dummy taking on the value of 1 if the perceived quality of the flood
33. Renter: A dummy variable taking the value of 1 if the resident rents their residence, and 0
34. Detached house: A dummy variable taking the value 1 (0 otherwise) if the building is a
detached house (this variable is the core base category for housing type)
35. Semi-detached house: A dummy variable taking the value 1 (0 otherwise) if the building
is a semi-detached house.
36. Town house: A dummy variable taking the value 1 (0 otherwise) if the building is a
detached house.
37. Multi-family house: A dummy variable taking the value 1 (0 otherwise) if the building is
a multi-family house.
38. Commercial building: A dummy variable taking the value 1 (0 otherwise) if the building
is a commercial building.
23
39. Secured documents: A dummy variable taking the value 1 (0 otherwise) if the responded
40. Move cars: A dummy variable taking the value 1 (0 otherwise) if the respondent moved
41. Move animals: A dummy variable taking the value 1 (0 otherwise) if the respondent
42. Turn off gas/electric: A dummy variable taking the value 1 (0 otherwise) if the
43. Evacuation: A dummy variable taking the value 1 (0 otherwise) if the respondent had to
References
Akerlof, G. A.. The market for ‘lemons’: quality uncertainty and the market mechanism. The
Arnott, R.J. and Stiglitz, J.E. The basic analytics of moral hazard. Scandinavian Journal of
Barros, P., Machado, M., and Sanz-de-Galdeano, A.. Moral hazard and the demand for health
1025
Botzen, W.J.W. Managing Extreme Climate Change Risks through Insurance. Cambridge
24
Botzen, W.J.W., Aerts, J.C.J.H. and van den Bergh, J.C.J.M.. Dependence of flood risk
(2009), W10440.
Botzen, W.J.W. and van den Bergh, J.C.J.M.. Monetary valuation of insurance against flood risk
Botzen, W.J.W. and van den Bergh, J.C.J.M.. Risk attitudes to low-probability climate change
risks: WTP for flood insurance. Journal of Economic Behavior and Organization, 82 (2012b),
pp.151-166.
Cardon, J.H. and Hendel, I. Asymmetric Information in Health Insurance: Evidence from the
National Health Expenditure Survey. The RAND Journal of Economics, 32 (2001), pp.408-
427
Chiappori, P.A. and Salanie, B. Testing for asymmetric information in insurance markets.
Cohen, A. and Einav. L. Estimating risk preferences from deductible choice. American
France. The Geneva Papers on Risk and Insurance – Issues and Practice, 33 (2008),
pp.645-658.
Cutler, David M., Amy Finkelstein, and Kathleen McGarry. Preference Heterogeneity and
(2008), pp.157-62.
DEFRA. R&D Outputs: Flood Risks to People Phase 2, FD2321/TR1 The Flood Risks to People
25
de Meza, D. and Webb, D.C. Advantageous Selection in Insurance Markets. The RAND Journal
Einav, L., Finkelstein, A., Ryan, S.P., Schrimpf, P. and Cullen, M.R. . Selection on moral hazard
Finkelstein, A., McGarry, K. and Sufi, A. Dynamic inefficiencies in insurance markets: evidence
Finkelstein, A. and McGarry, K. Multiple dimensions of private information: evidence from the
Grossi, P. and Kunreuther, H.C. Catastrophe Modeling: A New Approach of Managing Risk.
Hudson, P., Botzen, W.J.W., Kreibich, H., Bubeck, P. and Aerts, J.C.J.H. Evaluating the
Matching. Natural Hazards and Earth System Sciences, Discussion Paper 2 (2014), pp.681-
723. doi:10.5194/nhessd-2-681-2014
IPCC Field C.B., Barros, V., Stocker, T.F., Qin, D., Dokken, D.J., Ebi, K.L., Mastrandrea, M.D.,
Mach, K.J., Plattner, G.K., Allen, S.K., Tignor, M., and Midgley, P.M. (eds.), Managing the
Risks of Extreme Events and Disasters to Advance Climate Change Adaptation, Cambridge:
Kleindorfer, P.R. and Kunreuther, H.C. The complementary roles of mitigation and insurance in
Kreibich, H., Thieken, A. H., Petrow, T., Müller, M., Merz, B. Flood loss reduction of private
households due to building precautionary measures - Lessons Learned from the Elbe flood
in August 2002. Natural Hazards and Earth System Sciences, 5 (2005), pp.117-126.
26
Kreibich, H., Christenberger, S., Schwarze, R. Economic motivation of households to undertake
private precautionary measures against floods. Natural Hazards and Earth System Sciences,
11 (2011), pp.309-321.
Kunreuther, H.C. Mitigating disaster losses through insurance. Journal of Risk and Uncertainty,
12 (1996), pp.171-187.
Kunreuther, H.C. and Pauly, M. Neglecting disaster: Why don’t people insure against large
Improving Decisions in the Most Misunderstood Industry. Cambridge University Press, New
York, 2013
Kunreuther, H.C., Novemsky, N. and Kahneman, D. . Making low probabilities useful. Journal
Merz, B., Kreibich, H., Schwarze, R., Thieken, A. Assessment of economic flood damage.
Meyer, R., Baker, E., Broad, K., Orlove, B., Czajkowski, J. . The dynamics of hurricane risk
perception: Real-time evidence from the 2012 Atlantic hurricane season. Bulletin of the
Michel-Kerjan, E.O. and Kunreuther, H.C. Redesigning flood insurance. Science, 333 (2011),
pp.408-409.
27
Munich Re. Topics Geo − Natural catastrophes 2012: Analyses, assessments, positions. Munich
07742_en.pdf
Peterson, J.R. and Schneider, H.S. Adverse selection in the used-car market: evidence from
purchase and repair patterns in the Consumer Expenditure Survey. RAND Journal of
Pomerantz, F.J. and Suglia, R.P. Use of hurricane deductibles in the northeast states. FORC
Pirracchio, R., Resche-Rigon, M., Chevret, S., Evaluation of the Propensity score methods for
estimating marginal odds ratios in case of small sample, BMC Medical Research
Rosenbaum, P.R. and Rubin, D.B. The central role of the propensity score in observational
Rothschild, M. and Stiglitz, J.E. Equilibrium in competitive insurance markets: An essay on the
Schwarze, R. and Wagner, G.G. The political economy of natural disaster insurance: Lessons
Seifert, I., Botzen, W.J.W., Aerts, J.C.J.H. and Krebich. H. Influence of flood risk characteristics
on flood insurance demand: A comparison between Germany and the Netherlands. Natural
28
Sloan, F.A. and Norton, E.C. Adverse selection, bequests, crowing out, and private demand for
insurance: Evidence from the long-term care insurance market. Journal of Risk and
Thieken, A. H., Petrow, T., Kreibich, H., Merz, B. Insurability and mitigation of flood losses in
29
Tables
Table 1: Estimates of the difference in average flood damages due to having a flood insurance
policy (in EUR). The average treatment effect on the treated (ATT) is estimated using Propensity
flood insurance
method:
(2241) (5852)
(1684) (4490)
(1874) (4668)
Notes: *,**,*** stand for statistical significance at the 10%, 5% and 1% levels. The numbers in
parentheses are standard errors. Where analytical standard errors are not available, they have been
calculated via bootstrapping with 2000 repetitions. The ATT estimates above have been rounded to the
nearest whole Euro. For a list of variables used in the PS function refer to Appendix A.
30
Table 2: Natural hazard summary statistics
Panel A
Panel B
flood hazard
mitigation measure
31
Table 3: Estimates of the difference in average flood damages due to having a flood insurance
policy (in EUR) for households located in the Elbe and Danube River catchment areas
separately. The average treatment effect on the treated (ATT) is estimated using Propensity
flood insurance
ATT based on PSM using as Contents damage Building damage Contents damage Building damage
matching method:
32
Table 4: Difference in damage mitigation measure usage between the insured and non-insured groups
within the Elbe and Danube River catchment areas. Raw difference between the proportions of the
insured and non-insured population who employ a specific damage mitigation measure.
Note: See Kreibich et al. (2005) for definitions of the above damage mitigation measures.
33
Table 5: Probit model results of the relationship between any hurricane risk reducing behavior and
insurance coverage
Risk
No
-0.8 *** [-.27] -0.48 ** [-.13] -0.28[-.06] -0.35 [-.14] -1.02*** [-.28] -0.34 [-.09]
preparation
No
window -0.29** [-.074] -0.48*** [-.16] -0.14 [-.03] -0.13 [-.05] -0.09 [-.02] -0.29* [-.09]
protection
No
-0.02 [-.01] -0.34*** [-.11] -0.31 [-.05] -0.42*** [-.16] -0.32 [-.05] -0.25 [-.08]
mitigation
No
evacuation 0.27** [.08] 0.03 [.01] 0.18 [.03] -0.02 [-.01] 0.22 [.04] -0.18 [-.06]
plans
Safety 0.07 [.002] -0.05** [-.01] 0.004 [.001] 0.003 [.001] 0.003 [.00] -0.001 [-.00]
Log-
-361.77 -408.44 -109.64 -216.02 -163.6 -276.3
likelihood
LR chi2
(prob > 35.62 ( 0.000) 48.32 (0.000) 8.16 (0.15) 13.71 (0.02) 20.25 (0.001) 10.74 (0.06)
chi2)
Notes: *,**,*** stand for statistical significance at the 10%, 5% and 1% levels. Standard errors have been
34
Table 6: For those with homeowners insurance the relationship between the likelihood and number of preparation activities undertaken and
deductible coverage
Preparation outcome variable (1) (2) (3) (1) (2) (3) (1) (2) (3)
Flood insurance 0.19 0.55*** 0.40*** 0.69*** 0.13 0.38** 0.15 0.34** 0.28*
Experienced damage 0.39*** 0.25** 0.52*** 0.39 0.23 0.34 0.21 -0.20 0.19
Safety -.044** 0.01 -0.02 0.01* 0.003 0.002 -0.002 -0.002 0.004
$0 to $500 deductible 0.22 0.24 0.16 -0.4 0.09 0.1 0.01 -0.32 0.03
$501 to $1000 deductible 0.07 -0.17 0.15 0.08 0.48* -0.00 -0.20 -0.1 0.15
$1001 to $2500 deductible 0.31 0.42** 0.25 0.08 -0.24 -0.05 -0.79*** -0.04 0.44
> $2500 deductible 0.29 0.37 0.23 0.08 0.29 0.22 0.44 0.18 0.48
Constant 3.29*** -0.95*** -1.24*** 2.39*** -0.31 -1.07*** 3.16*** -0.76*** -1.38***
Selection Stage
35
N 612 612 612 294 294 294 469 469 469
Censored observations 40 8 18
Log likelihood -1085.3 -356.3 -276.9 -533.47 -193.79 -187.6 -750.89 -213.84 -222.94
Wald/LR chi2 (prob > chi2) 28.30 (0.00) 40.77 (0.00) 37.59 (0.00) 24.3 (0.001) 10.00 ( 0.18) 10.15 ( 0.18) 15.07 (0.035) 10.14 (0.18) 9.86 (0.19)
AIC 2200.73 728.6 569.81 1096.9 403.5 391.3 1531.78 443.69 461.8
Notes: *,**,*** stand for statistical significance at the 10%, 5% and 1% levels. (1) are the Heckman sample selection model results, (2) and (3) are probit models; each model was
36
i
Adverse selection may not only cause market failures in insurance markets, but since Akerlof (1970) it has been
realised that adverse selection can affect markets for durable goods. For example, Peterson and Schneider (2014)
behaviour of policyholders to make sure that premiums adequately reflect risks. Moreover, insurers can set
deductibles to provide incentives to take measures that limit risks, which reduces moral hazard problems
providing a small incentive for taking damage mitigation measures. An average annual benefit between €5 and €50
act has been partly modified by the Homeowner Flood Insurance Affordability Act signed by President Obama on
damage are removed as these are outlying values. Second, sample is trimmed to only observations within the
coastal counties in Florida, Alabama, Mississippi, and Louisiana, and Sandy respondents were from coastal counties
short-term behavior in regard to questions could have changed after the survey contact. Responses were not ex-post
verified. See Meyer et al. (2014) for more information on the survey application.
viii
The related safety question for Hurricane Irene, an earlier version of the field survey, was slightly different
utilizing a scale of 0 to 10 and not specifically indicating the consideration of both wind and water.
ix
It could be argued that the group employing damage mitigation measures while being insured is richer than their
non-insured counterparts. However, using the same method as employed in Hudson et al. (2014) no important
differences appear to exist in income between the insured and non-insured groups.
37
x
Adverse selection or moral hazard effects are regarded as being absent if, respectively, the mean comparison
estimate and the ATT are insignificant, meaning that their point estimates cannot be statistically distinguished from
zero.
xi
Percent determined from [1 – exp(coefficient value)].
xii
The surveys also collected information on whether respondents had ever previously experienced damage from a
hurricane. Model results for all three hurricanes using this damage information as the dependent variable suggest
that more risky behavior is correlated with less experience. Results can be obtained from the authors upon request.
xiii
See results in Table 6 for the role of feeling safer on the likelihood of undertaking preparation, window
qualitatively similar to those reported in Table 6, especially in regard to the lack of statistical significance on the
deductible levels.
xvi
Separate bivariate probit estimations were run for each storm model (1) to (3) with having a known deductible as
the dependent variable of the other equation. Chi-squared results on rho consistently indicated these as independent
equations.
xvii
To further judge the reliability of inference of the statistical models a post-hoc power analysis was conducted for
the models presented in Tables 1, 3 and 6. The results of the post-hoc analysis indicated that there was sufficient
38