Regional Heterogeneity and Spatial Spillovers in the Italian Insurance Market

Transcription

1 R E S E A R C H D E P A R T M E N T Regional Heterogeneity and Spatial Spillovers in the Italian Insurance Market 2005 Annapaola Lenzi Giovanni Millo W O R K I N G P A P E R W P 1 / 0 5

2 This publication is available at the site The Working Papers Series hosts the preliminary results of research projects carried out at the Research Department, published to stimulate discussion and elicit comments. The views expressed in the paper are those of the authors and do not involve responsibility of Assicurazioni Generali SpA. Any errors or omissions are responsibility of the authors.

3 Regional Heterogeneity and Spatial Spillovers in the Italian Insurance Market Annapaola Lenzi Giovanni Millo 15th April 2005 Abstract Italian insurance shows a high degree of regional differentiation in per capita expenditure. Of course this is, at least partially, due to heterogeneity in the economic development of Italian territory. This research aims at assessing the drivers of insurance consumption from this new perspective and testing whether these observable economic, social and demographic factors are able to fully account for regional variability in insurance density or, on the converse, diffusion effects of some kind, such as cross-border or global spillovers, are present. We find evidence of spatial effects and try to assess their nature. We highlight the importance of taking the spatial perspective into account when doing inference on regional data. JEL classification: C21, D12, G22 Keywords: Insurance Consumption, Spatial Econometrics, Heterogeneity, Spillovers Business implications Differences in per capita insurance consumption between the 103 Italian NUTS3 administrative units (province) are huge, both in the welldeveloped life sector and in the relatively underdeveloped non-life sector. They can be only partly explained by the heterogeneity in income, wealth, demography and other observable factors. Insurance density is proportionally increasing with disposable per capita income both in the life and in the non-life sector. Life insurance density, all the rest being equal, tends to be higher in younger regions; non-life in regions with the lowest dependency ratios. Non-observable factors, though, are shaping demand at the macroregional level. In both sectors, the South and Islands are underdeveloped with respect to the rest of the country beyond what explained by observable factors. In non-life insurance the North-West is overdeveloped with respect to North-East and the Centre. Such differences, possibly due to widespread cultural factors, could be indicating strong long-run growth potentials in now-lagging regions. This paper was prepared while the author was visiting the Research Dept. of Assicurazioni Generali SpA. Research Dept., Assicurazioni Generali SpA; corresponding author. Via Machiavelli 4, Trieste, Italy. Tel , giovanni millo@generali.com. 1

4 Contents 1 Introduction 3 2 The theory: background on insurance consumption Non-life Life A brief survey of empirical literature 6 4 Methodology and data issues Measuring insurance consumption Further data issues Administrative boundaries in Italy The information set Some stylized facts about the Italian insurance market Non-Life Life Local disparities and the development of Italian insurance 14 7 Spatial dependence analysis 17 8 Econometric estimates The models Non-Life Cross-section results Spatial heterogeneity and spatial dependence analysis Life Cross-section results Spatial heterogeneity and spatial dependence analysis Conclusions 31 2

5 1 Introduction The theory of insurance demand is well understood and developed. Yet, empirical tests have to overcome significant limitations in data availability, the most binding of which is the difficulty in determining, let alone in observing, what has to be meant by prices and quantities; this often makes estimating supply and demand systems an overly difficult task. Most researchers therefore concentrate on estimating single-equation models of the market equilibrium in which the observed insurance consumption, the product between unobservable equilibrium values for price and quantity, depends on a set of drivers of supply and demand. It is often (more or less implicitly) assumed in studies of direct insurance that supply is very elastic, and that therefore one can concentrate on the drivers of insurance demand by households and firms. Empirical investigations of insurance demand have up to now taken one of two perspectives: either on national aggregates or on (cross-sectional or panel) microdata. This work takes a different perspective, between national aggregates and microdata on households, in order to exploit available imat on aggregated insurance consumption in sub-national regions 1. While losing that part of the theoretically important information that is regionally invariant (e.g. tax rates on premiums or, to a certain extent, loadings on policies and interest rates) and can therefore be of no aid in explaining regional variability, this perspective gains on another front: that of location. The belief that location matters that is at the heart of the current wave of literature about the use spatial methods in applied econometrics (see, e.g., the introduction in Anselin, Florax and Rey [3]) could well apply to insurance consumption. Controlling for spatial effects, when present, is of the utmost importance in order to get unbiased and efficient estimates for the parameters of interest and for doing valid inference and diagnostics. Consistent estimation of regional models of insurance consumption is of interest for practitioners and firms for the purposes of market potential assessment, commercial planning and forecasting; relying on predictions distorted by omitted variable bias can be seriously misleading. From the point of view of the researcher, it can be useful to combine the regional perspective with those of cross-country and micro data in investigating consumers behaviour. Unlike cross-country data, where observational units are separated by borders which are to a great extent coincident with the spatial diffusion of the phenomenon of interest, regional data are aggregated according to more or less arbitrary administrative subdivisions of a territory which is homogeneous from the point of view of language, culture and regulation, and where insurance is sold by the same players at much the same conditions, possibly across borders. Thus, spatial effects are likely to be present and have to be tested for before relying on cross-section estimation, and included in the specification when appropriate. In the following we present a simple example of misjudgment deriving from neglecting the spatial perspective. Besides being a possible nuisance to be controlled for, we contend that assessing the nature of spatial spillovers may also shed some light on the presence of latent, unobservable soft factors underlying insurance purchases: global spatial spillovers, if still found after controlling for drivers of insurance needs and resources, may be reflecting the effects of the attitude towards risk or the 1 To our knowledge, the only regional study on Italian insurance is due to Prosperetti ([21]), who analyses regional data from Italy s province in a cross-sectional fashion. 3

6 awareness of insurance products by the population. Local spillovers, on the contrary, could come from what is called aggregation bias, the overlapping of the boundaries of an economic phenomenon with the artificial borders of administrative regions; in this case, from cross-border purchases or territory-based rate structures. Omission of the former kind in the model specification results in bias and inconsistency, of the latter in inefficiency; the consequences of neglecting local spillovers are therefore less serious, but still invalidating inference on parameters. Our work is introduced by sketching the main results in insurance demand theory, focusing on which are expected to be the main drivers of consumption and realizing that some are difficult to observe. A brief survey of existing empirical literature highlights data limitations to joint modeling of supply and demand and provides a further basis for selecting the relevant information set. Microdata also give evidence of spatial heterogeneity of some kind, in that the behaviour of residents in southern (and, in some cases, northern) Italy is often found significantly different from the rest. We address some further methodological questions: why we focus on demand and feel safe in assuming perfect elasticity of supply with respect to quantity; what we mean by insurance consumption and which measures have been proposed in the literature and by practitioners; we define the two macroclasses of insurance we will concentrate upon, life and non-life excluding mandatory motor third party liability. Then we give some stylized facts about the Italian insurance market, observing the underdevelopment of the non-life (non-motor) class by European standards and the rapid development and reshuffling of distribution channels in the life sector. We highlight the composition of both classes in order to assess the prevailing economic motives of purchase. An analysis of the heterogeneity of Italian territory at large and from the point of view of insurance consumption is followed by visual assessment and formal tests for spatial dependence in the response variables as well as in the explanatory ones. High differentiation and strong spatial dependence among neighbouring regions are found both for insurance consumption and for the relevant drivers. A taxonomy of the relevant spatial alternatives to the familiar cross-section model introduces the econometric estimates: the spatial lag and spatial error models are presented and their properties briefly sketched. Crosssection specifications are estimated for both classes, then tested for the omission of either type of spatial structure. An alternative specification accounting for regional heterogeneity by means of variable intercept and coefficients is searched for and in turn tested against augmented versions incorporating either a spatial lag or a spatial error structure. Variable coefficients are deemed not significant, with the exception of macroregional dummies accounting for intercept shifts. The best models are selected based on measures of model fit and misspecification tests, concluding that location matters for insurance consumption too. While retaining responsibility for all possible errors and misjudgments, we thank: Gaetano Carmeci and Matilde Trevisani for help and guidance; Roger Bivand, Ray Brownrigg and Denis White for kind assistance and advice in overcoming the technical difficulties of displaying georeferenced data; all R developers for their effort. Computations have been performed by R ([24]), in particular using the spdep spatial econometrics package ([5])and the maps package. This paper has been prepared combining R and L A TEX in a dynamic statistical document through the Sweave utility for reproducibility of the results, according to 4

7 the principles of literate statistical practice ([22]). 2 The theory: background on insurance consumption 2.1 Non-life The economic rationale behind life and non-life insurance is much different; so are theoretical models of consumption. Purchasing non-life insurance, a customer buys an indemnity for future losses against paying a fixed price, the premium, today, thus transferring future wealth from an uncertain to a certain state. Theoretical models of non-life insurance demand, starting from the seminal paper of Mossin ([20]), predict that for a given level of risk exposure insurance demand is increasing with risk aversion, probability of loss and total wealth (even though whether the propensity to insure (i.e., the desired coverage as a percentage of the wealth at stake) should increase or not, depends on the behaviour of risk aversion). Moreover, while (by Mossin s Theorem) full coverage is optimal under the fair actuarial price, the degree of coverage decreases with the loadings (Schlesinger, in [9]). The effect of uncertainty of non-insurable wealth is less clear-cut. Guiso and Jappelli ([13]) find that background uncertainty about income has a positive effect on the decision to insure, which becomes less evident as wealth increases. Under the condition of risk-aversion decreasing with wealth, Falciglia ([10]) shows that higher market interest rates lead to a shrinking in insurance demand 2. As far as commercial business is concerned, little-to-medium sized firms can be expected to behave similarly to households regarding the motivations of the insurance purchase. Large firms, on the contrary, might want to bear some risks by themselves, self-insuring through capital reserves, or transfer them through financial markets. Big firms also have contracting power in negotiating premium rates. Firm size could therefore matter Life In the words of Villeneuve (in [9]) life insurance serves to guarantee a periodic revenue or a capital to dependents of the policyholder (the spouse, the children, sometimes the parents or any other person) in case of his death (term life), or to himself, in case he survives. Thus, while the primary rationale behind the purchase of term life lies with the bequest motive, buying an endowment policy 4 or an annuity is mainly an investment choice, and is therefore best viewed as a problem of saving and asset allocation. In the unifying framework first developed by Yaari ([26]) and Hakansson ([14]), the demand for life insurance 2 The explanation lies with the so-called inverted economic cycle of insurance, in which one pays first, then, in the event of loss, receives his dues. Not insuring gives an opportunitygain to invest the spared amount of the premium on financial markets, which increases along with the prevailing rates of return. 3 See Yamori ([27]), Hoyt and Khang ([15]) and Main ([18]). 4 An endowment policy actually has an important term life component (it entitles the beneficiaries to payment of its face value upon death of the insured); the main scope of the cover, nevertheless, is saving. 5

8 is attributed to a person s desire to bequeath funds to dependents and provide income for retirement. Beck and Webb ([4]) synthesize it as follows: the consumer maximizes lifetime utility subject to a vector of interest rates and a vector of prices including insurance premium rates. This framework posits the demand for life insurance to be a function of wealth, expected income over an individual s lifetime, the level of interest rates, the cost of life insurance policies (administrative costs), and the assumed subjective discount rate for current over future consumption. In the case of term life, of course, also of the number, personal characteristics and preferences of the beneficiaries (see the extension of this scheme by Lewis ([17]), that is, in most cases, of family composition. 3 A brief survey of empirical literature When trying to put theories at the test, many variables need to be proxied because of data limitations. Empirical studies identify some observable counterparts to total wealth, risk exposure, probability of loss and risk aversion. Wealth, when not observable, is generally proxied by means of income; so is risk exposure, which is in turn related to total wealth and the level of economic activity. Loss probability may too be related to income as a measure of economic activity; urbanization has also been suggested for this purpose ([6]). Loss ratios of previous periods have also been suggested as a proxy for the probability of loss. Aspects of risk aversion may be captured by education or the age structure of the population, even though the expected sign of the effect is unclear (see [7] and the discussion in [6]). Summaries of the existing empirical literature can be found in [6] (propertycasualty), [12] (health), [4] and [19] (life). We will focus on the practical implications of previous studies on the preliminary choice of the set of variables to be included in the model. Cross-country comparisons by Grace and Skipper and Browne, Chung and Frees find a positive relationship between non-life insurance demand and income, literacy, religion and the type of legal system. For the life class, Beck and Webb ([4]) add inflation and the degree of development of the banking sector. Education, young dependency ratio, life expectancy and size of social security do not prove significant in this setting. Studies on microdata (usually household income or consumer expenditure surveys) have emerged in recent years, both in Italy and abroad, focusing mostly on the probability of purchasing an insurance policy (see, e.g., [25], [23]). The Italian reality has been analyzed in a number of recent papers drawing on data from the Bank of Italy s Survey on Household Income and Wealth (SHIW). Guiso and Jappelli ([13]) find that the effect of household resources, expected income, income risk, self-employment, education and urbanization on the decision to purchase casualty insurance is positive and significant, while that of age is non-linear and family size is not significant. They also find that macroregional dummies for North and South are, respectively, significantly positive and negative. Their results are quite similar for the amounts insured, but the role of age, macroregional dummies and education is weaker. Focarelli, Savino and Zanghieri ([12]) investigate separately the probability of buying health and other kinds of non-life insurance (excluding mandatory motor TPL). They find evidence of positive effects from income, financial wealth, education and residence 6

9 in the North or Centre for both classes, while age is not significant. Male, selfemployed and homeowners are more likely to have property-liability-casualty insurance, while managers are more likely to have both. Prosperetti ([21]), in what is to our knowledge the only disaggregated analysis so far on the Italian insurance market, decomposes non-life insurance into personal and commercial lines and into the main subcategories (liability, fire, theft, accident and health). He estimates cross-section models for insurance penetration in each and finds significant effects from (among demand side drivers) per capita consumption or income and, for some classes, from added value composition (shares of agriculture and industry) and firm size. He also includes (as supply side drivers) the structure of distribution (shares of big-, mediumand small-sized agencies and of brokers) and finds positive correlation between agency size and insurance penetration. As far as life insurance is concerned, Michielin and Billari ([19]) analyze term life and pension insurance in the same probit fashion, finding that the decision to buy pensions policies is positively influenced by income, residence in the North (and, to a lesser extent, in the Centre) and, in a non-linear way, by age; term life purchase is positively correlated with family size, education, self-employment, income and home ownership, while urbanization and macroregion of residence are deemed insignificant. Cannata, Menegato and Millo (in [8]) analyze possession of term life and endowment policies. They find evidence of positive correlation of term life ownership with income and number of children, while the coefficients for age, number of family members excluding children, financial wealth and residence in the South are significantly negative; for endowment policies, the effect of income, education and marriage are positive; of children and other family members, not significant; of age, residence in the South and, notably, (other forms of) financial wealth, negative. 4 Methodology and data issues Insurance data pose an inherent limitation to joint estimation of a supply and demand system. As Schlesinger (in [9]) notes, it is often difficult to determine what is meant by the price and the quantity of insurance. [...] the fundamental two building blocks of economic theory have no direct counterparts for insurance. In practice we can usually only observe insurance consumption, the product between equilibrium price and quantity, jointly determined by the interplay of supply and demand. In most of the empirical literature the authors concentrate on the determinants of demand, more or less implicitly assuming that supply is very elastic and that adjustment happens through quantities at a given price. 5 A tentative interpretation scheme that seems consistent with common knowledge about the Italian market is that supply is oligopolistic and the price is determined as a mark-up on costs, i.e. on claims plus commissions and administration expenses. Most of the Italian market is dominated by comparatively 5 Price proxies related to costs, notably the loss ratio, are often used in empirical work on non-life insurance; these are nevertheless available only for national aggregates. Browne, Chung and Frees ([6]), in a cross-country comparison perspective, suggest proxying insurance prices by the market share held by foreign insurers. In the case of life insurance, prices have been proxied by means of the ratio of premiums to business in force ([7]); Beck and Webb ([4]) discuss this issue and suggest a number of proxies. 7

10 few big nationwide players (the market share of the first five groups is over 60 percent both in life and in non-life). At regional level, we can therefore safely assume that shifts in demand do not affect the cost structure of supply too much, thus price is mainly determined by the supply side and the supply curve is very flat, the insurers being willing to offer as much cover as needed at the given price. Our strategy is therefore to concentrate on the drivers of demand, estimating a single-equation relationship for insurance consumption in each of the two sectors. Care shall be taken, in assessing the results, that according to anecdotal evidence the price structure, while quite uniform for the life sector, is indeed locally differentiated for some classes of non-life insurance. As we are unable to include it explicitly in the model for lack of data, this price structure could reflect in spatial dependence in the error terms deriving from omitted variables bias Measuring insurance consumption Measuring insurance consumption across administrative regions of different economic and demographic size requires resorting to some kind of relativization. Two common normalized measures are used in the literature as well as among practitioners: insurance penetration, defined as the ratio of insurance premiums on GDP, measures the importance of the insurance sector with respect to the total economy; insurance density, defined as premiums per capita, measures average per capita expenditure. Further measures specific to the life sector have been suggested ([4]). Unlike Prosperetti ([21]), we focus henceforth on premiums per capita. In the same fashion, all variables subject to a size bias in the information set have been normalized with respect to the relevant benchmark. Data on insurance premiums are collected on a regional basis by ISVAP, the Italian insurance Authority, divided into three categories: life, compulsory third party liability, the vast majority of which regarding motor vehicles, (henceforth MTPL) and other non-life. While MTPL is a homogeneous class, both life and other non-life comprise very different kinds of policies. In the following we give a description of the relative importance of each class in the aggregate (see Tables 3 and 5). We do not consider MTPL as, being a mandatory cover, it does not fit into the theoretical scheme sketched above 7 ; we focus on the other two macroclasses, life and non-mtpl non-life (henceforth, non-life tout court). 4.2 Further data issues Premium data are registered according to the location of salespoint as communicated by the companies. Besides the inevitable aggregation bias due to the arbitraryness of administrative boundaries with respect to the geographic dimension of economic phenomena (see [1]), some important additional biases may arise if the location of salespoint is different from the actual location of the insured. 6 Of course there are some important variables that have no cross-regional variability and are therefore omitted from the specification: examples are taxation, but also loadings that are uniform for all regions, as is the case for most life policies. 7 The decision to purchase an MTPL policy is a consequence of that of buying a car, thus the appropriate drivers of consumption have to be assessed differently. 8

11 Figure 1: Map of Italian macroregions First, mostly for big contracts negotiated by brokers but also for some distribution agreements, e.g., in bancassurance, some big units, usually located in an important industrial or financial centre, are accountable for all business nationwide. This happens, for example, for marine insurance premiums collected by business units located in the main harbours for customers located and doing business elsewhere, or for some nationwide salesmen network whose business goes through a single agency, typically located at the company headquarters. Second, collective policies purchased by the firms as a mandatory cover or as a fringe benefit for their employees, most typically in the accident, health and life classes, are bound to one salespoint location even if they are actually insuring risks spread over a wider territory. 4.3 Administrative boundaries in Italy In the following, we refer to the Italian administrative units called province, corresponding to level 3 in the NUTS (Nomenclature of Territorial Units for Statistics) classification by Eurostat, using the generic name of regions, and to the classification used by Istat, the Italian statistical office, when speaking of macroregions. Macroregions divide the 20 NUTS2 Italian regions (regioni) into 5 aggregates: North-West, North-East, Centre, South and Islands (see Tables 21, 22, 23 and Figure 1). 9

12 4.4 The information set The information set consists of an excerpt for the year 2000 from the GeoStarter database provided by Istituto Tagliacarne, an institution inside SiStaN (the Italian national statistical system). It provides both first-hand data and an organized collection of data from various institutional sources. Insurance data, in particular, are provided by Isvap, the Italian regulatory body. The original information set is huge, with variables in the hundreds. In order to reduce it to a tractable number of non-collinear potential regressors, we have restrained the original database by a combination of theoretical assumptions, heuristics and data mining procedures to less than 50 variables. These are representative of: population, age structure and territorial organization; productive tissue of the region; income and wealth; other. See the full list in Table 1. 5 Some stylized facts about the Italian insurance market 5.1 Non-Life The Italian non-life insurance market is still underdeveloped with respect to those of the main European countries. The penetration ratio (premiums/gdp) is lower than in the other four big economies (Germany, France, United Kingdom and Spain), especially in non-motor business. The class is dominated by MTPL, accounting for more than a half of non-life business. Its penetration is higher than in the rest of Europe both because of the high number of vehicles on the road and because of the steady, cost-driven increase in tariffs of the last years. Non-mandatory classes, on the contrary, are far less developed, with total penetration less than half that of our major European partners. The composition of non-mtpl non-life is balanced, with property as the leading class, at 12 percent of total non-life revenue, and non-mandatory motor, general liability and accident between 8 and 9 percent. As Focarelli, Savino and Zanghieri ([12]) note, Health insurance is most underdeveloped with respect to the rest of Europe, despite high private health expenditure. Marine, aviation and transit and credit and suretyship, both at little above 2 percent, play a minor role (Table 3). There are no data available about the share of personal and commercial lines in the revenues of every class, but according to common wisdom this is quite balanced, maybe slightly biased towards personal lines, in property; balanced in accident and health, with comparatively few but huge collective contracts purchased by the firms; and definitely leaning towards the commercial side in liability insurance. Non-life insurance in Italy is mostly distributed through tied agents, collecting about 85 percent of revenues (see Table 4) 8. The remainder is sold through brokers and, with lesser shares, through bank counters and direct channels (telephone, Internet). The direct channel still accounts only for about 3 percent of total revenues, though its importance is steadily increasing. 8 Data are comprehensive of MTPL. 10

13 Code Description 1 codistat Istat code of province 2 den Population density 3 numcompfam Average number of family members 4 popmaschi Male population 5 pop Total population 6 pop25.29 Population aged.. 7 pop60.64 Population aged.. 8 pop65.69 Population aged.. 9 pop70 Population aged.. 10 pop30.34 Population aged.. 11 pop35.39 Population aged.. 12 pop40.44 Population aged.. 13 pop45.49 Population aged.. 14 pop50.54 Population aged.. 15 malprof Professional disease cases 16 inflav Working accident cases 17 Ydproc Per capita disposable income 18 Cproc Per capita consumption 19 irpef Income tax 20 addunlocproc Employed in productive units per capita 21 addperunloc Average number of employed per productive unit 22 unlocagr Productive units, agriculture 23 unlocpesca Productive units, fishing 24 unlocestraz Productive units, mining 25 unlocmanif Productive units, manifacturing 26 unlocenergia Productive units, energy sector 27 unloccostr Productive units, building 28 unloccomm Productive units, commerce 29 unlocalbergh Productive units, hotels and tourism 30 unloctrasp Productive units, transports 31 unloccrediass Productive units, banking and insurance 32 unlocservimprese Productive units, business services 33 addgrmag Workers at big retailers 34 autocircproc Vehicles on the road per capita 35 autopercoltre2000 Percent of cars over 2 litres displacement 36 depproc Bank deposits per capita 37 impproc Bank loans per capita 38 pop25.54 Population aged.. (elaboration) 39 pop30.49 Population aged.. (elaboration) 40 popover60 Population aged.. (elaboration) 41 add.agrpesca Workers per sector, agriculture and fishing 42 add.estraman Workers per sector, mining and manufacturing 43 add.costr Workers per sector, building 44 add.commserv Workers per sector, commerce and services 45 add.crediass Workers per sector, banking and insurance 46 add.trasp Workers per sector, transport 47 sportpermille Bank counters per 1000 residents 48 ppcv Insurance density, Life 49 ppcd Insurance density, Non Life Table 1: Final stage variables 11

14 Total Life Non-Life Motor Non-Motor 1 United Kingdom France Italy Germany Spain Table 2: Insurance penetration in Europe, 2003 Class Premiums Share 1 Accident Health Motor other risks Marine aviation transit Fire Other damage to property Motor TPL General TPL Credit and suretyship Others Total Non life Table 3: Composition of Non-Life insurance Channel Tied agents Brokers Direct Company staff Banks Financial promoters Table 4: Distribution channels of Non-Life insurance 12

15 Class Premiums Share 1 Class Class Class Class Total Life Total Life + Non Life Table 5: Composition of Life insurance Channel Tied agents Brokers Post Company staff Banks Financial promoters Others Table 6: Distribution channels of Life insurance 5.2 Life The development of Italian life insurance in the last years has been spectacular, driven by the explosion of bancassurance, the use of bank counters as a distribution channel, and the development of new types of contract besides the traditional ones. Unit- and index-linked contracts in fact fluorished during the stockmarket boom of the late Nineties, pushed through the bank channel together with other financial products, their distinction with the latter becoming even more blurred. Later on, with the bursting of the stockmarket bubble, the product mix again shifted towards policies with capital guarantees of some form; in the meantime, a new player, the Italian Post Office, which had been a provider of very traditional, low-return and riskless savings accounts, entered the life market reconverting its many salespoints to the sale of financial products, including life policies, quickly gaining a significant market share and prompting another shuffle in the market (see Table 5 9 ). Nowadays the banks share in the distruibution of life policies (Table 6) is steady at 50 percent of premiums, while the Post Office is quickly gaining ground at the expense of tied agents. Financial promoters and company staff hold a minor and quite steady slice. The strategies of the supply side play a major role in driving revenues of one channel over the other or those of life insurance over competing financial products from the same groups. Italian life insurance is therefore a far less stable market than non-life. It has also recently become a much bigger one, at 4.9 percent of GDP against 2.6 percent of non-life in 2003, after lagging behind for many years. Most of the 9 Classes are reported according to the Italian classification; they correspond approximately to traditional endowment and annuities (plus term life, accounting for less than 3 percent of total) in class 1, unit- and index-linked policies in class 3, capitalization (life-independent endowment) in class 5, pension plans in class 6. 13

16 Min. 1st Qu. Median Mean 3rd Qu. Max. Gini Inhab./Km Family size Perc. aged Disp. income Pr.p.c., life Pr.p.c., nonlife Penetr. life Penetr. nonlife Table 7: Distribution across NUTS3 regions and inequality measures of some characteristics of Italian territory revenues come from endowment and annuities, either guaranteed or linked, while term life plays a residual role and long term care and dread disease policies are almost negligible. The vast majority of business is believed to belong to personal lines, though important collective policies are often sold to the firms. Thus, from an economic point of view, the relevant framework is that of saving theory. 6 Local disparities and the development of Italian insurance Italy is well known to be in many respects a heterogeneous country. Italian regions are highly differentiated both from the social, cultural and demographic point of view and from the more strictly economic one. The age structure of the population leans towards the older classes in the North-West and the Centre-North, while the residents in central North and the South are youngest. The structure of the family is also very differentiated: the average number of members of a family goes from 2 in Trieste to over 3 in Naples, constantly decreasing with latitude. Per capita income is highest in the North and the capital, lowest in the South, with all kinds of nuances in between (Table 7). Indicators of economic development like, e.g., registered cars per capita are similarly distributed, but with a higher concentration in the North-West and Centre-North than in the north-east. Unemployment, while on average close to that of the largest European countries, is dramatically higher in the South and negligible in most regions of the North and part of the Centre. The North is most industrialized, even though the share of industry in the productive tissue of the land is high also in most of the Centre, with the notable exception of Rome and its surroundings. Services, most notably touristic ones, dominate in the Islands and are important in the South as a whole. Insurance consumption is no exception. All of the last 20 regions in the overall ranking, both in the life and non-life classes, come from the South and Islands; all but three (in non-life) and one (in life) of the first 20 are northern regions (Tables 8 and 9). A comparison with the other commonly used measure of insurance consumption shows (Tables 10 and 11) that the situation is little changed by considering 14

17 Ranking Province Macroregion Ins. density 15 1 MILANO North West GENOVA North West PRATO Centre BOLOGNA North East TORINO North West BIELLA North West VERCELLI North West MODENA North East BOLZANO North East PARMA North East VARESE North West ROMA Centre PIACENZA North East BRESCIA North West FIRENZE Centre TRENTO North East ALESSANDRIA North West NOVARA North West AOSTA North West REGGIO EMILIA North East CATANIA Islands SIRACUSA Islands CROTONE South SALERNO South BRINDISI South CAMPOBASSO South RAGUSA Islands POTENZA South NUORO Islands TARANTO South AVELLINO South LECCE South CASERTA South BENEVENTO South REGGIO CALABRIA South CALTANISSETTA Islands COSENZA South TRAPANI Islands AGRIGENTO Islands ENNA Islands VIBO VALENTIA South Table 8: Ranking of Italian provinces by Non-Life insurance density 15

18 Ranking Province Macroregion Ins. density 17 1 BRESCIA North West MILANO North West MANTOVA North West VERONA North East BIELLA North West REGGIO EMILIA North East PARMA North East SONDRIO North West MODENA North East NOVARA North West FIRENZE Centre TREVISO North East BERGAMO North West CREMONA North West GENOVA North West VICENZA North East VARESE North West VERBANIA North West BOLOGNA North East LECCO North West TERAMO South TRAPANI Islands CAGLIARI Islands CATANZARO South BENEVENTO South PALERMO Islands AVELLINO South LECCE South SIRACUSA Islands ORISTANO Islands RAGUSA Islands CATANIA Islands BRINDISI South AGRIGENTO Islands CALTANISSETTA Islands POTENZA South ENNA Islands CROTONE South COSENZA South NUORO Islands VIBO VALENTIA South Table 9: Ranking of Italian provinces by Life insurance density 16

19 insurance penetration 10, though this statistic washes out the effect of the differences in income distribution. In other words, at a first glance heterogeneity in insurance consumption doesn t seem to be only due to difference in available resources, as the average propensity to buy insurance out of one s income is almost as differentiated. 7 Spatial dependence analysis Besides high macroregional differentiation, insurance penetration shows a high degree of spatial correlation 11. Moran plots (see Figures 2 and 3) show an evident cluster of low-density regions highly correlated with their neighbours, but the same applies to most observations. Influential outliers (indicated by a different symbol and labeled by their Istat code) also situate in the positive correlation sector, exception made, in the life class, for some relatively low-density regions with high-density neighbourhoods (Aosta, Trento, Rovigo and Lodi). In the non-life class Rome is a notable outlier in the second sector, that is, it is much more developed than its neighbourhood. The situation is much alike as far as the explanatory variables are concerned. Moran plots (not reported) clearly indicate spatial dependence. More formal statistical tests (Table 12) confirm the visual impression. In the following we try to assess whether the local disparities in explanatory factors are sufficient in explaining local variability; whether variable coefficients are needed to account for spatial heterogeneity; and whether global or local spillovers are significant, and should enter a model specification. Consistent with the idea that insurance covers are complicated and littleunderstood goods, that the Italian market is underdeveloped with respect to the overall economic development of the country, at least as far as the nonlife sector is concerned, and that the different degree of insurance penetration throughout the Italian territory may be due to ummeasurable, soft factors of rather diffuse nature (such as cultural propensity to insure, understanding of the products and awareness of insurance needs on the public s part) 12 as well as to differences in resources and needs, we expect to find some evidence of 10 Here, given the consumption-orientation of the analysis, we employed a slightly different measure of insurance penetration, normalizing on disposable income instead of GDP. 11 Tests and diagnostic plots for spatial correlation as well as spatial models are based on a spatial weights matrix constructed according to the principle of queen contiguity (that is, regions are considered neighbours if they share a common border or vertex; see [16]). According to common practice, the matrix has been row-standardized. Reggio Calabria and Messina, divided by the Messina Strait, have been considered contiguous. 12 There is evidence ([13]) that better educated households purchase more insurance, even when controlling for the higher income levels that may be expected to go together with higher schooling. The authors find strong evidence of positive correlation between education and the insurance decision and weaker evidence of correlation with the amounts insured. We argue that this is another sign of the importance of cultural effects of some sort on the insurance purchase, though we are not able to test the influence of schooling in this context (the only data available to us in this respect are ratios of alumni to the total corresponding age cohort, which have proved not significant. Data on the shares of different education levels in the population would probably be useful, but the only such information accessible to us were 1991 census data). From a geographic point of view, though, insurance culture might pertain more to people s attitude towards risk in general than to awareness and understanding of insurance products that may come from higher schooling. 17

20 Ranking Province Macroregion Ins. penetration 15 1 MILANO North West GENOVA North West PRATO Centre BOLZANO North East BIELLA North West TORINO North West TRENTO North East BRESCIA North West VARESE North West SONDRIO North West ROMA Centre BOLOGNA North East BERGAMO North West VERCELLI North West PAVIA North West CREMONA North West LODI North West MANTOVA North West COMO North West VICENZA North East CATANIA Islands MESSINA Islands SIRACUSA Islands CALTANISSETTA Islands POTENZA South BRINDISI South LECCE South RAGUSA Islands NUORO Islands CASERTA South SALERNO South TRAPANI Islands TARANTO South CAMPOBASSO South BENEVENTO South COSENZA South REGGIO CALABRIA South AVELLINO South AGRIGENTO Islands ENNA Islands VIBO VALENTIA South 0.55 Table 10: Ranking of Italian provinces by non-life insurance penetration 18

21 Ranking Province Macroregion Ins. penetration 17 1 BRESCIA North West MANTOVA North West SONDRIO North West VERONA North East MILANO North West BIELLA North West REGGIO EMILIA North East TREVISO North East BERGAMO North West NOVARA North West VERBANIA North West CREMONA North West VICENZA North East PARMA North East MODENA North East VARESE North West PESARO E URBINO Centre SIENA Centre AREZZO Centre LECCO North West CATANZARO South PALERMO Islands ORISTANO Islands AOSTA North West CROTONE South CAGLIARI Islands CATANIA Islands BENEVENTO South RAGUSA Islands SIRACUSA Islands CHIETI South ROVIGO North East BRINDISI South ENNA Islands SASSARI Islands POTENZA South TERAMO South AVELLINO South NUORO Islands COSENZA South VIBO VALENTIA South 1.76 Table 11: Ranking of Italian provinces by life insurance penetration 19

22 Moran s I P-value Geary s C p-value ppcd ppcv Ydproc depproc impproc Table 12: Spatial dependence tests spatially lagged ppcd ppcd Figure 2: Moran plot of Non-Life insurance density 20

23 ppcv spatially lagged ppcv Figure 3: Moran plot of Life insurance density 21

24 unmodeled global effects. If, on the contrary, the cross-section specification were able to fully account for the observed variability, this would bring evidence in favour of the view that less developed regions are such simply because of budget constraints and lower insurance needs. This would not rule out the possibility of local spatial effects due to aggregation biases of some kind: for example, due to the overlapping of administrative boundaries with operational areas of the sales force or any other kind of cross-border purchase. 8 Econometric estimates We estimate a standard cross-section specification, in which all variables have been log-transformed in order to remove heteroskedasticity. A general-to-specific specification procedure à la Hendry leads to two very simple cross-section models for the respective sectors (see Tables 13, 17). We proceed testing for spatial dependence in the cross-sectional model and then trying to account for macroregional heterogeneity implementing variable intercept and coefficients in the cross-sectional specification and finally testing whether spatial effects of some kind persist. Spatial specification tests are conducted according to the specific to general procedure recommended in [11]. Assessments of model fit and the main specification test results are presented in the text, parameter estimates and significance tests in Tables 13 to The models The simplest and most widely used tool in cross-section analysis is the linear regression model: Y = Xβ + ɛ (1) which, as is well known, can be estimated consistently and efficiently by OLS provided that 1. E[Y/X] = Xβ 2. E[ ɛɛ /X] = σ 2 I n Such conditions do rarely hold in spatial data, which often exhibit spatial heterogeneity, spatial dependence or a combination of the two (see [1]). Spatial heterogeneity pertains to coefficient variability in space, and can be dealt with by standard variable coefficients techniques. Spatial dependence may come in two forms, global and local, pertaining either to observed (modeled effects) or to unobserved variables (unmodeled effects) (for a taxonomy of spatial models, see [2]). For reasons given above, we concentrate on unmodeled effects. Global spatial effects are related to the exclusion from the model of some unobservable latent variable that can be diffused globally. Omission of the latter reflects in a spatial autoregressive process (SAR) in the errors: all errors at different locations in space influence all others, with an intensity decaying with distance. The appropriate model is then the spatial lag model: Y = Xβ + ɛ ɛ = ρw ɛ + u (2) 22

25 Estimate Std. Error t value Pr(> t ) (Intercept) log(ydproc) log(depproc) log(pop25.54/popover60) log(sportpermille) Table 13: Cross-section model, Non-Life or, in reduced form, Y = Xβ + (I ρw ) 1 u (3) The consequences on estimation of omitting a globally diffused spatial structure are inconsistency and biasedness of parameter estimates. Local effects mean the errors influence each other only between neighbouring regions. Errors follow a spatial moving average process (SMA), reflecting in the so-called spatial error specification: or, in reduced form, Y = Xβ + ɛ ɛ = u + γw u (4) Y = Xβ + (I + γw )u (5) Neglecting a spatial error structure has less serious consequences: estimates, while still consistent, are inefficient, thus invalidating inference. 8.2 Non-Life Cross-section results Non-life density is found to be significantly and positively dependent from income and wealth (as proxied by bank deposits), from an indicator related to the age structure of the population 13 and to the number of bank counters per 1000 inhabitants (Table 13). The overall fit is good (R 2 is 0.93) and the diagnostic tests do not reject normality and omoskedasticity of residuals. While the ratio of working age people to over-60s has a direct interpretation in terms of insurance needs for production-related policies (such as professional liability, accident and covers for every kind of entrepreneurial risk), the role of the bank counter density is unclear as this distribution channel holds an almost negligible share of the non-life sector (see Table 4). We suspect this to be due to spurious correlation Spatial heterogeneity and spatial dependence analysis Visual inspection of the choropleth map of residuals (Figure 4; positive residuals are colored in red, negative ones in blue) reveals a large cluster in which the 13 This is defined as the ratio of individuals aged between 25 and 54 to those over 60 years. It can be viewed as inversely related to the dependency ratio. 23