The Income Elasticity of Nonlife Insurance: A Reassessment

• Date: 01 June 2016
• Author: Giovanni Millo
• DOI: http://doi.org/10.1111/jori.12051
• Published date: 01 June 2016

THE INCOME ELASTICITY OF NONLIFE INSURANCE:

AR

EASSESSMENT

Giovanni Millo

ABSTRACT

In aggregate insurance regressions at the country level, the question whether

insurance is a normal or superior good translates into whether income

elasticity is significantly greater than one or not. Twenty-five years after a

seminal article, I reassess the income elasticity of nonlife insurance by means

of homogeneous and heterogeneous versions of the common correlated

effects estimator, controlling for common factors and individual trends and

characterizing the average behavior of insurance markets while allowing

for individual heterogeneity. The evidence supports the existence of a

cointegrating behavior between insurance consumption and GDP and the

view of nonlife insurance as a normal good.

INTRODUCTION

In aggregate insurance regressions at the country level, GDP shows up as by far the

most important driver of growth, proving positive, and significant, for example, in all

studies considered in Outreville’s (2013a) review. Hence, characterizing the long-run

elasiticity of aggregate premiums to GDP is of great value to practitioners trying to

forecast development trends in the medium to longer term. On the other hand, the

theoretical debate has long been going on whether insurance is a normal or a superior

good, translating in an aggregate setting into whether said elasticity is significantly

greater than one or not. I address the question of whether insurance is a normal or a

superior good from an aggregate perspective; that is, do market premiums grow less

or more than proportionally with economic development?

Giovanni Millo is at the Generali Research & Development at Generali S.p.A., via Machiavelli 3,

34132 Trieste, Italy. Millo can be contacted via e-mail: giovanni.millo@generali.com. The article

has benefited greatly from discussions with Gaetano Carmeci; nevertheless, all the errors are

the author’s responsibility. The views expressed are solely his own and do not necessarily

reflect those of his employer. The author is grateful to Swiss Re Research and Consulting for

providing missing data from back issues of sigma. All the computations in the article are done

inside the R open-source environment for statistical computing (R Development Core

Team, 2012), generally using the plm add-on package for panel data econometrics (Croissant

and Millo, 2008). This article has been prepared as a dynamic document with the Sweave utility

(Leisch, 2002) according to the principles of literate statistical practice.

© 2014 The Journal of Risk and Insurance. Vol. 83, No. 2, 335–361 (2016).

DOI: 10.1111/jori.12051

335

In two seminal articles, Beenstock, Dickinson, and Khajuria (1986, 1988) were the

first to consider the behavior of, respectively, life and nonlife

1

insurance with

respect to economic growth by pooling a number of time series from different

countries. Twenty-five years from Beenstock, Dickinson, and Khajuria’s influential

twin articles, methodology has progressed to the point of successfully tackling a

number of then-unresolved problems, already acknowledged by the original

authors: in particular, cross-sectional between-country heterogeneity in the

coefficients of interest (Beenstock, Dickinson, and Khajuria, 1988, p. 259) and

serial correlation (Beenstock, Dickinson, and Khajuria, 1988, p. 267).

2

New

methodological concerns have emerged since, to which the scientific community

paid scant attention or of which it was even scarcely aware back then: most

notably, cross-sectional and spatial correlation, and nonstationarity of variables in

panel data. The latter is possibly leading to spurious regressions in the sense

of Granger and Newbold (1974), so that results of regressions between nonstation-

ary data must be taken with care, at least unless cointegration is proved (see

Phillips and Moon, 1999). The former, cross-sectional correlation, can assume

different forms, essentially based on whether its scope is local and distance

decaying (spatial correlation) or globally affecting every country in the cross-

section (as in common factor models), as formalized in Pesaran and Tosetti (2011),

this second case being the most problematic, as symptomatic of a specification

flaw possibly leading to inconsistent estimates if the common factors’ influence

is not accounted for. National insurance markets are notoriously affected by

common, international factors, like shifts in the global price of reinsurance,

single catastrophal losses of more-than-national scope or global changes in

risk conditions, as most notably happened in 2001 after the World Trade

Center attack, and therefore controlling for this kind of dependence is of the

utmost importance. Fortunately, new estimators are available that are able to

effectively account for unobserved common factors, as will be detailed in the

following.

From the point of view of information, time has again healed some problematic

aspects acknowledged in the original study, like the short time span available

(Beenstock, Dickinson, and Khajuria, 1988, p. 260). Today we draw on a new Sigma

data set, beginning from the year 1970 as in Beenstock, Dickinson, and Khajuria (1988)

but now extending until 2010, thus spanning 40 years of insurance history (Sigma,

various issues). In turn, the geographical scope of the database has been constantly

extended, so that now with respect to the 1981 version in Beenstock, Dickinson, and

Khajuria it comprises a much larger number of countries and completely new areas

like Eastern Europe. Although many “new” countries have been added only recently,

on average the Sigma data set has become “long” enough to allow employing modern

panel time series methods.

1

Beenstock, Dickinson, and Khajuria (1986, 1988) use the term property–liability as in standard

insurance parlance, meaning total nonlife premiums and hence including other lines,

notably—but not only—accident, health, and transport.

2

The use of the Durbin–Watson statistic in a dynamic model is problematic (see Dezhbakhsh,

1990).

336 THE JOURNAL OF RISK AND INSURANCE