Economic growth and crime: Is there an asymmetric relationship?

Economic Modelling 49 (2015) 286–295

Contents lists available at ScienceDirect

Economic Modelling journal homepage: www.elsevier.com/locate/ecmod

Economic growth and crime: Is there an asymmetric relationship?☆ Eleftherios Goulas a,⁎, Athina Zervoyianni b a b

Department of Banking & Finance, Bursa Orhangazi University, Turkey Department of Economics, University of Patras, Greece

a r t i c l e

i n f o

Article history: Accepted 27 April 2015 Available online xxxx Keywords: Economic growth Crime

a b s t r a c t We examine the relationship between crime and per-capita output growth in a panel of 26 countries for 1995– 2009, focusing on the various channels through which crime can constrain growth and exploring the extent to which these channels are influenced by economic conditions. A simple structural growth model serves as a guide for the empirical specification and a reference point for the interpretation of the empirical results. Our estimates suggest significant potential gains from reducing crime during periods of worsening economic conditions, when market sentiment is pessimistic, and thus uncertainty regarding the return to saving is above average, employment is low, and the strain on government-sector resources through high public-safety spending is already sizable. Crime does not seem to be so harmful to growth when economic conditions are sufficiently satisfactory. In this respect, our results provide an explanation for the inconclusive empirical evidence, based on reduced-form models, regarding the strength of the growth–crime relationship. © 2015 Elsevier B.V. All rights reserved.

1. Introduction Crime imposes a burden on society and an extensive literature currently exists suggesting that the socio-economic costs of crime can be sizable (Czabanski, 2008; European Commission, 2010; World Bank, 2006, 2007). And while in most parts of the world crime rates are today lower compared to those recorded a few decades ago, a large fraction of the population in many countries still experiences crime every year. Moreover, public expenditures on crime prevention and law enforcement remain at high levels, crowding out other, more productive, types of government spending. At the same time, the decline in crimerelated activity may not continue at the same pace in the coming years, given the reduction in incomes due to the recent fall in economic activity worldwide. In view of these developments, how crime impacts on economic growth becomes particularly important. Although the importance of crime in determining a country's progress has long been recognized in the economic-policy literature, empirical studies have not yet produced a definite conclusion regarding the effect of crime on growth. Existing findings are contradictory, with some studies suggesting a strong adverse effect of crime on economic growth while other studies report evidence of no statistically significant impact. A recent World Bank study (World Bank, 2006), using a panel of 43 countries for 1975–2000, reports strong growth-reducing effects from higher crime rates even after controlling for a number of other

☆ We are grateful to two anonymous referees for valuable comments on earlier versions of the paper. ⁎ Corresponding author. E-mail addresses: [email protected] (E. Goulas), [email protected] (A. Zervoyianni).

http://dx.doi.org/10.1016/j.econmod.2015.04.014 0264-9993/© 2015 Elsevier B.V. All rights reserved.

factors affecting growth, including income inequality which is likely to be causally linked to crime. Càrdenas (2007) also finds a statistically significant negative association between per-capita-output growth and crime in a panel of 65 countries, after allowing for unobserved country-fixed effects and controlling for education and public infrastructure. On the other hand, Peri (2004), using provincial-level crime data from Italy, reports results indicating non-linearities in the growth–crime relationship, with modest- and low-crime showing no statistically significant adverse impact on growth. (Burnham et al., 2004), in exploring the effect of central-city crime on US county-level (per-capita) income growth, report results in the same direction, indicating no clear overall growth–crime relationship, with the growth effect of property crime appearing to be weak or perverse. At the same time, Mauro and Carmeci (2007), using data from 19 Italian regions for 1963–1995 and pooled-mean-group estimation techniques, find that crime impacts negatively on income levels but exerts no statistically significant adverse influence on growth rates. Chatterjee and Ray (2009), using a large cross-country dataset for 1991, 1995, 1999 and 2003 and controlling for human capital and institutional quality, report similar results, as they find no strong evidence of a uniformly negative association between growth and crime. Detotto and Otranto (2010), applying an autoregressive model, in which GDP growth is explained by past GDP and a crime proxy, to monthly crime data for Italy for 1979–2002, also find a small annualized real-GDP-growth reduction due to crime, with their estimates indicating cyclical components in the growth–crime relationship. These results suggest that, despite the growing empirical literature, the effects of crime on economic growth still are not well understood and that the growth–crime relationship is more complex than often assumed in existing studies. Crime may affect growth through four key

E. Goulas, A. Zervoyianni / Economic Modelling 49 (2015) 286–295

channels: (i) through lower physical- and human-capital productivity, by undermining confidence in the rule of law and thus discouraging innovation and entrepreneurship and the accumulation of knowledge via education; (ii) through the opportunity cost of public control of crime, as government-sector resources that could be used for productive activities, including education, health and infrastructure, are directed to crime prevention and law enforcement; (iii) through reduced labour supply, to the extent that some individuals are inclined to believe that income can be earned through illegal activities while others deliberately reject certain job types or job locations due to the fear of criminal victimization; and (iv) through reduced savings due to less secure property rights, as high crime rates contribute to a general perception of instability and bad business climate. Much of the existing empirical literature uses reduced-form models that cannot shed light on the different channels via which crime impacts on growth and on the extent to which the strength of these different channels is influenced by current economic conditions. This paper adds to the growth literature by distinguishing between the various mechanisms through which crime may have an effect on economic growth and by exploring the sensitivity of the growth– crime relationship to changing economic conditions in an attempt to identify possible asymmetric effects. Using panel data from 26 countries covering the period 1995–2009, we find that the effect of crime on growth is indeed asymmetric. The growth–crime relationship is found to be strongly negative in bad times, when market sentiment is pessimistic and thus uncertainty is high, employment is low and the strain on public-sector resources through public-safety spending is already sizable, and insignificant in good times. In this respect, our results provide an explanation for the inconclusive empirical evidence regarding the strength of the growth–crime relationship when using reducedform models. The rest of the paper is organized as follows. In Section 2.1 we analyse a simple structural growth model, which serves as a guide for the empirical specification and a reference point for the interpretation of the empirical results, while in Section 2.2 we describe the empirical specification. Section 3 describes the data and presents the estimation results. Section 4 contains concluding comments.

287

where A is a technology variable (assumed exogenous), gp* = (Gp/Y) represents productive public-sector spending, measured by the share of the corresponding government expenditures in GDP, and θ measures the return from such spending. RL and RK are labour-productivity- and capital-productivity-reducing factors, potentially related to the crime rate cr, to the extent that a high-crime environment is likely to reduce workers' incentives to accumulate knowledge and enhance skills as well as firms' incentives to engage in innovative entrepreneurial activities. cr is defined as the number of crime incidents, CR, to total population N, while—β and—δ reflect the potentially negative return to output arising from the adverse impact of crime on private-input factors' productivity. Denoting by y = (Y / N) and k = (K / N) per capita output and per-capita capital respectively, output supplied can be expressed in per capita terms as: yðt Þ ¼ Agp θ cr −γ lp kðt Þð1−aÞ with γ ¼ βa þ δð1−aÞ≥0 a

ð1bÞ

where lp = (L/N) is the labour-force participation rate.2 To the extent that in a high-crime environment some individuals are likely to perceive that they can make a living by engaging in crime-related activities while others are likely to be reluctant to accept late-night jobs or activities and locations associated with high crime-victimization rates, lp may fall as cr rises. Thus, lp ¼ ½1−ϕðcrÞ with ϕ0 ≥0:

ð1cÞ

At the same time, total government spending as percent of GDP, g*, is the sum of productive expenditures, gp*, and non-productive expenditures, gnp*,which include expenditures on crime prevention & lawenforcement that are likely to be positively related to the level of crime activity cr 3: g ¼ g p  þg np  g np  ¼ qðcr Þ;

ð1dÞ

q0 ≥0:

ð1eÞ

2. Growth and crime

Accordingly, on the supply side, combining (1b) with (1c)–(1e), percapita output is given as:

2.1. A simple structural model

y ðt Þ ¼ A ½ g  −q ðcrÞ θ cr−γ ½1−ϕðcr Þa k ðt Þ

Insights into how growth may be related to crime can be obtained by examining a simple growth model, with two private input factors, labour, L, and capital, K, along the lines suggested by Agénor (2008, 2010), Barro (1990), Bayraktar and Moreno-Dodson (2010), and Blankenau et al. (2007). In particular, resources claimed by the government can be put into productive uses, such as education, health and infrastructure, which enter into the production function by having the potential to improve the quality of all private input factors, and into non-productive uses, such as expenditures on crime prevention and law enforcement, which do not enter into the production function. Thus, assuming a constant-returns-to-scale technology with respect to L and K, output produced, Y, can be taken to be given by (1a) 1:

On the demand side, in the absence of unexpected events, y(t) is the sum of planned private consumption c(t), total planned private investment i(t), and overall government spending g(t), all defined in per capita terms (i.e. c = C/N, g = G/N, i = I/N):

Y ðt Þ ¼ Ag p θ ðRL Lðt ÞÞa ðRK K ðt ÞÞð1−aÞ with RL ¼ cr −β ; RK ¼ cr −δ ; cr ¼ ðCR=NÞ;

1

θN0; β; δ ≥0

ð1−aÞ

:

ð2Þ

yðt Þ ¼ cðt Þ þ iðt Þ þ gðt Þ:

ð3aÞ

The excess of households' income over consumption, y(t) − c(t), equals private savings, s(t), plus tax payments τ(t), while planned private investment consists of replacement investment and net additions to the (per capita) capital stock, i.e. iðt Þ ¼ ðn þ δÞkðt Þþ k ðt Þ, where δ is the rate of capital depreciation, n = (dN/dt)(1/N) is the rate of population growth (assumed exogenous), and k ðt Þ≡dk=dt. Assuming further 



ð1aÞ

Following much of the recent growth literature, we model productive government spending as a flow variable. Alternatively, it could be specified as a stock variable, in which case gp in (1a) would correspond to e.g. public investment as percent of GDP and a publiccapital accumulation function would have to be added. This would complicate the model, while there would be little difference as far as steady-states were concerned (see e.g. Futagami et al., 1993).

2 We abstract from equilibrium unemployment resulting from wage bargaining or other labour-market frictions. 3 To the extent that the size of government, as measured by the share of overall public spending in GDP, reflects socio-economic considerations and elements related to the decision-making process at the political level, g* is treated as a policy variable, and so it is specified as time-invariant. Over time the government sets g to grow at the same rate as y, so g* is constant.

288

E. Goulas, A. Zervoyianni / Economic Modelling 49 (2015) 286–295

that agents save a proportion sy of their after-tax income y(t) − τ(t), the equilibrium condition in the goods markets can be expressed as 4:

where




d ln yðt Þ ¼

 sy ð1−τÞ þ τ  −g yðt Þ ¼ ðn þ δÞkðt Þþ k ðt Þ 

ð3bÞ

where both overall public spending, g*, and government revenue, τ*, are scaled in terms of (per capita) GDP. With no public debt,5 increases in overall public spending to GDP must be financed by higher government revenue, so that g ¼ τ  :

ð3cÞ

At the same time, a high-crime environment may reduce private savings, by undermining the security of property rights and by contributing to negative market sentiments and a general perception of uncertainty regarding the proceeds from savings. Thus, letting π be the probability that the return to savings will be insecure and taking π to be potentially related to the crime rate, cr, we can write: sy ¼ σ y ð1−πðcr ÞÞ with σ y N 0; 0 ≤ π b 1; π 0 ≥ 0

ð3dÞ

From (3b)–(3d) and (2), it follows that the rate of capital accumula

tion γ k ðt Þ≡ kkððtt ÞÞ will be given as: γ k ðt Þ ¼

σ y ð1−τÞ½1−πðcr ÞA½τ  −qðcr Þθ cr−γ ½1−ϕðcrÞa −ðn þ δÞ ð4Þ kðt Þa 0 0 0 with θ; a N 0; π ; q ; ϕ ; γ≥0:

Along the balanced-growth path γk(t) = 0. Imposing this condition and using the resulting expression in (4) to substitute out k from (2), steady-state (per-capita) output, y∞ is given as: " y∞ ¼

# A1=a ½τ  −qðcr Þθ=a cr −γ=a ½1−ϕðcr Þ  ð1−aÞ=a

ðn þ δÞ

σ y ð1−τÞ½1−πðcr Þ

ð1−aÞ=a

:

ð5aÞ Outside steady states, the path of (per capita) output is determined by the path of k. Letting ψ(t)be the rate at which the (log of) percapita capital, ln k(t), approaches (the log of) its steady-state value, ln k∞, and denoting by ln yo an initial steady-state (per-capita) output, then outside steady states we can write as an approximation (see e.g. Bassanini and Scarpetta, 2001; Mankiw et al., 1992): lnyðt Þ− ln y∞ ¼ e−ð1−aÞψðt Þ ð ln yo − ln y∞ Þ

ð5bÞ

or   lnyðt Þ− ln yo ¼ 1−e−ð1−aÞψðt Þ ð ln y∞ − ln yo Þ:

ð5cÞ

Upon substitution into (5c) of a linearized version of (5a) by taking derivatives, one can derive an output-growth equation of the form given by (6):   d ln yðt Þ ¼ −υðt Þ ln y þ υ ðt Þ F A; σ y ; n þ δ; τ ; cr

ð6Þ

4 Note that since the excess of households' income over consumption  equals private saving plus tax payments, we can write (3a) as sy(y − τ) + τ = i + g or sy y− τy y þ τy y ¼ i þ gy y. Then, denoting by τ* = τ/y the ratio of tax revenues to GDP and substituting out i, we obtain (3b). 5 The no-public debt assumption can easily be dropped without causing any substantive change in the results as long as debt sustainability is assumed. At the same time, for a number of countries in our sample, the option of debt-financed increases in government expenditures is severely constrained through the Maastricht-Treaty rules, or through national laws (e.g. the ‘golden rule’ in the UK, see Chote et al., 2009).

  1−e−ð1−aÞψðt Þ ln yðt Þ− lny0 N0 ; υðt Þ ¼ t t

and ! 1 a1 a1 1 θ ; Fσy ¼ ; F τ ¼ − − a1 ; ; F nþδ ¼ −  aA σy ðn þ δÞ ð1−τ Þ gp  ð1−aÞ !  1 1 θ1 ; F cr2 ¼ −π 0 ¼ F cr1 þ F cr2 þ F cr3 þ F cr4 ; F cr1 ¼ −q0  a1 ; τ −g np  1−π

FA ¼ F cr

F cr3 ¼ −ϕ0 γ1 ¼

  1 ð1−aÞ θ N 0; θ1 ¼ ; F cr4 ¼ −γ 1; a1 ¼ N 0; lp a a

β þ δα 1 ≥0; q0 ; π 0 ; ϕ0 ≥0: cr

FA and F σ y , the partial derivatives of output growth with respect to technology and the saving rate respectively, are positive, while Fn + δ has a negative sign. Fτ⁎, the partial derivative of per-capita growth with respect to increases in tax revenue, will also have a negative sign if the adverse impact on capital accumulation of the fall in savings, due to the tax-induced reduction in disposable income, is large enough to outweigh any output-increasing effect resulting from the higher level of productive public-spending which the additional government revenue can finance. Fcr, the partial derivative of output growth with respect to the crime rate, reflects the four key channels through which the determinants of the growth–crime relationship operate. Firstly, for an unchanged level of government revenue τ*, increased crime leading to higher public-safety expenditures will divert resources away from productive uses, causing a fall in steady-state capital and lowering output growth between steady states (F cr1 b 0). Other things equal, the growth reduction will be larger the greater is the existing strain on public-sector resources (i.e. the larger is the initial level of non-productive government spending gnp*, including crime-related expenditures, relative to the overall government revenue τ*) and the greater is the opportunity cost of allocating public-sector resources to crime prevention & law enforcement (i.e. the higher is the return to productive public expenditures, θ). Secondly, a higher level of crime activity, by contributing to a general perception of instability and negative market sentiments regarding the security of the return to saving, may cause a fall in growth through a reduced incentive to save (F cr2 b 0). The higher is the existing level of uncertainty or negative market sentiments, as measured by the size of π, the more insecure will be the perceived proceeds from savings following a rise in crime, and thus the larger the absolute magnitude of F cr2 . Thirdly, increased crime is likely to induce individuals to devote a smaller fraction of their time to work, causing a reduction in output growth for an unchanged (per-capita) stock of physical-capital ( F cr3 b 0 ). This growth-reducing effect will be more pronounced the greater is the existing shortage of labour resources in the economy and thus the smaller is the initial level of employment relative to overall population (i.e. the lower is lp in F cr3 ). Fourthly, there may be an additional unfavourable influence, captured by F cr4 b 0, arising from the productivity-reducing effect of crime, via disincentives to built up knowledge and produce innovative entrepreneurial ideas. Accordingly, what (6) implies is that the strength of the growth– crime relationship may not be independent of the state of the economy. If the state of the economy is not particularly satisfactory, so that the strain on public-sector resources is already sizable, market pessimism and uncertainty is already high and employment is low, one can expect the terms (τ * − gnp*), (1 − π) and lp to be relatively small, in which case the partial derivatives F cr1 , F cr2 and F cr3 will be large in absolute magnitude and the overall effect of crime on growth, measured by F cr ¼ F cr1 þ F cr2 þ F cr3 þ F cr4 , will be strongly negative. In the opposite case, i.e. when economic conditions are satisfactory enough to imply relatively large magnitudes of (τ * − gnp*), (1 − π) and (lp), the partial

E. Goulas, A. Zervoyianni / Economic Modelling 49 (2015) 286–295

289

Table 1 Descriptive statistics. Variable

Obs.

Mean

Std. dev.

Min

Max

GDP per capita growth Savings rate Government revenue to GDP ratio Human capital Crime rate Public-order & safety spending to GDP ratio Annual change in the ESI Employment in industry to population ratio

285 285 285 285 285 285 285 285

0.0236 0.2294 0.4281 0.6043 2.1677 0.0172 −0.0130 0.1439

0.0342 0.0641 0.0800 0.1306 2.4083 0.0042 0.0735 0.0236

−0.1427 0.0919 0.2791 0.2658 0.4000 0.0062 −0.2332 0.0953

0.1207 0.4175 0.5853 0.9507 15.1000 0.0293 0.1728 0.2129

Notes: The sample consists of 26 countries over the period 1995–2009. All variables are expressed in percentage terms except for the ESI and the crime rate, which is defined as intentional homicides per 100,000 persons.

derivatives F cr1 , F cr2 and F cr3 may well be small in size, in which case higher crime will operate mainly through the productivity-reducing effect F cr4 , thus causing a much more limited overall response of growth. 2.2. Empirical specification Given (6), the empirical specification we use corresponds to the following model: ðgrowthÞ j;t ¼ δ1 ln ðyÞ j;t−1 þ δ2 ðsaving Þ j;t þ δ3 ðrevenueÞ j;t þ δ4 ðhumanÞ j;t þ δ5 ðcrimeÞ j;t þδ6 ðcrime  pessimismÞ j;t þ þδ7 ðcrime  low employmentÞ j;t 2009 X τ t ðyearÞ þ μ j þ ε j;t :

ð7Þ

þδ8 ðcrime  high spendingÞ j;t þ

t¼1995

The δi 0 s; τi 0 s are unknown constant parameters to be estimated, μj represents unobserved country-fixed effects and ε is an unobserved spherical disturbance term. The dependent variable, (growth)j,t, is real (per-capita) output growth, while ln(y)j,t − 1, the lagged value of (the logarithm of) GDP per capita, will enter the regression with a negative coefficient δ1 if conditional convergence applies.6 (saving)j,t and (revenue)j,t represent, respectively, the savings rate and the government revenue-to-GDP ratio. Thus, we expect δ2 N 0, δb3N0. In line with much of the empirical growth literature, human capital, (human)j,t, is also added as a separate explanatory variable, with the corresponding coefficient δ4 expected to be positive. We further include time dummies to control for world-wide growth of technology, as well as for other common shocks across countries that might have taken place during the period under consideration, such as monetary-policy changes, including the circulation of the euro. (crime) j,t is the crime-activity proxy, while (crime ∗ pessimism)j,t, (crime ∗ low employment)j,t and (crime ∗ high spending)j,t are interaction terms representing, respectively, the case of negative market sentiment and thus higher-thanaverage uncertainty regarding the proceeds from saving, low employment and high-strain on public-sector resources. In particular, to examine possible asymmetries in the growth–crime relationship, we proceed by constructing three dummies,7 defined as follows: i) (pessimism)j,t, takes the value of 1 when the annual change of an economic sentiment indicator of a country j in year t is negative, implying that market sentiment and thus uncertainty in the current period is deteriorating relative to the previous period; ii) (low employment)j,t, equals the value of 1 when the employment (in industry) to population ratio of a country j in year t is below the median value obtained from the distribution of all countries; and iii) (high spending)j,t, attains the value of 1 when the public-order & safety spending to GDP ratio of a country j in year t is above the median value obtained from the distribution of all countries. 6 According to the conditional convergence hypothesis, when macroeconomic policies and other key characteristics across countries and over time are accounted for, low/high levels of income per capita are associated with higher/lower growth rates in subsequent years. 7 For studies measuring asymmetries in the same way (see e.g. Drakos and Goulas, 2006; Drakos and Kallandranis, 2007; Guariglia et al., 2013).

Taking into account the key channels through which the growth– crime relationship may operate noted in Sections 1 and 2.1, we anticipate the coefficients on the three interaction terms in (7) to be negative. Thus, under a switch to market sentiments of pessimism, low employment and high public-order & safety spending, the overall responsiveness of growth to increased crime will be given by the sum δ5 + δ6 + δ7 + δ8. Rejecting the joint hypothesis Ho : δ6 = δ7 = δ8 = 0 in favour of the alternative that at least one parameter is significantly negative would provide evidence of asymmetries in the growth–crime relationship, depending on current economic conditions. At the same time, rejecting the hypothesis Ho : δ5 + δ6 + δ7 + δ8 = 0 in favour of the alternative that the sum of these four coefficients is strongly negative would imply that, under the operation of all four channels, increased crime can severely constrain economic growth.8

3. Description of the data and estimation results We employ a dataset covering 26 countries9 over the period 1995– 2009. Following a common practice in the literature (Càrdenas, 2007; Detotto and Otranto, 2010; Peri, 2004; World Bank, 2006), we proxy crime activity using data on intentional homicides per 100,000 population.10 The data are obtained from UNODC (United Nations Office on Drugs and Crime). This crime proxy, in addition to being less likely to be subjected to under-reporting bias in a cross-country context, also possesses the advantage of being the most reliable among other UN crime indicators as it has the same definition throughout the period under consideration.11 Data from the World Bank (World Development Indicators) are used to construct real (per capita) output growth (annual percentage change of GDP per capita, US$ constant (2005, PPP) prices), the savings variable, which is proxied by the share of gross domestic saving to GDP, and the human-capital variable, proxied by the ratio of gross tertiary enrolment 8 Population growth has not been included as an explanatory variable in (7) as the corresponding series for the sample we consider shows little variation across countries and over time. 9 Australia, Austria, Belgium, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Israel, Italy, Japan, Lithuania, Netherlands, Norway, Poland, Portugal, Romania, Slovak Republic, Spain, Sweden, Switzerland, United Kingdom, and United States. The country sample follows from data availability for all variables and the objective to have economies with different per-capita income levels but not very dissimilar institutional characteristics. 10 Data on other types of crime based on national crime statistics are not sufficiently reliable for cross-country analysis as nations often differ in how they classify different offences, while the extent of under-reporting of sub-categories of crime other than homicide varies considerably across countries (Heiskanen, 2010). Different levels of under-reporting and/or misclassification of certain crime types is a serious problem in cross-country empirical analyses as it can lead to biased estimates regarding the overall impact of crime. This has led many researchers to rely on homicides as a proxy for crime activity in cross-country or cross-region crime studies, as homicide series are in general less likely to suffer from under-reporting bias (see e.g. the references cited in the text). 11 UNODC homicide data are based on victimization surveys and so are more suitable for cross-country empirical analysis than data based on national crime & criminal justice statistics. Complete annual series for other types of crime covering the period 1995-2009 are not available in UNODC.

290

E. Goulas, A. Zervoyianni / Economic Modelling 49 (2015) 286–295

(b) Public-order & safety spending (% GDP)

50

Density

0

0

5

Density

10

100

15

(a) Employment to population (%)

.1

.15

.2

.25

.005

kernel = epanechnikov, bandwidth = 0.0068

.01

.015

.02

.025

.03

kernel = epanechnikov, bandwidth = 0.0010

(c)

4 0

2

Density

6

8

Annual change in the ESI

-.3

-.2

-.1

0

.1

.2

kernel = epanechnikov, bandwidth = 0.0138

Graph 1. (a)–(c). Kernel densities.

to the population of the corresponding age group. Data on (general) government revenue come from the IMF (World Economic Outlook), which includes tax receipts, social security contributions, grants receivable, and other forms of revenue. Series on public-order & safety spending are constructed from IMF data (Government Financial Statistics database, GFS), while the employment (in industry) to population ratio is calculated using data from the International Labour Organization (Key Indicators of the Labour Market, KILM). Finally, data on the (seasonally-adjusted) Economic Sentiment Indicator (ESI) provided by Eurostat (Business & Consumer Surveys, Economic and Financial Affairs of the EU) are employed to construct a proxy for changing market-sentiments and thus perceived uncertainty regarding the proceeds from savings. The ESI is a surveybased composite expectations index, reflecting opinions regarding the state of the economy, and negative (positive) changes in this index can be taken to represent pessimistic (optimistic) expectations regarding future macroeconomic conditions.12 This measure of perceived changes in economic conditions can be considered preferable to other indicators to the extent that it is highly forward-looking by being based on surveys of market opinions.13 Table 1 reports the statistical properties of the dataset, while Graph 1(a)–(c) show Kernel densities for the three variables upon which the construction of the asymmetric terms is based (i.e. changes in the ESI, employment-to-population ratio and public order & safety spending to GDP). The graphs suggest that whatever asymmetries are

12 For Australia, Ireland, Israel, Japan, Norway, Switzerland and the USA, we resort to the (amplitude-adjusted) Composite Leading Indicator (CLI) obtained from the monthly indicators of the OECD database. 13 For alternative measures of perceived changes in economic conditions in a growthcrime context (see e.g. Goulas and Zervoyianni, 2013).

Table 2 Mean values by country. Country

Crime ratea

Economic sentiment indicatorb

Employment to population ratio

Public-order & safety spending to GDP

Australia Austria Belgium Denmark Estonia Finland France Germany Greece Hungary Ireland Israel Italy Japan Lithuania Netherlands Norway Poland Portugal Romania Slovak Republic Spain Sweden Switzerland United Kingdom United States

1.4273 0.7643 2.1636 0.9538 9.6786 2.5429 0.7923 1.4500 1.0417 2.0538 1.1857 2.7100 1.2357 0.5375 8.1000 1.1571 0.8100 1.5556 1.2462 2.6833 1.9143 0.9300 1.0769 0.9286 1.5714 5.5111

0.9996 0.9983 1.0007 1.0037 1.0131 1.0135 1.0363 0.9919 1.0530 1.0111 1.0006 1.0002 1.0169 1.0022 1.0328 1.0160 0.9999 0.9676 1.0043 1.0700 1.0212 1.0447 1.0202 0.9987 1.0214 0.9943

0.1283 0.1624 0.1218 0.1500 0.1782 0.1446 0.1224 0.1796 0.1009 0.1496 0.1640 0.1141 0.1370 0.1669 0.1456 0.1192 0.1330 0.1369 0.1830 0.1537 0.1968 0.1470 0.1349 0.1604 0.1366 0.1229

0.0163 0.0151 0.0167 0.0101 0.0249 0.0136 0.0152 0.0164 0.0114 0.0204 0.0166 0.0172 0.0194 0.0134 0.0191 0.0180 0.0101 0.0190 0.0185 0.0231 0.0199 0.0187 0.0135 0.0164 0.0237 0.0211

Numbers denote mean values from 1995 to 2009. a Defined as intentional homicides per 100,000 persons. b For Australia, Ireland, Israel, Japan, Norway, Switzerland and the USA, we use the (amplitude-adjusted) Composite Leading Indicator obtained from the OECD database.

E. Goulas, A. Zervoyianni / Economic Modelling 49 (2015) 286–295

sentiments from 1995 until 2008, corresponding to short-run fluctuations in economic activity, but deteriorating market sentiments from 2008 onwards. There is also evidence of a falling labour–force–participation ratio after 2007. At the same time, Graph 2(e) indicates that public-order & safety-spending as percent of GDP has significantly increased in the last few years. Eq. (7) has been estimated by applying the system-GMM technique (Arellano and Bover, 1995; Blundell and Bond, 1998, 2000). This technique is extensively used in panel-data growth studies to allow for unobserved panel heterogeneity and simultaneously control for endogeneity bias arising from the possibility that one or more of the explanatory variables in growth regressions may not be strictly exogenous (see e.g. Aisen and Veiga, 2013; Bond et al., 2001; Christiansen et al., 2013; Guariglia and Poncet, 2008; Hoeffler, 2002; Rooth and Stenberg, 2012; Saidi and Aloui, 2010; Yamarik, 2010). Indeed, a common feature of most empirical growth models is that causation between the dependent and the right-hand-side variables may run in both directions, leading to endogeneity bias. In the system-GMM, suitably lagged levels and lagged first-differences of right-hand-side variables are used as instruments, ensuring that the estimates reflect causation running from the right-hand-side variables to the dependent variables and not vice versa. In our growth regressions, given that most of the explanatory

obtained will not be the result of few outlying observations. On the other hand, country averages over the period 1995–2009 for the ESI, the employment-to-population ratio and public-order & safety spending to GDP are reported in Table 2, along with corresponding countryaverages for crime rates. As is evident from Table 2, cross-country crime varies, with the US showing the highest level of crime activity among the non-EU countries and Estonia and Lithuania showing the highest crime rates among the EU member states. Further, Graph 2(a) shows the evolution over time of the mean crime rate for the countries in our sample, while Graph 2(b) shows the corresponding series for mean GDP per capita. Crime has fallen on average since 1995, with the sharpest decline occurring in 1999, although the declining trend has been reversed between 2000 and 2001 and from 2008 onwards. From Graph 2(a) and (b), there is evidence of a roughly negative relationship between increases in mean per capita GDP and mean crime, with periods of falling crime appearing to correspond on average to periods of rising GDP per capita and vice versa. Graph 2(c) shows mean changes in the ESI by year for the countries in our sample during the period 1995–2009, while Graph 2(d) and (e) show, respectively, the evolution over time of the mean employment-to-population ratio and the mean public-order & safety-spending to GDP. Graph 2(c) indicates volatile market

2.4 2.2 2 1.8

2000

2005

2010

22000 24000 26000 28000 30000

(b) GDP per capita (in US dollars)

2.6

(a) Crime rate (per 100,000 population)

1995

2000

2005

(d) Employment to population (%)

2010

.15

(c) Economic Sentiment Indicator

.85

.13

.9

.135

.95

.14

1

.145

1.05

1.1

1995

291

2000

2005

2010

1995

2000

(e)

.017

.018

.019

Public-order & safety spending (% GDP)

.016

1995

1995

2000

2005

Graph 2. (a)–(e). Mean values by year.

2010

2005

2010

292

E. Goulas, A. Zervoyianni / Economic Modelling 49 (2015) 286–295

Table 3 System-GMM estimates of the growth model. Dependent variable (growth)j,t. Regressor

Model 1

Model 2

Model 3

Model 4

Model 5

ln(y)j,t − 1

−0.0735⁎⁎⁎ (−3.94) 0.4663⁎⁎⁎ (3.65) −0.3468⁎⁎⁎ (−2.62) 0.1411⁎⁎⁎

−0.0725⁎⁎⁎ (−4.03) 0.4709⁎⁎⁎ (3.68) −0.3478⁎⁎⁎ (−2.75) 0.1403⁎⁎⁎

−0.0679⁎⁎⁎ (−3.34) 0.4129⁎⁎⁎ (3.19) −0.3451⁎⁎⁎ (−2.72) 0.1383⁎⁎⁎

−0.0767⁎⁎⁎ (−3.77) 0.4647⁎⁎⁎ (3.40) −0.3677⁎⁎⁎ (−2.66) 0.1481⁎⁎⁎

−0.0623⁎⁎⁎ (−3.36) 0.3724⁎⁎⁎ (3.10) −0.3213⁎⁎⁎ (−2.70) 0.1192⁎⁎⁎

(crime ∗ pessimism)j,t

(3.99) −0.0041⁎⁎ (−2.28) –

(3.60) −0.0037⁎ (−1.80) –

(3.63) −0.0001 (−0.06) –

(crime ∗ low employment)j,t

–

(4.11) −0.0037⁎⁎ (−2.39) −0.0041⁎⁎⁎ (−4.74) –

–

(crime ∗ high spending)j,t

–

–

−0.0037⁎⁎⁎ (−2.89) –

(3.76) 0.0018 (0.63) −0.0044⁎⁎⁎ (−4.78) −0.0047⁎⁎⁎ (−4.48) −0.0040⁎⁎

Observations m1 m2 Sargan Test

285 −2.48⁎⁎ 0.40 121.62 (p-val. 0.94) – – – – –

285 −2.57⁎⁎ 1.20 144.26 (p-val. 0.52) x2 = 15.94⁎⁎⁎

(saving)j,t (revenue)j,t (human)j,t (crime)j,t

(1) (2) (3) (4) (5)

H0 : δ5 + δ6 = 0 H0 : δ5 + δ7 = 0 H0 : δ5 + δ8 = 0 H0 : δ5 + δ6 + δ7 + δ8 = 0 H0 : δ6 = δ7 = δ8 = 0

– – – –

285 −2.37⁎⁎ 0.27 135.65 (p-val. 0.72) – x2 = 15.52⁎⁎⁎ – – –

−0.0044⁎⁎ (−2.17) 285 −2.43⁎⁎ 0.99 121.63 (p-val. 0.93) – – x2 = 6.56⁎⁎ – –

(−2.05) 285 −2.49⁎⁎ 0.97 161.03 (p-val. 0.16) – – – x2 = 23.76⁎⁎⁎ x2 = 41.01⁎⁎⁎

Numbers in parentheses denote z-scores, m1 and m2 are residual first and second order serial correlation tests, while Sargan stands for the over-identifying restrictions test. One, two, and three asterisks denote significance at the 10, 5, and 1 percent levels respectively. All models allow for robust standard errors. Time dummies are included in all specifications.

variables, including the crime rate, (crime), may in principle be affected by per-capita output growth, all right-hand-side variables, except the time dummies, have been treated as potentially endogenous and have been accordingly instrumented. The statistical adequacy of the model is established when the generated residuals do not exhibit secondorder autocorrelation and the over-identifying restrictions are not rejected. Estimation results for the entire sample are shown in Table 3 below. Model (1) shows estimates without controlling for influences arising from the state of the economy, while Models (2)–(4) report estimates after controlling for such influences. In all models, the estimates show a statistically significant positive effect on growth of savings and human capital and a negative effect of higher government revenue, consistent with the results of other studies (Afonso and Furceri, 2010; Barro and Redlick, 2011; Bassanini and Scarpetta, 2001; Gemmell et al., 2011; Morgese-Borys et al., 2008). The coefficient on lagged per-capita GDP is also negative and significant, indicating conditional convergence for the set of countries and time period considered. Moreover, in all columns, the Sargan test of overidentifying restrictions confirms the joint validity of the instruments used, indicating that the model is well specified. The hypothesis of no second-order serial correlation is also not rejected. As far as the effect of crime is concerned, Model (1) indicates a negative, although not particularly strong, relationship between per capita output growth and crime activity, with the coefficient on (crime) being significant only at the 5% level. The growth effect of crime is also small compared to the other determinants of growth in Table 3, with the estimated coefficient in Model (1) implying that a 1 point increase in crime activity leads to about 0.004 point reduction in per-capita economic growth. This small crime-effect is consistent with much of the empirical literature based on reduced form models, which suggests a modest or statistically insignificant negative effect of crime on growth (Burnham et al., 2004; Chatterjee and Ray, 2009; Detotto and Otranto, 2010; Mauro and Carmeci, 2007). However, reduced-form models cannot shed light on the different channels via which crime impacts on growth and on the extent to which the strength of these different channels is influenced by the state of the economy. Indeed, the question to be asked is to what extent the growth–crime relationship is sensitive to

changing economic conditions. Is the growth–crime relationship asymmetric as Section 2.1 seems to suggest, being strongly negative in bad times and weakly negative or statistically insignificant in good times? To what extent are the various channels through which crime can constraint growth influenced by the prevailing economic conditions? To address this issue, in Models (2)–(4) we examine the occurrence of asymmetric effects regarding the growth–crime relationship stemming from different assumptions as to: i) market sentiments regarding future macroeconomic conditions and thus the degree of uncertainty regarding the proceeds from saving, ii) the employment-to-population ratio in the economy, and iii) the strain on public-sector resources. We thus augment Model (1) by the interaction terms (crime ∗ pessimism), (crime ∗ low employment) and (crime ∗ high spending). In Model (2), where the influence of changing market sentiments is accounted for, the coefficient on the interaction term δ6 is negative and highly significant suggesting that the growth–crime relationship depends on the degree of pessimism regarding future economic conditions. In particular, under a switch to more pessimistic expectations regarding the state of the economy, and thus the proceeds from saving, the overall effect of crime on growth is given by the sum δ5 + δ6 and we emphatically reject the hypothesis that the sum of these coefficients equals zero (Hypothesis 1). On the other hand, under a switch to more optimistic expectations, the overall effect of crime is given by the coefficient on (crime), δ5. Determining the relative magnitude of the two coefficients, our results indicate that the growth reduction due to crime conditional on pessimistic expectations is 47% higher than the corresponding reduction conditional on optimistic expectations. Thus, pessimistic market sentiments provide a significant amplification mechanism for the adverse effect of crime on growth through the riskiness of savings. In Model (3), where the effect of low employment is accounted for, both the coefficient on (crime) and the coefficient on the interaction term (crime ∗ low employment) are statistically significant, giving an overall growth-effect of crime of δ5 + δ7 (= − 0.0074), with the hypothesis that the sum of these coefficients equals zero again being strongly rejected (Hypothesis 2). On the other hand, when the employment-to-population ratio in the economy is higher than average, the overall effect of crime on growth is given by the coefficient on

E. Goulas, A. Zervoyianni / Economic Modelling 49 (2015) 286–295

293

Table 4 System-GMM estimates of the growth model for EU countries. Dependent variable (growth)j,t. Regressor

Model 1

Model 2

Model 3

Model 4

Model 5

ln(y)j,t − 1

−0.0791⁎⁎⁎ (−2.65) 0.4659⁎⁎⁎ (3.44) −0.3164⁎⁎ (−2.58) 0.1190⁎⁎⁎

−0.0734⁎⁎ (−2.45) 0.4637⁎⁎⁎ (3.64) −0.3150⁎⁎ (−2.48) 0.1170⁎⁎⁎

−0.0823⁎⁎ (−2.57) 0.4206⁎⁎⁎ (3.13) −0.2856⁎⁎ (−2.45) 0.1117⁎⁎⁎

−0.0806⁎⁎⁎ (−2.60) 0.4515⁎⁎⁎ (3.13) −0.3373⁎⁎⁎ (−2.60) 0.1190⁎⁎⁎

−0.0756⁎⁎ (−2.48) 0.3998⁎⁎⁎ (3.12) −0.2769⁎⁎ (−2.32) 0.0999⁎⁎⁎

(crime ∗ pessimism)j,t

(3.94) −0.0046⁎⁎ (−2.27) –

(3.60) −0.0045⁎ (−1.86) –

(3.48) −0.0001 (−0.03) –

(crime ∗ low employment)j,t

–

(4.31) −0.0036⁎ (−1.93) −0.0050⁎⁎⁎ (−8.91) –

–

(crime ∗ high spending)j,t

–

–

−0.0047⁎⁎⁎ (−4.50) –

Observationsa m1 m2 Sargan Test

230 −2.49⁎⁎ 0.76 127.25 (p-val. 0.87) – – – – –

230 −2.18⁎⁎ 1.09 161.13 (p-val. 0.18) x2 = 20.83⁎⁎⁎

(3.79) 0.00004 (0.01) −0.0050⁎⁎⁎ (−8.11) −0.0049⁎⁎⁎ (−4.24) −0.0030 (−1.43) 230 −2.15⁎⁎ 0.90 158.56 (p-val. 0.19) – – – x2 = 25.09⁎⁎⁎ x2 = 78.58⁎⁎⁎

(saving)j,t (revenue)j,t (human)j,t (crime)j,t

(1) (2) (3) (4) (5)

H0 : δ5 + δ6 = 0 H0 : δ5 + δ7 = 0 H0 : δ5 + δ8 = 0 H0 : δ5 + δ6 + δ7 + δ8 = 0 H0 : δ6 = δ7 = δ8 = 0

– – – –

230 −2.41⁎⁎ 0.18 138.53 (p-val. 0.65) – x2 = 11.86⁎⁎⁎ – – –

−0.0048⁎ (−1.95) 230 −2.22⁎⁎ 0.94 132.67 (p-val. 0.77) – – x2 = 6.06⁎⁎ – –

Numbers in parentheses denote z-scores, m1 and m2 are residual first and second order serial correlation tests, while Sargan stands for the over-identifying restrictions test. One, two, and three asterisks denote significance at the 10, 5, and 1 percent levels respectively. All models allow for robust standard errors. Time dummies are included in all specifications. a The number of the EU countries included in the sample is 20.

(crime). This coefficient is significant only at the 10% level and equals − 0.0037. The growth-reducing effect of crime conditional on low employment is therefore twice as large as the corresponding effect conditional on high employment, indicating that there is an asymmetric reaction of economic growth to crime depending on the use of labour resources in the economy. In Model (4), where public-sector-resource strain is taken into account, the coefficient on (crime), while it loses its significance at standard levels, still has a negative sign. Moving to the interaction term, we document a negative and significant coefficient at the 5% level, rejecting the hypothesis that the sum of coefficients δ5 + δ8 (= − 0.0045) equals zero (Hypothesis 3). Thus, when public-safety spending is higher than average and the strain on public-sector resources due to crime is significant, the overall effect of crime on growth is 44 times larger, signifying another asymmetric source through which crime could be particularly harmful to growth. Finally, in Model (5) we present estimation results after controlling for all sources of potential asymmetry, i.e. all three interaction terms, (crime ∗ pessimism), (crime ∗ low employment) and (crime ∗ high spending), reflecting unfavourable economic conditions, are included. The coefficients on the interaction terms are all statistically significant implying that in bad times, when all sources of asymmetry are in operation, the effect of crime on growth is highly significant and the growth reduction due to crime is much amplified. In particular, assuming a switch to market sentiments of pessimism and thus higher uncertainty, a low level of employment and a high level of public-order & safety expenditures relative to GDP, the overall growth-effect of crime amounts to δ5 + δ6 + δ7 + δ8, which implies that a 1 point increase in crime activity can lead to up to 0.011 point reduction in economic growth. Accordingly, in bad times and under the operation of all asymmetric factors, the growth reduction due to crime can be large. Here, we emphatically reject the hypothesis that the sum of coefficients δ5 + δ6 + δ7 + δ8 equals zero (Hypothesis 4) as well as the hypothesis H0 : δ6 = δ7 = δ8 = 0, i.e. that jointly the coefficients of the asymmetric terms equal zero (Hypothesis 5). In contrast, in good times, i.e. under the assumption of optimistic market sentiments, high employment and low public-order & safety expenditures to GDP, the overall growth effect of crime is positive but

statistically insignificant at conventional levels, given by the coefficient δ5 (=0.0018). Confining our sample to European countries does not change these findings. This is shown in Table 4, where estimation results for only EU member states are reported. Model (1) still indicates a significant, although not particularly large, negative effect of crime on growth. In Models (2)–(4), the interaction terms always enter with negative signs and are statistically significant at conventional levels, confirming the existence of asymmetries in the growth–crime relationship even in this sub-sample. In Model (5), the coefficients of all three interaction terms have negative signs, with the first two being significant at 1% and the third being only marginally insignificant at conventional levels.Moreover, like in the corresponding model of Table 3, in Model (5) of Table 4 we reject the hypothesis that the coefficient sum δ5 + δ6 + δ7 + δ8 equals zero (Hypothesis 4) as well as the hypothesis H0 : δ6 = δ7 = δ8 = 0, i.e. that jointly the coefficients of the asymmetric terms equal zero (Hypothesis 5).14 Overall, the results in Tables 3 and 4 suggest that the growth effect of crime is strongly asymmetric in that the state of the economy matters. Crime reduces growth mainly in times of unfavourable economic conditions, when expectations regarding the future state of the economy worsen, aggregate employment is low and the strain on public-sector resources resulting from public-safety spending is already sizable. Crime does not seem to be harmful to growth when economic conditions are sufficiently satisfactory, and thus when the growth–crime relationship operates mainly through private-input factors' productivity. Indeed, the asymmetric nature of the growth–crime relationship can 14 One could in principle look at other sub-samples, such as the G7 sub-sample. However, restricting the estimation to the G7 group would imply a dramatically reduced sample size (sample size would be reduced by almost 70%, from 284 observation to less than 90 observations). Given the number of explanatory variables and the degrees of freedom in a system-GMM context, the statistical power of the results would be much reduced. Moreover, restricting the estimation to the G7 group would make the country-dimension of the panel (the i's) smaller than the time-dimension (the t's). In panels with the i's less than the ts, the system-GMM technique, which is the standard methodology used in much of the recent empirical literature on growth to control for possible endogeneities and reverse causality, is not recommended (see e.g. Roodman 2006 and the papers cited on p.11).

294

E. Goulas, A. Zervoyianni / Economic Modelling 49 (2015) 286–295

explain the inconclusive evidence in the existing literature regarding the strength of the crime–growth relationship when using reducedfrom models. On the other hand, given Section 2.1, an intuitive explanation for the sizable growth-reducing effect of crime in bad times is not difficult to find. Firstly, the opportunity cost of allocating additional resources to crime prevention & law enforcement can be expected to be greater in bad times, when the strain on public-sector resources is already sizable. Secondly, increased crime is likely to induce individuals to devote a smaller fraction of their time to work and the resulting negative growth effect can be expected to be more pronounced the greater is the shortage of labour resources in the economy and thus the smaller is the existing employment-to-population ratio. And, thirdly, a high crime rate contributes to a general perception of instability and bad business climate and the corresponding negative effect on savings, and therefore on growth, can be expected to be more pronounced when economic conditions are already unfavourable and thus the return to saving is already insecure. Our findings have important policy implications. Since the global economic crisis of 2008–2009 and the European debt crisis of 2009– 2011, market pessimism regarding macroeconomic performance, and thus perceived uncertainty regarding the return to savings, has increased in many countries. Also, labour-market performance remains fragile, with many economies currently showing lower employment levels compared to those in previous years. Moreover, the strain on public-sector resources has recently become more pronounced, as many countries have set constraints on overall public expenditures, either as part of area agreements, such as the Stability & Growth Pact in Europe, or in an attempt to avoid rising interest rates on public debt and speculative attacks on their currencies. At the same time, given the recent slowdown in economic activity worldwide, the opportunity cost of engaging in crime-related activities has fallen for a number of individuals who have experienced a reduction in income, so increased crime is a possibility. The combined effect of a pessimistic market environment, lower-than-average employment and higher-than-average strain on public-sector resources may well be a further slowdown in economic growth due to crime in the coming years. 4. Concluding comments Despite the fact that the importance of crime in determining a country's economic progress has long been recognized both among policymakers and in academic circles, the existing empirical evidence on the growth–crime relationship is inconclusive. In particular, while some studies present results suggesting a strong adverse influence of crime on growth, other studies report evidence indicating a weak negative effect or no effect at all. Much of this literature, however, is based on reduced-form models that cannot shed light on the different channels via which crime impacts on growth and the extent to which the strength of these different channels is influenced by the state of the economy. This paper has examined empirically the growth–crime relationship, using a simple structural growth model to identify possible channels through which crime may have an impact on per-capita output growth. We do not find a general strong negative relationship between percapita output growth and crime. This is because, in addition to potentially adverse effects on private-input factors' productivity, increased crime raises the level of insecurity in the economy and this is more likely to reduce growth: i) the higher is the initial level of uncertainty regarding the perceived return to savings, and thus the more pessimistic are market sentiments regarding the future state of the economy; ii) the higher is the existing opportunity cost of financing the required crime-prevention & law-enforcement expenditures; and iii) the smaller is workers' opportunity cost of engaging in crime-related activities and therefore the lower is aggregate employment. Indeed, taking explicitly into account the major channels linking growth to crime, we find evidence suggesting significant potential gains from reducing crime in

‘bad times’, i.e. during periods of worsening economic conditions, when the strain on public-sector resources resulting from already large public-safety expenditures is significant, when the existing employment-to-population ratio is below average and when the state of expectations is getting worse. Under such circumstances, our estimates for the entire sample imply that countries could raise per capita output growth by about one percent per year if they were to reduce crime rates by 10%. The empirical analysis in this paper could be extended in a number of directions. The country sample could be extended to include other non-EU and/or non-OECD economies in order to investigate whether different institutional characteristics among countries, including public-sector institutional quality and labour-market policies and institutions, can change the growth–crime relationship. Uncertaintyoriginated asymmetries in the growth–crime relationship could also be linked to specific variables, such as unemployment and inflation. At the same time, obtaining comparable cross-country annual data on different sub-categories of crime for the period under consideration would enable one to examine the sensitivity of the asymmetric elements of the growth–crime relationship to changes over time in different crime types. Extending the paper in the above directions is among our plans for future work.

References Afonso, A., Furceri, D., 2010. Government size, composition, volatility and economic growth. Eur. J. Polit. Econ. 26 (4), 517–532. Agénor, P.R., 2008. Health and infrastructure in a model of endogenous growth. J. Macroecon. 30 (4), 1407–1422. Agénor, P.R., 2010. A theory of infrastructure-led development. J. Econ. Dyn. Control. 34 (5), 932–950. Aisen, A., Veiga, F.J., 2013. How does political instability affect economic growth? Eur. J. Polit. Econ. 30 (1), 151–167. Arellano, M., Bover, O., 1995. Another look at the instrumental variable estimation of error-component models. J. Econ. 68 (1), 29–45. Barro, R., 1990. Government spending in a simple model of endogenous growth. J. Polit. Econ. 98 (5), 103–125. Barro, R.J., Redlick, C.J., 2011. Macroeconomic effects from government purchases and taxes. Q. J. Econ. 126 (1), 51–102. Bassanini, A., Scarpetta, S., 2001. The driving forces of economic growth: panel data evidence for the OECD countries. OECD Economic Studies No. 33. Bayraktar, N., Moreno-Dodson, B., 2010. How can public spending help you grow? An empirical analysis for developing countries. Policy Research Working Paper No. 5367. World Bank. Blankenau, W., Simpson, N., Tomljanovich, M., 2007. Public education expenditure, taxation, and growth: linking data to theory. Am. Econ. Rev. 97 (2), 393–397. Blundell, R., Bond, S., 1998. Initial conditions and moment restrictions in dynamic panel data models. J. Econ. 87 (1), 115–143. Blundell, R., Bond, S., 2000. GMM estimation with persistent panel data: an application to production functions. Econ. Rev. 19 (3), 321–340. Bond, S.R., Hoeffler, A., Temple, J., 2001. GMM estimation of empirical growth models. CEPR Discussion Paper No. 3048. Centre for Economic Policy Research, London, UK. Burnham, R., Feinberg, R., Husted, T., 2004. Central city crime and suburban economic growth. Appl. Econ. 36 (9), 917–922. Càrdenas, M., 2007. Economic growth in Columbia: a reversal of fortune? Ensayos Sobre Poltíica Econ. 25 (53), 220–259. Chatterjee, I., Ray, R., 2009. Crime, corruption and institutions. Working Paper No. 20. Monash University. Chote, R., Emmerson, C., Tetlow, G., 2009. The fiscal rules and policy framework. In: Chote, R., Emmerson, C., Miles, D., Shaw, J. (Eds.), The IFS Green Budget, January 2009. Institute for Fiscal Studies, UK. Christiansen, L., Schindler, M., Tressel, T., 2013. Growth and structural reforms: a new assessment. J. Int. Econ. 89 (2), 347–356. Czabanski, J., 2008. Estimates of Costs of Crime: History, Methodologies and Implications. Springer, Berlin. Detotto, C., Otranto, E., 2010. Does crime affect economic growth? Kyklos 63 (3), 330–345. Drakos, K., Goulas, E., 2006. Investment and conditional uncertainty: the role of market power, irreversibility and returns-to-scale. Econ. Lett. 93 (2), 169–175. Drakos, K., Kallandranis, C., 2007. State-dependent response of investment to cash flow: evidence for asymmetry in European manufacturing. Appl. Financ. Econ. 17 (14), 1191–1200. European Commission, 2010. Review of costs of crime literature. Available at. http:// www.costsofcrime.org/AnnotedBibliography.php. Futagami, K., Morita, Y., Shibata, A., 1993. Dynamic analysis of an endogenous growth model with public capital. Scand. J. Econ. 95 (4), 607–625. Gemmell, N., Kneller, R., Sanz, I., 2011. The timing and persistence of fiscal policy impacts on growth: evidence from OECD countries. Econ. J. 121 (1), 33–58.

E. Goulas, A. Zervoyianni / Economic Modelling 49 (2015) 286–295 Goulas, E., Zervoyianni, A., 2013. Economic growth and crime: does uncertainty matters? Appl. Econ. Lett. 20 (5), 420–427. Guariglia, A., Poncet, S., 2008. Could financial distortions be no impediment to economic growth after all? Evidence from China. J. Comp. Econ. 36 (4), 633–657. Guariglia, A., Tsoukalas, J., Tsoukas, S., 2013. Investment, irreversibility and financing constraints: evidence from a panel of transition economies. Econ. Lett. 117 (3), 582–584. Heiskanen, M., 2010. Trends in policy-recorded crime. In: Harrendorf, S., Heiskanen, M., Malby, S. (Eds.), International Statistics on Crime and Justice, Publication Series No.64. European Institute for Crime Prevention & Control, HEUNI & UNODC. Hoeffler, A., 2002. The augmented Solow model and the African growth debate. Oxf. Bull. Econ. Stat. 64 (2), 135–158. Mankiw, G., Romer, D., Weil, D., 1992. A contribution to the empirics of economic growth. Q. J. Econ. 107 (2), 407–437. Mauro, L., Carmeci, G., 2007. A poverty trap of crime and unemployment. Rev. Dev. Econ. 110 (3), 681–712. Morgese-Borys, M., Polgàr, É.K., Zlate, A., 2008. Real convergence and the determinants of growth in EU candidate and potential candidate countries—a panel data approach. ECB Occasional Paper No. 86.

295

Peri, G., 2004. Socio-cultural variables and economic success: evidence from Italian provinces 1951–1991. B.E. J. Macroecon. 4 (1), 1–34. Roodman, D., 2006. An introduction to difference and system GMM. Working Paper No. 103. Centre of Global Development. Rooth, D., Stenberg, A., 2012. The shape of the income distribution and economic growth: evidence from Swedish labor market regions. Scott. J. Polit. Econ. 59 (2), 196–223. Saidi, H., Aloui, C., 2010. Capital account liberalization and economic growth: GMM system analysis. Int. J. Econ. Financ. 2 (5), 122–131. World Bank, 2006. Crime, violence and economic development in Brazil: elements for effective public policy. Report No. 36525. World Bank, Washington D.C. World Bank, 2007. Crime, violence and development: trends, costs and policy options in Caribbean. Report No. 37820. World Bank, Washington D.C. Yamarik, S., 2010. Human capital and state-level economic growth: what is the contribution of schooling? Ann. Reg. Sci. 47 (1), 195–211.