2.1 Model specification
Let
\({\mathrm{WF}}_{it}\) be the wildfire crime count in region
i at time
t. Consistent with the relevant literature (see Ganteaume et al.
2013 and the references therein), we assume that
\({\mathrm{WF}}_{it}\) is a function of a
\(k\times 1\) vector of the risk factors,
Rit, and partition
Rit as follows:
$$R_{it}^{\prime } = \left[ {X_{it}^{\prime } Y_{it}^{\prime } } \right],$$
(1)
where the entry elements of the vector
\({X}_{it}\) are the covariates for social-economic risk factors, and
\({Y}_{it}\) is a vector of the control variables that include the demographic, environmental, and crime deterrence variables. In particular, vector
\({X}_{it}\) includes the risk factors related to income inequality, material deprivation, violence, educational attainments, and labour market conditions. Below, we describe the risk factors in Eq. (
1) in relation to the relevant literature.
Starting with income inequality, the proxy for this covariate under consideration is the disposable income inequality, (INER
it). The variable INER
t is defined as the ratio of the total income received by 20% of the population with the highest income to that received by 20% of the population with the lowest income. As for the expected sign of the estimated coefficients, theoretical models in the general crime literature suggest a positive relationship between income inequality and crime. For example, in his seminal paper, Becker (
1968) argues that for a given probability of apprehension and expected punishment, higher levels of inequality increase the expected benefit of committing a crime for the relatively disadvantaged. In a similar theoretical framework, Fajnzylber et al. (
2002) suggests that the effect of income inequality in society is strongly related to an individual’s relative income. They argue that in the case of the wealthy, it is most unlikely that the rising rate in inequality induces them to commit crime. However, for poorer social actors, an increasing rate of inequality may be crime-inducing because such an increase implies a larger gap between the wages of the poor and the rich, reflecting a larger difference in income from criminal and legal activities. The idea that inequality causes crime is supported by several theoretical works (Ehrlich
1973; Imrohoroglu et al.
2000). However, strong theoretical models are not always supported by empirical evidence. The results of empirical studies reveal mixed evidence regarding the positive relation between crime rate and inequality. Whilst some have found evidence of a positive relationship between inequality and crime (Enamorado et al.
2016; Harris and Vermaak
2015; Coccia
2017; Fajnzylber et al.
2002), others have failed to find any significant relationship (Bourguignon et al.
2000; Neumayer
2005).
Closely related to income inequality is the material deprivation risk, MATDEP
it. In the literature, economic theory suggests that individuals are more likely to get involved in criminal activity when they experience a negative income shock. In his seminal work, Grossman (
1991) established a relation between crime and material deprivation in terms of the opportunity cost framework. The author argues that decreasing income levels reduce the opportunity cost of engaging in crime with respect of other legal economic activities (Seter
2016).
Proxy variables for household wealth and the labour market conditions were also considered in the model. Namely, covariates for unemployment rate, UNEM
it, per capita disposable income, INC
it, and employment rate in non-agricultural sector, EMPL
it, were considered as potential risk factors. With regard to the sign of the estimated parameters, theoretical models in the general crime literature suggest a strong positive relationship between a worsening of the labour market conditions and crime. For example, Grogger (
1998) estimates a structural model of time allocation between criminality, the labour market, and other non-market activities, suggesting a negative relation between wages and criminal activity. Grogger’s (
1998) empirical findings show that young men are responsive to wage incentives and that the racial difference in crime rate in part can attribute to the labour market. Similarly, Gould et al. (
2002) found that the labour market conditions, especially wages, to be strongly related to crime for those who are most likely to commit crime (less educated men). However, concerning wildfires, the empirical studies have found mixed results. For example, Maingi and Henry (
2007) found that there to be no relationship between fire occurrence and unemployment (Sebastian-Lopez et al.
2008; Martınez et al.
2009; Lovreglio et al.
2010), whereas Prestemon and Butry (
2005) showed that arson fires and unemployment were related. A recent strand of the literature focuses on the indirect relation between the labour market and wildfire crime. This literature suggests that forests have been voluntarily set on fire to create firefighting jobs or to gain land for agriculture and pastures which have been retained due to being more valuable than logging (Leone et al.
2002).
To capture the effect of educational attainment on wildfire crime, two covariates were used: (i) the rate of population with an upper secondary level of education, EDUCit, and (ii) the rate of population with a tertiary education, UNIVit. The rationale for including the two proxy variables for education attainment is that we expected the return of education on income to be higher for individuals with a university degree.
With respect of the causal relationship between crime and educational attainment, in the literature, it has been argued that an individual’s education level may impact their decision to commit a crime in several ways. First, the effect of income is positively related to education. This is because higher levels of education attainment can be associated with increasing returns of legitimate work and a rise in the opportunity costs of illegal behaviour (Lochner
2004; Lochner and Moretti
2004). Second, the resources allocated to education create time constraints that deter criminal offences. Tauchen et al. (
1994) investigated this “self-incapacitation effect” and found that the time spent on pursuing education is negatively correlated with the probability conviction amongst youngsters. Hjalmarsson (
2008) focused on the impact of getting arrested before finishing school on the probability of graduating from high school and found that the probability of a young person being convicted for committing a crime greatly increases their likelihood of becoming a high-school dropout. Third, a stream of literature also associates a greater education with higher life satisfaction which, in turn, reduces the probability of committing a crime. For example, Oreopoulos (
2006) and Lochner (
2004) suggest that higher levels of education can increase risk aversion, lowering the crime rate. In an interesting paper, Usher (
1997) argues that education promotes a “civilisation effect” that contributes to reducing the incidence of criminal activity. The author argues that education conveys a civic externality, a benefit to society over and above the benefit to the student in terms of enhancing his future earning power. Michetti et al. (
2019) analysed the determinants of the monthly variations in forest fires for Italian regions between 2000 and 2011 and concluded that education attainment plays an important role in preventing fraudulent activity. Similarly, Torres et al. (
2012) found that the areas with fewer fires are characterised by a population with higher levels of education.
The last risk factor considered is the level of violence. Also, in this case, two proxies for violence were considered: (i) the homicide rate (HOMRit) and (ii) organised crime, (ORGCit), defined as the conviction rate for organised and mafia-related crime. The rationale for including the homicide rate as well as organised crime convictions is that the latter variable likely suffers from a significant measurement error. Organised crime is a difficult phenomenon to capture and using the number of trials for organised crime as the sole covariate to assess the impact of organised crime on wildfire crime may not be informative. In this respect, the inclusion of the covariate “homicide rate” may be useful to signal the significant presence of organised crime in a region. Clearly, the overall homicide rate does not distinguish between homicides committed by criminal organisations and other homicides. On the other hand, it is unlikely to suffer from a measurement error and allows us to test the hypothesis that the degree of violence in a region has an effect on the occurrence of wildfire crime.
As far as the sign of the expected estimated coefficients is concerned, in the literature, studies on the relationship between organised crime and wildfire crime are rare. One of the few empirical works that considers this relationship is the EFFACE (
2016) report where the evidence found there to be a positive relation between organised crime (mafia-like organisations) and the rate of fire crimes. The influence of organised crime is reported to be stronger in Italy’s southern regions where the government’s ability to enforce the law is weaker. The literature on environmental crime mainly focuses on the growing role of organised crime in relation to other types of environmental crime. This is particularly the case for illegal dumping and the international illegal trafficking of hazardous waste where it was found that organised mafia-like criminals play a significant role in environmental criminality (Germani et al.
2018). Overall, being the socio-economic determinant of other types of environmental crime similar to wildfire crime, we expect a positive relationship between organised crime and arson.
Regarding the control variables, the entry elements of vector
Yit in Eq. (
1) are (i) deterrence factors, (ii) weather-related factors, and (iii) demographic risk factors:
$$Y_{it} { } = \left[ {{\text{TRIAL}}_{it} , {\text{EQ}}i_{it} ,{\text{ RAIN}}_{it} , {\text{TEMP}}_{it} ,{\text{ AGRI}}_{it} , DEN_{it} } \right].$$
(2)
In Eq. (
2), the covariate
\({\mathrm{TRIAL}}_{it}\) is a proxy that captures the probability of apprehension. According to the literature, a higher deterrence level reduces the level of wildfire crime (Canepa and Drogo
2021 and the references therein), and we therefore expect a negative estimated sign between WF
it and
\({\mathrm{TRIAL}}_{it}\).
The variable
\({\mathrm{EQ}i}_{it}\) is a proxy for the quality of institutions. The quality of governance and institutions is another important deterrence factor which not only affects wildfire crime directly but also affects the socio-economic outcomes such as education, poverty, and inequality; hence, it has an indirect effect on forest fires through its effect on these socio-economic factors.
2 We therefore expect there to be a negative correlation between this covariate and wildfires (Charron et al.
2019).
Concerning weather-related factors, in Eq. (
2), the variable RAIN
it is the annual precipitation in mm and
\({\mathrm{TEMP}}_{it}\) is the temperature, measured as the mean temperature. In general, weather conditions that cause downward changes in fuel moisture and, consequently, upward changes in fuel availability are expected to increase the probability of wildfire occurrence (Albertson et al.
2009; Plucinski
2014; Guo et al.
2016). Similarly, higher mean and maximum temperatures are expected to exhibit a positive relation with wildfires (Preisler et al.
2004; Carvalho et al.
2008).
Looking at the population density, DEN
it, in the fire related literature, an increase in population density has been found to be positively related to wildfire crime. For example, Catry et al. (
2007) observed that a large majority of the fire ignitions in Portugal occurred in the municipalities with the highest population densities. Gonzalez-Olabarria et al. (
2015) found that the distribution of arson in north-eastern Spain occurred near coastal areas where the population density was higher. Similarly, Romero-Calcerrada et al. (
2008) found there to be a positive relationship between the intensive use of the territory and ignitions in the forest areas in Spain (Padilla and Vega-Garcıa
2011).
Finally, AGRI
it is the proportion of population in agricultural employment. Socio-economic transformations in rural areas such as rural exodus, a reduction in agricultural employment, and the abandonment of agricultural land may contribute to wildfire crime. In the related literature, the impact of this covariate is controversial with several empirical studies reporting a positive relation between fire occurrence and agricultural activities (Martınez et al.
2009; Rodrigues et al.
2016 amongst others), whilst others do not find that a decreased human impact associated with agricultural land abandonment leads to a statistically significant decrease in fire ignition probability (Ricotta et al.
2012).
2.2 Estimation methods
To investigate the effect of the risk factors in Eq. (
1), we considered a quantile regression model with a nonadditive fixed effect as suggested by Powell (
2022) (see also Chernozhukov and Hansen
2008; Powell
2020). The adopted model specification is particularly convenient as it allows us to shed some light on two related questions. First, is there a causal relation between the socio-economic risk factors and wildfire crime? Second, does one unit increase in a given risk factor of vector
\({R}_{it}\) in Eq. (
1) affect the regions with a lower wildfire crime rate differently from the regions with a higher wildfire crime rate?
To introduce the model under consideration, we relate it to the linear regression specification that is usually adopted in the literature. Most of the empirical works consider the conditional distribution of
\(E\left(\mathrm{WF}|R\right)\) using a structural model:
$${\text{WF}} = \alpha + R^{\prime } \beta + \varepsilon ,$$
(3)
where
\(\alpha\) and
\(\beta\) are unknown constant parameters, and
\(\upvarepsilon\) is an error term.
3 In Eq. (
3), the parameter
β has a causal interpretation as some effect of the risk factor
R on WF. However, for the Italian case, assuming that the marginal effects of the risk factors are the same in all regions may not be informative due to the heterogeneity of the regional economy. Rather than letting the differences in marginal effects go into the error term
\(\upvarepsilon\) and interpret β as some sort of average, we try to learn about the marginal effect heterogeneity.
In this paper, we allow the parameter
\(\beta\) to be a “random coefficient”, and in other words, the structural model we consider allows for region-specific parameters. Since in our model,
\(\beta\) is a vector of region-specific coefficients, they are no longer constants like in Eq. (
3), but rather random variables in a “random coefficients” model. The additive error term
\(\varepsilon\) in Eq. (
3) is now redundant since it is absorbed into the random intercept.
To make the model more tractable, we follow Powell (
2022) and assume that
\(\beta\) is a deterministic (but unknown) vector-valued function
β(·) applied to an unobserved random variable
U:
$$WF_{it} = R_{it}^{^{\prime}} \beta_{j} \left( {U_{it}^{*} } \right),$$
(4)
where
\({U}_{it}^{*}=f\left({\alpha }_{i}, {\varepsilon }_{it}\right)\) for some fixed effect
\({\alpha }_{i},\) and unknown function
f. Assuming that
\({U}_{it}^{*}\) is continuous, it can be normalised to have a Unif(0, 1) distribution. The model in Eq. (
4) allows the parameters to vary based on an unspecified function of the fixed effect and an observation-specific disturbance term whilst permitting individual-specific heterogeneity. Under the monotonicity assumption,
\({R}_{it}^{^{\prime}}{\beta }_{j}({U}_{it}^{*})\) is an increasing function of
U, then
$$P\left[ {{\text{WF}}_{it} \le R_{it}^{^{\prime}} \beta \left( \tau \right)|Z_{it} } \right] = \tau ,$$
where
\(\tau \in \left(\mathrm{0,1}\right)\) and
\({Z}_{it}=\left({Z}_{i1},\dots ,{Z}_{i,T}\right)\) is a set of instruments that can be arbitrarily correlated with the nonadditive fixed effect.
The model in Eq. (
4) can be estimated using the GMM method in an instrumental variable context under the following moment condition
$$E\left\{\left({Z}_{it}-{Z}_{is}\right)\left[1\left({WF}_{it}\le {R}_{it}^{^{\prime}}\beta \left(\tau \right)\right)-1\left({WF}_{is}\le {R}_{is}^{^{\prime}}\beta \left(\tau \right)\right)\right]\right\}=0,\mathrm{ for \,all\, }s,t,$$
$$E\left[ {1\left( {WF_{it} \le R_{it}^{\prime } \beta \left( \tau \right)} \right) - \tau } \right] = 0.$$
(5)
Under condition in Eq. (
5), in addition to other standard regularity conditions, the GMM estimation method produces consistent and asymptotically normal estimators (for more details, see Powell
2022).
2.2.1 Model specification issues
In the model specification procedure, two issues were taken into consideration. First, in Eq. (
1), more than one proxy variable for a given socio-economic risk factor was considered and some of them may be highly correlated. Accordingly, a number of models were estimated using the subset of the variables included in Eq. (
1). In the model selection procedure, the possible presence of multicollinearity was assessed using the eigenvalues of the different covariance matrices and computing the conditioning index (CI). Following Pena and Renegar (
2000), the CI index was calculated as
$${\text{CI}} = \sqrt {\frac{{\lambda_{{{\text{max}}}} }}{{\lambda_{{{\text{min}}}} }}}$$
(6)
where
\({\lambda }_{\mathrm{max}}\) is the largest eigenvalue and
\({\lambda }_{\mathrm{min}}\) the smallest eigenvalue of the variance covariance of the risk factor matrix
Rit. If
\(\mathrm{CI}=1\), there is no evidence of collinearity between the covariates in the estimated model. However, as the collinearity increases, the eigenvalues in Eq. (
6) become both greater and smaller than 1 and the CI number increases. Pena and Renegar (
2000) suggest that
\(\mathrm{CI}<10\) for a well-defined matrix as opposed to
\(11<\mathrm{CI}<30\) for moderate multicollinearity. Accordingly, the condition we imposed for introducing a covariate to the final model specification was a calculated CI number that was less than 10.
The adopted model specification procedure resulted in the selection of the best fitting models from a set of candidate specifications. We labelled these models M1, M2, M3, and M4, respectively. Models M1 and M2 include MATDEP
it as the risk factor, whereas, in models M3 and M4, the risk factor UNEM
it is included. However, the covariate MATDEP
it is excluded. The reason for this specification is that these two covariates were found to be highly correlated and, therefore, only one risk factor at the time was considered. Similarly, the two pairs of covariates of EDUC
it and UNIV
it, and INC
it and EMPL
it produced a high CI index. Accordingly, the EDUC
it risk factor was included in models M1 and M2 but excluded from models M3 and M4 where the covariate was replaced by UNIV
it. To avoid multicollinearity, INC
it was included in model M1 but not in model M2, and it was excluded from the model specification when the risk factor UNEM
it was included as the covariate (models M3 and M4). Surprisingly, the two proxies for violence, ORGC
it and HOMR
it, were not found to be highly correlated, and the CI index in all estimated models was found to be less than 10. For this reason, both risk factors were included in the models. The condition index for the estimated models is presented in Table
6 along with the correlation matrix of the covariates under consideration (Table
5).
The second issue in relation to the estimation of Eq. (
4) is the choice of instrumental variables for the potentially endogenous covariates. The related literature suggests that the inequality variable may be endogenous since income inequality may be correlated with crime (see Enamorado et al.
2016; Harris and Vermaak
2015; Fajnzylber et al.
2002). However, the reverse may also be true. To mitigate concerns about this form of reverse causality, we constructed an instrumental variable that is correlated with changes in regional inequality but not associated with regional wildfire crime rates. Specifically, we followed Enamorado et al. (
2016) (see also Bartik
1991; Blanchard and Katz
1992; Boustan et al.
2013) and constructed the instrument by calculating the predicted inequality growth rate by interacting the regional inequality shares with the national inequality growth rate. In particular, for each year, we estimated the percentile share of inequality. In doing so, we estimated to which national percentile of inequality each region belonged to in the initial year. We then used the estimated lagged shares of inequality and interacted them with the change in inequality at the national level. By design, this instrument allowed us to isolate the changes in inequality at the regional level that is driven by national shifts which should be correlated with regional welfare indicators but not with the wildfire crime observed in each region. In addition to this instrument, we followed Aldieri and Vinci (
2017) (see also Powell
2022) and used lagged explanatory variables as instruments. The idea is that the lagged level of the covariates may impact the trends in risk factors but do not share a statistically significant relationship with the current wildfire crime rate. This should also control for the potential endogeneity of the other covariates.
4