Forest Fires in Italy: An Econometric Analysis of Major Driving Factors

Despite the relevant fire risk to which Italy is subject from the north to the south, very few analyses focus on this topic. This article investigates the causes of forest fires frequency and intensity in Italy during the first decade of the twenty-first century. The dynamical aspects of fire danger are explored through the use of panel data techniques, which fully capture the impacts on forest fires regarding changes in both socioeconomic and climatic conditions. Italy is treated as a unique region in an initial model specification, and is then split into 3 geographical areas (north, center, and south) to capture locally specific aspects. Two different dependent variables are alternatively employed and a number of ad hoc tests are performed to corroborate the robustness of our estimates. The results highlight the importance of considering the fire situation separately for the northern, central, and southern parts of Italy. While the presence of railway networks positively affects fire risk, the impact of livestock depends on its specific composition. Favorable effects in fire reduction are represented by the increase in education levels (north and center) and touristic flows (north and south), and by the containment of illegal activities (south). Weather patterns appear to be important determinants throughout the Italian peninsula.


Introduction and Motivation
Forestland and trees offer vital services such as commercial and recreational uses, water and climate regulation services, and carbon sequestration activity. Unfortunately, there are several forest disturbances that undermine these service provisions. Compared to other calamities such as pests, plant diseases, wind, frost etc., fire represents the most threatening one in the Mediterranean area (Alexandrian et al., 1999).
Forest fires recur year after year, with a devastating intensity. There is no natural and vegetative landscape, which has not been altered by fire. Despite the fact that during the last decades scarce precipitations along with high temperature levels have been impacting the fire risk (Moriondo et al., 2006), these two variables represent just some of the possible causes of forest fires. Fire risk depends on a number of regional specific factors (Moreno et al., 1998and Martinez et al., 2008Westerling et al. 2006 for US, among others), human attitudes (Barbero et al., 1990;Martinez et al., 2008 andPausas and Keeley, 2009), and weather patterns (Westerling et al. 2006;Pausas and Fernández-Muñoz, 2012). The combination of all these aspects is therefore responsible for generating the final scenario of fire danger.
Recent studies acknowledge that forest fire risk is common to the overall southern Mediterranean area (Pausas et al., 2008 for Mediterranean area; Miranda et al., 2008 for Southern Europe; Dimitrakopoulos et al., 2011 for Greece;Costa et al., 2011 for Portugal). Italy is not an exception since it is affected by relevant fire risk from the north to the south. A consistent number of fire events occur not only during the warm seasons, as one would expect, but also during the winter. In the last decade, 360 fire events were registered on average each year. This frequency is associated with around 40 km 2 of land burnt, per year, and per region. In the first quarter of 2012, 3900 forest fire events occurred corresponding to 190 km 2 of forest area covered by fires. Compared to the same period in 2011, Italy experienced a 165% increase in the number of fire events and a 196% enlargement of the area affected (Corpo Forestale dello Stato, 2012). 2 This article fills several gaps in the literature by investigating the causes of forest fire frequency and intensity examining fire regimes during the first decade of the twenty-first century.
First, it focuses on the Italian situation, which, so far, has been analyzed by very few investigations.
It highlights signs and magnitudes of the effects considered, which are sometimes controversial in the literature. The present study initially considers the overall Italian peninsula. 3 It then analyses the fire situation by classifying the Italian territory into three geographical areas (north, center , and south) to highlight differences in fire patterns across regions and to provide a more in depth investigation regarding the role of humans and the characteristics of forest fire occurrences.
Third, on the methodological side, we implement a panel data approach fully capturing the dynamics of fire occurrence and the change in socio-economic and weather conditions in time. To corroborate the robustness of our estimates we use two, rather than just one, dependent variables, in addition to a number of ad hoc tests. Apart from a few exceptions, existing studies for Italy are based on the observation of simple historical trends and do not apply accurate econometric techniques to corroborate obtained results. Additionally, geographical-oriented approaches to forest fire analysis tend to account for fire realizations at some point in time (e.g., one specific year) neglecting dynamic aspects. This paper is organized as follows. Section 2 illustrates the historical situation of forest fires in Italy while section 3 discusses main driving factors of forest fire frequency and intensity. Section 4, in addition to describing the data construction process, draws descriptive statistics on both dependent and independent variables. Section 5 offers a brief survey on methodologies useful to analyze the fire regimes, and then describes the chosen theoretical framework. Sections 6 and 7 report findings derived from different model specifications while section 8 concludes.

A Brief History of Forest Fires in Italy
Among the Mediterranean countries Italy represents one of the most prone areas to forest-fires. 4 For example, Moriondo et al. (2006) find that Italy experiences the highest increase in annual extreme events in future scenarios conducted through return period analysis. Fire events cover the entire Italian peninsula, from the north to the south, however the larger wildfire events normally occur in the south (Figure 1). This explains why the few existing Italian-based analyses on fire events mostly focus on southern Italy (Pazienza and Beraldo, 2004;Bajocco and Ricotta, 2008).  The fire situation in Italy has a seasonal component. Events take place not only during the hotter periods but also during the colder ones. The most affected months appear to be February-March, July-August-September ( Figure 2).   Finally, using a ten-year moving average, a downward trend in fire frequency is shown for most regions in the north and the center . 5 For the remaining regions, especially in the south, a steady or mixed trend applies ( Figure 4). 5 Each interval represents a 10-year average of forest fire events for each region. For example, 1 is associated to the average of fire frequency between 1990 and 1999, while 2 relates to the 1991-2000 period, and so on. Information used (from JRC) lacks data for Sardinia during 1990-1996, and for Liguria in 1996. These missing data are accounted for when the moving average trends are obtained.

Forest Fires Driving Forces
Notwithstanding a high number of fire events remains without a well-identified cause (Alexandrian et al., 1999;World Wide Fund, WWF, 2004), 6 the literature typically distinguishes between natural and socio-economic conditions. This has originated two main streams of research. The first places more attention on natural-climatic or weather patterns (Chou et al., 1993;Pinol et al., 1998;Pausas, 2004;Moriondo et al., 2006;Westerling et al. 2006;Running, 2006;Pausas and Bradstock, 2007;Pausas and Fernández-Muñoz, 2012). The second emphasizes the social-economic variables that are assumed to generate higher impacts on fire ignition (Cardille et al., 2001;Leone et al. 2002;Yang et al., 2007;Martinez et al. 2008Martinez et al. & 2009Leone et al., 2009;Lovreglio et al., 2010;de Torres Curth et al. 2012). Figure 5 summarizes frequently used drivers in forest fire analyses . Higher humidity and precipitation slow down the process of forest fuels dry out, reducing fire risk. Pausas, 2004;Trouet et al., 2006;Westerling et al., 2006;Pausas and Bradstock, 2007;Trouet et al., 2010 Temperature Positive A rise in temperature leads to a drop in fuel moisture, and increase flammability of live and dead fuels, therefore raising the likelihood of observing fire events. Pinol et al., 1998;Wotton et al., 2003;Pausas, 2004;Trouet et al., 2006;Westerling et al., 2006;Running 2006;Pausas and Bradstock, 2007 Drought Negative It depends on the combination of precipitation amounts and temperature: it causes plant mortality and impact long-run forests flammability. Lloret et al. 2002;Breda et al. 2006;Aguado et al., 2007 population density/ growth rate Mixed A population enlargement may increase possible ignition causes due to human accidents. On the other hand, higher land demand, following the raise in population, could slow down the land abandonment process, e.g., the forest spontaneous re-growth. Donoughue and M ain, 1985;Serneels and Lambin, 2001;Leone et al. 2002;M ercer and Prestemon, 2005;Vadrevu et al., 2006;Syphard et al., 2007;Gellrich et al., 2007;M artinez et al., 2009;de Torres Curth et al. 2012; Infrastructures, connections Mixed A greater number of roads and railways may put more pressure on wild lands raising possible ignition causes. Nevertheless, good communication routes may help fire prevention and suppression. Cardille et al., 2001;Pew and Larsen, 2001;M artinez et al., 2009; Agriculture and pasture intensification Mixed Fire is often used by shepherds and farmers to i) maintain herbaceous vegetation only; or ii) eliminate wasting harvest in borders of croplands, iii) remove pests. Bolgiano, 1998;Kuhlken, 1999;Chuvieco et al., 1999;Leone et al. 2002;Pazienza and Beraldo, 2004;Vigilante et al., 2004;Whalley, 2005;M artinez et al., 2009 Education Negative More educated people may have a higher civic sense wich helps containing the number of fires due to human perverse behaviour or accidents. Butry et al., 2002; Unemployment Poverty level Positive Higher wellbeing and employment levels may discourage people from setting forests on fire for profit reasons. Chuvieco et al., 1999;Leone et al., 2002;Pazienza and Beraldo, 2004;M ercer and Prestemon, 2005;M aingi and Hendry, 2007;M artinez et al., 2009;de Torres Curth et al., 2012 Depopulation of rural areas Positive It implies land abandonment and spontaneous colonization of natural vegetation. This, translates in additional forest biomass, and consequently in greater forestland flammability. Le Houérou, 1987;M acDonald et al., 2000;Romero-Calcerrada and Perry, 2004;Koutsias et al. 2005;M artinez et al. 2009.

Touristic migration Mixed
The touristic use of forests for recreation could raise the probability of ignition by accident or negligence (campfires, smokers, etc.); Nevertheless, the parallel forest preservation for recreational scopes could impact the same probability with opposite sign. Atkinson and Therneau, 2000;Leone et al. 2002;M artinez et al., 2009 Presence of illegal organizations Positive Illegal organizations can control economical activities connected with land; set forests on fires to gain land for agriculture or pasture, retained more valuable than preserving forests for recreational use or logging. Leone et. al, 2002;Gonzalez-Olabarria and Pukkala, 2011
Although the data on forest fires is available for all 20 Italian regions, we eliminated the Aosta Valley region from our analysis due to the lack of data on most of the socio-economic and climatic variables. We believe that this exclusion has no impact on final results, since this region is the least prone area to forest fires in Italy. 7

Dependent Variables
Two different dependent variables are used for the analysis. The first represents the frequency of forest fires, namely, the number of fire events in a given region in a given year (lnfn_tot). The second (lnburntot) is the area burnt in each region, in each year, measured in square kilometers.
Both variables are expressed in natural logarithms, a convenient transformation which does not alter final results. The two variables are highly and positively correlated, but capture different aspects of fire dynamics. Therefore, using them both allows a richer interpretation of the results and also provides a robustness check. As shown in Table 1  and south) we notice that southern Italy is affected the most, followed by central and northern Italy.   Table   7 Aosta Valley records on average 15 events, which translate into 6.5ha size of area burnt per year.
8 2). To control for yearly fluctuations effects on fire occurrence Preisler et al., 2004) we add year dummies in our panel-based models.  The monthly average of forest fire size suggests that the largest fire events take place in July and August. June and September also show high figures although 4 times lower than the mean of July and August. Interestingly, despite what one would expect, a significant number of forest fires occur during low-temperature periods, especially in February and March (Table 3). Pooling the information obtained on the size and frequency at monthly averages per region and year (Tables 3 and 4), and given the outcomes of a correlation analysis, we acknowledge that the two variables are highly correlated. Yet, in some cases, their comparison helps us to draw interesting information. For instance, while in February the number of events is in line with that of the remaining months, the total amount of forest area covered by fires is relatively high, suggesting a large average size.

Explanatory Variables: Weather Patterns
Weather patterns in the analysis are represented by precipitation and temperature, temperature excursion (i.e., the difference between the annual maximum and minimum temperature), and the number of consecutive hot days in one year. The interactions of some of these variables are also included among the explanatory ones.
Over the period considered, the annual average number of consecutive hot days for each region ranges between 0 and 43.25 (Sardinia in 2003). In the north, center, and south, on average, 2.80, 6.40, and 11.65 consecutive hot days are recorded respectively. The annual excursion in temperature ranges between 7.3 °C and 12.8 °C. Despite yearly fluctuations and inter-annual variability, general paths can be identified in precipitation levels (Figure 7 (a) and Table 5): they are higher during the winter period (September-December) and notably decrease from May to August, especially in July. Temperature (Figure 7(b) and Table 6) shows a clear seasonal effect. Its level progressively raises starting in January, peaks around July, and then decreases slowly until the January values again. This seasonal effect is normally reflected in fire events (see e.g., Bajocco and Ricotta, 2008). Regarding temperature, we focus on the summer data instead, constructing a variable related to the average June to August period (35.4°C).   Concerning economic conditions, the households' wellbeing varies significantly between the north and the south. For example, the ISTAT indicator of relative poverty reports, for 2008, shows that the percentage of population below the poverty level ranges between 3.9% (Emilia Romagna) and

Explanatory Variables: Socio-economic Factors
28.8% (Basilicata and Sicily). During the whole period analyzed, on average, around 22%, 5.92%, and 6.01% of the population was declared to be "poor" respectively in southern, central and northern Italy. According to our data, the poverty indicator is negatively correlated with the portion of population that has a tertiary degree in a given region, which varies significantly from the north to the south and normally increases over time. The use of land is closely linked with the employment rate in the agriculture sector. The percentage of population employed in this sector in 2008 varies from 1.70% (Lazio) to 9.12% (Calabria).
Italy is characterized by the presence of illegal associations, which control economical activities connected to agriculture and pasture practices. Their control may impact land allocation and therefore, forest fire occurrences. 8 Especially in the past, forests have been voluntarily set on fire to create firefighting jobs or to gain land for agriculture and pasture which were retained more valuable than logging (Leone et. al, 2002). Indeed, low timber returns along with negligent forest management can result in higher fire risk (Gonzalez-Olabarria and Pukkala, 2011). 9 To capture the effect of the presence of illegal activities/associations on human-induced fires we use the proxy variable 'number of extortions per 1000 inhabitants', which shows an increasing trend in time in many Italian regions. This variable may also reflect the different incidence across regions. For example, in 2008, extortions per 1000 inhabitants ranged between 0.04 (Friuli-Venezia Giulia) and 0.21 (Campania). Table 7 below summarizes this information. Notes: Variable labels and description is provided in the Appendix.

Brief Survey on Existing Methodologies
Several methodologies have been used within the existing literature to explain (and/or predict) the main factors of fire risk. The choice of methodology strictly depends on the objective of the study, which can either be that of predicting future fire incidences or explaining the variation of the dependent variable. Therefore, it depends on the characteristic of the dependent variable, which can be discrete and generally dichotomous, or continuous. If, for example, we refer to fire occurrence (yes or no expressed as a 1-0 variable) or ignition index (high ignition and low ignition respectively receive the value of 1 and 0), the dependent variable is a dichotomous one. On the other hand, if we refer to the total number of fires in a given period (e.g., year) and in a given region, the dependent variable is a discrete one. Finally, if we consider the total area burnt in km 2 , the dependent variable is a continuous one.
One of the most frequently implemented techniques within fire risk analysis is the binary response regression, which takes different specifications according to the behavior of the dependent variable (i.e., logistic regression: Martinez et al. 2008Martinez et al. & 2009Poisson regression: Wotton et al. 2003; Poisson, binomial or negative binomial: Venables and Ripley, 1997;Cardille et al. 2001). If its variance is greater than the mean, negative binomial regression can offer a better fit, whereas Poisson regression is normally used when the variance equals the mean.
Other methodologies use stepwise-multiple linear regressions (i.e., multivariate OLS) when the dependent variable is continuous (e.g., Vadrevu et al., 2006). In most applications, Geographic Information Systems software (GIS) is then used to gather information on fire events or visualize results through mapping.
The use of binary response regressions and of stepwise multiple linear regressions does not prevent the researcher from incurring problems of biasedness and inconsistency of the estimates due to unobserved heterogeneity across regions/areas, which are not accounted for in the analysis (Greene, 2008). The spatial heterogeneity might rise due to a lack of structural stability or homogeneity of unit observation across space. Supposedly, Italian regions are heterogeneous in terms of cultural differences (Naveh, 1995), different land cover types (Bajocco and Ricotta, 2008) process is required to avoid potential multicollinearity problems among the explanatory variables prior to regressions. However, all exogenous variables could be used with the CRT analysis even though some of them are highly correlated (see De'ath and Fabricius for further information on the advantages of using the CRT approach). This methodology has recently been implemented by Sturtevant and Cleland (2007) to individuate the main exogenous determinants of forest fires in northern Wisconsin between 1985 and 2000. They classify the dependent variable according to different intensities of fire events. The relative importance of the different explanatory variables expresses the extent to which a specific variable is able to improve the model, namely to decrease the misclassification of forest fires or forest intensities for a given split, represented by their inclusion into the wrong groups (Atkinson and Therneau, 2000).
While there is a variety of cross sectional analyses investigating forest fire events, according to our knowledge, there are very few articles providing a panel data specification. Although Vadrevu et al. (2006) attempt to head towards this direction by using several years of data, they do not fully exploit the advantages of a panel data approach given their use of yearly data in a cross-sectional fashion. A panel approach has been offered by Pazienza and Beraldo (2004), which examine the effects of forestry legislation on the frequency of fires in southern Italy (National Park of Gargano).
However, they only focus on the southern part of Italy and account for very few socio-economic variables (unemployment rate, number of civil and industrial vehicles), disregarding many timevariant factors and totally neglecting weather patterns. By looking at municipalities, they find a low level of heterogeneity across them suggesting the use of pooled OLS regression estimation methods. Gonzalez (2007) analyzes the potential link between house pricing and the population density for the regions of Spain to explain the potential link between the house pricing and the fire events after accounting for the region and year fixed effects. Dogandjieva (2008), on the other hand, examines the relationship between land, wheat, and timber prices and forest fire incidences in four countries: Spain, Greece, Italy, and Bulgaria. It is found that the fixed effects estimation yields mixed results, confirming the existence of a link between profit motives and forest fires. Both in Dogandjieva (2008) and Gonzalez (2007), the majority of the time varying effects were excluded from their studies, potentially causing omitted variable bias.

Theoretical Framework
Aiming to overcome some of the estimation problems described in section 5.1, we propose a panel data approach to investigate the influence of both socio-economic and weather patterns on fire occurrence in Italian forests during the last decade. This approach allows us to capture not only the dynamics but also the potential changes in the relative importance of such driving factors of fire occurrences.
Both fixed and random effects models are used to account for the unobserved heterogeneity across regions and, most importantly, to obtain unbiased and consistent estimates for the time-varying explanatory variables. As explained in Wooldridge (2002), the fixed effects can be performed either by i) adding region and year dummies to the regressions (least squares dummy variable regression, or LSDV); ii) transforming variables by subtracting the regional average from both the dependent and independent variables (entity demeaned OLS regression). Given that these two approaches lead to the same coefficients -the entity demeaned OLS regression is identical to within-region estimation (Baltagi, 2008) -we choose to use the second one (equation 3) not to reduce the available degrees of freedom.
Specifically, a linear panel-data one-way model is employed to capture either the unobserved regional or time effects in the error term (time effects are tested by adding year dummies in the fixed and random effect models). This takes the following form: In (1) where :   (Greene, 2008): where α is the common intercept and i v are assumed to be random processes.
Given that the relation between i µ and itm X is not known a priori we use the Hausman specification test (Hausman, 1978) to test the correctness of the fixed effects assumptions and the Lagrange Multiplier test or LM test (Breusch and Pagan, 1980), is conducted to detect the presence of random effects.
Given the panel-data nature of our analysis, one of the main issues to be checked is the existence of an autoregressive (AR) process in the residuals. We test the presence of serial correlation performing a Wooldridge test (Wooldridge, 2002). Relying on its results we include AR (1) disturbances in both the RE and FE models. In order to test if region-specific errors are independent across regions, we further run the Pesaran cross section test (Pesaran, 2004) and find no dependence.
For each of the two dependent variables (fire frequency or fire size in logarithmic form), we estimate the model considering first all the Italian regions together (2 unrestricted model), and then separating Italy into north, center , and south (3 region-specific models), (Table 8). In doing so, we maintain the regressors constant on the right-hand side, although we then eliminate the variables resulting not statistically significant in each of the regional model. Additionally, for each of the two sets, results are reported for pooled OLS, FE, and RE models in which year dummies have been added to capture the possible time-specific shocks, common to all the regions (e.g., a policy that is effective over the entire Italian peninsula).  Table 9 and Table 10 in which pooled OLS (column 2), FE (column 3), and RE (column 4) approaches are compared.

Unrestricted Model & Logarithm of the Number of Fire Events as Dependent Variable
Testing jointly for fixed and random effects we find the existence of regional fixed effects rather than a year effect for the FE model; year dummies are jointly significant for all other specifications in Results show that increasing railway density by 1% raises the total number of fire events by 0.8% across regions and time. Consistent with Cardille et al. (2001) and Martinez et al. (2009), we find that railway density puts pressure on wild lands increasing the number of fire events. On the other hand, a percentage increase in relative poverty fosters fire frequency by 4.18%. In accordance with 13 Although the literature does not provide any specific critical value, the Baltagi-Wu LBI statistic detects a serious serial correlation when it is far below 2 (see Baltagi and Wu, 1999). As reported in Table 9, this statistic is 1.99, very close to 2 indicating that auto-correlated residuals (fourth column) disappear when an AR(1) term is included in the regression (column 5). de Torres Curth et al. (2012), although in opposition to Leone et al. (2002) and Martinez et al. (2009), we report that the fire events are higher in the areas where poverty levels are high. Similarly to Leone et al. (2002) and Martinez et al. (2009) we conclude that touristic flows impact positively the number of forest fires: 1% increase in tourist arrivals per km 2 leads to 0.24% raise in the total number of fire events.
Finally, we find that depending on the extent to which land is used, intensification in the agriculture activity either puts pressure on, or generates protection against, forest fires (Chuvieco et al., 1999;Leone et al. 2002;Vigilante et al. 2004;Pazienza and Beraldo, 2004;Martinez et al., 2009). For example, a percentage increase in the number of people working in the agriculture sector has the effect of decreasing fire frequency by 11.4%. Nevertheless, grazing activity generates mixed results depending on animal species. Specifically, assuming that both bovine and caprine elements per km 2 increase by 1%, the number of fires respectively decreases by 0.18% and increases by 1.16%. We are not aware of similar empirical evidence from other studies. In fact, despite the fact that livestock intensity has been considered a major driver of forest fires (Chuvieco et al. 1999;Leone et al. 2002;Martinez et al., 2009), to our knowledge there is no separate analysis on different livestock compositions. This highlights the need of additional investigation useful to build tailored firecontainment policies.
As for climatic factors, the only variable found to be significant is the average number of consecutive hot days in a year. Indeed, a 1% increase in this variable leads to a 0.24% increase in the yearly frequency of forest fires. This finding is in line with other studies supposing that drier seasons boost fuel flammability, thereby fostering the likelihood of observing fire events (Pinol et al., 1998;Pausas, 2004;Westerling et al., 2006). Both temperature and precipitation do not seem to explain the variation in the total number of forest fires.
Overall, forest fire variation across Italian regions is mostly driven by socio-economic factors rather than by climatic ones. Using the best specification, according to the empirical evidence (RE model), these variables altogether are capable of explaining 81% of the total variability in forest fire frequency.  Pesaran's CD 0.1515 Robust standard errors in parentheses, where *** p<0.01, ** p<0.05, * p<0.1 Table 10 reports results achieved by maintaining regressors unvaried, but using the natural logarithm of the total area burnt as dependent variable. In this case we find that the year effects are jointly significant in the pooled OLS, FE and RE estimations (columns 2 to 4 respectively). The

Unrestricted Model & Logarithm of the Total Area Burnt as Dependent Variable
Hausman test results suggest that the unobserved region specific effects embody elements, which are correlated with regressors. As a result, the FE model is the model of choice. This is in contrast with conclusions previously drawn using the alternative dependent variable. A similarity with the previous model relates to an AR (1) serial correlation in residuals resulting from the Wooldridge test. As a consequence, we propose FE estimates with AR(1) disturbances in column 5.
Similar to the previous model, railway and tourist density positively impact the expansion of the area burnt. Specifically, the railway density shows an impact, which is 5-times larger on this dependent variable than on fire frequency. Conversely to previous results, the use of land here, seems to generate a mono-directional impact, given that the higher the percentage of people working in agriculture, the amount of area covered by fire is smaller. Therefore, in this case more intense land management translates into a greater containment of fire propagation. Although the animal presence is less impacting than above, we find an interesting difference. A bigger bovine population reduces fire frequency but enlarges the total area burnt. Therefore, even though the presence of bovine elements leads to fewer fire events, if the fire takes place, we would experience an increase in the area burnt.
In opposition to the first model, population density plays a role in explaining the total area burnt. Its 1% increase drops the fire expansion by 7.3%. This could be due to a greater amount of land used for population settlement, which reduces the availability and continuity of fuel (Syphard et al., 2007) and leaves limited area for forests and fire propagation.
What deserves specific attention is the fact that while weather patterns do not explain variability in the number of fires, they play an important role in influencing the size of land burnt by fire events.
As before, we report that a percentage increase in the average number of consecutive hot days per year boosts the total area burnt by 0.38%. In addition, an increase in temperature from January to September positively impacts fire extension, although with low statistical significance.
Nevertheless, restricting the analysis to June, July, and August temperature levels notably strengthens the statistical significance. Specifically, an additional degree Celsius in the average temperature in this period is associated with an 18.26% increase in the total area burnt. 14 Additionally, as one could have expected, an increase of one milliliter in the monthly average precipitation between January-April and May-August decreases the total area burnt by 0.41% and 0.33% respectively. These findings are consistent with the branch of literature which supports the relevance of precipitation rates regarding fire size (Pausas 2004;Westerling et al. 2006).
Overall, our findings support Sturtevant and Clelabd (2007) and Pinol et al., (1998) statements that although the socio-economic factors are determinants of the fire frequency, larger fire events are mostly the consequence of biophysical and climatic factors. Concluding, with this dependentvariable specification the within-region variation of the socio-economic and climatic regressors explains 78% of the variability in the total area burnt.

Empirical Findings: Regional Models
There are convincing reasons to think that regions located in different geographical areas in Italy show specific fires patterns. This can be derived, for example, by simply observing the variability in both types of dependent variables. At the same time, we must recall that regional effects have been found to be significant in previous model specifications (see results in Table 9 and 10 on region effects). In order to check the existence of differences in estimated parameters for different geographical regions, we perform a likelihood ratio (LR) test comparing the full sample (unrestricted model) with the sub-samples of the northern, central, and southern regions of Italy (region-specific models). 15 The documented systematic difference between individual subgroup models and the full sample models (Table 11 and 12) supports the need of performing estimates separately for the north, center, and south of Italy. 16 Using two dependent variables for each of the subgroups, we run 6 regressions (2x3). 15 We used the same regressors, which are in the unrestricted model, meaning that the likelihood ratio has a degree of freedom of k=15 16 Note that the log-likelihoods for the full sample and the sub-samples are obtained with the pooled random effects model by using the maximum-likelihood estimator. The degrees of freedom for the likelihood ratio test is the number of parameters (variables) used for all regressions, k=15. In the following tables we report results for the three sub-geographical samples, for the two dependent variables. Only statistically significant variables are reported. Remaining variables, although used as controls, have been discarded from the three models by following the backward elimination technique (for further discussion on the backward elimination procedure see Agresti and Finlay, 1997;Menard, 2002). According to this method, starting from all the variables in the model the non-significant ones will be sequentially removed. In particular, by looking at the variance inflation factor (VIF) and the Pearson correlation coefficient, we eliminate the variables, which are both insignificant and correlated with the significant variables. 17 We are aware that, although widely used within applied econometrics, this approach may be challenged by statisticians. They would claim that a t-test resulting lower than a critical value, may be due to two different reasons: either the value of the coefficient is actually 0 or the empirical evidence is not enough to reject the null hypothesis (i.e., the hypothesis is not "accepted", yet "not rejected"). If we consider the second case, eliminating the variables retained not statistically significant, would entail the risk of producing distorted estimates.
In light of this we argue that maintaining all the regressors in place would have, on the one hand, avoided the aforementioned problems, yet on the other hand, would have produced other inconveniences. Specifically, including also variables whose coefficients result equal to 0 reduces the estimation accuracy, since more degrees of freedom than what is strictly needed are used. This problem is even more relevant when the number of observations is not as high as the number of explanatory variables whose coefficients have to be estimated. This is precisely our case, given that disaggregating the full sample of regions fractionates the total number of observations into three groups. In this respect, to decide how big a sample size should be, relative to the number of observations, the literature offers a ratio measure, which compares the sample size (nominator) with the number of free parameters (denominator). Tanaka (1987) suggests 20 as the rule of thumb while Bentler and Chou (1987) propose the ratio 5:1. Using the second source as a reference (the first appears unrealistically high), our ratios show that we are at the border of sufficiency in terms of sample size. Therefore, we chose to follow the backward elimination approach rather than incurring different, yet significant problems.

Dependent Variable
We develop the same estimation methods (pooled OLS, FE, and RE) to check which model is the best specification for the sub-sample models. When the dependent variable is the natural logarithm of the total number of events from the Hausman and Breusch Pagan tests, the pooled OLS is found to be the best specification for all geographical aggregations. 18 Results are shown in Table 13 where we only report the significant variables. These regressions, therefore, include the variables, which remain after the backward elimination technique mentioned in the previous section.
A common result across regional models, is that both railway density and the presence of caprine elements put pressure on forest land. Other factors influence fire frequency in an expected direction (Table 13). In particular, the presence of bovine animals, the level of education (as reported in de Torres Curth et al., 2012), tourism, and precipitation levels are relevant for the northern region; education and temperature for the center; and precipitation levels as well as extortion activity for the south. Therefore, in southern regions illegal activities represent an additional driver, providing the opportunity to create firefighting jobs or gain land for agriculture, or pasture, which are retained more valuable activities (Leone et al., 2002).

18
The Hausman test suggests using the random effects estimation, meaning that the unobserved heterogeneity across regions is not correlated with explanatory variables. However, a Breusch Pagan LM test suggests that we cannot reject the null hypothesis that the variance of residuals across regions is zero at 10% level of significance. This entails that there is no significant difference across regions within a given geographical classification.

Dependent Variable
We conduct similar regressions and tests as in section 7.1 when the dependent variable is the natural logarithm of the total area burnt and regions are grouped into north, center and south. Once more, for all the geographical aggregations we find that the best specification is the pooled OLS. Given that within geographical groups no unobserved heterogeneity is found, the pooled regressions show similar results for both dependent-variables models (Table 14). 19 For northern regions the statistically significant regressors are mostly the same. There is only one additional significant variable: a percent increase in the bovine elements per km 2 drops the fire frequency by 0.11%. Therefore, we can conclude that our results on the northern area are robust and not sensitive to a change in the dependent variable. 19 The Pearson correlation coefficients between the natural logarithm of total number of events and the natural logarithm of total number of events are 0.87, 0.88 and 0.88 for the northern, central and southern geographical specifications respectively. Despite the differences in magnitude we can reach the same conclusion for the central part of Italy.
Specifically, climatic factors slightly differ in the two models: a percent raise in the average number of consecutive hot days increases the total number of events by 0.61%. Nevertheless, a onemillimeter increase in the average precipitation rate between May and August is associated with around a 1% decrease in the total area burnt. Also, an additional degree Celsius in the average temperature between June and August translates to a 32.79% rise in the total area burnt.
Finally, for the southern part of Italy illegal activities explain the fire frequency, yet they have no impact on the area burnt. In contrast to the northern area, touristic flows positively affect the number of events suggesting a possibly insufficient risk management and forestland protection, or a law-enforcement problem. As for climatic variables, precipitation influences the area burnt with expected signs.

Concluding Remarks and Discussion
Using panel data techniques we analyze the determinants of forest fires frequency and size for the Italian regions during the 2000-2008 period. We first investigate the situation for all the regions pulled together. Then, lead by evidence on their heterogeneity, we specify three regional models for north, center, and south, according to the incidence of forest fires and their geographical location.
We also investigate two dependent variables in each of the abovementioned models: the number of fire events and total km 2 of burnt area regressed on both socio-economic and climatic factors. The most appropriate regression technique for each model (fixed and random effects, pooled OLS, etc.) has been selected through appropriate testing.
Despite the heterogeneity across Italian regions and the existence of locally-specific driving factors, some general conclusions and policy implications can be discussed.
First, improving safety in the railway network is expected to have a supportive impact in reducing fire events and size throughout Italy. Weather patterns also appear to be important determinants.
Additionally, the composition of livestock seems to be relevant to address the problem of forest fire.
In opposition to the negative effect of bovine presence reported in some cases, caprine grazing is an element of pressure on forestland, from the north to the south. In particular, bovine grazing, which has almost no impact on reducing fire frequency, seems to help in containing the spread of fire. This conclusion, normally referred to as the "grazing reduces blazing" argument, is well supported by the literature. Indeed, it is claimed that the introduction of livestock in forest areas functions as a natural undergrowth eliminator, i.e., prevents fire amplification by removing plant biomass (Whalley, 2005). A less obvious effect on fire is generated by the presence of caprine animals affecting positively both fire frequency and extension. This effect may find its explanation in more than one reason. For example, in order to gather minor forests for pasture, shepherds normally introduce nonnative grass increasing fuel loads. Alternatively, they may start small fires to obtain better grazing grass. On the other hand, caprine excessive pasture may imply the consumption of the youngest and greenest components of vegetation thereby increasing the portion of drier, mature, and dead material which is therefore more inflammable (Blasi et al., 2005;Bernetti, 2005). This is especially true in the case of a mineral shortage in soil, also resulting from adverse weather conditions. In fact, this leads animals to strip bark from trees and browse upper branches causing tree deterioration and death, increasing, in turn, fire prone materials. Finally, recent studies recognize that the incidence of grazing on fire may depend on vegetation structure, and grazing intensity (e.g., Leonard et al., 2010). Specifically, in places where part of the vegetation is unpalatable, selective consumption by herbivores may increase the proportion of dead fuel and therefore fire occurrence.
Apart from the mentioned ones, the remaining socio-economic conditions result to be more locally specific. A good analysis requires regional details. By dividing Italy into north, center, and south we find that results from the two dependent-variable specifications are comparable and sometimes homogeneous within the same geographical area. For example, increasing education levels decreases frequency and intensity of fires in the north and the center. As for the south, where the worst Italian burning situation takes place, we acknowledge i) illegal activities as a cause of fire frequency and ii) tourism presence as an element of explanation for the variation in fire size.
In sum, our results suggest that given this concept of "locality" for a driving factor, an accurate and ex-ante definition of fire risk in a specific area becomes crucial. This would help implementing an ad hoc weather forecasting, taking into account seasonality issues, and developing a tailored monitoring and prevention plan. This process entails the consideration of past policies and related results to design more effective actions in the future.
Furthermore, land management, and in particular, grazing and fuel control could concretely contain the incidence of forest fires. Further research is needed to shed light on the unknown impact of different grazing levels on fire. Clarifying this relation would make a valuable contribution to the understanding of forest conservation in Italy.