1 Introduction

Recently, an increasing number of studies have provided alarmist figures related to climate change. In this regard, most of them have estimated that global warming will range between 2.0 °C and 4.8 °C by the end of this century (IPCC 2013). North African countries (NACs) are particularly concerned with climate change, first because their geographical position (close to deserts) makes them very sensitive to temperature increase and rainfall decrease. In this regard, Péridy et al. (2012) show that since the 1970s, temperature has already increased by about 0.5 °C, and precipitation has significantly decreased. The problem is all the more important because these countries remain dependent on agriculture in terms of contribution to both GDP and exports. This reinforces the relevance of the question related to the impact of climate on agriculture in these countries.

The existing literature on this topic is quite extensive. In recent years, an increasing number of articles have been based on the Ricardian model. They state that the land value, which reflects the profitability of land, depends on climate, mainly temperature and rainfall. This approach, initiated by Mendelsohn and Shaw (1994), has been widely applied to a significant number of countries. For instance, Temesgen (2007) shows the negative income impact of temperature increase and precipitation decrease in Ethiopia. Similar results are found by Thapa and Joshi (2010) for Nepal. Some other studies focus on the impact of climate on land productivity and agricultural production (Ajewole et al. 2010 for Nigeria; Kabubo & Kabubo and Karanja 2007 for Kenya; Muchena 1994 for Zimbabwe; Lee et al. 2012 for Asia). Using US data on agricultural output and climate variables, Deschênes and Greenstone (2007, 2012) examine the impact of climate change on production. They conclude that climate change may increase farm profits up to $1.3 billion, depending on the choice of the functional form assumed for climatic and control variables. In this regard, Fisher et al. (2012) show that problems related to the calculation of profit, the choice of climate change projections and the presence of missing or incorrect data, are the main sources of divergence between results and conclusions regarding climate change impacts on agriculture.

The literature dedicated to Mediterranean countries is still scarce. For example, Giannakopoulos et al. (2005) show that the decrease in precipitation in European Mediterranean countries has a detrimental impact on yields. With regard to NACs, Nefzi and Bouzidi (2009) also suggest a negative impact of the decrease in rainfall but they also show that temperature increase may have a positive effect on the yield of specific products. More recently, Ben Zaied and Ben Cheikh (2015) use the panel data cointegration method in order to test the impact of annual temperature and precipitation of two production crops in Tunisia: cereal and dates, over the period 1979–2011. Results show that the long-run effect of temperature on the production of these two crops is generally negative, whereas the effect of precipitation is positive for cereal, but rain shortages in the south affect production negatively. They also test the interaction effect between temperature and precipitation. This test shows that rainfall has a weak positive effect that is counterbalanced by the threat of brutal temperature increases over the last decades.

One characteristic of these studies is that they disregard the spatial interactions of the data. Indeed, when estimating a production function applied to agriculture, one must consider the fact that the production in a region is not independent from the production in other regions (especially neighboring regions). Several reasons can explain this spatial interaction, such as the existence of dams, irrigation, an appropriate organization of farmers and scale economies. Disregarding these spatial interactions of agricultural production may produce estimation biases (Debarsy and Ertur 2010).

One additional lack in the existing literature is that it disregards the existence of dams and groundwater which can modify the impact of climate on agricultural production. For instance, a reduction in precipitation may have both a negative direct effect on production but also a negative indirect effect through the decrease in the stock of water in dams and groundwater.

This article is aimed at providing additional insights into the impact of climate in Tunisian agriculture. Indeed, agriculture is still a major contributor to GDP in Tunisia. For instance, it amounted to 11 % of GDP in 2012, even if this share has slightly declined in the past decade. Agriculture also accounts for 8 % of Tunisian external trade, 10 % of domestic investment and more that 15 % of total employment.

The main contributions of this paper are the following. Using data at the regional level (the 24 administrative regions in Tunisia), both the time and spatial dimensions of the data are explored through spatial panel data analysis. Second, the model includes crucial climate-related variables such as the stock of water in dams and groundwater. This allows the test of both direct and indirect effects of the main climate variables. Finally, the production function is estimated for the main crops in Tunisia, i.e., cereals, olives, citrus fruit, tomatoes, potatoes and palm trees. The advantage of this choice is that it makes it possible to distinguish specific effects of climate on irrigated and non-irrigated crops.

It is organized in 4 sections. Section 1 provides key features about agricultural production in Tunisia at the regional level. It also proposes a spatial exploratory analysis of the data through global and local spatial autocorrelation tests. Section 2 is dedicated to the presentation of the model and data. In this regard, a Cobb-Douglas production function is developed for Tunisian agricultural products based on regional data. In order to take into account the spatial interaction of production, and the time dimension of the data, several spatial panel data models are developed, especially Spatially Autoregressive Regressions (SAR) and Spatial Durbin Models (SDM). We discuss results in Section 3 for both irrigated and non-irrigated crops. Particular attention is paid to the spatial spillovers of the variables. Section 4 concludes and discusses the policy implications of the results.

2 Key features and spatial exploratory analysis

The main crops in Tunisia include both irrigated and non-irrigated crops. These are mainly tomatoes, cereals and olives (non-irrigated) as well as tomatoes, potatoes, citrus fruit and palm trees (irrigated) (Fig. 1).

Fig. 1
figure 1

Breakdown of the main crops in Tunisia (% of total agricultural production, average 1990–2012)

Full size image

The breakdown by regional areas shows that most crops are spatially concentrated (Fig. 2a and b). For instance, the production of cereals is mainly located in the North and Northwest areas (Beja, Jendouba, Siliana, Bizerte, Zaghouan and Kef). In these areas, the quantity of cereals produced has almost doubled in the past decades. As another example, the production of tomatoes is mainly located in the North and Northeast areas. Two regions (Nabeul and Kairouan) account together for 75 % of the total amount produced in 2012 (refer to Annex 1 for the map of regional areas in Tunisia).

Fig. 2
figure 2

a: The production of cereals in Tunisia. b: The production of tomatoes in Tunisia

Full size image

The spatial dimension of the geographical localization of agricultural production can be further investigated through the global correlation statistics (Moran statistics). This test makes it possible to identify a spatial correlation of the production in Tunisia. Annex 2 shows additional details on the computation method. Table 1 shows the test value (I), the standard deviation (Sd(I)) as well as the significance statistics Z. It is striking to observe that for all agricultural products, the Moran test is positive and significant. This means that agricultural production is positively spatially correlated in Tunisia. In other words, neighboring areas that have high production surround regional areas that also have high production.

Table 1 Global spatial correlation test (Moran)
Full size table

This correlation can be explained by several reasons such as the land quality and appropriateness for a specific production, common climate characteristics for neighboring areas, an appropriate organization of farmers, and scale economies.

One drawback of the Moran test is that it is a global test that does not say anything about the local structure of this spatial correlation. One way to overcome this problem is to use the LISA index (Local Indicator of Spatial Association) that makes it possible to identify four possible cases:

  • HH: a region with a high production is surrounded by other regions with high production

  • LL: a region with a low production is surrounded by other regions with low production

  • HL: a region with a high production is surrounded by regions with low production

  • LH: a region with a low production is surrounded by regions with high production

In the first two cases (HH and LL), the local autocorrelation is positive, whereas in the other cases (HL and LH), the local correlation is negative. Figure 3 provides the maps corresponding to the LISA index for each agricultural production. Hence, the LISA index goes further than the Moran test since it identifies the local characteristics of spatial associations (refer to Annex 2 for the computation method).

Fig. 3
figure 3

The local spatial correlation of Tunisian agricultural production (LISA index)

Full size image

Interestingly, the local test suggests that the local spatial correlation is generally positive. This result correlates with the global test. This conclusion is particularly striking for cereals, for which high-production areas in the North are surrounded by neighboring high-production areas. Similarly, low-production areas in the South are surrounded by neighboring low-production areas. This can be mainly explained by differences in climate conditions between the North (with enough rainfall) and the South (dry), as the cereal production is not irrigated. In other words, there is a climate border between the Northern part of Tunisia on the one hand, characterized by a Mediterranean climate for which precipitation is appropriate for cereal, and the South on the other hand, characterized by a semidesert climate which does not suit the production of cereal. These physical characteristics explain why Northern areas with high cereal production levels are surrounded by similar high-production areas, whereas Southern areas with low production levels are surrounded by similar low-production areas. A positive local correlation is also found for palm trees, olives and, to a lesser extent, citrus fruit. Again, this may be explained by the fact that there are climatic areas that are appropriate for these crops, even if other climatic areas are not appropriate. However, the local spatial correlation is negative for potatoes and tomatoes. One reason is that the production of these irrigated products is very much geographically concentrated. This explains that these regions with high production are surrounded by regions with low production.

3 The model and data

This section establishes a Cobb-Douglas production function for Tunisian agricultural products based on regional data. In order to take into account the spatial interaction of production and the time dimension of the data, we develop several spatial panel data specifications.

The model proposed here is close to that used in Bsaies and Mokadem (2009). However, instead of using CO2 emissions as a proxy for climate change, we directly use data on temperature and precipitation. This makes it possible to investigate more precisely the impact of these climate variables on each crop in each region for each year.

The basic Cobb-Douglas production function can be written as follows:

$$ {\mathrm{Y}}_{\mathrm{it}}=\mathrm{F}\left({\mathrm{K}}_{\mathrm{it}},{\mathrm{L}}_{\mathrm{it}},{\mathrm{T}}_{\mathrm{it}},{\mathrm{P}}_{\mathrm{it}}\right)={\mathrm{A}}_{\mathrm{it}}{{\mathrm{K}}^{\alpha}}_{\mathrm{it}}{{\mathrm{L}}^{\beta}}_{\mathrm{it}}{{\mathrm{T}}^{\theta}}_{\mathrm{it}}{{\mathrm{P}}^{\gamma}}_{\mathrm{it}}, $$
(1)

where Yit denotes the production of a given agricultural product in region i at year t, Kit and Lit respectively reflect the capital and labor used in agriculture, Tit and Pit correspond respectively to the average temperature and precipitation, whereas Ait includes unobserved variables (such as land quality, labor skills, and technical progress) that may influence the agricultural production.

These unobserved effects can be considered in the following equation:

$$ {\mathrm{A}}_{\mathrm{i}\mathrm{t}} = {{\mathrm{A}\mathrm{e}}^{\mathbf{u}}}^{\mathrm{i} + \mathbf{v}\mathrm{t} + \eta \mathrm{it}} $$
(2)

With ui: regional specific effect : vt time specific effects; ηit: standard error term.

Substituting (2) into (1) provides:

$$ {\mathrm{Y}}_{\mathrm{i}\mathrm{t}} = {\mathrm{Ae}}^{\mathbf{u}}{{{}^{\mathrm{i} + \mathbf{v}\mathrm{t} +}}^{\eta}}^{\mathrm{i}\mathrm{t}}{{\mathrm{K}}^{\alpha}}_{\mathrm{i}\mathrm{t}}{{\mathrm{L}}^{\beta}}_{\mathrm{i}\mathrm{t}}{{\mathrm{T}}^{\theta}}_{\mathrm{i}\mathrm{t}}{{\mathrm{P}}^{\gamma}}_{\mathrm{i}\mathrm{t}} $$
(3)

After log transformation, the final basic model is:

$$ \mathrm{Log}\ {\mathrm{Y}}_{\mathrm{i}\mathrm{t}} = \mathrm{Log}\ \mathrm{A} + \alpha \mathrm{Log}\ {\mathrm{K}}_{\mathrm{i}\mathrm{t}}+\beta \mathrm{Log}\ {\mathrm{L}}_{\mathrm{i}\mathrm{t}}+\theta \mathrm{Log}\ \mathrm{Tit} + \gamma \mathrm{Log}\ {\mathrm{P}}_{\mathrm{i}\mathrm{t}} + {\mathrm{u}}_{\mathrm{i}} + {\mathrm{v}}_{\mathrm{t}}+{\eta}_{\mathrm{i}\mathrm{t}} $$
(4)

This panel data model can be specified to take into account irrigated crops. In this case, we must include the impact of two additional variables, i.e., groundwater and dams. The effects of these variables on production crops can be considered either directly or combined with temperature and precipitation. For instance, a reduction in precipitation is likely to reduce the water stock in dams and the stock of groundwater that in turn will impact irrigation and thus production.

Hence, the model applied to irrigated crops with both direct and indirect effects of dams and groundwater is the following:

$$ Log\ {Y}_{it} = Log\ A + {\alpha}_1 Log\ {K}_{it}+{\alpha}_2 Log\ {L}_{it}+{\alpha}_3 Log\ {T}_{it}+{\alpha}_4 Log\ {P}_{it}+{\alpha}_5 log{D}_{it}+{\alpha}_6 logD{T}_{it}+{\alpha}_7 LogD{P}_{it}+{\alpha}_8 Log\ {W}_{it}+{\alpha}_9 LogW{T}_{it}+{\alpha}_{10} Log W{P}_{it} + {u}_i + {v}_t+{\eta}_{it} $$
(6)

With:

D:

Gross water stock in dams

DT:

The indirect effect of temperature on production via dams

DP:

The indirect effect of precipitation on production via dams

W:

Gross water stock in groundwater

WT:

The indirect effect of temperature on production via groundwater

WP:

The indirect effect of precipitation on production via groundwater

It is also worthwhile to test a modified specification that includes the combined effect of temperature and precipitation (Log (P*T)). This interaction effect makes it possible to test whether the effect of one variable can cancel the effect of the other variable, as suggested by Ben Zaied and Ben Cheikh (2015).

The measurement of the variables and the statistical sources are the following: Data concerning economic variables (production, labor, capital) are derived from national sources: Commissariats Régionaux au Développement Agricole (CRDA)Footnote 1 and Institut Tunisien de la compétitivité et des études quantitatives (ITCEQ).Footnote 2 Production is measures in tons. Labor and capital are measured as the employment for each crop and the value of physical capital (including transport equipment) in local currency, respectively. Agricultural data are provided by the Institut national de la statistique (INS),Footnote 3 the Observatoire national de l’agriculture (ONAGRI)Footnote 4 and the Ministère de l’agriculture et ressource hydrique (MARH).Footnote 5 Finally, climatic data are derived from the Institut national de la météorologie (INM).Footnote 6

Additional information about data summary is also provided in Annex 3. In particular, data on groundwater use water stocked in all 24 regions. However, dams are available in 11 regions only (Béja, Bizerte, Ben Arous, Jendouba, Kairouan, Nabeul, Zaghouan, Siliana, El Kef, Gabes and Gafsa). These dams play a crucial role since they can be used for irrigation not only in these regions but also in neighbor regions. For instance, the dam located in the Zaghouan region can be used for irrigating crops in the Sousse region. This means that spatial interactions and spillovers are existing and must be taken into account in our model.

In order to consider the spatial interaction of the data, two spatial econometric models are estimated. The first is a SAR model applied to panel data. This model, also called the spatial lag model, can be written as:

$$ {\mathrm{Y}}_{\mathrm{i}\mathrm{t}}\delta \kern0.5em {\displaystyle {\sum}_{\mathrm{j}=1}^{\mathrm{N}}{\mathrm{w}}_{\mathrm{i}\mathrm{t}}{\mathrm{y}}_{\mathrm{j}\mathrm{t}}}+{\mathrm{x}}_{\mathrm{i}\mathrm{t}}\upbeta +{\mathrm{u}}_{\mathrm{i}}+{\mathrm{v}}_{\mathrm{t}}+{\upeta}_{\mathrm{i}\mathrm{t}} $$
(7)
with Yit :

agricultural production for product i at year t

∑ N j = 1 wijyjt :

agricultural production of neighbor regions

xit :

vector of independent variables

wij :

spatial weight matrix (nearest neighbor matrix)Footnote 7

δ measures the intensity of the spatial interaction between the regional units. If it is positive, this suggests a positive spatial correlation of the production between the regional units.

The second model is a Spatial Durbin Model (SDM) which considers that the spatial interaction not only concerns the dependent variable, but also the independent variables:

$$ {\mathrm{Y}}_{\mathrm{i}\mathrm{t}}=\delta \kern0.5em {\displaystyle {\sum}_{j=1}^N{\mathrm{w}}_{\mathrm{i}\mathrm{j}}}{\mathrm{y}}_{\mathrm{i}\mathrm{t}}+{\mathrm{x}}_{\mathrm{i}\mathrm{t}}\upbeta +{\displaystyle {\sum}_{j=1}^N{\mathrm{w}}_{\mathrm{i}\mathrm{j}}{\mathrm{x}}_{\mathrm{i}\mathrm{t}}\upgamma}+{\mathrm{u}}_{\mathrm{i}}+{\mathrm{v}}_{\mathrm{t}}+{\upeta}_{\mathrm{i}\mathrm{t}} $$
(8)

In this case, ∑ N j = 1 witxjt is the vector of the independent variables in neighbor regions and γ measures the intensity of the spatial interaction due to the independent variables.

It must finally be noted that for:

H0 :

γ = 0, then the SDM (8) is equivalent to SAR (7)

H0 :

γ + δβ = 0, then the SDM is equivalent to a SEM (spatial error model) where errors are spatially correlated.

However, if we reject both γ = 0 and γ + δβ = 0, then the SDM specification is the most appropriate.

Whatever the choice between these three models (SAR, SEM or SDM), the OLS estimator cannot be implemented. Specific estimators must be used. In this paper, the Maximum Likelihhod estimator used by Elhorst (2003) will be chosen.

Data are based on Tunisian national statistics applied to the 24 regional territories (gouvernorats). These include cereals, citrus fruit, olives, potatoes, tomatoes and palm trees. The time period considered ranges from 2000 to 2012, except for palm trees (2000–2008) due to data limitation. The number of observations available ranges from 45 to 312 depending of the product considered. Preliminary tests on panel unit roots are provided in Annex 3. Using the Breitung test (2000), results show that most variables are stationary. Those that are not stationary are integrated of order 1. Basically, preliminary estimations of the variables in level or in first difference show very little difference in terms of sign and significance of the parameter estimates. Beyond that, the following estimations focus on spatial interaction of the variables in order to capture the spatial autocorrelation of these variables.

4 Estimation and results

The choice of the appropriate spatial panel estimator can be made through LM tests (Anselin 2008; Elhorst 2003) applied on the error terms (LMρ) and the spatial lags (LMδ). Table 2 shows these various tests applied to both SAR and SDM models.

Table 2 Results of LM tests
Full size table

Basically, the LMρ and LMδ tests generally reject the hypothesis of no spatial correlation for both error terms and spatial lags. In addition, the LM SAC tests are also significant. This leads us to reject the hypothesis of the absence of global autocorrelation. In addition, the tests applied to the SDM specifications are significant when applied to random effects, with the exception of olives where it is significant for fixed-effects.

All these tests suggest that the SDM signification is the most appropriate, as they show that the spatial correlation not only concerns the dependent variable, but also the independent variables. Given the results of the tests applied to FE and RE models, the final specification considered is SDM-RE for cereals, tomatoes and citrus fruit, whereas the SDM-FE seems more appropriate for olives.

Table 3 shows the estimation results for non-irrigated productions (cereals and olives) and Table 4 focuses on irrigated plants. Due to data limitations, results concerning palm trees and potatoes are not presented. In both tables, results are presented with the appropriate selected estimator (SDM-FE or SDM RE). However, as a sensitivity analysis, other estimators are also presented (a-spatial FE and RE models as well as SAR-FE and SAR-RE models).

Table 3 Results for non-irrigated crops
Full size table
Table 4 Results for irrigated crops
Full size table

Table 3 shows that the production of cereals significantly depends on capital and labor: a 1 % increase in these factors leads to a production increase by respectively 0.51 and 0.13 % using the SDM-RE specification. In addition, production is very sensitive to rainfall. As a matter of fact, a 1 % decrease in precipitation leads to 0.79 % decrease in production. However, cereal crops seem to be less sensitive to temperature. Indeed, although negative, the corresponding parameter estimate is barely significant. This can be explained by the fact that the harvest period for wheat generally ends in June. This is why production is less affected by severe summer heat waves that could be very much detrimental otherwise. It is interesting to observe that all the results are very robust whatever the model chosen. Table 3 also shows that a rise in precipitation or a decrease in temperature in neighboring regions leads to a rise in the production of cereals in a given region (variables W1x-P, WIx-T). This can be explained by the correlation between temperature and precipitation in one region and its neighbors.

Results concerning olives are not very different with regard to climate variables. For example, olive crops are very sensitive to precipitation (a 1 % decrease in rainfall gives rise to a 1.01 % decrease in production in the SDM-FE specification). The temperature variable is barely significant. Interestingly, when significant, the parameter estimate is positive. This suggests that olive trees like warm weather conditions and thus are not very sensitive to global warming. Moreover, the impact of the climate variables related to neighboring regions is similar (a rise in precipitation and a rise in temperature in the neighboring regions lead to an increase in the production of olives in the region considered).

An increase in capital has a positive effect on olive production but unexpectedly, labor seems to have a limited but negative impact for some specifications. This result seems to be unexpected at first sight. However, one explanation is that harvesting olives needs specific skills that workers generally lack in Tunisia, especially because they need appropriate training. In addition, workers often use inappropriate or poor quality working tools which can also be detrimental to production. Finally, workers are generally employed in farms where olives are not the main production. As a result, workers who harvest olives are less efficient than if they were to harvest the main crops (cereal). This points out the very low labor productivity for olives and the necessity for improvement in the tools used by the workers as well as for improvement in their skills.

With regard to irrigated crops, Table 4 shows that an increase in capital, precipitation and stock of water in the ground have a positive impact on the production of tomatoes. Identically, the combined impact of rainfall and groundwater is also positive (W*P). This suggests that the existence of groundwater is more favorable to the production of tomatoes in a given region when precipitation is not too low in this region. Looking at the impact of temperature, it is strongly negative in the SDM-RE model. The impact of labor is also negative. Again, this raises the question of the appropriate tools and skills of workers. Finally, the impact of dams (direct or indirect) is insignificant. To sum up, the production of tomatoes in Tunisia is strongly dependent on weather conditions, as a 1 % decrease in precipitation or a 1 % increase in temperature leads to a decrease in production by 7.1 and 16.7 % respectively. The impact of groundwater is positive but also depends on precipitation.

Results concerning citrus fruit are quite similar, i.e., production is very dependent on the direct effects of rainfall and temperature. The impact of groundwater is also positive (direct effect) but this positive impact is strongly reduced when precipitation decreases or when temperature increases (indirect effect). The impact of dams is generally less significant. It is strongly reduced when temperature increases (indirect effect).

As a sensitivity analysis, Table 5 tests the interaction effect P*T in order to check whether the impact of temperature can be offset by the impact of precipitation and vice-versa.Footnote 8 Results show that this combined effect is positive and statistically significant for cereals and olives (non-irrigated crops). This result suggests that the positive effect of precipitation on climate is more important than the negative effect of temperature. This predominant effect of precipitation correlates the results found in Table 3, because the magnitude of the parameter estimate for precipitation is more important than that for temperature (in absolute value). Basically, these results show that the negative effect of global warming can be reversed in the case where precipitation is increasing. However, if global warming is also characterized by a fall in precipitation (which is expected by IPPC 2013 in the case of Tunisia), these two negative effects reinforce each other. We will get back to the policy implications of these results in the conclusion.

Table 5 Testing the interaction effect temperature*precipitation
Full size table
Table 6 The direct and indirect effects of the variables
Full size table

However, the interaction effect on irrigated crops (tomatoes and citrus fruit) is not significant. This tends to show that for irrigated crops, the precipitation effect is not predominant, and that the effects of temperature and precipitation cancel each other out. However, if global warming is coexixting with a decrease in precipitation, these two effects are individually significant and lead to a strong decrease in production as shown in Table 4.

A final set of results is derived from the calculation of the spatial spillover effects of the variables on the neighboring regions. Indeed, the indirect effects calculated previously only reflect the impact of one variable (e.g., groundwater) via a different variable (e.g., precipitation) in the same region. Investigating the impact of a variable in region i via a different variable in region j (neighboring regions) can be done through the results of SDM models. In this regard, considering xitβ + ∑ N j = 1 witxjtγ in Equation (8) makes it possible to calculate these indirect spatial spillover effects.

Table 6 summarizes the direct and indirect impact of climate variables on region i and neighboring regions, respectively.

The most striking features that emerge from this table are first, that the direct and indirect impacts of dams are positive for tomatoes. This means that an increase in the water stock in a dam located in region i makes it possible to increase the production of tomatoes in this region and also the neighboring regions (spillover effect). These effects are, however, undermined by the combined effect of temperatures/dam (D*T). As a matter of fact, a rise in temperature, by influencing the quantity of water in the dam, has a negative impact on the production of tomatoes, both in region i and the neighboring regions. Another interesting spillover effect can be found through the combined effect of groundwater and precipitation. Indeed, an increase in the groundwater stock made possible by the rise in precipitation in region i has a positive effect on the production of tomatoes in neighboring regions.

5 Conclusion and policy implications

This article shows at regional level that the Tunisian agriculture is strongly dependent on the direct effects of temperature and precipitation for all the products considered. These results are basically in line with related studies, especially Ben Zaied and Ben Cheikh (2015). However, our results go beyond that study in several aspects. First, the crops taken into account cover not only cereal but also olives, tomatoes, and citrus fruit, as well as potatoes and palm trees. This makes it possible to distinguish specific effects of irrigated and non- irrigated crops. The inclusion of direct and indirect effects of dams and groundwater also provides interesting new results for irrigated crops. In particular, we show that dams and groundwater have generally a positive effect on agricultural production for irrigated crops. However, this impact is still considerably lessened in case of detrimental climate conditions (increase in temperature or decrease in precipitation).

Finally, the spatial analysis also provides some new results. For example, we show that an increase in the water stock in a dam located in region i makes it possible to increase the production of tomatoes in this region but also the neighboring regions (spillover effect). Similarly, another interesting spillover effect is that an increase in the groundwater stock made possible by the rise in precipitation in region i has a positive effect on the production of tomatoes in neighboring regions.

Given that recent studies have shown that in past decades, climate change has already started in Tunisia with a significant decrease in rainfall and an increase in temperature, the results found in the present article raise the question of the sustainability of the growth in agricultural production in Tunisia. In this regard, appropriate policies can be implemented in order to lessen the impact of climate on production. These include water saving, research on less water-demanding agricultural varieties, specialization on less water- demanding products (such as dates and olives), improvement of labor skills and capital efficiency, appropriate fiscal and environmental policies, and the development of regional policies dedicated to improve spillover effects related to dams and groundwater.

Such policies should be differentiated at regional level depending on climate characteristics and the climate impact on the crops highlighted above. For example, in Southern areas characterized by a semi-desert climate appropriate for olives or palm trees, specific policies dedicated to the implementation of more drought-resistant seeds must be implemented as well as water-saving policies. In Northern areas, which are suited for the production of cereals but also irrigated crops (potatoes, citrus fruit, tomatoes), more efficient water-management policies should be promoted as an adaptation policy. In addition, appropriate infrastructure should be developed in order to deliver water to central areas. As another example, land that is currently not used for agricultural purposes in the North should be dedicated to agricultural production in order to take advantage of significant precipitation.

Special policies should also be promoted in central areas that are located around the climatic border between Mediterranean and semi-desert climate areas, as production in this central region (cereal, olives and potatoes) can be affected by strong climate variations (especially droughts). This requires adaptation toward more irrigation or the use of water-saving seeds.