Impact Of Rural Development Policy And Less Favored Areas Scheme: A Difference In Difference Matching Approach

Transcription

1 Impact Of Rural Development Policy And Less Favored Areas Scheme: A Difference In Difference Matching Approach Salvioni Cristina and Sciulli Dario salvioni@unich.it Paper prepared for presentation at the EAAE 2011 Congress Change and Uncertainty Challenges for Agriculture, Food and Natural Resources August 30 to September 2, 2011 ETH Zurich, Zurich, Switzerland Copyright 2011 by [Salvioni and Sciulli]. All rights reserved. Readers may make verbatim copies of this document for non-commercial purposes by any means, provided that this copyright notice appears on all such copies.

2 Impact Of Rural Development Policy And Less Favored Areas Scheme: A Difference In Difference Matching Approach 1. INTRODUCTION The interest in impact assessment of agricultural and rural development policies is progressively growing. This is partly due to the increasing competition in the use of diminishing public funds. A key issue in policy evaluation is the establishment of a baseline or counter-factual scenario to determine additionality, i.e. the additional net impact that particular policy measures have had on a variable of interest. Matching methods can provide a tool to identify whether significant and causal differences in outcome variables occur between farms receiving and those not receiving a subsidy. In an impact assessment study, simply comparing mean outcomes may not reveal the actual treatment effect, as participants and non-participants typically differ even in the absence of treatment. A few empirical studies have been looking at the impact of farm and rural policy measure controlling for the non-random assignment of subjects to treatment, and the selection bias (Lynch and Liu, 2007; Lynch, Gray, and Geoghegan, 2007; Pufahl and Weiss, 2009; Chabé-Ferret and Subervie, 2010). Our paper aims to contribute to this literature by providing a micro perspective on the impact on structural characteristics and economic performance of farms participating in the first Italian RDP as a whole, and the impact produced by participation in the Less Favoured Areas (LFAs) scheme. The remainder of the paper is organized as follows. Section two provides overview about the EU RDP and LFAs scheme and their implementation in Italy. Section three presents the data and the PSM method, i.e. the semi-parametric econometric approach used to compare the performance of farmers participating and non-participating to the first ( ) Italian RDP and to LFAs scheme by accounting for their inherent differences. Section four presents the estimation results and in section five conclusions are drawn. 2. BACKGROUND 2.1. The EU s Rural Development Policy Rural Development (RD) policy aims at a) improving the competitiveness of agriculture and forestry by encouraging farmers to structural changes (Axis 1); b) improving the environment and the countryside (Axis 2); c) improving the quality of life in rural areas and encouraging diversification of economic activity (Axis 3). More in general, emphasis has been put on the potential of Rural Development (RD) measures to contribute to the Lisbon Strategy (growth and employment) and to the Goteborg Strategy (sustainable development). The RD policy framework offers a menu measures. Member States choose from this menu those measures that suit the needs of their rural areas best. These are then included in their national or regional programmes. In Italy, 51 different regional programmes were prepared. The Centre-North Regions had one RD programme for rural development measures funded mainly through Pillar 2 of the CAP. They may, in principle, contain all the rural development measures.

3 In the South, that is in Objective 1 regions, the RD programmes cover only the 8 accompanying measures (early retirement, less favoured areas, agri-environment, afforestation of agricultural land, 2 quality measures and 2 meeting standards measures) while the remaining measures are integrated into the Objective 1 programmes, that is into the Regional Operational Programmes (POR) under the Community Support Framework. The financial resources of the 49 different Italian RDPs were mainly concentrated on the measures aiming at enhancing agricultural competitiveness 1 and those directed at preserving and, where possible, enhancing the environment, biodiversity and the amenity value of the countryside. By contrast, policy measures devoted to measures promoting non agricultural rural development, for example diversification in non farming activities and rural infrastructure development, covered 10% of total resources The LFAs support scheme In many areas across Europe, agricultural productivity is economically and geographically marginalised due to natural disadvantages. Such areas are defined as Less Favoured Areas (LFAs), this term identifies an area with natural handicaps (lack of water, climate, short crop season and tendencies of depopulation), or that is mountainous or hilly, as defined by its altitude and slope. Due to the handicap to farming there is a significant risk of agricultural land abandonment and thus a possibility of loss of biodiversity, desertification, forest fires and the loss of highly valuable rural landscape. High Nature Value (HNV) farming systems often occur in LFAs, they are characterised by traditional agricultural practices that maintain countryside features and support high levels of biodiversity. As a consequences, the LFAs compensatory allowances have been one of the first measures of the Common Agricultural Policy aimed to address the environmental benefits associated to traditional or extensive farming systems, and to ensure the provision of public goods, mainly ecosystem services, that would otherwise be under provided or disappear. The LFAs policy, originally conceived in the 70s as a structural policy with the aim to contribute, through continued use of agricultural land, to maintaining the countryside as well as to maintaining and promoting sustainable farming systems (EC No. 1698/2005), is now part of Axis 2 of the Rural Development Policy along with agrienvironmental payments. A considerable part of the utilised agricultural area in the EU-27 is currently located in regions where conditions are difficult for this activity, Despite the wide percentage of surface designated as LFAs 2, only a limited proportion of farmers benefit from a compensatory allowance. In 2005 approximately 1.4 million farms, representing about 13% of the total number of farms in the EU25, received support under all LFAs schemes. The financial support to LFAs amounted to 8 billion, approximately 18 % of the Community funding for Rural Development for In Italy the LFAs, according to the definition used in the RDP, cover 61% of the territory and they are mainly located in mountain areas (70%). The financial 1 Investments in agricultural holdings (12.6%), the setting-up of young farmers (5.8%), improvement of processing and marketing of farm products (6.5%). 2 Around 15% in mountains, 35% in areas in danger of abandonment of land use and around 3% in areas affected by specific handicaps.

4 resources devoted to these compensatory allowances amounted to the 6.7% of total RDP expenditures. 3. DATA AND METHODOLOGY 3.1. Data The analysis is based on a panel of 3159 commercial family 3 Italian farms drawn from the Farm Accountancy Data Network (FADN) sample 4. For the analysis we used data from 2003, the first year in which the information started to be collected for a representative sample, and 2007, the latest year for which information are currently available. We drew a 5 waves balanced panel of farms containing only those holdings for which information were collected in both 2003 and Given the available data and the model requirements, the 2003 wave information are used to define the pre-treatment control variables, while the 2007 wave information are used to define our outcomes (measured at the end of the year). Finally, the waves are used to identify farms receiving or not the RDP or LFA payments. More in detail, in the dataset 341 farms (corresponding to 13.32% of total observations) received at least a RDP payment over the period. These farms represent the treatment group. As for the outcome variables, we focussed our attention on those variable that capture the ability of RDP to contribute to the overall goals of increasing rural employment and economic growth (Lisbon Strategy). More in detail, in order to check whether the RD policy induced a structural change effect in the group of treated farms we selected as outcome variables total and familiar annual working units (respectively AWU and FAWU) and total and utilized agricultural area (respectively TAA and UAA). In addition, in order to check if the policy improved the economic performance of treated farms we evaluated the ATT on labour productivity and on profits per family working unit. The first indicator refers to the value added 5 net of RDP payment per working unit. As for the LFAs scheme, we have taken into account those farms that are eligible for compensatory allowances, that is those farms located in LFAs. The sample contains 1343 farms, of which 140 were treated, i.e. received a compensatory allowance. Given the aim of this measure is maintaining extensive farming systems in these areas, In addition to check the impact on farm employment and income, in the LFAs case we also check the impact on farm total output per hectare in order to evaluate whether payments contributed to maintain and promote the use of extensive farming system or, on the contrary, if they were associated to intensification of farming. Tables 1 and 2 inform about the descriptive statistics respectively in the RDP and LFAs samples. 3 We define sole ownership farms as family farms. This is consistent with what usually done by DGAGRI. 4 The Italian FADN survey started to be conducted on statistically representative sample drawn from the census in The sample is stratified according to criteria of geographical region, economic size (ESU) and farm type (FT). The field of observation is the total of commercial farms, that is farms of size greater than 4 ESU (4,800 euro). 5 Farm net value added is the sum which is available to reward all of the factors of production, that is, all the labour, land, and capital used on the farm, irrespective of who owns them.

5 The data reported in table 1 reveal that the mean differences in outcome variables of non participants in RDP are lower than those related to participants. More in detail, the comparison suggests that RDP induced a change in structures and an improvement in the economic performance of recipients farms. The comparison of mean differences of treated and untreated farms in the LFA sample (table 2) gives the impression that the participation to this measure produced an impact on total and family labour. Table 1: Descriptive statistics in the RDP sample Treated (obs. 341) Untreated (obs. 2220) T-test Type Variables Mean Std. Dev. Mean Std. Dev. Outcomes ( ) Covariates (2003) D FAWU ** D AWU D Profit/FAWU ratio *** D UAA * D TAA *** D Added value*/awu Age of the operator *** Male operator ** North-West North-East *** Centre * South Islands * Plane * ESU < FT olive FT wine FT field crops *** FT citrus FT livestock *** Environ. protected areas *** Pluriactivity *** Source: own elaboration on FADN data Note: * added value net of the payment

6 Table 2: Descriptive statistics in the LFA sample Treated (obs. 140) Untreated (obs. 1203) T-test Type Variables Mean Std. Dev. Mean Std. Dev. D FAWU *** D AWU * D Profit/FAWU ratio Outcomes D UAA ( ) D TAA D added value* D Output D Output /AUU Covariates (2003) Age of the operator *** Male operator North-West ** North-East Centre *** South ** Islands *** Plane Mountain *** ESU < FT olive FT wine FT field crops *** FT citrus *** FT livestock *** Environ.l protected areas Pluriactivity *** Diversified farms Source: own elaboration on FADN data Note: * added value net of the payment 3.2. The model We are estimating the causal effect of a payment from RDP on various firms outcomes (AWU, FAWU, UAA etc.). Ideally, we like to compare the outcomes of firms receiving PSR (the treatment group) to the same firms not receiving a payment (the control group) to determine the average treatment effect (ATEj): ( j D = 1 Y j D = 0) = E( Y j D = 1) E( Y j = 0) ATE j = E Y D (1) where the subscript j indicates the 2007 outcomes analysed (AWU, FAWU, Profit/FAWU ratio, UAA, TAA and Corrected added value). (Y1j D=1) is the outcome of treated Y1j if firm has received a payment (D=1), and (Y0j D=0), the outcome of untreated (Y0j) if firm has not received a payment from RDP (D=0).

7 However, as we can observe each firm only in one state, the outcomes for treated had they not been treated is an unobserved counterfactual. To solve this puzzle, microeconometricians proposed to estimate the average treatment effect on the treated (ATTj): ( Y D = 1) = E( Y D = 1) E( Y = 1) ATTj = E Yj j j j D (2) That is, the mean effect of receiving a payment from RDP rather than not on the firms that received a payment from RDP (the impact of treatment on the treated). In any case, Y0j D=1 is not observable and, as Becker and Ichino (2002) underlined, since in observational studies assignment of subject to the treatment and control groups is not random, the estimation of the effect of treatment may be biased because of the existence of confounding factors 6. An unbiased estimate of ATT can be obtained if treatment satisfies the Conditional Independence Assumption (CIA): ( Y D) X 0 (3) The outcome of untreated is independent of the treatment conditional on some set of observed covariates X. In other words, according to CIA, conditioning on a suitable set of covariates, it is possible to remove all systematic differences in outcomes in the untreated state. Unfortunately, there may be systematic differences between treated and untreated outcomes, even after conditioning on observables, because of unobservable factors and/or level differences in outcomes. To solve these problems, Heckman, Ichimura and Todd (1997) suggest a conditional difference-in-difference matching estimator (CDID) 7, for which both before and after treatment outcome information is used. Specifically, let t 1 represent a time period after the treatment start date and t 0 a time period before the treatment. The CDID (see Pufahl and Weiss 2009 for an application) compares the conditional before and after outcomes of treated with those of untreated: E ( Y Y D = 1, X ) E( Y Y D 0 X ) t t t t, = Rosenbaum and Rubin (1983), to reduce the estimation bias in the estimation of treatment effects with observational data, proposed the Propensity Score Matching (PSM) method. PSM method has two main advantages when compared with standard econometric techniques. First, it preserves us from making strong assumptions on functional form, like linearity and additivity of regressors, which characterize standard econometric models. Second, PSM is based on the idea that the bias is reduced when the comparison of outcomes is performed using treated and control firms who are as similar as possible. This is allowed applying the matching procedure based on the propensity score, i.e. the conditional probability of receiving a treatment given pre-treatment characteristics: 6 ATT corresponds to the ATE only if the occurrence of conviction is unrelated to outcomes. 7 While CDID solves the problem of time-invariant unobservable factors, time variant unobserved heterogeneity possibly remains unidentified.

8 ( X ) Pr ( D = 1 X ) = E( D X ) p (4) When observations with the same propensity score have the same distribution of observable characteristics independently of treatment status 8, the balancing property is satisfied 9 and, hence, the common support condition holds (Caliendo and Kopeinig, 2008). Moreover, satisfying the balancing property means that exposure to treatment may be considered to be random and therefore treated and control units should be on average observationally identical (CIA or selection on observables). To better examine the common support condition the propensity scores of the groups examined are plotted in Figure 2. In the graph, the top histogram reports observations that received a payment from RDP, while the bottom histogram represents those not receiving a payment from RDP. The horizontal axis defines intervals of the propensity score and the height of each bar on the vertical axis indicates the fraction of the relevant sample with scores in the corresponding interval. Fortunately, the figure shows that in all cases the overlapped region is quite wide and it is not needed to eliminate a relevant number of observations. Figure 2. Propensity score histograms by treatment status RDP LFA Propensity Score Propensity Score Untreated Treated Untreated Treated Source: own elaboration on FADN data Matching may be implemented with a variety of different methods. All methods construct an estimate of the expected unobserved counterfactual for each treated observation by taking a weighted average of the outcomes of the untreated observations. What differs is the specific form of the weights. In order to check that our results are not driven by the kind of PSM technique chosen, we use two widely used methods that deal very differently with the trade-off between bias and variance: Gaussian Kernel Matching 8 For a complete discussion on matching methods, see Dehejia and Wahba (2002). 9 If the balancing property is not satisfied this means that the two groups are too different in terms of observables and additional information would be needed.

9 (GKM) and Nearest Neighbour Matching (NNM). The first is a non-parametric matching estimator that uses weighted averages of all firms in the control group to generate the counterfactual outcome. One major advantage of these approaches is the smaller variance that is achieved because more information is used. A drawback of these methods is that also observations that are bad matches may be used. GKM can be seen as a weighted regression of the counterfactual outcome on an intercept with weights given by the Kernel weights. Weights depend on the distance between each firm from the control group and the treated observation for which the counterfactual is estimated (Smith and Todd, 2005). The second method is the most straightforward matching estimator. A farm from the comparison group is chosen as a matching partner for a treated farm that is closest in terms of propensity score MAIN RESULTS As well known, PSM technique requires a first step, in which the probability of receiving a treatment is estimated with respect to pre-treatment control variables to remove systematic differences between treated and untreated observations. In our application of PSM, we first estimate a logit regression in which the dependent variable equals one if the farm was treated and zero otherwise. We matched participants and non-participant observations by two PSM techniques as discussed earlier. The standard errors of the impact estimates are calculated by bootstrap using 500 replications for each estimate ATT for RDP We tried alternative specifications of the logit model, for example we tried to exploit the information about regional location of the farm, but the balancing test failed. The specification used in this paper (Table 3) is the most complete and robust specification that satisfied the balancing property. The logit model correctly classifies 87.13% per cent of observations. In general, farms located in environmentally protected areas and those specialized in breeding animals are more likely to benefit of RDP payments. The probability of participation decreases when the agricultural family is pluriactive, that is when some members of the household work off farm, when operators are young and when they are female, when the farm is specialized in the production of field crops and when it is located in the plain. We then matched participants and non-participant observations by the two PSM techniques, namely the GKM and the NNM method, described in the previous paragraph (Table 4). Our analysis reveals that participation in RDP has a significant positive causal impact on family labour, while it does not have a significant impact on other structural indicators such as total labour units and farm land, either total and cropped. These findings suggest family labour has been substituted to external waged labour. This may be due to the use of cost reduction strategies based on self-exploitation (accepting returns to owned labour and land lower than the market wage and rent) to cope with external economic pressures and survive economic crisis. A second possible explanation for the increase in family labour is that the diversification of farms in non agricultural activities 10 For a detailed discussion, see Caliendo and Kopeining (2008)

10 promoted by RDP increased the opportunity to employ on farm skilled family labour trained in fields different from farming. Table 3: results of the logit regression on the RDP sample Coef. Std. Err. Age of the operator -0,021 0,005 *** Sex of the operator -0,264 0,144 * North-West -0,193 0,217 North-East -0,504 0,233 ** Centre -0,264 0,217 South -0,477 0,207 ** Plane -0,787 0,155 *** ESU < 8-0,224 0,167 FT olive -0,026 0,244 FT wine -0,459 0,285 FT field crops -0,518 0,160 *** FT citrus -0,334 0,250 FT livestock 0,108 0,172 Environmental protected areas 0,808 0,263 *** Pluriactivity -0,622 0,216 *** Intercept 40,419 8,945 *** Number of obs 2561 LR chi2(15) 114,86 Prob > chi2 0,000 Pseudo R2 0,057 Log likelihood -959,82 Source: own elaboration Table 4: Average Treatment Effect (ATT) of the treated RDP sample Gaussian Kernel Matching Nearest Neighbour Matching ATT Std. Err. t ATT Std. Err. t FAWU AWU UAA TAA Added value* Added value*/awu 1769, ,202 0, , ,874 0,324 Profit/FAWU ratio Source: own elaboration Note: * added value net of the payment More in general, given there was not a significant change in total labour unit we can conclude that the RDP did not produce a direct impact on rural employment. In addition, farms that participated in the RDP present a better economic performance than non-participant farms. It is interesting to note that in the case of corrected value added, i.e. net of RDP payments, the average treatment effect on the treated is significant in the case of the CDID estimator based on the GKM, while it is not significant in the case of NNM. This difference is possibly due to the less information used in the second method. In previous paragraphs we already mentioned that NNM is

11 the most straightforward matching estimator, as a consequence the signal given by this estimator may be more reliable than the one produced by the GKM. Finally, the increase in profits per unit of family labour is positive and significant both in the case of the estimator based on the GKM and that of the NNM ATT for LFA support scheme The results of the logit model on the sample of farms located in LFAs are reported in table 5. The model correctly predicts 90.36% of the observations. The probability to participate in LFA measure increases when the farm is specialized in livestock production, while it decreases when farms are specialized in the production of fruits and filed crops. The probability lowers among young and female farmers, as well as when some member of the family farm is employed off farm (pluriactivity). As expected the probability of participation increases when farms are located in mountain areas and it is lower in farms located in Central and Southern regions. Table 5: results of the logit regression on the LFA sample Coef. Std. Err. Age of the operator ** Sex of the operator * North-West North-East Centre *** South *** Plane Mountain *** ESU < * FT olive FT wine FT field crops *** FT friut&citrus *** FT livestock * Environmental protected Pluriactivity ** Diversified farms Intercept ** Number of obs 1343 LR chi2(17) Prob > chi Pseudo R Log likelihood Source: own elaboration In the case of LFAs scheme the average treatment effects on the treated are all positive but not statistically significant apart from that associated to total output per hectare of land, this ATT is negative and significant (Table 6). These results suggest that the scheme was effective in lowering land use intensity, while maintaining farm income.

12 Table 6: Average Treatment Effect (ATT) of the treated LFAs sample Gaussian Kernel Matching Nearest Neighbour Matching ATT Std. Err. t ATT Std. Err. t AWU FAWU UAA TAA Total output Total output per ha Added value*/awu Profit per FAWU Source: own elaboration Note: * added value net of the payment 5. CONCLUSIVE REMARKS We assess the impact of RDP as a whole and of the measure aimed to compensate farmers in LFAs accounting for all systematic differences in outcomes in the untreated state. In the case of RDP as a whole, we found that recipient farms increased family labour units more than non-recipient ones, while no significant changes have been estimated in the case of total labour units. This suggests that family labour has been substituted to waged external labour force, and that there was not direct net impact of the policy on rural employment. On the contrary, we find the programme had a positive net impact on value added. Hence, it contributed to meeting the GDP growth objective. As for the LFA scheme, we have found a negative statistically significant effect on the output per hectare of land, while no significant change on other selected structural and economic indicators. These results suggest that the measure promoted the extensivation of farm production, while not negatively affecting farmers income. These evidences appear to be encouraging in respect with the use of payments for ecosystem services, that is payments, such as the compensatory allowances paid under the LFAs scheme, to compensate farmers for maintaining or promoting the use of sustainable farming systems in environmentally sensitive areas, in view of preserving farmland landscapes and conserving fragile environments. Our future work could go in several directions. A natural extension is to update the analysis in order to capture the long-standing effect of RD policy. A further direction of research is to enlarge the set of outcome variables in order to better evaluate the environmental impact of the LFAs scheme. In addition, it would be of interest to disentangle the complementarities between LFAs compensatory allowances and agrienvironmental measures. REFERENCES Becker S. and A. Ichino (2002). Estimation of average treatment effects based on propensity scores. Stata Journal, vol. 2(4): Caliendo M. and S. Kopeinig (2008). Some practical guidance for the implementation of propensity score matching. Journal of Economic Surveys, vol. 22(1): Chabé-Ferret S. and J. Subervie (2010). Evaluating Agro-Environmental Schemes by DID Matching: Theoretical Justification, Robustness Tests and Application to a French Program, mimeo.

13 Dehejia, R. and S. Wahba (2002). Propensity score matching methods for nonexperimental causal studies. Review of Economics and Statistics, 84 (1): Lynch, L., W. Gray, and J. Geoghegan (2007). Are Farmland Preservation Program Easement Restrictions Capitalized into Farmland Prices? What Can a Propensity Score Matching AnalysisTell Us?. Review of Agricultural Economics, 29(3): Lynch, L., and X. Liu (2007): Impact of Designated Preservation Areas on Rate of Preservation and Rate of Conversion: Preliminary Evidence. American Journal of Agricultural Economics. 89(5): Pufahl, A., and C. R. Weiss (2009). Evaluating the Effects of Farm Programmes: Results from Propensity Score Matching. European Review of Agricultural Economics, 36(1): Rosenbaum P. and D. Rubin (1983). The central role of the propensity score in observational studies for causal effect, Biometrika, 70: Smith, J A. and Todd, P.E. (2005). Does matching overcome Lalonde s critique of nonexperimental estimators? Journal of Econometrics, vol. 125(1-2):