Green productivity in agriculture: A critical synthesis

ing from the material balance context, our interpretation of the previous material balance argument is the following. Suppose the production model takes explicitly into account all factors that influence production, including such aspects as skill and motivation. In the theoretical case where all apects that affect production are explicitly taken into account, the set of input factors will explain the observed outputs perfectly (in regression terminology, we have a perfect fit with the coefficient of determination equal to 1). Hence, there is no inefficiency or unexplained residuals. This implies that all firms or countries are equally productive, and we cannot draw a distinction between inefficient or efficient producers. The efficiency indices that are constructed from the unexplained residuals ultimately depend on the omitted variables: the variables that are excluded from the analysis. In the worst case, efficiency differences capture unobserved random variation, and thus the efficiency analysis produces nothing but random noise. If the objective of the analysis is to assess green productivity, we may make a deliberate choice to exclude some of the input or output variables from the analysis. This is exactly the reasoning behind the environmental performance approach by Kuosmanen and Kortelainen (2005). By using value added as an aggregate measure of inputs and outputs, they deliberately depart from the conventional production models by excluding labour and capital inputs from the analysis. If the focus is on environmental performance, it can be argued that efficient use of labour is not a valid excuse to pollute. Another issue worh noting concerns the dynamics of the material balance. In practice, the material balance equation can be used for estimating the quantities of pollutants that are not directly observable or difficult to measure. In agriculture, it is standard to use the nutrient balance method to estimate the surplus of nitrogen and phosphorus. OECD applies this method for calculating nitrogen and phosphorus balances at country level (see OECD, 2007a,b), and subsequently uses the nutrient balances to construct indicators of environmental performance (OECD, 2008). However, it is important to note that these surplus measures are flow variables that ignore the nutrient flows in the previous periods and the accumulation of nutrients to the soil. The cycle of nutrients in the soil is a very slow process, particularly in the case of phosphorus. The recent study by Kuosmanen and Kuosmanen (2012) suggests it would be more appropriate to consider the stocks of nutrients rather than the flows. They propose a simple dynamic model of material balance, which takes the time and the accumulation of nutrients explicitly into account. Specifically, stock of nutrient (Z) in country i in period t can stated as , 1 , 1 (1 ) (1 ) it i i t it i i t it it Z Z s Z a x b y where i is the decay rate of nutrient in country i. In practice, the nutrient stocks could be constructed from the available data of nutrient flows in a similar manner to the capital stocks constructed from Green productivity in agriculture: a critical synthesis 18 investments: the nutrient surplus is analogous to capital investment. In the measurement and analysis of productivity, it is standard to use the capital stock (or a flow of services from the stock) rather than the investment flow. By the same argument, it would be appropriate to use in environmental performance analysis or environmentally sensitive TFP measurement the stocks of nutrients or the flow of nutrients from the stock to water, air, and soil. For practical purposes, the flow of nutrients from the stock (i.e., the decay of the nutrient stock) is usually assumed to be proportionate to the stock, and hence the choice to use stock or flow variables does not matter. Whether one is interested in flow or stock variables, explicit modelling of the nutrient stock is in our view important both from the conceptual and practical point of view. From the conceptual point of view, taking the past nutrient inputs and the accumulation of nutrients in the long term explicitly into account provides a better proxy of the environmental pressure from nutrients than the nutrient surplus or deficit during a single period. Note that the nutrient balance can be positive (surplus) or negative (deficit). Nutrient deficit can be good or bad for the society, depending on the current stock of nutrients. In general, the socially optimal level of nutrient balance depends on the current level of the nutrient stock. From a practical point of view, the nutrient deficit is problematic in the assessment of green productivity. Nutrient deficit can sometimes occur at the country level (see, e.g., OECD, 2011a; Section 1.4), but it is very common at the farm level. Consider, for example, a simple ratio of agricultural value added to the nitrogen balance as a partial productivity indicator (the same issue concerns TFP indices as well). If a country has nitrogen deficit, this ratio becomes negative. To avoid the problem of negative productivity indices, a common practice is to add a sufficiently large number (say M) to the nutrient balance figures such that all data points become positive. But in this case, what is the meaning of the value added divided by the nitrogen balance plus M? Technically, the problem with the rescaled nutrient balance figures is that they are defined on the interval scale, and hence mathematical operations such as multiplication and division are not meaningful. The nutrient stock effectively resolves the problems caused by temporary nutrient deficit: the nutrient stock is always positive, and so is the nutrient flow from the stock (i.e., the decay of stock). 5.3 Production risk Agricultural productivity tends to fluctuate over time due to random variations in temperature, rainfall, and other weather conditions. Farmers can take different measures to manage the production risk. For example, they can diversify their activities over various production lines (e.g., joint production of crops and dairy products) and use the land for cultivating different crops. The damage control inputs discussed above are a way to reduce the risk of pest damages. The input use (e.g., capital intensity) can also influence the production risk. Taking production risk into account in productivity and efficiency analysis is a challenge that goes beyond the data and measurement problems. For example, consider rainfall as a random variable that is beyond the control of the farmer. Even though farmers cannot control the rain, they can affect their risk exposure through their input choices (consider, e.g., capital investment to an irrigation system). A farmer makes his production decisions based on some expected amount of rainfall. Even if the input choices were perfectly rational in light of the ex ante expectations, if the realized amount of rainfall differs from the expected value, then the farmer’s input choices may appear to be inefficient in light of the ex post evaluation. Just and Pope (1978) were the first to take the production risk explicitly into account as a part of the production model. Kumbhakar (1993, 2002) has adapted the production risk model to the SFA setting. Technically, the treatment of production risk resembles the standard econometric approaches to heteroscedasticity (e.g., Greene, 2011). In practice, we can make the variance of the inefficiency term u, the noise term v, or possibly both, depend on inputs x and potentially some other factors (e.g., weather Green productivity in agriculture: a critical synthesis 19 variables) to take the production risk explicitly into account. If the inefficiency term u is heteroscedastic, then this needs to be taken into account in the efficiency analysis or performance measurement as well. The conventional Just and Pope model of production risk assumes some specific continuous probability distribution for the output loss due to the risk factors. In contrast, Chambers and Quiggin (1998, 2000) model the risk by assuming a discrete set of states that occur with certain probabilities. They apply the state-contingent approach to analyse principal-agent problems and the pollution control, among other applications. Recently, O’Donnell et al. (2010) examine how the state-contingent approach could be applied in productive efficiency analysis to take into account the production risk. 5.4 Critical synthesis The previous sub-sections review some econometric modelling issues that are particularly relevant to agricultural productivity analysis. Most of the previous studies known in the literature tend to focus on dealing with one of these issues in isolation, ignoring the other issues noted above. Ideally, it would be important to consider the damage control inputs, material balance, and production risk simultaneously, provided that necessary data are available. Further methodological research is clearly needed to achieve a better synthesis of these issues. The main purpose of this section was to recognize importance of these issues in the present context. We must emphasize that the empirical study reported in Section 7 does not take all of the methodological issues discussed in this section explicitly into account. We do apply the material balance accounting to construct the nutrient stocks, which are used as environmental inputs, and we try to capture random idiosyncratic risks implicitly by assuming a stochastic noise term in the model. We recognize that more thorough investigation of the role of damage control inputs and production risk would be worthwhile. In practice, one of the major constraints concerns the data availability, which will be discussed next. 6. Environmental issues specific to agriculture In this section we examine the main environmental questions related to agriculture. Our discussion is focused on the current situation in the OECD countries. Globally, agriculture is closely related to environmental problems such as deforestation or water stress, particularly in developing countries. In the OECD countries, clearing of forests for agricultural lands is a less important problem, and the current environmental discussion is focused more on the undesirable emissions from agricultural activities to air, water and soil. In the context of high income countries, commonly used indicators for the environmental pressures caused by agriculture include the following: Nitrogen emissions to air, soil, and water Phosphorus emissions to air, soil, and water GHG emissions (particularly CO2) Consumption of fossil fuels The use of toxic pesticides Land use diversity We will next briefly discuss the above environmental issues, commenting the measurement problems and data availability. The nitrogen and phosphorus balances can be calculated using the nutrient balance method (OECD, 2007a,b). The data for nitrogen and phosphorus surpluses (or deficits) are available. However, it is not Green productivity in agriculture: a critical synthesis 20 directly obvious how these variables should enter productivity analysis. One problem with the nutrient balance measures concerns the negative values (deficits). The nutrient balance is defined on the interval scale, whereas the productivity ratios and more sophisticated TFP measures would require the ratio scale. One possibility to avoid this problem is to apply the stocks of nutrients as suggested by Kuosmanen and Kuosmanen (2012). Alternatively, we could utilize the nutrient stocks to estimate the emissions of nutrients to air, soil, and water separately. The main sources of GHG emissions from agriculture include the methane from manure and the use of fossil fuels for energy. Different GHGs could be aggregated to CO2 equivalents by using the coefficient representing the global warming potential, used and published by the United Nations Framework Convention on Climate Change (UNFCCC). Precise estimation of the GHG emissions from agriculture is challenging, but proxy indicators could be calculated based on available data on energy use and the livestock. Pesticides are both an environmental and health concern. However, the measurement of the damage caused by pesticides is challenging. Even if we ignore the social welfare considerations, the problem with pesticides is that there exists a very large number of different pesticides products, with different toxic properties. In practice, we would have to aggregate the pesticide products to some pesticides index in one way or another. However, it would be misleading to just add up the quantities of pesticides, or use some cost based aggregate. For example, it is possible that less expensive pesticides are also more harmful to the environment. If we are interested in the environmental pressure, the aggregation of different pesticides products should somehow reflect the relative harmfulness of different pesticides products. While it would be interesting and important to take the pesticides use explicitly into account, we leave this issue for future research due to data problems. In contrast to monoculture, diversity of land use is beneficial both for biodiversity and for recreational uses. Land use diversity could be measured, for example, by the standard Shannon-Weaver diversity index. While land use diversity could be easily incorporated to the productivity analysis at the farm level, it is not necessarily appropriate to apply the same approach at the country level. More specifically, the ShannonWeaver diversity indices calculated at the country level cannot be disaggregated to the farm level, or vice versa. It would seem necessary to calculate the diversity indices first at disaggregate level (e.g., farm level), and then use the country average or median. However, this would be both tedious and costly. 7. Empirical application 7.1 Data and model specification In this application we compare three alternative orientations to productivity measurement: purely economic (ECON), purely environmental (ENV), and the mixed economic and environmental (MIX). The total factor productivity (TFP) index used in the application is the conventional Malmquist index, consisting of the technical change (TECH) and efficiency change (EFF) components. The frontier is estimated by the panel data version of stochastic semi-nonparametric envelopment of data (StoNED: Kuosmanen and Kortelainen, 2012) assuming monotonicity, convexity, and constant returns to scale. Technical change is captured by a parametric, linear time trend. For comparison, we will also consider the conventional SFA and DEA approaches. In Section 7.3 we estimate the conventional Cobb-Douglas production function by SFA, specifying the most productive country as the benchmark as in Schmidt and Sickles (1984), and capturing the technical change by a linear trend. In Section 7.4 we apply the panel data version of DEA suggested by Ruggiero (2004), which is capable to assimilate stochastic noise. The empirical analysis has been conducted using the following software packages. The StoNED model was Green productivity in agriculture: a critical synthesis 21 estimated using the GAMS (General Algebraic Modeling System) software and its MINOS solver (see http://nomepre.net/index.php/computations for further details). The SFA model was estimated using the Stata software (see http://www.stata.com/). The DEA model using the EMS software (Efficiency Measurement System: see http://www.holger-scheel.de/ems/). See Appendix 1 for a more detailed description of the assumptions and properties of the DEA, SFA, and StoNED methods. Comparable sector level data on economic inputs and outputs as well as environmental variables are not available for all OECD countries. Based on data availability, we have included in the cross country comparison 13 OECD countries: AUT, DEN, FIN, FRA, GER, GRE, ITA, NED, NOR, POR, SPA, SWE, and UK. The time span of the study is 15 years, from 1990 to 2004. Unfortunately, more recent data were unavailable for many key variables. In future research, it would be interesting to extend the present study to cover a larger set of countries (including non-European countries) and more recent years, provided that comparable data are available. As the output variable, one could use the gross output or the value added; see OECD (2001) for a detailed discussion of these two approaches. In the present context, it is convenient to resort to the value added approach where the use of intermediate inputs is subtracted from the gross output. It is somewhat challenging to obtain comparable data of the use of intermediate inputs at the sectoral level. Further, a large number of input variables can cause problems for estimation, in particular, possible multicollinearity in the parametric approaches, and the so-called ‘curse of dimensionality’ in the non-parametric approaches. Therefore, in this study we use the net production value (constant 2004-2006 prices, 1 000 Int. $) reported by FAOSTAT as the output variable y in all three alternative models. Based on the description provided by FAOSTAT, it seems that this output variable excludes the intermediate inputs (seed and feed are explicitly mentioned in the description), which are included in the gross production value reported by FAOSTAT. The three model specifications differ in terms of the specification of the inputs. The ECON model includes the following inputs: Capital K (gross capital stock, constant 2005 prices, FAOSTAT), Labour L (primary agriculture employment, number employed, OECD) Land LA (total agricultural land area, hectares, OECD) The ENV model includes the following variables (treated as input factors): Agricultural total GHGs (Tonnes CO2 equivalent, OECD) Nitrogen stock N (own calculations, based on the nitrogen surplus reported by OECD) Phosphorus stock P (own calculations, based on the phosphorus surplus reported by OECD) Land LA (total agricultural land area, hectares, OECD) The model MIX contains all six input variables of both the ECON and ENV models. The inputs of the ECON model are rather standard. Note that the land capital could be modelled as a part of the capital stock. Based on the description of FAOSTAT, the gross capital stock reported by FAOSTAT only includes the physical assets in use. Therefore, we model the land area as a separate input variable. Further, it is common to express all input and output variables proportional to the land area (e.g., output per hectare). In fact, dividing all input variables and the output by one of the input variables effectively imposes the CRS assumption; we will utilize this property in Section 7.3. The ENV model includes all environmental variables for which comparable data are available. The GHG emissions take into account the use of fossil fuels for energy, but also methane emissions from livestock. Based on the discussion in Section 5.2, we prefer to use the stocks of nitrogen and phosphorus as indicators of environmental pressure from nutrients use. Finally, the ENV model also includes the land area, which Green productivity in agriculture: a critical synthesis 22 can be considered as a part of the natural capital. A practical reason for including the land area is that we can impose CRS consistently across all models by dividing all input variables and the output by the same input variable, the land area. In other words, we can express all variables on a per hectare basis. The MIX model combines the economic and environmental perspectives, allowing us to model trade-offs and substitution possibilities between economic inputs and environmental resources explicitly. Note that both the ECON and ENV models are nested within the more general MIX model. The ECON model is obtained from the MIX model by restricting the shadow prices of the environmental variables as equal to zero. Similarly, the ENV model is obtained from the MIX model by setting the shadow prices of the economic inputs equal to zero. As a result, the empirical fit of the MIX model (measured, e.g., by the coefficient of determination R) is always better than that of the ECON model or that of the ENV model. A limitation of the MIX model is that does not distinguish whether high TFP level or TFP growth is due to good performance in economic or environmental criteria. Therefore, it is useful to consider and compare the results of all three approaches. To gain intuition to the levels and changes in productivity in this sample of countries during the time period, we first examine partial productivity measures, the average labour and capital productivity, presented in Tables 1 and 2. Table 1: Labour productivity (y/L) in the sample countries; averages of 1990 – 2004 Country Mean y/L Efficiency rank y/L rank AUT 18.82 41 % 7 4.48 % 5 DEN 46.28 100 % 1 4.55 % 4 FIN 11.82 26 % 10 3.63 % 7 FRA 39.40 85 % 3 4.69 % 3 GER 26.52 57 % 5 4.30 % 6 GRE 9.82 21 % 12 3.41 % 8 ITA 20.03 43 % 6 5.76 % 1 NED 45.22 98 % 2 2.33 % 11 NOR 11.27 24 % 11 1.82 % 12 POR 5.68 12 % 13 3.37 % 9 SPA 18.77 41 % 8 5.39 % 2 SWE 17.24 37 % 9 2.72 % 10 UK 30.61 66 % 4 0.94 % 13 Notes: The unit of measurement for labour productivity y/L is $1,000 per worker. Efficiency is calculated as the ratio of mean labour productivity y/L of country and the mean labour productivity of Denmark, which has the highest y/L ratio in the sample. y/L is the geometric mean of the annual changes in labour productivity y/L. Countries are sorted in alphabetical order in Table 1: the first column indicates the country abbreviation. The second column indicates the average labour productivity in $1 000 per worker. The third column indicates the relative labour efficiency, calculated as the ratio of country’s average labour productivity and that of Denmark, the country with the highest labour productivity in this sample. To help a reader to quickly recognize the most productive and least productive countries, the fourth column indicates the relative rank of a country in terms of labour productivity. Observe the large differences in the level of labour productivity across countries. For example, Greece and Portugal achieve on the average only 21% and 12% of the labour productivity of Denmark during this time period. The large differences in labour productivity across countries are not surprising given the intensive use of capital intensive production technologies in some countries in contrast to more traditional labour intensive agriculture in others. Green productivity in agriculture: a critical synthesis 23 Column y/L in Table 1 indicates the change in labour productivity, calculated as the geometric mean of the annual changes of y/L. The last column indicates the relative rank of a country in terms of labour productivity growth. All countries achieved positive labour productivity growth during this period. The highest growth occurred in Italy, followed by Spain and France. Table 2 presents the analogous statistics for capital productivity y/K. Observe the large differences in the level of capital productivity across countries. No other country in the sample comes even close to the capital productivity of the Netherlands. We should add that the measurement of the capital stock is a challenging task, and the data of capital stocks reported by FAOSTAT might not be fully comparable across countries, but in this study, we assume the FAOSTAT data are correct. As for the changes, the growth of capital productivity y/K is rather modest compared to the labour productivity growth. For Norway, Sweden, and the United Kingdom, capital productivity declined during this time period. The comparison of Tables 1 and 2 reveals a rather different picture both in terms of the levels of productivity and the productivity growth. This highlights the need to resort to TFP measures that can accommodate multiple input factors, including environmental factors. Table 2: Capital productivity (y/K) in the sample countries; average of 1990 – 2004 Country Mean y/K Efficiency rank y/K rank AUT 244.6 26 % 10 1.318 % 3 DEN 410.1 44 % 2 1.763 % 2 FIN 136.9 15 % 12 1.059 % 5 FRA 387.9 42 % 4 0.932 % 7 GER 304.5 33 % 8 2.454 % 1 GRE 399.7 43 % 3 0.308 % 10 ITA 365.2 39 % 5 0.548 % 9 NED 925.5 100 % 1 1.000 % 6 NOR 132.4 14 % 13 -0.049 % 11 POR 248.2 27 % 9 1.152 % 4 SPA 346.1 37 % 6 0.603 % 8 SWE 187.9 20 % 11 -0.275 % 12 UK 338.2 37 % 7 -0.300 % 13 Notes: Efficiency is calculated as the ratio of mean capital productivity y/K of country and the mean capital productivity of the Netherlands, which has the highest y/K ratio in the sample. y/K is the geometric mean of the annual changes in capital productivity y/K. Analogous to labour and capital productivity, environmental partial productivity measures could be calculated using the environmental indicators noted above. Since such environmental partial productivity indicators have been reported elsewhere (see, e.g., OECD, 2008, 2011a), we turn our focus on the TFP measures. 7.2 Stochastic semi-nonparametric envelopment of data We first estimate the following log-transformed production models by convex nonparametric least squares (CNLS: Kuosmanen, 2008; Kuosmanen and Kortelainen, 2012) ECON: ln ln ( , , ) it it it it it y f K L LA Trend t ENV: ln ln ( , , , ) it it it it it it y g GHG N P LA Trend t MIX: ln ln ( , , , , , ) it it it it it it it it y h K L LA GHG N P Trend t Green productivity in agriculture: a critical synthesis 24 where f, g, and h are unknown monotonic increasing and concave functions that exhibits constant returns to scale, Trend is a parameter that represents Hicks neutral technical change, and it is the disturbance term that captures both inefficiency and noise. Note that in the panel data setting the idiosyncratic errors cancel out over time. Thus, we follow Schmidt and Sickles (1984) and use the average of the residuals of the most efficient country as the benchmark level to define the frontier. The country specific efficiency levels are estimated as the geometric mean of the residuals of country i divided by the geometric mean of the residuals in the most efficient country