Evaluating environmental effects of adopting automatic milking systems on Norwegian dairy farms

Abstract

We present a novel procedure based on eco-efficiency for assessing farm-level effects of technology adoption while considering secondary effects. Secondary effects are defined as structural and behavioural adaptations to technology that may impact environmental, social or economic outcomes. We apply the procedure to automatic milking systems (AMS) in Norway and find that AMS induces secondary effects, most strongly by decreasing labour per cow and increasing herd sizes. For estimating effects of AMS we employ a novel causal machine learning approach. AMS induce heterogenous effects on eco-efficiency, negatively associated with herd expansion and labour per cow.

1. Introduction

Technical improvements at the farm level are one crucial path to improved environmental sustainability (Messerli et al., 2019). However, new technologies may induce additional changes that positively or negatively affect environmental outcomes. Overlooking these effects increases the risk that novel technologies lead to maladaptation (Pörtner et al., 2022). We refer to these changes as ‘secondary effects’, defined as structural and behavioural adaptations to novel technology, which may impact environmental, social or economic outcomes. It is crucial to consider secondary effects when assessing how (novel) technology affects farm sustainability. Awareness of secondary effects enables steering technology adoption and usage through regulatory changes or technological advancements to promote sustainable development. Yet, secondary effects are often overlooked when assessing farm-level effects of novel technology in terms of economic or environmental impacts. Previous literature has indicated secondary effects of, for example, smart farming technologies on the likely responsiveness to greenhouse gas (GHG) taxes (Schieffer and Dillon, 2015), the social impacts of GHG mitigating policy (Harrison et al., 2021) or as rebound effects where efficiency improvements can lead to increased resource usage (Herring and Roy, 2007; Sears et al., 2018; Paul et al., 2019). Smart farming technology is predicted to improve the sustainability of agriculture (Balafoutis et al., 2017; Duckett et al., 2018; Finger et al., 2019). Nevertheless, as the extensive usage of most smart farming technologies and robotics still lays in the future, empirical evaluations are scarce (Lieder and Schröter-Schlaack, 2021), and the inclusion of secondary effects is rare.

As European agriculture accounts for approximately one-tenth of global GHG emissions (FAO, 2020), the technological development must contribute to lowering this environmental impact. One type to robotic technology that is already widely adopted by farmers are automatic milking systems (AMS) which provide an interesting study case for secondary effects. In the livestock sector, accounting for a large share of agricultural GHG emissions, there is considerable potential for mitigating emissions by reducing the emission per unit of product (Mbow et al., 2019). In dairy farming, one way to reduce emissions per unit of product is to increase the milk yield per cow (Zehetmeier et al., 2012), which can be achieved by adopting more efficient technology, such as AMS. One of the countries with the highest implementation of AMS is Norway. In 2018, cows milked with AMS produced 47 per cent of the milk (Vik et al., 2019). In 2020, this had increased to 57 per cent (Mikalsen, Österås and Roalkvam, 2021). Norwegian farmers adopt AMS to increase their work-time flexibility and thus quality of life, and to reduce farm labour requirements (Hansen, 2015; Stræte, Vik and Hansen, 2017; Vik et al., 2019). However, previous studies have also indicated that AMS adoption in Norway is coupled with structural changes on farms, specifically farm expansion (Vik et al., 2019; Rønningen, Magnus Fuglestad and Burton, 2021). AMS has also been found to be associated with changes in feeding patterns towards more high-energy feed (Bijl, Kooistra and Hogeveen, 2007; Oudshoorn et al., 2012; Schewe and Stuart, 2015) and less grazing allowing for cows to be milked more frequently (Oudshoorn et al., 2012; Gołaś et al., 2020; Lessire et al., 2020). Furthermore, using AMS is associated with increased energy consumption (Steeneveld et al., 2012). Consequently, AMS can generate secondary effects as it is coupled with several farm-level changes. The implications of those changes for farms’ environmental performance, particularly GHG emissions, remain an open question. To this background, the question arises of how AMS relate to farms’ structural development and environmental performance.

In this paper, we provide novel insights on the effects of AMS adoption on farm-level environmental performance, specifically focusing on the effects of AMS on GHG emissions efficiency. GHG emissions efficiency refers to an eco-efficiency measure focusing specifically on GHG emissions as the environmental outcome (Stetter, Wimmer and Sauer, 2022). Eco-efficiency is expressed as a ratio between value-added and indicators for GHG emissions. Integrating economic and environmental factors into one efficiency measure is crucial to managing trade-offs between environmental objectives and production (Huppes and Ishikawa, 2005). Using data envelopment analysis (DEA), scores are generated describing farms’ ability to produce output while inducing minimal environmental damage (Kuosmanen and Kortelainen, 2012). We aim to assess what structural and behavioural factors can be identified as secondary effects of AMS adoption and how this affect farms’ GHG emissions efficiency. Our aim is formulated as two research questions:

1. What structural and behavioural factors can be identified as secondary effects of AMS adoption?

2. Does AMS adoption generate changes in farms’ GHG emissions efficiency, which can be associated with the structural and behavioural changes?

We contribute to the literature evaluating the secondary effects of AMS through our empirical results. The procedure we provide in this paper can also be applied in other settings when evaluating the secondary effects of novel technology. Using our novel approach, we combine results on how AMS generate secondary effects with results on how AMS induces changes in GHG emissions efficiency. Linking AMS adoption to eco-efficiency is already a novel contribution. This allows to understand the secondary effects of AMS adoption in the form of structural and behavioural changes and in terms of farms’ environmental performance. By attributing the effect of AMS on GHG emissions efficiency to the identified secondary effects, insights for policy and extension can be provided on what aspects to target to achieve sustainable development of farms when adopting AMS.

We find that AMS adoption is associated with increased herd sizes, increased share of feed concentrates and increased milk yields per cow. Further, we find largely heterogenous effects of AMS adoption on GHG emissions efficiency with a negative effect on average. The effect of AMS adoption on GHG emissions efficiency highlights the importance of evaluating how new technology affects farms environmental outcomes.

The remainder of the paper is organised as follows: First, we present our novel procedure to evaluate secondary effects of novel technology and the methodologies we use. Second, we present the dataset we use to conduct the empirical evaluation. Third, we present the findings, and discuss the conclusions that can be drawn from using this approach in the context of Norway and AMS adoption. Finally, we provide some suggestions for future research.

2. Method

We employ a novel four-step procedure to identify structural and behavioural factors as secondary effects of AMS and to assess whether AMS adoption generates changes in farms’ GHG emissions efficiency which can be associated with the structural and behavioural adaptations. An illustration of this procedure is provided in Figure 1.

In steps one and three, we assess the effect of AMS adoption on structural and behavioural factors and on GHG emissions efficiency, respectively. We identify six factors as important in terms of AMS adoption and GHG emissions efficiency based on previous research. These factors are labour per cow, number of cows, share of feed concentrates, arable land per milk output, milk per cow and off-farm income. To assess GHG emissions efficiency we include value-added, energy consumption, fertiliser consumption and enteric fermentation. We present these variables and motivate their inclusion in Section 3. We obtain the effect of AMS adoption on each of the factors by calculating their counterfactual development if the farms had not adopted AMS by using a matrix completion approach by Athey et al. (2021). From this counterfactual, we can calculate an average treatment effect for the treated (ATT). The advantage of the matrix completion approach is that it basically performs a matching based on the pre-treatment trend, for example matching adopting farms that grow in herd size prior to adoption to non-adopting farms with similar growth in herd size. Additionally, it allows for controlling of individual time-invariant unobserved factors as well as time individual-invariant factors, similar as a fixed effects (FE) regression. Nevertheless, it is important to underline the possibility of reverse causality between the factors and AMS adoption. For example, we cannot detangle if the motivation to increase herd size leads to adopting AMS or if the availability of AMS motivates a farm to increase herd size. Similarly, AMS adoption might increase value-added, which would affect farms’ GHG emissions efficiency, but it might also be that changes in value-added make adoption affordable. Our approach cannot resolve this potential reversed causality, which is also conceptually difficult to detangle. However, by matching observations on pre-treatment development, we can compare farms on a similar development trajectory. Further, even if adopting AMS is part of an expansion or intensification strategy, if AMS allows the farmer to realise this strategy is already a secondary effect according to our definition. Nevertheless, we are careful in concluding directions of causality. Another aspect to consider is that the factors included as potential secondary effects likely interact. We do not account for these interactions when assessing the effects of AMS adoption as we want to obtain estimates of the changes in each factor independently. However, in the OLS regression (step four), the interaction between the factors is controlled for to enable explaining the changes in GHG emissions efficiency.

For our second question, regarding how AMS adoption affects GHG emissions efficiency and whether we can associate this to the secondary effects identified in step one, we employ steps two to four. In step two, we assess farm-level GHG emissions efficiency using the methodology developed by Kuosmanen and Kortelainen (2005). Previous evaluations of efficiency in livestock farming focusing explicitly on GHG emission include the works of Dakpo, Jeanneaux and Latruffe (2017) and Stetter, Wimmer and Sauer (2022). In the fourth step, we determine which variables correlate with GHG emissions efficiency using linear OLS regression. Finally, we seek to identify how the factors we identify as secondary effects can be associated with the relation between AMS adoption and GHG emissions efficiency. We multiply the effect of AMS adoption on each factor as assessed in step one with the marginal effects obtained in the regression in step three. This indicates how structural and behavioural change can explain the changes in GHG emissions efficiency generated by AMS adoption. We dedicate the rest of this section to outlining the details of the methods applied in each step of our procedure.

2.1. Steps one and three: assessing the impact of automatic milking systems

In steps one and three, we evaluate the effects of AMS on the structural and behavioural factors and GHG emissions efficiency. In these steps, we require a method to deal with unbalanced panel data and staggered adoption. For this, we rely on a novel machine-learning approach in line with an increasing stream of literature in recent economic research that has also turned to machine learning for causal questions (Storm, Baylis and Heckelei, 2019). Specifically, we employ a matrix completion approach for causal panel data models (Athey et al., 2021) that allows estimating a treatment effect in cases of staggered adoption and an unbalanced panel dataset. This approach can be seen as nesting a two-way fixed effect approach with synthetic control approaches (Abadie and Gardeazabal, 2003). The two-way FE allow to control for time-invariant as well as unit-invariant unobservables. The synthetic control approach constructs a synthetic counterfactual by matching on pre-treatment trends over time. Specifically, the approach considers treatment effect estimation as a missing data problem, where we lack the counterfactual outcomes that need to be predicted in order to compute treatment effects. In the matrix completion approach, the missing counterfactual observations are predicted by learning a low-rank representation of the observed non-treated outcomes using nuclear norm regularisation (Athey et al., 2021). Based on this low-rank representation, the counterfactual observations can be predicted. As the approach nests the FE and the synthetic control approach, it allows to combine both and determine their relative weighting in a data-driven way. Previously, researchers needed to decide a priori which of the two approaches to use.

The matrix completion method allows for including time- and farm-specific covariates, which are utilised in this paper. Including time- and farm-specific covariates adds to the FE already included in the model as it allows to account for the interaction between farm and time fixed factors. For example, being located in a remote region (a farm-fixed covariate) while diesel prices rise (a time-fixed covariate) might play a role in the effect of AMS on GHG emissions efficiency and is considered by adding the covariates. The included farm characteristics are a binary indicator for adoption, the farmer’s year of birth, the farm’s location and the year of AMS adoption (set to zero for non-adopters). Furthermore, we include 370 agricultural input and output prices (The Budget Committee for Agriculture, 2022) as time-specific variables, such as prices of various crops and vegetables, livestock and livestock products (such as milk) and inputs like fertiliser and subsidies. One advantage of the approach is that it uses regularisation to avoid overfitting, allowing to include a larger number of control variables.

Having derived counterfactual outcomes for adopting farms if they did not adopt AMS, we estimate the average effect of treatment for the treated (ATT) as: |$\tau = \mathop \sum \limits_{i,t:\,{W_{it}} = 1} \left[ {{Y_{it}}\left( 1 \right) - \,{Y_{it}}\left( 0 \right)} \right]\,/\mathop \sum \limits_{i,t} {W_{it}},$| where |${W_{it}}$|= 1 if a farm has adopted and |${W_{it}}\,$|= 0 otherwise. Y_it(1) is the observed outcome for the observations with AMS, and for Y_it(0) we estimate their counterfactual outcome as |$Y\overline {_{it}} \left( 0 \right)$|⁠. Finally, we calculate the difference between the realised outcome and the counterfactual for each observation to gain insights on the distributions of the effects.

To increase transparency of the results and to provide an understanding of how this method compares to more commonly used econometric procedures, we conduct a two-period propensity score weighted difference-in-difference (PS-DID) regression and a FE regression in  Appendix 1.

2.2. Step two: assessing GHG emissions efficiency

We assess GHG emissions efficiency, a measure of eco-efficiency only considering indicators for GHG emissions. Throughout this section, we use the term ‘eco-efficiency’ when describing the procedure, as this is the most commonly used terminology. Eco-efficiency considers farms’ ability to minimise the environmental damage caused at a given amount of production and is defined as the ratio of economic value-added to emissions or other environmental damage (Kuosmanen and Kortelainen, 2005). Thus, it is a relative measure where eco-efficiency is achieved when production compensates the environmental harm it generates with sufficient value-added. What is considered sufficient value-added is determined by the structure of the sample, with the most eco-efficient farms having the highest ratio of economic value to environmental damage. Thus, if observations are added to a sample, the eco-efficiency of an individual unit can change if the efficiency frontier is affected. Although this paper focuses on GHG emissions, Kuosmanen and Kortelainen’s (2005) approach has the potential to simultaneously examine multiple environmental factors. Another term for eco-efficiency is sustainable intensification (Firbank et al., 2013; Gadanakis et al., 2015; Smith et al., 2017). The most common application of eco-efficiency in agriculture is at the farm-level (Zhou et al., 2018), which is the focus of this paper. The definition of an eco-efficiency only focusing on GHG emissions as ‘GHG emissions efficiency’ was initially made by Stetter, Wimmer and Sauer (2022).

We use DEA (Charnes, Cooper and Rhodes, 1978) to assess eco-efficiency, followed by bootstrapping of the efficiency scores to reduce sample bias (Simar and Wilson, 2000). DEA is a deterministic approach that evaluates each unit towards an efficiency frontier constructed from the most efficient units in the sample. Following the approach by Kuosmanen and Kortelainen (2005), our analysis rests on the assumption of a pollution-generating technology set which states that ‘value-added v can be generated with environmental damage z’. We consider that this pollution-generating technology set can improve over time, such that more value-added can be generated with less environmental damage, by assuming irreversible technical change. The assumption of irreversible technical change is based on the rationale that the technology available in each year consists of the technology in previous years and new technology developed in the year under evaluation (Lansink, Pietola and Bäckman, 2002). This assumption implies that observations are only compared to other observations with at least as good technology as themselves. Thus, observations of farms in earlier years can have the opportunity to achieve higher efficiency, despite not having access to as advanced technology as the observations made in later years. Practically, this is implemented by including all observations in the current year and all previous years in the comparison group when computing the eco-efficiency (Lansink, Pietola and Bäckman, 2002). Apart from implementing the assumption of irreversible technical change, the same methodology as in previous papers using cross-sectional observations can be applied (Gómez-Limón, Andrés and Reig-Martínez, 2012; Martinsson and Hansson, 2021; Pérez Urdiales, Lansink and Wall, 2016).

Following the notation used by Kuosmanen and Kortelainen (2005), we express the eco-efficiency for farm n as EE_n= V_n/D(Z_n), where V_n denotes the value-added and D(Z_n) is a damage function of the environmental pressures Z of farm n. The function D of the M environmental pressures for farm n can be approximated linearly as D(Z_n) = w₁z₁ + w₂z₂ … + w_Mz_M, where there are z_M environmental pressures, each with its own weight w_m. Weights are determined using DEA to produce the highest eco-efficiency score possible for each observation. The inverse of the maximisation problem is calculated to obtain linearity:

s.t.

We evaluate (1) for each year subsetting the data such that the observations are compared to observations made in the same year or in previous years. That is, when evaluating farms observed in 2013 (our earliest year of observation), we only compare these to other observations made in 2013. Evaluating farms in 2014, we include observations from 2013 and 2014, and so on. Eco-efficiency is measured against a frontier estimated from the sample, where the addition or omission of observations may alter the frontier and, consequently, the farms’ estimated efficiencies. Eco-efficiency is a further development of assessing technical efficiency. Thus, this paper’s eco-efficiency formulation corresponds to an input-oriented Farrell (1957) efficiency model, with environmental pressures as inputs and value-added as the only output (Bonfiglio, Arzeni and Bodini, 2017). Furthermore, the method utilised in this paper is based on a radial assumption, which can be a limitation in a short run perspective if not all variables can be varied at the same rate. Nevertheless, with a longer time-frame, this assumption becomes more feasible.

Using bootstrapping, pseudo samples are generated from which efficiency estimators are derived. From this, Monte Carlo realisations of the estimated efficiency can approximate the bias. The final step is calculating bias-corrected efficiency scores by subtracting the bias from the estimated efficiency. The bootstrapping is done as a Shephard input distance function (Simar and Wilson, 1998; Bogetoft, Otto Maintainer and Otto, 2022). Following Simar and Wilson (1998), we set the number of bootstraps in this application to 1,000. Furthermore, DEA is sensitive to outliers in the data, and it is common to drop observations identified as outliers (see e.g. Latruffe and Desjeux (2016); Weltin and Hüttel (2019)). Applying the procedure by Wilson based on log-ratios, we identify seven observations as outliers. By inspecting the data, we cannot find anything atypical about these observations identified as outliers that would indicate mistakes in the data recording. Thus, we chose to keep all observations.¹ The eco-efficiency scores are generated using the R software, and the package Benchmarking (Bogetoft and Otto, 2011) which draws heavily on the FEAR package (Wilson, 2008).

V and Z values in equation (1) must be positive to obtain a finite solution. This is known as the DEA’s positivity property (Bowlin, 1998). Thus, we replace values of Z that are zero or less with a small arbitrary number (Bowlin, 1998). Farms with negative V values are omitted from the analysis because they would generate eco-efficiency scores close to zero. Thus, we remove one farm with negative value-added. Meanwhile, energy and fertilisers both exhibit negative values for environmental pressures. This indicates that nothing was consumed of that indicator that year, which encourages the replacement of a small number to enable computation of the farm’s eco-efficiency for that year.²

2.3. Step four: assessing the association between GHG emissions efficiency and the structural and behavioural factors

We use a FE regression to assess the drivers of the GHG emissions efficiency. This constitutes the fourth step in our procedure (see Figure 1) and allows us to establish a connection between the changes in secondary effects and those in GHG emissions efficiency. A set of factors with hypothesised relationship to AMS adoption and GHG emissions efficiency are included in a linear regression as

where |${\alpha _i}$| and |${\gamma _t}$| are farm- and year-FE, respectively. The result from this step shows the marginal effect of each factor on GHG emissions efficiency. We use the results to link the GHG emissions efficiency to the factors included as potential secondary effects of AMS. A FE approach allows us to control for unobserved farm- and time-invariant heterogeneity, which is beneficial as we want to use the result of equation 2 to assess how farms’ GHG emissions efficiency changes through potential changes in the factors. However, it comes at the cost that only within-variation can be exploited for identification, producing less precise estimates. An alternative to including FE would be to pool the observations in a cross-sectional regression. This would increase the estimation efficiency and enable using all data rather than only the within-variation, but would not provide estimates on within changes on farms.

Theoretically, linear regression can be used to explain DEA efficiency scores (Hoff, 2007; McDonald, 2009). DEA scores are bounded between 0 and 1. However, by applying a bootstrapping procedure, few farms are evaluated as fully efficient (obtaining a score of 1) and thus fewer corner solutions are realised, which further supports the usage of OLS. As argued and further explained by Hoff (2007) and McDonald (2009), DEA scores are best described as fractional data generated from a normalisation process rather than censored data for which a Tobit model would be appropriate. However, using OLS in the second stage has also drawn criticism due to the efficiency scores’ serial relationship (Simar and Wilson, 2007). The method proposed by Simar and Wilson (2007) tends to generate similar results as linear OLS regression when explaining eco-efficiency (Latruffe, Davidova and Balcombe, 2008). Banker and Natarajan (2008) demonstrated that two-stage DEA provides a consistent estimator when data are generated by a monotone increasing and concave production function, further supporting the use of OLS regression in the second stage. OLS regression also has the advantage of being widely used and recognised by many, which offers an advantage in transparency and understandability, as also pointed out by McDonald (2009). Nevertheless, as Tobit regression has also been pointed out as a feasible method for this purpose (Hoff, 2007), we conduct a Tobit regression as a robustness check to the OLS. The Tobit is displayed in  Appendix 2, and the results are close to what we obtain with the OLS.

3. Data and descriptive statistics

We focus on conventional dairy farms using the Norwegian account results in agriculture and forestry, comparable to the EU’s Farm Accountancy Data Network, between 2013 and 2019. The dataset includes information on AMS usage, providing a unique opportunity to study AMS usage at the farm level. The dataset is an unbalanced panel. We filter the data such that farms adopting AMS are all observed for at least one year before adoption, allowing the time of adoption to be determined. Thus, our analysis includes 273 farms that did not adopt AMS and 47 that adopted AMS between 2014 and 2019, adding up to 1,594 observations. Table 1 presents some descriptive statistics for each variable used in the analysis splitting the data between farms adopting AMS before and after adoption and non-adopters. Table 1 distinguishes between structural and behavioural factors and the variables used to calculate GHG emissions efficiency. We deflate value-added and off-farm income to 2015’s consumer price index (Totalkalkylen, NIBIO).

We can obtain some first indications of the effects of AMS by testing whether there are differences in means between the farms before compared to after adopting AMS. We test for differences in means using a two-sample t-test after concluding normality using QQ plots. We do not find strong evidence to reject the null hypothesis of equal means for energy consumption, fertiliser usage and off-farm income. For the other variables, the null hypothesis can be rejected. This indicates changes after AMS is adopted and motivates further investigation.

Table 1 shows that farms that adopting AMS are different in the considered parameters than those that do not adopt. For example, farms observed before adoption already have higher energy usage, lower labour per cow and are slightly larger in the number of cows. Further, there are likely other aspects of farms which we do not consider here that determine whether farms adopt AMS or not. In this paper, we do not study the determinants of adoption, but recognising that farms adopting AMS are different from farms not adopting AMS is important for interpreting our results as effects of AMS among the farms that adopt.

3.1. Operationalisation of GHG emissions efficiency

In Norway, GHG emissions from livestock production are primarily caused by field emissions, forage production and intrinsic animal emissions (Oort van and Andrew, 2016). This is measured from enteric fermentation, manure management, feed production and energy consumption. The indicators selected for this efficiency evaluation reflect this to the extent possible, given the available data. Energy expenditures and fertiliser expenditures are divided by the price of diesel and mineral fertiliser, respectively, for each year obtained from NIBIO (Totalkalkylen, NIBIO) and are thus expressed as quantities. Enteric fermentation is calculated using IPCC methods (Eggleston et al., 2006) using values adapted to the Norwegian context from the national inventory report (NIR) from 2021 (Bjønness, 2021). Details on the calculation of the emission factors are provided in  Appendix 3.

Despite being a major contributor to farms’ GHG emissions, we do not consider manure management. Arguably, manure management systems do not differ significantly regarding GHG emissions (Soteriades et al., 2019). However, IPCC methods for calculating GHG emission coefficients vary based on manure management. The lack of knowledge regarding the manure management system employed is a deficiency, and incorporating this information into the Norwegian account results dataset would allow for a more accurate evaluation. In 2018, three of four farms used a blade spreader for manure application, indicating limited variation between farms (Kolle and Oguz-Alper, 2020).

Finally, we use value-added as the economic indicator in our GHG emission efficiency formulation. Value-added is formulated as the total value of production from agriculture, including subsidies and minus intermediate consumption. Intermediate consumption includes purchased animals, feed, seeds, fertilisers, machinery maintenance, fuel and hired labour. The animals included are all animals purchased to a farm in a year, including cattle such as calves which are bread on the farm for meat or milk production. Thus, it does not include the value of the permanent livestock. Value-added measures the remuneration to own labour, capital and land.

3.2. Operationalisation of structural and behavioural factors

This section provides definitions of the variables and a discussion about their potential relation to AMS adoption and GHG emissions efficiency. We include the structural and behavioural factors based on the hypothesis that they are secondary effects of AMS adoption. Figure 2 illustrates the relationship between the factors and AMS adoption and GHG emissions efficiency, respectively.

In the remainder of this chapter, we provide motivation for inclusion of each of the factors.

3.2.1. Herd size

Adopting AMS can be part of an expansion strategy and sometimes an unintended consequence where farmers expand to utilise the machine and recoup their investment fully (Vik et al., 2019). Previous eco-efficiency evaluations have shown a positive relation with the number of cows (Soteriades et al., 2020; Martinsson and Hansson, 2021). On the one hand, the number of cows enables farms to generate higher value-added by increasing the size of production, but, on the other hand, more cows also generate higher (total) enteric fermentation and higher energy costs as the size of production is larger. Associating herd size to GHG emission efficiency provides indications of whether it is best to have more small farms or fewer larger ones from a GHG emissions efficiency perspective. How herd size relates to GHG emissions efficiency depends on, for example, how well the farmer can manage the herd and on contextual constraints to produce as efficiently as possible and generating high value-added while keeping the GHG emissions to a minimum. It is important to emphasise that herd size and eco-efficiency are not correlated by construction, as both farms with large and small herds have opportunities to produce eco-efficiently.

3.2.2. Labour per cow

A primary motive among Norwegian farmers for investing in AMS is reducing labour and increasing work-time flexibility (Stræte, Vik and Hansen, 2017). The indicator labour per cow captures the total labour per cow, including both hired and family labour. Previous eco-efficiency assessments have studied the relation to hired labour, finding that hired labour is negatively associated with eco-efficiency (Bonfiglio, Arzeni and Bodini, 2017; Martinsson and Hansson, 2021). We hypothesise that higher family labour per cow can be negatively related to GHG emissions efficiency, as the farmer can spend less time on farm management when more physical work is required in the production. Thus, we hypothesise that our indicator of labour per cow shows a negative association to GHG emission efficiency.

3.2.3. Milk per cow and share of feed concentrates

Milk per cow and the share of feed concentrates are included as secondary effects as they can change with AMS adoption, as AMS allows for increased milking frequencies (Oudshoorn et al., 2012). The variables are related as the cows need to consume more feed concentrates if milking intervals are to be increased, which reduces grazing (Lessire et al., 2020) but increases milk yield per cow. There are conflicting findings on whether farms with AMS import more high-energy feed (Oudshoorn et al., 2012). Both factors reflect managerial decisions on the level of intensity and potentially affect GHG emissions efficiency: Milk per cow enables farmers to generate more agricultural value relative to the number of cows, and the share of feed concentrates fed to the cows affects their enteric fermentation.

3.2.4. Off-farm income and arable land

The share of income derived from non-farm sources and the farmland used for grain and cash crop production relative to milk output are measures of specialisation, which might change when the farmer invests in AMS. Many Norwegian farmers seek income elsewhere because their farms are typically small and produce little economic value (Oort van and Andrew, 2016). Off-farm income reflects the farmers’ focus on farming relative to other income-generating activities, and the farmland used for grain and cash crops reflects the degree of farm specialisation. We hypothesise that investing in AMS will increase the farmers’ dairy-focus and thus decrease off-farm income and arable production relative to milk. Specialisation has been found to be negatively related to environmental performance on a sector-level considering the milk yield relative to beef output on dairy farms (Soteriades et al., 2019). Thus, we hypothesise that arable land per milk output can have a positive association to GHG emissions efficiency. Farmers engaged in off-farm activities have higher labour opportunity costs, which could create an incentive for managing the farm more efficiently. To our knowledge, specialisation in terms of off-farm income and share of arable production has not been related to eco-efficiency previously.

4. Results and discussion

In this section, we present the results of our four-step procedure. First, we answer the question of what structural and behavioural factors can be attributed as secondary effects of AMS adoption. Second, we answer whether AMS adoption generates changes in farms’ GHG emissions efficiency, which can be associated with structural and behavioural changes. In the analysis, we use the standardised form of the structural and behavioural variables obtained by subtracting the mean and dividing by the standard deviation. As the variables are expressed in different units, using their standardised version enables an easier comparison of the effects. We explain the interpretation of the standardised variables in the following sections. 33.9 per cent of our data is imputed to obtain counterfactual outcomes using matrix completion, comparable to 25 per cent missing entries in the example provided by Athey et al. (2021) to illustrate their method.

4.1. Identifying structural and behavioural factors as secondary effects of AMS

From the matrix completion procedure, we obtain ATT estimates for each factor. As we use the variables’ standardised form, the effects express average changes in standard deviations with AMS adoption. The standard deviations for each factor are displayed in Table 1. For example, the number of cows is increased by 0.76 standard deviations on average when AMS is adopted, corresponding to 8.7 cows (0.76*11.5). Beside ATT effects, we also compute the effect of AMS for each observation as the difference between the observed and the predicted outcome. Results indicate a substantial heterogeneity in the estimated effects around the ATT effects (Figure 3).

Earlier studies already documented that farms with AMS have larger herd sizes than farms without (Rønningen, Magnus Fuglestad and Burton, 2021). We can add to these findings that differences between farms with and without AMS are not only due to baseline differences but that adopting AMS is associated with an enlargement of the herd. Vik et al. (2019) previously found that farms with AMS expand through a qualitative study, and our results support this finding. Moreover, we find that the adoption of AMS positively correlates with increased milk production per cow, as expected, given that AMS adoption offers the potential to increase milking frequency (Oudshoorn et al., 2012). Further, the share of feed concentrates in the diet is rising. The ATT is negative for labour per cow and arable land per milk output. This indicates a higher intensification and specialisation in dairy production compared to crops. At the same time, the negative ATT of areal production seems to be driven by some negative outlier observations. Off-farm income displays a small negative ATT highly centred around zero. Considering the effect for each observation indicates that for most factors, the direction of the effect is clear, while the magnitude of the effect is heterogeneous. Our findings indicate that farms develop in similar directions in the structural and behavioural variables after AMS adoption.

Nevertheless, interpreting the effects as causal should be done with caution. We do not account for potential interrelation between the change in these variables, as we estimate the effect in separate models. This must be considered when concluding, as the change in one variable from AMS adoption might be affected or driven by the change in another.

4.2. Identifying the effect of AMS adoption on GHG emissions efficiency and the relation to the structural and behavioural factors

We investigate how farms’ GHG emissions efficiency is affected by AMS adoption and to what extent we can attribute this to structural and behavioural factors. We first present the results of the GHG emissions efficiency evaluation and how this is affected by AMS adoption. This is followed by the estimated association between the GHG emissions efficiency and the structural and behavioural factors, coming together in Table 3 to answer our question of how AMS affects GHG emissions efficiency and if this can be associated with the effects of AMS adoption observed in the structural and behavioural variables.

The average bias-corrected GHG emissions efficiency for the complete sample is 0.47, indicating room for improvement. Farms could reduce environmental pressures by 53 per cent while maintaining value-added to become fully efficient. Table 2 shows the results of the GHG emissions efficiency assessment for the entire sample and when separating adopters and non-adopters. Note that no observation is on the frontier, resulting from bootstrapping.

Non-adopting farms have the highest mean efficiency, whereas the adopting farms after adoption have the lowest. We test the differences in means between the three groups in Table 2 using t-tests. We can reject the null hypothesis of equal means between AMS adopters and non-adopters before and after adoption. Comparing the adopters before AMS with the adopters after AMS using the same t-test yields a p-value of 0.2.

Figure 4 displays the effect of AMS on GHG emissions efficiency generated from the matrix completion. The root mean squared error (RMSE) is 0.17. The ATT of AMS adoption on GHG emissions efficiency is −0.015, indicating that the average effect of adopting AMS among the adopters is to decrease GHG emissions efficiency by 0.015. However, computing the effect of each observation, it is evident that there are large heterogeneities in the results. Figure 4 shows that the small average effect is not due to an absence of impact but rather to the large heterogeneity in responses.³

As the matrix completion procedure does not consider dynamic effects (Athey et al., 2021), Figure 4 does not display differences in the effects given time since AMS adoption.

Finally, we aim to attribute the effects of AMS adoption on GHG emissions efficiency to the structural and behavioural factors to provide insights into the mechanisms behind the changes in GHG emissions efficiency. Correlating the changes in GHG emissions efficiency to the structural and behavioural factors can provide explanations for the large heterogeneity between farms, and potentially contradictory effects can explain the small average effect.

We obtain the contributing power of each factor to the association between GHG emissions efficiency and AMS by multiplying the relationship between each factor and GHG emissions efficiency (step three) by the change in each factor associated with AMS (step one). Still, we use the standardised forms of structural and behavioural factors. For the number of cows, this would indicate that when increasing the number of cows by one standard deviation (11.5, Table 1), GHG emissions efficiency is expected to change by −0.02. Thus, the increase in herd size induced by AMS (0.77 standard deviations) contributes to a decrease in GHG emissions efficiency by 0.77 * (−0.02) =  −0.015. We conduct this procedure for each factor listed in Table 3.

The total effect displayed in Table 3 is obtained by summing up the contributing power of all factors. This total effect of −0.011 is smaller in absolute terms than the estimated effect of AMS on GHG emissions efficiency in step 3, which we estimate to be −0.015. Thus, other processes may be at work which are not reflected by the structural and behavioural variables we include in the model. Nevertheless, our procedure demonstrates the relative importance of the factors as secondary effects affecting farms’ environmental performance. In Table 3, the structural and behavioural factors show contradicting relations to GHG emissions efficiency, which can explain the small average effect of AMS adoption on GHG emissions efficiency.

The second column displays the result from the linear regression outlined in Equation 2. We generate the results using a two-way FE regression. Despite the contributing power of each factor being relatively small, they are not irrelevant. Small decreases in farm-level GHG emissions efficiency can substantially increase total emissions. However, as displayed in the regression plot in Figure 5, the confidence intervals are large indicating that more data would be required to increase the precision of the result. For example, access to a balanced panel data of the farms between 2013 and 2019, or a panel data covering a longer period of time, could help narrowing the confidence intervals. Not being able to reject the null hypothesis that there is no relationship between changes in a factor and changes in the GHG emissions efficiency calls for further investigation.

Decreasing the labour per cow is associated with higher GHG emissions efficiency and constitutes the largest contribution to the effect of AMS adoption on GHG emissions efficiency. The negative association between labour per cow and GHG emissions efficiency is consistent with prior findings (Bonfiglio, Arzeni and Bodini, 2017; Martinsson and Hansson, 2021). The strong association between labour per cow and changes in GHG emission efficiency generated by AMS comes both from a relatively large effect of adopting AMS, but also from a large estimated association with GHG emissions efficiency. Thus, decreasing the labour per cow on a farm due to AMS contributes positively to farms’ GHG emissions efficiency. Whereas the previous findings of a negative relation between labour per cow and eco-efficiency has focused on a cross-sectional setting, our findings indicate that it is not only the case that farms with more labour per cow also have lower eco-efficiency, but that there is also an association between these variables when considering changes within farms.

The number of cows also displays a large contribution to the changes in GHG emissions efficiency, mainly through the effect generated by adopting AMS. The result of a negative association between number of cows and GHG emission efficiency contradicts previous findings such as Martinsson and Hansson (2021). One reason for the negative association with GHG emissions efficiency could be that increasing the number of cows increases enteric fermentation. Another reason could be that diesel consumption is higher on larger Norwegian farms (Oort van and Andrew, 2016), possibly due to long distances between the farm centre and the most distant fields. Nevertheless, further research is required to investigate why larger farms are less GHG emissions efficient. Extension services, farmers and researchers need to find ways to achieve herd expansion combined with higher GHG emissions efficiency.

The contradicting effect of labour per cow and herd sizes illustrates that the small estimated effect of AMS adoption on GHG emissions efficiency (Figure 4) is not due to a lack of processes, but rather that the adoption of AMS is associated with several processes that have contradicting effects on GHG emissions efficiency.

Milk yield per cow and the share of feed concentrates have contradicting effects and are nearly cancelled out, as seen in Table 3. The changes in these variables are likely a result of increased milking frequency, which is enabled by AMS adoption (Oudshoorn et al., 2012). Previous research has identified milk yield per cow to be positively associated with eco-efficiency (Soteriades et al., 2020), which the findings of this study support. Achieving the maximum milk yield with the smallest proportion of feed concentrates can support increasing GHG emissions efficiency. Thus, we must identify ways to increase milk yield without increasing feed concentrates to affect GHG emissions efficiency positively.

Off-farm income and areal production per milk output have little influence on the association between GHG emissions efficiency and AMS adoption, as they are estimated to have very low correlation to GHG emissions efficiency and change little with the adoption of AMS.

To sum up, the most GHG-emissions-efficient farms in the sample have lower labour per cow, smaller herds, higher milk yield and a lower proportion of feed concentrates in the dairy cow feed ratio. These results indicate a challenge for the Norwegian dairy industry in maintaining GHG emissions efficiency as demand for dairy products increases and mechanisation drive farms to expand in size (Vik et al., 2019). Understanding the importance of different structural and behavioural factors that could drive the overall change in GHG emissions efficiency provides valuable insight into the significance of various underlying processes triggered by the adoption of novel technology. Based on this, future research efforts and extension services can be more precisely targeted. Investigating the correlation between the GHG emissions efficiency and the structural and behavioural changes, we do not find any indications that the included variables can explain the considerable heterogeneity in farms’ GHG emissions efficiency responses to AMS adoption.

5. Conclusion

We present a novel analytical approach for evaluating farm-level effects of new technology adoption, including secondary effects. We use matrix completion and DEA to assess the impact of adopting AMS on structural and behavioural factors and farm GHG emission efficiency. To our knowledge, this paper is the first to empirically demonstrate the presence of secondary effects and use eco-efficiency evaluation to assess the effects of new farm technology. This procedure can be used in future research to evaluate other technologies.

We identify that AMS adoption generates secondary effects in our sample of Norwegian dairy farms. After AMS adoption, farms increase herd sizes, the share of feed concentrates in the cows’ diet and milk yield per cow while decreasing labour per cow. The average effect of AMS adoption on GHG emission efficiency is small (−0.015) but with significant heterogeneities across observations. The negative and heterogenous relationship between AMS adoption and GHG emissions efficiency is a novel finding that highlights the importance of evaluating farm-level effects of novel technology, including the possibility of secondary effects, as they can have unexpected effects on GHG emissions efficiency. For agriculture to develop towards increased environmental sustainability while ensuring economic viability, novel technology must contribute to improved farm-level eco-efficiency.

By providing a link between AMS adoption and structural and behavioural factors, our study enables more precise steering of the development of farms to achieve desired policy targets given AMS adoption. Furthermore, linking structural and behavioural factors to GHG emissions efficiency allows for deriving insights into which processes to target for eco-efficient AMS implementation. We recommend providing extension services to enable farms to expand while maintaining or increasing their GHG emissions efficiency and increasing milk yield per cow without increasing the share of feed concentrates. Further, it is interesting to note the considerable heterogeneity in the effects of AMS adoption on GHG emissions efficiency, which calls for further research to determine what drives farms to increase or decrease their efficiency after adopting AMS in particular and novel technology in general. Future research should also invest in understanding the mechanisms behind the secondary effects of novel technology to enable forecasting the development of farms as more autonomous technology is developed and made available to farmers. For example, expanding the procedure to consider interrelations between the variables as AMS is adopted would be a helpful extension adding further understanding to the effects of AMS adoption. Further research should also expand the approach by Kuosmanen and Kortelainen (2005) to implement more realistic assumptions of non-radial changes of variables in the very short run.

Acknowledgements

This research is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy - EXC 2070–390732324.

Footnotes

References

Appendix A: PS DID and fixed effects regression

We apply PS-DID and FE regression to estimate how AMS adoption affects GHG emissions efficiency. Comparing the more unconventional machine learning approach to classic econometric methods improves the transparency and understanding of the results. The main differences are that the econometric methods provide statistical significance and confidence intervals, which the matrix completion currently lacks, and that the matrix completion can make use of the full dataset and utilise the variations there, while the econometric methods are more limited in this respect. We apply the econometric methods to assess the effect of AMS adoption on GHG emissions efficiency to illustrate the differences between the methods. In this appendix, we briefly outline the PS-DID and FE regression. Both methods produce similar results as the matrix completion approach, showing that our results are robust across different methodologies.

PS DID. As AMS adoption is not randomly distributed, we use propensity scores to control for factors which impact the adoption decision. Including propensity scores help to realise the parallel trends assumption, which is a prerequisite for a DID regression. The variables used for calculating the propensity scores are displayed in Table A.1.

Further, DID requires treated and control groups to be observed in two periods—one period before adoption and one after. Thus, we subset the dataset to only keep observations the first and the last time they are observed. The average treatment effect of adoption can be determined by comparing the outcomes of adopters and non-adopters (Khandker, Koolwal and Samad, 2009). The DID regression can be expressed as |$\Delta E{E_i} = \,{\alpha _i} + {\beta _1}(adoptio{n_i}) + \,{e_i}$|⁠. Adoption is a binary indicator of whether the farm is an adopter or not. PS are included as weights in the regression.

To test the robustness of the PS and the sensitivity of their specification to the result, we consider two versions of the DID using PS calculated through two different methods: Full matching and inverse probability weights (IPW). We use the Matchit package in R (Stuart et al., 2011) to obtain propensity scores and conduct the full matching. We calculate the IPW weights as |$IP{W_i} = \,\frac{{Adoptio{n_i}}}{{P{{\left( X \right)}_i}}} + \,\frac{{1 - Adoptio{n_i}}}{{1 - P{{\left( X \right)}_i}}}$|⁠. |$P{\left( X \right)_i}$| is the propensity score and |$Adoptio{n_i}$| is, as before, a binary indicator of whether farm i is an adopter or not. The matching satisfies the condition of common support as depicted in Figure A1.

Recent methodological contributions have suggested ways to expand the method to allow for staggered adoption by e.g. formulating different groups of adopters adopting in different times (Callaway and Sant’Anna, 2021). However, this comes with the issue that it requires sufficiently large groups of adopters in the different years. In our application, this would result in small subsets of farms adopting in each year.

FE regression. The other alternative we consider to assess the impact of an intervention is to use a two-way FE regression, including a dummy variable for the change in adoption status. The dependent variable is GHG emissions efficiency, and the explanatory variables are years since adoption, with one dummy for each year relative to adoption. We can create dummy variables of whether farms are in the years before adoption, as the actual adoption is preceded by a decision to adopt which can generate adaptive changes already before the AMS is implemented on the farm. However, the estimates do not reflect any counterfactual outcomes, but only consider how AMS affects the adopting farms. We include dummies for up to four years before adoption and four years after adoption, with an omitted reference period of five years before adoption. The regression can be written as |$E{E_{it}} = {\rm{\,}}{\alpha _i} + {\varphi _t} + {\rm{\,}}{\beta _1}(yr\_since\_adop{t_{it{\rm{\,}}}}) + {\rm{\,}}\mathop \sum \limits_{n = - 4}^4 {\beta _n}dummy\_yr\_since\_ado{p_{it}} + {\rm{\,}}{e_{it}}$|⁠. Farm and year-fixed effects are included, denoted by α and φ in the FE regression equation, respectively. This method only considers variation among farms whose adoption status changes, i.e. farms that adopt AMS; non-adopting farms are excluded. The results of the PS DID and FE regression are displayed in Figure A2.

Appendix B: Second stage DEA using Tobit regression

A commonly used alternative to OLS regression in DEA efficiency assessments is Tobit (Hoff, 2007; McDonald, 2009). As Tobit has been identified as a feasible alternative to OLS, we apply it as a robustness check to our OLS estimation used in the paper. While McDonald (2009) reached the conclusion to not use Tobit in the second stage DEA, Hoff (2007) identified Tobit as sufficient in representing second stage DEA assessments. Tobit regression is best used with censored data or where corner solutions are present, which is the case with DEA efficiency scores ranging between 0 and 1 (Hoff, 2007). The results from the Tobit regression are displayed alongside the results from the fixed effects OLS regression and shown in Table A.2.

Appendix C: Emissions factors calculation

To calculate methane emissions from enteric fermentation, we use equations provided by the IPCC (Eggleston et al., 2006) adapted to the Norwegian context in the NIR (Bjønness, 2021). We use the following equation to calculate enteric fermentation:

where |$E{F_i}$| denotes emissions factors for animal category i, and |${N_i}$| is the number of individuals in that category. To calculate enteric fermentation, we use the variables number of cows (including heifers and bulls), litre milk yield (including milk sold, milk consumed on the farm and waste), feed concentrates and roughage expressed in feed units from the Norwegian account results dataset. Emissions factors are calculated using the IPCC equation:

where GE is gross energy intake and |${Y_m}$| is the methane conversion rate and depend on the climatic conditions and geographical area considered. To enable calculating these factors in the Norwegian context, we use estimates from the Norwegian NIR (Bjønness, 2021) which provides equations for calculating the GE and |${Y_m}$| for dairy cattle, and provide estimations of the emissions factors for other livestock. Following the Norwegian NIR, we use the following equations to calculate GE, and |${Y_m}$|⁠:

|${\rm{Milk}}305$| is the energy corrected milk yield during the 305 days long lactation period. Due to data availability, we use the total yearly milk yield without correcting for energy content. In our calculation of enteric fermentation, we also include heifers and bulls, by multiplying the number of individuals in each category with emissions factors calculated using gross feed intake and methane conversion rate calculated for the Norwegian context from the NIR report (Bjønness, 2021). Thus, in total there are five categories of livestock included when we calculate a farms’ enteric fermentation; dairy cows, heifer >1 year, heifers <1 year, bulls >1 year and bulls <1 year.