Risk attitudes and innovation in Dutch arable farming

Abstract

On-farm innovations have the potential to increase both agricultural productivity and income while mitigating environmental impacts. However, as adopting innovations can lead to risks, it is important to understand the role of farmers’ risk attitudes. We quantified Dutch arable farmers’ risk attitudes based on the cumulative prospect theory (CPT) and used the benefit of the doubt approach to obtain an innovation index based on expert elicitations and adopted innovations. Subsequently, we used a fractional response model to test our pre-registered hypothesis and investigate the association between the parameters of farmers’ risk attitude and farm-level innovation. We find no statistically significant association between CPT parameters and the innovation index. Our results therefore cast doubts on the potential of revealed risk preferences to understand real-world behaviour.

1. Introduction

Faced with the challenges of climate change, resource scarcity, and growing food demand, the development, diffusion, and adoption of agricultural innovations¹ provides a much-needed opportunity for sustainable intensification (Balaine et al. 2020; Alston and Pardey 2021; Finger 2023). One key variable for understanding agricultural innovation is farmers’ risk attitude (Ghadim et al. 2005; Dessart et al. 2019). To better represent on-farm innovation and improve economic farm-level models, an accurate measure of farmers’ risk attitude is needed. This knowledge is also essential when designing policies to promote innovation. While sharing many similar aspects, the literature on agricultural innovation and risk attitudes has seldomly been connected (Lemessa et al. 2019). In the innovation adoption literature, risk attitudes are often measured based on the expected utility theory, assuming that farmers decide by calculating the utility of each option, given a certain degree of risk aversion (or seeking) (Marra et al. 2003; Chavas and Nauges 2020; Streletskaya et al. 2020). Most findings suggest that risk aversion implies lower levels of adoption (e.g. Marra et al. 2003; Brick and Visser 2015). However, recent research has found that farmers’ risk attitudes are better explained by the cumulative prospect theory (CPT) (Tversky and Kahneman, 1992; Rommel et al. 2023). Human biases assumed in the prospect theory, i.e. loss aversion, status quo bias, and the overweighting of small probabilities, may provide potential additional explanations for a low uptake of innovations (Streletskaya et al. 2020).

We used country data from a multi-country replication of the widely cited study of farmers’ risk attitudes by Bocquého et al. (2014, for the replication, see Rommel et al. 2023). The elicitation of farmers’ risk attitudes in the replication study followed the approach developed by Tanaka et al. (2010) and was based on multiple price lists. The Dutch version of the survey in the replication study focussed on arable farms and included several Dutch-farmer-specific questions on innovations. Läpple et al. (2016) report that regional factors play a role in innovation activities; arable farming in the Netherlands is an interesting example, as highly intensive agriculture has led to environmental problems that can be tackled by promoting agricultural innovations (Government of the Netherlands 2020), while creating a situation of policy uncertainty (Yanore et al. 2022; Sok and Hoestra 2023). This paper addresses the question to which degree risk attitudes and on-farm innovation are associated with each other. Our pre-registered hypothesis is that farmers who are more willing to take risks are also more innovative when it comes to farm-level innovation, meaning they adopt more innovative and harder-to-implement innovations that originate from developers and initiators with a high innovation capacity. In particular, we expect that higher levels of loss aversion and the overweighting of low probabilities will result in less innovation.

Currently, we lack a systematic understanding of how risk attitudes map into farm-level innovation. Existing work on innovation in agriculture often focusses on the adoption of individual innovations rather than on the farm level and thus neglects the complexity of innovations beyond adoption decisions (Diederen et al. 2003; Klerkx et al. 2012; Läpple et al. 2015; Groher et al. 2020; Spiegel et al. 2021). This complexity includes, for instance, who initiated an innovation and who developed it, but also how innovative and difficult to implement a technology is. The few empirical studies on CPT and individual innovation adoption behaviour show mixed results. Liu (2013) reports that more risk-averse or more loss-averse Chinese cotton farmers adopt Bt cotton later, while farmers who overweight small probabilities adopt earlier. Holden and Quiggin (2017) investigated the adoption of drought-tolerant (DT) maize by Malawian farmer households. Both more risk-averse and more loss-averse households were more likely to adopt DT maize, while subjective probability weights were not significantly associated with the adoption decision. Anand et al. (2019) used a simulation to show the importance of loss aversion for the adoption of bioenergy crops in the USA. Transferring these results to farmers in the Netherlands is complicated by the substantially different context, both in terms of the country and the level of the decision (Höhler and Müller 2021). Moreover, the relevance of risk attitudes for predicting real-world behaviour has recently been challenged (Hertwig et al. 2019; Charness et al. 2020).

Our contribution is twofold: (1) Using a multidimensional farm innovation index based on the benefit of the doubt (BoD) approach (Cherchye et al. 2007), we advance knowledge of the association between risk attitudes and innovation from the level of individual innovation adoption to the farm level. Indexing allows us to look at not just one but a variety of innovations and to consider not just the adoption but also the process leading to adoption and the innovativeness of the technology. (2) We add to the CPT and innovation literature by advancing the understanding of the role of risk aversion, loss aversion, and probability overweighting in innovation adoption. Policy programmes to promote innovation, as in the case of the Dutch policy to reduce nitrogen pollution, can benefit by being able to set more targeted incentives and to foster innovation at the farm level. In Section 2, we explain CPT, the measurement of innovation, and our empirical approach. This is followed by a description of the replication study data collection and by the presentation of descriptive statistics. The results section is divided into results of estimating risk attitudes, calculating the innovation index, and examining the association of the two sets of results. We end with a discussion of the results and policy implications.

2. Methods

2.1 Behaviour under risk

CPT has three parameters: (1) σ is an anti-index of concavity for gains, where σ < 1 indicates risk aversion in the gain domain; (2) λ is the coefficient of loss aversion, where values of λ > 1 imply that the decision-maker is more sensitive to losses than gains; and (3) γ reflects the curvature of the probability weighting function, with values of γ < 1 implying probability distortion (overweighting of small probabilities and underweighting of large probabilities). Following the replication study by Rommel et al. (2023), we assume the status quo as the only reference point. The utility function of CPT is defined as follows:

with y as the payoffs from a risky tradeoff.

The decision weights are defined over cumulative probabilities. The value of any binary lottery with payoffs y₁ and y₂ and a probability P of getting y₁ is defined as:

with ω as the probability weighting function, which is strictly increasing from the unit interval into itself, satisfying ω (0) = P₀ = 0 and ω (1) = P₁ = 1. We followed Tanaka et al. (2010) and used the one-parameter weighting function of Prelec (1998), i.e. for any P  > 0 and P  ≤ 1 the probability weights ω are defined as follows:

with γ as the parameter that controls the curvature of the probability function (γ > 0). Bocquého et al. (2014) used maximum likelihood for the estimation, which we used in our analysis.

2.2 Innovation index

We use an adapted version of the innovation index of Läpple et al. (2015) that combines farm data with expert elicitations to account for the complexity of innovation. The index comprises assessments of the innovations with regard to different dimensions and thus simplifies a comparability of the different innovations. Our innovation index, Innov_f, for a specific farm f consists of three terms and is defined as follows:

The first term includes P_g as the level of innovativeness of innovation g, the difficulty of its implementation, q_g, and z_gf as a dummy variable that indicates whether it was adopted on farm f in 2020. The second and third terms capture the acquisition of knowledge, a crucial factor in the innovation process (Läpple et al. 2015). The terms describe the innovation capacity x of the developer d and the innovation capacity z of the initiator r. The weights w are determined endogenously with the BoD approach, which results in the advantage of unit invariance. The lower and upper bounds of the weights are determined by expert elicitations (Cheryche et al. 2007).

2.3 Fractional response model

Läpple et al. (2015) use a Tobit regression to investigate the drivers of their agricultural innovation index. The use of Tobit regressions and the underlying assumption of a censored dependent variable have been a much-disputed subject within the field of efficiency analysis (Simar and Wilson 2007; Ramalho et al. 2010). We follow the suggestion of Ramalho et al. (2010) and use a fractional response model (FRM) instead of a Tobit model. First introduced by Papke and Wooldridge (1996), the model is based on the assumption:

with G(·) as a known non-linear function that satisfies 0 < G(|${x}_i\beta $|) < 1.

The parameters are estimated by a quasi-maximum likelihood estimation based on the following Bernoulli log-likelihood function:

In the first specification, we only include the BoD estimate as a dependent variable and the three parameters on behaviour under risk, namely risk aversion, loss aversion, and probability weighting, as independent variables. In the next step, we include selected covariates from the replication study. For robustness checks, we change the dependent variable to a dummy variable, indicating innovative farms with a BoD estimate larger than 0.5. The robustness check uses a logistic regression model.

3. Data collection

Data on risk attitudes and innovations were collected in an online survey in June 2021. The participants were Dutch arable farmers who own, rent, or lease agricultural land for which they make land use and investment decisions. Participants were recruited via a specialized agency and were randomly selected from their database. The invitation was sent to 5,000 arable farmers; the final sample consists of 154 farmers (response rate = 3 per cent). Following the original study of Bocquého et al. (2014), the replication study featured three multiple price lists, of which two included only positive payoffs while the third also contained negative payoffs. In addition to the survey from the replication study, we provided the sample of Dutch farmers with a list of eleven different innovations, including their definitions.² The innovations and related survey questions were selected on the basis of the annual Dutch innovation monitor among Dutch farmers (Agrimatie.nl 2023). We asked participants to indicate which innovations they had adopted on their farm in 2020. We also asked for the developer of the innovation; whether the innovation was developed on their farm, in a separate enterprise, in cooperation with other enterprises, or mainly or exclusively by other enterprises. In addition, participants had to indicate on whose initiative it was introduced (the initiator). For this question, the choice was between their own farm, a fellow farmer, a supplier, a customer, an advisor, a research institute, or another initiator (see Appendix A2 for the additional Dutch survey instrument).

Table 1 shows an overview of the covariates: farmer age, farm size in terms of total arable area, education, and owned land. The average arable farm in our sample is larger than the 2021 national average of 41.4 ha (Agrimatie.nl 2022). From the variables collected by Rommel et al. (2023), we further used education, general trust towards other people (dummy), and the proportion of land owned. Moreover, we use information on farmers self-assessed risk attitudes for an additional robustness check. This self-assessment included four items: one on risk attitude in general and three on the domains of farm production, financing and investment, and marketing and selling agricultural products.

The study design was pre-registered at aspredicted.org (Anonymous 2021). In the lottery task, participants had to select which row in each of the three multiple price lists they desired to switch from the safer lottery (Option A) to the riskier lottery (Option B). The farmers received an initial sum of money (expressed in points) and could lose or earn money as a result of their decisions in the lottery task. The Dutch farmers were paid via bank transfer and earned an average of 16.09 Euro, ranging from 4.35 Euro to 97.50 Euro (see also Rommel et al. 2023).

The values for the BoD approach as shown in equation (3) were based on interviews with eleven experts on Dutch arable farming and farm innovation recruited via e-mail from our networks. Using Likert scales, the experts (1) estimated the weights of the sub-indices in the innovation index, (2) evaluated the innovativeness and difficulty of implementation of the innovations, and (3) the innovation capacity of their innovators and developers.³

4. Results

4.1 Behaviour under risk

The coefficient estimates for the Dutch sample were taken from the original study by Rommel et al. (2023). Based on the mid-point approach by Tanaka et al. (2010), CPT parameters were estimated per respondent. Table 2 shows the estimates and 95 per cent confidence intervals in brackets, as well as statistics of model fit for the whole sample and for the Dutch sub-sample. Information on the distribution of estimates for the three parameters under the mid-point approach can be found in Appendix A3.

Compared to the complete dataset, the Dutch sample shows a lower degree of loss aversion, as expressed by lambda. Farmers in both samples are risk averse—sigma is <1—and overweigh small probabilities, indicated by gamma <1. The distribution of the parameters is displayed in Rommel et al. (2023).

4.2 Innovation index

On average, farmers adopted 1.64 of the mentioned innovations in 2020. The maximum number of adopted innovations in 2020 was 7. Most of the farmers indicated that these innovations were developed by other enterprises or institutions (52 of 160), by themselves (24), or mainly by others (22). The initiative was taken by farmers (92 of 160), suppliers (5), or others (4).

Table 3 provides an overview of the type and frequency of adoption. The category ‘Other’ was an open text field. Responses included a washing area to prevent yard emissions and CO₂ cooling.

Adoption of most innovations is only very weakly correlated (correlation coefficients <0.2) with each other, while some innovations show moderate correlation coefficients (CO₂ cooling and storage computers: 0.56, Global Positioning System (GPS) and precision farming: 0.46, sensor and precision farming: 0.41).

Table 4 presents the descriptive statistics for the composite indicator computed with the BoD approach. The average composite score is 0.598 with some farmers reaching the maximum of 1 and others with a composite score of 0.

4.3 Fractional response model

We tested the association between the BoD estimate and the estimated CPT parameters. As shown in Table 5, none of the three CPT parameters shows a statistically significant coefficient.

A larger farm size and a lower share of owned land appear positively associated with the BoD estimate. As a robustness check, we used a dummy variable for a BoD estimate larger than 0.5 ((5), (6)) as a dependent variable. The association of farm size with farm-level innovation is stable. The CPT parameters are not statistically significant in any model.

An additional robustness check included four items on farmers’ self-assessed risk attitudes as independent variables, one on risk attitude in general, and three on the domains of farm production, financing and investment, and marketing and selling agricultural products (see Appendix A1). We ran the model with the BoD index and a dummy for an index value larger than 0.5 as dependent variables. None of the items showed a statistically significant coefficient.

5. Discussion

Our study was designed to determine the association between risk attitudes and farm-level innovation using data on Dutch arable farmers’ revealed risk attitudes and their adopted innovations in 2020. Risk attitudes are seen as a critical factor in understanding farmer behaviour (Bocquého et al. 2014). Our findings do not confirm most of the empirical evidence on the role of CPT parameters in the agricultural innovation adoption literature. In contrast to the studies by Holden and Quiggin (2017) and Liu (2013), risk aversion and loss aversion are not significantly associated with innovation. In line with Holden and Quiggin’s (2017) findings on individual adoption decisions, subjective probability weights were not significantly associated with innovation. Our results reflect those of Groher et al. (2020) and Läpple et al. (2015), who also found that large farm sizes are associated with higher adoption rates and innovation. Overall, our findings support the notion of the limited ability of risk attitudes to predict real-world behaviour (Hertwig et al. 2019; Charness et al. 2020).

There are several potential explanations for our null results: (1) Since risk attitudes are highly domain-specific, the predictive ability of generic measures is limited (Weber et al. 2002). Self-reported, domain-specific risk measures have been shown to be a stronger predictor of real-world behaviour (Weber et al. 2002; Dohmen et al. 2011; Hertwig et al. 2019). However, we did not find a statistically significant association between self-reported risk attitudes and estimated index values. In our case, the predictive ability could be further diminished by the fact that we were not looking at the innovation level but at the farm level, and that we also included other factors of the innovation process beyond the adoption decision, such as the innovation capacity of initiators and developers. (2) Other drivers of innovation, such as ambiguity aversion, environmental concerns, farming objectives, credit constraints, or risk perception could have a higher explanatory power (Marra et al. 2003; Anand et al. 2019; Dessart et al. 2019; Chavas and Nauges 2020). (3) Since CPT rests on strong assumptions, its empirical adequacy to describe human choice under risk might be limited (Harrison and Ross 2017; Charness et al. 2020). We have followed the standard assumption that the status quo is the only reference point; we further assumed that σ and γ do not differ between the loss and gain domains. Moreover, different estimation strategies can lead to very different parameter estimates for the same data (Rommel et al. 2023). (4) Our sample size may prevent us from detecting small effects. (5) The role of risk attitudes in farm-level innovation is ambiguous. Some of the innovations will increase risks, but others might reduce risk, providing risk-averse farmers with an incentive for adoption.

Our results suggest the need for caution in the use of elicited risk attitudes, both self-reported and lottery-based, in policy making. The findings do not support the idea of segmenting farmers according to their risk attitudes for designing effective policy measures (Dessart et al. 2019), at least not based on a generic measure of their risk attitude from a single elicitation method (see also Finger et al. 2023). Similarly, model simulations of innovation uptake support programmes should not be based on such measures. Farm size appears to have a decisive lever in farm-level innovation. In this context, it is worth discussing whether our results argue in favour of supporting smaller farms or further supporting the already more innovative larger farms. Beyond that, it is valuable to reflect on what the larger farm size represents; economies of scale, a particular business strategy, or a type of farmer? The farm size in our sample is more than twice the national average. Self-selection of interested farmers in the sample cannot be excluded. A representative sample or weighted estimates could provide additional insights.

Combining multiple innovations into one index allowed us to make more robust statements. At the same time, we grouped together very different innovations whose innovation processes can vary greatly depending on the characteristics of the farmers and those of the innovations themselves (Diederen et al. 2003). As a consequence, risk attitudes may well play a large role for the adoption of some of the innovations, but not for others. In addition, some farmers may adopt different portfolios of innovations than others. As discussed in Rommel et al. (2023), different estimation strategies for the CPT parameters result in different estimates. The mid-point approach provides insights into the distribution of the parameters. Using an interval regression could make use of these values and provide further insights.

6. Conclusion

Our aim was to investigate the association of risk attitudes specified by three CPT parameters and farm-level innovation expressed by a BoD index. We were unable to show a statistically significant association of risk parameters and BoD estimates. Larger farm sizes were associated with higher index values, suggesting that they are more innovative. Our findings were robust for different model specifications (see Appendix A1).

Endnotes

Acknowledgements

We thank the farmers for their participation in the survey.

Data availability

Data and code available on request.

References

Appendix

A1. Robustness check with self-assessed risk attitude

A2. Survey instrument on innovation

Part IIIrealisedinnovations

At last, we want to find out whether you realised any process innovations at your farm in 2020. Please read the following definitions and answer the questions below.

Process innovations are new or clearly improved technologies and new or clearly improved methods of producing and delivering products. Examples are new machines, installations or barns and other supplies, or computer systems. The innovation must be new for your farm, but not necessarily new for the market or sector.

* Whether something is innovative does not depend on the amount of the investment.

* Whether something is innovative does not depend on the method of financing.

* Whether something is innovative does not depend on the ownership situation.

Which process innovation(s) did you introduce on your farm in 2020?

Who developed this process innovation(s)? Tick the most appropriate answer.

Whose initiative was it to bring about this innovation(s)? Tick the most appropriate answer.

A3. Kernel density estimates and boxplots of the distribution of CPT parameters, using the mid-point approach (see also Rommel et al. 2023)