Heterogeneity of European farmers’ risk preferences: an individual participant data meta-analysis

Abstract

We present a new approach to establish an empirical overview of farmers’ risk preferences and the characteristics associated with these preferences. We rely on an Individual Participant Data Meta-analysis whereby we identify studies eliciting risk preferences through self-assessments and Holt and Laury lotteries, and construct and analyse a unique dataset of 5,157 farmers from 19 studies in 13 European countries. Our results reveal significant heterogeneity in risk preferences across elicitation methods, within and across studies, risk domains and farm and farmer characteristics. European farmers are on average risk-averse when risk preferences are elicited by lotteries, and on average risk-neutral when elicited by self-assessments. Beyond differences in the average risk aversion, there are distributional differences pointing to a limited convergence between elicitation methods and a larger explanatory power of farm and farmer characteristics to determine risk preferences elicited with self-assessments compared to lotteries.

1. Introduction

The wide range of uncertainties farmers face and their increasing nature (e.g. pest pressure, extreme weather events and food price volatility) have increased researchers’ interest in farmers’ risk preferences. Risk preferences have been considered a key factor for farmers’ management, marketing, investment and adoption decisions. Moreover, their recognition is crucial for policy design (e.g. to assess welfare changes, risk management support and diffusion of technologies). While there is a large range of individual studies eliciting farmer risk preferences and some systematic reviews (e.g. Iyer et al., 2020), there have been no attempts to combine the individual participant data from various studies. This, however, is needed to overcome limitations from individual studies such as small sample sizes and therefore low statistical power, uncertainty regarding the external validity of results based on specific contexts and limitations from conventional meta-analysis such as the reliance on results published in the manuscripts (see Ferraro and Shukla, 2023).

This paper aims to overcome these limitations and provide a new approach to study farmers’ risk preferences, their distribution and the factors behind their heterogeneity. Our empirical analysis is motivated by a set of questions regarding risk preferences and their heterogeneity: how important is the deviation from risk neutrality among European farmers? Are there systematic differences in the categorisation of farmers’ risk preferences depending on the elicitation method used? Do risk preferences vary with farmers’ characteristics such as gender, age, farm size and production system? And, to what extent these differences are study specific or generalisable?

We provide a new approach to the agricultural economics literature, relying on an individual participant data meta-analysis. Our analysis focuses on European farmers, who are subject to similar economic, institutional and policy environments (e.g. the Common Agricultural Policy of the European Union). In a first stage, we identified 50 studies conducted in 17 European countries where risk preferences were elicited with two prevailing approaches, namely Holt and Laury (2002) lotteries and self-assessments. After seeking the data, we recovered complete raw data about farmers’ risk preferences from 19 studies in 13 European countries. Our compiled database contains the risk preferences of 5,157 farmers together with information about farm and farmer characteristics. The meta-data allows us to understand how heterogeneous farmers’ risk preferences are, and to construct a risk profile of farmers based on the factors associated to their risk preferences.

There is a growing empirical literature that elicits and studies risk preferences of the general population globally (see Von Gaudecker, Van Soest and Wengstrom, 2011; Falk et al., 2018; Brown et al., 2023) and among European farmers (Spiegel et al., 2021; Rommel et al., 2023).¹ In this literature, the average farmer tends to be risk averse (Iyer et al., 2020). Nevertheless, risk preferences have been found to range from very risk averse to risk loving.² The implication of this heterogeneity is attached to the extent to which risk preferences drive economic decisions at the farm level and respond to objective elements such as farm and farmer characteristics. Four aspects have been considered regarding farmers’ risk preferences: (i) sociodemographic characteristics, (ii) farm characteristics, (iii) decision-making and (iv) risk environment (e.g. policy changes and weather shocks). In this paper, we focus on the first two aspects, investigating the heterogeneity of risk preferences in light of measurable factors associated to the context surrounding the farmer.

Several farmers’ demographic characteristics have been explored in previous research. For example, there is evidence that older farmers have lower risk tolerance and higher risk aversion (i.e. Gómez-Limón, Guerrero-Baena and Sánchez-Cañizares, 2020; Maart‐Noelck and Musshoff, 2014; Reynaud and Couture, 2012). Consistent with non-agricultural contexts, female farmers are found to be more risk averse than male farmers (e.g. Maart‐Noelck and Musshoff, 2014; Nielsen, Keil and Zeller, 2013). Other aspects studied include education, farm size, farm orientation, off-farm income and land ownership (e.g. Cerroni, 2020; Liu, 2013; Nielsen, Keil and Zeller, 2013). Despite the abovementioned evidence, we currently lack understanding of the overarching evidence base, e.g. with respect to magnitudes and drivers of the heterogeneity in risk preferences. Furthermore, concerns regarding the generalisation of findings remain, partly due to cross-sectional variation and limited comparability of risk preferences among different groups of farmers (e.g. arable, livestock, mixed farmers).³ For example, the risk profiles associated to risk preferences are often embedded in specific contexts and are focused on small samples, limiting the external validity and the statistical power of the analyses (see e.g. Ferraro and Shukla, 2023). Iyer et al. (2020) performed a systematic review on the measurement of risk preferences of European farmers. This study, however, only presented key findings from papers (e.g. average risk aversion across all farms), without exploring the underlying heterogeneities in the individual observations in each study. Furthermore, there is no systematic analysis of the connection of determinants/correlates of farmers’ risk preferences.

We aim to overcome these limitations and systematically and quantitatively analyse risk preferences and the attributes associated with them. We perform an individual participant data (IPD) meta-analysis where the relations between farmers’ characteristics and risk preferences are estimated across studies. There are several attractive features of our setting. First, the application of IPD meta-analysis allows us to not only identify the average risk aversion among farmers but also its distribution, i.e. how heterogenous farmers are in their preferences. Second, maximising the commonalities across studies increases the coverage of risk preferences across farmers’ characteristics beyond the results reported in published manuscripts. This allows us to identify the attributes of farmers with specific risk preferences and improve the assessments regarding the external validity of the different analyses. This is relevant because conventional meta-analyses are constrained by the objectives of the publications, the heterogeneity of methods used and the reported coefficients. In contrast, our statistical analysis utilises raw data from each study, enabling us to broaden our analysis to include papers whose objectives may not be directly linked to the correlation between risk preferences and farm/farmer characteristics. Third, we overcome the challenges posed by conventional meta-analysis, wherein researchers exercise discretion regarding methodological approaches and the inclusion of variables in econometric analysis. This discretion often results in limited comparability of correlation coefficients across surveys, leading to restricted or erroneous inference.

Our results suggest that European farmers are on average risk-averse when risk preferences are elicited using Holt and Laury (2002) lotteries, and on average risk-neutral when elicited using self-assessments based on Likert scales (following e.g. Meuwissen, Huirne and Hardaker, 2001). For the latter, there are significant differences in farmers’ risk preferences depending on the risk domain. For example, farmers are most risk averse in the financial domain and least risk averse in the production domain. We find, moreover, that the risk preferences are highly heterogeneous. Using lotteries, 30 per cent of farmers are found to be very risk averse and 21 per cent are risk loving.⁴ For self-assessments, we find that 12 per cent of farmers are very risk averse and 18 per cent are very risk loving. This heterogeneity implies that neither risk neutrality nor risk aversion represents the risk preferences of all farmers.

While exploring the determinants of the heterogeneity of farmer risk preferences, we find that only a few of the farmer and farm characteristics are significantly associated with risk preferences elicited through lotteries. On the contrary, when risk preferences are elicited with self-assessments, several patterns arise. For example, we find that male farmers and farmers with higher education are less risk averse than their counterparts. Farmers that have larger farms have lower levels of risk aversion while farmers that employ a larger workforce are more risk averse. Despite these associations, we find that the documented heterogeneity of farmers’ risk preferences cannot be fully accounted for with the rich number of socioeconomic and farm-specific factors considered. This is consistent with previous efforts to characterise risk preferences in broader populations and points to a limited understanding of the context surrounding risk preferences (see e.g. Von Gaudecker, Van Soest and Wengstrom, 2011).

We discuss potential reasons for the larger explanatory power of farm and farmer characteristics to determine risk preferences elicited with self-assessments compared to lotteries, including confounding factors in the identification of risk aversion such as learning abilities, numeracy skills and complexity of the elicitation tasks in the case of lotteries (e.g. Dave et al., 2010; Oprea, 2022) and risk perceptions in the case of self-assessments (Schildberg-Hörisch, 2018). Differences in the elicitation context can also explain the different patterns between elicitation methods (Finger, Wüpper and McCallum 2023a; Mata et al., 2018). For instance, Holt and Laury (2002) lotteries operate in the gain domain of income risk (i.e. to win money), whereas self-assessments are by construction broad and may rather capture the loss domain (Mata et al., 2018).

Our analysis provides useful benchmarks and comparable estimates to evaluate risk preferences and the magnitude and significance of attributes associated to them, relevant for policy, industry and researchers. We offer a discussion on the analysis of risk preferences based on the empirical estimates, and discuss why, despite the challenges behind their measurement, they should be considered in policy making. We provide the first application of IPD meta-analysis in agricultural economics and discuss the challenges for future implementation including selection bias due to non-availability of IPD for all the literature identified and differences in survey protocols. The method applied in this setting can be replicated to other contexts to address policy questions in a systematic and quantitative manner, while offering an avenue for open science and a robust safeguard against research practices such as p-hacking.

The remainder of the article is organised as follows. Section 2 discusses the literature on risk preferences. Section 3 describes the method comprising the data collection procedure, the econometric analysis and robustness checks. Section 4 describes the compiled data and section 5 presents the results. Finally, section 6 concludes.

2. Background

2.1. Risk preference elicitation

Preferences towards risky decisions are not directly observable and require the administration of elicitation methods. Both psychologists and economists have developed a wide set of elicitation methods. For example, methods based on the observed economic behaviour and revealed preference over lottery decisions, as well as methods based on self-assessments using Likert scales and reported risky behaviour. There are two main frameworks that conceptualise risk preferences, i.e. Expected Utility Theory and Cumulative Prospect Theory. Risk preferences elicited under the Expected Utility framework mostly follow the experimental design of Holt and Laury (2002).⁵

Self-assessments vary in the way they are administered. The most used method is the domain-specific willingness to take risks measured with Likert scales. An example of such assessments include the adaptation of Meuwissen, Huirne and Hardaker (2001), where farmers assess whether they agree to a statement presented to them in a 5-point Likert scale where 1 refers to ‘Do not agree’ and 5 ‘Fully agree’, as follows: ‘I am willing to take more risks than my colleagues with respect to: (i) Production, (ii) Marketing, (iii) Financial issues, iv) Farming in general’ (Meuwissen, Huirne and Hardaker, 2001).

If risk preferences elicited using lotteries and self-assessments are measures of the same trait (i.e. are rank-order stable), both measures should, to some extent, exhibit similar patterns in relation to key correlates. However, differences in farmers’ risk preferences and risk profiles under different elicitation methods are expected to arise mainly because the measurement of risk preferences according to the psychometric and economic approaches differs. Lotteries are administered in a risk-controlled environment in which the actual choices of individuals are observed, often relying on incentivisation of participants so that observed choices are incentive compatible (Holt and Laury, 2002). This provides a consistent setting for comparing risk aversion coefficients across individuals. Nevertheless, the complexity of the tasks and the presence of compounded lotteries under economic incentives entail potential confounding factors that hinder the identification of a risk aversion coefficient, including learning abilities and numeracy skills (e.g. Dave et al., 2010; Oprea, 2022). Self-assessments, on the other hand, are easy to understand, but tend to capture more influences such as risk perceptions (Schildberg-Hörisch, 2018) and have been argued to be a composite measure that captures both risk coping ability and willingness to bear that risk (see Young et al., 1979 for a discussion). Despite these differences, risk preferences are expected to differ across farm and farmer characteristics and empirical evidence consistently demonstrates that risk preferences exhibit systematic variations across contextual factors (e.g. life cycle risk aversion).

2.2. Potential correlates of risk preferences

The literature exploring the correlates of risk preferences together with the intuition behind risk aversion provides a good ground to identify the type of correlation expected from risk preferences and certain attributes.⁶ Age, gender and education are often found to be significantly associated with risk preferences. Older and female farmers are expected to be more risk averse than young and male farmers (see e.g. Falk et al., 2018; Von Gaudecker, Van Soest and Wengstrom, 2011). As for broader populations, education is expected to be negatively correlated with risk aversion (see e.g. Nielsen, Keil and Zeller, 2013; Von Gaudecker, Van Soest and Wengstrom, 2011).

In terms of farm characteristics, there is some evidence of a negative association between risk aversion and farm size (e.g. Cerroni, 2020) although the relationship is found to be insignificant in most studies (e.g. Meuwissen, Huirne and Hardaker, 2001; Sulewski and Kłoczko-Gajewska, 2014). In addition, the literature on whether individual risk preferences vary with off-farm income and land ownership has been limited and inconclusive. However, there are indications that risk aversion is positively associated with off-farm income and negatively with land ownership (see e.g. Zhao and Yue, 2020).

Consistently with the diversification intuition, it is expected that the percentage of time spent in agricultural activities and having off-farm income are positively and negatively correlated with risk aversion, respectively. The intuition behind the relationship between succession and risk preferences is not well documented. However, there is evidence suggesting that if the farm continues after the retirement of the producer, the farmer is more risk averse (Picazo‐Tadeo and Wall, 2011). The productive orientation of the farmer (e.g. arable farmers, livestock farmers) and the type of production (i.e. organic or conventional) can be important correlates of risk preferences. While evidence of a negative association between risk aversion and adoption of organic agriculture is well documented (e.g. Flaten et al., 2005), evidence on production orientation is rather scarce mainly due to samples focusing on a single farm type (e.g. Roe, 2015; Spicka, 2020).

All in all, we can expect risk preferences to exhibit covariation with farm and farmers’ characteristics, some of which are observable. The extent of this variation can depend on the context surrounding the elicitation of risk preferences. In the case of self-assessments, risk preferences are often elicited for different risk domains that can be correlated with farm and farmers’ characteristics differently depending on the domain. For example, the non-separability between labour and consumption for some farmers can imply a higher exposure to financial risks and therefore lower willingness to take risks in this domain (e.g. Dercon and Christiaensen, 2011). Similarly, the level of control over production or market risks (e.g. with financial buffers, savings, farm size) can determine the willingness to take risks in the production and market domains (e.g. Mariano, Villano and Fleming, 2012).⁷

We follow an explorative approach by looking at the patterns that arise from the analysis. In what follows, we present the methods used to identify the correlates of risk preferences elicited through lotteries and self-assessments.

3. Method

3.1. Data collection

Our search for primary studies eliciting European farmers’ risk preferences involved three steps.⁸ First, we searched for academic papers on the scientific citation databases Scopus, Web of Science and Wiley Online Library based on the pre-defined eligibility criteria. Studies were included if they: (i) were peer-reviewed and written in English, (ii) elicited farmers’ risk preferences through self-assessments or Holt and Laury (2002) lotteries, (iii) included risk preferences as a variable of interest, (iv) were published in the period 2000–2021 and (v) focused on a European country.⁹ The PRISMA flow diagram of this process is detailed in the Figure A1 in Appendix and a summary is shown in Figure 1. The initial search returned 986 articles. After removing duplicates, we screened the titles and abstracts of 612 papers which resulted in the identification of 70 articles for full paper screening. After full screening, we narrowed the selection down to 50 relevant studies and extracted information related to the survey design, estimator of risk preferences, the correlates of risk preferences and risk perceptions. We then contacted the authors to get the original datasets used in the analysis.

Out of the 50 papers, 8 papers had their datasets publicly available either with the published article or on other platforms (i.e. GitHub, Figshare, DANS-EASY). We contacted the corresponding authors of the remaining 42 papers via e-mail and received the original datasets and documentation for 11 articles—4 articles under a data agreement between the parties. For the remaining 31 papers we either received no reply from authors (n = 23), the author replied but the dataset could not be made available (n = 4)¹⁰ or the e-mail provided by authors was not functional (n = 4).¹¹ In total, 19 datasets were available to us (rate of positive responses: 38 per cent).¹² With the datasets at the individual level, we harmonised the risk preferences and correlates available in the surveys. The harmonisation process implied maximising the amount of information extracted from the different surveys without compromising on the comparability of variables. For example, risk preferences elicited using self-assessments from 10-point and 7-point Likert scales were re-scaled to 5-point Likert scales.¹³ For risk preferences elicited using lotteries, we follow Holt and Laury (2002) and define the (constant) coefficient of relative risk aversion (CRRA) based on the number of safe choices. The full harmonisation protocol is available with the pre-registration of the meta-analysis.

3.2. Econometric analysis

We analyse the risk preferences of European farmers and investigate their distributional aspects within and across studies. We conduct our analysis in three stages. First, we show and test how the risk preferences differ across farmers, accounting for different elicitation methods. Second, we test and quantify if and how risk preferences vary according to farmer and farm characteristics. Third, we test how robust the results are to several alternative specifications. In a first analysis, we descriptively analyse the meta-data and compute the average risk aversion under self-assessments and lotteries. We use the sample mean and coefficient of variation to evaluate the heterogeneity of risk preferences, and test for linear and monotonic relationships between the measures using Pearson and Spearman correlation tests.

To investigate what explains the heterogeneity of risk preferences, we estimate a one-stage IPD meta-analysis. This approach consists of analysing the relations of interest at the farmer level, considering the variation that comes from the studies. We relate the risk preference of farmers to characteristics that are present in the studies (e.g. age, gender, education, off-farm income) through a linear model as specified in Equation 2. The model is estimated for each of the elicitation methods independently (i.e. self-assessments and lottery-choice tasks).¹⁴

where |$R{P_{ik}}$| is a measure of risk preferences of producer |$i$| from study |$k$|⁠. When risk preferences are elicited using self-assessments, |$RP$| is the harmonised scale (i.e. 5-point scale) and when they are elicited using lotteries, it refers to the average CRRA. The vector |${X_{ik}}$| includes farmers’ characteristics (e.g. type of production–livestock, arable farming, mixed farming, age, gender, farm size, off-farm income). Our sample comprises surveys from single countries and multi-country surveys (i.e. each country following a similar survey design). We expect farmers’ risk preferences to depend on both the study and the country where the elicitation was performed. Therefore, the vector |${\gamma _k}$| includes study-country fixed effects that account for the unobserved contextual characteristics of the studies and countries. Additionally, a few characteristics can exhibit cross-study variation but not within-study variation. To identify these parameters, the vector |${Z_k}$| comprises characteristics that are present in more than one study and therefore are not captured by |${\gamma _k}$| (e.g. framing of lotteries, economic incentives). Finally, |${\varepsilon _{ik}}$| corresponds to the idiosyncratic error term clustered at the study and country level to control for potential correlation across farmers within a study-country combination. We report the p-values after the wild bootstrap clustering to account for few clusters as suggested by Cameron, Gelbach and Miller (2008). The cluster unit is given by farmers within the same study and country.¹⁵

The coefficients of interest are given by |${\beta _1}$|⁠. These estimators can be fully understood when the set of correlates are equal across contexts. However, given that in meta-analyses this is rarely the case, we follow the most prevalent approach of opting for a one-to-one correlation (e.g. Brown et al., 2023), with the difference that we compute the correlation with regression analysis.¹⁶ The coefficients capture the correlation between farm characteristics and risk preferences without the role of mediating or moderating variables. For example, age and experience are correlated and when considered together in a model, their coefficients might show a wide range potentially capturing that more experienced farmers might be older. In a one-to-one analysis, their correlation is evaluated in isolation from other potentially relevant variables. All through the analysis, we carefully interpret our results as correlations and not as causal relations and provide extensive checks to test the robustness of our results.

3.3. Robustness checks

To test for the robustness of our results, we perform eight additional analyses, addressing aspects of selection bias, and robustness of findings for lotteries and self-assessments. First, we evaluate the differences in risk preferences retrieved from IPD and the published papers with funnel plots to visually inspect potential biases. Second, given our interest in risk preferences elicited using lotteries, issues of inconsistencies might be prevalent (Meraner, Musshoff and Finger, 2018). In Holt and Laury (2002) lotteries, multiple switching behaviour and subjects who select with certainty a lottery with lower payoff have been argued to reflect either indifference between two lotteries or a lack of comprehension of the task (Holt and Laury, 2002; Ihli, Chiputwa and Musshoff, 2016). Therefore, to rule out spurious correlations, we exclude farmers with inconsistent behaviour, i.e. farmers who have multiple switching behaviour or select with certainty a smaller payoff (e.g. Rommel et al., 2017). Third, we consider an alternative measure of risk preferences where instead of the average CRRA computed based on the number of safe choices, we rely on the first switch to the risky lottery. Fourth, if inconsistent behaviour is random across farmers, concerns regarding the internal consistency of the measure of risk preferences can be reduced. To address this aspect, we analyse whether the inconsistent behaviour in lotteries is correlated with farmer and farm characteristics. We re-estimate Equation 1 specifying as the dependent variable, a dummy variable taking the value of 1 if the farmer has at least one inconsistent switch.

Equation 1 presents a model that establishes a one-to-one correlation to address the tradeoff between sample size and the number of variables in the model. An alternative approach involves examining how the correlates of risk preferences vary systematically as other correlates are introduced into the model. In our fifth exercise, we employ the Extreme Bounds methodology (Leamer, 1985) and Sala-i-Martin’s procedure (Sala-i-Martin, 1997), wherein multiple models, encompassing all possible combinations of predictors, are estimated to capture the distribution of coefficients representing the correlation between farm and farmer characteristics and risk preferences. The Extreme Bounds method provides a conservative criterion for robustness; a correlate is deemed robust if the upper and lower extreme bounds of the empirical distribution of coefficients share the same sign (i.e. zero is not included in the estimates). In line with Sala-i-Martin (1997), we calculate the weighted average coefficients and standard deviations. We also report the percentage of coefficients above or below zero relative to the mean coefficient and the percentage of coefficients significant at a 5 per cent level (e.g. Carmignani et al., 2014).¹⁷

Six, to address issues related to methodological heterogeneity in IPD meta-analysis (Christensen, 2003) and to shed light on the external validity of individual studies, we estimate a two-stage IPD meta-analysis. Compared to the one-stage meta-analysis where all studies are analysed simultaneously, in the two-stage meta-analysis, each study is analysed separately, and the resulting coefficients combined across studies. Although none of the approaches outperforms the other, a two-stage meta-analysis can reveal if the correlation coefficients are disproportionately influenced by single studies. We test whether the patterns arising in the regressions hold when the alternative methodological approaches are applied (e.g. when the coefficients at the study level are weighted and pooled).

Seven, the literature on risk preferences globally suggests that risk preferences are country-specific due to several factors including institutional, geographic and cultural differences (Falk et al., 2018; Rieger, Wang and Hens, 2015; Vieider et al., 2015b). We, here, analyse the meta-data to identify systematic differences between countries and within countries with t-tests on the equality of means. Eight, to acknowledge that the dependent variable is bounded for self-assessments (e.g. is measured at a scale of 1–5), we test whether our results are robust to the econometric specification by estimating an Ordered Probit model.

4. Data

Our dataset comprises risk preferences of farmers from 19 studies in 13 European countries (see Figure 2). Note that multiple countries can be covered in one study. The elicitation of preferences through lotteries is available for three countries, while the use of self-assessments is available for all 13 countries. The country with the largest number of datasets in our sample is Germany, with nine datasets. Countries in Central and East Europe are underrepresented in the literature and therefore in our analysis.

Our database contains the risk preferences of 5,157 farmers, with farm and farmer characteristics. Ten studies elicited preferences through lotteries following the approach of Holt and Laury (2002) (n = 10) and 14 studies elicited preferences using self-assessments (see Table 1).

Preferences elicited using lotteries have two main variations across studies based on whether they are: (i) economically incentivised (i.e. monetary payments), and (ii) framed under an agricultural context. Table 2 shows the descriptive statistics of the risk preferences in the sample. For 84 per cent of farmers, lotteries used economic incentives, and 49 per cent had an agricultural framing. In the standard Holt and Laury (2002) approach, a common proxy of risk preferences is the number of safe choices. A risk neutral farmer would select four choices and a risk-averse farmer would select more than five safe choices. Following this standard setting, we find that on average, farmers select 5.34 safe choices before switching to the risky lottery, which means that they are moderately risk averse. From the number of safe choices, it is possible to derive the relative Arrow-Pratt coefficient of risk aversion range for each farmer. The average CRRA aversion is 0.35 (N = 1,475), indicating that the average farmer is slightly risk averse.

Self-assessments consist of a series of statements or questions for which the farmer reports on a scale, either the level of agreement (e.g. from do not agree to fully agree) or the level of their willingness to take risks. The reasoning behind this measure is that individuals identify the central point in the scale and rate themselves above or below this point. Moreover, differences between individuals can be identified at other points on the scale, for example, individuals that choose the extreme values of the scale. Studies in our sample mostly follow Dohmen et al. (2011) for a general measure of risk willingness and Meuwissen, Huirne and Hardaker (2001) for a domain specific measure (see Table 1). The original measure has a scale 0–10 with a middle point of 5. We transform this scale to range between 1 and 5 to increase comparability between the self-assessments and find that the average is 2.77 (N = 2,982), meaning that on average farmers in our sample are risk-neutral tending towards risk-loving. Domain-specific preferences have important heterogeneity around risk neutrality (defined at the cutoff of 3, as the scale ranges from 1 to 5). While, on average, farmers are risk-neutral tending towards risk-loving for agriculture in general and in the production domain, they are risk-neutral tending towards risk aversion on average in the marketing and finance domain. In the sample, 83 per cent of the farmers have domain-specific preferences phrased in relation to other farmers.¹⁹

Table 3 shows that on average farmers in the sample are 48 years old, 92 per cent are male and 33 per cent have higher education. The succession of the farm is arranged in 43 per cent of the cases, 37 per cent of farmers rent more than 50 per cent of their land, 38 per cent of farmers have off-farm work or income, while 57 per cent rely mostly on agriculture as their main income. On average, farmers have 0.35 units of workforce in the farm and 26 years of experience in farming.²⁰ Most farmers have less than 75 hectares of land, 54 per cent are devoted to arable and horticulture and 32 per cent to livestock, while the remaining 14 per cent are mixed (i.e. livestock and arable). Additionally, 13 per cent of farmers produce under organic requirements. Note the varying number of observations across the farm and farmer characteristics. This variability limits the possibilities to analyse all factors at once and leads our analysis towards identifying the significant correlates in single specifications as shown in the next section.

5. Results

We present our results in three stages. First, we show how the risk preferences of farmers differ across elicitation methods between farmers. Second, we show how risk preferences vary according to farmer and farm characteristics. Third, we show the robustness of the results to several specifications.

5.1. Heterogeneity of farmer’s risk preferences

First, we focus on risk preferences elicited through self-assessments. We aim to identify the heterogeneity of risk preferences between and within studies. For this purpose, Figure 3 shows the association between the mean of the self-assessments in four domains (x-axis) and the coefficient of variation (y-axis) for each of the studies and countries in the sample.²¹ The vertical grey line shows the point of risk neutrality and values higher than 3 refer to stated risk-averse preferences. Three aspects are relevant. First, there is a negative relationship between the coefficient of variation and the mean risk aversion that is consistent for all domains of risk preferences. This means that when the average farmer is found to be slightly risk loving, the heterogeneity around the mean is large (i.e. large within-sample variation). The personal risk aversion, for example, shows less heterogeneity than risk aversion in the agricultural domain, evidencing in addition the domain-specificity of risk preferences. Second, on average, farmers’ preferences are located around risk neutrality, between the values 2.5 and 3.5.²² Third, compared to the risk aversion in agriculture in general, the risk aversion associated to the finance, market and production domain is larger. This is supported by the mean of these preferences in all the sample, where farmers are most risk averse in the financial domain, followed by market, then production and lastly agriculture in general (see Table 2). In sum, there is a large heterogeneity in farmers’ risk preferences, ranging from slightly risk loving to risk averse. In the case of the agricultural domain, 18 per cent of farmers report having high willingness to take risks while 12 per cent report a very risk-averse preference (see Figure 4).

Figure 5 shows the distribution of the CRRA for all studies that elicit preferences using lotteries. At the right, the mean of CRRA and sample size are reported. The vertical grey area signals the interval of risk neutrality. As per Holt and Laury’s (2002) classification, coefficients of risk aversion σ<−0.15 indicate risk loving preferences, and σ >0.15 risk aversion. The individual distributions suggest that risk aversion is a prevalent preference among farmers, although there is large heterogeneity between and within the studies. The coefficient of risk aversion ranges from −0.17 (risk loving) to 0.59 (risk averse), and on average, farmers are slightly risk averse with a CRRA of 0.35. Risk aversion is prevalent for 62 per cent of farmers, although similar to self-assessments, there is important heterogeneity with 21 per cent of farmers having risk loving preferences and 30 per cent being very risk averse (see Figure 6).

Adjustments made to the Holt and Laury (2002) lottery do not seem to lead to significant differences in the CRRA when compared to other studies. For example, Cerroni (2020) adjusts the lottery by setting a series of choices between a certain payoff and a lottery, and Hermann, Sauthoff and Mußhoff (2017) increase the number of lotteries to 20, increasing the accuracy of the estimates of risk preferences. In both cases, farmers are found to be risk averse on average. Consistent with previous literature, farmers are less risk averse in self-assessments compared to lotteries (see e.g. Nielsen, Keil and Zeller, 2013) and there is a limited convergence of the two methods. This implies that beyond differences in the average risk aversion, there are distributional differences in risk preferences elicited with the two methods. This variation, moreover, is expected given that each measure is a compound measure of other constructs, e.g. risk perceptions and risk coping ability in the case of self-assessments and learning abilities and numeracy skills in the case of lotteries.

With a subsample of farmers for which self-assessments and lotteries are available (N= 886 farmers), we estimate parametric and non-parametric correlation tests and find that the CRRA is positively correlated with self-assessments in the personal domain and in the domain of agriculture (see Table A5 in Appendix). The correlations range from 0.15 to 0.52 (i.e. from a very weak relationship to a positive moderate), showing limited convergence of risk preferences from self-assessments and lotteries. This finding is consistent with previous literature for broader populations (e.g. Mata et al., 2018; Pedroni et al., 2017). Overall, we find evidence of large heterogeneity in risk preferences elicited using lotteries and risk assessments between and within the studies. Nevertheless, this evidence is solely descriptive. To compare risk preferences across studies, it is necessary to identify the multiple sources of variation, i.e. the risk profiles of farmers. Next, we explore whether farmer and farm characteristics are systematically and significantly associated with risk preferences.

5.2. Risk profiles: farmer and farm’s characteristics

We estimate linear models and identify the key attributes significantly associated with farmers’ risk preferences. Figure 7 shows a summary of the results of this exercise for three measures of risk preferences: the self-assessment for personal risk aversion, self-assessment for agriculture in general and the average CRRA. The y-axis shows the regression coefficients for the correlates considered in the analysis in relation to the dependent variable, and the x-axis shows the associated p-value. Coefficients falling in the grey area imply a significant association (p < 0.05) of a correlate (i.e. farm and farmer characteristics) with risk preferences. Risk preferences measured with self-assessments capture more patterns than preferences elicited using lotteries. For the latter, no coefficient is significantly associated with risk preferences at 5 per cent significance levels, while more patterns arise when self-assessments are considered, especially for the domain of agriculture in general. This suggests that risk preferences elicited using lotteries have a weak relationship with farmer and farm characteristics.

In what follows, we describe the results more in detail. Table 4 presents the results of the estimation of Equation 1 for the measures of risk preferences using self-assessments in two variations: risk aversion in general following Dohmen et al. (2011) and risk aversion in the domain of agriculture in general.²³ The econometric model in Equation 1 is estimated for each characteristic independently due to the variation in the number of observations available. In squared parenthesis we report the p-value after wild cluster bootstrapping to account for small number of clusters. The table shows the number of observations and in round parenthesis the number of clusters comprised in the regression.²⁴

For the agricultural domain of risk preferences, we find that young, male farmers and farmers with higher education are less risk averse than their counterparts. Higher levels of risk aversion are positively correlated with the number of workers in the farm. Farmers that have succession of the farm arranged are less risk averse (i.e. more risk loving) than farmers that have no succession defined. Organic farmers are less risk averse than conventional farmers. Agricultural area is significantly related with risk preferences. Larger farms are associated with less risk aversion. This relationship, moreover, is consistent across farm sizes in terms of magnitude, especially for the domain of agriculture in general. While farmers with 25−50 hectares of land are 0.24 units less risk averse than farmers of the reference group (i.e. having between zero and 25 hectares), farmers with more than 250 hectares of land are 0.99 units (i.e. 1 point on the self-assessed scale) less risk averse than the reference group (see column 3). Off-farm income, type of farm (i.e. horticulture, arable and livestock) and land ownership are not significantly associated with risk preferences.

For personal risk aversion, the direction and significance of characteristics such as gender, education, workforce and agricultural area are the same as the agricultural domain. However, the coefficients of succession and organic agriculture are not significant for the personal risk aversion. This shows that despite the positive correlation between the personal risk willingness scale and the scale for the agriculture domain, the contextualisation of risk preferences in self-assessments matters. Furthermore, the relevance of the domain of risk preferences is supported by the consistency in the direction and significance of associations between farm and farmer characteristics and the risk preferences in the production, market and financial domains (see Table A6). This implies that while farm-specific characteristics provide useful tools to understand risk preferences in the agricultural domain, other characteristics outside this domain could better explain the self-reported personal risk aversion.

The estimation using the CRRA is presented in Table A7 in the Appendix. The regression controls for the lottery setup (i.e. whether the lottery was incentivised or had an agricultural framing) and dummies for study and country. The results from this exercise display a lack of significant associations between risk preferences elicited using lotteries and farm and farmers characteristics. One exception is the coefficient for workforce that is positive and significant at 10 per cent. This same overall pattern is found for other farm populations and elicitation procedures using lotteries (e.g. Rommel et al., 2023; Zhao and Yue, 2020), ruling out the possibility that this is due to small sample sizes and raising questions regarding the sources of heterogeneity of risk preferences in lotteries.

All in all, we find that male farmers, farmers with higher education and farmers with larger farms up to 25 hectares and organic farmers are less risk averse than their counterparts. Farmers that have a successor arranged in their farms are less risk averse than farmers without succession, while workforce is positively associated with risk aversion.²⁵ Despite the intuition behind land ownership, off-farm work and farm income reliance, they are not significantly correlated with risk preferences (for the lotteries or the self-assessments).

5.3. Robustness checks

First, we assess the selection bias of risk preferences between IPD meta-analysis and the published information in papers (see Figure A2 in Appendix). We observe that all measures of risk preference exhibit symmetry around the threshold identified in the main analysis, except for the CRRA coefficient. The CRRA coefficient shows asymmetry, with three studies displaying notably large coefficients above 2. However, when we consider the number of safe choices, the most commonly reported measure of risk preferences, this asymmetry diminishes. This observation suggests that the risk preferences obtained through IPD meta-analysis do not deviate significantly from the existing literature.

Second, we explore potential biases arising from the inclusion of observations with inconsistent behaviour. In the sample, 11 per cent of farmers presented multiple switching between the safe and the risky lottery. In addition, 8 per cent of farmers chose the lottery that provided with certainty a lower expected payoff. Taken together, this behaviour represents 16 per cent of the sample. We estimate Equation 1 excluding farmers with inconsistent behaviour (see Table A8 in Appendix). Results remain robust with the difference that this model identifies significant differences between livestock farmers and arable and horticulture farmers. Third, we compute the (constant) relative correlation of risk aversion based on the first switch to the risky lottery and find that results are robust with two exceptions namely the coefficient for workforce loses significance and the coefficient of agricultural area between 100 and 150 hectares becomes significant at 10 per cent. This suggests that besides the lack of significant associations between risk preferences and farmer’s characteristics, the identified patterns are not stable to different specifications. Therefore, the primary model, which does not identify any substantial correlation between risk preferences and farmer characteristics, is favoured.²⁶

Fifth, in line with Leamer (1985) and Sala-i-Martin (1997), we conduct 6,141 regressions considering all possible combinations of regressors for each of our three primary dependent variables (Tables A10 and A11 in the Appendix). The method’s systematic approach mitigates selective reporting, providing an overview of likely values for the correlates of risk preferences. Our findings indicate that the magnitude of coefficients aligns closely with those in the main analysis, and the identified patterns prove robust across all model specifications considering observable correlates. Notably, variables such as age, gender, education, workforce, succession, organic agriculture and farm size persist as significant predictors of risk preferences in the domain of agriculture. This is evident in the substantial proportion of coefficients maintaining sign stability, a high share remain significant at the 5 per cent level, and robust upper and lower bounds. As in the main analysis, the personal domain exhibits weaker correlations which are less consistent than the domain of agriculture, and lotteries show no significant correlation between observed characteristics and the CRRA coefficient. For the CRRA coefficient, for example, the extreme bounds always have different signs and the share of models where the p-value lower than 5 per cent is never higher than 41 per cent.

Six, to address aspects related to the methodological approach, we perform a two-stage meta-analysis and report the results for two dependent variables: self-assessments under the agricultural domain (Figure A3) and the relative coefficients of risk aversion (Figure A4). Due to the high heterogeneity between studies, we estimate a random-effects inverse-variance weighted pooled effect. We find a high consistency between the one-stage and two-stage meta-analysis results regarding the direction, magnitude and significance of the relationship between the different correlates and risk preferences with a few exceptions. These include the coefficients of workforce and organic agriculture for self-assessments that retain the direction but lose significance. Additionally, despite the observed heterogeneity in the magnitude of the relationship between farm/farmer characteristics and risk preferences, the sign and significance of the coefficients correspond to the overall averages suggesting that individual studies have some level of external validity compared to other studies. Overall, the patterns identified in the main analysis are consistent when varying the specification of the meta-regressions. This, moreover, supports the methodological observation that no approach outperforms the other (see e.g. Burke, Ensor and Riley, 2017).

Seven, Table A12 in Appendix suggests that there are significant differences in risk preferences across countries. Particularly, German farmers, which are the largest farmer population in our sample (26 per cent), are significantly more risk averse than the average European farmer when risk preferences are elicited with self-assessments but are no more or less risk averse than the average farmer when the CRRA is considered. Moreover, there are significant within-country differences as evidenced by the heterogeneity of risk preferences under self-assessments and lotteries for German farmers. Both insights point to limited external validity of each study eliciting risk preferences, within and across countries.

Eight, the results are robust to the specification of an Ordered Probit for the self-assessments (Table A13). All in all, the additional analyses corroborate the main results by considering inconsistent behaviour in lotteries, the methodological choice of the meta-regression (i.e. two-stage meta-analysis) and the nature of the dependent variables in the econometric models.

6. Conclusions

In this paper we present a new approach to establish an empirical overview of farmers’ risk preferences and their distribution, using an IPD meta-analysis. Our results suggest that risk preferences of farmers are highly heterogenous within and across elicitation methods. This implies that beyond differences in the average risk aversion, there are distributional differences in risk preferences. While on average the risk preferences of farmers range between slightly risk-averse and risk averse with Holt and Laury (2002) lotteries, they fall around the point of risk neutrality when elicited by self-assessments. We observe a significant share of farmers that are very risk averse, between 12 per cent (for self-assessments) and 30 per cent (for lotteries) of farmers, and risk loving, between 18 per cent (for self-assessments) and 21 per cent (for lotteries) of farmers. Moreover, we highlight the significant heterogeneity of risk preferences within and across studies, risk domains and farm and farmer characteristics. We find, for example, that the heterogeneity of risk preferences elicited using self-assessments can be explained to some extent by a broad set of characteristics such as age, gender, education and farm size. In contrast, we find no robust patterns arising from risk preferences elicited with lotteries.

Our analysis has implications for policy makers and researchers. Agricultural policies, such as the Common Agricultural Policy of the European Union, aim to support farmers risk management capacities, e.g. by subsidising insurance schemes (e.g. Bucheli et al., 2023). Moreover, various policy-relevant decisions taken by farmers, such as the participation in voluntary agri-environmental schemes, input use and the investment in new technologies, are related to farmers’ risk preferences (e.g. Schaub et al., 2023; Wuepper et al., 2023). From a perspective of policy design, learning that a significant share of farmers is not highly risk averse thus informs the evaluation of welfare interventions and economic models and allows resources to be allocated appropriately (Peterson and Boisvert, 2004). Incorrect assumptions regarding risk preferences can have unintended consequences such as low participation rates and premiums in the insurance market that are too low or too high (Peterson and Boisvert, 2004). Moreover, the observation that risk preferences are related to the farmers’ context and characteristics and can potentially interact with them for decision making could elucidate open questions regarding low uptakes of agri-environmental schemes, crop insurances and new technologies.

For economic models, our findings imply applications could consider risk aversion according to farmer profiles rather than assuming risk aversion or risk neutrality. Ex-ante assessments provide a useful tool for policy evaluation with recent applications specifically considering behavioural factors such as risk preferences to improve the design of targeted and tailored agricultural policy measures (e.g. Huber et al., 2024; Dessart et al., 2019). Our analysis provides this type of assessments with a reasonable range of risk aversion among European farmers. Moreover, our results show that the elicitation approaches matter, i.e. overarching conclusions whether European farmers are risk averse or not depends on the elicitation approach. Moreover, what explains preferences (e.g. farmer characteristics) depends on the elicitation approach. Thus, policy decision shall be based on a wide evidence base on farmers’ risk preferences.

There are several avenues for future research. First, future studies shall analyse farmers’ risk preference elicitation with a wider geographic coverage of studies systematically, e.g. covering European countries and farming systems, using coherent replication studies (e.g. Rommel et al., 2023) and including less researched countries (e.g. Eastern European countries) as well as at global levels. Second, it is relevant to systematically assess the predictive power of farmers risk preferences elicited with different approaches to explain real farm decisions, e.g. regarding production, investment and risk management choices. Third, future work could further investigate the correlation between Likert scales and lotteries which capture probability weighting, loss aversion and ambiguity preferences and explore the patterns behind contextualised risk preferences. This could include the consideration of farm structural factors such as the non-separability of labour and consumption and behavioural factors such as farming goals (e.g. preference for agricultural activity), risk perceptions and perceived level of control over risks. Finally, concerns of omitted variable bias remain to be a challenge to explain the heterogeneity of risk preferences. We here consider a large set of farm and farmers’ characteristics and implement Extreme bound analysis, a procedure to test how the correlates of risk preferences vary systematically as other correlates are introduced into the model. This implementation helps mitigate these concerns, but future research can focus on aspects studied to a limited extent in the agricultural context such as shocks (e.g. illness, weather events, pests), personality traits and cognitive abilities.

Despite the inherent constraints of meta-analyses and systematic reviews on observed research and specific geographical units, the analysis used here overcomes issues of existing studies, e.g. related to small sample sizes and low statistical power and uncertainty regarding the external validity. In this context, IPD meta-analyses remain underused in the economics and agricultural economics literature. While our analysis provides a first application in this literature by analysing risk preferences, the method can be applied to a variety of contexts where gaps regarding the external validity of empirical analyses remain. This type of analysis, however, is only possible when the documentation for replicability of results is available to researchers. We obtained a positive response rate of 38 per cent to documentation availability, meaning that replication is possible for less than half of the evidence in the empirical literature about farmers’ risk preferences. To increase the transparency and credibility of empirical research and untap the potential of IPD meta-analysis, open research practices need to be strengthened, i.e. researchers should make data, codebooks and codes available or address confidentiality issues with alternative replication methods (e.g. synthetic data, Wimmer and Finger, 2023).

Acknowledgements

The authors would like to thank the editor, Salvatore Di Falco and the anonymous journal reviewers for their comments and suggestions. Special thanks go to Charles Rees, Andrea Fürholz and Louissa Wyss for their support in harmonising the diverse databases, essential for the completion of this study.

Supplementary data

Supplementary data are available at ERAE online.

Footnotes

References

We identified a total of 50 studies, out of which we were able to retrieve the IPD for 19 studies. In terms of sample size, there is no statistical difference between the 19 surveys included in our analysis and the 31 surveys that were excluded due to unavailability of IPD. The literature encompasses various sampling procedures, including the use of sampling frames based on groups such as farmers involved in extension services, machinery cooperatives, agricultural forums, alumni networks, farmers’ associations and cooperatives, agricultural publishers and farmers listed in administrative data sources like the Farm Accountancy Data Network and registries of organic farming. Among the identified surveys, only 11 of them discuss response rates, which range from 10 per cent to 61 per cent.

Search terms example

Scopus

TITLE-ABS-KEY (((‘risk preference*’) OR (‘risk attitude*’) OR (‘risk aversion’) OR (‘risk tolerance’)) AND ‘elicit*’ AND ((‘farmer*’) OR (‘producer*’) OR (‘grower*’)))

TITLE-ABS-KEY (((‘risk preference*’) OR (‘risk attitude*’) OR (‘risk aversion’) OR (‘risk tolerance’)) AND ‘measure*’ AND ((‘farmer*’) OR (‘producer*’) OR (‘grower*’)))

TITLE-ABS-KEY (((‘risk preference*’) OR (‘risk attitude*’) OR (‘risk aversion’) OR (‘risk tolerance’)) AND ‘case stud*’ AND ((‘farmer*’) OR (‘producer*’) OR (‘grower*’)))

Robustness 1

The selection bias between IPD and published information could pose a significant concern in IPD meta-analysis. The funnel plots below display risk preference estimates for four measures. In Panel A, the CRRA estimates and the number of safe choices under HL (Holt and Laury, 2002) differentiate between IPD meta-analysis estimates (red triangles) and estimates from the systematic review where no IPD data were available (blue circles). The CRRA coefficient plot exhibits an asymmetry, suggesting a potential downward bias in the average CRRA estimated through IPD meta-analysis. However, when examining the number of safe choices under HL, both samples (IPD meta-analysis and systematic review) show similar estimates symmetrically distributed around the threshold. Panel B demonstrates the self-assessments based on 10-point and 5-point Likert scales as reported by authors and estimated in the IPD meta-analysis, displaying a similar distribution across the threshold. In conclusion, while caution is warranted when interpreting the CRRA coefficient due to potential downward bias, the interpretation of other risk preference measures aligns with existing literature.

Robustness 2 and 3

Robustness 4

To identify any systematic bias coming from inconsistent behaviour we estimate a model that relates this behaviour with the correlates considered in the analysis (see Table A5). Inconsistent behaviour is measured with a dummy variable signalling whether the farmer had any inconsistent behaviour in the lottery, including farmers that have more than one switch in a lottery task or choose with certainty a lower payoff. We find that there are no significant associations between inconsistencies and the farm and farmer characteristics considered in the analysis that could hint at a systematic bias coming from inconsistent behaviour.

Robustness 5

Robustness 6

Two-stage meta-analysis

In the following we show the forest plots of each of the correlates for three dependent variables: self-assessments under agriculture in general, for lotteries, CRRA coefficients. The squared area shows the contribution of the study to the combined estimate and the horizontal lines the 95 per cent confidence interval. The combined estimate and its confidence interval are shown by a diamond. We use the STATA command ipdmetan to produce the forest plots. The I-squared represents the level of heterogeneity in the studies considered and the p-value is associated to the Cochran’s Q test (i.e. significance of heterogeneity between surveys). Given that the studies considered are very heterogeneous—following the criterion of I-squared <50 per cent—we estimate a random-effects inverse-variance weighted pooled effect. This weight incorporates both the between and within study variance. We follow the most used estimator for between study variance DerSimonian-Laird (DL) (see Fisher (2015) for more details on the estimation procedure).

Self-assessments: Agriculture in general

Lotteries: CRRA coefficient

Robustness 7

Robustness 8