The role of peer effects on farmers’ decision to adopt unmanned aerial vehicles: evidence from Missouri

ABSTRACT

This paper examines the role of peer effects on farmers’ decision to adopt Unmanned Aerial Vehicles (UAVs). This relationship is analysed by means of a Bayesian spatial autoregressive probit model applied to spatially explicit data describing the awareness, attitudes, and adoption of UAVs by 809 Missouri farmers. Results show that it is not only the farmers’ characteristics, awareness, and attitudes towards UAVs that affect adoption, but also the adoption behaviour and traits of neighbouring peers. Peer effects arise both from UAV adoption of nearby farmers and from spatial spillovers of farmers’ characteristics, awareness of UAV agricultural applications, expectations of economic and environmental benefits from UAV use and perceived neighbour privacy concerns in the use of UAVs.

KEYWORDS:

• Spatial dependence
• technology adoption
• unmanned aerial vehicles
• bayesian modelling
• spatial probit model

JEL CLASSIFICATION:

• Q12
• Q16
• C11
• C21

1. Introduction

Unmanned Aerial Vehicles (UAVs) are likely to play an important role in the future of precision farming. Being able to survey large areas of land and collect high-resolution remotely sensed imagery under diverse weather conditions and without the inherent safety risks and high costs involved with the use of manned aerial vehicles (e.g. fixed-wing airplanes, helicopters), UAVs are a new potential weapon in the farmer’s arsenal to improve decision-making and resource allocation (Pritt 2014; Floreano and Wood 2015; Schirrmann et al. 2016; Walter et al. 2017). Recently, a few studies have investigated farmers’ UAV adoption decisions, and the factors influencing this decision (Zheng, Wang, and Wachenheim 2019; Thompson et al. 2019; Michels, von Hobe, and Musshoff 2020; Skevas and Kalaitzandonakes 2020). Despite providing valuable insights into the factors affecting UAV adoption (such as ease of use, farmer expectations for economic and environmental benefits of UAVs, household income, farm size, precision agricultural technology literacy, etc.), these studies ignore peer effects, despite their potential influence on UAV adoption decisions.

Peer effects can be defined as the influence of the behaviour and characteristics of an individual’s peers on that individual. Possible channels through which such effects may arise are strategic interaction (e.g. crop clustering for mutual benefit), and spatial correlations in the environment in which farmers operate, such as knowledge spillovers, social conformity concerns, and perceived positive or negative external effects of the adoption decision (Skevas, Skevas, and Swinton 2018). The role of peer effects on agricultural technology adoption is a topic of increasing importance. Empirical studies in this area focus on the effect of peer influence on the adoption of high-yielding rice varieties (Holloway, Shankar, and Rahmanb 2002), milk recording (Läpple et al. 2017), bioenergy crops (Skevas, Skevas, and Swinton 2018), organic agriculture (Lewis, Barham, and Robinson 2011; Läpple and Kelley 2015; Wollni and Andersson 2014), and conservation practices (Kolady et al. 2020), and provide evidence of the importance of neighbourhood effects on technology adoption.

In the context of UAVs, peer effects or interactions among neighbouring farmers are important for several reasons. First, UAVs are a new agricultural technology at an early stage of diffusion, and hence it is likely that there will be a higher level of interaction (and possibly dependence) between nearby farms on issues related to technology use than in a case where technology is already well-established (e.g. organic farming). This argument is supported by findings in the literature showing that information on the use of new agricultural technologies becomes less important as the technology is adopted by a relatively large percentage of the farmer population (Chatzimichael, Genius, and Tzouvelekas 2014). Second, several characteristics of the UAV technology make it extremely interesting for spatial decision-making analysis. For example, even though UAVs can be fully automated (i.e. take off, fly, and land on their own), farm operators will still need to acquire the knowledge on how to operate them successfully. For instance, flight experience is a skill that will be needed to successfully operate a UAV in the field. The need for farmers to develop a whole new set of skills to use a UAV may lead them to gather information or cooperate with neighbouring peers that have experience with UAVs. As another example, some farmers may raise privacy concerns to the extent that neighbouring farmers’ UAV operations may involve collection, retention and use of data from nearby properties. Again, this may influence neighbouring farmers' UAV adoption decisions. All these reasons make UAV adoption an interesting study subject in the context of spatial decision-making analysis.

Against this background, the objective of this work is to empirically examine the role of peer effects in UAV adoption decisions. We apply a Bayesian spatial autoregressive probit model to data from a rich sample of more than 800 crop farmers in Missouri, which seeks to identify farmer awareness, perception, and adoption of UAVs. We contribute to the literature by being the first to examine the role of peer effects on farmer decision to adopt UAVs. The remainder of this article is organized as follows. Section 2 presents the methods we used to study the role of peer influence on farmers’ decision to adopt UAVs. In section 3 we present the data and discuss the empirical strategy. Section 4 reports and discusses our estimation results, and section 5 concludes.

2. Methodology

In order to explain farmers’ decision to adopt UAVs and determine if neighbourhood effects impact UAV adoption, a spatial probit model is used. Recognizing that both neighbours’ characteristics (e.g. large farmers are more likely to adopt innovations that spillover to their neighbours) and outcomes (i.e. UAV adoption in our study) may influence a farmer’s decision to adopt UAVs, we first considered a spatial Durbin (SD) probit model. This model, which can be obtained from a simple rearrangement of the spatial error model, allows for the neighbours’ UAV adoption decisions, the farmer’s characteristics, as well as a linear combination of neighbouring farmers’ characteristics to affect UAV adoption decisions. Because the SD model resulted in insignificant estimates for almost all neighbours’ characteristics and was outperformed by a model that only includes farmer’s characteristics and his/her neighbours’ outcomes as adoption determinants (for more details, see the supplementary material, sections 1 and 2), the latter model was chosen for further analysis. This model is known as Spatial Autoregressive Model (SAR). Using the vector $\mathbf{y}$ to denote farmers’ UAV adoption decisions, the SAR probit model is defined as follows:

(1) ${\mathbf{y}}^{\ast }=\rho \mathbf{W}{\mathbf{y}}^{\ast }+\mathbf{X}\beta +\epsilon$(1)

with its corresponding Data Generating Process (DGP) being:

(2) ${\mathbf{y}}^{\ast }={\left(\mathbf{I}-\rho \mathbf{W}\right)}^{-1}\left(\mathbf{X}\beta +\epsilon \right)$(2)

and the observed equation written as:

(3) $\mathit{y}=\left\{\begin{array}{c}1,if{\mathbf{y}}^{\ast }>0\\ 0,if{\mathbf{y}}^{\ast }\le 0\end{array}\right.$(3)

where ${\mathbf{y}}^{\ast }$ is a $N\times 1$ vector of continuous unobserved variables (one for each unit in the sample ($i=1,..N))$, whose values determine whether the observed vector $\mathbf{y}$ is equal to zero (not adopt) or one (adopt), $\mathbf{X}$ is a $N\times K$ matrix of covariates that may affect farmer’s decision to adopt UAVs (e.g. farmers’ socioeconomic characteristics), $\mathbf{W}$ is a spatial $N\times N$ weighting matrix, $\rho$ and $\beta$ are scalar and $N\times 1$ vector of parameters, respectively, to be estimated and $\epsilon$ is a $N\times 1$ vector of error components that capture statistical noise. We assume that $\epsilon$ follows a standard-normal distribution and therefore the latent-variable formulation presented above leads to a probit model.

The SAR model suggests that farmers’ decisions to adopt UAVs do not only depend on their own characteristics captured by $\mathbf{X}$ but also on a linear combination of neighbouring farmers’ adoption decisions represented by $\mathbf{W}{\mathbf{y}}^{\ast }$. The latter relationship is quantified by the parameter $\rho$, which is called the spatial autoregressive parameter. The $\mathbf{W}{\mathbf{y}}^{\ast }$ variable can be endogenous since unobserved farmer or regional characteristics can simultaneously affect both on-farm and neighbouring farms’ adoption decisions. To address the potential endogeneity of $\mathbf{W}{\mathbf{y}}^{\ast }$ we follow Läpple and Kelley (2015) and assume that the adoption decision of the farmer $i$ is influenced by past adoption choices of his/her neighbours. More formally:

(4) ${\mathbf{y}}_t^{\ast }=\rho {\mathbf{W}\mathbf{y}}_{t-1}^{\ast }+{\mathbf{X}}_{\mathbf{t}}\beta +\epsilon$(4)

where $\mathbf{W}{\mathbf{y}}_{\mathbf{t}\mathbf{-}\mathbf{1}}^{\mathbf{\ast }}$ is the time lag of the adoption decision of neighbouring farmers in the previous year. This assumption is not trivial because often farmers wait to become familiar and see the effectiveness of an innovation on other farms’ performance before they adopt it on their own farms. Recursive substitution of past values for the vector ${\mathbf{y}}_{t-1}^{\ast }$ on the right-hand side of eq. 4(4) ${\mathbf{y}}_t^{\ast }=\rho {\mathbf{W}\mathbf{y}}_{t-1}^{\ast }+{\mathbf{X}}_{\mathbf{t}}\beta +\epsilon$(4) over $T$ time periods implies that we can interpret the cross-sectional SAR model as the expectation of a long-run steady state relationship or equilibrium (LeSage and Pace 2009):

(5) $li{m}_{T\to \mathrm{\infty }}E\left({\mathbf{y}}_t^{\ast }\right)={\left(\mathbf{I}-\rho \mathbf{W}\right)}^{-1}\mathbf{X}\beta$(5)

The estimate of $\rho$ indicates the existence of spatial dependence in farmers’ adoption decisions but not the economic significance (Elhorst 2014). Instead, quantitative statements can only be made by deriving the so-called direct and indirect (or spillover) marginal effects of the independent variables $\mathbf{X}$. To do so, one needs to focus on equation (2)(2) ${\mathbf{y}}^{\ast }={\left(\mathbf{I}-\rho \mathbf{W}\right)}^{-1}\left(\mathbf{X}\beta +\epsilon \right)$(2) and calculate the partial derivative of ${\mathbf{y}}^{\mathbf{\ast }}$ with respect to each explanatory variable included in $\mathbf{X}$. This results in the following matrix: ${\left(\mathbf{I}-\rho \mathbf{W}\right)}^{-1}\beta \phi \left(\mathbf{X}\beta \right)$ where $\phi \left(\cdot \right)$ stands for the standard normal density function. This matrix allows the computation of three marginal effects: the direct, indirect, and total marginal effects. The direct effects correspond to the diagonal elements of the above matrix and reflect the impact of changing a particular explanatory variable for a particular farmer on the expected long-run value of the dependent variable for that farmer. The indirect effects correspond to the row sums of the off-diagonal elements of the above matrix and capture the change on farmers’ long-run value of the dependent variable as a result of a unit change in a particular explanatory variable of their neighbours.¹ Finally, the total effects are simply the sum of the direct and indirect effects.

3. Data and estimation procedures

3.1. Data

The dataset used in this study comes from a mail survey of Missouri farmers conducted in 2018. In the absence of a list of all farmers in Missouri, the individuals contacted were drawn from a mailing list consisted of landowners who owned 100 or more acres of agricultural land. The process used to create the mailing list first entailed selecting a stratified random sample of 12 Missouri counties and then continued with a random selection of 250 landowners who owned 100 or more acres of agricultural land from each sampled county.² The identification of this type of landowners relied on property tax records obtained from county assessor offices. Dillman’s Total Design Method was used to administer the survey (Dillman, Smyth, and Melani 2011). This method involved a series of mail contacts with all prospective participants, using a prenotification contact, two survey mailings, and reminder postcards. The 809 respondents who provided usable data yielded an adjusted response rate of 28%.

The survey included questions on current land use and management practices, use of UAVs in agricultural operations, attitudes and awareness of UAV agricultural applications, and demographics. In the land use and management section, participants were asked how many acres of land they owned and, whether they currently rented out any portion of their land, whether they raise livestock on their farm, and how many acres they enrolled under the conservation reserve programme. The section on UAV usage asked respondents whether they were using a UAV in their agricultural operations. UAV users were further asked to state the year they first used a UAV for agricultural purposes. The attitudinal and awareness variables include both directly observable and constructed variables. The former group of variables includes farmers’ beliefs about UAVs’ ability to reduce input costs and protect the environment. The constructed variables were derived by applying factor analysis on a set of 13 Likert-scale and dichotomous questions related to awareness about UAV agricultural application, and concerns about UAV costs and operational knowledge and training, UAV data collection and interpretation, and UAV-related issues with neighbours.³ This process resulted in three constructed attitudinal variables: UAV awareness, UAV usage concerns, and concerns with neighbours. Finally, demographic questions on gender, age, cooperative membership, income, education, and presence of successor completed the survey. Table 1 presents summary statistics of the data used.

Table 1. Descriptive statistics of surveyed farmers

Variable name	Definition	Units	Mean	SD
rent out	rent out land to other farmers	(0/1)	0.45	0.50
livestock	raise livestock on the farm	(0/1)	0.44	0.50
farmland	acres of farmland	acres	852.42	1070.28
crp	land in Conservation Reserve Programme	(0/1)	0.24	0.43
age	age	years	65.45	13.06
male	male gender	(0/1)	0.86	0.35
coop	farmer is a member of a cooperative	(0/1)	0.47	0.50
income	household income	1-6^a	3.49	1.31
educ	education	1-6^b	3.20	1.33
successor	availability of successor	(0/1)	0.50	0.50
UAV use	uses a UAV for farm operations	(0/1)	0.08	0.27
UAV usage concerns	concerned with UAV usage	score	-2.88e^−^9	1.00
UAV awareness	awareness of UAV ag applications	score	-3.62e^−^9	1.00
concerns with neighbors	neighborhood-related concerns of using UAVs	score	2.03e^−^9	1.00
UAV_environment	believes UAVs can reduce environ. footprint	(0/1)	0.35	0.48
UAV_input cost	believes UAVs can decrease input cost	(0/1)	0.17	0.38

Source: Skevas and Kalaitzandonakes (2020)

3.2. Empirical specification

The variable $\mathbf{y}$ reflects the decision of farmers to adopt UAVs and is captured by the UAV_use variable. This variable reflects whether or not respondents were using a UAV for their agricultural operations in 2018. In order to avoid an endogeneity bias, a time-dependence formulation is used by assuming that the adoption decisions of farmers in time $t$ depend on past adoption choices (i.e. $t-1$) of their neighbours.⁴ The spatial weighting matrix (i.e. $\mathbf{W}$), which approximates the neighbouring relationship between farmers, is specified using an inverse distance matrix. Such a matrix is specified because it places a higher weight on closer than more distant neighbours. This is in line with the First Law of Geography that states ‘everything is related to everything else, but near things are more related than distant things’ (Tobler 1970, 236).⁵ The information that closer neighbours exert a stronger influence than more distant ones would be lost by employing a k-nearest neighbour or contiguity matrix. The use of an inverse distance matrix to specify the neighbourhood links is a common practice in spatial econometric studies of agricultural decision-making (Läpple and Kelley 2015; Storm, Mittenzwei, and Heckelei 2015; Skevas, Skevas, and Swinton 2018). The diagonal elements of $\mathbf{W}$ are set to zero by assumption since no farmer can be viewed as its neighbour. The off-diagonal elements $w_{ij}$ are set equal to $1/d_{ij}$, where $d_{ij}$ is the Euclidean distance between farmers $i$ and $j$, if those two farmers operate within a certain distance $d^{\ast }$ (known as the distance cut-off), and 0 otherwise. Following Skevas and Grashuis (2020), $d^{\ast }$ is set to the minimum distance that all producers in our sample have at least one neighbour, which is 50 km in our case. Sensitivity analysis with respect to the distance cut-off point was performed using a variety of alternative cut-off points (i.e. 20, 40, 60 and 70 km) as in Bell and Dalton (2007). As in Skevas and Grashuis (2020), we normalize $\mathbf{W}$ by dividing each of its elements by its largest characteristic root. Unlike row normalization, this process allows the mutual proportions between the elements of $\mathbf{W}$ to remain unchanged, thus permitting $\mathbf{W}$ to maintain its economic interpretation in terms of distance decay (Kelejian and Prucha 2010; Elhorst, Lacombe, and Piras 2012).

Lastly, the covariates in $\mathbf{X}$ include farmers’ land use and production practices (i.e. rent_out, livestock, farmland and crp), socioeconomic characteristics (i.e. age, male, educ, income, coop and successor), and level of awareness and attitudes towards the use of UAVs (i.e. UAV awareness, UAV usage concerns, concerns with neighbours, UAV_environment and UAV_input cost).

3.3. Bayesian estimation

The SAR probit model is estimated using Bayesian techniques as in LeSage and Pace (2009). Bayesian estimation is based on the model’s likelihood function, which is denoted as p(y, y*│X, β, W, ρ) and the imposition of prior probabilities on the parameters to be estimated (Assaf 2011). A multivariate-normal prior is placed on β. The mean vector is set equal to 0, and the precision matrix is diagonal with a value of 0.001 in its diagonal entries. This prior is denoted as p(β). Multivariate normal prior distributions are almost exclusively used in the Bayesian econometrics literature for parameters that can take any value on the real line Skevas (2020b). Furthermore, the fact that the precision of the utilized prior distribution is set equal to such a low number makes this prior rather uninformative, manifesting our ignorance about the true values of the parameters and our intention to allow the data to speak for themselves about these values. A Beta prior is imposed on ρ to restrict it to the unit interval.⁶ The first shape parameter equals 2 and the second shape parameter equals 4. This prior is denoted as p(ρ). Given that the mean of the Beta distribution equals the ratio of the first shape parameter and the second shape parameter, this parameterization results in a prior value for ρ that is close to the findings of Skevas, Skevas, and Swinton (2018). Having specified the likelihood of the SAR model and the priors on the parameters to be estimated, application of Bayes’ theorem leads to the following posterior:

p(y*, β, ρ│y, X, W) ∝ p (y, y*│X, β, W, ρ) × p(β) × p(ρ) (6)

Drawing samples from the posterior presented in equation (6) involves the use of Markov Chain Monte Carlo (MCMC) and data augmentation techniques.

4. Results and discussion

Before presenting our model results, we briefly discuss neighbourhood relationships. Sample farmers have on average 112 neighbours within a 50 km radius. This number is larger than the respective figures reported in the literature that assesses spatial interdependence in farm decision-making in developed countries (Läpple and Kelley 2015; Läpple et al. 2017). Minimum distances to neighbouring farmers are, on average, 4.5 km, with the shortest distance being 124 m. Läpple et al. (2017), in their study of peer effects in sustainable technology adoption of Irish dairy farms, reported an average minimum distance between neighbouring farms of 6.9 km.

Moving on to the results of estimating the model in equations (1)(1) ${\mathbf{y}}^{\ast }=\rho \mathbf{W}{\mathbf{y}}^{\ast }+\mathbf{X}\beta +\epsilon$(1) to (3), first, we present evidence of spatial dependence in UAV adoption decisions. Then, we discuss the direct, indirect and total marginal effects.⁷ Summary statistics of the spatial autoregressive parameter ($\rho$) is presented in table 2. Since there is no significant difference in the economic interpretation of the spatial autoregressive parameter and the marginal effects across different weight matrix specifications, the results of the estimates with a 50 km distance cut-off are presented. Table 2 shows that the estimated $\rho$ parameter is positive and significant. This implies that, ceteris paribus, the adoption decision of neighbouring farmers influences a farmer’s decision to adopt UAVs. In terms of the magnitude of the spatial dependence, our finding is slightly lower than that of Läpple and Kelley (2015) who report spatial autoregressive parameter estimates between 0.093 and 0.135 when modelling the adoption of organic drystock farming in Ireland. On the other hand, higher estimates of spatial dependence were reported by Skevas, Skevas, and Swinton (2018) (between 0.180 and 0.290), Läpple et al. (2017) (between 0.284 and 0.342), and Wollni and Andersson (2014) (0.321) in their adoption studies of bioenergy crops, milk recording, and organic farming, respectively.

Table 2. Spatial autoregressive parameter estimate for the SAR model

	90% CI
ρ	Lower bound	Upper bound
0.054	0.01	0.121

Moving to the remaining results of the estimated model, table 3 provides marginal effect estimates which are separated into direct, indirect and total effects.⁸ If a farmer $i$ rents land to other farmers, then his/her UAV adoption probability is reduced by 4.8% (direct effect). A likely explanation for this finding is that farmers that rent out their land are less motivated to farm improvement and adoption of new innovations or choose to leave UAV technology adoption to their tenants. When farmer’s $i$ neighbours rent land to other farmers, this reduces farmer’s $i$ adoption probability by 0.3% (interpreted as the average row effect). The total effect is a reduction of the adoption probability by 5.1%.

Table 3. Marginal effects of the spatial probit model

	Marginal effects
	Direct effect	Indirect effect	Total effect
rent_out	-0.048	-0.003	-0.051
livestock	-0.003	-2.0E-04	-0.003
farmland	0.006	3.3E-04	0.006
crp	-0.022	-0.001	-0.023
age	-0.122	-0.007	-0.129
male	0.002	1.2E-04	0.002
coop	-0.002	-0.0001	-0.002
income	0.016	0.001	0.017
educ	-0.013	-0.001	-0.014
successor	0.021	0.001	0.022
UAV usage conserns	-0.006	-3.7E-04	-0.006
UAV awareness	0.020	0.001	0.021
Concerns with neighbors	-0.037	-0.002	-0.039
UAV_environment	0.056	0.003	0.059
UAV_input cost	0.055	0.003	0.058

Note: Credible intervals are not presented to ease visual interpretation, but are available from the authors upon request. Bold values denotes that the variable significantly affects the probability to adopt UAVs (i.e. the 90% credible interval does not cross zero).

Each additional year of increase in age of farmer $i$ reduces the probability of adoption of the $i^{th}$ farmer by 12.2%, while a one-year increase in age of neighbouring farmers leads to a 0.7% decrease in the probability of adoption of farmer $i$ .The combined total effect on all farms is 12.9% reduction of adoption probability. Older farmers are less likely to adopt UAVs, possibly because they are close to retirement and have less time to reap the benefits of an investment in a new technology. Other studies in the literature have found a similar relationship between farmer age and adoption of UAVs (Michels, von Hobe, and Musshoff 2020; Skevas and Kalaitzandonakes 2020), and other agricultural innovations (see for e.g. Barham et al. 2004; D’Antoni, Mishra, and Joo 2012; Leng et al. 2020). A likely explanation for the negative spillover effect of age and rent_out variables is that farmers’ motivation to adopt new technologies such as UAVs is negatively impacted by the reduced work motivation of neighbouring older farmers and those preferring to be landlords rather than farmers.

An increase in income increases farmer’s i adoption probability by 1.6%. When a farmer $i$ is surrounded by neighbours that experience a higher income then this increases farmer’s $i$ adoption probability by 0.1%. The total effect is an increase in adoption probability by 1.7%. High income can help reduce the risk of adopting new technologies and allow farmers to experiment with these technologies. These technologies can then spill over to neighbouring farmers, for example, through direct communication of the benefits of the technologies or a desire to conform with peers’ choices (Wollni and Andersson 2014).

The attitudinal variables UAV_awaremess, concerns with neighbours, UAV_environment and UAV_input cost have a significant impact on the probability to adopt UAVs. An increase in the level of UAV awareness by one standard deviation increases the own farmer’s (+2%) as well as neighbours’ adoption probability (+0.1%). The indirect effect of this attitudinal variable is interpreted as the average column effect and provides evidence of a likely knowledge spillover due to communication among farmers.

An increase in the level of concerns with neighbours by one standard deviation on farm $i$ reduces the probability of adoption on the same farm by 3.7%. Such behaviour implies that there is a perceived social cost to the adoption of UAVs. It is not uncommon that farmers care about the acceptance of their technology choices in their social environment. For example, Wollni and Andersson (2014) found that social conformity has a positive effect on the adoption of organic farming in Honduras. A one standard deviation increase in concerns with neighbours of all neighbouring farmers reduces farmer’s $i$ adoption probability by 0.2%. The total effect is a 3.9% decrease in probability.

Moving to the UAV_environment and UAV_input cost variables, an increase by one standard deviation in farmer’s $i$ belief that the use of UAVs can help protect the environment and lower total input cost increases the probability of adoption by the same farmer by 5.6% and 5.5%, respectively, and also increases the adoption probability of neighbouring farmers by a cumulative 0.3%. In relation to the direct effect of the belief variables, perceived economic and environmental benefits of new technologies are important for their uptake (Reimer, Weinkauf, and Prokopy 2012; Skevas et al. 2013) and UAVs are therefore no different in that respect. The indirect effect of the belief variables is interpreted as the average column effect and likely suggests that farmers share their knowledge or beliefs regarding the potential economic and environmental benefits of using UAVs with their neighbours.

In sum, the indirect or spillover effects of the attitudinal variables provide some evidence, albeit small, that UAV adoption decisions of farmers are influenced by their neighbours’ concerns and attitudes towards the use of UAVs. Läpple and Kelley (2015) and Skevas, Skevas, and Swinton (2018) also found a significant effect of neighbours’ attitudes on the uptake of organic agriculture in Ireland and the supply of land for bioenergy production in Michigan, respectively.

The relatively small indirect effects found in this study are an expected finding given the magnitude of the spatial autoregressive parameter $\rho$, presented in Table 2 (note that the indirect effects are derived by scaling the estimated parameters by $\rho W$). In addition, previous similar studies also reported small indirect effects (Läpple and Kelley 2015; Läpple et al. 2017), indicating that it is common to find small peer effects in an agricultural technology adoption context. However, we argue that although small, the reported indirect effects are not negligible. The importance of indirect effects is reflected in the fact that if one ignores peer influences, inference is simply based on direct effects, which are deflated when compared to the total effects that account for spatial spillovers (see Table 3). This holds for most variables and especially for age. The relatively small indirect effects can also be justified from a theoretical viewpoint. A farmer’s decision to adopt a new technology (like UAVs) would be primarily affected by his/her own characteristics, and less by the adoption decision and characteristics of his/her neighbours. For example, even if a liquidity- or credit-constrained farmer is convinced by his/her neighbouring peers to adopt an innovation, his/her financial position would most likely hamper the adoption process.

5. Conclusions

This study uses a spatial autoregressive (SAR) probit model under a Bayesian framework to examine the link between peer effects and UAV adoption by farmers. The empirical application uses cross-section data of 809 Missouri farmers surveyed in 2018 and provides evidence of the importance of peer effects in farmers’ UAV adoption decisions.

Our study provides two important conclusions. First, farmers’ UAV adoption decisions depend not only on their own characteristics and attitudes towards the UAV technology, but also on those of their neighbouring farmers. Second, our findings provide some empirical evidence on the channels of farmer dependencies. These channels are as follows: a) indirect effects of farmers’ age, income, and renting out land to other farmers, and b) neighbours’ attitudes and concerns regarding the use of UAVs. The proximity of older farmers and farmers who prefer to be landlords is associated with a negative influence on UAV adoption decisions, perhaps because they are less motivated to farm and refrain from investments in new technologies. High-income farmers are found to be more likely to adopt UAVs, and this has a positive spatial spillover effect on neighbours’ adoption decisions. This result likely suggests that farmers follow the practices of their more successful neighbours – a result that has been confirmed by other studies as well (e.g. Chatzimichael, Genius, and Tzouvelekas 2014). Farmers expecting economic and environmental benefits from the adoption of UAVs and those being more aware of UAV applications are more likely to adopt this technology and this increases their neighbours’ probability to adopt UAVs. On the other hand, perceptions about neighbour privacy concerns reduce the farmer’s and neighbours’ probability to adopt UAVs. These results imply that neighbours’ attitudes and concerns towards the use of UAVs correlate with farmers’ UAV adoption decisions. The finding that neighbours’ attitudes towards UAVs affect UAV adoption decisions is likely a sign that farmers communicate with each other and influence each other’s behaviour regarding the adoption of agricultural innovations.

Future research could extend this work by examining if peer effects in UAV adoption would be found in different states or countries where production systems and conditions and socioeconomic characteristics of the farmer population might be different; taking into account both geographic and socioeconomic criteria (e.g. farm profitability as in Skevas, Skevas, and Cabrera 2021) when defining the neighbourhood space; and providing a behavioural identification of why spatial spillovers arise in the adoption of UAVs.

Declarations and ethics statements

The authors have no competing interests to declare.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by the USDA National Institute of Food and Agriculture, Hatch project 1015805.

Notes

¹ Alternatively, the indirect effects can be computed as the column sums of the off-diagonal elements of the ${\left(\mathbf{I}-\rho \mathbf{W}\right)}^{-1}\beta \phi \left(\mathbf{X}\beta \right)$ matrix, which results in a different interpretation. The column effect shows the impact of changing a particular element of an exogenous variable $x_{ki}$ on the long-run value of the dependent variable of all farmers. The two ways of computing the indirect effect result in the same value and, as a result, it does not matter which one is used in the interpretation of indirect effects (Elhorst 2014). As in the study by Läpple et al. (2017), we discuss either the row or column effect, depending on which one makes more sense in the interpretation.

² For more information on the sampling strategy and data collection and preparation for analysis, see Skevas Kalaitzandonakes (2020).

³ See Section 3 in the supplementary material for the questions used in the factor analysis. More information on the performed factor analysis could be found in Skevas and Kalaitzandonakes (2020).

⁴ Details on how a time-dependence SAR formulation overcomes the endogeneity bias can be found in Skevas (2020a) on an efficiency analysis setting.

⁵ In the context of farming, Maertens and Barrett (2012) show that operating near another farmer increases the likelihood of a link.

⁶ Restricting ρ in the unit interval is not uncommon in the farm spatial dependence literature (see for e.g. Läpple et al. 2017; Skevas, Skevas, and Swinton 2018; Pede et al. 2018; Skevas 2020a) and lies on the argument that if a particular variable (UAV adoption in our case) increases (decreases) in one area, it also tends to increase (decrease) in neighbouring areas (Elhorst 2014). This is true in our case because farmers tend to adjust their behaviour to that of their successful neighbouring peers (Conley and Udry 2010; Chatzimichael, Genius, and Tzouvelekas 2014) or in an effort to conform to communal norms and professional ideals (Wollni and Andersson 2014).

⁷ The results are based on the following MCMC sampling scheme: the first 20,000 draws are discarded to remove the influence of the initial values, 1 out of 2 draws are retained to remove potential autocorrelations and a total of 60,000 draws are retained from the posterior density.

⁸ The posterior moments with respect to the covariates in X are presented in Table A1 in the appendix, for completeness.

Appendix

Table A1. Posterior moments of the determinants of UAV adoption

		90% CI
	Coef.	Lower bound	Upper bound
Constant	2.088	-0.080	4.275
rent_out	-0.425	-0.726	-0.135
livestock	-0.030	-0.279	0.219
farmland	0.051	-0.080	0.183
crp	-0.195	-0.510	0.110
age	-1.077	-1.575	-0.587
male	0.019	-0.430	0.499
coop	-0.016	-0.265	0.235
income	0.140	0.040	0.241
educ	-0.110	-0.216	0.001
successor	0.186	-0.071	0.447
UAV usage conserns	-0.057	-0.182	0.070
UAV awareness	0.174	0.005	0.353
Concerns with neighbors	-0.327	-0.485	-0.174
UAV_environment	0.497	0.232	0.767
UAV_input cost	0.486	0.219	0.755
ρ	0.054	0.010	0.121
Log marginal likelihood	-243.52	-	-
Posterior probability	0.000	-	-