As mentioned above, our archetype definition is based on a Q methodological study by Braito et al. (2020). These authors conducted their study with 33 Austrian crop farmers (selected from a range of backgrounds, e.g., with/without livestock, organic/conventional, male/female, AES/no AES, different regions) and identified four different archetypical farmer viewpoints (hereafter: archetypes) in winter 2017/18. The 34 statements (“Q set”) used by Braito et al. (2020) reflect the potential determinants of soil management: aspects relating to farm, farmer, socio-institutional context, and natural context that may determine farmers’ management choices. The statements were gathered through a literature review and six stakeholder interviews. The question to which interviewees sorted the statements on a scale ranging from -4 (disagreement) to + 4 (agreement) was “What determines how you manage your soil?”. Braito et al. (2020) identified four soil management archetypes: Nature Participants (NP), driven by their relationship with nature and having a focus on innovation in soil management; Pleasure Seekers (PS), sharing a focus on nature but considering personal freedom and joy as essential; Traditional Food Providers (TFP), prioritizing food production and valuing traditions in managing their soil, and Profit Maximizers (PM), motivated by their farms’ economic viability and profitability.
We use these four archetypes to group respondents of a questionnaire survey into four types and then model respondents’ participation in AES in an econometric model. While the archetypes primarily relate to soil management and AES also cover additional aspects, we deem it appropriate to link soil management archetypes to AES participation for two reasons: First, Braito et al.’s (2020) archetypes resemble broader archetypes from previous studies (see, e.g., Davies and Hodge 2007) and the statements used comprise almost all aspects that are relevant for the (environmental) management of a crop farm, beyond mere soil management. Farms that have a different focus (esp. grassland, permanent crops or intensive livestock production) may not be well represented by these archetypes, but our questionnaire focuses on crop farms and mixed farms only. Second, The AES relevant to such farms largely target soil management. At the time of our study, the five schemes (out of 23) that received most subsidies were all at least partly related to soil management on cropland,Footnote 1 and among the top 10 schemes in terms of total subsidies spent there was only one that was relevant for crop farms but not soil management relatedFootnote 2 (BMLFUW 2020).
Assigning survey respondents to farmer archetypes
We transfer the Q set used by Braito et al. (2020) to our questionnaire survey in the following way (Danielson 2009). Questionnaire respondents were presented with 31 statements of the Q set and asked to indicate their agreement with each statement on a five-point Likert-type scale, ranging from “strongly agree” to “strongly disagree”. The statements were grouped into sets of 10–11 to ease respondents’ evaluation. In their German original, the first 21 statements started with the phrase “When dealing with my soil…”, which was written out only once per set to reduce the reading load for respondents. The corresponding statements were then restricted to their second half (e.g.: “…I rely on my own education and experience”). The remaining 10 statements were presented as full sentences. Compared to the original Q study, we removed three statements that had clearly been identified as consensus statements by Braito et al. (2020), i.e., statements that all archetypes had ranked similarly. Table 1 lists all statements in the order they were presented to the survey respondents; respondents’ mean responses, as well as the statements’ respective ranks (-4 to + 4) by the four archetypes.
Table 1 Statements from the Q set as presented to survey respondents, statement rankings by archetypes (see Braito et al. 2020), and mean survey response We apply two different methods to group our survey respondents according to the archetypes: the “scale creation method” (SC method) (Danielson, 2009; also presented by Brown (2002) and Baker et al. (2010) as “standardized factor index score”) as well as the “profile correlation method” (PC method) (Danielson 2009). To avoid confusion between the archetypes as identified by Braito et al. (2020) and the individual survey respondents (partly) sharing these archetypical views, we will refer to the latter as a farmer’s “type” hereafter.
For the SC method, we select two defining statements for each archetype. These selected statements need to fulfill two criteria (Danielson 2009): salience (i.e., the respective archetype agreed or disagreed strongly with these statements) and distinction (i.e., the respective archetype differed (significantly) from other archetypes in its agreement with these statements). Whether a statement is ‘distinguishing’ for one archetype to satisfy the latter criterion can be determined by statistical significance (this is also used in Q methodology itself). However, in some cases, the statistically-determined “distinguishing statements” for an archetype in Braito et al. (2020) do not satisfy the salience criterion. In these cases, we select statements that are salient and clearly representative of the respective archetype in a more qualitative sense. For example, we choose the statement “managing my soil ought to give me pleasure” as a defining statement for the Pleasure Seeker archetype because it is at the core of the archetype, even if it only weakly distinguishes the archetype from others. In Table 1, all defining statements are printed in bold.
After determining these defining statements, we create a score for each survey participant on each archetype. Table 2 illustrates this process by means of an example; participant 58, who is defined as a Nature Participant type according to his/her responses and the resulting maximum (normalized) viewpoint score. This process involves the following steps: 1) reverse code participant responses (PR) to those statements that the archetypes placed on the negative side of the Q distribution, creating PR’, 2) multiply PR’ with the absolute value of this statements’ archetype ranking (AR) to create the participant score (PS) for each statement, 3) sum the PS values per archetype to obtain an archetype score (AS), and 4) normalize the AS into T-scores (mean: 50, standard deviation: 10) to account for differences in the attainable maximum scores. We then assign to each participant the type that she/he scores highest on.
Table 2 Example for determining one respondents' type based on the scale creation method For the PC method, Danielson (2009) again suggests presenting a number of representative statements per archetype to survey respondents and then correlate each participant’s responses with the rankings of these statements by each archetype. Compared to the SC method, each archetype needs to be represented by a larger number of statements to allow for meaningful correlation results. These statements do not (all) need to be salient but can also be located in the middle of the Q distribution. We utilize all 31 statements presented to survey respondents. This slightly modified version of Danielson’s method (he suggests selecting only a subset of statements) reduces the subjective judgment required for selecting representative statements. Aside from this modification, we proceed as suggested and correlate each participant’s responses with each archetype’s Q rankings, using a Spearman correlation. In essence, we correlate each row of our dataset with one row per archetype that contains this archetype’s ranking. This produces correlation scores for each survey respondent with each archetype that we directly use in our further analysis. Therefore, an individual participant may correlate positively with each or multiple of the four archetypes to some degree.
The econometric model
The farmer types determined by the SC and the PC methods then serve as our explanatory variables of interest in econometric models of AES participation. AES participation consists of two decisions that we can model conjointly or separately: a farmers’ decision to participate in any AES at all, and a farmers’ decision on the level of participation in AES; i.e., the decision on the number of schemes to participate in or the intensity of these schemes (e.g., schemes that require substantial changes to the farming operation vs. schemes that require little change). We define both decisions in terms of (the existence of) per-hectare AES income. As non-participants have zero AES income the dependent variable is censored at zero.
Depending on theoretical and statistical considerations, several modeling options for zero-censored dependent variables exist (for helpful discussions of these options, see for example Madden (2008), Humphreys (2010), and Carlevaro et al. (2009)). Our model choice is based on the following considerations. First, we consider all zeros as true zeros that arise from one mechanism: non-participation as a matter of principle (as opposed to, e.g., non-participation due to AES payments being too low). This appears reasonable, given Austria’s ‘broad and shallow’ approach to AES that results in very easy access to several low-level schemes for all potentially interested farmers (the Austrian agri-environmental program explicitly aims at achieving comprehensive AES coverage of all agricultural land). Second, we wish to investigate actual (not potential) outcomes, and to consider the participation and level of participation outcomes separately, since we suspect that farmer types may play different roles in the corresponding decisions. This leads us to the use of a two-part model, which essentially consists of a Probit model to model participation, combined with an OLS regression model of the level of participation for participants only (Belotti et al. 2015; Madden 2008).
The Probit model (first part of the two-part model) is used to estimate the probability of a positive outcome \(Y\), i.e., an AES income above zero, ϕ (Y > 0) = Pr(Y > 0 | X, T), where \(\boldsymbol{T}\) is either a set of dummies representing survey respondents’ farmer types based on the SC method or the set of correlation coefficients for each type based on the PC method, and \(\boldsymbol{X}\) is a vector of control variables (see below). To model the participation level decision in the second part, we model ϕ (Y|Y > 0, X, T), again using the same \(\boldsymbol{T}\) and \(\boldsymbol{X}\) as above, in an OLS regression specified as \( \text{for\;all}\;Y|Y\; > \;0,\;Y\; = \;\alpha \; + \;{\varvec{\beta T}}\; + \;{\varvec{\gamma X}}\; + \;\varepsilon,\)
where \(\alpha \) is an intercept; \(\varepsilon \) is the error term; and \(\boldsymbol{\beta }\) and \(\boldsymbol{\gamma }\) are vectors of parameters to be estimated.
For comparison, we also estimate a linear OLS regression model where we treat the two decisions (participation and participation level) as one. All calculations were done in R (R Core Team 2018).
As outlined in the introduction, we expect that a farm’s production portfolio and characteristics of the farm(er) are related to AES participation (Arata and Sckokai 2016; Pufahl and Weiss 2009; Zimmermann and Britz 2016). We therefore include the following control variables \(\boldsymbol{X}\) in all models: the log of farm size (utilized agricultural area (UAA) in ha), cattle density and the density of pigs and poultry (both in livestock units (LU) per ha), the farms’ rental share (share of rented UAA), productivity (all outputs/all inputs), whether the farm receives any payments for being situated in a least favored area (LFA, dummy variable), whether the farmer has finished education of ‘Matura’ (graduation exam from secondary school, permitting university entrance) or higher (dummy variable), and the farmer’s age (in years).
Data and variables
The implementation of our model draws on two main data sources: Austrian data from the EU’s farm accountancy data network (FADN), and a survey with Austrian farmers that participate in the FADN. The FADN collects annual harmonized micro-economic data on commercial farms in all EU countries to evaluate their income and the impact of the Common Agricultural Policy (CAP). Data are gathered via stratified samples by national agencies. While aggregated data are freely available online, these agencies (in Austria the Federal Ministry of Agriculture, Regions and Tourism (BMLFUW)) provide farm-level data to scientists for research purposes. We use these farm-level data as control variables on farm structure and economic indicators and for our dependent variable on AES income.
In Austria, a vast majority of farmers participate in AES. Correspondingly, only 19 (6.6%) of the farmers in our sample have an AES income of zero (“zero participants”). This is partly due to the existence of a scheme that has farming requirements almost identical to the Austrian ‘greening’ requirements for the CAP’s first-pillar payments (BMLFUW 2015) and that is therefore accessible to almost all farms with little additional effort. To account for this, we subtract the potential payments for this most basic scheme, ‘environmentally sound and biodiversity-promoting management’, from the total sum of payments. This corresponds to approximately 45€ per ha, depending on total UAA and type of farmland. In our case, AES participants are therefore defined as farmers who participate in more than just this basic scheme, raising the number of non-participants in our sample to 40 (13.9%). In terms of summary statistics, the group of ‘non participants’ does not differ fundamentally from the group of ‘zero participants’: both have a UAA that is significantly smaller, and an average number of pig/poultry LU per ha that is significantly higher than the total sample.
To determine farmers’ types (as described above) and to include information on respondents’ age and education level in our model we use the data collected in an online questionnaire survey. The survey was conducted in spring 2018 and was sent out to the 1,147 FADN farmers (out of a total of 1,879 FADN farms) who farmed at least 5 ha of cropland and rented part of this land. The survey consisted of three sections: the first section was part of a study on agricultural land renting (not used here), a second section contained the Q statements presented in Table 1, and a third section asked for additional sociodemographic information (only information not included in the FADN data). To connect the survey data to the economic FADN data, respondents were required to enter their FADN farm ID at the beginning of the survey. The entire questionnaire took about 20–30 min to complete, with the first section being the most time-consuming. No debriefing questions for reliability checks were included; however, we did not see any obviously unreliable results (such as identical responses to an entire block of statements). The tax and accountancy consultancy firm that administers the FADN data collection on behalf of the Austrian federal ministry pre-tested the questionnaire, identified and contacted farmers, sent out the survey invitations, and encouraged farmers to participate via e-mail, phone calls, and during their annual farm visits.
We attained a response rate of 31% with 344 fully completed questionnaires. Considering that the survey was a lengthy and voluntary online survey with no incentives or compensation attached, this is a reasonable response rate for a social science study (Sauermann and Roach 2013), and is comparable to other farmer studies (Avemegah et al. 2020). Since contact details of the farmers remained with the consultancy managing the questionnaire, we did not conduct non-response bias checks. However, since we have FADN data for all farms that were contacted for the survey, we can compare respondents and non-respondents to some extent (see below). A total of 300 respondents provided a correct FADN ID, enabling us to use their data for our analysis. We further excluded permanent crop farms and farms with an output share of > 49% of vegetables from our analyses, since their structure and AES income differs considerably from other farms. We additionally excluded one farm with very high leverage from the model, as we could not determine why that farm had an unusually high AES income.
Table 3 compares FADN data for respondents and eligible non-respondents to evaluate the representability of the sample. It shows that significant differences exist with respect to farm size, cattle density, and rental share. All these differences are likely a result of the study focus on farms with cropland and rented land. Farms with cattle and with little rented land are therefore underrepresented and results may not be fully transferrable to such farms. In the outcome variable – AES payments per ha – respondents do not differ significantly from non-respondents.
Table 3 Comparison of survey participants and nonparticipants (arithmetic means) For the 288 survey respondents, Table 4 shows descriptive statistics of the variables used in the model; for the full sample and by farmer type as determined by the SC method. For variables taken/computed from the FADN, Table 4 also provides the variable names as defined and used by the European Commission (2020).
Table 4 Summary of the variables used in the regression model (arithmetic means and percentages), including FADN variable names One observation in Table 4 worth mentioning concerns the prevalence of and relationship between the four farmer types defined by the two different methods. The first row of the Table shows the number of farms per type in our sample (as defined via the SC method). Here, the different farmer types appear to be distributed rather evenly among the general survey population. As the second row shows, this is also true for the AES non-participants, with the exception of the TFP type. The bottom four lines of the Table show the mean correlation coefficients of respondents with the archetypes as calculated by the PC method. Here we see that overall the correlation with archetypes varies, and survey respondents’ mean correlation with the Traditional Food Provider and Pleasure Seeker archetypes is lower than with the other archetypes (correlation scores for non-participants – not shown here – are very similar). The Table also shows that the different ways of identifying types do not lead to identical results, as PC correlation scores with one archetype are not necessarily highest for those assigned to the same type according to the SC method.
From Table 4 it also becomes evident that some substantial differences between types exist concerning AES payments, but also concerning other farm characteristics such as UAA and the presence of livestock. Therefore, it is essential to include farm structural variables as controls in our analysis of the relationship between farmer types and AES participation.