Measuring heterogeneous preferences for the preservation of prime farmland with and without agrivoltaics

Abstract

This study addresses a gap in cultural ecosystem service (CES) assessment of prime farmland located in peri-urban areas by presenting results from a choice experiment recently conducted in Utah’s Wasatch Front region. The choice experiment was designed to account for heterogeneous effects associated with a wide array of socio-demographic and attitudinal characteristics on household preferences for farmland preservation, including farmland used for the joint production of solar power and agricultural products. We apply a mixed-logit model to our data that controls for preference heterogeneity among Wasatch Front households along two dimensions – at the individual household level and according to different household types. We find that the typical household is willing to pay a non-trivial annual fee to preserve the region’s existing peri-urban farmland, and to a lesser extent is willing to pay for agrivoltaics on that land. We also find some evidence of preference heterogeneity among different types of households for farmland preservation and agrivoltaics; heterogeneity based upon traditional socio-demographic characteristics such as household income and location, as well as unique attitudinal differences related to how households view themselves in relation to agriculture, farmland preservation, and the extent to which taxation is an appropriate mechanism to fund local public goods. These findings can serve as crucial components of broader land-use studies designed to account for the full range of agri-environmental ecosystem services.

Keywords:

• cultural ecosystem services
• choice experiment
• preference heterogeneity

JEL Classifications:

• Q15
• Q24
• Q57

Disclosure statement

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

Notes

1 See de Groot, Wilson, and Boumans (2002, de Groot et al. 2012), Millennium Ecosystem Assessment (MEA)) (2005), Gómez-Baggethun and Barton (2013), and Crossman et al. (2013) for background on the common classifications of ecosystem services.

2 Nahuelhual et al. (2014) observe that among households in more highly developed countries CESs are frequently ranked ahead of other types of ecosystem services in terms of their importance to human welfare. Although commonly overlooked in terms of determining their value monetarily (due to their intangibility and the inherent difficulties associated with their measurement), CESs are more accessible and intuitively appreciated by people in their daily lives than other agroecosystem services (Plieninger et al. 2015; Willcock, Peh, and Camp 2015). See Woods et al. (2020) for further exploration of these types of issues surrounding the valuation of CESs. See Baulcomb et al. (2015) for methods used to identify different types of CESs.

3 Lopez et al. (2012) propose that joint solar production is most relevant for regions such as the Wasatch Front given its prodigious supply of solar energy.

4 See Adamowicz et al. (1998), Boxall et al. (1996), and Hensher, Rose, and Greene (2015) for broad perspectives on the choice-experiment method. Johnston et al. (2017) and Weng et al. (2021) provide more recent assessments and guidance on choice experiments, particularly with respect to the number of different alternatives in a given choice situation. In designing and administering our survey, we adhered as closely as possible to Johnston et al. (2017) recommendations pertaining to choice experiments.

5 We also created dummy variables for each farming practice, e.g., practice1 = 1 if the existing peri-urban farmland is managed as cut hay, zero otherwise; practice2 = 1 if the land is managed for an early-season dryland crop, zero otherwise; etc. Further, we created a dummy variable to indicate pasturage or orchard, i.e., animal_orchard =1 if the land is managed for pasturage or orchard, zero otherwise). Running our regressions with these dummy variables in place of animal produced statistically insignificant results for each of the alternative coefficients.

6 See also Bulte et al. (2005) and Carson, Groves, and List (2014) regarding consequentiality in general. Measuring the extent of payment consequentiality and these types of interactions is admittedly beyond the scope of our study.

7 See Carson and Groves (2007) and Hensher (2001) for further discussion on the role of information in reducing potential hypothetical bias.

8 We applied Stata’s Dcreate routine to create the design (Stata/IC 14.2 for Windows (64-bit x86-64). It is commonly believed that an efficiency score above one indicates an efficient design.

9 Duke et al. (2012) and Rewitzer et al. (2017) report response rates of 47% and 43%, respectively, while Gaur et al. (2023) report a 24% response rate. Wang and Swallow (2016) do not report their survey’s response rate. In their meta-analysis of response rates from choice experiment studies in general, Watson, Becker, and de Bekker-Grob (2016) report an average response rate of roughly 50% for experiments consisting of at most four attributes. More recently, Davies, Wu, and Schaafsma (2023) and Lindberg, Swearingen, and White (2019) report response rates of 6.4% and 17%, respectively. As Davies, Wu, and Schaafsma (2023) point out, “the literature suggests that response rates to online surveys are likely to be lower than other survey methods (Manfreda et al. 2008). For example, a response rate of just 5% was found by Marta-Pedroso, Freitas, and Domingos (2007). Using postal mail to advertise a web-based survey (i.e., combining two methods) is likely to have reduced the response rate further. Needham et al. (2018) obtained a slightly higher response rate of 14.8% using the same approach, although they had a higher dropout rate, at 20.8% (footnote 6, page 6).”

10 Potential sample selection bias also exists because even though 98% of residents in both Layton and Spanish Fork have access to wired broadband of 25 mpbs or faster, respondents must nevertheless be computer-literate to participate in online surveys (US Census Bureau 2019; Willcock, Peh, and Camp 2015).

11 Based upon their empirical evidence, Lancsar and Louviere (2008) conclude that more than 20 respondents per block in a CE design are rarely required to estimate reliable models, but undertaking significant post hoc analysis to identify and estimate co-variate invariably requires larger numbers of respondents. Given that our experimental design consists of three blocks, our per-block sample size is 30 respondents. These findings generally concur with Louviere, Hensher, and Swait (2000). In their exhaustive review of the choice-experiment literature, de Bekker-Grob et al. (2015) found that 32% of 69 studies reviewed reported sample sizes (number of respondents) less than 100. Along with de Bekker-Grob et al. (2015), Rose and Bliemer (2013) suggest that sample sizes of 100 generate statistically significant parameter estimates in choice experiments. Using Johnson and Orme’s (2003) heuristic, a minimum number of respondents contingent upon the dimensions of our study (in particular, the number of choice situations per respondent and the largest number of attribute levels among our attributes) is 139. Our sample size therefore falls between these rules of thumb.

12 See Woods et al. (2020) for further information about the outcomes of these statistical tests.

13 As mentioned in Section 5, we also estimate the ML model assuming a triangular distribution for error terms ${\mathit{\upsilon }}_i$ (McFadden and Train 2000; Hensher, Rose, and Greene 2015).

14 MWTP can be calculated as the ratio of the two coefficients due to the linear-preference assumption expressed in Equation (2)(2) $$U_{\mathit{\text{ijk}}}\left({\mathit{w}}_{\mathit{\text{ijk}}}\right)={\mathit{\alpha }}_i{\mathit{x}}_{\mathit{\text{ijk}}}+\mathit{\beta }\left({\mathit{x}}_{\mathit{\text{ijk}}}\bullet {\mathit{z}}_i\right)+{\epsilon }_{\mathit{\text{ijk}}},$$(2) . See Hensher, Rose, and Greene (2015) and Alpizar, Carlsson, and Martinsson (2003) for further details.

15 The absence of a specific variable defined for low-income households in Table 2 by default relegates this class of households as the baseline household income category.

16 In theoretical terms, i.e., in relation to utility function $U\left(\bullet \right)$ specified in Equation (1)(1) $$\begin{array}{c}\mathrm{M}\mathrm{\text{ax}}\\ \left\{{\mathbf{w}}_{\mathit{\text{ijk}}}\right\}\end{array}\mathrm{ }{\mathrm{U}}_{\mathit{\text{ijk}}}\left({\mathbf{w}}_{\mathit{\text{ijk}}}\right)+{\mathrm{\epsilon }}_{\mathit{\text{ijk}}},\mathit{i}=1,\dots ,\mathit{N},\mathit{j}=1,\dots 3,\mathit{k}=1,\dots ,6.$$(1) of Section 3.1, ${\mathit{\text{CMRS}}}_{\mathit{\text{az}}}=-\frac{{\left.\left(\partial U\left(\bullet \right)/\partial a\right)\right|}_z}{\partial U\left(\bullet \right)/\partial c},$ where ${\left.\left(\partial U\left(\bullet \right)/\partial a\right)\right|}_z$ denotes the change in $U\left(\bullet \right)$ with respect to an incremental change in attribute $a$’s level for a given level of variable $z,$ and $\partial U\left(\bullet \right)/\partial c$ denotes the change in $U\left(\bullet \right)$ with respect to an incremental change in cost level $c.$ In our case, attribute level $a$ represents the level of preserve and $z$ represents energy = 1.

17 The MNL specification can be considered a special case of the ML specification, where estimation of the coefficient matrix ${\mathit{\alpha }}_i$ in Equation (2)(2) $$U_{\mathit{\text{ijk}}}\left({\mathit{w}}_{\mathit{\text{ijk}}}\right)={\mathit{\alpha }}_i{\mathit{x}}_{\mathit{\text{ijk}}}+\mathit{\beta }\left({\mathit{x}}_{\mathit{\text{ijk}}}\bullet {\mathit{z}}_i\right)+{\epsilon }_{\mathit{\text{ijk}}},$$(2) is effectively restricted to equal a vector of constants $\mathit{\alpha }={\mathit{\alpha }}_1=\bullet \bullet \bullet ={\mathit{\alpha }}_N.$

18 We also ran a specification with choiceA interacted with the amount of time a respondent spent completing the survey (choiceA*duration). This term was meant to further refine the potential effect of choiceA on respondents’ idiosyncratic choices, the hypothesis being that the shorter the survey’s duration, the more likely a respondent will have chosen Alternative A in any given choice situation. However, choiceA*duration was found to be statistically insignificant.

19 In addition to its poorer fit of the data, we are unable to reject the MNL model’s underlying behavioral association with the independence of irrelevant alternatives (IIA) using a standard Hausman test (Hausman and McFadden 1984).

20 The standard deviation for animal’s coefficient estimate was found to be statistically insignificant in a prior run of the ML model. Hence, in the model’s final version we estimated a constant coefficient for animal.

21 We conducted an internal validity test on our ML model, whereby the respondents’ actual choices in the experiment were compared with their stated attitudes toward farmland preservation and proximity to an agrivoltaic installation. We found that (1) respondents who indicate positive support for agrivoltaics situated on nearby farmland are, all else being equal, more likely to choose an alternative including agrivoltaics, and (2) respondents who either “strongly” or “somewhat” disagree with the statement, “There is enough farmland in Utah”, were, all else being equal, more likely to choose an alternative including full preservation of farmland. Results of these tests are available from the authors upon request.

22 The full suite of results for the regressions run to produce the results contained in Tables 5 and 6 are available upon request from the authors.

23 In our empirical calculations, we only include an estimate for in the denominator of Equation (4)(4) $${\mathrm{\text{MWTP}}}_{\mathrm{\text{az}}}=-\frac{{\mathrm{\beta }}_{\mathrm{z}}^{\mathrm{a}}}{{\overline{\mathrm{\alpha }}}^{\mathrm{c}}+{\mathrm{\beta }}_{\mathrm{z}}^{\mathrm{c}}},$$(4) for the variables listed in Table 6, i.e., the variables whose interaction with cost was found to be statistically significant in the regression whose results are presented in Table 6. This occurred solely for the interaction between cost and nonrural in the case of the preserve attribute, i.e., the estimate of was included solely in the calculation of the Differential WTP for the nonrural interaction with preserve.

24 These percentages are each relative to the reference term value for preserve of $125.44 per acre, e.g., −21% for lowmidinc equals (−25.81 ÷ 125.44) × 100, etc.

25 Layton’s estimated population density of 3,135 people per square mile is roughly 25% larger than Spanish Fork’s density of 2,360 (USA.com 2023).

Additional information

Funding

This work was supported by Utah Agricultural Experiment Station [UTA0-1646].