2. Background

The choice experiment in our study presents participants with an opportunity to adopt turfgrass that require less water and fertilizer than traditional turfgrass. Hard fescue (Festuca brevipila Tracey), chewings fescue (Festuca rubra L. ssp. commutata Gaudin), and other similar fine fescues are alternatives to conventional species such as Kentucky bluegrass (Poa pratensis L.) and perennial ryegrass (Lolium perenne L.) for temperate climates commonly found in the northern United States, where lawns typically are an important part of urban landscapes and cover large areas (Milesi et al., Reference Milesi, Elvidge, Dietz, Tuttle, Nemani and Running2005). Kentucky bluegrass (Poa pratensis L.) is a winter-hardy turfgrass that requires relatively large amounts of water, fertilizer, and pesticides to meet consumers’ esthetic and functional requirements.

Public pressure to reduce the amount of water and chemicals required for lawns in the United States is increasing. Previous studies have examined the factors and interventions that may impact household water conservation behavior (Addo, Thoms, and Parsons, Reference Addo, Thoms and Parsons2018; Fielding et al., Reference Fielding, Russell, Spinks and Mankad2012; Goette, Leong, and Qian, Reference Goette, Leong and Qian2019; Qaiser et al., Reference Qaiser, Ahmad, Johnson and Batista2011; Willis et al., Reference Willis, Stewart, Panuwatwanich, Williams and Hollingsworth2011). Many states, including Minnesota, have recently imposed restrictions on phosphorus fertilizer (Lee and McCann, Reference Lee and McCann2018), and outdoor watering restrictions are common throughout the country (Milman and Polsky, Reference Milman and Polsky2016). As concern about droughts associated with climate change (Sheffield and Wood, Reference Sheffield and Wood2008) intensifies, pressure to reduce water dedicated to home lawns will likewise increase. Previous studies have established that social awareness and exposure to drought affect residents’ attitudes toward water saving (Beal, Stewart, and Fielding, Reference Beal, Stewart and Fielding2013; Knuth et al., Reference Knuth, Behe, Hall, Huddleston and Fernandez2018). Knuth et al. (Reference Knuth, Behe, Huddleston, Hall, Fernandez and Khachatryan2020) showed that plant watering requirement plays a more important role for residents of states who frequently experienced drought.

Reducing inputs required for lawns can be accomplished using various approaches, but perhaps the most effective is using alternative turfgrass species. Previous studies have demonstrated that low-input turfgrass can perform well with limited water and no fertilizer or pesticides and require less mowing (Diesburg et al., Reference Diesburg, Christians, Moore, Branham, Danneberger, Reicher, Voigt, Minner and Newman1997; Hugie and Watkins, Reference Hugie and Watkins2016; Watkins et al., Reference Watkins, Fei, Gardner, Stier, Bughrara, Li, Bigelow, Schleicher, Horgan and Diesburg2011). Several studies have shown that consumers are willing to pay premium prices for certain attributes of low-input turfgrass. Yue, Hugie, and Watkins (Reference Yue, Hugie and Watkins2012) conducted a choice experiment targeting Minnesota consumers and found that they are willing to pay premiums for turfgrass that requires less irrigation, fertilization, and mowing. Similarly, Yue et al. (Reference Yue, Wang, Watkins, Bonos, Nelson, Murphy, Meyer and Horgan2017) found that consumers in the United States and Canada are willing to pay a statistically significant price premium for turfgrass that leads to less watering, fertilizer inputs, and mowing. Hugie, Yue, and Watkins (Reference Hugie, Yue and Watkins2012) conducted a conjoint analysis and found that maintenance attributes in general and irrigation requirements in particular have significant effects on consumers’ willingness to purchase low-input turfgrass seeds. Ghimire et al. (Reference Ghimire, Boyer, Chung and Moss2016) found that consumers in the southern United States rank low maintenance, drought tolerance, and saline tolerance as the most desirable lawn attributes.

Despite the numerous benefits of low-input turfgrass and evidence that consumers are willing to pay a premium for those benefits, adoption of low-input turfgrass for lawns in the United States has been limited (Braun et al., Reference Braun, Patton, Watkins, Koch, Anderson, Bonos and Brilman2020). All of the previously mentioned studies assumed that the alternative turfgrasses would perform as well as conventional ones and did not consider potential adoption risks. As with any new product, adoption of a new turfgrass involves uncertainty about its performance. Low-input turfgrass could be susceptible to diseases and/or weeds that would not be problematic for conventional species, for example, and consumers likely would know less about the new species’ specific management requirements. Our study investigates consumer decision-making about low-input turfgrass from a behavioral perspective to shed light on factors that prevent consumers from adopting this useful technology.

3. Experiment Design

To investigate barriers to adoption of low-input turfgrass, we first convened a focus groupFootnote ² involving a small number of homeowners. We asked them about their turfgrass maintenance practices, the attributes of the grasses that were most important to them, and their awareness of alternative turfgrass requiring fewer inputs. We also showed participants a sample square of sod planted with a low-input turfgrass and asked them about their preferences regarding the grass and their willingness to switch to it. We found both limited adoption of the low-input turfgrass and little awareness of such grasses as an alternative for lawns. The focus group discussion revealed several prevalent concerns that potentially prevent participants from adopting low-input turfgrass: the amount of effort required to replace the entire lawn at once, uncertainty about whether the new turfgrass would survive, and whether the long-term cost and time savings would cover the initial investment cost.

We used the results of the focus group to design our choice experiment analyzing homeowners’ decisions regarding adoption of low-input turfgrass. Specifically, since the effort required to completely convert a lawn to low-input turfgrass was a concern, we included three options in the choice experiment: one option replaced all of the conventional grass at once using sod and two options replaced the conventional grass gradually using overseeding. Note that the one-time full replacement option provides an immediate return on the investment while the gradual replacement options provide delayed returns. To address expressed concerns about the alternative turfgrass failing to survive, we included the probability of success for the low-input turfgrass and the additional expense required to replace it if it failed. We assumed that participants’ reference points were to the status quo of conventional turfgrass and that all of their calculations of time saved, costs incurred, and potential gains would be based on their current investments in their lawns.

In the choice experiment, participants were first given an introduction to low-input turfgrass:

The experiment administrators then explained the context of the choice experiment, which presented participants with scenarios simulating a situation in which a lawn service company provided a service converting conventional lawn grasses to low-input turfgrasses. The lawn service setting was designed to eliminate potential confusion among participants regarding the amount of effort required to make the switch. Participants then received a detailed written explanation of the choice scenarios. In each choice task, participants would be presented with three replacement options (one full replacement and two gradual replacements), and the instruction stressed that there was no “correct answer” and that their task was to determine which option provided them with the greatest overall value and worked best for their households.

Each choice scenario consisted of four options—the three replacement options (A, B, and C) and the option to opt-out and decline conversion (D). Option A represented full immediate replacement using sod and delivered returns immediately since there was no need to wait for seeds to germinate and become established. This option not only represents the most rapid way to adopt the technology but also involves the greatest upfront investment of time and money. Option B represented conversion over an intermediate length of time of 2 years using overseeding. During those two transitional years, there was a cost per year for seed and participants would begin to obtain the benefit of conversion at the beginning of the third year. Option C represented a slower overseeding conversion of 3 years, a fixed investment cost for seed, and benefits obtained at the beginning of the fourth year. As defined by prospect theory, the losses and gains from conversion and their associated probabilities are relative to the status quo, which is the participant’s current lawn, and Option D is the status quo.

Each option presented in the experiment was characterized by five attributes. The first was the upfront cost (sod, seed) per 1,000 ft², which was $600, $800, and $1000 under Option A; $300, $400, and $500 under Option B; and $240, $320, and $400 under Option C. The second attribute was time savings of 20%, 40%, and 60% provided by low-input turfgrass during the growing season (less mowing, watering, and fertilizing). The third attribute was potential cost saved per year after completing the conversion, thanks to reduced needs for water and fertilizer, which was set at $200, $250, and $300. The fourth attribute was probability of success of the new turfgrass (that it would grow as expected and require no further investment), which was set at 70%, 80%, and 90%. The fifth attribute was the additional investment required if the alternative product failed; it was set at $150, $175, and $200.Footnote ³ We designed the choice experiment with optimal D-efficiency.

The choice experiment was designed to capture intertemporal discounting, present bias, probability weighting, risk aversion, and loss aversion. To capture present bias, we asked participants to choose between immediate adoption with a high one-time investment cost and slower adoption with a lower per-period investment cost over multiple periods. Gradual adoption under Options B and C, which have different transition periods, was designed to capture the discounting factor. The varying probabilities of gains and losses were designed to capture probability weighting, and the varying amounts of losses and gains were designed to capture loss aversion and risk aversion.

We conducted the experiment in 2019 and used Qualtrics™, a professional survey company, to obtain a representative sample of U.S. homeowners 18 years of age or older. Interested participants were first asked a screening question: “Do you have a home lawn?” Individuals who answered yes were allowed to participate in the data-collection survey conducted prior to the choice experiment. Since the experiment design involved a 3-year transition period under Option C, the survey asked participants whether they expected to live at their current residences for less than 3 years and excluded all participants who answered affirmatively. The final sample for the choice experiment consisted of 866 participants.

4. Model Setup and Estimation Strategy

To analyze the results from the choice experiment, we set up a model that incorporates both prospect theory and intertemporal choice with present bias to capture how homeowners make decisions about adopting low-input turfgrasses. Let i = {1, …, I} denote homeowner i; j = {0, 1, 2, 3} denote opt-out Option D, Option A, Option B, and Option C, respectively, and ν _ij the utility function of individual i choosing option j with opt-out utility normalized to 0 (ν _i0 = 0). Assuming that the value function of the options, u _ij , has the separable additive form,

(1) $${{u_{ij}} = {{\rm{\nu }}_{ij}} + {\varepsilon _{ij}}}$$

where u _ij is the value function from option j and ϵ _ij is the random utility error term assumed to be type-I extreme-value distributed. Thus, by McFadden (Reference McFadden1974), the probability of individual i choosing option j can be defined as

(2) $${{s_{ij}} = P{r_i}\left( {{u_{ij}} \ge \mathop {\max }\limits_{k \in \left\{ {0,1,2,3} \right\}} \{ {u_{ik}}\} } \right) = {{exp\left( {{\nu _{ij}}} \right)} \over {1 + \mathop \sum \nolimits_{k = 1}^3 exp\left( {{{\rm{\nu }}_{ik}}} \right)}}}.$$

Individual i evaluates the utility from option j at a finite horizon that is assumed to be the expected duration of individual i’s residency at the current location. Let T ₁ denote the length of the transition period, which is one of the predefined attributes, and T ₂ denote the expected duration of residence. We directly asked participants how long they expected to live at their current addresses. The utility from option j should consist of three components: (1) utility from average time savings per period (&#xi; _j ), (2) the net present monetary value of the investment (η _j ), and (3) the alternative-specific constant term (a _i ). Assuming utility takes the separable additive form,

(3) $${{{\rm{\nu }}_{ij}} = {{\rm{a}}_j} + {{\rm{\omega }}_2}{{\rm{\xi }}_j} + {{\rm{\omega }}_1}{{\rm{\eta }}_j}}.$$

In equation (3), ω₁ and ω₂ are the marginal utility from average time savings and the net present monetary value of the investment, respectively.

When assuming exponential discounting, expected utility ν _ij should take the form

(4) $$\eqalign{ {\nu _{ij}} = \,& {a_j} + {{\rm{\omega }}_2}\mathop \sum \nolimits_{t = {T_1} + 1}^{{T_2}} {{\rm{\beta }}_2}^t{{\rm{\tau }}^{{{\rm{\gamma }}_2}}}/\left( {{T_2} - {T_1}} \right) \cr & \!+ {{\rm{\omega }}_1}\left\{ \mathop \sum \nolimits_{t = {T_1} + 1}^{{T_2}} {{\rm{\beta }}_1}^t\left[ {{\rm{\pi }}{g^{{{\rm{\gamma }}_1}}} - \left( {1 - {\rm{\pi }}} \right){l^{{{\rm{\gamma }}_1}}}} \right] - \mathop \sum \nolimits_{t = 0}^{{T_1}} {{\rm{\beta }}_1}^t{c^{{{\rm{\gamma }}_1}}}\right\}. \cr} $$

In equation (4), β is the exponential discounting factor, τ is the time savings per period relative to current practice (percentage), γ ₁ and γ ₂ are the risk-aversion parameters, π is the probability of gain, and c _t , g _t , and l _t are the cost, potential gain, and potential loss in period t, respectively. The average present value of time saving is $\xi _{j}=\sum _{t=T_{1}+1}^{T_{2}}{\beta _{2}}^{t}\tau ^{{\gamma _{2}}}/(T_{2}-T_{1}),$ the expected total present value due to loss and gain generated by low-input turfgrass is $\sum _{t=T_{1}+1}^{T_{2}}{\beta _{1}}^{t}[\pi g^{{\gamma _{1}}}-(1-\pi )l^{{\gamma _{1}}}],$ and the present value of the total cost is $\sum _{t=0}^{T_{1}}\beta ^{t}c^{{\gamma _{1}}}.$ Thus,

$$\eta _{j}=\sum _{t=T_{1}+1}^{T_{2}}\beta ^{t}[\pi g^{{\gamma _{1}}}-(1-\pi )l^{{\gamma _{1}}}]-\sum _{t=0}^{T_{1}}\beta ^{t}c^{{\gamma _{1}}}.$$

To account for present bias, loss aversion, and probability weighting, the model takes the form

(5) $$\eqalign{ {\nu _{ij}} = & \, {a_j} + {{\rm{\omega }}_2}\mathop \sum \nolimits_{t = {T_1} + 1}^{{T_2}} {\beta _2}^t{{\rm{\tau }}^{{\gamma _2}}}/\left( {{T_2} - {T_1}} \right) \cr & \!+ {{\rm{\omega }}_1}\left\{ \mathop \sum \nolimits_{t = {T_1} + 1}^{{T_2}} \alpha {\beta _1}^t\left[ {w\left( \pi \right){g^{{\gamma _1}}} - w\left( {1 - \pi } \right)\lambda {l^{{\gamma _1}}}} \right] - \mathop \sum \nolimits_{t = 0}^{{T_1}} {\alpha ^{{1_{t \gt 0}}}}{\beta _1}^t\lambda {c^{{\gamma _1}}}\right\}. \cr} $$

The difference between equations (4) and (5) is three components: the quasi-hyperbolic discounting (Laibson, Reference Laibson1997) parameter α used to capture present bias, the probability weighting function w(⋅), and the loss-aversion parameter λ. The functional form suggested by prospect theory is adapted from Tversky and Kahneman (Reference Tversky and Kahneman1992) with the individual evaluating losses and gains according to the function ${V}(\cdot )$ ,

(6) $${V\left( {\rm{x}} \right) = \left\{ {\matrix{ {{x^{{{\rm{\gamma }}_1}}}} & {if\;x \ge 0} \cr { - {\rm{\lambda }}{{\left( { - x} \right)}^{{\gamma _1}}}} & {if\;x \lt 0} \cr } } \right.,}$$

and the probability weighting function,

(7) $${w\left( {\rm{\pi }} \right) = {{{{\rm{\pi }}^{\rm{\mu }}}} \over {{{\left( {{{\rm{\pi }}^{\rm{\mu }}} + {{\left( {1 - {\rm{\pi }}} \right)}^{\rm{\mu }}}} \right)}^{1/{\rm{\mu }}}}}}.}$$

It is straightforward to see that equation (5) is a generalization of equation (4); that is, equation (5) becomes equation (4) when α = 1, μ = 1, and λ = 1. We assume an individual’s reference point is the status quo.

Having specified an individual’s utility and likelihood, we estimate the distribution of the parameters using Bayesian inference. The choices made by the participants are represented by ${Y}=\{y_{1},\ldots,y_{I}\}$ , a vector in which each element represents one of the four options in a choice scenario (opt-out, A, B, and C). We are interested in estimating the cumulative distribution function of the parameters (e.g., F(Θ|Y)). However, the nonlinear nature of the value function makes it difficult to solve using optimization algorithms so we use Bayesian inference instead. We assume that the prior cumulative distribution of the parameters was

(8) $$\left\{ {\quad \matrix{ {{\beta _1} \sim U\left( {0,1} \right)} \hfill \cr {{\beta _2} \sim U\left( {0,1} \right)} \hfill \cr {\alpha \sim U\left( {0,1} \right)} \hfill \cr {\mu \sim N\left( {1,1} \right)} \hfill \cr {\lambda \sim N\left( {3,2} \right)} \hfill \cr {{\gamma _1} \sim U\left( {0,1} \right)} \hfill \cr {{\gamma _2} \sim U\left( {0,1} \right)} \hfill \cr {{\tau _1} \sim N\left( {0,3} \right)} \hfill \cr {{\tau _2} \sim N\left( {0,3} \right)} \hfill \cr {{a_j} \sim N\left( {0,3} \right)} \hfill \cr } } \right.$$

where U(⋅,⋅) refers to a uniform distribution and N(⋅,⋅) denotes a normal distribution. By Bayes’ theorem, the distribution of the parameters is proportional to ${F}({Y}|\Theta ){F}(\Theta )$ . In other words,

(9) $${F\left( {{\rm{\Theta |}}Y} \right) \propto F\left( {{\rm{Y|\Theta }}} \right)F\left( {\rm{\Theta }} \right)}$$

where F(Y|Θ) is the likelihood function and F(Θ) is the distribution of the parameters. The likelihood function F(Y|Θ) is given by

(10) $${F\left( {Y{\rm{|\Theta }}} \right) = \mathop \prod \limits_{i \in \left\{ {1,..,I} \right\}} \mathop \prod \limits_{j \in \left\{ {0,1,2,3} \right\}} s_{ij}^{{1_{\left\{ {{y_i} = j} \right\}}}}{{\left( {1 - {s_{ij}}} \right)}^{{1_{\left\{ {{y_i} \ne j} \right\}}}}}.}$$

The prior probability function is given by equation (8).

The posterior distribution F(Θ|Y) is estimated using a differential-evolution Markov chain Monte Carlo (MCMC) method (ter Braak and Vrugt, Reference ter Braak and Vrugt2008) and implemented using BayesianTools in R (Hartig et al., Reference Hartig, Minunno, Paul, Cameron, Ott and Pichler2019). We employed seven chains in the MCMC process, which allowed for testing of convergence. We ran each chain until the last 5,000 iterations indicated convergence. That is, until the potential scale reduction factor (PSRF) was less than 1.1 for all parameters, the convergence threshold suggested by Brooks and Gelman (Reference Brooks and Gelman1998).