Econometric models.

According to Balcombe et al. (2015), if a respondent has no more than two fixations to one attribute in a single choice task, the ANA problem exists; if respondents do not attend the attribute for at least half of the choice tasks, they are visual non-attenders. On the above basis, this study first examines the existence of ANA and visual non-attenders. To determine whether respondents’ attention during choice experiments affects sod-producers’ WTA estimates of sod attributes, we estimate conditional logit models (CLMs) including and excluding respondents’ fixation, and we compare the obtained WTAs between the two models using multiple testing procedures. We also estimate a scale heterogeneity multinomial logit model (SHLM) to examine the effect of potential changes in variance scale caused by attention changes. Finally, we discuss the factors between sod-producers’ WTAs and consumers’ WTPs, estimated from Ghimire et al. (2019), for the same sets of sod attributes. CLM and SHLM have been estimated using SAS 9.4 and R 3.5.1, respectively.

Conditional logit model.

The CLM focuses on respondents’ choices on product attributes rather than individual characteristics. Then, an individual i’s utility by choosing an alternative j from choice set k is formed by a deterministic component, $\mathbf{\beta }{\mathbf{X}}_{ijk}$, and a stochastic component, ${\epsilon }_{ijk}$, as:$$U_{ijk}=\mathbf{\beta }{\mathbf{X}}_{ijk}+{\epsilon }_{ijk},$$where X is a vector of choice-specific variables, $\mathbf{\beta }$ is a corresponding vector of parameters, and ${\epsilon }_{ijk}$ is an error component which has the iid Gumbel distribution. In the CLM, the unobservable scale parameter is set equal to $\frac{{\pi }^2}{6}$ to create a constant variance equal to one for all choice tasks. However, it is possible that the scale parameter varies across individuals and choice sets.

Scale heterogeneity multinomial logit model.

To address the potential scale change, the iid assumption for the random term is relaxed to allow the scale to vary following a particular distribution (Czajkowski et al., 2014). Drawing on Fiebig et al. (2010), the utility function of a scale heterogeneity model is specified as$$U_{ijk}=\mathrm{\beta }{\mathrm{X}}_{\mathrm{ijk}}+\frac{{\epsilon }_{ijk}}{{\sigma }_k},$$where ${\sigma }_k$ is the scale parameter as a function of a vector of variables, which is defined as an exponential function as ${\sigma }_k=\exp (\overline{\sigma }+\mathrm{\phi V})$, while $\overline{\sigma }$, $\mathrm{V}$, and $\mathrm{\phi }$ are the intercept of scale function, a vector of scale-related variables, and the vector of corresponding parameters, respectively. Because our study focuses on change in individuals’ attention during experiments, we specify the scale parameter as a function of change in eye fixations as$${\sigma }_k=\exp (\overline{\sigma }+\phi {\nu }_k),$$where ${\nu }_k$ is the specified fixation measurement in the kth choice task. If the parameter $\phi$ equals zero, we conclude no scale change; thus no learning or fatigue effect is present. If $\phi$ is positive and significant, the larger the fixation variable is, the smaller the error variance is; thus an increasing fixation variable across choice tasks indicates the existence of a learning effect, while a decreasing fixation variable indicates the fatigue effect. If $\phi$ is negative and significant, the larger the fixation variable is, the larger the error variance is; thus an increasing fixation variable confirms the existence of a fatigue effect, while a decreasing fixation variable confirms the learning effect.

Discrete choice experiment.

For the experimental design we started with 48 possible choice sets, divided into four blocks, and we used a fractional factorial design procedure to generate 12 choice tasks with a 98.5% D-efficiency with no priors. When presented to respondents, the orders of the choice tasks were randomized to eliminate starting-point effect. Even though randomization of choice sets addresses a problem of potentially violating iid assumption caused by choice set design, the potential violation of iid assumption due to a respondent’s attention changes during the experiment could still be problematic. Each choice task contains three alternatives: two “product purchase” options and one “no-purchase” option. The two purchase options each described a sod product with at least one of the valued attributes and a respective price. In consultation with a horticulture expert and professor at Oklahoma State University, and using estimates provided by two sod producers for farm gate prices in Kansas and Oklahoma, we determined the upper and lower bounds of the price attribute for warm-season sod. The prices presented to respondents varied at one of the four levels: $0.15/ft², $0.25/ft², $0.45/ft², and $0.55/ft². Producers were asked to consider, “Options A and B represent two different sets of sod/sod bermuda or zoysia marketing characteristics and reductions in buyers’ maintenance such as weed control, mowing, and fertilizer. Which option (A, B, or C) would you be most likely to produce to market to consumers?” Figure 1 presents a sample choice task as seen by respondents. To discuss factors between sod producers and consumers, our study focuses on the same set of attributes as in Ghimire et al. (2016, 2019): winterkill reduction, shade tolerance, drought tolerance, salinity tolerance, and maintenance cost reduction. Winterkill occurs when sod suffers irreversible damage due to low temperatures (Dionne et al., 2001; Trenholm, 2000). Because photosynthesis requires sunlight, the decrease in light intensity due to shade can reduce the quality of sod. Warm-season sod, in general, is less shade-tolerant than cool-season sod (Moss et al., 2009; Patton, 2008). The high amount of water usage from sod irrigation, particularly under restricted water usage, increases the need for drought-tolerant sod (Leksungnoen et al., 2012; Zhao et al., 1994). When the use of freshwater is restricted, groundwater is often in use for irrigation as an alternative water source. However, the high soluble salt levels in groundwater may cause salinity stress to grasses. Swift and Koski (2007) show that high salt levels in water may cause grass to suffer from root damage and increase drought kill. Finally, maintenance cost and price are also considered as important factors that can affect consumer choice of sod varieties (Ghimire et al., 2016, 2019; Yue et al., 2017).

View Figure

• Download as Powerpoint
• Download Figure

Respondents in our sample were warm-season sod producers recruited in person at the Turfgrass Producers International Conference (TPIC) in San Diego, CA, in 2015. Collecting data in person is preferred because of eye-tracking technology, and the conference venue made recruitment easier and more complete compared with methods such as e-mail recruitment. The recruited respondents were asked to take the choice experiments with eye-tracking devices installed in the conference booth and information area. Participants first completed the demographic information on a tablet and subsequently proceeded to the two working stations with computers hooked up to eye-tracking devices. Participants were asked to sit comfortably and aligned with the monitor of the eye tracker, with a nine-point calibration first conducted to ensure proper eye-tracking metrics. Participants then responded to the choice tasks and received a compensation fee of $30 at the end of the experiment. To obtain more precise eye measurements from respondents, we used two Tobii TX 300 Hz devices. These are one of the highest resolution devices capable of recording at 0.4° accuracy (binocular) and 0.15° precision (unfiltered). We used the TX 300 mounted 23 in. screen (Tobii screen attached to the eye tracker) at a resolution of 1920 × 1080 pixels and a nine-point calibration. The ranges in fixation counts for each attribute were very large. For example, the numbers of fixations for the box corresponding to the winterkill reduction, shade tolerance, drought tolerance, salinity tolerance, 10% maintenance cost reduction, and price ranged between 0–24, 0–30, 0–24, 0–29, 0–17, and 0–14, respectively. Fixation ranges provide an indication that the high-resolution Tobii device is capturing differences in our specified areas of interest even though they are relatively small.

Due to the long duration of the experiment (mostly between forty-five minutes and an hour) and the limited number of producers who produced warm-season sod among conference participants, we have a relatively small sample size. In total, 32 subjects participated in our survey. Four subjects’ data were missing eye-tracking information because we were unable to calibrate the device to their eyes. Thus, the usable data are from 28 subjects with 12 choice sets for each subject, adding up to a total sample size of 336.

Our sample has 21 female participants and seven male participants. Producers who attended the TPIC at their own expense tend to be the more innovative producers and may indeed include more females who manage and direct sod farms and family businesses. The average age of our participants was around 48 years, ranging from 26 to 76 years old. Nine participants had attended some college, while 17 participants had earned a bachelor’s degree or higher. Six of our respondents majored in sod management or plant and soil science, and the rest majored in other subject areas. Our respondents were mainly from South Carolina (22.2%), Illinois (14.8%), Oklahoma (14.8%), Texas (14.8%), and North Carolina (11.1%). The major grass types they produce are bermuda grass (41.9%), tall fescue (22.7%), St. augustine (14.4%), and zoysia (13.6%). Our respondents’ total production area represented 50,075 acres, which accounted for about 15% of sod production area nationwide (USDA-NASS, 2019). Two of them had larger than 10,000 acres of production area, while five others had production areas between 1000 and 10,000 acres. The rest of the producers had less than 1,000 acres in production. Based on the categories of 2017 Census of Agriculture (USDA-NASS, 2019), most respondents are larger-scale sod producers with a production area larger than 399 acres. Because we recruited our respondents at a networking and research conference, the sample may include more innovative sod producers. Even though our sample may be biased in terms of which producers attended the conference, the larger-scale, more innovative sod producers are more likely to adopt, produce, and sell more newly developed cultivars at premium prices. The top three largest buyers of our respondents’ sod were landscape services, commercial or residential developers, and golf courses. The respondents’ total annual sales in 2014 added up to $254 million, with average sales of $9 million per producer. Warm-season sod sales in Oklahoma, Texas, North Carolina, Georgia, and Florida in 2015 were about $109 million (Chung et al., 2018). Comparing the 2017 sales amount (USDA-NASS, 2019), our sample counts for more than 22% of the sod value in the United States. Thus, our sample is likely to encompass a large proportion of the U.S. market. Sod producer data were collected contemporaneously with information on sod consumer preferences, allowing for a unique insights and comparisons between producer and consumer preferences.

Unattended attributes.

Figure 2 illustrates respondents’ non-attendance to attributes. For each attribute, a diamond assigned to one choice set indicates that the respondent passed over or “non-attended” the attribute. Eleven respondents attended to all attributes except for price in all choice tasks, while the rest (17 respondents) ignored at least one sod attribute in one choice set. Only one respondent fully attended to all six stimuli including the price. Respondent 16 is a visual non-attender of all attributes. Among all stimuli, drought tolerance is the attribute with the least visual-non-attenders, while price is the one with the most visual-non-attenders. Thirteen respondents are visual non-attenders for the price stimulus. We claim the existence of ANA in our choice experiment data.

View Figure

• Download as Powerpoint
• Download Figure

Visual data analysis.

For each area of interest (AOI), four measurements were recorded, including time to first fixation, total fixation duration, fixation counts, and total visit duration. Time to first fixation is the length of the time that a respondent first fixates on an AOI. Total fixation duration is the entire length of time that a respondent fixates on one AOI. Fixation count is the number of times that a respondent fixates on an AOI. Total visit duration is the entire length of time that a respondent visits one AOI. Behe et al. (2015) stated that total visit duration is the most influential predictor of product choice. To control for individuals’ different reading habits, the proportion of total visit duration (PTVD) for each choice task among all choice tasks was included as the fixation measurement variable in regression, i.e., PTVD = $\frac{totalvisitduratio{n}_k}{{{\displaystyle \sum }}_{n=1}^{12}totalvisitduratio{n}_n}$, where k and n indicate the number of choice tasks.

Table 1 presents the pairwise-comparison results of PTVD for the first two choice sets (1 and 2) with each of the remaining choice tasks. We compare mean differences using the Friedman’s test because PTVD data in our subsamples neither follow normal distributions nor have equal within-group variances. Therefore, standard parametric tests may not be appropriate to test for differences in PTVD between choice sets. We thus conduct the Friedman’s test, which first ranks the values of each matched set and then tests whether the rank totals for subsamples differ significantly (Dulce et al., 2015). Because choices in the first task might be more inconsistent compared with other choice tasks (Campbell et al., 2015), we also compared task 2 with the rest of the choice tasks. Table 1 shows that there is an overall decreasing trend of PTVD as respondents move toward later choice tasks, as well as a larger variability in PTVD between choice task 1 and the rest of the choice tasks. The fact that respondents pay more attention to the first choice task is likely a learning effect as respondents move through tasks. Figure 3 confirms that there is an overall decreasing trend in total visit duration as well as the proportion of total visit duration from the first choice task to the last choice task. Total fixation duration and fixation counts also show a similar decreasing trend.

Table 1.Pairwise comparison of the proportional total visit duration (PTVD) for the first two choice tasks using the Friedman test (n = 336).

View Fullscreen

View Figure

• Download as Powerpoint
• Download Figure

Estimation results.

Table 2 reports estimated parameters and corresponding WTAs from three regression models: CLM without eye-fixation variables (I), CLM with eye-fixation variables (II), and SHLM with an eye-fixation variable (III). [Referring to Lagarde (2013), we attempted to include “ANA × attribute” interaction terms in the regression to distinguish between the attribute irrelevance and ANA, as well as detecting the effect of ANA on respondents’ WTAs. Results show that when ANA is present, the estimated coefficients are not significantly different with their counterparts when ANA is absent, indicating that respondents ignored some attributes because of ANA instead of irrelevance. Furthermore, including ANA in the regression does not statistically change respondents’ WTA estimates. Thus, we still exclude the ANA variables for the following regressions.] From Model I, the price coefficient is positive and significant at the 5% level, implying that producers prefer to market sod products that are more expensive. The estimated coefficients of all sod attributes are positive as expected and significant at the 5% level, indicating that producers prefer to sell sod products with improved attributes. From Model I, producers’ WTA for an attribute is calculated as the attribute coefficient divided by the price coefficient. All WTA estimates from Model I are significant at the 5% level. Results show that sod producers value drought tolerance the most ($0.42/ft²), followed by shade tolerance ($0.38/ft²), winterkill reduction ($0.31/ft²), salinity tolerance ($0.23/ft²), and 10% maintenance cost reduction ($0.20/ft²).

Table 2.Estimates of the conditional logit (CLM) and scale heterogeneity multinomial logit (SHLM) models.

View Fullscreen

To consider the effects of a change in an individual’s attention on choice or preferences (Van Loo et al., 2015), Model II extends Model I by incorporating the PTVD × attribute interaction terms. Compared with Model I, Model II does not improve model performance at the 5% level (P = 0.118). Estimation results reveal the same signs and significance levels for all coefficients. Producers’ WTA of each attribute is calculated as $$\frac{{\beta }_{attribut{e}_t}+{\beta }_{attribut{e}_t\times PTVD}\overline{PTVD}}{{\beta }_{price}+{\beta }_{price\times PTVD}\overline{PTVD}},$$where ${\beta }_{attribut{e}_t\times PTVD}$ is the estimated parameter of the interaction term between the tth attribute and PTVD, ${\beta }_{price\times PTVD}$ is the estimated parameter of the interaction term between price and PTVD, and $\overline{PTVD}$ is the mean of PTVD from 12 choice tasks. The ranking of attributes from Model II stays the same with that in Model I, with drought tolerance as the highest ($0.42/ft²), followed by shade tolerance ($0.39/ft²), winterkill reduction ($0.32/ft²), salinity tolerance ($0.24/ft²), and 10% maintenance cost reduction ($0.20/ft²). Coefficients of all interaction terms are insignificant, except for the “PTVD × drought tolerance” variable. The negative coefficient of the “PTVD × drought tolerance” interaction term indicates that an individual’s longer attention in one specific choice task would lower the value of the drought tolerance attribute in that task. As our respondents are sod producers who know the industry well, the longer time they spend in one choice task reveals that they take more time to consider the attributes and price combination, and that they are not quite confident about their choice. Krajbich et al. (2010) indicate that more-difficult decisions usually take respondents more time. In this case, they may lower their WTA to match the attributes and price combinations. While producers’ WTA for drought tolerance is the highest among all five attributes, it is probable that producers would lower the value of drought tolerance. However, this may be an empirical conclusion from our study.

Model III extends Model I by allowing scale to vary as individuals’ attention level changes by choice task. The log-likelihood ratio test confirms that Model III is superior to Model I at the 5% level (P = 0.013). The estimated coefficients of the price and sod attributes are all positive and significant at the 5% level. The ranking of attributes remains the same as in previous models, with drought tolerance as the highest (0.44/ft²), followed by shade tolerance ($0.39/ft²), winterkill reduction ($0.31/ft²), salinity tolerance (0.23/ft²), and 10% maintenance cost reduction ($0.21/ft²). The intercept for the scale function is positive but insignificant. According to the negative sign of the fixation-related scale term and decreasing trend of PTVD in Table 1 (taking Choice 2 as the base task), we identify that respondents paid identical attention between choices 2 and 3, and then less attention all the way to Choice 12. Consistent with previous literature (Czajkowski et al., 2014), the results indicate an overall learning effect in this study.

Statistical tests of WTA between models.

As the simulated WTAs do not satisfy normality and homogeneity of variance assumptions for the parametric tests, we conducted three nonparametric tests to investigate the preference change by comparing WTAs from Models II and III with that from Model I: Kruskal–Wallis one-way rank sum test, the confidence interval of differences, and the complete computational resampling approach. We generated 1000 data sets for each model based on estimates and their variance and covariance matrices using a Monte Carlo simulation procedure. The Kruskal–Wallis rank sum test is a one-way analysis of variance approach that tests whether two samples originate from the same distribution or not. For the confidence interval approach, first, differences in WTAs are calculated between all possible one million combinations (1000 × 1000) from two models in comparison. (As WTA is a monetary value, we rounded all numbers to two decimals so that the smallest unit of WTA is in one cent.) Then, the 2.5% and 97.5% percentiles of the calculated one million differences are used to obtain the 95% confidence interval. For the complete computational resampling approach, we order the WTA differences calculated from the one million pairs (used for the confidence interval approach) and calculate the proportion of differences that are less than and equal to zero among all pairs. Then, the P value of the two-sided test is computed by doubling the proportion of differences that are less than and equal to zero. Therefore, the null hypothesis of equality is rejected at the 5% level if the P value of the two-sided test is smaller than 0.05 (Poe et al., 2005).

Table 3 summarizes results of the three nonparametric tests. The Kruskal–Wallis rank sum test identifies differences in WTAs between models. When including the fixation variables as interaction variables, the estimated WTAs for winterkill reduction, shade tolerance, drought tolerance, and salinity tolerance show significant differences between Models I and II. The calculated WTA for 10% maintenance cost reduction from Model II is not statistically different from its counterpart from Model I. The WTAs for all attributes except for shade tolerance differ between Model III and Model I at the 5% level. Results from the other two tests (95% confidence interval and complete computational resampling tests) indicate no statistical differences in estimated WTAs between Models II/III and Model I at the 5% level. The three tests used in this study show mixed results, an outcome that provides no clear evidence about the role of individuals’ attention change on their preferences during a choice experiment. A possible reason is that our respondents are sod producers who are quite knowledgeable on sod attributes and the overall industry, and thus their preferences for sod attributes might be relatively stable under different choice environments.

Table 3.Willingness-to-accept comparisons between models.

View Fullscreen

Discussion of sod producers’ WTA and consumers’ WTP.

Ghimire et al. (2019) stated that consumers are willing to pay the highest premium for drought tolerance ($0.45/ft² to $0.86/ft²), followed by winterkill reduction ($0.18/ft² to $0.39/ft²), shade tolerance ($0.33/ft²), maintenance cost reduction (−$0.26/ft² to $0.08/ft²), and salinity tolerance ($0.03/ft²). Producers’ WTA of winterkill reduction is between the ranges of consumers’ WTP, which is slightly lower than consumers’ maximum WTP. Producers expect slightly higher price premiums from providing the shade tolerance attribute than consumers are willing to pay, as consumers’ WTP is lower than producers’ WTA. Producers’ WTA of drought tolerance is even below the lowest level of consumers’ WTP, indicating that producers may be able to receive higher price for providing the drought tolerance attribute to consumers in the five states. Producers expect higher prices than consumers’ WTP for providing the salinity tolerance and maintenance cost reduction attributes. Estimates from the two studies indicate considerable differences between consumer WTPs and producer WTAs. In general, producers’ WTAs are higher than consumers’ WTPs for the improved sod attributes except for drought tolerance, but the rankings for these attributes are the same between consumers and producers.

There might be two possible reasons for the divergence between producers’ WTA and consumers’ WTP. First, producers may have not updated themselves with consumers’ preferences for new sod attributes. In other words, they may have not realized turf consumers’ real preferences. Second, producers know consumers’ preferences, but there may be limitations in providing sod with improved attributes that consumers want. The current grass species may not meet consumers’ demands, and/or the cost of providing grass species with improved attributes could be too high.