3. Choice Experiment Design and Hypotheses

The experiment features 750 ml cabernet sauvignon—the most widely produced and consumed varietal in the United States; it is also the most widely planted varietal in the OCP region (Karlsson and Karlsson, 2018; Outer Coastal Plain Vineyard Association, 2020; Thach, 2015).

Six appellations labels are included: OCP; New Jersey, USA; Pilesgrove, New Jersey (Figure 1); California, USA; Napa Valley; and Product of USA (Table 1). These appellations will henceforth be referred to as OCP, NJ, Pilesgrove, California, Napa, and USA, where the italic denotes appellations.

Table 1. Attributes of the choice experiment

^a Base Categories of Dummy Variables, Prices are continuous.

A Southern New Jersey winery in Pilesgrove could use the first three labels. OCP allows the wine to be marketed by its AVA, where the state name is omitted. Whereas NJ and Pilesgrove advertise the state name. As the same wine can be marketed under these labels, any differences in WTP (ceteris paribus) thus result from the appellation labels, which teases out the effect of the AVA.

The design leverages OCP’s fine wine production aspiration, which has been hindered by its state’s weaker reputation (Smith, Reference Smith2013; Thompson, Reference Thompson2018). In 2012, OCP wines competed against Bordeaux wines in a blind tasting, in a format similar to the Judgement of Paris that had propelled Napa into the first premier American wine region; the OCP wines were competitive, rank-sum tests found only 1 out of 12 OCP wines were deemed statistically inferior to the significantly higher-priced Bordeaux wines (Quandt et al., Reference Quandt, Taber, Quandt, Taber and Quandt2012).

Despite it, New Jersey has not been traditionally seen as a fine wine producing region. The state produces 2.7 million gallons of still wine in 2017, which is 0.17% of the U.S. production by volume (U.S. Department of the Treasury, 2018b). Its winemaking reputation may have been somewhat negative. Media depictions demonstrate such sentiment; for example, George Taber, a renowned wine journalist, points to the state’s history of “mass production of rotgut,” referring to the sweet Welch-like wines that the state is known for (Davidson, Reference Davidson2013). The negative reputation was said to be an obstacle to the region’s aspiring winemakers, prompting the creation of the OCP AVA as a means to shed the negative reputation of New Jersey wine (Chace and Smith, Reference Chace and Smith2013; Davidson, Reference Davidson2013; Smith, Reference Smith2013). The Wine Enthusiast magazine noted the rise of New Jersey wine quality, claiming that “New Jersey is ready for its time in the spotlight,” but its subtext imply that the state’s winemaking reputation is relatively lower (Thompson, Reference Thompson2018).

In this study, the degree to which OCP elevates the competitiveness of the New Jersey wine is a point of interest, specifically:

$$H_0^1{:}\ WT{P_{NJ}} - WT{P_{OCP}} = 0$$

Hypothesis 1 tests for the difference in WTP between the state appellation of New Jersey and OCP. Failure to reject the null indicates that the AVA is ineffective for price premium generation.

If hypothesis 1 is rejected with $WT{P_{NJ}} \lt WT{P_{OCP}}$ , hypothesis 2 would suggest whether the premium could be due to the omission of “New Jersey” from the label,

$$H_0^{2a}{:}\ WT{P_{Pilesgrove}} - WT{P_{OCP}} = 0$$

$$H_0^{2b}{:}\ WT{P_{Pilesgrove}} - WT{P_{NJ}} = 0$$

Pilesgrove is a town within the OCP; the Pilesgrove label thus represents a hypothetical sub-AVA of OCP in the vein of Calistoga AVA.Footnote ² Rejection of hypothesis 2a with $WT{P_{Pilesgrove\;}} \lt WT{P_{OCP}}$ , and the failure to reject hypothesis 2b would suggest that respondents fail to recognize the geographical relations of these New Jersey-based appellations. Also, the inclusion of Pilesgrove could inform producers of the potential of such townships sub-AVA.

As a rational and fully informed consumer would not have significantly different WTP for the same wine, the pattern above would indicate that the assumption of perfect information is violated. Therefore, the WTP for OCP is likely stemming from consumer’s lack of knowledge. This strategic use of AVA is reflected in real-world examplesFootnote ³ and as noted in Smith (Reference Smith2013).

As with all choice experiments, the number of required choice sets increases exponentially with the number of attributes, which renders a larger number of appellation labels impractical (Hensher et al., Reference Hensher, Rose and Greene2005). Nevertheless, the setting is sufficient to draw implications about marketing with relatively unknown AVAs.

In addition to appellations, the design also includes the Organic wine label and the Sustainable Winemaking label. These ecolabels are included to enhance realism, as it reflects wine sold in the market.Footnote ⁴ The role of these increasingly popular labels to consumer preference has been discussed, where there remains a general lack of consensus regarding whether the labels appeal to consumers (Abraben, Grogan, and Gao, Reference Abraben, Grogan and Gao2017; Delmas and Grant, Reference Delmas and Grant2014; Di Vita et al., Reference Di Vita, Pappalardo, Chinnici, La Via and D’Amico2019; Sogari et al., Reference Sogari, Corbo, Macconi, Menozzi and Mora2015; Waldrop, McCluskey, and Mittelhammer, Reference Waldrop, McCluskey and Mittelhammer2017). This is despite that the market for organic wine has been expanding steadily, and various claims of sustainability have been increasingly used in marketing (Berghoef and Dodds, Reference Berghoef and Dodds2011; de La Hamaide and Denis, Reference de La Hamaide and Denis2018; Ecolabel Index, 2019; Sogari, Mora, and Menozzi, Reference Sogari, Mora and Menozzi2016). Nevertheless, these ecolabels have been included here for realism, as the property is essential for data quality (Hensher et al., Reference Hensher, Rose and Greene2005; Louviere, Hensher, and Swait, Reference Louviere, Hensher and Swait2000; Train, Reference Train2003). Additionally, the inclusion may minimize experimenter demand bias, which arises from respondents being overly focused on the research objectives (Sawyer, Reference Sawyer1975).

Lastly, four levels of prices are used, which range from $7.99 to $16.99 in increments of $3.00. These prices correspond to the $10 per 750 ml bottle after-tax mean retail price nationally (Bekkerman and Brester, Reference Bekkerman and Brester2019; Quackenbush, 2017; Wine and Vines, 2017). Bekkerman and Brester (Reference Bekkerman and Brester2019) show that wine purchases above the price of $20 represent only a smaller segment of the market. Thus, the price range is sensible given the goal of charting the consumer’s preference at the mean level, rather than the narrower preferences of connoisseurs. The chosen range implies that the results might not reflect the market for higher-priced wines, where different evaluation criteria are likely used. Nevertheless, the proposed implementation satisfies our main objective, which is to understand how AVA affects most consumers.

The choice experiment features graphical representations of wine labels as depicted in Figure 2. In particular, the attributes appear as text on the wine labels, where the exact form is denoted in Table 1. In cases where a product is without Sustainable Winemaking or the organic grape claim, the space where the labels occupy are left blank, reflecting the common practice in the marketplace.

Figure 2. A sample choice set.

The study, as a positive analysis, intends to simulate the actual preference in the marketplace. Thus, respondents were not told the exact locations of the geographical labels and the definitions of Sustainable Winemaking and organic wine. This is so that consumers’ preference for the attributes is not altered with the provision of information, as previous experiments have shown that such information can alter preference (Uchida et al., Reference Uchida, Roheim, Wakamatsu and Anderson2013).

To reduce hypothetical bias, i.e., respondents’ tendency to inflate their stated WTP, a cheap talk script from Lusk (Reference Lusk2003) is used in the study. The script has been shown to reduce hypothetical bias (Lusk, Reference Lusk2003; Tonsor and Shupp, Reference Tonsor and Shupp2011).

The choice sets are generated with D-efficiency criteria with the Partial Factorial Algorithm in JMP 13 (Crabbe and Vandebroek, Reference Crabbe and Vandebroek2012). The final design scores highly at 93.84%, where all main effects can be estimated (Kuhfeld, Reference Kuhfeld2010). The algorithm yields 24 unique choice profiles; the choice profiles are distributed into 48 choice sets (Hensher et al., Reference Hensher, Rose and Greene2005). Then, the choice sets are distributed into six blocks. Each respondent completes eight choice sets, which is below the threshold of fatigue (Czajkowski, Giergiczny, and Greene, Reference Czajkowski, Giergiczny and Greene2014).

4. Model

A mixed logit (ML) model estimates the consumer preference. The model uses random coefficients to represent taste variation, allowing consumer WTP at different percentiles to be estimated (Train, Reference Train2003). Following the Random Utility Model,

$${U_{ijt}} = \alpha {p_{ijt}} + {\bf{\mu }}{{\bf{x}}_{ijt}} + {{\rm{\varepsilon }}_{ijt}}{\rm{\;\;\;}}\left( 1 \right)$$

where ${U_{ijt}}$ represents the utility level of respondent i from the wine in the choice set t’s alternative j. The variable p represents the price level; its fixed coefficient avoids unrealistic WTP estimates implied by a random price coefficient (Hensher et al., Reference Hensher, Rose and Greene2005). The model sets OCP, non-ecolabeled wines as the reference category. The dummy-coded vector x depicts the non-price attributes, x = [Opt Out, Organic, Sustainable, NJ, Pilesgrove, California, Napa, USA].Footnote ⁵ The coefficients associated with x are assumed to be normally distributed. Formally, ${\bf{\mu }}\sim{\rm{N}}\left( {{\bf{\beta }},{\rm{\;}}{{\bf{\sigma }}^2}} \right)$ : ${\bf{\beta }}$ corresponds to the mean; $ {\sigma ^2}\;$ corresponds to the variance term that captures unobserved taste variation. The random coefficients are specified correlated to account for correlated taste among the attributes (Hensher et al., Reference Hensher, Rose and Greene2005).

The model’s likelihood function involves an integral without a closed-formed solution. It is thus estimated with the Maximum Simulated Likelihood Method of STATA 15’s Mixlogit module (Hole, Reference Hole2007; Train, Reference Train2003). The results are simulated with 2000 Halton draws, where the stability of the estimates is established (Walker, Reference Walker2002).

WTP estimates are derived from the ML model. Means of the WTP are calculated as $ - {\beta _{attribute}}/\alpha $ , and their standard errors are derived with the delta method (Hensher et al., Reference Hensher, Rose and Greene2005). To illustrate taste heterogeneity, the median and the 25th percentile of the WTP are simulated with the Krinsky and Robb method; specifically, 1,000 vectors of ${\beta _{attribute}}$ are generated, wherein the random coefficients’ correlation is accounted (Hensher et al., Reference Hensher, Rose and Greene2005).

7. Econometric Estimates

The ML model converges with McFadden R ² of 26% (Table 4) (Hensher et al., Reference Hensher, Rose and Greene2005). The price coefficient ( $\alpha $ ) is negative (P < 0.01) as expected. The opt-out coefficient $\;({\beta _{opt - out}}$ ) is negative (P < 0.01)—a sign of successful sample screening, as the respondents are interested in the products. The positive coefficient $\;({\beta _{organic}}$ ) indicates that organic is a preferred attribute, although at the mean level, consumers are indifferent toward sustainable winemaking.Footnote ⁷ Six of eight diagonal elements of the random coefficient’s Cholesky matrix are statistically significant, suggesting substantial taste heterogeneity toward the attributes (Hensher et al., Reference Hensher, Rose and Greene2005). A joint test that all the tested appellations are equivalent to OCP is rejected (P < 0.01), indicating that the origin labels affect consumer choice.

Table 4. Mixed Logit (ML) estimates

Notes: Single, double, and triple asterisks (*, **, ***) indicate statistical significance at the 10%, 5%, and 1% level. Standard errors in brackets.

The utility estimates infer a hierarchy of appellations. Napa is the most preferred, followed by California; their difference is statistically significant (P < 0.01), suggesting Napa is more preferred among the two. The Product of USA comes third. OCP, the omitted category, ranks fourth. The last pair is the New Jersey and Pilesgrove wines, where they are not statistically different (P = 0.86). These estimates largely fit the expected hierarchy of appellations, and our key interest is that the utility associated with OCP is significantly different from all other examined appellations.

With the negative $WT{P_{NJ}}$ (P < 0.01), consumers prefer OCP over the NJ state appellation (Table 5). The average consumer is willing to pay about $1.00 more for the OCP wines. And 25% of the consumers are willing to pay $4.00 per bottle more for the OCP wines (Table 5). Thus, the result supports the benefit of switching from state appellation to the AVA.

Table 5. WTP derived from the ML model

Notes: Single, double, and triple asterisks (*, **, ***) indicate statistical significance at the 10%, 5%, and 1% level. Standard errors are calculated with the delta method. Percentile is calculated by Krinsky and Robb simulation that account for the correlated random coefficients. Standard errors in brackets.

$WT{P_{Pilesgrove}}$ is statistically equivalent to $WT{P_{NJ}}$ (P = 0.86), but it is lower against $WT{P_{OCP}}$ (P < 0.01). Consumers are indifferent between both wines explicitly noted as from New Jersey, despite one indicates the township and another only the state. However, OCP elicited a higher WTP than Pilesgrove when both could have been identical wines. Given that the state of origin is omitted from the OCP label, the omission may have contributed to the preference of OCP wines, as previously observed.

The OCP label affects competitiveness. There are substantial differences in WTP between Napa and California wines to New Jersey wines ($5.71 and $4.53, respectively). The OCP label lowers the gaps considerably. Switching from New Jersey to OCP reduces the gaps by $1.05 on average. The effect is pronounced at the niche level; 25% of the consumers are willing to pay only around $1.50 less for the OCP wine than the wines from Napa and California. These results again support that the AVA label is competitiveness enhancing.

Finally, the average consumer prefers Product of USA to OCP, for which they are willing to pay $1.25 higher. The preference is in line with observations that the average consumer would rely more on broader and more recognizable appellation than AVA (Atkin and Johnson, Reference Atkin and Johnson2010; Atkin and Newton, Reference Atkin and Newton2012; Bruwer and Johnson, Reference Bruwer and Johnson2010; Johnson and Bruwer, Reference Johnson and Bruwer2007b). However, the preference is not generalized to the whole sample; the WTP at the 25th percentile suggests that consumers are willing to pay more ($0.42) for OCP than for the USA-labeled wine. The contrast may be due to the differences in consumers’ characteristics, where the niche consumers may make use of AVAs more as a quality indicator. A question is nevertheless raised if denoting the American origin of an AVA label could be advantageous—e.g., OCP, USA—which future studies tailored for this purpose can verify.