II. Conceptual framework

We examine how increasing the number of AVAs, which is equivalent to more narrowly defining regional designations, affects the information available to consumers. We first present this process graphically in Figure 1. We begin with a finite two-dimensional plane (or main region) containing $N$ firms. If no subregions are defined within the main region ( $AV{A_1}$), the number of regional designations is one, as illustrated in Figure 1(a). Consumers’ perception of product quality in $AV{A_1}$, defined as an aggregate measure of the quality of all member firms’ products, is not informative in making choices between products within the region. It is costly for consumers to obtain information about each firm in $AV{A_1},$ and this is the only approach available when one AVA is designated.

Figure 1. Illustration of firms in dynamic AVA space with increasing regional specificity (i.e., more~narrowly defined regions) moving from (a) through (f).

The original space may be divided, which weakly adds information about product quality within the smaller area, also increases information search costs. A single division yields two regional designations. In Figure 1(b), the two regions are the broadest region, $AV{A_1}$, and a more narrowly defined subset of the original AVA space, $AV{A_2}$. Consumers retain their existing knowledge of individual firms’ reputations for quality, but now there is additional information about these two sets of products, each sharing a common regional designation. For simplicity, we assume that each firm utilizes only one designation based on the most narrowly defined subset of which the firm is a member. In Figure 1(b), any firms located within $AV{A_2}$ utilize the $AV{A_2}$ regional designation and any firms located outside of $AV{A_2}$ utilize the $AV{A_1}$ designation. In the context of wine, $AV{A_1}$ could be the State of California and $AV{A_2}$ could be the Napa AVA. This increased specificity allows consumers to make more informed purchase decisions.

As the number of regional designations increases, hence increasing specificity by narrowing the AVA space, consumers can obtain more information. A central authority can make as many divisions as it wishes pursuant to any regulations in place, but the number of regional designations is bounded by the number of firms, $N$. Once the number of regional designations, $J$, reaches the number of firms (such that $N = J$), AVAs impart no information to consumers that is unique from information about the individual firms contained therein. Thus, the technical constraint on the number of regional designations is $N$.

In the context of wine, this is equivalent to reducing the size of regional designations to comprise only a single winery. If winery information is also known to consumers, then the AVAs provide no additional information. Figure 1(c–f) demonstrates this complete narrowing of the AVA space to $J = N$. A stronger constraint on the number of regional designations concerns the number of firms in each designation. Any regional designation containing only one firm does not impart additional information to consumers. Hence, there is a more restrictive informational constraint dictating that the number of regional designations should not exceed half the number of firms ( $J \leq N/2$). With this restriction, the average number of firms per regional designation is at least two firms. We argue that this is a reasonable restriction because if $J \gt N/2$, on average, there would be fewer than two wineries per regional designation.

Our conceptual framework modifies Costanigro et al. (Reference Costanigro, Bond and McCluskey2012), where firm and collective reputations are derived recursively, and firms maximize profits in each period, subject to those dynamic reputation constraints. For simplicity, we assume that each firm produces a fixed quantity of output per period, normalized to one. By abstracting away from a joint decision over quantity and quality, we consider an individual firm's choice of quality ${x_{i,j,t}}$, where $i$ indexes the firm, $j$ indexes the region, and $t$ indicates the year. With experience goods, consumers cannot perceive quality prior to purchase (Nelson, Reference Nelson1970), so they must rely on reputations as a proxy for quality. Each firm has a time-specific individual reputation, ${r_{i,t}}$, which is based solely on its own quality.

Similarly, each region has a time-specific collective reputation, ${R_{j,t}}$, which is an aggregation of individual members’ product quality within the region. Both reputation variables are recursively constructed to account for consumers’ priors. Following Costanigro et al. (Reference Costanigro, Bond and McCluskey2012) and for simplicity, we consider a single lagged value of reputation, along with current quality. In the general case, firm reputation can be formulated as

(1)\begin{equation}{r_{i,t}} = \alpha _1^r{r_{i,t - 1}} + \alpha _2^r\mathop \sum \limits_j {x_{i,j,t}},\end{equation}

where $\alpha _1^r \in \left[ {0,1} \right]$ and $\alpha _2^r \in \left[ {0,1} \right]$ are weights assigned to priors and current quality across all the firm's products, respectively. Similarly, regional reputation can be constructed as

(2)\begin{equation}{R_{j,t}} = \alpha _1^R{R_{j,t - 1}} + \alpha _2^R\mathop \sum \limits_i {x_{i,j,t}},\end{equation}

where $\alpha _1^R \in \left[ {0,1} \right]$ and $\alpha _2^R \in \left[ {0,1} \right]$ are weights assigned to priors and current quality across all firms in the region, respectively. We assume that the reputation variables are bounded between zero and 1, which can be guaranteed by $\alpha _1^r + \alpha _2^r = 1$, $\alpha _1^R + \alpha _2^R = 1$, and ${x_{i,j,0}} \in \left[ {0,1} \right]$. As Shapiro (Reference Shapiro1983) notes, reputation is an asset that should be treated as a dynamic state variable, and qualitative results should be similar regardless of the reputation variable's construction.

Since costs are not directly observable, we focus exclusively on the revenue side. Price is an implicit function of the reputation variables such that

(3)\begin{equation}{p_{i,j,t}} = p\left( {{r_{i,t}},\phi \left( J \right),{R_{j,t}},{x_{i,j,t}}} \right),\end{equation}

where $\phi \left( J \right)$ is a familiarity parameter that is decreasing in the total number of regional designations, $J$; i.e., $\phi '\left( J \right) \lt 0$. The familiarity parameter encapsulates the decreasing ease with which consumers are able to disentangle information about collective reputations of each region as the number of regions becomes large.

Our goal is to assess the conjecture that as the number of regional designations in the market increases beyond a threshold, firms’ marginal effect of collective reputation on price is increasing along with number of regional designations to a point. Then, after the threshold number of regional designations is exceeded, the marginal effect of collective reputation on price begins to decrease. Formally, we examine whether there is some critical number of regional designations, $\bar J$, such that ${\left. {\frac{{\partial p}}{{\partial R}}} \right|_{J \lt \bar J}} \gt 0$ and ${\left. {\frac{{\partial p}}{{\partial R}}} \right|_{J \gt \bar J}} \lt 0$.

III. Data

To examine the model and to test our hypotheses in a real-world context, the wine industry provides a practical case study. We collected unbalanced panel data on ratings, regions, prices, production, vintages, and firms from Wine Spectator for the period 1985 to 2013. The expert rating act as a proxy for product quality. Wine Spectator employs a “single-blind” methodology, which means that their experts know general information—such as vintage, appellation, and variety—but not the individual winery or the price (Shanken and Matthews, Reference Shanken and Matthews2012). We understand that these ratings are made by human experts, so there can be a subjective component. However, we assume that consumers’ quality assessments are, on average, consistent with the Wine Spectator scores. In general, these ratings are consistently found to have a highly significant relationship with prices.

Worldwide data are available, but differences in GI definitions and rules, as well as country-specific petitioning processes for GI creation, may limit meaningful analysis on such a broad scale.Footnote ² To minimize these issues and isolate a specific GI space in which to work, the analysis herein focuses on the state of Washington. This framework could be extended to examine other regions individually or a broader aggregate, such as the Pacific Northwest. Washington is presently of particular interest because it is a newer region within the industry, and several regional designations were formed during the period of study.

Though we have data on the actual locations where wines are sourced, we are primarily interested in those wines for which the firm explicitly lists an AVA on its label. A specific AVA label may be used if and only if certain thresholds are met.Footnote ³ Incorporating only stated AVAs in the analysis is appropriate since general consumers will not have access to information on a wine's geographic source beyond what is available on the label or in published reviews.Footnote ⁴

Prices are those quoted directly from the firms rather than the secondary retail market. For uniformity, all bottles of unusual volume are excluded from the analysis. We begin with 9,601 rating observations, but of those only 9,243 contain the requisite information to warrant inclusion in the estimation procedure. We also exclude 120 non-vintage bottles that cannot be used to construct age and recursive reputation variables. A summary of the data grouped by vintage is available in Table 1, and summary statistics for key variables are presented in Table 2.

Table 1. Summary of unbalanced panel data with observations arranged by year

Table 2. Summary statistics of key variables

Individual firm-level data were collected manually from firm websites. Measurement is annual, and we use end-of-year AVA totals to account for issues of implementation and product-release dates. Wine prices are adjusted to 1982–1984 values by a consumer price index for alcohol. We build explicit recursive variables for firm and collective reputations below using expert-ratings data that are available to consumers.

IV. Empirical approach

Following Rosen (Reference Rosen1974), we assume that wine prices are implicitly determined through product attributes. Quality is assessed after production and cannot be changed ex post. We include proxy observations of quality, so we use price as the dependent variable, which is implicitly determined by perceived quality attributes. Under the hedonic-price approach of Rosen (Reference Rosen1974), price can be decomposed into its primary elements; in this case, the most basic elements are firm reputation, AVA reputation, quality, and number of AVAs. The analysis begins with a general price equation for each bottle of wine:

(4)\begin{equation}p = p\left( {r,R,x,J} \right),\end{equation}

where $p$ is price, $x$ represents a bottle's quality score, $r$ is firm reputation, $R$ is AVA reputation, and $J$ is the total number of AVAs.Footnote ⁵ This model allows for the calculation of marginal willingness to pay for specific product attributes; for example, $\partial p/\partial R$ indicates the effect on price of an increase in collective reputation. We consider a semi-log modelFootnote ⁶ such that

(5)\begin{equation}\log \left( {{p_{i,j,t}}} \right) = {\beta _0} + {\beta _1}{J_t} + {\beta _2}J_t^2 + {\beta _3}{r_{i,t}} + {\beta _4}{R_{j,t}} + {\beta _5}{x_{i,j,t}} + {\epsilon_{i,j,t}},\end{equation}

where $i$ indexes the firm, $j$ indexes the AVA, and $t$ indexes the vintage, and ${\epsilon_{i,j,t}}$ is the error term. We include the number of AVAs ( $J$) and the squared term ( ${J^2}$); therefore we can explicitly solve for the optimal number of AVAs for Washington, a key result of the study. Note that we limit this initial analysis to AVA, quality and reputation variable to focus on the impacts of the number of AVAs on prices.

We employ the narrowest AVA used in labeling. Prior to performing any regression analysis using this model, the variable proxies must be explicitly defined and constructed. The goal is to indicate heterogeneous, ordinal values associated with reputations that are not imparted simply with region indicator and age variables. To estimate the value of a collective reputation, one must disentangle its effects from the individual firm reputation effects on potential revenue.

Since quality is not directly observable, we use wine ratings as a proxy for quality. Reputations are dynamic, so firm and AVA reputations are calculated recursively, as modified from equation (1) for simplicity of programming. Firm $i$’s reputation is

(6)\begin{equation}{ \tilde r_{i,t}} = \frac{1}{{1 + {\rho ^r}}}\left[ {{\rho ^r}{{\tilde r}_{i,t - 1}} + \left( {J_{i,t}^{ - 1}\mathop \sum \limits_{j = 1}^{{J_{i,t}}} {x_{i,j,t}}} \right){\textrm{ }}} \right],\end{equation}

where $\left\{ {i,j,t} \right\}$ references a combination of firm $i$, AVA $j$, and vintage $t$ (i.e., $\left\{ {i,j,t} \right\}$ represents a bottle index), ${J_{i,t}}$ denotes the number of AVAs in which firm $i$ produces a bottle of vintage $t$, and ${\rho ^r}$ is a weighting parameter to account for the relative emphasis of consumer priors and current quality aggregates. AVA reputations are calculated similarly as

(7)\begin{equation}{\tilde R_{i,t}} = \frac{1}{{1 + {\rho ^R}}}\left[ {{\rho ^R}{{\tilde R}_{i,t - 1}} + \left( {I_{j,t}^{ - 1}\mathop \sum \limits_{i = 1}^{{I_{j,t}}} {x_{i,j,t}}} \right){\textrm{ }}} \right],\end{equation}

where ${\rho ^R}$ is a weighting parameter similar to that used in the firm reputation calculation and ${I_{j,t}}$ is the number of firms in AVA $j$ for vintage $t$. For both reputation variables, the initial value is mean quality in the first period of inclusion—for initial firm reputation, it is a mean across firm production in a firm's initial vintage, and for initial AVA reputation it is a mean across all production for that AVA's initial vintage. We begin the analysis with ${\rho ^r} = 1$ and ${\rho ^R} = 1$, in other words assuming that priors and current quality are equally weighted in reputation construction.Footnote ⁷

Since our primary research objective is to understand to examine how new division of the AVA space impacts prices over time, we also subset the data by date range based on the introduction of new AVAs. In these estimations, we include additional control variables in the hedonic regression. We do not include the number of AVAs since by design there is no variation during the time period. To this baseline model (equation 5), we add other controls ( ${Z_{i,j,t}}$, ${Z_{i,j}},$ and ${Z_t}$). Specifically, we estimate the following for our data that is partitioned by dates when new AVAs are added:

(8)\begin{equation}\log \left( {{p_{i,j,t}}} \right) = {\beta _0} + {\beta _1}{x_{i,j,t}} + {\beta _2}{\tilde r_{i,t}} + {\beta _3}{\tilde R_{j,t}} + + \mathop {\mathop \sum \nolimits}\limits^H_{h = 1} {\beta _{h + 3}}{Z_{h,i,j,t}} + \mathop {\mathop \sum \nolimits}\limits^{T - 1}_{t = 1} {\beta _{t + H + 3}}{Z_t} + {\epsilon_{i,j,t}},\end{equation}

where ${Z_{h,i,j,t}}$ is a set of $H$ controls for red wine, interactions between red wine and other variables, scarcity (defined as the inverse of cases produced at time t), age, “estate” and “reserve” indicators, an indicator for whether Washington is listed explicitly on the bottle, and vintage fixed effects ${Z_t}$. We also include controls for diversification within firms (i.e., the number of reviewed bottles by firm and vintage) and the annual share of Washington in the set of Wine Spectator reviews as a proxy for interest in the state's wine industry.

With these regressions, we are interested in whether the coefficient estimates on the firm- and AVA-reputation variables change significantly across these subset time ranges. As the number of AVAs increases, we expect the impact of AVA reputation to be weaker. The rationale is that with a large number of AVAs, consumers will be less familiar with each one, and the AVA reputation will matter less in terms of price. Furthermore, as the number of AVAs increases, we expect that firm reputation will be more important in its impact on price.

Finally, in order to understand how the creation of new AVAs is related to prices across price quantiles, we estimate the relationship between price and the reputation variables using quantile regression (Koenker and Bassett, Reference Koenker and Bassett1978). In order to achieve stable estimation in the quantile tails (Davino et al., Reference Davino, Furno and Vistocco2013), we estimate the regression at every fifth quantile.