II. Related literature

Our study contributes empirical evidence on knowledge spillovers in winemaking, but it also contributes to a broader literature on how knowledge spillovers contribute to localization economies. It most directly adds to the handful of studies that have documented knowledge transmission within winery clusters using detailed survey data. Prominent among those studies are Giuliani and Bell (Reference Giuliani and Bell2005), Morrison and Rabellotti (Reference Morrison and Rabellotti2005), Giuliani (Reference Giuliani2007), Farias and Tatsch (Reference Farias and Tatsch2014), Maghssudipour, Lazzeretti, and Capone (Reference Maghssudipour, Lazzeretti and Capone2020), and Choi and Gu (Reference Choi and Gu2020).

Giuliani and Bell (Reference Giuliani and Bell2005) surveyed 32 firms in Chile and analyzed the characteristics of the social network between them, emphasizing formally acknowledged knowledge-sharing ties. They also highlighted the role of “flying winemakers,” consultants who work for several wineries and are not bound by location. Similarly, Morrison and Rabellotti (Reference Morrison and Rabellotti2005) surveyed 26 wineries in an Italian cluster and mapped social proximity between the firms, finding that some firms practice a mutual exchange of knowledge while others are isolated. Giuliani (Reference Giuliani2007) analyzed survey data from 105 wineries in 3 clusters (2 in Italy, 1 in Chile), and mapped the self-reported linkages between firms in terms of both winemaking and business knowledge. She found that firms with more human capital tend to be more linked, but only with firms with a similar knowledge base due to the need for reciprocal knowledge sharing. Farias and Tatsch (Reference Farias and Tatsch2014) conducted a similar style of study as the previously mentioned of 20 wineries in a Brazilian winery cluster and found similar results.

Recent studies by Maghssudipour, Lazzeretti, and Capone (Reference Maghssudipour, Lazzeretti and Capone2020) and Choi and Gu (Reference Choi and Gu2020) focus on the multiplicity of knowledge networks that wineries are connected to. Maghssudipour, Lazzeretti, and Capone distinguished between economic and social ties in studying the Montefalco cluster in Italy, finding that economic ties are more important to knowledge transfer. Choi and Gu focused on the linkages between universities, government agencies, and private companies in innovation in China's nascent wine industry. They find that winemaking knowledge first clustered around China's coastal cities, then spread inland during the period 2007–2016.

Another impetus for starting a winery in an established wine region is to take advantage of the collective reputation that the region's wineries enjoy. Collective reputation has been previously studied by numerous authors, see Tirole (Reference Tirole1996) for a seminal theoretical treatment of it, and Costanigro, Bond, and McCluskey (Reference Costanigro, Bond and McCluskey2012) on the simultaneous effects of collective and private reputation. Yang, McCluskey, and Brady (Reference Yang, McCluskey and Brady2012) found evidence that clustered wineries are positively correlated in terms of prices and wine ratings, and they used hedonic spatial lag models to estimate the effect of proximity, weighted by quality, on wine prices. This provides detailed microeconomic evidence that proximity to high-quality wineries is a determinant of wine quality.

The Marshallian externality of intra-industry knowledge spillovers includes all knowledge transfer, not just those through formal channels, and studies that rely on survey data may miss the more indirect channels of knowledge transfer. In the wine industry, other channels include churn in the market for skilled labor (experienced winemakers leaving one firm for another, or starting their own), unofficial personal contacts, and trade conferences. The empirical method applied in our paper captures knowledge spillovers indirectly, inferring that the effect of proximity to other wineries must be due to knowledge transfer, once other factors are controlled for. A downside to indirect methods such as ours is that they may conflate knowledge spillovers with unobserved externalities driving location, a point made by Breschi and Lissoni (Reference Breschi and Lissoni2001). While we try to control for other factors driving localization, we acknowledge that there is no perfect way to do so.

As mentioned earlier, the idea that firms cluster near other firms in the same industry due to the “Marshallian externalities” of proximity to intermediate input suppliers, local labor pool, and intra-industry technological spillovers was first described by Marshall (Reference Marshall1920), later reemphasized by Krugman (Reference Krugman1991). Arrow (Reference Arrow and Nelson1962) outlined the economics of innovation, noting that an inherent disincentive for R&D activity is the potential for knowledge to spill to rival firms, especially through labor-market churn. That such intra-industry technological spillovers generate increasing returns to scale at the industry level later motivating the development of models of endogenous economic growth (e.g., Romer Reference Romer1986, Reference Romer1990).

Rosenthal and Strange (Reference Rosenthal, Strange, Henderson and Thisse2004) have a detailed discussion of the literature regarding empirical studies of the relative importance of the Marshallian external factors for industry concentration. Our study adds to this literature. Jaffe (Reference Jaffe1989) specified university and private R&D as inputs into state-level knowledge production functions. Jaffe, Trajtenberg, and Henderson (Reference Jaffe, Trajtenberg and Henderson1993) found that patents are more likely to be cited by other patents within the same state and metropolitan area. Audretsch and Feldman (Reference Audretsch and Feldman1996) found that the more knowledge-oriented an industry, the more geographically clustered its innovative activity, which is indicative of the presence of local knowledge spillovers. Rigby and Essletzbichler (Reference Rigby and Essletzbichler2002) examined national plant-level data and created direct measures of the external factors, finding positive effects of the concentration of supply linkages, productivity growth of suppliers (a proxy for knowledge), and labor-pool mix. Holmes (Reference Holmes1999) looks at a single industry (textiles) and finds that firms located in areas with high concentrations of other firms in the industry are more vertically disintegrated, that is, they purchase inputs from supplying firms more often than they vertically integrate the supply. Evidence shows that proximity to inputs is an important positive localization externality. More recently, Bloom, Schankerman, and Van Reenen (Reference Bloom, Schankerman and Van Reenen2013) argued that technology spillovers are offset by the effects of competitor firms carrying out R&D and taking market share from product market rivals. They estimated the technology spillover effect to be more than twice as high as the market share effect at the aggregate level. Another detailed empirical analysis was done by Greenstone, Hornbeck, and Moretti (Reference Greenstone, Hornbeck and Moretti2010), who found a large productivity effect on incumbent firms from the quasi-random placement of new plants nearby.

III. Empirical methodology

We test the relationship between the number of winemaking establishments (wineries) per zip code and measures of the strength of externalities operating in that location. To do so, we constructed a panel dataset at the zip code year level, covering 598 zip codes in Washington State and the years from 1994 to 2014. Our zip code coverage was limited to those for which we had access to data on the distances between them, and our panel only includes 83% of the 719 zip codes within Washington State. However, only 179 of the zip codes ever had one or more wineries, and almost all of these are included in the 598.

Analysis at the zip code level allows us to account for the effects of proximity at a highly granular level. The size of zip codes varies with population density, but for Washington, most zip codes are less than ten miles wide. This helps to limit the measurement error in our study: the variation in distances between zip codes (measured) is much larger than the variation in distances within zip codes (unmeasured and assumed to be zero).

Each proximity variable is constructed in the following manner:

$$index_{it} = \mathop \sum \limits_{\,j = 1}^J \displaystyle{{x_{\,jt}} \over {d_{ij}}}, \;$$

where i indicates zip code, t indicates year, x indicates a variable that is dispersed over a geographic unit j, and d is the distance between j and zip code i.

Knowledge spillovers in this context refer to the use of winemaking knowledge by other firms than those that generated it. It is usually the case that each winery has a team of winemakers. Presumably, each winemaker has idiosyncratic knowledge regarding the winemaking craft. Therefore, it is reasonable to presume that the quantity of winemaking knowledge within a zip code is an increasing function of the number of wineries near that zip code in a region. Our distance-weighted measure of winemaking knowledge is therefore:

$$WI_{it} = \mathop \sum \limits_{k = 1}^K \displaystyle{{w_{kt}} \over {d_{ik}}}, \;$$

where w _kt is the number of wineries in zip code k ≠ i in year t. This index applies distance weights linearly.

Locations close to vineyards provide a natural advantage for wineries, which can better monitor and protect the quality of grapes, as well as save on transport costs. We construct an index of proximity to vineyards in the form:

$$VI_{it} = \mathop \sum \limits_{\,j = 1}^J \displaystyle{{vines_{\,jt}} \over {d_{ij}}}$$

Vines_jt is the number of grapevines cultivated in AVA j in year t.Footnote ⁴ The factor d _ij is the Euclidean distance d between zip code i and AVA j. Ideally, this measure would be constructed using distances from each winery to each acre of wine grapes. As constructed here, however, we use the centroid distances from the approximate center of each zip code to the center of each AVA. Since Washington's AVA vary in size and shape, this measure has an upward measurement error for some zip codes and a downward measurement error for others.

Another factor we consider is the distance to customers, reflecting the influence of transportation costs. We constructed an income index using the income-weighted population sizes of counties in Washington.

$$II_{it} = \mathop \sum \limits_{c = 1}^C \displaystyle{{inc_{ct}} \over {d_{ic}}}$$

Here, inc indicates aggregate household income in county c and year t. Distances are again measured using centroid distances, but this time between zip codes and counties.

To measure the depth of winemaking knowledge based on winemaking history in an area, we define the variable Yearsprior _i as the number of years that early Washington wineries had collectively operated in zip code i prior to 1994. We define early Washington wineries as those in operation prior to 1985 that were still in business in 2017. In our models, we include the interaction between WI _it and Yearsprior _i, to allow the effect of proximity to wineries to depend on winemaking history.

As mentioned previously, a reputation for excellence among the wineries in a particular location may also attract new wineries. That is, critically acclaimed wineries in an area may raise the collective reputation of a winemaking region and attract more winemakers. This collective reputation effect is distinct from knowledge spillovers in that even wineries with no interaction with other winemakers can free-ride on their common appellation label. To examine the effect of the reputation of wineries on the growth of wineries, we focus on star wineries that have been highly critically acclaimed. We categorize a winery as a “best Washington winery” if it has produced at least one wine that is listed in Wine Searcher's “Best Washington Wines” list.Footnote ⁵ Details of the construction of this variable are given in the data section, but the definition of the variable is given as:

$$bestwa_{it}\equiv number \;of\; best\; wineries\; located\; in\; zip\; code\; i \;by\; year\; t$$

Our empirical approach starts with regression models estimating the effects of the variables on the change in the number of wineries per zip code. This model takes the form:

(1)$$\Delta w_{it} = \beta _0 + \beta _1VI_{it} + \beta _2II_{it} + \beta _3WI_{it} + \beta _4WI_{it}\ast YP_i + \beta _5bestwa_{it} + {\boldsymbol \alpha }_t + {\tf="P46114X" \char"27}_{it}, \;$$

where w is the number of wineries, VI, II, and WI are the proximity indexes for vineyards, income, and other wineries, YP is the history variable yearsprior, and alpha is a set of year indicators.

We also estimate the model in levels to examine the co-trending behavior of the number of wineries and the explanatory variables. Due to the high number of zip code year observations with zero wineries, we turn to negative binomial regression models. Negative binomial models are suitable for this data because the dependent variable is a discrete count with many zero-valued observations.

In general, negative binomial models are specified by

(2)$$f( {w_{it}{\rm \vert }{\boldsymbol x}_{{\boldsymbol it}}, \;u_{it}} ) = \displaystyle{{e^{-\lambda _{it}u_{it}}{( {\lambda_{it}u_{it}} ) }^{w_{it}}} \over {w_{it}!}}, \;$$

where x_it is a vector that includes all of the right-hand side variables in Equation (1) and

(3)$${\boldsymbol x}_{{\boldsymbol it}}{\boldsymbol ^{\prime}\beta } + {\tf="P46114X" \char"27}_{it} = \ln \lambda _{it} + \ln u_{it}.$$

Here, the parameter λ _it is the expected number of wineries per zip code and year, and u _it is a random effect that is usually assumed to have a gamma distribution.

IV. Data

Our primary source of data on the number and size of wineries is the County Business Patterns (CBP) survey, a publicly available dataset from the U.S. Census Department. The CBP contains data on the number of wine-producing establishments (plants) at the zip code level for the years 1994–2014. It also contains the number of wine establishments in each of many employment size categories, from which we can construct a rough measure of winemaking employment per zip code.

To construct our Yearsprior variable, we gathered information on early Washington wineries from winery licenses issued by the Washington State Liquor and Cannabis Board, and trade publications. To our knowledge, there are only 23 wineries that meet our criteria of having been established prior to 1985 and surviving to the present. The relative youth of the Washington wine industry means that early in its history, winemaking knowledge was scarce and sparsely distributed. According to our data, among Washington zip codes that have ever had a winery, the mean Yearsprior is 4.6. However, two zip codes have more than ten times that: 98072, Woodinville, and 99350, Prosser.Footnote ⁶

For the variable $vines\_index$, our index of proximity to grapes, we used data from the USDA National Agricultural Statistics Service (2011). This report provided data on the quantity of vines in cultivation in each AVA for the years 1991, 1995, 2000, 2005, 2007, 2009, and 2010. For years between those dates, we estimated the quantity using linear interpolation. For the four years of our panel post-2010, we assumed no growth, and so used the 2010 values throughout. This assumption is reasonable because growth in the number of vines in cultivation dropped off after 2007 with the onset of the Global Financial Crisis. The distance data were constructed using a GIS to calculate distances between the approximate center of each zip code and AVA. Similarly, the variable inc_index_it was constructed using data on aggregate personal income per county from the Bureau of Economic Analysis and weighted with distances between zip codes and counties. It was deflated and converted to real income using the consumer price index for all urban consumers (CPI-U) from the Bureau of Labor Statistics. For the full sample of 12,558 observations for which there was data on the vineyard proximity index, Table 1 contains a summary of all the data used in this analysis.

Table 1. Descriptive statistics (Sample size: N = 12,558)

To construct the bestwa (best Washington wineries) variable, we turned to Wine Searcher's list of “Best Washington Wines.” For each wine, Wine Searcher collected critic scores from dozens of leading publications. While there is a lot of noise in individual critics’ scores (as shown in Stuen, Miller, and Stone (Reference Stuen, Miller and Stone2015)), the averages provided by Wine Searcher are much more reliable. Wine Searcher's Best Washington Wines list includes 25 wines that average 93 or above on the common 100-point rating scale. The determination of the “best” wines and wineries is, like judging art, inherently subjective, and we know that many excellent wineries are left off this list. In defining “best” too narrowly, we risk limiting the explanatory power of the variable, but we do not find that to be the case in the empirical results.

Those 25 wines are produced by just 11 wineries, operating in just 5 zip codes. These wineries and their zip codes of operation are listed in Table 2.Footnote ⁷

Table 2. Best Washington wineries (based on average critic ratings)

Note: ^aA collaboration between the Chateau Ste. Michelle winery and winemaker Ernst Loosen, under the Eroica brand label.

Two of these are located in the Woodinville cluster (which includes nearby Snohomish), while seven are in Walla Walla. Three are co-categorized as “early wineries,” having begun operations prior to 1985 (these being Quilceda, Leonetti, and Chateau Ste. Michelle).

VI. Concluding remarks

This study examines the spatial distribution of winemaking establishments, using anonymized data on the number of wineries in Washington State. It examines several critical factors: proximity to the primary raw material (grapes), proximity to consumer income, proximity to other wineries, winemaking history, and the presence of elite wineries. We find evidence that the effect of proximity to other wineries hinges on the local history of winemaking. For zip codes with a substantial history of winemaking (at least 11 winery years prior to 1994), the elasticity of the number of wineries with respect to proximity to other wineries in the region is at least 0.44. On the other hand, we find no such proximity effect for zip codes with no history of winemaking.

The finding that winery clustering depends on the length of winemaking history in a location has interesting policy implications. It suggests that winery clustering is not just the result of pre-existing geographic features such as the distribution of land suitable for viticulture and the distribution of the population. Historical accidents, such as chance meetings between winemakers, the affinity of experienced winemakers for particular locations, or the availability of other winemaking inputs (equipment adapted from other industries, technical education programs), may seed new clusters. Governments may therefore be able to encourage cluster formation through subsidies for new firms and education. However, efforts to encourage cluster formation in already-established wine regions, outside of existing clusters, face a substantial hurdle, as most economic incentives favor locating in existing clusters.

We also find evidence that the presence of star wineries has an effect on growth in the number of wineries. This effect could arguably go both ways, as elite winemakers could be attracted to growing winery clusters. We leave it to the reader to decide whether elite wineries are more of a cause or effect of the growth of a winery cluster.