2.1. DTC agricultural production and online marketing

DTC agricultural sales traditionally occurred via farmers markets, on-farm stores, roadside stands, and community supported agriculture (CSA) programs. In 2017 for the first time, the Census of Agriculture’s report form listed online marketplaces as a DTC market channel. Online sales accounted for 6% of aggregate DTC sales made by U.S. farms in 2015 (USDA NASS, 2016).

Fifty-four percent of DTC farms are in metropolitan counties (Table 1). O’Hara and Lin (Reference O’Hara and Lin2019) found that greater population levels within 50 miles of a local food farm increase the probability that it undertakes DTC sales and, conditional on entrance, their level of DTC sales. A challenge with establishing farmers markets in low population density areas is that vendors prefer attending fewer, but larger markets, to reduce their average marketing costs (Schmit and Gomez, Reference Schmit and Gomez2011). Also, incomes may be higher in urban cities than in rural areas, and DTC production near cities is income elastic (O’Hara and Low, Reference O’Hara and Low2016). In addition to DTC sales, population density also facilitates direct sales of food by farms to retailers like supermarkets and restaurants (O’Hara and Lin, Reference O’Hara and Lin2019).

Table 1.

Descriptive statistics of DTC farms

Note: We calculated the statistics from 4,664 farms in LFMPS that participated in DTC markets.

We eliminated a select number of farms for which county-level internet statistics were not available.

Source: We use Census Bureau definitions from Tables 1, 2, and 3 when providing regional classifications.

DTC sales in the United States increased throughout the 1990s and 2000s in response to a greater consumer interest in purchasing local foods (O’Hara and Low, Reference O’Hara and Low2016). However, DTC sales have stagnated within the past decade (Low et al., Reference Low, Adalja, Beaulieu, Key, Martinez, Melton, Perez, Ralston, Stewart, Suttles, Vogel and Jablonski2015). The USDA NASS estimated that between 2015 and 2017, DTC sales declined by 10% from $3.1 billion (measured in 2017 USD) to $2.8 billion (USDA NASS, 2016, 2019).

Transaction costs incurred by producers and consumers are an impediment to in-person DTC agricultural venues. The marketing and distribution costs that farms incur for local sales can range between 13 and 62% of their revenue (King et al., Reference King, Hand, DiGiacomo, Clancy, Gomez, Hardesty, Lev and McLaughlin2010). Impediments that consumers experience with in-person DTC transactions include: operational hours of a neighborhood farmers’ market may present scheduling conflicts or occur during inclement weather, local farms may be too far away for consumers to visit on a consistent basis, and CSA farms may deliver products that are incompatible with what consumers want or are able to prepare that week. Efforts to reduce the transaction costs for consumers can exacerbate the costs incurred by farmers, which, in turn, could have contributed to the recent plateau in DTC sales. For instance, more farmers markets within an area can, in some instances, fragment the customer base of existing ones (Lohr et al., Reference Lohr, Diamond, Dicken and Marquardt2011). If so, DTC farmers must attend more markets, and, thus, incur greater costs, to maintain their existing level of in-person sales.

A more fundamental challenge confronting DTC farms is that consumers now have more opportunities to make online grocery purchases (OneSpace, 2018), and, thus, can more readily buy local products without leaving their home. The percentage of groceries that consumers purchase online is increasing annually, and industry reports are predicting that this trend will continue (CNBC, 2017; OneSpace, 2018). Similarly, online businesses that deliver meal kits (e.g., Blue Apron) and imperfect fruits and vegetables (e.g., Imperfect Produce) are, at least anecdotally, adversely impacting CSA programs (Huntley, Reference Huntley2018; TNFE, 2018; Wolnik, Reference Wolnik2019). Online delivery businesses compete with DTC agricultural producers since online produce buyers are younger, more educated, more likely to live in urban areas, and have greater local food expenditure levels than other market participants (Gumirakiza and VanZee, Reference Gumirakiza and VanZee2017).

E-commerce platforms could offer a way for DTC farms to compete for consumers that are increasing online grocery purchases. Online DTC sales can reduce the transportation costs of in-person DTC markets since, if the goods are shipped, neither producers nor consumers need to travel or coordinate schedules to fulfill the transaction. Since travel is not required, online sales are not geographically constrained. They can also reduce search costs for consumers. As an example, a consumer who wanted to subscribe to a CSA program while customizing their order could easily compare alternatives online.

At the same time, there are costs associated with online marketplaces. Customer acquisition costs may be higher since customers need to at least be aware of the website, and perhaps want detailed information about the farm. Consequently, some in-person interactions between farmers and consumers may be needed before online transactions occur. Lu and Reardon (Reference Lu and Reardon2018) show that once consumers are familiar with a food product through tactile observation, they may be comfortable making subsequent online purchases.

2.2. Farm technology and internet use

The rural household broadband adoption rate changed from 46 to 62% between 2009 and 2013, which was higher than the corresponding increase of 67 to 70% in urban households (Barnes and Coatney, Reference Barnes and Coatney2015). The percentage of farms with internet access increased from 57% in 2007 to 70% in 2012, and further increased to 75% in 2017 (USDA NASS, 2019; Whitacre, Mark, and Griffin, Reference Whitacre, Mark and Griffin2014). Broadband availability has increased farm sales and attracted new firms to rural areas (Kandilov et al., Reference Kandilov, Kandilov, Liu and Renkow2017; Kim and Orazem, Reference Kim and Orazem2017). However, despite improvements, high-speed internet deficiencies remain in rural areas (FCC, 2018). For instance, hosting an online marketplace requires high-speed internet access and not low-speed service. So, policymakers remain focused on expanding high-speed internet access in rural areas.

There is an extensive literature on agricultural technology adoption (e.g., Sunding and Zilberman, Reference Sunding, Zilberman, Gardner and Rausser2001). A farm’s decision to sell products online is based, in part, on time-management factors. For instance, working off the farm is positively related to farm technology adoption and human capital levels (Fernandez-Cornejo et al., Reference Fernandez-Cornejo, Mishra, Nehring, Hendricks, Southern and Gregory2007). In reviewing the literature, Briggeman and Whitacre (Reference Briggeman and Whitacre2010) show that farm internet adoption can be associated with the farmer’s age, education, farm size, regional location, whether they work off the farm, gender, and whether they have diversified operations. Internet access increases the probability that a farm makes DTC food sales (Low and Vogel, Reference Low and Vogel2011; Park, Mishra, and Wozniak, Reference Park, Mishra and Wozniak2014). Also, DTC farms with internet access have a greater level of sales (Detre et al., Reference Detre, Mark, Mishra and Adhikari2011; Low and Vogel, Reference Low and Vogel2011; Uematsu and Mishra, Reference Uematsu and Mishra2011).

Agricultural Resource Management Survey (ARMS) data has been used extensively in the farm-level internet-use literature. The ARMS has focused on whether farms have the internet, but not on online marketing (Stenberg et al., Reference Stenberg, Morehart, Vogel, Cromartie, Breneman and Brown2009). The ARMS data in these studies are from the mid-2000s; subsequent versions of the ARMS have de-emphasized questions about internet use. However, there has been considerable innovation in ways that the internet could be valuable to farms since that period. DTC farms can use smartphones to process payments, use social media to advertise their products, or partner with software developers to create smartphone apps that allow customers to search for local farms. For instance, the “WhatsGood” app allows customers in the Northeast to order foods directly from local farmers and have it delivered to their home or workplace.

To summarize, earlier studies have investigated farm internet adoption. However, there is a dearth of research examining the characteristics of farms that market food online to consumers directly. We contribute to this literature by using a unique national-level dataset to examine which farms use online marketplaces for DTC food sales.

3.1. LFMPS

In 2016, for the first time, the USDA’s National Agricultural Statistics Service (USDA NASS) administered the LFMPS (to collect 2015 data).Footnote ¹ In contrast to the ARMS, which annually solicits information about farm production practices and profitability, the LFMPS is the first USDA dataset with detailed data about specific market channels used by “local” food farms. In the LFMPS, “local” food farms are those that make direct sales of food to consumers, retailers, institutions, and/or intermediaries that market locally-branded food products. The USDA NASS used the phrase “local” even though online sales are not geographically constrained, and could include transactions between farmers and consumers located anywhere in the country. In our article, for consistency we refer to LFMPS respondents as “local” food farms. The data do not allow for comparisons between local food and non-local food farms since farms not selling local food are not surveyed in the LFMPS.

The USDA NASS administered the LFMPS in all 50 states to 44,272 farms that might have sold food through local market channels. NASS identified the farms through their pre-existing sampling frame and the internet via “web scraping.” The resulting cross-sectional dataset consists of 5,697 observations, which corresponds to a response rate of 13%. The USDA NASS obtained 42% of responses by mail, 39% by phone, 13% via face-to-face interviews, and 6% by web reporting. USDA NASS stratified the LFMPS sample by farm size, state, and marketing channels. Farms with greater sales levels had a greater probability of being sampled than smaller-sized farms. Since the resulting responses did not represent a random sample of local food farmers, the USDA NASS developed population weights to derive nationally representative totals from the data.

The LFMPS solicited sales data among various DTC channels, including online marketing, as well as the year that the farm commenced DTC marketing. The LFMPS defined an online marketplace as a “web-based platform designed for the selling of goods.” Farms with online sales were further asked the percentage of online sales that were sold to consumers within either the same state or a 400-mile radius of the farm. The LFMPS also contained background questions about whether farmers used the internet to access resources in the following ways: purchase input supplies/equipment, access peer-learning resources, access business products/services, identify funding opportunities, and obtain price/market information.

The LFMPS had binary questions about the agricultural commodities produced for local sale. Pertinent commodity classifications for DTC marketing included cattle/calves, vegetables/melons/potatoes, fruit/tree nuts/berries, poultry/eggs, and sheep/goats. The LFMPS did not ask which specific market channels the commodities are sold through or their respective sales levels. However, farms reported their aggregate sales level by market channels for unprocessed commodities and, for vertically integrated operations, value-added products. In the LFMPS, examples of value-added products included bottled milk, cheese, eggs in small cartons, meat, wine, and jam. Another distinguishing attribute of the LFMPS is that it was the first national-level survey by the USDA to collect sales data on value-added agricultural products.

3.2. Empirical model

We hypothesize that online marketplace use by a farm depends on its own characteristics and those of nearby customers. So, we test whether the probability that a DTC farm has an online marketplace that is influenced by the metropolitan classification of its county; socioeconomic attributes of its proximate customer base (i.e., income, education, and connectivity levels); and attributes of the farm operator, such as age and experience level.

We estimate a baseline regression with these variables that we refer to as Model 1. We then explore alternate specifications. In Model 2, we include interaction terms of the farm’s experience level and its metropolitan county classification. In Model 3, we control for variables that capture characteristics of the farm that could be important, but could also be simultaneous. In Model 4, we estimate our model on a subset of the data.

More specifically, we estimate a logit regression specified in equation (1):

(1)$$P(y = 1 | {x} ) = {{\exp \, ( {{\bf{x}}\beta })} \over {\rm{1} + \exp \, ( {{\bf{x}}\beta } )}}$$

The binary dependent variable in equation (1), y, is equal to one if a farm with DTC sales operates an online marketplace and is equal to zero for DTC farms without online marketplaces. In equation (1), we denote the independent variables by x and their corresponding coefficients by β. To correct for endogenous sampling, we estimate weighted regressions using population weights created by the USDA NASS.

In Model 1, we include two binary explanatory variables that control for whether the farm is in: (a) a metropolitan county, and (b) a non-metropolitan county that is adjacent to a metropolitan county (USDA ERS, 2016). The omitted variable, thus, is whether the farm is in a non-metropolitan non-adjacent county (15% of our sample). To improve the readability of this article, we refer to farms in non-metropolitan adjacent counties as “adjacent” farms and farms in non-metropolitan non-adjacent counties as “distant rural” farms. We include these variables because, as we described earlier, distant rural farms may be more likely to have an online marketplace to circumvent their greater transportation costs with accessing population centers. At the same time, metropolitan farms may be more likely to have online marketplaces if they have more in-person DTC marketing opportunities and can leverage in-person customer interactions to increase the visibility of their online platform.

We include an independent variable that represents the county-level ratio of residential high-speed internet connections (HSIC) to the total number of households (FCC, 2017a).Footnote ² This variable is a proxy for local internet adoption. We include this variable because farms in counties with greater internet usage among potential local clients may be more likely to have an online marketplace. We assume that county-level internet subscribership is exogenous regarding the farm-level decision to have an online marketplace, given the relatively modest size of the local foods sector.

We include two other county-level control variables that could be correlated with the propensity of customers to purchase foods online: per capita income and the percentage of the population with a bachelor’s degree. Income could influence the propensity of customers to make online purchases, as higher-income customers may have greater opportunity costs of time associated with non-internet shopping. The percentage of the population with a bachelor’s degree has been used as a control variable in other local food studies (e.g., Brown et al., Reference Brown, Goetz, Ahearn and Liang2014; O’Hara and Benson, Reference O’Hara and Benson2019). In our context, education could be associated with internet shopping if those with greater levels of formal education have received greater training with using the internet.

We control for the average age of operators on the farm and the square of average age. Previous studies have found that age has a quadratic impact on internet use (Mishra, Williams, and Detre, Reference Mishra, Williams and Detre2009; Briggeman and Whitacre, Reference Briggeman and Whitacre2010). So, these findings suggest older farmers are less likely to use the internet, perhaps since they are less accustomed to this technology. Since making sales online is a way that farmers use the internet, we expect the effect of age on online marketing to be like its impact on internet use.

We also include a binary variable that is equal to one if the farm operation began selling in DTC markets in 2009 or later. Hereafter for brevity, we refer to such farms as “post-2009” farms and similarly refer to farms that began DTC marketing before 2009 as “pre-2009” farms. Some internet-use studies have used age and experience variables as proxies for each other (Mishra, Williams, and Detre, Reference Mishra, Williams and Detre2009). We distinguish between the influences of age and experience on online marketing since we control for both attributes. To do so, we model experience as a discrete variable since the positive correlation between the two variables may otherwise make interpreting their coefficients challenging. Thus, the age variable coefficients are conditional on whether their farm began DTC marketing prior to 2009 or not. Nonetheless, the age and experience variables are still correlated with each other.

We select 2009 as the threshold year for entering DTC marketing because it corresponds to the year in which the American Recovery and Reinvestment Act (ARRA) broadband internet resources were first deployed. The ARRA is relevant because it provided an unprecedented amount of money ($7.2 billion) for broadband grant programs (Barnes and Coatney, Reference Barnes and Coatney2015). A positive coefficient on this variable could imply that post-2009 DTC farms are more likely to perceive online marketing as an effective way to reach consumers than pre-2009 DTC farms. This could occur if post-2009 farms have greater skill levels with online software and advertising techniques. In contrast, a negative coefficient could suggest that pre-2009 DTC farms are using online marketplaces as another market channel for pre-existing consumers. This latter explanation is plausible if pre-2009 farms have a loyal following among consumers or high brand awareness, and if post-2009 farms have challenges with advertising their online marketplace to consumers that are unfamiliar with their farm or products.

We include two race/ethnicity binary variables: farms with at least one Hispanic operator and farms with non-Hispanic white operators. Thus, we interpret these two race/ethnicity variables relative to farms that do not have a Hispanic operator but have at least one non-Hispanic non-white operator. We include race/ethnicity control variables because there could be language or cultural differences that influence the use of an online marketplace. We likewise test whether having a female operator influences online marketplace participation. Briggeman and Whitacre (Reference Briggeman and Whitacre2010) found that male farmers are more likely to use the internet. Also, O’Hara and Lin (Reference O’Hara and Lin2019) found that the gender and race/ethnicity of the farm operators influenced which local food market channels that the farm accessed.

Researchers have typically controlled for the farmer’s education level in internet-use studies, since it is a proxy for human capital (Mishra, Williams, and Detre, Reference Mishra, Williams and Detre2009; Briggeman and Whitacre, Reference Briggeman and Whitacre2010). The LFMPS does not contain questions about education. Instead, we include an independent variable that controls whether any farm operators have ever served on active duty in the U.S. Armed Forces, Reserves, or National Guard. This variable could be a proxy for human capital if veterans have received training on using the internet.

We control for farm size using a variable representing the farm’s total gross value of sales (GVS). The GVS code ranges between 1 for the smallest farms (i.e., those with sales less than $1,000) and 13 for the largest farms (i.e., those with sales greater than $5,000,000).Footnote ³ Larger farms are more capable of making non-DTC local sales (Low and Vogel, Reference Low and Vogel2011). Since it is likely that there are technological synergies between making DTC sales online and non-DTC sales online, larger farms may thus be more likely to develop an online platform to enable non-DTC transactions. Finally, we include state fixed effects due to variation in state policies, consumer preferences, and infrastructure that could influence the decision of a farmer to have an online marketplace. State fixed effects have been used in other LFMPS studies (O’Hara and Lin, Reference O’Hara and Lin2019).Footnote ⁴

In Model 2 we add two interaction terms that we did not include in Model 1. The two interaction variables represent post-2009 farms in (a) metropolitan counties and (b) non-metropolitan adjacent counties. The inclusion of the two interaction terms changes the interpretation of three variables that we include in Model 1. In Model 1, the post-2009 variable represents all post-2009 farms, whereas in Model 2, the post-2009 variable represents post-2009 farms in distant rural counties. The metropolitan variable in Model 1 represents all metropolitan farms, whereas in Model 2, it reflects pre-2009 metropolitan farms, and likewise regarding the non-metropolitan adjacent variable. The interpretation of these five coefficients (the two new interaction terms and the three variables that represent experience and metropolitan county classifications that we include in Model 1) in Model 2 are relative to the omitted variable, which represents pre-2009 farms in distant rural counties.

In Model 3, we include a binary variable that is equal to one if a farmer uses the internet to access resources to assist them with their operation. We combine the responses to the various categories to create one composite internet-use variable because responses to these internet-use questions are positively correlated with each other. We classify farmers as using the internet to access resources if they provide an affirmative response to any of the LFMPS internet-usequestions that we described previously. This variable could be interpreted as a proxy for the farm operator’s computer skills and human capital, which, in turn, could influence their decision to sell products online. Also, in Model 3, we include a binary variable that is equal to one for farms that sell value-added products through DTC marketplaces. Value-added products could be more conducive to online marketing if they are less perishable, more durable to ship, or have packaging that allows farms to promote their mission and enterprise identity without direct customer contact.

We include these two variables as a robustness check against the possibility of omitted variable bias. Nonetheless, we separately present regression results for Model 3 because the value-added sales variable and internet-use variable are both potentially endogenous. For instance, the decision to market value-added products could be made simultaneously with the decision to sell products via an online marketplace. Or, a farmer could train on the internet for the purpose of developing an online marketplace.

In Model 4, we estimate the same specification from Model 2 on the subsample of DTC farms that produce value-added products. The purpose of Model 4 is to assess whether the spatial dimensions of online marketplace participation varies among value-added DTC farms without including a potentially endogenous variable in the regression.

We report the marginal effects since the parameter estimates from logit regressions are challenging to interpret. For the binary independent variables, the marginal effect represents the change in the predicted probability of having an online marketplace as the independent variable changes from zero to one. For the continuous independent variables, the marginal effect represents the instantaneous rate of change. In both instances, we evaluate the other independent variables at their mean values.