3.1. Buyer Surveys and Instrumentation
Nontimber forest product (NTFP) harvesters in the eastern US forests are numerous, widely dispersed and often difficult to locate. Traditionally, they sell NTFPs to local buyers, who often pay cash. For this reason, the primary buyers of American ginseng were surveyed. These buyers are required by law to purchase a license to legally trade the species. Primary buyers also are required to report information about ginseng transactions, including the location and volume of raw material purchased. State agencies that oversee ginseng programs in their jurisdiction collect market-wide ginseng trade data from buyer reports, and hold and often publish lists of licensed ginseng buyers. Similar tracking programs do not exist for off-root species.
Not all ginseng buyers purchase off-root species, but virtually all off-root buyers with overlapping range purchase ginseng [33
]. Thus, the study area included every Appalachian state defined by the Appalachian Regional Commission (ARC) [34
] that maintains a ginseng program (exceptions: Mississippi and South Carolina do not permit ginseng harvests). Also included were adjacent eastern US states with ginseng programs, similar forest habitat and a history of commercial trade in off-root, medicinal NTFPs. This included the rest of the Ohio Valley in Indiana and Illinois, and the Ozark and Ouachita Plateaus in Arkansas and Missouri.
Three methods were used to collect and analyze trade data for eleven off-root species over two years (buying years: 2014 and 2015). The main method was a mail survey, but an option to complete the survey over the phone, or by using an online platform, also were offered. The mail survey followed Dalman’s Tailored Design Method [35
], and consisted of a presurvey recruitment mailing, a survey with instructions, a reminder postcard and a replacement survey.
The survey included four sections, which addressed four goals: (1) capture market participation and knowledge of market structure among respondents; (2) tally the volume of off-roots (non-ginseng) purchased by respondents; (3) determine the point-of-origin for off-roots purchased by respondents; and (4) measure the first-order value for off-roots purchased by respondents.
The first survey section asked buyers general questions about business structure, such as whether ginseng, off-root NTFPs, or other nonmedicinal NTFPs were purchased, if products were harvested in addition to purchasing from others, the number of employees, and to whom products were sold (product fate). Product fate was divided into five categories (consumers, fellow buyers, manufacturers, retailers and other). The second section asked buyers to report the amount purchased for 10 herbaceous plants typically harvested for roots and one tree harvested for bark, equaling eleven off-root species (Table 1
). Off-root plants and trees were selected if they were harvested in eastern US deciduous forests, valuable in the market, commonly harvested, and/or rare (i.e., of conservation interest) [36
The third section asked buyers to indicate from where their purchased material was harvested. It was important to avoid a level of resolution that could result in respondent identification. County-level analysis was desirable, but there would only be one large buyer at this scale in some cases. To preserve anonymity and make data compatible with the US Forest Service FIA program, the study area was partitioned into FIA research units. FIA units are multicounty areas composed of roughly the same amount and type of forest. Buyers were supplied with a map, as well as a list of FIA research unit regions and counties grouped into these units. They were asked to attribute percentages of their annual volume purchased for each product to FIA unit subdivisions. The final section asked buyers to report average prices paid to harvesters. Average prices were requested to avoid discomfort with respect to proprietary and sensitive payment transactions (after [17
]). Interviews were also solicited through the survey, and by using a chain referral method, to help refine the instrument in subsequent years.
A pilot study of primary buyers in Virginia and North Carolina’s ginseng registries was conducted in 2012 for the purchasing year 2011. One hundred and two surveys were successfully delivered, and the response rate was 10 percent. A revised pilot survey was tested in 2014 for the 2013 purchasing year. The second test included six southeastern states with ginseng programs: Alabama, Georgia, Kentucky, North Carolina, Tennessee and Virginia. Forty-five responses were received from 264 surveyed buyers, for a response rate of 17 percent.
Pilot findings were used to develop formal surveys, which were administered in 2015 and 2016 (with respect to the 2014 and 2015 buying years) for ginseng buyers registered in 15 states in the eastern US deciduous forest range. Response rates rose relative to pilot surveys as the program expanded, buyers developed familiarity, and the survey instrument matured.
To maximize analyzable data, 700 unique buyers hailing from 15 states (Alabama, Arkansas, Georgia, Illinois, Indiana, Kentucky, Maryland, Missouri, New York, North Carolina, Ohio, Pennsylvania, Tennessee, Virginia and West Virginia) were surveyed over the course of a two-year study, and their responses were combined.
Two-year means for purchased off-root NTFPs were used for buyers who responded in both rounds of surveying. If they did not purchase a particular off-root product in one year, the amount for the reported year was used. Aggregators, who also purchase volume from fellow primary buyers, were surveyed, but only the amount they purchased directly from the harvesters was included to avoid counting the same material twice. The majority of the material was dry, and the relatively small amount of fresh material was converted to dry pounds at a fresh-to-dry rate of 33 percent. These data were converted to kilograms for publication.
Two methods were used to measure nonresponse bias (after [37
]). In the ‘wave’ method, later respondents are assumed to resemble nonrespondents. Significant differences between date of response and variables of interest are considered a red flag. Wave analysis failed to find significant differences between the response date and purchasing tendencies of select species. A second method administered an abbreviated survey to a subsample of nonrespondents (n = 18), which uncovered no large outliers in weight and species purchased between respondents and nonrespondents. Taken together, respondents appear to be similar to nonrespondents. Nonresponse results were additionally evaluated by comparing industry figures and estimates of total market output obtained from aggregator input through open-ended comments and informal input.
3.2. Purchasing Probability
Habitat modeling methods were used to develop a model that predicts whether a respondent purchased in the study year: (1) no woodland medicinal NTFPs; (2) ginseng only; or (3) ginseng and off-root species. In many classic wildlife abundance studies, habitat modeling is used to identify limiting and enabling factors such as climate, site conditions, presence of food and potential nesting sites to develop predictive models that assign probability scores for the species in question across a landscape that is modeled, but not directly observed [38
]. A similar approach was used to model the probability of plant poaching in Shenandoah National Park [39
]. This study used a habitat modeling approach, and tested whether off-root purchasing can be predicated using environmental and socioeconomic variables associated with a business location. Probabilities in the predictive model derived from respondent data were associated with nonrespondent businesses to project off-root purchasing that has not been directly observed. Respondent values were combined with projected nonrespondent values to estimate the overall total annual trade volume, value and distribution for eleven species.
Multinomial logistic regression was used to model respondent off-root purchasing, because dependent variables were both parametric and nonparametric, as well as nonhierarchical. Model output includes a chi-square (χ2) likelihood ratio test, which indicates the overall significance of predictor variables, and a Nagelkerke R2 that approximates a measure of goodness of fit. Parameter estimates demonstrate the significance and effect of individual variables. Because the study segmented results into FIA research units, it was possible to identify and describe areas of high and low production across the study area. Exploratory factor analysis was used to optimize the number of variables, and collinearity diagnostics were run on hypothesized predictors to reduce spuriousness in the model.
To develop environmental variables for the buyer model, habitat data were determined using the US Forest Service FIA Program’s EVALIDator online tool [40
]. Total acreage and percentage of ‘ideal’ forest type group by county were used, which were defined as the forest-types where study species are commonly found: oak–hickory, maple–beech–birch and elm–ash–cottonwood. Total figures and percentages were coded as continuous variables, titled % Ideal Forest
To reduce model error and control the number of parameters, the many FIA research units were aggregated into regions based on physiographic areas, and one category that represented FIA zones located outside of the range for most species, and/or outside of the study area (Figure 1
The Northern Mountains included the Appalachian, Adirondack, Catskill and Pocono areas of New York and Pennsylvania. The flatter and more urbanized area to the east was designated as Mid-Atlantic/Northeast. The flatter, less forested regions north of the Ohio River and west of the Appalachian region were aggregated as the Central Lowlands. The Ozark Plateau in Arkansas and Missouri was treated as a distinct region, as was the rolling areas of mostly deciduous forest along the Ohio, Mississippi and Tennessee Valleys. The Piedmont area of the southeast was also considered as a distinct region. Areas that were either outside the range of ginseng, or outside the study area, were treated as one region. Locations were coded as a categorical variable: FIA Research Unit Regional Categories.
Secondary data sources were used to create the socioeconomic variables included in the buyer model. The US Department of Agriculture’s Economic Research Service publishes the Rural–Urban continuum (RU), which classifies counties based upon their urban population and whether they are part of or adjacent to large metropolitan areas [41
]. Data from the most recent RU dataset (2013) were used as an ordinal categorical variable in the model, called Rural
The US Forest Service’s National Woodland Owners Survey (NWOS) provided ownership data for use in the model [42
], to include percent corporate forest ownership as well as public versus private forest ownership. These data were summed to create a percentage of each FIA research unit that represents a potential ‘commons’ forest, on the basis that these large tracts of land may be more likely to be harvested. This continuous variable is referred to as % Commons
Additional socioeconomic data and data on population were obtained from the American Community Survey (ACS), an annual survey conducted by the US Census at the county level [43
]. Five-year averages from 2015 were used to code for median income, poverty rate, rate of Supplemental Nutrition Assistance Program (SNAP) benefit receivers (government assistance, sometimes referred to as ‘food stamps’) and a combination of unemployment rate with the percentage of the population out of the workforce. This ‘not working’ statistic is useful because many in the study area have claimed disability, stopped looking for work, or are participating in informal economic activity [44
The rate of drug overdose per 10,000 deaths was obtained from the County Health Rankings Program developed by the University of Wisconsin’s Public Health Initiative [45
]. Data were coded as continuous variables: Population, Median Income, % Below Poverty, % Not Working, % on SNAP Benefits, and Overdose Rate
Occupational data were gleaned from the ACS, specifically the percentage of the county working population employed in natural resources, forestry, mining and agricultural fields, subsequently coded to create a continuous variable, what is termed here % Relevant Employment.