Abstract
Many farmland valuation studies rely on survey estimates to form the dependent variable in a first-stage hedonic model. This study, based in New York State, provides a microscale comparison of transaction prices and producers’ market value estimates from the U.S. Department of Agriculture’s June Area Survey. Although we find similar weighted value distributions, regression results identify differences in marginal effect estimates and illustrate how market thinness plays a role in the comparability of observed transaction prices and self-reported values. The findings have implications for future hedonic studies, including insights into behavioral differences concerning how farmers and market participants perceive the value of farmland. (JEL Q11, Q51)
1. Introduction
Farmland is a nondepreciable asset, and farmland values, therefore, implicitly capture all potential future margins of adjustment in both production practices and land use (Mendelsohn, Nordhaus, and Shaw 1994; Capozza and Helsley 1989). As a result, a large volume of empirical research has used farmland values to infer implicit prices (or costs) of a host of nonpriced (or hedonic) market factors such as climate conditions (Deschenes and Greenstone 2007; Fezzi and Bateman 2015; Severen, Costello, and Deschenes 2018), development rights (Plantinga, Lubowski, and Stavins 2002), irrigation water supply (Buck, Auffhammer, and Sunding 2014), and renewable-energy policy (Towe and Tra 2013). Many studies, particularly in the area of Ricardian climate impact estimation (see review by Mendelsohn and Massetti 2017), rely on self-reported farmland value estimates from opinion surveys, as opposed to observed transaction prices. As a general rule, however, economists tend to favor transaction prices over opinion surveys because prices “reflect a free market’s expression of different individuals’ evaluations of property relative to other purchase options” (Darling 1973, 25). To this point, it has remained an open question whether sales prices or self-reported values are more appropriate in empirical studies of farmland values (Nickerson and Zhang 2014). This study examines the degree to which market transaction prices and self-reported market value estimates provide complementary information on farmland values and the implicit prices of nonpriced market factors. Specifically, we compare arm’s-length farmland transaction prices with hypothetical self-reported price estimates from the U.S. Department of Agriculture’s (USDA’s) June Area Survey.
While economists prefer transaction prices to values derived from opinion surveys, there are a number of situations in which the use of self-reported land values may be advantageous. In perfectly efficient markets (e.g., stock markets), transaction prices are equivalent to the fundamental value of the good being traded. However, market values can deviate from fundamental values because of market imperfections such as asymmetric information, transaction costs, bargaining power, and thin trading, all of which are well-documented characteristics of farmland markets (e.g., Just and Miranowski 1993; Chavas and Thomas 1999; Lence 2001; Sherrick and Barry 2003; Cotteleer, Gardebroek, and Luijt 2008).1 In addition, confidentiality (e.g., due to nondisclosure rules) may limit access to transaction records (Geltner, MacGregor, and Schwann 2003), which are often subject to recording and measurement errors (Doss and Taff 1996). Furthermore, the properties transacted in a given year may not constitute a representative sample, which could lead to sample selection bias and call into question the external validity of inferences based on observed farmland market transactions. That is, some properties may be more likely than others to be sold (Clapp and Giaccotto 1992), and as a result, the distribution of transaction prices may deviate from that of fundamental values for the population of interest. Finally, it has also been shown that market distortions such as the influence of real estate agents may drive a wedge between market prices and marginal costs (Cotteleer and van Kooten 2012; Doss and Taff 1996). Given these potential concerns, transaction prices may provide an imperfect signal of the underlying fundamental value of farmland.
Partly stemming from these perceived deficiencies of observed farmland transaction prices, a number of institutions (e.g., USDA, Federal Reserve banks, and various land-grant universities) conduct opinion surveys to monitor farmland market conditions (Kuethe and Ifft 2013). In contrast to sales prices, which reflect the outcome of bargaining between buyers and sellers in a local property market, self-reported values are based on respondents’ opinions of the likely settlement price of a particular or theoretical parcel of farmland. One potential issue with relying on hypothetical farmland values concerns the survey target population. Most USDA surveys (e.g., the JAS and Census of Agriculture) target farm operators, as opposed to farmland owners. Since approximately 40% of all U.S. farmland is rented, many farmers often do not own the land they operate, which may influence how they answer questions concerning its hypothetical market price.2 In addition, since self-assessed values do not represent the outcomes of actual market exchanges, they are theoretically unbounded and may suffer from a form of hypothetical bias. Specifically, although they are typically elicited as the price that would result from an arm’s-length market exchange, self-reported values may overstate (or understate) the true market value of the land if respondents do not accurately account for the compensating demand-side influence on the price they would be willing to accept.3 It is clear, then, that both data sources contain potential biases, prices from market imperfections and sample selection, and self-reported values from survey target population issues and hypothetical bias.
The use of opinion surveys to track real estate market conditions is not unique to the farm sector (Kuethe 2016; Geltner, MacGregor, and Schwann 2003). The broader real estate literature suggests that homeowners’ beliefs of the market value of their homes, in aggregate, result in estimates that are equal to those of professional appraisers (Kish and Lansing 1954; Kain and Quigley 1972) but higher than market transaction prices (Goodman and Ittner 1992; DiPasquale and Somerville 1995; Kiel and Zabel 1999). A more recent analysis by Banzhaf and Farooque (2013) shows that indices based on self-reported housing values from the U.S. Census microdata are highly correlated with those based on transaction prices, with observed sales prices performing slightly better in terms of how they reflect local public goods. In the context of farmland markets, previous studies have demonstrated a high correlation between farmland transaction prices and survey-based values, but find that survey values tend to understate observed price appreciation (Scott and Chicoine 1983; Barnard and Wunderlich 1984; Zakrzewicz, Brorsen, and Briggeman 2012). Stinn and Duffy (2012) further find that county averages of farmland value estimates based on opinion surveys tend to be higher than corresponding average sales prices, though not in a statistically distinguishable manner. Other studies have shown that aggregate state-level values tend to match well across data sources, but diverge in certain counties (Shultz 2006), such as those near urban areas (Gertel 1995).
A related strand of the literature takes a more fine-scale hedonic approach to examine the differences between individual farmland transaction prices and assessed values used for tax purposes. These studies are better able to account for parcel-level characteristics through hedonic price analysis, and they suggest that assessed values and sales prices tend to diverge in their ability to detect some determinants of farmland values. Ma and Swinton (2012) find that estimates of farmland values prepared for tax purposes consistently underestimate the value of surrounding natural amenities. In a study of water valuation in droughtprone regions, Grimes and Aitken (2008) find that assessed land values are highly correlated with transactions prices but underestimate the added value of irrigation. Thus, both Ma and Swinton (2012) and Grimes and Aitken (2008) suggest differences in attribute valuation between market participants (i.e., transaction prices) and the opinions of tax professionals.4 This finding is consistent with similar studies of other real estate classes, such as residential real estate (Berry and Bednarz 1975; Nicholls and Crompton 2007; Bowman, Thompson, and Colletti 2009; Cotteleer and van Kooten 2012) and industrial land (Kowalski and Colwell 1986).
This study makes several contributions and has implications for future hedonic valuation research involving agricultural land. First, while the existing literature documents deviations between transaction prices and opinion surveys in aggregate, data access has limited the ability of researchers to examine more granular differences between these sources of information. The present study, which takes place in New York State, is the first to compare observation-level farmland values derived from (1) an opinion survey of farmers administered by the USDA and (2) market transaction records. Our study is unique, given its use of a confidential parcel-scale dataset of self-assessed farmland values in conjunction with a comprehensive set of arm’s-length farmland transaction prices. A preliminary baseline comparison suggests minimal differences between the datasets when transaction prices are weighted by parcel acreage.
Second, a comparative regression analysis identifies differences in the marginal implicit values of certain farmland characteristics. The regression analysis builds on the previous studies of Ma and Swinton (2012) and Grimes and Aitken (2008) by analyzing differences in the marginal implicit prices of various agricultural property characteristics. However, instead of comparing transaction prices to assessed values, we are able to examine the differences between sales prices and values from opinion surveys, a distinction that has broader implications within the revealed preference literature. For example, the regression results suggest that development potential is capitalized differently into self-reported values and transaction prices. Overall, our results provide new insights into the nature of hypothetical bias associated with using self-reported land values in hedonic modeling applications, a topic that has received little attention to date in the existing literature.
Experience and salience represent two fundamental behavioral concepts that may explain why farm operators would place different values on farmland characteristics than is implied by market transactions. If farmers have recently purchased farmland or have observed many purchases, their farmland value estimates may be more likely to be informed by the same characteristics that are valued in market transactions. Farmers have varying, and unobservable, levels of experience with land markets, but basic intuition suggests that operators in areas with relatively more transactions would be more knowledgeable about potential sale outcomes. While some research in this area suggests that experience would lead to more consistency with market outcomes (List 2003; Palacios-Huerta and Volij 2009), recent studies from the psychology and economics literature suggest that it may not always be the case that experience leads to rational behavior (DellaVigna 2009). DellaVigna (2009) suggests that market thinness, in particular, may limit the role of experience in mitigating nonstandard economic behavior. To this end, we analyze how the extent of local farmland market activity influences the regression estimates and the resulting comparability of sales prices and self-reported values.
Related to experience is salience, which in some research has been shown to play a larger role than experience in explaining economic behavior (Haigh and List 2005). Salience, or the degree to which a factor stands out or is noticeable relative to other factors, can directly influence decision-making and has been shown to be relevant in many contexts (Bordalo, Gennaioli, and Shleifer 2012a, 2012b). For example, in areas with more recent droughts, surveyed farmers were more likely to report believing in a changing climate (Diggs 1991). Farm operators may overemphasize farmland characteristics that are more salient to them. For example, farm operators may place greater emphasis on soil characteristics and weather patterns than urban amenities or connectivity, because their day-to-day work involves agricultural production. The concepts of experience and salience inform our empirical framework and are useful for interpreting our results.
2. Data
New York State has a diverse agricultural sector, as well as large metropolitan centers and recreational areas that exert development pressure on farmland values. According to the 2012 Census of Agriculture, New York is home to over 35,000 farms and nearly 7.2 million acres of agricultural land (USDA NASS 2014). Approximately 30% of farms are beef cattle operations and 14% are dairy operations. New York ranks third in the United States for milk receipts and is in the top five for many fruits and vegetables (including apples, cabbage, grapes, and snap beans). While corn and soybeans are the most common field crops grown, many operations also produce tree crops, oats, and potatoes.
Farmland Value Data
Our empirical analysis draws on two primary data sources: (1) self-reported estimates of land values obtained from the USDA’s June Area Survey (JAS) and (2) farmland transactions obtained from the New York Office of Real Property Tax Services.
June Area Survey
The JAS is a multipurpose survey administered annually by the USDA’s National Agricultural Statistics Service (NASS). JAS data are the primary inputs for a variety of USDA publications, including the official annual land value estimates (e.g., USDA NASS 2016). The survey sample is drawn from a comprehensive probability- and area-based sampling frame that divides the United States into segments that are approximately one square mile in area. In intensively cultivated areas, segments are sampled at a rate of about 1 out of 125, and in areas that are less intensively cultivated, segments are sampled at a rate of about 1 out of 250 to 500. When a segment is sampled, enumerators contact all producers operating “tracts” within its boundaries. Tracts denote land inside a segment in a common farm operation.5
The JAS is administered using a rolling panel sampling frame, with approximately 20% of sampled segments rotating out in a given year.6 We also note that, compared to the Census of Agriculture, for which response is mandatory, the JAS microdata may suffer from a higher degree of nonresponse bias. NASS provides weights, which are adjusted each year based on the collected JAS responses to produce representative national and state-level estimates. The JAS elicits separate per-acre estimates of the market value of cropland and pastureland from each farmer-respondent (Figure 1).7 For our purposes, we combine the stated cropland and pastureland value responses into a single estimate of total farmland value for each tract. In contrast to surveys conducted by banking institutions and land grant universities, which typically ask respondents about the value of land for a given county or region, the JAS asks farmers about the specific land in their operation. One notable distinction between the JAS and transaction data (described below) concerns the value of buildings and structures. The transaction data include the price of any buildings or structures on the land, while the data we use from JAS survey questionnaires is meant to reflect the hypothetical sales price of the land alone.8 Prior studies that use the JAS land value data include those by Schlenker, Hanemann, and Fisher (2007), Towe and Tra (2013), Borchers, Ifft, and Kuethe (2014), Ifft, Kuethe, and Morehart (2015), Ifft, Bigelow, and Savage (2018), and Bigelow et al. (2019).9
Farmland Transactions
Farmland transaction records were obtained from the New York Office of Real Property Tax Services. The transaction records include the total acres sold, the sales price for each parcel, and several other variables we use to screen the data. We first eliminate all transactions between related parties to eliminate properties that may not have been sold for fair market value. Second, we retain only parcels that are classified as agricultural (property class 100-190). Third, we remove all transactions for parcels less than one acre in area (about 3% of transactions), since these are likely associated with homesteads or other farmland fragments not used for production, and transactions with prices less than $100/ acre or greater than $50,000/acre (about 2% of transactions). Finally, we remove observations from Long Island, as farmland in this region differs markedly from that of the rest of the state.10
The eight regions (omitting New York City and Long Island) used in our analysis are shown in Figure 2. Gray shading indicates the counties with an above-median number of transactions, which we use to demarcate thin and active markets. Farmland transactions in New York tend to take place in the eastern and southern portions of the state. The farmland transaction prices are mapped in Appendix Figure A1. Higher-priced parcels are clustered in the Hudson Valley and Capital regions (around Albany), both of which have relatively active markets and large urban areas that exert development pressure on farmland. Prices are also high in the Finger Lakes region, which has both high-valued agriculture and recreational areas that have the potential to increase the nonagricultural use value of farmland (Borchers, Ifft, and Kuethe 2014). High prices in northwest New York likely reflect the area’s high-quality soils associated with the region’s dairy and field crop sectors. As shown in Appendix Figure A2, transaction parcel size is fairly uniform across the state. The main exception is the concentration of relatively small parcels in the Hudson Valley region, which, as noted above, is also characterized by relatively high sales prices. Since smaller farmland parcels are generally more attractive for development, the combination of relatively high prices and small parcels in Hudson Valley likely reflects farmland conversion potential (Brorsen, Doye, and Neal 2015).
Additional Data Sources for Explanatory Variables
Both the survey and transaction data are georeferenced, allowing us to link observations with a number of additional land characteristics and controls. The survey data are georeferenced using JAS segment centroids obtained from the USDA NASS. Sales are georeferenced by linking the tax identification numbers for the sold parcels to statewide tax parcel maps. We are able to georeference almost 85% of sales using this procedure. The additional variables included in our analysis are described in Appendix Table A1.
We use several variables to control for differences in potential agricultural productivity. First, we construct a measure of soil quality derived from the 10-class soil quality index developed exclusively for farmland taxation in New York State (New York State Department of Agriculture and Markets 2017). The index is calculated using historic yield potential for hay and corn and is used by appraisers to value specific soil types. For our purposes, the soil quality index is grouped into three bands representing high-quality soil (classes 1–3), medium-quality soil (classes 4–6), and low-quality soil (7–10).11 Second, to account for cross-sectional differences in climate, we use the 800 m PRISM data to measure average precipitation and mean temperature over the 30-year period covering 1981–2010.12 The climate variables are measured at the start of the growing season, covering the months of April through June, the period most critical to agricultural production in New York. Specifically, cold weather and excess moisture, which often occur during this period, are historically the leading cause of loss and trigger of crop insurance indemnities (Ifft 2018).
Finally, we include additional characteristics representing nonagricultural factors that may affect farmland values. These variables include commuting time to towns with a population greater than 2,500 residents and large urban areas with a population greater than 1 million residents, distance to highway exit ramps, median household income, and a population interaction index. The population interaction index is measured as the inverse-distance-weighted sum of population within a 50-mile radius of the parcel centroid. The population data used in constructing the variable were developed by downcasting the 2010 U.S. Census block population counts to a 0.5 km grid covering the contiguous United States.13 All of these factors should, in theory, be related to capitalized development potential but could also signify access to agricultural markets. We also account for recreation potential with a variable representing proximity to recreational water bodies, a popular source of outdoor recreation in many parts of New York, such as the Finger Lakes region.
Preliminary Analysis
The study period for our analysis covers 2009–2014. In this period, the JAS includes 139 individual segments, yielding 389 usable segment-year observations. The transaction dataset includes 2,741 arm’s-length transactions. Table 1 provides the mean values of the dependent and explanatory variables used in the analysis.14 In order to gauge the spatial comparability of the survey and transaction values, we would ideally provide a map of the survey values, similar to that in Appendix Figure A1. However, sample respondent disclosure restrictions limit our ability to report a map of the survey values to compare where transactions occur with the locations of completed JAS surveys. We do note that the shares of each dataset falling within the eight New York economic regions are relatively balanced. The regional sample shares are as follows: Capital Region (6% of sale sample, 8% of survey sample), Central New York (12%, 15%), Finger Lakes (28%, 22%), Mid-Hudson (3%, 4%), Mohawk (14%, 14%), North Country (14%, 6%), Southern Tier (9%, 14%), and Western New York (13%, 17%).
The first two columns of Table 1 report the unweighted means of all variables in employed in our analysis from both the JAS (Survey) and transaction database (Sales). Normalized mean differences (NMDs) are reported in the third column.15 Casual inspection of these summary statistics suggests that the two datasets are relatively balanced, with only travel time to small towns and precipitation exceeding the rule-of-thumb NMD value of |0.25| (Imbens and Wooldridge 2009). The mean transaction price of $3,802 is roughly 32% higher than the corresponding JAS estimate of $2,889, but the corresponding NMD is a relatively modest –0.14.16 To some extent, the higher average value in the transaction data is expected since, as noted above, observed transactions may reflect the value of both land and buildings, while the survey data reflect only the value of land. Despite the small NMD, it is hard to argue that a 32% difference in value is not an economically meaningful indicator of the lack of comparability between the prices and survey values.
Columns (4) and (5) of Table 1 report weighted means of the same variables. The JAS observations are weighted using the survey weights, which are calibrated to yield representative state-level estimates of farmland values, acreage, and production. In contrast to the JAS, the transaction data do not have a built-in weighting variable. As Barnard and Wunderlich (1984), we weight sale observations by parcel acreage in an effort to more accurately represent the total area of farmland sold over our study period. After weighting, the survey value mean is reduced by 14%, while the mean transaction price declines by 30%. More importantly, the difference in per-acre land values is reduced substantially, with the average weighted sales price of $2,576 just $90 (4%) higher than the average JAS value of $2,486, producing a normalized weighted mean difference of –0.02. Aside from precipitation and acres, the NMDs indicate that the sale and survey samples are relatively balanced. The acreage difference of 155 acres stands out but is expected since agricultural land transactions may span multiple parcels owned by the same person. The difference in precipitation is fairly minimal at 0.8 inches, with the relatively high accompanying NMD being driven by the small standard deviations in precipitation for the two samples (1.1 inches; not shown).
Figure 3 plots the cumulative distribution functions (CDFs) of each sample with and without the weights. Without weights (Figure 3a), the CDF for the survey values approaches one faster than that of the transaction prices, which accords with the disparity in unweighted mean values in Table 1. This difference is also borne out by a two-sample Kolmorogov-Smirnov (KS) test of the equality of the sample distribution functions. The KS test results reject the null hypothesis of equality when using the sales prices (D = 0.67, p = 0.049) and the survey values (D = –0.086, p = 0.006) as the reference group. After weighting, the CDFs align much more closely based on visual inspection (Figure 3b). The KS test results bear this out as well, as we fail to reject the null hypothesis that the two distributions are equal when the survey values serve as the reference distribution (D = –0.049, p = 0.231), although the difference remains significant when the sales prices are used as the reference group (D = –0.105, p = 0.001).
The similarity between weighted self-reported land values and transaction prices is notable for at least two reasons. First, the weighted means reveal that the market for New York farmland is heavily influenced by small, high-valued land parcels. A “small parcel price premium” is well established in the existing literature, with recent work suggesting that it stems from the ease of developing smaller parcels (Brorsen, Doye, and Neal 2015). Second, the close correspondence between the weighted JAS values and weighted transaction prices reduces concerns regarding the comparability of the two data sets. A potential critique of the use of sales data in hedonic regression applications is that sold parcels are not representative of the broader stock of real estate, inducing a form of selection bias (Clapp and Giaccotto 1992). This concern is particularly relevant in the context of farm real estate markets, which are notoriously thin, with approximately 0.5% of U.S. farmland changing hands annually (Sherrick and Barry 2003). The similarity between the mean values of the sales prices and self-reported values suggests that farmers are relatively good predictors, on average, of the value of their land. The similarity of the two sources is also encouraging from a survey design perspective, since a primary goal of the JAS is to produce representative national and state-level farmland value estimates. Weighting farmland sales by acreage produces a more representative sample of transaction prices and provides a simple, replicable method that can be applied to future agricultural hedonic modeling applications.17
3. Estimation Strategy
We estimate a set of hedonic farmland value models to examine the degree to which the implicit prices of various farmland characteristics from the two data sources differ (Palmquist 1989).18 A generic version of the hedonic farmland price model can be expressed as [1] where Vi is the per-acre land value of parcel i, Ai is a vector of variables related to the agricultural productivity, Di is a vector of nonagricultural variables (such as urban proximity or recreation accessibility) that potentially impact farmland values, β1 and β2 are vectors of unknown parameters to be estimated, and εi is a standard independent and identically distributed residual term.
We estimate two versions of Equation [1], one using sales prices and another using self-reported land values from the JAS. To ensure that the models are comparable, both specifications include the same set of explanatory variables listed in Table 1. Following one of two general strategies for comparing the performance of competing datasets in property value modeling, we do not take advantage of any of the features of one dataset that are not available in the other (Banzhaf and Farooque 2013). For instance, we do not exploit the panel structure of the JAS data using, for example, segment fixed effects, nor do we include any supplemental covariates that could be drawn from the survey data, such as the age of the farm operator or other operation and production variables. Likewise, we do not include any information from the sales data, such as parcel property class, that is not available in the JAS dataset. We acknowledge that omitting some of the factors unique to either of the two datasets may induce omitted variable bias, a problem that has attracted much of the recent methodological attention in the farmland valuation literature (Nickerson and Zhang 2014). We emphasize that the goal of this exercise is to assess the comparative performance of the two data sources in terms of how they reflect a set of factors commonly used to explain farmland values, not to estimate the marginal effect for a particular policy variable or determine the data-generating process for each source individually. Given that weighting has such a substantial effect on the farmland value differences and limited effects on the explanatory variables, we apply survey and acreage weights in our preferred models.19
We first estimate a set of baseline results pertaining to [1] using the full JAS and sales samples. Our preliminary comparison suggested that, in aggregate, farmers’ self-assessed values accurately reflect observed market prices. The baseline hedonic price models examine the degree to which self-reported values accurately reflect the implicit market prices of various farmland attributes, including those related to agricultural productivity and development potential. In addition to the baseline JAS model, we also estimate a set of models stratified by market activity. Specifically, we estimate separate models for subsamples of parcels falling in counties above and below the median sales volume observed during our study period (2009–2014).20 We refer to the sample based on counties with an above-median sales volume as the “active-market” (or “thick-market”) sample, and the one based on counties below the median as the “inactive-market” (or “thin-market”) sample. Restricting the samples in this way allows us to gauge how market activity affects farmers’ perceptions of farmland values and the degree to which learning occurs as market information is transmitted to potential market participants. A priori, one would expect that survey responses and sales in thicker markets will be affected by a similar set of characteristics, as respondents are likely to have received relatively more information on the factors that influence the potential market value of their land. In addition to estimating separate models for the survey responses and transaction prices, we also estimate pooled models. The first pooled model contains the same set of covariates as the models for each individual source, plus a dummy variable that takes a value of 1 for the survey observations: [2] In [2], the parameter estimate for γ denotes the extent to which the survey values and sales prices differ after controlling for other factors that affect the value of farmland. The second pooled model expands on the first by including a complete set of interactions between the survey dummy variable and all other included covariates: [3] The specification in [3] allows for the capitalized marginal values of the agricultural (α1) and development-related (α2) factors to differ across transaction prices and survey responses. The source-specific models (i.e., the two versions of [1]) are nested in [3], providing a convenient way to formally test for statistical differences in the estimates between the two sources.
Equations [1], [2], and [3] are estimated in log-linear form using per-acre values. All models also include region and year dummy variables. Standard errors in the three equations are clustered at the county level. In equations [2] and [3], we cluster the standard errors by county and data source to replicate the standard error estimates generated from estimating the models separately.21 We note that clustering by county represents a conservative approach to inference. Standard error estimates generated using the spatial HAC estimator of Hsiang (2010) are slightly smaller and available upon request.
4. Results
Baseline Results
Table 2 reports the coefficient estimates from Equation [1] using the full samples of self-reported JAS values (column 1) and sales prices (column 2), with differences estimated via Equation [3] (column 3). Standard errors clustered by county are reported in parentheses. Results from the simple pooled model (Equation [2]), which includes only a dummy variable to distinguish sale and survey observations, indicate that there is not a significant difference in land values between the two datasets after controlling for the full set of model covariates (Appendix Table A2). This finding bolsters the results from the weighted mean comparison and implies that there is no detectable gap between sales prices and self-reported farmland value estimates after controlling for common observable farmland value drivers.
Parcel acreage has a positive effect on survey values and a small, weakly significant, negative effect on transaction prices. The difference is statistically meaningful and likely reflects the fact that farmers themselves place a higher premium on having large contiguous tracts of land to operate. Farmland transaction prices, however, appear to reflect the small parcel development premium, even after weighting by acreage (Brorsen, Doye, and Neal 2015).
Both sets of results indicate that farmland values are positively impacted by soil quality, with estimates of 0.6% to 0.7% for each percentage point increase in high-quality soils and 0.3% for medium-quality soil, though the latter is significant only in the sales price model. Given the larger number of observations in the transactions data (2,741 compared to 389), one would expect to find fewer statistically significant coefficients in the JAS regressions. Precipitation has a significant negative effect in the transaction model (–10% per inch), with no significant effect in the survey model. The negative effect may reflect the conventional belief that flooding and excess moisture have historically created more problems for New York farms than drought. Thus, relatively wetter areas tend to be associated with lower land values. Given that the magnitude of the precipitation effect is more than twice as big in the survey model, the lack of a statistical difference in the estimates may largely be the result of noise in the relatively small survey sample. Neither model produces a significant effect for mean temperature. Between the survey and sales model results, we find no evidence of statistically distinguishable effects for the main agronomic explanatory factors related to soils and climate.
The way urban development pressure (or market access) is capitalized into land values differs between the sale and JAS results. Land values decline with travel time to large cities in both models. An additional hour of travel time is associated with a 33% decline in peracre sales price and a (weakly significant) 20% reduction in self-reported values, though the difference is not statistically significant. Neither model produces a significant relationship between farmland values and the proximity of smaller population centers. The sale and survey models, however, diverge in how they capture the effects of nearby population, proximity to transportation infrastructure, and local income levels. The coefficient on the population interaction index is 1.5% in the JAS model but is indistinguishable from zero in the transaction model, and the difference is statistically significant. Distance to interstate highway ramps exerts a negative impact on sales prices (1.1% with each additional mile of distance). Sales prices also reflect a positive effect of median household income (1.8% for each additional thousand dollars). In addition, the implicit price of recreational amenities, as captured by the distance to recreational bodies of water, is capitalized into sales prices (1.8% decline for each additional mile). Survey values do not reflect any of these factors, and the resulting wide confidence intervals around the corresponding estimates result in the sale and survey marginal effects being indistinguishable.
In sum, the baseline estimates imply that different measures of development potential are salient to farmers compared to farmland market participants. Specifically, since farmers are generally noncommuters, it seems reasonable to suspect that they would attribute a lower share of their land’s value to how easy it will be to commute from their land to a major urban area. Similarly, farmers would seem to be less likely to account for influences related to recreational water accessibility and the negative externalities associated with highway ramp proximity. It also seems plausible that farmers, many of whom likely live in close proximity to the land they farm, may perceive at least some of the stream of future development rents through the level of nearby population, as this may convey information on local amenities, such as having multiple school choices and shopping areas, or access to local markets.
Subsample Results for Thin and Thick Local Farmland Markets
We now present results for a set of regression models estimated using subsamples of observations from what we refer to as active (or “thick”) and inactive (or “thin”) farmland markets, which we define as counties with a sales volume that is greater than or below the median observed during our study time frame. When compared to our baseline results, these estimates shed light on the degree to which the implicit hedonic values derived from survey responses (and sales prices) are influenced by the extent of local farmland market activity. The level of activity in a local farmland market may affect farmers’ perceptions of market prices, because thinner markets are generally characterized by lower levels of information exchange.22 As a result, we expect survey responses and sales prices in more active markets to be affected by a similar set of characteristics.23
The first three columns of Table 3 report the results for observations located in active-market counties, and the last three report the same for thin-market counties. More than half of the JAS observations (248; 63% of the baseline sample) are contained in thick-market counties, indicating that the JAS survey is concentrated in counties with relatively active farmland markets. The market-activity subsample models are also estimated using the simple pooled specification, results of which are reported in the final two columns of Appendix Table A2. In the pooled model for the active-market sample, we do find evidence that, after controlling for the other covariates in the model, the JAS values are higher than the sales prices, with the survey dummy variable producing a significant effect of 16.9%.24 The survey dummy variable produces a negative, but insignificant, effect in the thin-market model.
Parcel acreage exhibits a strong negative effect on active-market sales prices, and the opposite, a strong positive effect, on active-market survey values, reinforcing the idea that farmers and market participants have conflicting preferences for the amount of land involved in a transaction (either hypothetical or observed). While the signs of the acreage effects are the same in the thin-market results, the estimates are not different from zero.
The effects of the production-related explanatory factors exhibit several notable changes when the sample is restricted by market activity. First, the thick-market subset suggests that farmers place an equal premium of 1% on both high- and medium-quality soils, compared to the baseline model that showed that high-quality soils commanded a higher marginal value. In contrast, the sales prices in active markets convey roughly the same values for soil quality as the baseline estimates. The survey effect for medium-quality soil is more than double that of the sales prices, implying that farmers in active sales markets may overvalue soil quality. We do not find much evidence of soil quality being capitalized into land values for the thin-market results, apart from a small, positive effect of medium-quality soils in the transaction sample.
In terms of the climate variables, the effects are strikingly different in the active-market subsample compared to those of the full sample. The active-market survey results indicate that each additional inch of precipitation is associated with a 39% reduction in land values, while each additional degree Celsius is associated with a 32% reduction. These effects are larger than those of the sales model (at the 10% significance level) and suggest that farmers in active markets are more sensitive to factors associated with local climate. The large climate effects are likely the result of the narrow ranges of both precipitation and temperature observed in our data set, where a one-inch increase in precipitation corresponds to slightly less than one standard deviation (1.1 inches), and a one-degree Celsius change in temperature equals more than a one standard deviation change (0.7 degrees Celsius). For the sale observations, climate is capitalized into prices in roughly the same manner as the baseline sample. We find no evidence of climate effects in any of the thin-market estimation results.
Nonagricultural price determinants are a more significant driver of JAS values in the active-market sample. At 38% for each additional hour of travel time, the value gradient for urban area proximity in the restricted JAS sample is similar to the corresponding baseline sales estimate. The active-market sale sample, however, shows no effect of being closer to large cities. Household income has a positive effect of 5.5% in the restricted JAS model, contrasting with the null effect found in the active-market sales model. It also appears that some of the urban influence effect captured by nearby population in the baseline JAS model has shifted to local income, as the effect of the population interaction index is not significant when using the restricted JAS sample. The effect of highway ramp proximity also exhibits meaningful differences, indicating that being closer to a highway is a good thing from the perspective of farmers and the opposite for transaction prices. The recreational water access effects are relatively unchanged in the active-market models.
For thin markets, we find that large city proximity produces a large effect in both the survey and sales models. Survey values are also affected by nearby population levels (significant at the 10% level), while no such effect results for sales prices. The two largest differences for the thin-market results stem from (1) the large positive effect of being farther away from small towns in the survey sample and (2) the positive impact of local income levels on sales prices.
Contrary to our expectations, it is not clear that the self-reported land value model from the JAS compares more favorably to the sales model when we restrict our attention to active local farmland markets. However, it does appear that, qualitatively, the active-market JAS values are affected by a similar set of factors as the full and active-market samples of sales prices, although in several cases the magnitudes of the estimates are larger in the restricted JAS model. This apparent overcapitalization suggests that there may be upward bias in how JAS respondents perceive the value of factors that affect the hypothetical transaction price of their land. Furthermore, the larger implicit values come even though the active-market JAS sample has a slightly lower mean value ($2,217/acre) compared to baseline JAS sample ($2,487/acre) (not shown). One possible explanation for this pattern is that farmers located in more active markets have experience with a wider range of potential transaction outcomes and may be anchoring their estimates to higher prices. Farmers in urban-influenced areas may overstate the value of their land, which could be reflected in the active-market marginal effects (Gertel 1995).
Overall, the active-market results suggest that farmers in relatively thick local farmland markets (1) are relatively more market savvy in terms of the variety of attributes they account for, (2) overestimate the value of their land in general, as evidenced by the survey dummy variable in the pooled model, and (3) overestimate capitalization of the influence of several land attributes, particularly those that convey information about agricultural use potential. We also find that land values, either self-reported or based on observed prices, in active markets are influenced by a more cohesive set of factors compared to those derived from thin markets, where our estimation results are far noisier. While some of this is likely attributable to the naturally smaller samples in the thin-market areas, it is also in line with the notion that there is a greater amount of information transfer taking place in active markets, which results in more consistent capitalization of observable farmland value determinants.
Robustness Checks
We conduct a series of robustness checks to measure the sensitivity of our results to alternative estimators and specifications. The full results are reported in the Appendix but are briefly summarized here.
Unweighted Estimates
To gauge the effect of weighting on the regression results, we reestimate our regression models without weights (Solon, Haider, and Wooldridge 2015). The unweighted results are shown in Appendix Tables A3 (full sample) and A4 (market-activity subsamples). In the full samples, the same factors (acreage, nearby population, and highway ramp proximity), with and without weights, exhibit meaningful differences between sale and survey values, although, contrary to the weighted results, the unweighted sales prices do exhibit a large significant effect related to small-town proximity (1.25%/hour of commute time). For active farmland markets, there are minimal changes to the coefficient differences, except for a significant difference in recreational water proximity and a marginally significant difference for high-quality soils. Both measures of land value also capitalize a significant effect related to small-town proximity without weights. The thin-market results also yield the same general pattern of differences, aside from parcel acreage, which shows a meaningful difference. Generally, the model specification appears to be a better fit for the unweighted thin-market sales prices.
Nonlinear Model Specification
The main purpose of the regression application was to compare results between farmland transaction prices and survey values based on a single model specification. We now explore the sensitivity of the main results to the use of an alternative model specification that allows for nonlinearities in the effects of the explanatory factors (Appendix Tables A5 and A6). Specifically, we estimate a version of the model that includes second-order, squared terms for the distance and proximity variables, climate, nearby population, and income (as in Plantinga and Miller 2001). In addition, we also include a precipitation-temperature interaction variable (as in Fezzi and Bateman 2015). Overall, the nonlinear specification does not fit our data well. In the full sample, the same pattern of results emerges for acreage and soil quality, and there are a few marginally significant differences produced for the effects of median income (linear and squared) and the second-order term for precipitation. In active markets, we continue to find differences in the climate effects, particularly for precipitation (linear, squared, and the interaction term). We note that although the coefficients on the first-order climate variable terms are large, the implied marginal effect of a one-inch increase in precipitation, evaluated, for example, at the active-market survey means of 11.3 inches and 13 degrees Celsius, is 46%, which is comparable to the effect derived from the baseline active-market survey model. The thin-market results exhibit just one meaningful difference in coefficients, stemming from the positive and significant squared term on the population interaction index in the survey model.
Urban Area Proximity
In the baseline specification we included two measures of urban area proximity, one for small towns with at least 2,500 people and another for large urban cities (i.e., New York City) with a population of at least 1 million. Past research has shown the importance of accounting for multiple sources of urban influence (Zhang and Nickerson 2015). Using an alternative population cutoff of 10,000 for small towns produces no differences in the estimates for the full or active-market samples (Appendix Tables A7 and A8). For thin markets, however, the large positive effect found in the base specification for survey values goes away, while a negative gradient emerges for the sales prices, suggesting that the thin-market results are sensitive to how proximity to small urban centers is accounted for in the model. We also estimate a version of the model that includes proximity to intermediate cities with populations of at least 100,000 (Appendix Tables A9 and A10). Doing so yields no effect in the full sample results. However, the active-market sales prices exhibit a significant relationship with travel time to intermediate cities, which are associated with a negative gradient of 29% per hour. Sales prices in thin markets, however, yield a positive effect of being farther away from intermediate-sized urban areas (33% per hour). Overall, these results provide further evidence that the underlying first-stage hedonic price function differs for survey values and sales prices, particularly once market activity is considered.
Cross-Prediction Comparison
The hedonic regression analysis identifies a number of differences between sales prices and self-reported values in the implicit prices assigned to various farmland attributes. An alternative way of viewing the comparability of the opinion survey and sales data is to compare the conditional distributions of land value predictions generated with the hedonic models to the observed distributions of sale/ survey values. To this end, we generate distributions of cross-predicted land values, where we fit the sample of transaction prices with the survey model and, vice versa, use the sales model results to predict the distribution of self-reported survey values.25
We conduct this exercise using the full sample of JAS and transaction observations as well as the subsamples defined by market activity. After generating the model predictions, we then measure cross-prediction accuracy by computing the root mean square error (RMSE) from using each model to predict the counterpart set of land values (e.g., the fitted sales prices generated from the survey model estimates are compared to the observed sales prices). The RMSE is a commonly used metric in forecast evaluation. RMSE assumes a symmetric loss function (positive and negative prediction errors are equally weighted) that places a steeper loss on large errors. As a result, a prediction that minimizes RMSE is associated with few large deviations from zero. We calculate an overall RMSE for each of the four sets of predictions, as well as RMSEs for each decile of each underlying distribution in order to get a sense of where prediction accuracy is highest and lowest. To account for differences in sample size between the sale and survey estimation samples, we make two adjustments. First, when using the sales model to predict survey responses, we base the coefficient estimates on a random subsample of sales (no replacement) with sample size equivalent to that of the survey data (389, 248, or 141), estimate the model, and then fit the resulting parameter estimates to the survey observations. The resulting RMSEs represent an average over 1,000 iterations of this process. Second, when we use the survey model to predict the sales prices, to minimize the influence of outliers we generate the predictions for a random sample of (again, 389, 248, or 141) sold parcels, repeating the process 1,000 times and calculating average RMSEs.
The results of the cross-prediction exercise are presented in Table 4. There are two general observations that can be drawn from comparing cross-prediction accuracy across the different models and data samples. First, the transaction models are always a more accurate predictor of survey values than the survey models are of sales prices. In fact, the accuracy of the sales model as a predictor is at least twice that of the survey model in seven deciles for the active-market and full-sample models and nine deciles for the thin-market sample, where the survey model is an extremely poor predictor of sales prices. These results are somewhat expected, implying that survey values are implicitly based on local sales, rather than the opposite (i.e., sales prices being based on self-reported survey values). In other words, the relative accuracy of the sales-based predictions suggests that farmers may form their own land value expectations based on an implicit subjective hedonic model that compares their own parcel characteristics to those of nearby transacted parcels.
Second, the active-market predictions are generally more accurate than the full-sample predictions, and the thin-market predictions are generally the least accurate. This ordering is preserved for seven of the deciles (and the overall average) when using the survey model to predict sales prices. Similarly, eight of the deciles of RMSEs generated from the sales model, all but the first and last, exhibit this rank order. The thin-market predictions are consistently the least accurate of any sample considered. These results lend further credence to the notion that there is generally a higher degree of information transferred in more active markets and that sales prices and survey values in these areas tend to be driven by a similar set of characteristics.
Finally, it is worth noting that both the sales and survey models tend to underestimate values in the opposing data set. When predicting the survey values using the sales models, we find that 64%, 56%, and 65% of the full-sample, active-market, and thin-market cross-prediction errors, respectively, are negative. Similarly, using the survey models as predictors leads to underestimates of survey values in 62% and 63% of the full-sample and active-market cases, respectively. The cross-predictions for the thin-market survey values result in an underestimation rate of 44%, the only case with more positive than negative deviations. On balance, this implies that the factors unaccounted for in our specification are more likely to exert a positive influence on farmland value (e.g., investments that improve agricultural productivity).
5. Discussion and Conclusion
Hedonic studies of the determinants of farmland values principally draw insights from one of two possible sources of information: market transactions or opinion surveys. A number of important studies involving the value of agricultural land, particularly in the area of Ricardian climate impact estimation, examine self-reported land values. To date, researchers have produced little evidence speaking to which data source is more appropriate for hedonic analysis. This study is the first to compare observation-level individual market transactions with self-reported farmland values from the USDA’s JAS, which serves as the foundation for the USDA’s annual farmland price reporting and analysis.
The comparative assessment we conduct reveals a number of insights on the composition of farmland markets and the way in which different groups of farmer-respondents estimate the value of their land. First, we demonstrate that when survey responses and observed transaction prices are weighted by survey weights and parcel acreage, respectively, the two sources produce virtually identical statewide farmland value estimates. This finding illustrates the distortionary role played by small, development-ready parcels in farmland markets (Brorsen, Doye, and Neal 2015). Second, we find several sources of difference in the estimated implicit price of farmland characteristics obtained from sales and survey models that are consistent with recent discoveries in behavioral economics. For example, market participants and farmers have differing perceptions of the development value of agricultural land, with market prices being driven by commuting ease and income and survey values by local population level. These results are in line with those of Banzhaf and Farooque (2013), who find that, in the context of residential real estate markets, transaction prices do a better job, when compared with self-reported values, of reflecting the value of locational amenities. When we focus our analysis on surveys administered in active farmland markets, we find greater agreement in the factors that influence land values in both data sources. However, the implicit prices captured by active-market survey values are relatively high (e.g., soil quality and income) and significantly larger in magnitude than those derived from the sales data. This finding is consistent with the notion that self-reported market values may contain some degree of hypothetical bias. We leave for future research a more detailed analysis of the potential behavioral factors driving the disparity in active-market areas.
Applied economists tend to favor sales data over opinion surveys in hedonic price modeling. Our preliminary analysis and model results suggest that weighting transactions by parcel acreage may be a partial remedy for sample selection bias and is an important consideration in studies that aim to represent the broader population of farmland in a given area. The acreage-based weighting procedure is readily applicable in nearly all contexts and is less arbitrary than other methods for removing parcels not likely to be sold for agricultural use (e.g., dropping parcels that are close to cities or below a certain size threshold).
Our results have implications for applied economics research when price data are limited due to availability or market imperfections. In some cases, reconsideration of survey data may be merited. For example, if research is limited to areas in which detailed market data are available, economic knowledge may be limited or even biased toward such markets. For example, transaction price data are not available in many developing countries, but farmland market analysis still has high policy relevance (e.g., Wineman and Jayne 2018). In the case of the JAS, the panel data structure may mitigate potential bias while facilitating policy evaluation. Further, survey data may allow for analysis of factors not observable with transaction data. When feasible, comparison of price data with survey data can highlight potential biases or measurement errors associated with both sources. Our analysis provides some preliminary guidance on the use of survey data in hedonic studies. First, farmers may be better at predicting the market value of agricultural characteristics than nonagricultural characteristics. Second, the salience of different land characteristics to survey respondents should be considered. Third, farmers in active markets are better able to identify factors influencing farmland sales prices but may overestimate the influence of some characteristics.
The findings also have important implications for agricultural policy. The USDA estimates the total value of U.S. farm real estate to be $2.6 trillion, and farmland accounts for the vast majority (83%) of the value of the farm sector’s total asset base (USDA ERS 2017). It is the primary store of farm-sector wealth and an important source of collateral (Nickerson et al. 2012). There are many U.S. farm policies designed to improve access to land for beginning and economically disadvantaged farmers (e.g., direct farm ownership loans and the Transition Incentive Program), and the price of land is often cited as a significant barrier for small and beginning farmers to reach a viable scale of production. Thus, determining how and why alternative measures of farm real estate value differ from one another is critical to the agricultural sector.
Despite its novelty, there are several shortcomings and fruitful extensions of our work that should be acknowledged. Since farmland markets tend to be relatively thin, it is important to consider the applicability of our findings to the broader population of interest. This is particularly a concern for empirical analysis based on sales prices but may also apply to survey values if respondents anchor on (a few) recent transactions. In effect, this is similar to a sample selection problem, where sold parcels may differ from unsold parcels in some systematic way that is unobservable to researchers. The greatest challenge of estimating the presence of selection bias would be determining and measuring a valid exclusion restriction to use in the first-stage market participation model, or a variable that drives a parcel to be sold but does not influence the subsequent price of that parcel. Research to date provides little generalizable guidance on such an exclusion restriction, and we leave that as an important avenue for future research.
Another way to make survey and transaction data more directly comparable, and to identify the behavioral implications of price information dissemination, would be to restrict the analysis to survey responses for land parcels that are actually sold. In addition, such a survey design could consider the timing of transactions to examine the role of anchoring in survey opinion formation relative to transactions. For example, in the context of the JAS, we may expect recent transactions (around June) to receive a greater weight for survey respondents. However, the design would have to consider the lag between transaction agreements and consummation of each sales contract. With a larger data set one could also study more closely the salience of different farmland characteristics. For instance, when farmers report their land value estimates and put an implicit price on soil quality, one could determine the scale of variation (e.g., county or some prespecified areal buffer around each parcel) that governs each farmer’s heuristic hedonic function.
The external validity of our analysis may also be limited by the selection of our study area, New York State, which was driven to a large degree by data availability, its robust and diverse agricultural sector, and the array of different influences on the value of New York farmland. However, as with any individual state, New York agriculture is clearly not perfectly representative of the broader U.S. farm sector, and determining the external validity of our findings (e.g., concerning how farmers and sales market participants capitalize future development potential in different ways) will require an expanded empirical analysis. Given the increased availability of sales records from private real estate data vendors, future comparisons of alternative sources of farmland values should become more feasible.
Acknowledgments
We would like to thank Vince Breneman, Ryan Williams, Caren Kay, and Melanie Bruce for help in constructing key variables used in this analysis. We also thank Julie Mueller and conference participants at the 2018 World Congress of Environmental and Resource Economists, seminar participants at Economic Research Service, and David Just for valuable feedback on an earlier version of this paper. This research was supported in part by the USDA National Institute of Food and Agriculture, Multistate Project 1007199.
Footnotes
↵1 As an example of the generally thin nature of farmland markets, Bigelow, Borchers, and Hubbs (2016) note that just over 2% of farmland was anticipated to be exchanged in an arm’s-length transaction in the 2015–2019 period. For context, this compares to a turnover rate of roughly 2% per year in residential property markets. The thinness of farmland markets is also evidenced by the long holding periods (more than 20 years) for most farmland owners (Zhang, Plastina, and Sawadgo 2018).
↵1 The JAS and census questionnaires differ slightly in terms of how they solicit land value estimates from farm operators. For one, the JAS asks respondents to provide separate estimates for the value of land and total farm real estate, though the questions pertain to different geographic portions of the farm operation in question. For the JAS questions related solely to farmland, respondents provide separate estimates for the value of pasture and cropland (and, further, for dryland and irrigated cropland in some states). Importantly, the JAS questions relate to land in the farm operation within the JAS segment. The census questionnaire, in contrast, is split out by land tenure, with separate questions for land owned and operated, rented into the operation, and rented out by the farmer to a different operation. County-level farm real estate value estimates published by USDA, however, aggregate the values across these three tenure categories.
↵3 Hypothetical bias and the disparity between willingness to pay and willingness to accept have been studied extensively in the stated preference literature (e.g., Johnston et al. 2017; Tuncel and Hammitt 2014). We are not aware of any systematic studies of these issues in the context of hedonic property value analysis or revealed preference valuation more broadly.
↵4 Farmland is assessed at a preferential rate in all 50 states (Anderson and England 2014). Most states rely on use-value assessment, under which the assessed value of farmland is based solely on its ability to generate agricultural revenue. This limits the insight that can be gleaned from a comparison of assessed values and sales prices in the United States. In contrast to assessed values, the survey values used in this study are meant to represent the price that a parcel of farmland would be worth in an arm’s-length transaction.
↵5 Although the tract is the most disaggregated unit available within the JAS, we are not able to georeference individual tracts, which precludes us from linking them with external data sources (e.g., spatial data on soils). However, we do observe the centroid of each JAS segment. The segments, therefore, form the unit of analysis for this study. As noted by an anonymous reviewer, it is possible that the use of segments as opposed to tracts may create a form of measurement error, but it is difficult to gauge the magnitude of the potential bias this creates since we are unable to geolocate the specific tracts associated with each segment.
↵6 The sample size is also larger for years in which the Census of Agriculture is administered. As it pertains to this study, 2012 has a larger cross-sectional sample than other years (2009–2011 and 2013–2014).
↵7 The survey in Figure 1 is the 2014 version given to farm operators in Missouri. In states with a non-negligible amount of irrigated cropland, separate land value questions are asked about the irrigated and nonirrigated portions of the tract. In New York, and other states where a large share of cropland is nonirrigated, a single cropland value question is used.
↵8 The JAS also collects data on farm real estate value, which reflects the value of both land and buildings. However, the farm real estate value data are collected at the farm level, as opposed to the tract level at which the land value data are collected. Farms, as they pertain to the JAS, extend outside the boundaries of JAS segments, which hinders our ability to link spatial data sources to JAS farms with any degree of confidence.
↵9 More information on USDA-NASS’s JAS can be found at https://www.nass.usda.gov/Surveys/GuidetoNASSSurveys/JuneArea/.
↵10 To maintain comparability across data sets, we also remove JAS tracts with a reported land value of less than $100/acre or greater than $50,000/acre and JAS segments located on Long Island.
↵11 We note that the terms “high-quality,” “medium-quality,” and “low-quality” are defined relative to the soils throughout New York, which likely differ substantially, in terms of quality, from the soils found in other regions. In addition, although our use of a soil index specific to New York is an improvement over the coarser nationwide indices more commonly used, prior research has shown that perceptions of soil quality may vary within states (Zhang and Duffy 2017).
↵12 The PRISM climate data are available at www.prism.oregonstate.edu/.
↵13 Documentation for the population interaction index variable, which is based on a refinement to an earlier, coarser measure, is available at https://www.ers.usda.gov/data-products/population-interaction-zones-for-agriculture-piza/documentation/.
↵14 We are unable to report or compare minimum or maximum values in order to maintain JAS respondent confidentiality.
↵15 The normalized mean difference for variable j is measured as , where and denote the means for samples 1 and 2, respectively, and νj1 and νj2 represent the corresponding sample variances. In the context of matching applications, Imbens and Wooldridge (2009) note that NMDs in excess of |0.25| can be problematic.
↵16 All transaction prices and survey land values are adjusted to 2014 dollars using the GDP implicit price deflator compiled by the Bureau of Economic Analysis.
↵17 In the context of housing markets, Parsons (1990) argues that lot-size weights should be applied to explanatory locational attributes in order to properly account for the “user cost”associated with occupying bigger or smaller lots. For a comprehensive general treatment of weighting in econometric models, see Solon, Haider, and Wooldridge (2015).
↵18 Rosen (1974) provides what is generally considered to be the original derivation of the hedonic equilibrium for consumer goods, while Palmquist (1989) covers the case of derived demand for a differentiated factor of production, such as farmland. To avoid redundancy, we do not replicate the derivations from these two seminal papers here. We point the reader to Nickerson and Zhang (2014) for an overview of contemporary issues in hedonic models of farmland value.
↵19 As recommended by Solon, Haider, and Wooldridge (2015), we provide unweighted results for our main specification in Appendix Table A3, which we discuss briefly in the robustness checks section.
↵20 To create a sharper dichotomy based on local market activity, with a larger data set one could, for example, compare the upper and lower third of each sample based on county-level transaction volume.
↵21 An alternative would be to cluster the pooled model standard errors by county to allow for within-county cor-relation across the two data sources. We note that doing so results in no meaningful difference in the inferences based on the model. Treating the standard errors this way links the error terms from estimation of [1] separately for the sales and survey data, which is similar in spirit to the seemingly unrelated regressions (SUR) estimation technique of Zellner (1962). The main distinction between our approach and SUR is that we do not have any overlap between the sales and survey samples. As we discuss in the conclusions, a potentially fruitful area for future research would be to conduct an analysis similar to ours for surveyed parcels that are sold, which would allow for a direct application of the SUR estimator.
↵22 A higher degree of activity within a local farmland market does not, however, necessarily imply that the market is more competitive, as bargaining power and locational constraints (e.g., Cotteleer, Gardebroek, and Luijt 2008) can create market power opportunities for specific types of buyers and sellers. We thank an anonymous reviewer for pointing this out.
↵23 Roka and Palmquist (1997), however, provide evidence that direct market experience (selling land in the previous year) has no effect on self-reported farmland value opinions.
↵24 Since the dependent variable is in natural log form, we interpret the survey dummy variable coefficient using Kennedy’s (1982) method.
↵25 Since the dependent variable in all of our models is in natural log form, we follow Wooldridge (2012) in forming the predicted prices and correct the fitted values with an adjustment term derived from an auxiliary regression of the actual values on the predicted values in levels.