Abstract
Farmland provides agricultural products and natural amenities, as well as residential sites. The emergence of exurbanization appears to be changing the demand for natural amenities and their role in determining land values. To better understand how appraised value and sale price capture the determinants of farmland value in a region facing exurbanization, this study applies the hedonic method to land transaction data in southwestern Michigan during 2003–2007. Results suggest that appraised values are a poor substitute for sale prices if the research goal is to understand dynamically evolving determinants of land value in exurbanizing regions, especially the value of natural amenities. (JEL Q24, Q51)
I. INTRODUCTION
Agricultural land is increasingly recognized as offering more varied services than simply to produce food, fiber, and fuel. As an input for production, agricultural land is chiefly managed for the supply of crops and livestock. As a home to farm families, agricultural land has always offered certain natural amenities (e.g., forests, lakes, and wetlands). The role of these natural amenities is now widely recognized. In the public arena, farmers and their lands are called upon by governments or nongovernment organizations, through economic incentive programs, to provide environmental services that offer benefits on and off the farm (Swinton et al. 2007). In the private arena, agricultural land in the exurban fringe is often sought after for its natural amenities by individual land purchasers. These changing demands on agricultural land can be expected to be reflected in its value.
The economic value measured by land sale prices is an important indicator of the underlying value that people place on the services that land can provide. Hedonic analysis of rural land prices is an important tool for estimating the economic value of underlying natural and built features, as revealed by market transactions. If rural land values are indeed changing in response to evolving demand drivers, it becomes important to understand both how the determinants of land values are changing and what methods can be used to measure those changes efficiently and without bias.
Exurbanization is a particularly important driver of changes in the demand for rural land, including agricultural land. Exurbanization is a special form of urban development across the continental United States (Nelson and Dueker 1990). Unlike “urban sprawl” that occurs close to metropolitan areas, exurbanization or ‘‘rural sprawl’’ is a pattern of development that is “decreasingly linked by proximity to urban centers and increasingly driven by access to open space and recreational opportunities” (Brown et al. 2005, 1858). Exurbanized area grew nearly sevenfold from 1950 to 2000 in transitional metropolitan counties, and nearly tenfold in counties adjacent to metropolitan counties (Brown et al. 2005). According a Bureau of the Census report for 2000–2007, exurbs are growing much faster than the central counties (Mackun 2009).
Given the prevalence of amenity-driven exurbanization on agricultural lands, how can we best examine these processes with existing data sources? As the revealed expression of the economic value of land, sale price is the gold standard from which to analyze land value determinants. However, three major problems detract from the appeal of using land sale prices in hedonic analysis of the underlying values determining farmland value. First, farmland changes hands infrequently, so sale price data can be very sparse. Thus, the number of property sales needs to be fairly large in order to achieve reliable results with regression methods (Birch and Sunderman 2003). Second, the reported prices of some land transactions may be biased, further reducing the number of valid observations available. For example, when buyer and seller are related, farmland sales are not conducted at arm’s length, often resulting in sale prices significantly below market levels. Seller financing may also cause a downward bias in sale price when the price reported is only the amount of money changing hands at the time the deal is signed (Pollakowski 1995). Third, recording errors may occur even for arm’s-length transactions, such as recording the address of the seller rather than that of the property being sold, or reporting false land-use codes (Pollakowski 1995).
Appraised values prepared for property tax assessment offer a potential solution to the shortcomings of land sale price data. Almost every property in the United States is assessed for tax purposes, so appraised values are largely available and well maintained by the tax offices. Several studies used both sale price and appraised value to analyze land value determinants.
Most studies comparing real estate sale price with assessed value have analyzed residential properties. Kim and Goldsmith (2009) compared assessed property value and sale price in capturing the impact of swine production on rural residential property values in southeastern North Carolina. Their results revealed that assessed value more effectively captured locational and environmental effects, although physical attribute estimation was comparable. Schuler (1990) analyzed residential property values in southeastern Florida using both appraised value and sale price. While appraised value was found to reflect the market values of individual house variables (e.g., living and nonliving areas), it tended to overestimate neighborhood quality via the assessed value of comparable properties. Rush and Bruggink (2000) measured the value of ocean proximity on barrier island houses in New Jersey, using both assessed value and sale price as the dependent variable. They found that the assessed value results generally resembled the market model, except for underestimation of the contribution to value from central air conditioning and the number of bathrooms. Other studies combined sale price and appraised value data, assuming they are well correlated. Clapp and Giaccotto (1992) proposed an assessed value method to estimate the value of property in which appraised value at time zero adjusted by assessment equity effect was used as an independent variable in addition to year dummies for sale price variation in different years. Dornbusch and Barrager (1973) investigated the effect of water pollution control on residential, recreational, and rural waterfront property values in the United States. They used the difference of pre-control sale price and post-control tax assessment as the dependent variable, with only validated assessed values taken into account.
Only one study compared these two measures of land value for agricultural land. Grimes and Aitken (2008) used both farmland sale price and the land tax valuation as dependent variables, respectively, to investigate the value of water in a drought-prone farming region in New Zealand. They found that the tax valuation data was a reasonable proxy for market value but possibly understated the value of irrigation. However, this study concentrated on the production function of land in a stable agricultural region. Hence, it did not provide a sound empirical test of sale price versus appraised value in a region that is undergoing a process of land use change like exurbanization. Determining whether appraised value and sale price differ in their ability to capture the influential factors of farmland value in a region facing exurbanization requires a suitable dataset from such a region. In the remainder of this paper, we propose a broader conceptual model of agricultural land price determination and apply the hedonic method to sale price and appraised value data during 2003–2007 near metropolitan areas in southwestern Michigan.
II. CONCEPTUAL MODEL AND HYPOTHESES
Conceptual Model for Agricultural Land Valuation
Farmland is an asset whose economic value emerges from three broad classes of service flows that it can yield (Henneberry and Barrows 1990; Xu, Mittelhammer, and Barkley 1993). First, as a medium for agricultural production, farmland enables livelihoods and serves as a store of wealth. Second, as a home site, farmland offers access to open space and pursuit of a country lifestyle. Rural country living offers outdoor recreation opportunities, such as fishing, hunting, and scenic views. Third, apart from agricultural production and natural amenity consumption, agricultural lands also have asset option value to property developers interested in future nonfarm uses. Thus, the value of agricultural land can be generally estimated from these production, consumption, and asset roles. Given the important differences between natural capital and built capital as asset classes (Daily 1997; World Bank 2006), we partition the production and consumption roles into built assets (B) and natural assets (E).
We develop a conceptual model for farmland valuation in a market setting (Ma and Swinton 2011). The land purchaser chooses land parcel M to maximize a utility function whose arguments include built attributes with consumption appeal Bc (e.g., on-site residence), environmental amenity attributes Ec (e.g., recreational and aesthetic resources), and other purchased goods N (equation [1]). Let P(Z) be the parcel price, a function of hedonic price determinants, Z. The budget constraint requires that the land purchase expense P(Z)M(T) plus the expected present value of all other purchased goods expenditures, PNN, not exceed the expected present value of cumulative future income from agricultural production, πb, and nonfarm activities, NFI, plus the option value for future development, π(L,s), where L represents factors related to farm location, and the buyer’s planning horizon runs from t = 0 to t = s, where s is the period of land conversion from agricultural to nonagricultural use. T represents the land transaction attributes, and r is the discount rate.
[1]The buyer’s expected future profit from production in period t, πb, equals farm product revenue PyY minus variable cost PxX and fixed cost FC (equation [2]). The production of Y is facilitated by environmental resource attributes Ep (e.g., pest population regulation by natural predators), productivity-enhancing factors from prior human investments Bp (e.g., land improvements), and other variable inputs, X.
[2]The land seller (equation [3]) is assumed to maximize the net return from the land sale, with sale transaction costs depending on the land parcel (M).
[3]The agricultural land attributes Z are made up of consumption-related environmental characteristics (Ec), production-related environmental characteristics (Ep), consumption-oriented built characteristics (Bc), and production-oriented built characteristics (Bp) (equation [4]). Both buyer and seller are assumed to face location (L) and transaction (T) traits. Levels of all the above variables are conditioned in part by product and input prices.
[4]Hypotheses
Using this framework, the hypotheses to be empirically tested center on the relative importance of the determinants of sale prices, and the suitability of appraised values as proxy variables for sale prices. In null form, the hypotheses to be tested are these:
Hypothesis 1. All five categories of asset variables in equation [4] are drivers of land values, except for sale transaction type (T), which is absent from the appraised value (AV) model.
This hypothesis tests whether agricultural land value stems from its roles as a production input, source of natural amenities, or asset, and whether both built assets and natural assets are valued.
Hypothesis 2. The sale price (SP) and AV models do not differ, except in T.
a. The SP and AV models do not differ in explanatory power.
b. The SP and AV models do not differ in the driving variables, except for sale transaction type, which is absent from the AV model.
This hypothesis aims to test whether appraised value is as good as sale price in capturing underlying determinants, and whether the two land values are influenced by same factors.
III. STUDY AREA AND DATA
This study focuses on four counties (Allegan, Barry, Eaton, and Kalamazoo) in southwestern Michigan, where all properties lie within 50 miles of a metropolitan statistical area as defined in the 2000 census (Figure 1). Lake Michigan is located to the west of Allegan County. Major cities such as Grand Rapids, Lansing, and Kalamazoo are located in or close to the four counties (Figure 2).
Study Area in the State of Michigan
Detailed Locations of Parcels in the Four Study Counties
Source: Ma and Swinton 2011
This area of Michigan is not only suited for cropland and pasture for agricultural production, but also endowed with abundant natural recreational and aesthetic amenities. Its productive soils and microclimates, due to rolling topography and the surrounding Great Lakes, have created favorable conditions for agriculture (MDA 2003). However, the role of agriculture is changing. From 1950 to 2007, Michigan farmland decreased from 17.3 million acres to 10.0 million acres, while the number of farms dropped from 155,500 to 56,000. In the four counties studied, 0.44 million acres of farmland was lost between 1950 and 2007, of which 54% was in cropland. The average percentage of farmland in the four counties decreased from 79% in 1950 to 50% in 2007 (USDA 1950–2007).
Two major factors have contributed to the use of agricultural land for nonfarm development. First, population statistics indicate that the number of households in the area is increasing, while at the same time, the average size of a household is decreasing. The conversion of agricultural land from production to residential use has occurred as former urban dwellers move out to the suburbs and rural areas (MDA 2003). Second, the soil and water resources that result in the agricultural advantages also make this area a desirable place to live and to recreate. The construction of secondary residences for recreation and retirement has increased, and they directly compete with land for farming (MDA 2003).
Lands in the study area are used for a mix of agricultural production, residence, and recreation. The proximity of small cities creates opportunities to develop land for commercial and industrial use. Thus, this area is a good representation of the three functions of agricultural land, and hence a good area for testing the ability of alternative datasets to discern determinants of land value in a dynamic setting. According to the 2007 Michigan Land Value Survey, the major agricultural factors that influence land prices in southwestern Michigan are grain price and farm expansion, while the nonagricultural factors include home sites, hunting access, water access, and interest rates (Wittenberg and Harsh 2007).
Data for this study were collected from three major sources. Land transaction and parcel information, including land price, appraised value, sale date, contract type, and land class, were collected from county equalization offices, which administer local property taxes. Geo-referenced parcel maps were obtained from county GIS offices. Other soil and land cover variables were constructed with ArcGIS software using several GIS databases, including the U.S. Department of Agriculture Soil Survey Geographic Database (SSURGO), the Conservation and Recreation Lands dataset, the National Oceanic and Atmospheric Administration Coastal Change Analysis Program (C-CAP) database, and other Michigan GIS data on rivers, lakes, wetlands, cities, and major roads. From a total of 337 parcel transactions (see Figure 2 for the distribution of parcels) during 2003–2007 (Ma 2010), a common dataset of 203 parcel sales is used for comparing the determinants of sale price and appraised value.
IV. EMPIRICAL MODEL
Hedonic Method
Hedonic analysis is a powerful revealed preference method for nonmarket valuation of the environment and natural resources. Based on Lancaster’s (1966) view of utility as generated by the underlying characteristics of goods, Rosen (1974) provided the classic theoretical foundation for the hedonic model by exhibiting individual choices in market equilibrium. Let Z = (z1, z2, …, zn) denote n attributes of a differentiated market good. In a perfectly competitive market with a sufficient number of goods, the equilibrium price p can be determined by the interaction of utility-maximizing consumers and profit-maximizing producers. The fundamental hedonic equation is p = h(Z), where h(·) represents the relationship between a good’s price and its attributes. h(Z) can take diverse functional forms. Regressing the observed price p on all attributes of the good, ignoring differences in supplier or consumer characteristics, we can obtain an estimated marginal value 
of each attribute as depicted in the equation 
, which is the marginal price that people would pay for a small change in attribute i. The goal of this study is to compare the marginal value of each land production/consumption attribute revealed by estimating the hedonic price equations, using land sale price and appraised value, respectively.
Functional Form
There is little theoretical basis for choosing the functional form of a hedonic regression. The Box-Cox transformation, a general and flexible class of functions, is widely applied in empirical hedonic analysis. For both the sale price model and the appraised value model, Box-Cox tests clearly rejected the linear and reciprocal forms, with p-values close to zero, but could not reject the natural logarithmic transform at the 0.01 level. Hence, we use a semilog model with natural logarithm of farmland value per acre (P) regressed on the vectors of untransformed independent variables (Bp, Bc, Ep, Ec, L, T). The farmland value is changed from a single variable in the conceptual model to a vector in the empirical specification to represent individual land sales in the sample. Two models are estimated following equation [5], with real sale price and appraised value as the dependent variables, and the βj are vectors of coefficients to be estimated.
[5]Variables
As noted above, the per-acre sale price and appraised value are used as dependent variables P in two separate models. The sale price is the transaction price for all available arm’slength sales, including warranty deeds1 and land contract deeds,2 while excluding quitclaim deeds.3 The appraised value used in this study is known as the state equalized value (SEV), which is especially used in Michigan for property tax purposes. According to the 1994 constitutional amendment known as Proposal A, the assessed value of each real property is determined by local assessors, based on the condition of the property on December 31 of the previous year, and is normally at 50% of the estimated market value (referred to as true cash value). The SEV is adjusted from the assessed value following county and state equalization procedures. SEV is approximately equal to 50% of the sale price if there was a transfer of ownership in the previous year; otherwise it is mainly determined by local assessors following a mass appraisal technique. The SEV data was available only in year 2007 or 2008. Both sale price and SEV were deflated to 2007 constant prices using the Prices Paid by Farmer Index (NASS 2008).
Following the conceptual model, the explanatory variables include vectors of built features (Bp, Bc), environmental resources (Ep, Ec), location (L), and sale transaction (T) to estimate their marginal contribution to farmland value (Ma and Swinton 2011). Variables representing the level of environmental resources (Ep, Ec) were constructed from measures of natural resources and landscapes. Descriptive statistics for all variables can be found in Table 1. The production of crops and livestock on farmland is directly related to the tillable area on the land parcel, specifically the CULTIVATED LAND AREA AS PERCENTAGE OF PARCEL for crops and PAS-TURE AREA AS PERCENTAGE OF PARCEL for livestock. To measure the influence of proximate farm lands, the variable CULTIVATED LAND AREA AS PERCENTAGE OF SURROUNDING AREA is calculated from a 1.5 km radius from the parcel boundary. Nearby land not cultivated may provide natural habitat offering services of biological pest control (Gardiner et al. 2008; Thies, Steffan-Dewenter, and Tscharntke 2003) and pollination (Kremen et al. 2004; Steffan-Dewenter et al. 2002) that aid agricultural production. Nearby natural areas may also harbor game animal species that could provide recreational opportunities and/or cause crop destruction. Land productivity is proxied by binary variables for FARMLAND CLASSIFICATION from the SSURGO database. Class 1, the baseline omitted from the economic model, is “All areas prime farmland.” Class 2 is “Prime farmland if drained,” Class 3 is “Farmland local importance,” and Class 5 is “Not prime farmland.”4 Soil erosion potential is calculated as the parcel-area weighted average SOIL LOSS TOLERANCE FACTOR, the maximum average annual rate of soil erosion by wind and/or water. Soil drainage is categorized by intervals from the parcel-area weighted average value of a drainage index (1–99) (Schaetzl 1986). The WELL DRAINED dummy has an index value between 34 and 65, the POORLY DRAINED dummy has an index value above 65, and overdrained soils have an index value below 34. On-site water resources are captured by RIVER LENGTH and LAKE AREA AS PERCENTAGE OF PARCEL. Variables for water recreation and irrigation potential,5 DISTANCE TO RIVER and DISTANCE TO LAKE, are the straight-line distance from the parcel centroid to the edge of the nearest river or lake. WETLAND AREA AS PERCENTAGE OF SURROUNDING AREA indicates wetland roles for water-holding capacity and wildlife habitat, while CONSERVATION LAND PERCENTAGE OF SURROUNDING AREA6 serves a similar purpose. GRASSLAND AREA AS PERCENTAGE OF PARCEL and FOREST AREA AS PERCENTAGE OF PARCEL indicate recreational and habitat land use potential inside the parcel. The influence of managed recreational land7 is measured by its distance from the parcel centroid along roads, labeled as RECREATION LAND DISTANCE.
Summary Statistics for Sale Price (SP) Model and Appraised Value (AV) Model, 203 Land Parcels, Southwestern Michigan, 2003–2007
The production-oriented built attributes (Bp) include basic land properties like TOTAL ACRES and DEGREES OF SLOPE, which is the area-weighted average of representative land slope from SSURGO. The baseline LAND USE CLASS is crop production, with dummy variables for livestock production and residential or hobby farm. Building attributes (Bc) are constructed from aerial photographs8 combined with information from county tax offices. BUILDING PERCENTAGE is the proportion of parcel area covered by buildings and accessories. The NUMBER OF AGRICULTURAL BUILDINGS and NUMBER OF RESIDENTIAL BUILDINGS capture building numbers by type.
Attributes related to location (L) and transactions (T) serve as conditioning variables in the model for a complete specification. The option value of nonfarm development is in-dicated by measures of surrounding urban area and distance to cities. Binary dummy variables identify the closest major city to each parcel (GRAND RAPIDS, LANSING, KALAMAZOO, HOLLAND, or Battle Creek [omitted baseline]), each of which has a population greater than 35,000. The DISTANCE TO MAJOR CITY variable measures the straight-line distance to the closest city, while the DISTANCE TO MAJOR ROAD variable measures the straight-line distance from the parcel centroid to the edge of the nearest highway. The variable DEVELOPED LAND AREA AS PERCENTAGE OF SURROUNDING AREA indicates nearby urbanization within a 1.5 km radius. Binary variables are also included for sales year, month, and transaction instrument type (LAND CONTRACT rather than base case of warranty deed). See Table 1 for summary statistics.
Spatial Autocorrelation
As the land parcels are spatially ordered data, the spatial dependence among observations cannot be ignored. According to Anselin (1988), spatial dependence is “the existence of a functional relationship between what happens at one point in space and what happens elsewhere.”We test for spatial autocorrelation using Moran’s I. A positively significant I indicates the variable value at each location i tends to be similar to the values at spatially nearby locations, whereas a negatively significant I indicates spatially different trends among locations (Anselin and Hudak 1992). The global spatial autocorrelation by Moran’s I test indicates significant spatial autocorrelation for 37 of the 50 variables in the SP model and 29 of the 34 variables in the AV model. The local spatial autocorrelation test suggests that 12 observations in the SP model and 14 observations in the AV model have highly significant (p-value < 0.2) spatial autocorrelations.
We model spatial relationships using a spatial weights matrix, W, comprised of inverse distance weights. Let dij denote the distance between parcel i and parcel j, the matrix element wij = 1/dij if dij < c, and wij = 0 if I = j or if dij > c, where c is the cutoff point for spatial autocorrelation. This weighting approach implies that observations located closer to each other are more highly correlated than observations further apart. As the distance between parcels increases, the correlation weights become smaller. Beyond the cutoff point, no correlation is assumed (Lynch and Lovell 2002).
The Moran’s I spatial correlogram for these datasets suggests significant spatial correlation among observations in the distance bands of 200–800 m for the SP dataset and 400–1,000 m for the AV dataset. Hence, we generated an inverse distance weighting matrix with a cutoff point of 600 m from the centroid of each parcel. Similar cutoff points have been used in other hedonic studies. Lynch and Lovell (2002) studied the easement payment on farmland in Maryland and set the cutoff at 490 m from parcel centroid.
Diagnosis of the structure of spatial autocorrelation in the SP and AV datasets suggests spatial dependence is due only to correlation in the error terms of the two models, implying spatial error structure, rather than a spatial lag in the dependent variable (Ma 2010). Hence, to control for spatial autocorrelation, we estimate a spatial error model, as follows:
[6]
[7]where λ denotes the spatial autoregressive parameter, μ denotes a vector of homoskedastic and uncorrelated errors, and all the other terms are defined as above.
Based on the diagnosis, both maximum likelihood and generalized spatial two-stage least squares can be used to estimate the spatial error model (Anselin 1988; Kelejian and Prucha 1998). Since the eigenvalue matrix cannot be computed because more than a half number of parcels have no neighbors within 600 m, the spatial error model by maximum likelihood estimation cannot be implemented. Hence, we estimate the model by ordinary least squares using a Stata code by Conley.9 The results that follow are from models corrected to be robust to spatial autocorrelation.
Regression Diagnostics
Regression diagnostics led to some adjustments in the data and econometric model. Pairwise correlation and variance inflation factors showed evidence of multicollinearity, leading several variables to be dropped, including forest and grassland in parcel neighborhood, county dummies, and some soil productivity measurements (Ma 2010).We retain FOREST AREA AS PERCENTAGE OF PARCEL and WETLAND AREA AS PERCENTAGE OF PARCEL for complete specification, though they are negatively correlated with two other variables and they raise the variance inflation factor. The joint F-test for dropped variables was insignificant at the 0.05 level, indicating that the coefficient estimates on all dropped variables are jointly equal to zero. Breusch-Pagan tests of heteroskedasticity were insignificant (Ma 2010).
V. RESULTS AND DISCUSSION
The appraised values are positively correlated with the sale prices; however, the correlation coefficient is only 0.3. Although the two hedonic models share a core set of determinants, they differ both in overall determination power (R2 square) and significant explanatory variables (Table 2).10
Comparison of Results between Sale Price (SP) Model and Appraised Value (AV) Model, 203 Land Parcels, Southwestern Michigan, 2003–2007
Comparison of the two models first reveals that the AV model gives a higher R2 (0.7) than the SP model (0.3). The large difference leads to rejection of Hypothesis 2a of equal explanatory power between the two models. This result is consistent with previous studies. Using unequal samples of SP and AV data, Grimes and Aitken (2008) and Kim and Goldsmith (2005) found higher coefficients of determination for the AV models. Using the same dataset for residential properties in both AV and SP models, Schuler (1990) also found the assessed value as the dependent variable achieved higher R2 (0.88) than the model using sale price (0.80).
The higher coefficient of determination from the AV model suggests less variability among appraisers’ value estimates than among transactions in the land market. The first source of differential variability in our samples is that far fewer individuals were involved in land appraisal (33 township appraisers) than in land sales (buyers and sellers for 203 parcels). Those appraisers tend to conduct land tax assessment in the same manner, as they have gone through standard training. Fixed-effects regression of these recent, inflation-adjusted appraised values on sale prices also finds township dummy variables to be significant. The second source of variability is different determinants of value. Appraisal is based on the application of formulas that place weight on certain major factors, such as on-site buildings, soil productivity and erosion, and land use class, whereas more factors are taken into consideration by land buyers and sellers. This can be seen from the different significant variables discussed below.
Coefficient estimates for explanatory variables in the two models are compared in Table 2. The second column indicates whether the coefficient estimate is significant at the 0.10 level in the SP model, AV model, or both (SP/AV). Inspection of the table reveals significant coefficient estimates associated with all five categories of explanatory variables in both models, so Hypothesis 1 cannot be rejected.
Closer inspection of the number and types of coefficients across the two models is required to evaluate Hypothesis 2b that the explanatory variables will be the same between the SP and AV models. Certain variables have significant coefficient estimates with the same signs in both models. Larger total parcel acreage decreases both sale price and assessed value by 3% per 10 acres increase in area. The proportions of conservation land and pasture in the parcel neighborhood have positive effects in both models. However, the effects from those environmental variables are less in the AV model than in the SP model. The relative magnitudes of major city influence from Lansing, Kalamazoo, and Grand Rapids are the same in both models, but the absolute effects are larger in the SP model. Land with well-drained or poorly drained soil has higher value than land with over-drained soil in both models.
However, large price effects of environmental amenities that exist in the SP regression are not significant in the AV model. Of the Ep and Ec variables, 13 have significant coefficient estimates in the SP model versus only 6 in the AV model. In particular, the following five variables have significant price effects due to environmental amenities in the sale price model, none of which are significant at the 10% threshold in the appraised value model. On-site lakes are found to have the greatest effect, raising land price by 4% per 1% increase in parcel area made up of lake surface. This large effect may reflect aesthetics of scenic views and natural recreational opportunities for swimming, fishing, and boating. In contrast, on-site streams reduce land price by 1% per 10 m of stream length within the parcel, as disamenities such as flooding, erosion, parcel partition, and reduction of productive area outweigh possible benefits. Forested land increase price by 1% for each 1% increased in its on-site area share. Wetland in surrounding areas also increase land value by 3% per 1% increase in its proportional share. The price effects associated with forested lands and wetlands are likely due to recreational and aesthetic values. However, the amenity from on-site wetlands is likely to be offset by the unfavorable effect on crop production, leading to an insignificant effect in the SP model. Interestingly, land designated as not prime farmland fetched 50% higher sale prices than other land, presumably due to potential for nonfarm use. All in all, fewer environmental factors matter in the AV model, and their effects are less pronounced. These findings lead us to reject Hypothesis 2b and conclude that AV models underestimate the effect of environmental variables, which implies that using appraised values can lead to biased interpretation of land value determinants.
Land uses related to production and residence also vary between the two models. One percent increases in area of cultivated lands within the parcel and its neighborhood boost land sale price by 2% and 1% respectively, but have no significant effect on assessed value. The area and number of buildings play more important roles in the AV model, whereas estimated values are negligible for on-site forestlands, croplands, and pastures. For example, an additional residential building raises assessed value by 22% but has no statistically significant effect in the SP model.
Based on the comparison, it seems that assessed value mainly depends on physical productive and residential farmland features, such as total acres, buildings, and soil. By contrast, sale price also captures on-site and off-site resources and landscapes that reflect amenity values, such as water bodies and forests. This distinction is derived from the data generation process of sale price and appraised value. In this study, the appraised value is the tax assessment appraised from existing market data by the county equalization office.
Two features distinguish taxable value assessment from other types of farm appraisal. First, little time is spent on each parcel, because all parcel assessments in a county usually have to be completed by a certain date and because costs for detailed examination of each parcel would be prohibitive. Second, successful farm tax appraisal depends more on uniformity—the relative value of one parcel compared to another—rather than on each farm’s absolute value (Murray et al. 1983). Thus, assessed value is normally appraised on certain principal factors following a mass appraisal technique. A typical process includes (1) soil ratings, determined from soil surveys, multiplied by base values, which are generated from benchmark appraisals using the sales comparison approach; (2) building value, estimated by the cost of each building component; (3) further adjustments, taking into account factors such as natural resources, topography, erosion, drainage, location, roads, market, water supply, physical features, and nuisances. According to a Michigan farm appraiser, 11 land tax assessment is based on market data collected by the equalization department in every county and then is applied through a factor for each taxing division within the county. Therefore, it tends to reflect an overall picture of the market and principal land value determinants. In contrast, sale prices reflect individual market participants’ perceptions and expectations and, hence, can better capture subtler, more subjective effects, such as amenity values. A recent survey of land buyers and sellers supported this conjecture by finding that land with perceived amenity values and development potential tended to fetch a higher sale price (Deaton, Hoehn, and Norris 2007). It may also be that in an exurbanizing setting where the determinants of rural land purchase are changing, there is a lag between the value determinants applied by taxable value appraisers and current conditions in the land market.
VI. CONCLUSION
The ready availability and abundance of assessed value data makes a seemingly attractive proxy for sale price data in hedonic land value studies. Although the major categories of variables are captured in both AV and SP models, the AV model appears to underrepresent the role of environmental amenities. For agricultural land in southwestern Michigan, the appraised value places relatively greater emphasis on built and agricultural production–related attributes, such as farmland total acres, soil properties, and buildings. The sale price reflects these values as well, but it also captures more amenity values from on-site and off-site water resources and landscapes perceived by individual market participants. The appraised value from tax assessment appears not to be a good substitute for sale price if the research goal is to understand the determinants of land value, especially amenity value. These results contrast with the finding by Grimes and Aitken (2008) of negligible difference between sale price and appraised values in identifying the determinants of land value in a stable agricultural zone. The key factor explaining the difference between our findings and theirs appears to be that Michigan rural land use is not stable. Rather, it is evolving from agricultural to residential.
Comparison of our two hedonic land value models suggests that where land use is changing, land appraisal formulas may obsolesce and fail to capture fully the contemporary determinants of land price. Exurbanization is an example of changing land use, where agricultural lands previously valued for production attributes begin to be sought for residential purposes. As residential demand for rural land grows from nonfarmers, the implied values from natural amenities, such as nearby water bodies and forest, play a greater and greater role in shaping land prices. Given these evolving changes in the relative importance of land value determinants, appraised values may underestimate the contributions of natural amenities, due to the overreliance on past land prices that were more heavily influenced by agricultural productivity and investment factors.
Acknowledgments
The authors gratefully acknowledge financial support from the “KBS-LTER Project: Long-Term Ecological Research in Row-crop Agriculture” of the National Science Foundation (NSF #0423627), and a 2008 NSF social science supplement to the KBSLTER for “Estimating the Value of Ecosystem Services Embodied in Agricultural Lands.” The authors also thank Frank Lupi, Larry Leefers, and anonymous reviewers for valuable comments, and Sarah Ac-Moody for GIS assistance. This research was conducted while the first author was a graduate student, Department of Agricultural, Food, and Resource Economics, Michigan State University, East Lansing.
Footnotes
The authors are, respectively, postdoctoral research fellow, Woods Institute for the Environment, Stanford University, Palo Alto, California; and professor, Department of Agricultural, Food, and Resource Economics, Michigan State University, East Lansing.
↵1 In a warranty deed, the grantor promises that the grantor has good title to the land, can transfer title to the land, and can deliver possession of the land to the grantee.
↵2 In a land contract, the price is paid in periodic installments by the purchaser, who is in possession of the property. The vendor and vendee each have an interest in the property until final payment is made.
↵3 A quitclaim deed conveys all of the right, title, and interest that the grantor had in the land at the time of the transfer, without warranting or professing the validity of the grantor’s claim. It is often used for transfers between family members, gifts, placing personal property into a business entity, to eliminate clouds on title, or in other special or unusual circumstances.
↵4 Natural Resources Conservation Service policy and procedures on prime and unique farmlands are published in the Federal Register, Vol. 43, No. 21, January 31, 1978 (available at http://soils.usda.gov/technical/handbook/contents/ part622.html).
↵5 According to a 2009 Michigan Department of Agriculture report on irrigation water use, irrigation is needed for some high-value crops in Michigan during July and August, when rain-fed crops often suffer from a moisture deficit. The primary source of water for agricultural irrigation in Michigan is groundwater (75%), with the remainder withdrawn from surfacewater sources (available at www.michigan.gov/documents/MDA_Cranberry_GAAMP_133282_7.pdf).
↵6 Conservation lands include lands owned by federal agencies (U.S. Forest Service, U.S. Fish and Wildlife Service, National Park Service, and Natural Resources Conservation Service), state agencies (Michigan Department of Natural Resources, Michigan Department of Environmental Quality, and Michigan Department of Transportation), nongovernmental organizations, local governments (county, township, and municipal), and private land with conservation easements, long-term contracts, and similar conditions.
↵7 Recreational lands are open spaces used for recreation, with all types of ownership (federal, state, local, and private), such as parks, beaches, and camping sites.
↵8 Michigan Geographic Data Library, Digital Orthophoto Quads, 2005 series (available at www.mcgi.state.mi.us/mgdl/?rel_thext&action=thmname&cid=17&cat=Digital+Orthophoto+Quads+-+2005+series).
↵9 Stata code (V 6.0) by Professor Timothy G. Conley, Graduate School of Business, University of Chicago (available at http://faculty.chicagobooth.edu/timothy.conley/research/gmmcode/statacode.html; accessed June 19, 2009).
↵10 See Ma (2010) for comparison of model variants with all available data (220 for the SP model and 283 for the AV model, including some observations exclusive to one dataset or the other).
↵11 Personal communication via email with Douglas K. Hodge, an appraiser/realtor in the Michigan office of the Capstone Realty Resources real estate brokerage, during August 2009.
References
In this issue
Jump to section
Related Articles
Cited By...
- Water Market Participation and Agricultural Land Values
- Estimating the Impact of Critical-Habitat Designation on the Values of Developed and Undeveloped Parcels
- Utility-Scale Solar Farms and Agricultural Land Values
- Flooding and the Value of Agricultural Land
- Flooding and the Value of Agricultural Land
- Following the Market? Hedonic Farmland Valuation Using Sales Prices versus Self-reported Values
- How Does Urbanization Affect Water Withdrawals? Insights from an Econometric-Based Landscape Simulation
- Housing Market Bust and Farmland Values: Identifying the Changing Influence of Proximity to Urban Centers
- Agricultural Land and the Small Parcel Size Premium Puzzle