The Effect of Climate Change on Canadian Farmland Values: A Ricardian Approach

The Ricardian Approach: Climate and Nonagricultural Uses of Farmland

The Ricardian approach assumes that farmers are profit-maximizing and that their land use decisions reflect the most profitable land use. As the climate changes, farmers are assumed to alter their production practices (including their use of land and crop choice) to maximize the present value of net returns. This adaptive process is assumed to be reflected in the future stream of net returns associated with farmland, which are in turn revealed in the price of farmland. In their seminal article, Mendelsohn, Nordhaus, and Shaw (1994) provide a useful example of how a wheat farmer adapts to higher temperatures by planting corn when changing temperatures allow it to become more profitable than wheat. By focusing on the relationship between climate change and farmland prices, the Ricardian approach implicitly incorporates adaptivity. This stands in sharp contrast to a static production-function approach that examines how climate change affects the returns to a specific crop.

Authors often rely on the capitalization model to illustrate specific dimensions of the Ricardian approach (see Ortiz-Bobea 2020; Bareille and Chakir 2023). We emphasize the effect of climate and nonagricultural uses of farmland (e.g., residential housing). The capitalization model is

1

where LV_t corresponds to farmland values at time t, R_u,t is the rental price from using the most profitable land use (u) at time t, p and w are input and output prices, l is a measure of specific land attributes, c_t is a vector of climatic conditions at time t, and γ corresponds to a discount rate.

Emphasizing the importance of nonagricultural influences on land values, Ortiz-Bobea (2020) extends equation [1] to consider the influence of nonagricultural influences on land values. In its simplest form, Ortiz-Bobea (2020) describes farmland values as a weighted average between farm and nonfarm components. The weight is represented by the function a(t^*), which identifies the optimal time, t^*, for conversion from farm to nonfarm use. Equation [2] details this equation and, similar to Ortiz-Bobea (2020), omits the discount rate for clarity:

2

When conversion from agricultural to nonagricultural land use is distant (t^*→∞) and (a(t^*)→1), prices will equal the farm use component (LV_t^F). Similarly, when conversion from agricultural to nonagricultural land use is imminent (t^*→0) and (a(t^*)→0), price will equal the nonfarm component (LV_t^N). Importantly, this means that Ricardian estimates must consider the influence of nonagricultural benefits on farmland values. For this reason, our Ricardian model includes novel control variables for nonagricultural influences.

Stage 1: Hedonic Regression

We estimate the relationship between climate variables and farmland values using a pooled dataset spanning six years.⁷ Our regression framework closely resembles the pooled model approach used by Sanghi and Mendelsohn (2008) and Bareille and Chakir (2023). Equation [3] describes the hedonic regression used to estimate the climate coefficients:

3

LV_{i m p} represents the per-acre sale price of a parcel of farmland i, in county m and in province p. The matrix C includes climate characteristics for seasonal temperature and precipitation, while Q and N represent matrices for quality characteristics of the farmland and nonagricultural variables. Quality characteristics include measures for soil quality and slope. Nonagricultural variables include population density and median income at the census subdivision level. D represents the distance of a parcel to the nearest population center of 90,000 people or more.⁸ M and P correspond to census division level and provincial-level fixed effects, respectively. Results are presented for models that use only census division fixed effects, provincial fixed effects, and no fixed effects (referred to as the base model). Y represents year fixed effects and indicates which calendar year the transaction occurred. Last, u corresponds to unobservables at the parcel, census division, and provincial levels. β^′, θ^′, θ^′, and α^′ are all sets of parameters to be estimated.

Following previous studies, the per-acre value of farmland is log-transformed (Schlenker, Hanemann, and Fisher 2007; Massetti and Mendelsohn 2011; Van Passel, Massetti, and Mendelsohn 2017).⁹ In addition, we include seasonal climate variables measuring historical climate conditions in January, April, July, and October. Seasonal precipitation and temperature measures were selected over alternative climate measurements, such as growing degree days, because they better capture the effect of climate on farmland values outside of the growing season (Massetti, Mendelsohn, and Chonabayashi 2016).¹⁰ In addition to a linear term, climate variables include a squared term to capture nonlinearities. Including seasonal climate variables in linear and quadratic form is consistent with previous studies (see De Salvo, Diego, and Giovanni 2014, table 1).

Stage 2: Aggregate Impacts

The second stage of the Ricardian approach seeks to estimate the aggregate impacts of climate change on farmland values by multiplying the climate coefficient estimates from the first stage by the difference between historical and projected climate data. When Mendelsohn, Nordhaus, and Shaw (1994) first introduced this approach, the estimated coefficients were multiplied by a uniform climate change scenario. Other studies, including Canadian studies like Reinsborough (2003) and Weber and Hauer (2003), used a similar approach to determine aggregate impacts.

Rather than multiplying the estimated climate coefficients from stage 1 by a uniform climate change scenario, the strategy used here involves subtracting the predicted historical farmland value from the predicted future value. The predicted historical price uses the coefficient estimates from stage 1 and multiplies them by the historical value for each covariate. The predicted future price uses the same coefficient estimates but changes only the values of the climate variables (based on downscaled 1 km gridded climate projections), while keeping the values for the control variables constant. This allows for the estimate of the expected change in farmland values to be the result of changes in only the climate variables.

This procedure also allows for the impacts to be determined at the parcel level and then averaged across the whole sample. The granularity of the dataset and the availability of climate forecasts at high spatial resolutions make this procedure possible and enables an examination of regional impacts in addition to Canada-wide impacts. Equation [4] shows the calculation used to determine the expected per-acre change in farmland values resulting from climate change for each parcel in the sample:

4

where ΔLV_i corresponds to the change in land values for parcel FV_i corresponds to the predicted future log value of parcel i, and HV_i corresponds to the predicted historical log value of parcel i. RMSE is the root mean squared error, which estimates the error variance. This adjustment is needed because the error terms only have a mean of zero in log units. Simply exponentiating the predicted dependent variable would result in predictions of farmland values that are systematically underestimated (Newman 1993; Wooldridge 2009).

Farmland Values

Our unique parcel-level information on farmland prices and attributes is obtained from Canada’s largest agriculture and agri-food lender, Farm Credit Canada (FCC).¹¹ This dataset contains all of FCC’s market sales transactions between 2017 and 2022. For each transaction there is information about the farmland parcel, including the total price, price per acre, latitude and longitude coordinates, land size, land use, and irrigation status. If the sale includes only a single parcel, the per-acre value represents the total sale price divided by the total number of acres. However, it is common for a single sale to include multiple land uses, such as using some land for cultivation and some for pasture. In these cases, the sale is broken up into multiple parcels and the per-acre price represents an assessed value calculated by an FCC appraiser and based on local market conditions. In total, 76% of the parcel observations are for sales that include multiple parcels.¹² In addition, some sales include parcels with buildings or are potentially used for nonagricultural purposes. Although the value associated with the land is separated from the building value, our approach limits the sample to only land classified by FCC as cultivated, fruit or pastureland. After removing observations with incomplete data and non–arm’s length transactions and restricting the sample to ensure that the data were representative of true market values, the final dataset included 45,706 parcel observations.¹³

Table 1 provides summary statistics for these observations and presents farmland values in nominal dollars.¹⁴ The sample represents farmland parcels that were purchased or sold through FCC. The sales prices in the dataset are higher than the average per-acre price of farmland published by Statistics Canada. For example, for 2022, the average per-acre price of farmland in our dataset is $11,300 compared with the Statistics Canada published value of $4,527 (Statistics Canada 2023b).

Throughout North America, only a small portion of farms are responsible for the bulk of agricultural production. In Canada, the largest 10% of Canadian farms are responsible for more than two-thirds of all revenue (AAFC 2021). The higher prices in the FCC dataset likely reflect the relatively more productive Canadian farmland, which is actively engaged in generating farmland revenue. In addition, the 2017 mean per-acre price is higher than the 2018 mean, which does not accord with Statistics Canada trends. For this reason, we provide pooled results and results for each cross section. We examine sensitivity in our aggregate effects for pooled results with and without observations from 2017 in Appendix Table A7. As expected, our results are qualitatively robust across these different specifications.

Historical Climate Data

Historical weather data from 1985 to 2021 was extracted at a downscaled 1 km² spatial resolution using the ClimateNA software (Wang et al. 2016; Mahony et al. 2022). ArcGIS software was used to match the location of each farmland parcel with the respective historical downscaled weather data that account for geological features that can affect weather, such as elevation. We follow the approaches outlined in previous literature (e.g., Mendelsohn, Nordhaus, and Shaw [1994] and all previous Canadian studies) and construct 30-year climate normals for each year between 2017 and 2022. For instance, the 2022 climate-normal consists of the average weather conditions between 1992 and 2021. Specifically, climate variables used in this study measure average precipitation and temperature for January, April, July, and October.

Climate Forecasts

Forecasts about future climate conditions for the same monthly precipitation and temperature variables were obtained through AdaptWest’s database (Wang et al. 2016; AdaptWest Project 2022; Mahony et al. 2022). AdaptWest generates an ensemble climate forecast at a 1 km² spatial resolution using the ClimateNA software by downscaling data from projections based on the sixth IPCC report (Wang et al. 2016; AdaptWest Project 2022; Mahony et al. 2022). A unique feature of our dataset is that both the historical and the climate data are already appropriately downscaled.

The downscaled data prevent aggregation bias that arises when using gridded historical climate data and coarse spatial data from general circulation models (GCMs) (Auffhammer et al. 2013). To address concerns regarding aggregation bias (i.e., Auffhammer et al. 2013), previous Ricardian studies add the difference between GCM-predicted historical climate and observed historical climate to the coarse GCM predictions of future climate (e.g., Fisher et al. 2012; Severen, Costello, and Deschênes 2018). Our dataset is generated by ClimateNA software, which uses historical baseline climate data to downscale both historical and future climate variables and thereby harmonizes subregional variation in climate between parcels for both historical and future periods (Wang et al. 2016; Mahony et al. 2022).

Auffhammer et al. (2013) raised concerns about climate researchers selecting one GCM over alternatives. They argue that there is little evidence that one model should be preferred over others. Our climate forecasts are generated by calculating the ensemble mean of different GCMs for various emissions scenarios. Specifically, we present results using the ensemble mean of 13 GCMs under the Shared Socioeconomic Pathway (SSP) 2-4.5 scenario for 2041–2070.¹⁵ The SSP2-4.5 scenario is considered the “middle of the road” scenario, where CO₂ emissions following their historical patterns and do not shift considerably until the second half of the century (Riahi et al. 2017). Table 2 provides summary statistics for historical and predicted climate data.

Distance to Nearest Population Center

To provide improved controls for nonagricultural related influences on farmland values, a proximity variable measuring the distance of a parcel to the nearest Canadian population center of 90,000 or more inhabitants as of 2016 was created.¹⁶ Because conversion of farmland to nonagricultural uses is likely inversely related to the distance from a population center, the proximity variable controls for nonagricultural influences on farmland values (Chicoine 1981).

The proximity variable is a novel aspect of the dataset and is derived from the unique information regarding the precise latitude and longitude coordinates of each parcel. The near analysis feature in ArcGIS was used to create distance polygons at 10 km intervals for each population center with 90,000 or more inhabitants. The polygons were overlaid with the location of the parcels to determine the distance to the nearest population center.¹⁷ The regression uses the midpoint of each bin to create a continuous variable. For example, for the bin defined as 10–20 km from the nearest population center, the value used in the regression is 15 km. Table 3 provides summary statistics for selected control variables, including the proximity variable.

Soil Quality

To account for heterogeneity in land quality, soil quality and composition data from the Canadian National Soil Database (NSDB) was matched with farmland price data (AAFC 2013). A circular buffer (with the same area of each farmland parcel) was created to intersect the farmland parcel with a map of soil polygons published by the NSDB. The buffer for each parcel was overlaid with the soil polygon map to determine the soil characteristics and components of a particular parcel. In some cases, the parcel buffer intersected multiple soil polygons. For these occurrences, a weighted average of the soil polygons was used based on the percentage of the buffer intersecting each polygon.

Socioeconomic Data

Following previous Ricardian studies, data were collected from the Canadian Census of Population on median income and population density at the census subdivision level. The Canadian Census of Population is conducted every five years, with the most recent version containing 2021 values. This article merges 2021 census data with farmland sales occurring in 2021 and 2022 (Statistics Canada 2023a). The remaining sales data are merged with the 2016 census data (Statistics Canada 2019).

Marginal Effects of Climate on Farmland Values

The discussion of the results emphasizes the results from the pooled model. For completeness, we provide results using yearly cross sections with county fixed effects in Appendix Table A5. Table 4 provides selected coefficient estimates across six pooled models with robust standard errors.¹⁸ Models 1 and 2 include census division fixed effects, models 3 and 4 include provincial fixed effects, and models 5 and 6 include no geographic unit fixed effects. Hereafter, models 5 and 6 are referred to as “base” models. Models 2, 4, and 6 include a linear and a squared term of the proximity variable.

Table 5 presents estimates of the marginal effect for the climate variables in different model specifications. These effects account for the marginal effect of the linear and squared terms of each climate variable on farmland values. The following equation is the marginal effect calculation:

5

where C_i is the value of the individual climate variable, β₁ is the estimated coefficient for the linear term, and β₂ is the estimated coefficient for the quadratic term. For additional ease of interpretation, equation [5] is evaluated at the mean of each climate variable and multiplied by 100.¹⁹ For example, the estimated marginal value of April temperature in model 2 is 28.34, implying that 1°C increase in April temperature above the historical sample average would result in a 28.34% increase in farmland values.

Similar to Mendelsohn and Reinsborough’s (2007) model for Canada, the results in Tables 4 and Table 5 show a positive association between farmland values and temperature and precipitation, meaning that for a uniform one-unit increase in temperate and precipitation across all seasons, the aggregate impact on farmland values is positive. Consistent with findings from Reinsborough (2003) and Weber and Hauer (2003), warming in April is associated with positive increases in farmland values. This result may reflect the benefits of a longer growing season. Last, for all models in Table 5, the marginal value of October precipitation is negative. Weber and Hauer (2003) obtained a similar result and argue that the harmful marginal value of October precipitation could be explained by an increase in frost risk due to wet weather during harvest.

Models 2, 4, and 6 include a linear and squared term of the proximity (in kilometers) of each parcel to the nearest population center. Across all models where it was included, the coefficient estimate for the linear term of proximity variable is negative and statistically significant. The quadratic term is positive and statistically significant across all models, implying that the negative effective decreases as proximity to a population center increases. This observation is consistent with expectations of farmland values being inversely related to the distance to a population center (e.g., the farther a parcel is from a population center, the lower its value, up to a certain threshold).

Table 4 confirms that the proximity variable is statistically significant and negatively associated with the price of farmland. By this measure, farms farther from population centers of 90,000 are associated with lower prices.²⁰ Hence, as previous literature discusses, omitting variables relating to nonagricultural influences could potentially bias Ricardian estimates of the association between climate and farmland prices. (The bias would depend on the correlation between climate measures and urban proximity measures.) Table 5 compares models that include and omit the proximity measure. Generally these results indicate that including the urban proximity variable mainly influences the magnitude of the marginal impacts, rather than the direction of the effect. However, in some cases, the effect appears relatively large. For example, a 1°C increase in July temperature is associated with only 4.53% decrease in farmland values in model 1, but a 2.15% increase in model 2 when the proximity variables are included. These results emphasize the ongoing need to consider measures of urban pressure when using the Ricardian approach.

We control for nonagricultural influences and other fixed unobservable influences. In Tables 4 and 5, models 1 and 2 include census division fixed effects, and models 3 and 4 include provincial fixed effects. The impact of including geographic fixed effects appears to have a more noticeable effect on the marginal effects compared with including the proximity variable. This effect is noticeable for the marginal effect of October temperature. In the base model without the inclusion of the proximity variable, a 1°C increase in average October temperature estimates a 20.29% increase in farmland values. When census division fixed effects and the proximity variable are included, the effect becomes negative and estimates an 11.22% decrease in farmland values for a 1°C increase in average October temperatures. A possible explanation is that the base model captures the positive effect that buyers are willing to pay for warmer fall temperatures occurring in specific census divisions, unrelated to the agricultural productivity of a parcel. Once census division fixed effects are included, the nonagricultural related component is controlled for and thereby diminishes the positive marginal effect of warmer October temperatures observed in the base model. This result may signal the negative effect that a reduced number of “chilling days” (the required number of days between a certain temperature threshold before flowering can occur) has on perennial crops (Hatfield and Prueger 2015). A similar observation was found by Weber and Hauer (2003), who estimate a significant negative effect on farmland values for an increase in October temperatures above the median.

Aggregate Impacts of Climate Change on Farmland Values

As a general matter, predicting the future is precarious because there is a great deal of uncertainty that economic models cannot address. The Ricardian model has been criticized (e.g., Cline 1996; Kaufmann 1998; Darwin 1999) for a host of theoretical and empirical limitations (we discuss some of these). For this reason, contemporary research continues to explore empirical methods to enhance the Ricardian approach and its methodology (Schlenker, Hanemann, and Fisher 2006; Severen, Costello, and Deschênes 2018; Ortiz-Bobea 2020; Bareille and Chakir 2023). Like Ortiz-Bobea (2020), we address omitted variable bias; like Severen, Costello, and Deschênes (2018), we are concerned about variation in perceptions of climate change across the regions we study. In this regard, we believe the capacity to include census division fixed effects is an improvement on many previous studies.

Perhaps one of the more notable assumptions associated with the Ricardian method is that the estimated hedonic price function remains stable for the predicted nonmarginal changes associated with future climate. This assumption is a long-standing issue of concern in the hedonic literature (e.g., Freeman 1993). That said, it undergirds previous Ricardian estimates (e.g., Mendelsohn, Nordhaus, and Shaw 1994; Reinsborough 2003; Weber and Hauer 2003).

As mentioned, previous research has raised concerns regarding the sensitivity of the estimated marginal and total effects to functional form assumptions (e.g., Bareille and Chakir 2023). For these reasons, we view all Ricardian estimates with some caution. However, these estimates serve as an ongoing benchmark for the projected effect of climate change on farmland values and enable comparisons to regions in Canada and throughout the world, where similar methods have been applied. With this caveat in mind, we provide estimates of the future effect of climate change on farmland prices. We limit our results to the sample of Canadian farms in our dataset and the climate expected to characterize our sample of farms in 2070.

We use the coefficient estimates from Table 4 and follow the procedure outlined in Section 2 to calculate aggregate impacts. Our results use AdaptWest’s climate projections for each climate variable under the SSP2-4.5 warming scenario for 2041–2070 (Wang et al. 2016; AdaptWest Project 2022).²¹ The SSP2-4.5 climate change scenario is described by the IPCC as the “middle of the road scenario,” where social, economic, and technological trends follow historical patterns (Riahi et al. 2017). The time horizon of approximately 50 years is consistent with Weber and Hauer’s (2003) application of the Ricardian approach in Canada and allows for a more direct comparison of results between studies.

Table 6 presents full results across the six pooled models. Ninety-five percent confidence intervals around the sample mean were calculated using 1,000 bootstrap replications for the predicted historical per-acre price and future per-acre price. The final two rows in Table 6 detail the annualized per-acre change (calculated using a 5% discount rate) and the percentage change from the mean predicted historical per-acre value to the mean predicted future per-acre value.²² Our estimates of the aggregate impacts of climate change imply a 30%–73% increase in the price of farmland by 2070. Alternatively, when the mean per- acre impact is annualized with a discount rate of 5%, it ranges from $151 per acre to $377 per acre. Presently, the average per-acre price of farmland in the dataset is $9,526 (in 2017 Canadian dollars). In addition to aggregate impacts across the entire Canadian sample, we provide a breakdown of average impacts per province from model 2 in Table 7. All provinces are estimated to experience increases in average farmland values due to climate change.

Our aggregate estimates seem consistent with studies estimating the effects of climate change on farmland values in northern U.S. counties and regions of northern Europe (Mendelsohn, Nordhaus, and Shaw 1994; Van Passel, Massetti, and Mendelsohn 2017; Bareille and Chakir 2023). In France, Bareille and Chakir (2023) estimate a 54%–110% increase in French farmland values depending on the climate scenario. In a separate study of farmland values across Europe, Van Passel, Massetti, and Mendelsohn (2017) find future climate change to have beneficial effects on farmland values in countries in the north and harmful effects on farmland values in the south. Farmland values in northern European countries, including Denmark, Finland, Ireland, Sweden, and the United Kingdom, are expected to increase under the “moderate” warming scenario (Van Passel, Massetti, and Mendelsohn 2017). Under this scenario, Swedish farmland values are expected to experience the largest increase, rising by approximately 65% by 2100 (Van Passel, Massetti, and Mendelsohn 2017). Similarly, Mendelsohn, Nordhaus, and Shaw (1994) estimate increases of $200–$1,300 per acre (in 1982 dollars) for some northern U.S. counties in their cropland and crop revenue models. Generally, northern latitude areas share similar climate conditions and are expected to experience similar future changes in climate. Thus, Ricardian impacts should be relatively homogeneous across regions with similar current and predicted future climate. Our results are consistent with this expectation.

The empirical results in Table 6 are based on unweighted regressions and rely on the quadratic functional form. Previous research has found that underlying model assumptions influence both aggregate impacts and marginal effects. Bareille and Chakir (2023) find that a log-linear specification, compared with a log-quadratic function significantly reduces the marginal effects of climate. We also find sensitivity. That said, across the various alternative specifications we examine, the price effect of climate change is expected to be positive.

Table 8 provides a summary of the aggregate impacts across a range of sensitivity analyses we conducted. Column (2) identifies the aggregate per-acre results and restricted to only cultivated, fruit and pastureland. The model in column (3) uses the same sample restrictions as our primary model but uses a log-linear function form specification rather than a log-quadratic. This specification for the census division fixed effects (proximity) regression indicates an average increase of 43%, compared with 58% in the log-quadratic model.

Column (4) uses weighted least squares for our preferred, log-quadratic functional form specification. Each parcel is weighted by acreage. These results are generally in line with our current results. In column (5), we provide results for a very inclusive model with no restrictions on land uses. As discussed, a sale can include parcels used for different activities and can include buildings. While the value used in the regression represents the value of the underlying land a building is on, there may be an omitted variable problem. For this reason, we include a dummy variable to categorize any land that is potentially used for nonagriculture-related activities. This is the most inclusive model and a reminder that most farms are a combination of parcels whose current uses vary.²³ In this model, the effects are positive and generally larger relative to our current model.

As a benchmark, our results consistently suggest that climate change will have a positive and significant effect across our sample of Canadian farmland values. That said, as emphasized in previous literature and in our own discussion, the exact magnitude of the positive prediction is difficult to determine precisely and appears sensitive to underlying assumptions.