By using a panel of county level data and including county and state by year fixed effects, we rely on across county variation in county-specific deviations in weather within states. This means that our estimates are identified from comparisons of counties that had positive weather shocks with ones that had negative weather shocks, within the same state. Put in another way, this approach non-parametrically adjusts for all factors that are common across counties within a state by year, such as crop price levels. If production in individual counties affects the overall price level, which would be the case if a few counties determine crop prices, or there are segmented local markets for agricultural outputs, then this identification strategy will not be able to hold prices constant. The assumption that our approach fully adjusts for price differences seems reasonable for most agricultural products for at least two reasons. First, production of the most important crops is spread out across the country and not concentrated in a small number of counties. For example, McLean County, Illinois and Whitman County, Washington are the largest producers of corn and wheat, respectively, but they only account for 0.58% and 1.39% of total production of these crops in the US. Second, our results are robust to adjusting for price changes in a number of different ways. In particular, the qualitative findings are similar whether we control for shocks with year or state by year fixed effects.6 Returning to equation , consider the second term, which is the change in profits due to the weather-induced change in quantities. We would like to obtain an estimate of this term based on long run variation in climate, since this is the essence of climate change. Instead, our approach exploits short run variation in weather. Since farmers have a more circumscribed set of available responses to weather shocks than to changes in climate, it seems reasonable to assume that Short Run > Long Run. 7 For example, farmers may be able to change a limited set of inputs in response to weather shocks. But in response to climate change,plastic pots for planting they can change their crop mix and even convert their land to non-agricultural uses .
Consequently, our method to measure the impact of climate change is likely to be downward biased relative to the preferred long run effect. In summary, the use of weather shocks to estimate the costs of climate change may provide an appealing alternative to the traditional production function and hedonic approaches. Its appeal is that it provides a means to control for time invariant confounders, while also allowing for farmers’ short run behavioral responses to climate change. Its weakness is that it is likely to produce downward biased estimates of the long run effect of climate change.Agricultural Production. The data on agricultural production come from the 1978, 1982, 1987, 1992, and 1997 Censuses of Agriculture. The Census has been conducted roughly every 5 years since 1925. The operators of all farms and ranches from which $1,000 or more of agricultural products are produced and sold, or normally would have been sold, during the census year, are required to respond to the census forms. For confidentiality reasons, counties are the finest geographic unit of observation in these data. In much of the subsequent regression analysis, county-level agricultural profits are the dependent variable. This is calculated as the sum of the Censuses’ “Net Cash Returns from Agricultural Sales for the Farm Unit” across all farms in a county. This variable is the difference between the market value of agricultural products sold and total production expenses. This variable was not collected in 1978 or 1982, so the 1987, 1992, and 1997 data are the basis for our analysis. The revenues component measures the gross market value before taxes of all agricultural products sold or removed from the farm, regardless of who received the payment. Importantly, it does not include income from participation in federal farm programs, labor earnings off the farm , or income from non-farm sources. Thus, it is a measure of the revenue produced with the land. Total production expenses are the measure of costs. It includes expenditures by landowners, contractors, and partners in the operation of the farm business.
Importantly, it covers all variable costs . It also includes measures of interest paid on debts and the amount spent on repair and maintenance of buildings, motor vehicles, and farm equipment used for farm business. The primary limitation of this measure of expenditures is that it does not account for the rental rate of the portion of the capital stock that is not secured by a loan so it is only a partial measure of farms’ cost of capital. Just as with the revenue variable, the measure of expenses is limited to those that are incurred in the operation of the farm so, for example, any expenses associated with contract work for other farms is excluded.Data on production expenses were not collected before 1987. The Census data also contain some other variables that are used for the subsequent analysis. In particular, there are variables for most of the sub-categories of expenditures . These variables are used to measure the extent of adaptation to annual changes in temperature and precipitation. The data also separately report the number of acres devoted to crops, pasture, and grazing. Finally, we utilize the variable on the value of land and buildings to replicate the hedonic approach. This variable is available in all five Censuses. Soil Quality Data. No study of agricultural land values would be complete without data on soil quality and we rely on the National Resource Inventory for our measures of these variables. The NRI is a massive survey of soil samples and land characteristics from roughly 800,000 sites that is conducted in Census years. We follow the convention in the literature and use the measures of susceptibility to floods, soil erosion , slope length, sand content, clay content, irrigation, and permeability as determinants of land prices and agricultural profits. We create county-level measures by taking weighted averages from the sites that are used for agriculture, where the weight is the amount of land the sample represents in the county. Since the composition of the land devoted to agriculture varies within counties across Censuses, we use these variables as covariates. Although these data provide a rich portrait of soil quality, we suspect that they are not comprehensive. It is this possibility of omitted measures of soil quality and other determinants of profits that motivate our approach.
Climate Data. The climate and weather data are derived from the Parameter-elevation Regressions on Independent Slopes Model .This model generates estimates of precipitation and temperature at 4 x 4 kilometers grid cells for the entire US. The data that are used to derive these estimates are from the more than 20,000 weather stations in the National Climatic Data Center’s Summary of the Month Cooperative Files. The PRISM model is used by NASA,drainage for plants in pots the Weather Channel, and almost all other professional weather services. It is regarded as one of the most reliable interpolation procedures for climatic data on a small scale. This model and data are used to develop month by year measures of precipitation and temperature for the agricultural land in each county for the 1970 – 1997 period. This was accomplished by overlaying a map of land uses on the PRISM predictions for each grid cell and then by taking the simple average across all agricultural land grid cells.To replicate the previous literature’s application of the hedonic approach, we calculated the climate normals as the simple average of each county’s annual monthly temperature and precipitation estimates between 1970 and two years before the relevant Census year. Furthermore, we follow the convention in the literature and include the January, April, July, and October estimates in our specifications so there is a single measure of weather from each season. Table 1 reports county-level summary statistics from the three data sources for 1978, 1982, 1987, 1992, and 1997. The sample is limited to the 2,860 counties in our primary sample.Over the period, the number of farms per county declined from approximately 765 to 625. The total number of acres devoted to farming declined by roughly 8%. At the same time, the acreage devoted to cropland was roughly constant implying that the decline was due to reduced land for livestock, dairy, and poultry farming. The mean average value of land and buildings per acre in the Census years ranged between $1,258 and $1,814 in this period, with the highest average occurring in 1978. The second panel details annual financial information about farms. We focus on 1987-97, since complete data is only available for these years. During this period the mean county-level sale of agricultural products increased from $60 to $67 million. The share of revenue from crop products increased from 43.5% to 50.2% in this period. Farm production expenses grew from $48 million to $51 million. Based on the “net cash returns from agricultural sales” variable, which is our measure of profits, the mean county profit from farming operations was $11.8 million, $11.5 million, and $14.6 million or $38, $38, and $50 per acre in 1987, 1992, and 1997, respectively. The third panel lists the means of the available measures of soil quality, which are key determinants of lands’ productivity in agriculture. These variables are essentially unchanged across years since soil and land types at a given site are generally time-invariant. The small time-series variation in these variables is due to changes in the composition of land that is used for farming.
Notably, the only measure of salinity is from 1982, so we use this measure for all years. The final panels report the mean of the 8 primary weather variables for each year across counties. The precipitation variables are measured in inches and the temperature variables are reported in Fahrenheit degrees. On average, July is the wettest month and October is the driest. The average precipitation in these months in the five census years is 3.9 inches and 2.0 inches, respectively.Table 2 explores the magnitude of the deviations between counties’ yearly weather realizations and their long run averages. We calculate the long run average variables as the simple average of all yearly county-level measurements from 1970 through two years before the examined year. Each row reports information on the deviation between the relevant year by month’s realization of temperature or precipitation and the corresponding long run average. The first column presents the yearly average deviation for the temperature and precipitation variables across the 2,860 counties in our balanced panel. The remaining columns report the proportion of counties with deviations at least as large as the one reported in the column heading. For example, consider the January 1987 row. The entries indicate that 73% of counties had a mean January 1987 temperature that was at least 1 degree above or below their long run average January temperature . Analogously in October 1997, precipitation was 10% above or below the long run average in 95% of all counties. Our baseline estimates of the effect of climate change follow the convention in the literature and assume a uniform five degree Fahrenheit increase in temperature and eight percent increase in precipitation associated with a doubling of atmospheric concentrations of greenhouse gases .It would be ideal if a meaningful fraction of the observations have deviations from long run averages as large as 5 degrees and 8% of mean precipitation. If this is the case, our predicted economic impacts will be identified from the data, rather than by extrapolation due to functional form assumptions. In both the temperature and precipitation panels, it is clear that deviations of the magnitudes predicted by the climate change models occur in the data. It is evident that for all four months there will be little difficulty identifying the 8% change in precipitation. However in the cases of temperature, deviations as large as +/- 5 degree occur less frequently, especially in July. Consequently, the effects of the predicted temperature changes in these months will be identified from a small number of observations and functional form assumptions will play a larger role than is ideal.