The cost is that it may measure something different from the impact of climate on long-run profit

One possible measure is farmland value, which presumably reflects long run profitability, and which has been used quite extensively since Mendelsohn et al. introduced the so-called Ricardian approach for assessing the impact of climate change on agriculture. With this approach, one estimates a Ricardian or hedonic regression relating farmland value to climate variables and to other variables that control for non-climate factors which might also affect farmland value. This approach is intended to capture the long-run impact of climate on farmland value, and it allows for farm-level adaptations that might be undertaken as climate varies. In their paper, DG question whether there are unmeasured and omitted variables that also influence farmland value and that might be correlated with the climate variables in such a way as to bias their coefficient estimates in the hedonic regression. DG’s important methodological innovation is to propose an alternative approach using fixed-effects to control for time-invariant idiosyncratic features of the county within a panel data setting. However, for this approach to work, one cannot use climate variables as regressors since, in practice,blueberry production these are likely to be fixed over the duration of the panel and, hence, perfectly collinear with the county fixed-effects.

For this reason, DG use annual weather rather than long-run climate, since weather does vary over the course of the panel. Also, because of their annual focus, DG use annual “profit” as their metric of value rather than farmland value; this becomes the dependent variable in their regression. They are thus measuring the effect of weather on short-run profit rather than that of climate on long-run farming profitability or land value.They find no statistically significant relationship between U.S. agricultural profits, proxied by sales-minus-costs as reported in county-level data of the 1987, 1992, 1997, and 2002 agricultural censuses, and weather variables in the same years. They also find no statistically significant relationship between yields of the major field crops corn and soybeans and weather. They conclude that if short-run weather fluctuations have no influence on agricultural profits or output, then in the long-run, when adaptations are possible, climate change is likely to have no impact or will even prove beneficial. With any measurement strategy there are benefits and costs, and the ultimate effectiveness of the strategy is an empirical question.The benefit of DG’s strategy is that it is less vulnerable to unmeasured and omitted time-invariant factors.

DG argue that their measure overstates any possible long-run adverse impact of climate because it reflects the short-run response to fluctuations in weather and therefore does not allow for longer-run adaptation, which could only be less costly. But there are also many ways farmers cope with short-run shocks that would be more costly and/or less sustainable in the long run. For example, if the short-run response to a sudden increase in temperature is to pump more groundwater, this strategy may be less sustainable and/or more costly over the long run with a permanent increase in temperature than in the short run, due to depletion of the groundwater resource. In that case, the short-run impact of a fluctuation in weather would understate the long-run impact of a permanent shift in climate. Besides the conceptual issues associated with DG’s measurement strategy, there are serious questions about how they implement it and whether it actually produces the results they claim. Perhaps most importantly, there are some unusual features of the data used by DG and their representation of climate change scenarios that appear to influence their results, in each case in a direction away from finding any potential negative impact of the change. In our own research we have considered regression models that use both cross-sectional climate variations and time-series weather variations. In Schlenker et al. we show that a better-specified hedonic model that accounts for the influence of irrigation on farmland values is robust and predicts large negative impacts from projected climate changes.

In Schlenker and Roberts we find a strong relationship between corn, soybean, and cotton yields and weather, a relationship that indicates extremely warm temperatures sharply reduce yields for all three crops. Adaptation to warmer, or even extreme, temperatures, is suggested by DG and others, and this is of course possible, especially over time with the development of new crop varieties, but it is worth noting that we find no evidence of greater heat tolerance in yield regressions in warmer regions in the South as compared to cooler regions in the North, and no evidence that relative heat tolerance has grown over time. The relationship is strong and robust and very similar whether derived from time-series variations in weather or cross-sectional variations in climate and comparable in the cross-section of farmland values. Thus, while one cannot predict whether adaptations to extreme heat may occur in the future, there appears to have been little or no adaptation in the past . Climate models, in turn, project that the frequency of extremely warm temperatures will increase significantly. Holding fixed the locations where crops are grown, we predict potential losses in yields for key crops of approximately 30-40% by the end of the century under a slow- warming scenario and 60-80% under the fastest warming, “business as usual”, scenario. These predictions also accord with our research that uses the hedonic approach, where potential losses in farmland value range from approximately 30% to 70% for the same scenarios over the same time period. What explains the stark differences between our empirical findings and those of DG? With regard to DG’s results on profits and yields, we present evidence showing the difference stems from several sources: coding and data errors in the weather data that magnify within state temperature fluctuations by a factor of seven; an unusual and in our judgment incorrect characterization of climate change across the units of observation; differences in underlying climate change scenarios, in particular reliance by DG on an earlier and more optimistic climate projection than that used in the Fourth IPCC assessment and in our analysis; and DG’s omission of storage, and perhaps other financial or technological mechanisms, that smooth their measure of short-run profits in the presence of weather-induced output fluctuations and cause the short-run impact of weather on profit to understate the long-run impact of a permanent shift in climate.

To investigate differences we downloaded DG’s data and STATA code from the AER website. We found several irregularities in their weather and climate data. These data irregularities explain a large portion of the differences in findings. DG have two weather variables in their data set: the variable dd89, which measures growing degree days for each year and county, and dd89 7000, which measures the average number of degree days in each county between 1970 and 2000.3 These two variables do not appear consistent with each other. The correlation of the county-level average of the four-year panel and the 31-year average given in their data is only 0.39. Given the wide variation in temperatures in the cross-section,blueberry in container one would expect a stronger correlation between the 4-year and 31-year averages across counties. We reconstruct the same weather variables from raw data sources and find a correlation of 0.99. We also find the average of dd89 is much lower and the standard deviation much higher than in our replication. Second, DG’s baseline climate measure has a value of zero degree days for 163 counties. If correct, this measure implies temperatures do not exceed 8C in those counties during the growing season . Temperatures this low would seem implausible in any state, yet many of these counties are in warm southern states . Anomalies caused by missing or incorrect measurements, which as we shall show have an important influence on estimated impacts of climate change, are illustrated in Figures 1 and 2. We independently calculate the degree days variable dd89 7000 used by DG and display it in the bottom panel of Figure 1.4 Note the much smoother pattern as compared to the large discontinuous changes in the top panel. Average temperatures vary smoothly across space, where counties of the same latitude tend to have comparable average temperatures that increase as one moves southward. Exceptions to this rule are mountain chains like the Rockies in the West or the Appalachians in the East, where temperatures are cooler due to gains in altitude. The discontinuous pattern induces incorrect weather variation, which has an especially large influence on parameter estimates in regression models that use state-by year fixed effects. Within-state temperature deviations in our replicated data set are roughly one seventh of DG’s. Third, DG’s predicted changes under warming scenarios are discontinuous in space and range from a decrease of 880 growing degree days to a 6572 growing degree days increase . This pattern is odd given that the underlying climate model does not predict cooling anywhere in the U.S. and the variance of the projected changes far exceeds that of any climate model. Predicted changes in DG’s model and in our replication are shown in Figure 2. Again, compare the discontinuities in the top with the more coherent patterns in the bottom. The large variability of DG’s predicted climate changes stems from the way they combine observed weather and climate-change forecasts. The difficulty arises from the fact that general circulation models generate climate predictions on a coarser geographic scale than data available in historic records. DG use historic county-level data as a baseline combined with climate predictions that are uniform across each state. Thus, after climate change, Los Angeles and San Francisco, Salinas and San Joaquin Valleys, Mount Whitney and Death Valley, are all assumed to have the same climate since all are in California. Much of the within-state variation, however, is maintained in the baseline values, which are county-level averages. Such a representation of climate change therefore displays regression towards the mean, with cooler counties becoming much warmer and some very warm counties becoming cooler. This regression-toward-the-mean effect is accentuated by apparent errors in the baseline degree-day measure. Consider for example Fresno, Kings, and Tulare counties in the southern San Joaquin Valley of California. In DG’s data, Fresno is predicted to have a decrease of 414 degree days ; Kings county has an increase of 403 degree days and Tulare an increase of 4685 degree days . Tulare’s large increase is the result of a zero baseline. But even for Kings and Fresno counties, for which there are no missing baselines, predicted climate changes are too different for bordering counties. This treatment of climate change is unusual. We are not aware of any other application of the Hadley GCM model that predicts decreasing average temperatures by the end of the century in any U.S. location. Rather, the standard approach is to add regional predicted changes from the climate models to the sub-regional baseli While there are differences between DG’s and our own model of yields, much of the difference in our predicted impacts stems from the data issues described above. We generally find large negative projected climate impacts from replicated profit regressions as well, though results here are somewhat mixed and less likely to be significant, for reasons we discuss in the next section. Comparisons of the original and replicated yield and profit regressions are summarized in Tables 1 and 2. In our replications we fix the observations so they exactly match those used by DG. This excludes some agriculturally important counties, which are missing in DG. For example, 66 of Iowa’s 99 counties are missing from their data set, yet Iowa is the largest producer of corn and soybeans, in turn the nation’s two largest crops. On the other hand, most of Nevada’s counties are included, which are highly irrigated. Irrigation poses a problem for estimation of the effects of climate variables both in a cross-section and a panel. In a cross-section such as the Ricardian or hedonic approach the problem is that since irrigation tends to be correlated with temperature and precipitation it can bias estimates if omitted, as we discuss later in the section on robustness. In a panel, the effect of weather fluctuations depends on water availability. DG deal with this problem by estimating regressions with separate coefficients for irrigated and rainfed counties.