Sunday, December 30, 2012

The Economic Impacts of Climate Change: Evidence from Agricultural Output and Random Fluctuations in Weather

In "The Economic Impacts of Climate Change: Evidence from Agricultural Output and Random Fluctuations in Weather" which appeared in 2007 (Volume 97, Issue 1) of the American Economic Review at Olivier Deschênes and Michael Greenstone measured the economic impact of climate change on US agricultural land by estimating the effect of random year-to-year variation in temperature and precipitation on agricultural profits. Their preferred estimates indicate that climate change would increase annual profits by $1.3 billion in 2002 dollars (2002$) or 4 percent. They claimed that this estimate was robust to numerous specification checks and relatively precise, so large negative or positive effects are unlikely. They also found the hedonic approach—which was the standard in the previous literature—to be unreliable because it produces estimates that are extremely sensitive to seemingly minor choices about control variables, sample, and weighting.

In a 2012 comment on the paper Anthony C. Fisher, W. Michael Hanemann, Michael J. Roberts, and Wolfram Schlenker  American Economic Review, Volume 102, Issue 7 pages 3749-60. utilizing a series of studies employing a variety of approaches found that the potential impact of climate change on US agriculture is likely negative.  They note that Deschanes and Greenstone (2007) report dramatically different results based on regressions of agricultural profits and yields on weather variables. The divergence is explained by (1) missing and incorrect weather and climate data in their study; (2) their use of older climate change projections rather than the more recent and less optimistic projections from the Fourth Assessment Report; and (3) difficulties in their profit measure due to the confounding effects of storage.

In a reply Olivier Deschênes and Michael Greenstone (American Economic Review 2012, Volume 102, Issue 7, pages 3761–3773 or ) admit that Fisher et al. (2012) (hereafter, FHRS) uncovered coding and data errors in the 2007 paper. They "acknowledge and are embarrassed by these mistakes".
Deschenes and Greenstone summarize FHRS’ main critiques of the 2007 study (DG) as follows:
(i) there are errors in the weather data and climate change projections used by DG; 
(ii) the climate change projections are based on the Hadley 2 model and scenarios, rather than the more recent Hadley 3 model and scenarios;
(iii) standard errors are biased due to spatial correlation;
(iv) the inclusion of state by year fixed effects does not leave enough weather variation to obtain meaningful estimates of the relationship between agriculture profits and weather;
(v) storage and inventory adjustment in response to yield shocks invalidate the use of annual profit data; and
(vi) FHRS argue that a better-specified hedonic model produces robust estimates, unlike the results reported in DG.

DG claims that four of these critiques have little basis. Nevertheless, in their reply they report estimates based on corrections and the climate model used in 2007 and a more recent one.
The New DG reanalysis of agricultural profits with corrected data leads to three primary findings. First, contrary to the results in DG (2007), the corrected data suggest that an immediate shift to the projected end-of-the-century climate would reduce agricultural profits. This impact is larger when projections from more recent climate models are used and smaller in econometric models that allow for local shocks to input and output prices and productivity.
Second, the PDV over the remainder of the century of the projected impacts from a recent climate model is roughly $164 billion, or about 5 years of current annual profits. This estimate is likely to overestimate the loss, because it fails to allow for any technological advances or adaptation in response to higher temperatures. Third, the estimated losses are more than 50 percent smaller than those from the standard approach and generally statistically insignificant when one uses a textbook distributed lag model and calculates the dynamic cumulative effects that account for farmers’ dynamic inventory adjustments in response to temperature realizations.

The resulting change in per-acre profits is multiplied by the number of acres of farmland in the county and then the national effect is obtained by summing across all 2,342 counties in the “REPLY” sample. The same calculation is applied to contemporaneous and lagged weather variables. Average annual aggregate profits in the 2,342 counties in the sample are US$(2002) 32.8 billion. Standard errors are clustered at the county level.
The meaning of these results lies in the eyes of the beholder.  DG 2012 believe that these results fail to make a convincing case for large negative impacts of climate change on aggregate profits in the US agricultural sector. In this respect, they lead to a different conclusion than some of FHRS’s important work on the likely impacts of climate change on the yields of particular crops .... This difference may simply reflect the difference between crop yields and profits as outcomes, where the latter is more amenable to adaptation even in the short run, and provides a more complete indicator of productivity in the agricultural sector than crop yields alone. In contrast, recent research suggests that there may be substantial negative impacts on agriculture and health in poorer countries, especially those with already intemperate climates (Guiteras 2009; Burgess et al. 2011). Much uncertainty remains, however, about the likely economic impacts of climate change, and so further research is necessary.
Results again demonstrate that predicted impacts of climate change are heavily dependent on functional form, choices of the temperature variables, the covariates used for adjustment, and the particular year of data used to fit the model. Therefore, they maintain the conclusion from DG that the hedonic approach is unlikely to provide credible estimates of the impact of climate change on the agricultural sector due to problems of omitted variables. 

Using a corrected version of their data file, a variety of specifications that do and do not account for local shocks, and the climate model (i.e., Hadley 2) available in 2007 they find that climate change is projected to reduce annual agricultural sector profits by about US$(2002) 4.5 billion by the end of the century.  They obtain similar results when they apply the same specifications to a data file ... provided by FHRS. These results contrast with the 2007 finding of a statistically insignificant increase of roughly $1.3 billion. Using a 3 percent discount rate and annual projections of climate changes, the present discounted value of the change in agricultural profits between 2010 and 2100 is −$66 billion. To put this in context, historical annual agricultural sector profits are about $33 billion.

Notably, more recent climate model projections (i.e., the Community Climate System Model 3 (CCSM 3) and A2 scenario indicate greater warming.  The application of these projections lead to larger damage estimates. The use of such climate change predictions causes the change in annual agricultural sector profits to increase in magnitude to about $9.9 billion by the end of the century. The present discounted value of projected profit changes with these projections over the next 90 years is $164 billion.The remaining point raised by FHRS pertains to the fact that the farm revenue measure in the census of agriculture includes products sold, regardless of their year of production. Thus, the relationship between annual profits and annual weather realizations may be confounded by inventory adjustments. The textbook solution to such issues of dynamic inventory adjustment in agricultural and other settings is to use a distributed lag model and compute cumulative effects. Thus, the impact of a year’s weather realization is captured over several years. In this setting, the coefficients on the lag of temperature tend to have the opposite sign as the contemporaneous temperature variables. Since the “full” impact of temperature from a distributed lag model is the sum of the coefficients, the projected impacts of climate change from this model are more than 50 percent smaller than those described above. This approach is more demanding of the data and the estimates are less precise than is ideal, however. 

Finally, they underscore the point that all of these estimates are derived under the unrealistic assumption of no technological progress and adaptation over the remainder of the century. It seems reasonable to assume these economic forces will contribute to reducing the predicted damages.
The biggest impact of the data errors in 2007 is that growing season degree-days were too low: the farmland-weighted average in the DG data is 2,561 while in the corrected sample, the corresponding average is 3,821.  In addition, 79 counties were incorrectly dropped due to the errors in the weather data.  In contrast, the growing season rainfall variable in 2007 was error-free.

They utilize two sets of daily predicted climate change data. The first one is from the Hadley 2 model coupled with the IS92a scenario (which they label for simplicity Hadley 2), the same used in DG. The variables contained in this file are daily precipitation and daily minimum and maximum temperatures. ... The second is from the National Center for Atmospheric Research’s CCSM 3 under the A2 scenario, which together predicts larger temperature increases than Hadley 2....
The farmland-weighted predicted change in average growing season degree-days over 2070–2099 for Hadley 2 for the counties in their sample is 673, which corresponds to an 18 percent increase over the 1970–2000 average of 3,821.7. The predicted change from CCSM 3 A2 is 1,441, which corresponds to a 35 percent increase over the 1970–2000 baseline.
Among the eight estimates (four specifications and two sets of counties), the degree-day marginal effects are only statistically significant for nonirrigated counties in the column (1a) specification with year fixed effects. This specification indicates that 100 additional growing season degreedays reduces profits by $1.27 per acre; mean profits per acre in this sample are $31.30.
The next section develops corrected predicted impacts of climate change on US agricultural profits. Specifically, they combine the estimates from the estimation of equation (1) with the projected differences in growing season weather from Hadley 2 and CCSM 3 A2.
With respect to the substantive issue of predicted climate change impacts, the row 4 Hadley 2 estimates range from predicted losses of $7.5 to $1.7 billion, corresponding to −23 percent to −5 percent of current annual agricultural sector profits.

The weighted average of these estimates is a change in annual agricultural profits of −$4.5 billion or 14 percent at the end of the century, when the weights are the inverse of the standard errors.

... The corresponding end-of-century annual climate damage estimates using the CCSM 3 A2 predictions are ... larger in absolute value since this climate model predicts a significantly larger increase in growing season temperatures, as well as a reduction in growing season rainfall. The predicted losses range from $14.8 to $4.8 billion. The inverse standard error weighted average of the predicted impacts under CCSM 3 A2 is −$9.9 billion, more than twice as large as under the Hadley 2 model; this is about 30 percent of current annual agricultural profits.

... With respect to the present discounted value (PDV) of the predicted annual impact of climate change on aggregate farm profits over 2010–2099, based on a discount rate of 3 percent the mean of the PDV estimates from the Hadley 2 projections and the four specifications indicates that the US agricultural sector is predicted to suffer losses of $66 billion over the remainder of the twenty-first century. The corresponding figure from the CCSM3 A2 predictions is a loss of $164 billion. Thus, these estimates, which do not allow for long-run adaptation or directed technical change, imply that climate change will cause a loss of about 2 and 5 years of current profits in the agricultural sector, respectively.

FHRS make the important point that in a given year farmers are able to store some of their grain output and sell it in future years.

However, ... the absolute value of the change in inventories accounts for only 1.4 percent of “total cash receipts from marketings” during the period 1994–2003. In 2007, this statistic caused DG to conclude that storage was not a major factor. FHRS, however, show convincingly that among farms without livestock, the value of a year’s production exceeds sales in bountiful years and is below it in lean years. FHRS argue that this invalidates the use of annual profit data to learn about the impacts of weather realizations.
In the context of making projections about climate change, the key finding is that models that account for lagged weather generally predict smaller cumulative losses than the ones that account only for contemporaneous weather. For example, the weighted average (again using the inverse of the standard errors as the weight) of the Hadley 2 baseline model’s estimates in panel A is −$1.3 billion, compared to −$4.5 billion from the model that just includes contemporaneous weather variables. In the case of the CCSM 3 A2 estimates, the corresponding estimates are −$3.4 billion and −$9.9 billion from the models with and without the lag, respectively.
By Olivier Deschênes 1 and Michael Greenstone 2
1. Department of Economics, University of California, Santa Barbara, 2127 North Hall, Santa
Barbara, CA 93106, IZA, and NBER (e-mail:; 
2. MIT Department of Economics, E52–359, 50 Memorial Drive, Cambridge, MA 02142–1347, and NBER (e-mail:
American Economic Review 
Volume 102, Issue 7; 2012; pages 3761–3773

No comments:

Post a Comment