This
post discusses research from a paper coauthored with Andrew Barkley and Lanier
Nalley in the Proceedings of the National
Academy of Sciences. The paper can be found here. We utilize Kansas
field-trial data for dryland winter wheat yields. A major strength of this data
is that we were able to match yield data with daily temperature observations
across eleven locations for the years 1985-2013.
So,
there is a lot of variation in the data, and we can accurately measure local
temperature exposure. Max, Sol, Wolfram, and Adam Sobel have a nice paper on
the importance of such accuracy here, and Wolfram has
blogged on the importance of daily versus more aggregate (e.g. monthly)
measures here.
Although
not the main focus of our paper, we find that the frequency at which
temperature exposures are measured has a large impact on simulated warming
impacts (see the supplementary information here). Any stats geek
– myself included – will tell you that accurate identification requires
sufficient variation, and the more variation the better! Mike and Wolfram have
some great posts on constructing temperature measures here and here.
We
follow their prescribed method for interpolating temperature exposures and
constructing degree days. However, it is still common in many empirical
analyses to use minimum and maximum temperatures to construct a measure of
average temperature and call it a day. Don’t do this! You are missing so much
important variation in temperature exposure that can be measured using the
interpolation approach outlined by Mike and Wolfram.
Another
consideration not often taken into account in climate change impact studies is
that warming temperatures can have both positive and negative yield impacts.
Extreme temperatures on both the low (cold) and high (heat) end of the temperature
distribution are typically bad for crops. So if we think of warming as a
shifting of the distribution to the right, the result is fewer of the former
(positive effect) and more of the latter (negative effect).
So
what? Well, we find that the net warming impact is negative for winter wheat in
Kansas (more heat trumps less freeze), but omitting the beneficial effects of
freeze reduction leads to vastly overestimated impacts (Figure 1).
Figure 1. Predicted warming impacts under
alternative uniform temperature changes across the entire Fall-Winter-Spring
growing season. Impacts are reported as the percentage change in yield relative
to historical climate. The preferred model includes the effects from a
reduction in freezing temperatures, while the alternative holds freeze effects at
zero. Bars show 95% confidence intervals using standard errors clustered by
year and variety.
The
upshot here is that an accurate identification of warming impacts for winter
wheat requires accounting for both ends of the temperature distribution. It
would be interesting to know if this finding applies to other crops as well.
An
additional strength of our data is that we observe 268 wheat varieties
in-sample, which allows us to estimate heterogeneous heat resistance. As with
other crops, winter wheat has experienced a steady increase in yields over time
due to successful breeding efforts. Much of this increase is driven by a lengthened
grain-filling stage, which increases yield potential under ideal weather
conditions but introduces additional susceptibility to high temperature
exposure during this critical period. David has some great posts on evolving
weather sensitivities here, here, and here.
Essentially,
if this line of reasoning holds we should expect to see a tradeoff between
average yields and heat resistance across varieties. We group varieties by the
year in which they were released to the public and allow the effect of extreme
heat to vary across this grouping. [Aside: there are practical reasons why we
group by release year that are discussed in the paper, we are experimenting
with other grouping schemes in on-going projects].
We
find that there does indeed exist a tradeoff between heat resistance and average
yield, with higher yielding varieties less able to resist temperatures above 34°C
(Figure 2). If the least resistant variety is switched to the most resistant
variety, average yield is reduced by 6.6% and heat resistance is increased by
17.1%. We also find that newer varieties are less heat resistant than older
varieties. Linear regressions using estimates for the 268 varieties indicate
that these relationships are statistically significant (P-values < 0.05).
Figure 2. Mean (average) yields and heat
resistance are summarized by release year. Heat resistance is measured as the
percentage impact on mean yield from an additional degree day above 34°C. The
smaller the number in absolute value the more heat resistant the variety is.
These
findings point to a need for future breeding efforts to focus on heat
resistance, and there is currently much work being done in this area. Check out
the Kansas State University Wheat Genetics Resource Center (WGRC) and the
International Maize and Wheat Improvement Center (CIMMYT) here and here.
From
a historical perspective, our results indicate that such advancements will
likely come at the expense of higher average yields. However, there is
potentially a huge upside to developing a new variety that combines high yields
with improved heat resistance. Under such a scenario, reduced freeze exposure
could outweigh increased heat, leading to a net positive warming effect.
In
the absence of such a silver bullet variety, the average-yield/heat-resistance
tradeoff presents an interesting challenge for producer adaptation, which will
ultimately be driven by some economic decision-making process. Producers are
individuals, or families, and as such they have a certain tolerance for
exposing themselves to risk. Much work has been done showing that farmers enjoy
smoothing their consumption over time, which is akin to reducing profit
variation. Farrell Jensen and Rulon Pope have a nice paper on this here.
So
from a climate change adaptation perspective, it is important ask whether producers
prefer a variety that offers high average yield but low heat resistance, or a
variety with lower average yields coupled with high resistance? Are there
important risk preference differences across producers, or are they a fairly
homogeneous group? Currently, we don’t have a firm answer for these pertinent
questions.
There
has been much work in the agricultural economics literature on risk preference
heterogeneity and the extent to which producers will trade off average yield
for a reduction in yield variance. However, yield variance captures deviations
both above and below the average, which might not be the relevant measure of
risk under a warming climate since we are largely concerned with negative (i.e.
downside) yield effects.
Martin
Weitzman refers to this as fat-tailed uncertainty, and has done some really
interesting work in this area (e.g. here). Jean Paul
Chavas and John Antle are agricultural economists that seem to be working in
this direction using the partial moments framework that John developed, see here, here, and here.
Knowledge
about the willingness of producers to trade off yield for risk reduction should
clearly be an important focus of future breeding efforts. Historically, plant
physiologists and geneticists have worked independent of agricultural
economists, but this should change as climate change presents a clear need for
well-conceived interdisciplinary research.
In
closing, it is worth pointing out that public policy will also likely have a
strong effect on the welfare implications for producers under warming. Direct
funding support for research provides one linkage, but another often overlooked
linkage arrives in the form of subsidized agricultural production. For example,
do policies that protect producers against large-scale crop losses provide a
disincentive to adopt heat resistant varieties? Wolfram and Francis Annan have
looked at this issue here and find that U.S.
corn and soybean producers’ adaptation potential is skewed by government
programs, in turn implying that producers will choose subsidized yield
guarantees over costly adaptation measures.
Thus,
even if we come to know what the
optimal adaptation path is, it is not clear how we will get there. Economists love to talk of the unintended consequences
of public policy. Sometimes it seems that every good policy has a dark side. It’s
called the dismal science for a reason ;-)
Thanks Jesse. Great paper and blog post. Since part of the idea of guest posts is to encourage discussion, here's a question I had after reading this. You say the benefits of warming aren't often taken into account. i think it's true that freeze effects aren't usually explicit like you did. but to the extent that freeze is correlated with other things it drives some of the relationship b/w yields and temperature. i don't think anyone actually estimates a model with separate terms for cold effects and then omits it when they project, do they?
ReplyDeletein general since all the predictors are correlated (I presume, but it's not clear how much) it's not really fair to fix some changes at zero and let others change. i realize the figure was meant to emphasize how the benefits of less frost offset the losses somewhat, and it does, but it strikes me that if the goal is to say how much "ignoring" frost altogether could bias the inferred sensitivities or projections, it might have been fairer to re-estimate a model without frost instead of just setting those changes to zero.
again, thanks for posting.
David, thanks for the opportunity to post on GFEED and your comment, we really appreciate it.
ReplyDeleteIn general, it’s an open question as to whether omitting freezing temperatures from the regression model will lead to biased warming impacts. What it will likely do is bias the estimated impact of extreme heat, as freeze is correlated with both yields and heat. This is omitted variable bias in action. Note that for locations/crops in which freezing temps are not a driver of yields, this bias is not a concern.
In the present case, it is not until we control for freeze that we get at the “true” heat effect. So I see your point but I think that the comparison we offer is valid from the standpoint of
(i) if you correctly identified the heat effect, and
(ii) ignored the benefits of freeze reduction.
Your right of course that omitting freeze and re-estimating the model might produce similar warming impacts, but I hope that we agree that the heat effect used to simulate those impacts is likely biased. Often models with biased parameter estimates forecast the same, or sometimes better than, models with unbiased estimates. I definitely concede that, but I think it's fair to question whether omitting freeze can affect the warming impacts.
The main point that we are trying to get across could have been more clearly stated. When conducting these types of analyses, the results are often subjected to a variety of robustness checks. If freeze is not accounted for in the regression, it is worth considering whether the impacts differ from a model that does include freeze since it is easy to do (you already have the data needed to construct the measure) and pertinent (omitted variable bias is a major potential confounder).
The upshot is that our finding is for a single dataset on a single crop. I personally am curious as to whether this is a valid concern for other datasets/locations/crops as well. As noted above, we don’t consider this a closed book, but an open question for the scientific community to consider.
In the US, winter wheat is planted in the Fall and harvested in May/June. So there are lots of opportunities for inopportune freezes to affect yields, and perhaps more importantly, planting cannot be delayed in the Spring to avoid freeze exposure as with say corn or soybeans. Thus, it would be unwise to extend this finding to other crops without first conducting the necessary research.
Great comment! Thanks for the opportunity to clarify our findings and their implications.