The paper I replicated is “The economic Impacts of Climate Change: Evidence from Agricultural Output and Random Fluctuations in Weather: Reply” by \citet{d53692} (hereafter, DG2012). This is a reply to the comment from \citet{Fisher_2012} (hereafter, FHRS) to the original paper in 2007 \cite{198593} (hereafter, DG2007).
All three papers mentioned above followed the debate of the impact of climate change on U.S. agriculture starting from the first hedonic paper by \citet{a827bf}. Arguing that the traditional hedonic approach suffered from omitted variable issues, DG2007 proposed the fixed effect approach with panel data on land profitability and DG2012 corrected and re-presented their results addressing for the data and methodology issues as pointed out by FHRS.
This topic is controversial due to the fact that perfect information we need to unveil the true relationship between climate change and agriculture is far beyond availability in reality. In the other word, the ideal research design for this question goes a little bit beyond than saying we want to manipulate the weather and observe the corresponding profit changes.
It is not easy to find the correct and comprehensive climate/weather covariates conveying the impact of climate change on agriculture. Temperature and precipitation are the most commonly used and easy-to-get data. But climate change also impacts agricultural productivities in other ways such as altering the frequency of extreme weather events such as drought.
Farmers adapt to climate change by land use change and farm management practices. As a result, the production function approach tends to overestimate the impact of climate change. Previous studies have analyzed the impact of climate change on both agricultural yield and profit. It is hard to say which one is superior than the other. Apparently profit data is more amendable with climate change. However, if profit data is collected, additional concerns are needed to address for the long-term v.s. short term adaptation and the storage issue \cite{f76951}.
I would like to elaborate on the aforementioned fruitful discussion of climate change on agriculture in the first chapter of my dissertation. I am interested in spatial correlations of crop yields and the potential regional variations of climate change impact. Thus I think this replication practice will give me a good start.
Addressing for the two major comments from FHRS on 1) data and coding issue and 2) the potential storage, DG2012 provided the corrected results for:
Impacts of weather fluctuations on agricultural profits
Estimates of climate change on farm profits using new Global Circulation Model (GCM) projections
Potential Storage issue and its impact on the estimates
The empirical model for the weather fluctuation impacts is as follows: \[\textbf{Y}_{ct} = \alpha_{c}+\gamma_{t}+\textbf{X}_{ct}\pi+\sum_{i} \beta_{i}f_{i}(W_{ict})+ u_{ct}\] where c indicates for county and t references a year. Quatratic terms are allowed in the weather variables. Estimation results is given in Table 1 below.
It is noteworthy that though the model allow for variations between irrigated and dryland county, irrigated county are ones having more than 10% of the farmland irrigated. As a result, it might be not surprising that they cannot reject the equality of weather variables across irrigated and dryland counties (Panel C).
In this section DG2012 took the estimates from the previous section to predict the impact of climate change on agricultural profits based on GCM data. Table 2 of DG2012 presented four results based on different weather data sources which leaded to similar conclusion. I am going the replicate the one with their own data source (the 4th row in Table 2, DG2012).
One argument made by DG2012 is that “allowing for local shocks tends to reduce the magnitude of the predicted loss, although this does NOT always come at the expense of reduced statistical precision” observing that the estimated standard error clustered by state is 3.03 for year fixed effect (column 1a) and 2.29 for USDA region*year fixed effects. Personally I think this argument is tricky from two perspectives: 1) it is the only observation supporting this argument, and 2) we are observing a larger change in the estimated mean, so it is not the case in the coefficient of variation perspective.
DG2007 calculated the agricultural profit by subtracting sales revenue by the production cost. Storage was concern since farmers might store their products in good weather years for future sales and do the contrary in bad weather years. As a result, sales did not reveal only current year weather information and would bias our estimation towards zero. DG2012 address this issue by introducing the lag-one weather variables in the estimation. Again I am only going to replicate the results using climate data from the Hadley 2 models instead of both GCM models (Panel A of Table 2, DG2007).
A natural extension from the DG2012 would be the cross terms for growing season days and precipitation. Conditional on the level of precipitation the effects of temperature on agriculture might be totally different. For example, warmer conditions generally helps crop growth. However, when drought presents high temperature and evaporation might cause severe damage. And we would expect such effect differs between irrigation land and dry land. Table 4 compares the impacts when the cross term of weather variables are included (we only did the comparison for the year fixed effect specification, i.e. column 1a and 2a in Table 1). It is interesting to see that this changed the estimates for the irrigated land significantly while the dry land estimates remained similar. However, note that the irrigated land estimates were not statistically significant thus we might not draw any noteworthy conclusion based on this. Furthermore, we still cannot reject the equality hypothesis for the weather variables across dry land and irrigated land.