The authors do not spend much time focusing on the statistical significance of the climate variables they identify as producing economic damage, perhaps since the way they calculated the measures of statistical significance showed that most variables were statistically significant. The paper calculates the statistical significance of the estimated parameters by assuming that the unobserved influences on real income are correlated across geographical areas. That’s a sensible assumption. The technical term for that assumption is that the standard errors are “clustered by region.” However, the paper assumes that the unobserved influences on real income are not correlated over time, an assumption that is almost certainly false. To correct the unobserved factors affecting real income being correlated across regions and over time we “double cluster,” i.e., we cluster standard errors in the regression model over regions and time. Other academic papers that study the effect of climate change on economic growth double cluster their standard errors.

The table below shows my re-estimates of the paper’s regression model using the paper’s original assumption—clustering standard errors by region—and using the new double clustering of the standard errors. I’ve circled in red the one temperature variable that is responsible for almost all of the economic damage. As can be verified, the statistical significance of the important temperature variable and most of the other variables evaporates when we double cluster the standard errors. The statistical significance of the model is not robust to a reasonable alternative method.

The principal mechanism producing the startling high economic damage in the paper is the empirical finding that climate change from many years ago accumulates into loss of real income today. The authors use a standard statistical technique, the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) to justify using ten years of lags for the temperature variables and four years of lags for the precipitation variables. You use the AIC and BIC by finding the number of lags in the model that produce the smallest values. The authors claim that the AIC and BIC are minimized when there are either eight, nine, or ten lags of temperature variables and four lags of temperature variables, where each lag is one year. We can see their argument in Extended Data Fig 2. from the paper.

The authors appear to have made a serious mistake. As the table below shows, re-estimating the paper’s regression model using their R code with fewer lags shows that the AIC and BIC are smallest (i.e., most negative) when there is only one temperature lag.

Using only one lag of the temperature variable makes a profound difference in the model’s estimates. If we re-simulate the paper’s statistical model using the RCP 8.5 climate scenario with one lag of the temperature variable, economic damages get to about 10% by 2100 rather than 60% as claimed in the paper.

To produce their estimates of economic damage resulting from climate change, the authors simulate the model’s estimation results using two standard climate scenarios, RCP 2.6 and RCP 8.5. RCP 2.6 and RCP 8.5 project the ten temperature and precipitation variables to 2100. Using those scenarios, the authors calculate the median loss in global real income to 2100 under the two climate scenarios. As we see, the real economic damage is in the 19% range for RCP 2.6 and RCP 8.5.

Under RCP 8.5, a climate scenario in which greenhouse gas emissions increase without any mitigation measures, loss in global real income is a whopping 60% by 2100.

Looking at their estimates it appeared that one temperature variable dominated. I therefore zeroed out the coefficients of all the other variables, only keeping the one temperature variable, and re-simulated their model, using their python code. The chart below shows that only one variable in the model actually matters, producing almost all the losses.

Economists make the distinction between statistical significance and economic significance. The chart above shows that even if the omitted temperature and precipitation climate variables are statistically significant (in fact most aren’t), they are not economically significant.