Data Science, Econometrics and Climate Adaptation. OH MY!

A review of  Hsiang and Narita's  Adaptation to Cyclone Risk: Evidence From the Global Cross Section from 2012

This month, a world leaders and scientists are  engaging in a new round of climate talks in Paris. I personally believe climate change is one of the most important policy issue of the times, and hope you all have been following these talks in the news. When most people think about climate change, especially within the U.S. political context, they think about mitigation. Mitigation refers to efforts to curb carbon emissions and prevent further climate change. This is contentious for two reasons. First, some people like to claim that carbon emissions from humans are not the cause of climate change. Second, given our current technology it is legitimately expensive curb emissions, and in one way or another this will impose real costs on everybody in society. Of course, I agree with scientific consensus that human emissions are the cause of climate change. Most importantly, I have a great fear that the costs of not preventing climate change will be even more expensive in the long run than the costs of curbing them.

However, there is another part of the climate equation—one that is less controversial and sexy: adapting to climate change. Even if we stopped emitting all carbon today, warming would continue for decades.  Instead, we need to figure out ways to live with it. We can change our lives, our cities, and our resource planning to adjust for the uncertain future that comes with climate change. This is a topic that I could go on at length about. Wait, I did for nearly 200 pages in my PhD dissertation. I did some interesting (well, I guess the word "interesting" is debatable) research proposing planning approaches to adapt.

There is a whole other line of empirical research, examining the extent to which people already adapt to their climate. This research is suggestive of how some countries, states or cities may adapt in the future. To keep myself sharp on this topic, in this post I review paper estimating the extent to which countries have adapted to the threats of tropical cyclones (commonly known as hurricanes or typhoons). This is important for climate change, because all the evidence suggests that climate change will increase the frequency and intensity of tropical cyclones. The paper is Hsiang and Narita's  Adaptation to Cyclone Risk: Evidence From the Global Cross Section from 2012. And its really freaking cool, even if you don't have a PhD studying climate change!

What's so cool about this paper?

I'll be completely honest, the paper is a little dense and times confusing. I wouldn't exactly call it an accessible piece of research. But, what I really love about this paper is that it pushes some new frontiers in interdisciplinary research and highlights the value of cutting edge computing power. Even better this, it hides this in the context of an otherwise conventional econometric analysis.

The authors create a model to reconstruct the wind speed and energy of all 6,712 tropical cyclone between 1950 and 2008. While they don't get into the guts of this model too much in this paper, its pretty clear that this in not econometrics as usual. Instead, it seems to be a mix of meteorology and geography, with a dash of data science. They use a database tracking the trajectory of historical cyclones. Utilizing statistical relationships that describe the dynamics of a storm, they estimate the windspeed and energy of each storm on a geographic grid — think of a pixelated map, where each pixel represents a square with 0.1 degree of latitude and  0.1 degree of longitude (if you're interested, he describes this model in greater detail here, but be warned this is really in the weeds). He describes that the storms create a total of 6,443,520 data points. I kind of wish he'd write a paper specifically about the model, or better yet share the code with the rest of us. The results of these reconstructions are the dependent variables in a regression analysis, which I will get to soon.

Before I dive into the econometrics, I want to explain why I love this paper. First, it seems to be a completely novel approach to traditional econometrics. Pretty much every econometrics paper I have read (prior to this) depends on observational data, things that were actually recorded somewhere. This approach that approximates observational data with simulations and statistical inference. Doing this is no small task. These types of weather models are pretty fascinating in their own right. In Nate Silver's book, The Signal and The Noise, he describes the evolution of weather models from being based on pretty simple statistical models (a combustion of previous days weather and historical averages) to complex computational models that simulate the physics of weather system. I imagine this is equally fascinating. I also suspect the computational complexity and statistical techniques extend this research beyond econometrics and has much in common with modern data science methods. In fact, Hsiang is part of the data science faculty at U.C. Berkeley and has spoken at Strata (a major data science conference). I am quite obviously interested in the intersection of data science and economics and would love to learn a little bit about the computing tools it took to pull this off.

Economists do it with models

Enough gushing about the research methods, lets get to the paper itself. The authors begin with a model drawn from economic theory. This model calculates the expected capital of a region in a given period of time. I think its a little complicated to think about this in terms of capital, if that trips you up, just think wealth. In this model capital (wealth) in the current period is, is a function of capital (wealth) in the prior period, some growth rate, and the probability that some capital (wealth)  is  destroyed in tropical cyclone.

Next, he breaks down the probability that capital (wealth) is destroyed in a little more detail. Ther are three things that matter:

1. The risk of a cyclone hitting the region
2. The probability distribution of the intensity of the cyclone
3. The level of adaptation that nation undertakes, to decrease the damage, should storms of a given intensity hit. 

In this model, nations can choose their level of adaption. Countries choose this level to maximize their expected economic output* (a function of capital), less the cost of adaptation.

With this model in places, the authors can mathematically walk through some testable scientific hypotheses, using a method called comparative statics. I'll save you the math, but check out the paper if you are interested. He generates a number of hypotheses which he confirms in the data. I'll just talk about one, which he clearly things is the most interesting (and so do I).  For every country, you can calculate the relationship between damages and cyclone intensity. Countries that have a historical record of stronger cyclones will adapt. Stronger storms will generate relatively less damages in these countries than they do in countries with a record of weaker cyclones, all else equal.

Give me the numbers

In this paper, The authors do not directly measure wealth. Instead, they measures changes in wealth, by looking at a database of self-reported statistics, including deaths and economic damages. It may seem weird at first to think of deaths as a proxy for wealth. To me, it seems pretty reasonable to consider the lives' of citizens as a very important measure of wealth when considering decisions that nations make.

Before testing the hypothesis described above, the paper provides some useful summary statistics about tropical cyclones.  It contains a nice "rule-of-thumb": for the average country, the difference in being exposed to a Category 2 and Category 1 hurricane is double the economic damages and triple the deaths. 

But the analysis is not interested in the average country. It focuses on the differences between countries, particularly based on their historical record of cyclones. One figure in the paper, shown below, does a nice job illustrating the main hypothesis of the paper.  Each panel is a different country (Japan, Philippines, and Vietnam), with same graph. Each point on a graph represents the storms in that country in a given year. The X-axis measures the intensity of the storm, using the maximum windspeed (averaged across the land area of country). The Y-axis measures the number of deaths (normalized by population on logarithmic scale).

To read this, first focus on where each point falls on the X-axis, in each country. Vietnam has the lowest intensity storms over time, with an average maximum windspeed of just 10 meters per second (the vertical line). Japan is the highest, with just over 20, and Philippines is somewhere between them. Next, look at the slope of the relationship between deaths and windspeed. Japan has the shallowest slope, suggesting that as storms get stronger, the rate of deaths increases more slowly than it does in Vietnam. This relationship doesn't just hold for these three examples. Its statistically significant, across both measures (deaths and economic damages), fora couple of different model specification. 

This suggests that countries do, in fact, adapt. Countries with more stronger storms are better at preventing damages. More importantly, the statical analysis calculates the rate at which countries adapt. It's pretty small, countries only prevent about 3% of the expected damages for every 1 m/s increase in windspeed by adapting. This is important, some other climate change projections, called integrated assessment models, assume that countries can adapt 30% of damages away in the future. And these integrated assessment models are the basis of a lot of important climate analyses.

The authors offer a suggestion for why this may be the case. They say, the statistical analysis suggests cost of adaptation is high. This is in contrast to integrated assessment models, which think there is a lot of low hanging fruit and that we can adapt a fair amount for relatively cheaply. By giving it us some insight into how countries have adapted in the past, the authors suggest we can learn something about the cost of adaptation.

I've got a bone to pick with you

While I am impressed by the statistical analysis, I worry the suggestion that adaptation is expensive interprets their own model a little too literally. The theoretical model behind this assumes that countries are rational actors that can just optimize their level of investment to increase their wealth. Of course, we know that a country is not simply one rational actor, but instead represents a lot of people and groups. Some investment in adaptation will come from governments. However,  governments don't necessarily optimize investments: they are made up of "agents" who may have other incentives. Other investment in adaptation will come from individuals, households, of companies. But some of this investment has positive externalities, which means some of the benefits will go people other than those who bought it. In these situations, firms and individuals notoriously invest less in than would be optimal for society. 

This is all a long way of saying,  there is not any guarantee that this 3 percent rate of adaptation represents some optimal level. With attention turning to climate change and ever increasing budgets to address climate change, there may be political will to adapt more. 

This statistical approach doesn't get into any of the guts about what adaption might mean.  Sometimes, it is physical investments in levees in New Orleans, or it could be a government department creating a better evacuation plan. New technologies, or investments in new technologies can make adaptation cheaper. For example, people I used to work with at RAND used a ton of analytical fire power to suggest robust investments in Louisiana. Better decision tools make investments cheaper, by preventing spending on the least valuable forms of adaptation. These decision-making tools wouldn't have been possible without state-of-the-art computing infrastructure. Increased focus on climate change can generate investments in the technology to make adaptation cheaper. 

Maybe the integrated assessment models are too optimistic about the costs of adaptation. But, I don't believe that simply seeing that countries have adapted far less in the past suggests that future adaptation is too expensive.

* For those of you that don't speak economics nerd, expected with wealth is the amount is essentially the amount some one would earn on average from a risky bet. So, if there is a 50 percent chance of winning $100 and a 50 percent chance of losing $100, there expected payoff is zero.