This post was prompted by the comments on my proposed updated 80K Hours Climate Change Problem Profile.
It’s important to make it clear up front that the surprising truth is that there is genuinely very little quantitative research into the impacts of climate change of 4C and above. The research which does exist is necessarily limited in scope and makes a large number of assumptions - many of which will tend to undervalue the overall impact of climate change.
In this post I examine four previous attempts to examine aspects of the impact of climate change and/or the cost-effectiveness of climate change interventions. Full details of these analyses are included below, but their headline figures are summarised here:
- 2016 GWWC estimate - 2.8C temperature increase by 2100 produces mortality estimates that [with proposed model fixes applied by me] suggest Cool Earth can save a life for ~$6,000, compared to $3,461 to save a life with Against Malaria Foundation [with enormous uncertainty about this estimate]
- 2018 Halstead Extinction Risk - <1% - 3.5% extinction risk (>10C of warming)
- 2019 Bressler Mortality Estimate - 4.1C temperature increase by 2100 results in 76 million deaths [provisional results from an in-progress PhD]
- 2019 Hillebrandt Cost-Effectiveness - 2.2C temperature increase by 2100 produces a SCC that [with proposed model fixes applied by me] suggests Cool Earth is 1.15x more effective than global health interventions [Range: 0.0000003x - 4,041x]
Based on currently announced national commitments, greenhouse emissions are likely to lead to global temperature increases of 2.3ºC-3.7ºC by 2100 with a 25% chance of exceeding 4°C based on current national policies. This suggests that (1) and (4) are undervaluing action on climate change since they are based on much lower levels of projected warming. Furthermore, (1) and (4) both have very large flaws in their methodology which are likely to dramatically under-value climate action - see below for full details.
(3) projects 76 million deaths over the period 2020-2100. This is of a similar magnitude to the total deaths caused by the second world war (70-85 million people over 6 years). This is also of a similar magnitude to the largest famines seen in the 20th century (1-2M people/year). These kinds of numbers give an idea of the scale of impact which we can expect if climate change of 4C happens.
(2) computed an existential risk of <1% - 3.5%. This risk is not accounted for in any of the existing cost-effectiveness analyses which only focus on the average case along with a high/low estimate of impact.
One of the central ideas in effective altruism is that some interventions are orders of magnitude more effective than others. There remain huge uncertainties and unknowns which make any attempt to compute the cost effectiveness of climate change extremely challenging. However, the estimates which have been completed so far don’t make a compelling case that mitigating climate change is actually order(s) of magnitude less effective compared to global health interventions, with many of the remaining uncertainties making it very plausible that climate change interventions are indeed much more effective.
Moreover, this result is reached when only considering the impact of deaths attributed to climate change. This seems like an enormously narrow lens through which to consider a problem which risks displacing hundreds of millions of people, threatening global food systems, causing massive species extinction, and could trigger climate tipping points that amplify all of these projected impacts. Given all of this, it seems extremely likely that climate change mitigation is actually at least an order of magnitude more cost-effective than the best available global health interventions.
Discounting of future values is a common practice in economics which has a huge impact on the forecast impact of climate change. Climate change is already impacting the world today and, if emissions continue, the impacts are expected to continue to get much worse. Many forecasts only choose to consider impacts within the 21st century, and hence the worst of these impacts will be at the end of that period. Taking a couple of exemplar years - 2050 and 2100 - these are roughly 30 and 80 years away. The impact of different levels of discounting is as follows:
- 1% -> 30 years: 74%, 80 years -> 45%
- 2% -> 30 years: 55%, 80 years -> 20%
- 3% -> 30 years: 40%, 80 years -> 9%
This means that if you choose to discount the future by 2%/year, then you are choosing to value impacts in 2100 as only 20% as important as if they were happening today. Therefore it’s important to ask what level of discounting is being applied when you look at climate impact forecasts.
If you believe that lives in the future are also valuable, perhaps even just as valuable as lives today, then you may choose a very low or even zero discount rate and this will have a very large impact on your resulting valuation of climate change impact.
1.2. Global Mortality
Some of the estimates below are expressed in terms of number of deaths per year. To put these numbers in context, it’s useful to have a few points of comparison.
- Globally, there are currently ~60 million deaths/year across all causes, including those related to age related deaths. This is forecast to grow to ~120 million deaths/year by 2100 due to population growth and an aging population [source]
- In the 20th century, the largest famines killed 10-20M people/decade, so 1-2M people/year, all of which happened when the world had fewer than 4 billion people [source]
- Since the 1960s, wars have killed at most 300K people/year [source]
- World War II killed 70-85 million people over 6 years, which is 11.7-14.2 million people/year, at a time when the world population was ~2.3 billion [source]
- In 2017, 437K people died from Malaria [source]
1.3. IAM Validity Concerns
Two of the estimates below are based on Integrated Assessment Models (IAMs). Serious concerns have been raised with the use of these models.
"In a recent article, I argued that integrated assessment models (IAMs) “have crucial flaws that make them close to useless as tools for policy analysis.” In fact, I would argue that calling these models “close to useless” is generous: IAM-based analyses of climate policy create a perception of knowledge and precision that is illusory, and can fool policy-makers into thinking that the forecasts the models generate have some kind of scientific legitimacy. IAMs can be misleading – and are inappropriate – as guides for policy, and yet they have been used by the government to estimate the social cost of carbon (SCC) and evaluate tax and abatement policies. What are the crucial flaws that make IAMs so unsuitable for policy analysis? They are discussed in detail in Pindyck (2013b), but the most important ones can be briefly summarized as follows:
1. Certain inputs – functional forms and parameter values – are arbitrary, but have huge effects on the results the models produce. An example is the discount rate. There is no consensus among economists as to the “correct” discount rate, but different rates will yield wildly different estimates of the SCC and the optimal amount of abatement that any IAM generates. For example, these differences in inputs largely explain why the IAM based analyses of Nordhaus (2008) and Stern (2007) come to such strikingly different conclusions regarding optimal abatement. Because the modeler has so much freedom in choosing functional forms, parameter values, and other inputs, the model can be used to obtain almost any result one desires, and thereby legitimize what is essentially a subjective opinion about climate policy.
2. We know very little about climate sensitivity, i.e., the temperature increase that would eventually result from a doubling of the atmospheric CO2 concentration, but this is a key input to any IAM. The problem is that the physical mechanisms that determine climate sensitivity involve crucial feedback loops, and the parameter values that determine the strength (and even the sign) of those feedback loops are largely unknown, and are likely to remain unknown for the foreseeable future.
3. One of the most important parts of an IAM is the damage function, i.e., the relationship between an increase in temperature and GDP (or the growth rate of GDP). When assessing climate sensitivity, we can at least draw on the underlying physical science and argue coherently about the relevant probability distributions. But when it comes to the damage function, we know virtually nothing – there is no theory and no data that we can draw from.
4. IAMs can tell us nothing about the likelihood or possible impact of a catastrophic climate outcome, e.g., a temperature increase above 5°C that has a very large impact on GDP. And yet it is the possibility of a climate catastrophe that is (or should be) the main driving force behind a stringent abatement policy."
[Pindyck, 2017, The Use and Misuse of Models for Climate Policy]
Further relevant criticism can be read in [Weitzman, 2011, Fat-Tailed Uncertainty in the Economics of Catastrophic Climate Change].
2. Climate Change Impact / Cost-Effectiveness Estimates
2.1. 2016 Giving What We Can Cost-Effectiveness
Giving What We Can (GWWC) describe their approach and results here. The approach can be summarised as follows.
- Social Cost of Carbon (SCC) rejected as an appropriate measure of impacts.
- WHO’s 2014 report “Quantitative risk assessment of the effects of climate change on selected causes of death, 2030s and 2050s” [source] selected as key source of mortality estimates. This report estimates incremental climate change related mortality in 2030 and 2050 for heat-related mortality; coastal flood mortality; diarrhoeal disease; malaria; Dengue fever; and undernutrition. The A1B emissions scenario is used which predicts 2.8C temperature increase by 2100 [source].
- The estimates in 2030 and 2050 are assumed to define a linear relationship between year and number of deaths. The central value is an increase of 201.2 extra deaths/year on top of a baseline of 241K/year incremental deaths in 2030.
- The causes of deaths in the WHO report only account for 5.1% of total mortality, so as a conservative estimate, all causes of death are assumed to scale by the same amount, so 201.2*(100/5.1)=3931 extra total deaths/year.
- Reducing emissions in a year delays some fraction of these extra deaths/year.
- Hence, the cost of an emissions reduction can be multiplied through to reach a cost per life saved. The central estimate is $97,300, and the most generous estimate is $32,700.
- No discounting is applied in the final reported figures, although the spreadsheet allows this to be added on at the end.
126.96.36.199. WHO Report Limitations
The projections in the WHO report [source] come with a large number of limitations and caveats. Many of these are described in the report itself - the authors are clearly aware of the great difficulty involved in producing these kinds of estimates. However, these limitations are so severe that the resulting numbers must be used with extreme caution. Let’s consider a couple of the sections.
Malnutrition modeling doesn’t account for: increases in extreme weather events, sea-level rise, changes in water demand, increases in pests and diseases, loss of income from land which becomes unproductive. [WHO 2014, p70]
“We believe our estimates should be considered very conservative for the following two reasons [...] our modelling does not include the impact of shocks; it considers stunting due only to expected average conditions” [WHO 2014, p96]
- There is no change in storm-surge frequency and intensity from baseline (but floodwaters are deeper with sea-level rise).
- People flooded on average once a year autonomously leave the area and are not at risk of flooding and hence mortality
- Sea level rise: Average global warming was 2.4°C by the 2050s and 3.8°C by the 2090s. This corresponds to global mean sea-level rises of 0.22 m by 2050 and 0.37 m by 2080.
This last assumption is clearly out of date with the IPCC forecasting 0.52m of sea level rise in a 1.5C of warming world by 2100 - “Model-based projections of global mean sea level rise (relative to 1986–2005) suggest an indicative range of 0.26 to 0.77 m by 2100 for 1.5°C of global warming” [IPPC SR15 - Summary for Policymakers]
The GWWC estimate doesn’t use the coastal flooding mortality estimates as the WHO report only forecasts these within broad bands (e.g. 10K-30K) [WHO 2014, p35] and the estimates don’t turn out to change between bands between 2030 and 2080. Given more recent estimates of much greater sea level rise, this no longer seems plausible.
188.8.131.52. Linear Assumption
The GWWC model relies on the assumption that the point estimates given for mortality in 2030 and 2050 can be extrapolated into a linear relationship. This seems like a deeply flawed assumption which is contrary to academic work such as this 2015 nature paper - Global non-linear effect of temperature on economic production [slides]. It also fails some basic sanity tests as the presented numbers claim that climate change is causing excess malaria and diarrhoeal disease deaths, but that as climate change worsens, it causes fewer of these deaths.
184.108.40.206. Expanding To All Causes Of Death
The GWWC model asserts “we can quite roughly estimate that mortality due to climate change might grow proportionally to current levels of mortality - that is, that these diseases which currently make up 5.117% of global mortality will make up 5.117% of additional mortality due to climate change and, hence, that deaths due to climate change are 19.54 times higher than estimated in the WHO’s assessment.”
This assertion is weak as the resulting estimate is dependent on the five estimates taken from the WHO report. The two largest terms are (1) “excessive heat”, rising at 2851 deaths/year between 2030 and 2050, from a baseline of 37K in 2030, (2) “malaria”, declining at 1369 deaths/year between 2030 and 2050, from a baseline of 60K in 2030. If the GWWC estimate had not included Malaria (by choice, or if Malaria had not been in the 2014 WHO report), then the change in deaths/year between 2030 and 2050 would have risen from 201/year to 1571/year. There were 435K malaria deaths in 2017 [source], which is ~0.7% of global deaths. 1571*100/(5.117-0.7) = 35.6K/year, rather than the original 201/year estimate. So the estimate of change in deaths/year is very sensitive to the choice of estimates to include before multiplying out.
Finally, it seems wrong to count projected reductions in malaria deaths against climate change action when the reduction in deaths is presumably primarily because of direct action against malaria. If the climate was not warming, you would expect malaria to be declining more rapidly, but the GWWC model seems to imply the reverse. In fact, there is a campaign to eliminate malaria by 2040 [source], that if successful, would further invalidate the GWWC model which attributes malaria death reductions to climate change until long after this date.
220.127.116.11. Lives Are Saved Every Year
This appears to be one of the biggest flaws with the GWWC estimate. The GWWC estimate works on the basis that reducing emissions saves some fraction of the increase in deaths that would have happened as a result of those emissions. However, this saving actually applies for every year after the emissions were reduced.
The world currently emits 37Gt CO2/year. Ignoring longer term CO2 absorption processes, assuming these emissions continued at that rate indefinitely, if emissions are reduced by 1Gt in one year, then atmospheric CO2 concentrations will be lower every subsequent year than they would have been otherwise.
So the question is, how many years of saved lives should be included in the calculation? In theory the correct number should be the time until a given emission of CO2 has later been recaptured and sequestered. We don’t expect to be able to recapture most emitted CO2, so a very conservative value to use would be to attribute 50 years of increased deaths to each emission. Hence, this increases the estimate of lives saved by a factor of 50x. This also ignores any other impacts of a given CO2 emission, some of which are actually or effectively irreversible, such as triggering climate tipping points, species extinction, and sea level rise.
18.104.22.168. Use Of Central WHO Estimates
Cells C46 - E50 contain the estimates of lives saved for a given emissions reduction. These cells follow the same format as the rest of the sheet, with a central, low, and high estimate. However, these estimated are all based on the central WHO estimates. The only variation comes from use of a (central, low, high) estimate for the cost per acre of land protected by Cool Earth and the downward effect of adaption.
22.214.171.124. Handling Of Projected Declines
In the areas of Undernutrition, Malaria, and Diarrhoeal deaths, the WHO estimates showed declining climate change attributed mortality between 2030 and 2050. Cells C48-C50 reverses the sign of these estimates, which means they add to the lives saved rather than subtracting from them. I can’t see any rationale for this.
2.1.3. Updated Estimate
I have attempted to produce an updated estimate with the following changes:
- I have removed consideration of malaria deaths which may have been entirely eliminated by 2040 and have adjusted the “Percentage of total deaths” figures downward by the approx 0.7% of global deaths caused by malaria today.
- Number of lives saved are taken to be 50x the reduction in per-year incremental climate attributed deaths.
- I have updated the low/high estimates to actually use the low/high estimates of climate mortality.
- Removing sign reversal from Undernutrition, and Diarrhoeal deaths.
The resulting central estimate is $5,886 per life saved, which is the same order of magnitude as the $3,461 quoted to save a life by the Against Malaria Foundation.
The low and high estimates end up being weird due to the methodology used in the original estimate. For example, the low estimates for malnutrition are that in 2030 there are 119,807 fewer deaths, which drops to 29,203 fewer deaths by 2050. This produces a “low” estimate that climate change increases malnutrition mortality by 4530 lives a year, compared to the median estimate of climate change reducing malnutrition mortality by 524 lives a year. These kinds of numbers lead to a revised range of cost per life saved of between a low of $3,819/life saved and a high of -$701/life saved. This seems entirely nonsensical to me.
My updated model is available here.
2.2. 2018 Halstead Extinction Risk
Halstead posted to the EA forums about his 2018 paper “Stratospheric aerosol injection research and existential risk”. This paper estimates the risk of human extinction from climate change by combining the following estimates.
10C of warming is chosen as the threshold above which climate change will cause human extinction.
Table 1 - Atmospheric CO2 Concentration in 2100 -> Probability
- 400 - 1%
- 500 - 5%
- 600 - 20%
- 700 - 30%
- 800 - 20%
- 900 - 15%
Note, the probabilities don’t sum to 100% - the 9% chance of >900 is ignored. The paper doesn’t explain why.
Table 2 - Probability of warming >10C, at each CO2 concentration -> Probability
- 400 - 0.2%
- 500 - 0.83%
- 600 - 1.9%
- 700 - 3.2%
- 800 - 4.5%
- 900 - 6.6%
Deducing from the estimates in Tables 1 and 2, the unconditional probability of existential catastrophe-level warming is ∼3.5%. I use Weitzman’s estimate of climate sensitivity because it attempts to account for climate feedbacks which are important from the point of view of existential risk reduction. However, Weitzman’s ECS estimate is highly controversial, and there are a few reasons to think it may be too high. Nordhaus (2011a, 2011b) has criticised Weitzman’s analysis of the sample of IPCC model probability distributions across ECS. Weitzman (2009a) has defended his approach and noted that even if Nordhaus’ approach is correct, the probabilities in Table 2 would be reduced by around 60%, which still suggests that the risk of existential catastrophe is ∼1.5%.
Thus, the headline estimate I have produced in this section is highly controversial and some lines of argument suggest that the existential risks of climate change are (much) lower, plausibly < 1%. This controversy should be borne in mind in what follows.
[Halstead, 2018, p5]
So the range <1% - 3.5% is the existential risk predicted by this paper.
The probabilities in the tables above come from a 2011 Weitzman paper “Fat-Tailed Uncertainty in the Economics of Catastrophic Climate Change”. This paper also included estimates of the probability of >5C of warming.
Table 3 - Probability of warming >5C, at each CO2 concentration -> Probability
- 400 - 1.5%
- 500 - 6.5%
- 600 - 15%
- 700 - 25%
- 800 - 38%
- 900 - 52%
Multiplying this through in the same way as before gives a 26.2% chance of greater than 5C of temperature increase. This is reassuring as it is (very roughly) in line with the 25% chance of greater than 4C temperature increase predicted here.
2.3. 2019 Bressler Mortality Estimate
Bressler is a Sustainable Development PhD candidate who is working on accounting for climate mortality in the calculation of the Social Cost of Carbon (SCC). This work extends William Nordhaus’ DICE Integrated Assessment Model (IAM). Bressler gave a public talk with some early results from his work in July 2019 and posted it to the EA forum.
This estimate is based on examining how climate change impacts global mortality in a future with 4.1C of temperature increase by 2100. The model predicts that over the next 80 years, 76 million cumulative additional deaths are caused. These are deaths from health impacts, increased murder, and intergroup conflict response. I have reached out to Bressler to find out more details about what specifics are included/excluded from these estimates. Accounting for these deaths triples the SCC estimate. It should be noted that these are all preliminary numbers.
The video shows a bar chart with the total deaths in each 5 year period between 2020 and 2100. The death rate is projected to have reached 2.18 million deaths/5 years by 2100.
2.4. 2019 Hillebrandt Cost-Effectiveness
Hillebrandt posted this estimate to the EA forum in October 2019. After posting, the estimate underwent a major update which changed the conclusion. I will only be discussing the updated version.
The model takes as an input an estimate of the SCC from a 2018 paper “Country-level social cost of carbon” of US$417 per tonne of CO2 (66% CI: US$177–805). This is computed on the basis of a 2% pure time preference discounting rate along with a 1.5% elasticity of marginal utility [See this for details on growth-adjusted discounting]. The paper uses RCP6.0 which is projected to result in 2.2C of warming by 2100.
The SCC is then normalised by the relative utility of $1 in a poor country versus the US - using a range of three values (13,610x, 1,260x, 120x). The result is multiplied further by a range of three values for the relative effectiveness of the very best interventions versus direct cash transfers (17.5x, 7.95x, 0.83x). Finally, the range of costs for offsetting/reducing emissions is taken to be ($232, $10, $0.02) based on a selection of scalable solutions.
The result is that climate change interventions are predicted to be X times as effective than global development: (0.0000003x, 0.004x, 4,041x).
126.96.36.199. Use of IAM based SCC
As per section 1.3. there are serious validity concerns with the IAM models which underly estimates of the SCC. It’s unclear to me whether these concerns apply entirely to the 2018 paper underlying this analysis as it implements its IAM differently.
“we used country-level climate projections taken directly from gridded ensemble climate model simulation data as well as country-level economic damage rela-tionships taken directly from empirical macroeconomic analyses. As climate and economic quantities are empirical in this analysis, these uncertainties are probabilistic in our output.”
The post on the EA forum does note in Appendix 3 that the validity of IAM models are questioned. However, the median SCC used from this paper of $477 is actually higher than other estimates which are often in the $50-$200 range.
188.8.131.52. SCC Excluded Costs
Appendix 1 of the EA forum post notes that the SCC used in the calculation excludes a number of factors which may turn out to be very important, such as tipping points, ocean acidification, sea level rise, and biodiversity loss. This is used to justify the use of a 10x higher SCC in the “pessimistic” case.
184.108.40.206. Use of High Cost Per Tonne of CO2 Averted
The cost per tonne of CO2 averted is taken from a sample of highly scalable interventions, with the lowest cost being $0.02. The median case is taken to be $10. However, this seems like a surprising choice given that an individual choosing to donate their money towards climate change today would definitely be able to find an intervention which was cheaper than this. The 2016 GWWC estimate used a cost $0.38/tonne for donations to Cool Earth.
In the section about the cost of abatement, the forum post quotes the v2.0 GHG abatement which was published in 2009 by McKinsey. The latest version is v2.1 from 2010, which is still very old at this point. A more recent paper from 2018, The Cost of Reducing Greenhouse Gas Emissions, computes an updated estimate. This paper says:
One sobering insight from the estimates in Table 2 is that many of the least-expensive interventions cover a small amount of CO2 reductions, whereas the scalable technologies that are at the center of discussions about a transformation to a low-carbon economy - electric vehicles, solar photovoltaic panels, and offshore wind turbines - are among the most expensive on the list.
However, the paper goes on to examine two case studies of solar power and electric cars and proposes that the initially high costs come down dramatically with deployment scale, and so using today’s prices is misleading.
Another datapoint to consider is the Drawdown Project, described on wikipedia
Project Drawdown is a climate change mitigation project initiated by Paul Hawken and climate activist Amanda Joy Ravenhill. Central to the project is the compilation of a list of the “100 most substantive solutions to global warming.” The list, encompassing only technologically viable, existing solutions, was compiled by a team of over 200 scholars, scientists, policymakers, business leaders and activists; The team measured and modeled each solution's carbon impact through the year 2050, its total and net cost to society, and its total lifetime savings.
The results were published in a 2017 book and all the writeups for the solutions are available online. The 80 solutions that it examined that use established technology, have an overall cost/tonne of $28.61. However, the estimated savings are $71.87/tonne, for a net saving/tonne of $43.25. The savings largely come from lower operating costs, so financing will likely be required to cover the initial capital costs of these solutions, which will in many cases pay for themselves over time.
Finally, a May 2019 EA forum post promoted research by “Let’s Fund” which promoted funding a thinktank to advocate for increasing government funding for clean energy R&D. The median projected financial return calculated by their fermi estimate was 28x.
2.4.3. Updated Estimate
I have produced an updated estimate with the following assumptions:
- Median SCC: $477 - no adjustment for over/under-estimation
- Income adjustment: 120x - this is conservative about how much more valuable $1 is in a developing country
- Cost per tonne: $0.38 - this is taken from the 2016 GWWC estimate
- GiveDirectly vs. global health interventions: 7.95x - Median Givewell charity effectiveness vs. cash
On the basis of these assumptions, climate change intervention is 1.15x more effective than global health intervention.
There is clearly considerable uncertainty in this result, given that the original estimate had a range of 0.0000003x - 4,041x, which is 10 orders of magnitude. However, I claim that the title claim of the original EA forum post, that “Global development interventions are generally more effective than Climate change interventions” is far too strongly worded.
My updated model is available here.