The Paper That Breaks Climate Economics

Part 1 of 2 on a damning new paper that takes on the top-down climate-economics literature — “The empirically inscrutable climate-economy relationship”

Today, I discuss a new pre-print by Finbar Curtin and Matt Burgess of the University of Wyoming which is by far the most important climate paper I’ve read in quite some time — “The empirically inscrutable climate-economy relationship.”¹

Curtin-Burgess (CB26) ask a straightforward question: Can we actually measure how climate affects the economy from the historical record?

Their answer is no.

Economists have sought to identify a relationship of climate and the economy from the historical record as a basis for projecting into the future how changes in climate might affect economic growth. CB26 explain why meaningfully identifying that relationship is simply not possible.

In today’s post, part one of two, I explain the analysis of CB26 and take it further. I argue that the question they ask was never really answerable to begin with — the entire project of regressing aggregate economic output on aggregate climate variables is conceptually and fatally flawed.

This post briefly overviews the macroeconomics literature that has sought to connect climate and the economy, describes the theoretical framing of CB26, explains how I take their arguments even further, and sets the stage for Part 2 — which looks at CB26 replications of key climate-economics papers.

Grab a cuppa and settle in . . .

Bottom Up and Top Down

Two parallel traditions in climate economics ask two different questions.

The older tradition, which dates to William Nordhaus’s 1991 paper “To slow or not to slow” and the DICE model he introduced in 1992 asked: what is the right global carbon price?

Answers to that question estimate what has come to be known as the social cost of carbon — the dollar value of the damage of an additional emitted ton of CO₂.

This work typically follows a bottom-up approach: Nordhaus and those who followed him estimate climate damage sector by sector, such as agriculture, human health, energy demand and so on. Add up the sectoral damages, and get a total damage estimate.

Nordhaus won the Nobel Prize in 2018 for this work, and DICE remains the reference integrated assessment model for U.S. government analyses.

The bottom-up approach combines — sector-by-sector — various damage functions that relate temperature (typically) to economic outcomes. Sectoral damage functions make explicit the causal pathway from climate to economic impact.

The newer tradition — top-down — emerged in the 2010s and asks a related but different question: how does climate actually affect economic output in the historical record as a basis for projecting how changes in climate might affect the economy in the future?

The top-down approach seeks to bypass creating sectoral damage functions and instead looks to establish a relationship between climate variables (typically temperature) and aggregate economic output. The core idea is that once that relationship is established, it can be used to explore how future economic output may change based on changes in climate variables.

Dell, Jones and Olken (2012) is an early example of a panel regression of GDP growth on temperature. Burke, Hsiang and Miguel (2015) took this approach forward and produced the influential claim that unmitigated warming would cut global GDP per capita by 23 percent by 2100.

Many papers followed — Kahn et al. 2021, Kalkuhl and Wenz 2020, Kotz et al. 2024 (since retracted), Nath, Ramey and Klenow 2024, Bilal and Känzig 2026 — each relating an aggregate climate index to an aggregate economic index and interpreting the resulting coefficient as a causal damage function, which could then be used to project future damage as a function of future changes in climate.

Curtin and Burgess challenge whether any such top-down approach can answer the core question it asks and conclude that top-down approaches are fundamentally incapable of answering that core question.

CB26 take on an entire literature.

The Curtin and Burgess Critique

Curtin and Burgess ask what it would take for an aggregate top-down regression to actually recover a meaningful causal relationship, and they explore whether econometric methods allow for meaningful results.

CB26 argue that they do not: The climate-economy relationship, as measured from aggregate historical data, is — in their word — inscrutable.

To be clear, CB26 explicitly accept that human activity changes the climate, posing risks. In fact, their critique has nothing to do with the physics of climate science. Their critique is about economics, and specifically methods of econometrics.

CB26 show — rigorously and comprehensively — that the magnitude of any climate effects on economic growth, resulting from the top-down approach, cannot come out of the historical record with any confidence, no matter how sophisticated the econometrics.

The CB26 theoretical argument is straightforward but technically involved. To illustrate, I will use panel-data regression of country GDP on country temperature as a running example, because that is the approach most of the top-down literature uses. The general conclusion applies to any top-down approach.

The data are a panel — a table with one row per country per year. For instance, Rwanda 1994 is a row. India 1980 is a row. The United States 2010 is a row. Each row has a temperature, a precipitation, a GDP growth rate, and, depending on the approach, a handful of other variables. For example, 170 countries across 50 years gives 6,500 rows (170 x 50). That panel has two dimensions, country and year.

Climate economists want to know how a change in temperature causes a change in GDP growth. The methodological challenge is that temperature correlates with nearly everything. Countries with different temperatures have different institutions, religions, colonial histories, natural resources, neighbors, etc. Similarly, years with different temperatures have different oil prices, wars, technology, financial crises, etc. A naive regression of GDP on temperature risks mistaking spurious influences for a temperature effect.

Economists handle this challenge with a tool called fixed effects.

The Princeton University Research Guide on panel data explains:

“The fixed effects model assumes that the omitted effects of the model can be arbitrarily correlated with the included variables. . . . Fixed effects explore the relationship between predictor and outcome variables within an entity (country, person, company, etc.).”

In plain language: fixed effects tell the regression to throw away every kind of variation that is either specific to a country (and therefore potentially confounded with the country’s average climate) or specific to a year (and therefore potentially confounded with global shocks that hit every country the same way).

Country fixed effects remove everything specific to a country that stays constant over time — culture, institutions, geography, history. Year fixed effects remove everything that hit every country similarly in a given year — the 2008 financial crisis, the 1973 oil shock, global technology trends. What remains is a residual: India ran hotter than usual in 1987, cooler than usual in 1988 — and the regression uses that residual to identify the temperature effect.

This approach works only if the temperature-GDP relationship is the same everywhere, at every time. The regression pools every country’s residual and fits a single function — or a single quadratic, in the Burke-Hsiang-Miguel case — through the data.

CB26 explain:

“Many climate-econometric studies assume one global climate-economy relationship. They employ panel methods which treat every individual time-period observation as variation over this global function (e.g., Dell et al. (2012); Burke et al. (2015c); Kahn et al. (2021)).”

The relationship cannot be the same everywhere. Take two countries that look similar on key dimensions. CB26 point to El Salvador and Iraq: about the same average temperature and the same level of affluence. When the authors estimate a temperature effect for each country separately, they find a positive relationship for El Salvador and a negative one for Iraq. The average conceals a tropical climate (El Salvador) and a desert oil economy (Iraq), a remittance-dependent labor market and a war-scarred institutional environment.

Pooling El Salvador and Iraq into one regression is, according to CB26, methodologically unsound:

“Assuming a global climate-economy relationship means that points from one country implicitly affect the estimate of the climate-economy relationship for all countries. Fixed-effects models result in countries’ relationships being implicitly weighted by their variations in temperature (and/or other climate variables used), with larger-variance countries given higher weight. . . . The result is that relatively small countries at temperature extremes can have a large influence on global estimates.”

Not only is it problematic to pool relationships across space, it is also problematic to pool them across time. For instance, a climate-economy relationship in India in 1965 is not the relationship in India in 2015. Over that time India’s GDP increased by a factor of ten and saw expansion of air conditioning, irrigation, and a service sector.

Economic development changes how a country experiences weather: Kahn (2005) and Toya & Skidmore (2007) each found that a one-percent increase in real GDP per capita cuts death and damage rates from natural disasters by roughly half a percent. Rich countries and poor countries respond to identical weather differently, and the same country responds to identical weather differently as it grows rich. Pooling five decades of Indian data and looking for a temperature effect on GDP is, again, according to CB26, methodologically suspect.

Any relationship of climate and the economy varies across both of the panel’s dimensions and CB26 argue that assuming otherwise, as the top-down literature does, introduces bias.

An obvious fix would be to let the relationship vary — give every country its own coefficient, give every year its own coefficient, and let the coefficient depend on income and baseline climate. But that approach is also methodologically problematic because it turns a search for a fixed relationship into a need for as many relationships as there are specific times and places.

CB26 explain:

“Each data point in a climate-economy time series has a time index (t) and a space index (i, often country, sometimes region). If the climate-economy relationship has qualitatively meaningful variation across both space and time — as we argue it does — there are not enough degrees of freedom to estimate it. Assuming away some of spatial and/or temporal variation to preserve degrees of freedom creates qualitatively meaningful bias. This is the core, irreducible estimation challenge.”

So the top-down researcher has no way out. Let the climate-economy relationship vary as much as it actually varies, and it cannot be modeled. Hold it fixed, and the model can generate a headline-producing result, but that result has no real-world meaning in any specific place.

Curtin and Burgess describe this as the ”core, irreducible estimation challenge” that better data or cleverer methods cannot fix.

CB26 go to some length to explain that theirs is a statistical argument, not a climate argument. It is a problem any panel-data researcher faces when the when the relationship of interest varies along every dimension of the data. Climate econometrics sits in an especially bad spot because both dimensions of heterogeneity are large.

The methodological problem that Curtin and Burgess describe is well-established in modern econometrics — they cite Chernozhukov et al. 2013 from that literature.

What is new in the Curtin-Burgess paper is the argument that this weighting problem is not simply a minor methodological issue — it is a problem basic to the structure of every pooled estimate in the field.

Spatial and temporal heterogeneity are the core of the identification failure, but they are not the only problems, and Curtin and Burgess discuss three others.

• The first is the problem of growth miracles and growth disasters, terms CB26 use to describe a small number of extreme economic events in the historical record that have nothing to do with climate — such as the collapse of the Soviet Union, the Rwandan genocide, Iraq’s post-war rebound after 2003, or Oman’s 1968 oil boom. Such events produce single-year GDP changes ten-to-thirty-times larger than a country’s typical growth rate, up or down. When one of them happens to fall in a year of unusual weather in a small country, the regression treats the weather as the cause of the GDP swing. Curtin and Burgess show that six or nine such observations — out of roughly 6,500 in the dataset — drive a quarter to a third of the estimated climate damage in prominent papers (and will be discussed in Part 2).

• The second is that the climate a country has already adapted to is not observable in the data. An economist cannot distinuish damage from novel events from damage caused by expected events in a country being poorly matched to its environment. A cold country with poor winter infrastructure will suffer losses when winter happens, but calling that climate damage would obviously be misleading, because the cause is poor adaptation to documented climate variability.

• The third is that “climate” is not one number. Temperature, precipitation, humidity, extreme events, ocean circulation, and wind patterns all matter for how society functions and these variables interact with each other. Nearly every top-down study collapses all of that into annual-average temperature and total precipitation.

Why I Think the Problems Go Even Deeper

Here go beyond Curtin and Burgess — not to dispute their argument, which is solid, but to push it further than they do.

The authors treat the problems above as technical challenges in econometrics: Spatial heterogeneity introduces bias. Temporal heterogeneity introduces bias. Influential observations introduce fragility. Reducing climate to one variable introduces omitted-variable bias.

Their framing suggests that, in principle, a top-down regression done carefully enough with good enough data would recover a meaningful relationship if only researchers could overcome these obstacles. CB26 argue that in practice researchers cannot overcome the obstacles, so the relationship is “empirically inscrutable.”

Their theoretical model presumes that a well-defined climate-economy relationship exists and that the problem is aggregating it into a single estimable function based on available data.

I go further — the aggregate relationship, as the top-down literature conceptualizes it, simply does not exist as a coherent causal object at all.

Consider what the top-down regression actually does when seeking to relate an average temperature to GDP. Both are indices: No person, no crop, no factory, no piece of infrastructure ever experiences average temperature. A farm in Iowa experiences the temperature at that farm, minute by minute. Similarly, a factory in Bangalore experiences the temperature at that factory. An average temperature in these studies is built up by averaging gridded observations across a country’s land area — usually population-weighted — and then averaging again over the year. What enters the regression is a statistical summary of a summary.

GDP is a similar kind of index. Country GDP aggregates the output of millions of firms, each operating under local conditions, across sectors whose relationships to weather differ enormously. Global GDP aggregates country GDPs. No person, company, or government experiences GDP.

A defender of the top-down approach might reply that GDP is the variable we actually care about for policy, so treating it as downstream of the real mechanisms misses the point. That defense fails for a simple reason. GDP is the outcome we care about, and climate change is the cause we care about, but neither aggregate index is where the causal process actually plays out. The question is what causes changes in GDP — and those causes operate at the level where people, firms, and physical systems respond to their local conditions.

Let me illustrate the problem with an analogy. Imagine trying to learn whether paying baseball players more makes them hit better by comparing team payrolls to team batting averages. You would find a correlation — higher-payroll teams do tend to hit better on average, because they pay for talent — but the correlation tells you almost nothing about the underlying question about how pay influences a batter’s hitting.

A team’s payroll is the sum of what every player on the roster earns, spread across superstars and journeymen in different proportions on every team. A team’s batting average is a blend of what every hitter contributed, spread across starters and bench players in different proportions on every team across a long season. Whatever mechanism links pay to performance operates at the level of individual players in individual at-bats, and neither aggregate metric provides the data necessary to establish a causal relationship.

Correlating the two team averages can give a number, but it is a statistical artifact of thirty different roster structures over 162 games per team, with forty or so different players suiting up over a season, playing in different stadiums against different opponents.

Fancy econometrics can certainly produce a number that relates an index of batting to an index of salaries. The number will reflect some real pattern — higher-paying teams do get better hitters on average — but it will not answer the question a team owner actually cares about, which is whether spending more money on this roster or on new players would improve this team’s performance. That counterfactual lives at the level of individual players, individual contracts, and individual at-bats, not at the level of team or season aggregates.

The climate-economy regression is worse off, because the indices it correlates sit even farther from the underlying mechanisms than payroll and batting average do.

The bottom-up approach, for all its problems, at least tries to include plausible causal pathways between changes in weather variables and outcomes. You can look at a Nordhaus damage function and see exactly what the author thinks is happening in agriculture, on coastlines, in the energy sector as climate changes.

The top-down approach skips the causal mechanism step and assumes that the aggregate statistical relationship somehow reflects the sum of all causal mechanisms.

CB26 provide strong evidence that the assumption fails. The sensitivities they document across the top-down literature — 20, 30, 50 percent swings in headline estimates from minor specification changes; wildly different answers from papers using the same data; results driven by a handful of observations that have nothing to do with the weather — are exactly what we should expect from a regression that is correlating two indices with no stable causal pathway between them. If a real aggregate relationship existed and had a stable causal interpretation, it would show up consistently across specifications that reasonably isolate it. The top-down climate literature’s estimates do not show any such consistency.

The field has mistaken the inscrutability of a statistical artifact — a quantitative relationship between two abstract indexes — for uncertainty about a presumed-real quantity. The damage function the top-down literature claims to estimate simply does not exist as a coherent object.

If the problem were merely technical — better data, longer panels, more clever econometrics — the field could reasonably expect to close in on the truth over time. If the problem is conceptual — that the top-down approach is correlating two downstream summaries with no causal structure between them — then more data and cleverer methods will not help. The field will continue producing numbers and those numbers will generate headlines, but those numbers, whatever they mean, don’t tell us anything meaningful about how future changes in weather might affect future economic activity.

In Part 2, I walk through what Curtin and Burgess found when they reproduced the three most influential papers in the top-down literature — and what they document is devastating.

Read their pre-print here.

Comments welcomed!

Before you go: If you appreciate being kept up-to-date on the latest climate research, please let me know by clicking that “❤️ Like.” More likes means more readers of THB. Thanks!

THB exists because of its paid subscribers. Please consider an upgrade!

1. To state the obvious, this post reflects my interpretation of Curtin/Burgess.