Debating the Economic Cost of Climate Change

 Two prominent young economists have published a famous paper in the esteemed Quarterly Journal of Economics.

Bilal, Adrien, and Diego R. Känzig. “The macroeconomic impact of climate change: Global versus local temperature.” The Quarterly Journal of Economics 141, no. 2 (2026): 889-944.

I fed their paper to RefineInk. This is an AI that produces high quality referee reports. Note that their paper has already been published BUT the AI wrote a new 28 page report listing its concerns with their paper. Here is the Refine Ink report.

Here is my academic paper critiquing the 2026 QJE paper.

Here is the Burgess and Curtin paper that provides an excellent critique of this entire macro/climate change empirical literature.

In the Substack below, I dictated my thoughts to Zoom and I took my zoom transcript and fed this to Chat GPT. I also fed Chat GPT the Refine Ink report linked to above. Chat GPT crafted this essay based on my dictation and the Refine Ink report.

Can Climate Econometrics Take the Heat?

A $1,200 social cost of carbon rests on a view of the economy in which people learn too little and adapt too slowly

Two prominent young macroeconomists, Adrien Bilal and Diego Känzig, have written an important and now highly influential paper on the long-run macroeconomic effects of climate change. Published in the Quarterly Journal of Economics, their work estimates that a 1°C rise in global temperature produces large and persistent losses in world GDP. Their calculations imply a social cost of carbon exceeding $1,200 per ton.

That is an enormous number—far above the estimates associated with Michael Greenstone’s work or the Biden administration’s regulatory calculations. If Bilal and Känzig are right, the implications for energy, climate, and growth policy are profound.

But a large headline number is also a reason to look carefully under the hood.

Their paper has been refereed, presented widely, and debated by famous economists. I do not dispute that their statistical method can identify a particular kind of filtered temperature shock under their maintained assumptions. My concern is different: Is the estimated relationship a stable law that we can project far into the future? Or will people, firms, and governments learn, innovate, and adapt as the climate changes?

This is where the Lucas critique enters climate economics.

Only 67 Observations Carry a Great Deal of Weight

Bilal and Känzig work in the tradition of macroeconomic integrated-assessment models. Rather than studying how local weather affects particular crops, workers, cities, or firms, they use aggregate global time-series data. In the sample I analyze, the common period is 1950 to 2019—just 67 annual observations.

They use a version of James Hamilton’s time-series filter to construct global temperature “shocks.” They then estimate local projections asking how a temperature shock today predicts cumulative world GDP-per-capita growth several years later.

This is sophisticated econometrics built on a very short time series. With only about 65 years of usable data, every modeling choice matters. The definition of a “shock” matters. The assumption that the damage coefficient is stable over time matters. And the distinction between what the econometrician knows after seeing the full sample and what economic agents knew in real time matters enormously.

Under their statistical assumptions, the authors recover a large negative relationship. But those assumptions embed a surprisingly passive view of economic behavior.

The Boiling-Frog Economy

From roughly 1960 to the present, the world has not warmed in complete secrecy. Households, firms, governments, scientists, and entrepreneurs have gradually observed the trend. They have not known its exact speed or severity, but climate change has increasingly become a known unknown.

Economic agents facing a known unknown have incentives to learn. Risk-averse households seek information and backup options. Firms search for threats to profits and opportunities for new products. Governments build infrastructure. Entrepreneurs respond to emerging demand.

The alternative is what I call the boiling-frog economy. In that view, we sit passively in the pot, failing to notice that the water is warming. We make no precautionary investments, develop no new technologies, change no production methods, and never reconsider where or how we live.

That worldview conflicts with what I learned at the University of Chicago about rational expectations and investment under uncertainty. It also sits uneasily with Robert Lucas’s work on changing rules, Ed Prescott’s emphasis on equilibrium responses, and Paul Romer’s theory of endogenous innovation.

Over 65 years, people will notice a trend. Some adaptation efforts will fail, but others will succeed. Successful ideas can diffuse widely because knowledge is a public good. Better air conditioning, heat-resistant crops, improved medicines, water storage, backup power, tougher building materials, better forecasting, and movement toward safer locations all reduce exposure. As the market for climate safety expands, innovators have stronger incentives to supply it.

None of this means climate change is harmless. It means that the damage function is unlikely to be a timeless constant. I claim that the Social Cost of Carbon declines over time because of adaptation but that is not my main point here.

What Did People Know at the Time?

The central empirical issue is easy to state but requires some care.

Bilal and Känzig construct their temperature shocks using forecasting coefficients estimated from the entire sample. Their forecasting equation contains only lagged temperature values, so they are not literally placing future temperatures on the right-hand side. Estimating those coefficients with the full sample is a standard retrospective econometric procedure. If the temperature process is stable, this choice does not by itself invalidate their estimate of the historical relationship between temperature and economic activity.

But a retrospective statistical innovation is not necessarily the same thing as the news that households, firms, and governments perceived in real time.

Imagine an economic decision-maker in 1984. That person knows the temperature history through 1983 and forms beliefs using the information and forecasting models available at the time. In 1985, another observation arrives, and beliefs are updated. As evidence of a warming trend accumulates, the forecasting process itself may change. This is recursive learning.

The modern econometrician can instead use the entire subsequent temperature history to estimate a single forecasting relationship and then apply it retrospectively to 1984. That procedure can identify a useful statistical residual. But interpreting that residual as the surprise actually experienced by economic agents requires an additional assumption: that the econometrician’s stable forecasting law corresponds reasonably well to the beliefs people held at the time.

This distinction matters because the QJE paper eventually moves from estimating a historical reduced-form relationship to modeling expectations, investment, and the future economic consequences of permanent warming. If households and firms gradually learned about climate change, then their perceived innovations, expected future temperature paths, and incentives to adapt may have differed from those implied by a full-sample filter.

Neither the full-sample nor the recursive method perfectly reconstructs what people actually believed. But comparing them provides a useful diagnostic of how much the results depend on the assumed information structure.

My Re-Estimation

I conduct a preliminary re-estimation using recursive temperature forecast errors. Each year’s shock is constructed using only temperature information available through the previous year. I then estimate simplified local-projection regressions at three- and five-year horizons.

This is not yet a perfect apples-to-apples replication. My recursive and full-sample specifications differ in some sample choices, filter lag lengths, recession controls, and other details. A definitive comparison should hold the filter order, initialization period, estimation sample, and local-projection controls constant while changing only whether the forecasting coefficients are estimated recursively or over the full sample.

Even with that qualification, the results are informative.

At the five-year horizon, the full-sample Hamilton specification produces a climate coefficient of −9.273. The recursive specification produces a smaller coefficient of −7.073. The standard errors are quite similar, but the smaller recursive point estimate is no longer statistically significant at conventional levels.

This comparison does not establish that full-sample estimation creates an upward bias. It does show that the magnitude and statistical significance of the estimated damage are sensitive to how the underlying temperature innovation is constructed.

I also allow the climate effect to change over time. Under the adaptation hypothesis, the interaction between the climate shock and a time trend should be positive: the initial effect of heat may be negative, but the marginal damage should diminish as adaptation capital accumulates.

The interaction coefficients are positive, but the evidence is not decisive. At the five-year horizon, the interaction is 0.338 under recursive filtering and 0.337 under full-sample filtering. Recursive estimation therefore does not uncover an adaptation gradient that the full-sample method had concealed. Both specifications are directionally consistent with diminishing damages, but the short time series provides limited statistical power to distinguish adaptation from stability.

My claim is therefore modest. The recursive exercise does not overturn the Bilal and Känzig results. It shows that their headline estimate is not invariant to the assumed information structure, while neither approach directly recovers the beliefs or adaptation decisions of actual economic agents.

What Can the Full-Sample Method Miss?

My Monte Carlo simulations help clarify the conceptual concern. I simulate economies in which people learn and adapt and then compare full-sample and recursive methods for constructing temperature shocks.

In these simulations, the full-sample method can generate larger damage estimates and can make the response appear more stable over time even when the underlying economy is adapting. This does not prove that the same distortion drives the Bilal and Känzig findings. A simulation demonstrates a possibility, not an empirical verdict.

Recursive shocks bring their own problems. They are estimated using short and changing historical samples and may be noisy. More fundamentally, an econometrician’s recursive forecast is still not necessarily the forecast used by households, firms, or governments. People may have possessed scientific information, private signals, or subjective models that are absent from a univariate temperature equation.

The relevant choice is therefore not between a correct recursive shock and an incorrect full-sample shock. The deeper issue is whether either statistical residual can be interpreted as the information innovation that caused people to revise their expectations and behavior.

This creates a trilemma for reduced-form climate econometrics. It is extremely difficult to do all three of the following at once:

1. Estimate a credible causal effect of historical temperature variation.

2. Convert that estimate into the effect of permanent future climate change.

3. Determine how learning, innovation, and endogenous adaptation alter the damage function.

Bilal and Känzig make an important contribution to the first task. Their evidence is less decisive concerning the second and third.

From a Historical Coefficient to a $1,200 Social Cost of Carbon

The information-set question is only one part of the larger problem. Several additional steps separate the historical temperature coefficient from a social cost of carbon exceeding $1,200 per ton.

First, the paper uses the ratio of cumulative temperature and output responses over roughly ten years to infer the economic effect of a permanent 1°C increase. But global temperature remains elevated at the end of that window, and output has not fully returned to its baseline. A finite-horizon cumulative ratio is not automatically the exact long-run effect of permanent warming. What happens beyond the observed ten-year response matters enormously.

The structural model fills in that unobserved tail by imposing a particular parametric damage function. The chosen function eventually decays, ruling out a permanent effect on the economic growth rate. But alternative damage functions could fit the observed first ten years and imply different cumulative losses over the next century. The $1,200 estimate therefore depends not only on the historical data but also on how the model extrapolates beyond them.

Second, the reported confidence interval does not appear to propagate every relevant source of uncertainty. The calculations vary the scale of the damage function while holding its shape and persistence parameters fixed. Yet those persistence parameters help determine both the eventual size and timing of the economic loss. The reported interval of $399 to $2,015 per ton should therefore be interpreted as uncertainty conditional on the estimated shape of the damage function, not as a complete accounting of uncertainty surrounding the SCC.

Third, the physical transportability assumption deserves more attention. Historical innovations in global temperature may be exogenous to economic activity, but natural temperature variability and sustained greenhouse warming need not produce identical changes in oceans, storms, droughts, heat extremes, and geographic exposure for every additional degree of global temperature. Establishing the historical relationship does not, by itself, establish that the same coefficient applies to twenty-first-century anthropogenic warming.

None of these concerns proves that climate damages are small. They show that the $1,200 number is not delivered directly by a transparent historical experiment. It emerges from a chain of empirical and structural assumptions concerning the temperature innovation, the unobserved tail of the response, the physical consequences of greenhouse warming, future adaptation, and the propagation of statistical uncertainty.

Mother Nature Can Trigger the Lucas Critique

Lucas’s original point was that when government changes the rules of the game, households and firms re-optimize. A reduced-form relationship estimated under the old rules may not survive the new policy regime.

The same reasoning applies when Mother Nature changes the rules.

As climate risk becomes more visible, forward-looking people revise their beliefs. They invest in protection, demand new technologies, change production methods, move capital, and sometimes relocate. Entrepreneurs respond to that demand. In general equilibrium, those choices can flatten the relationship between climate change and subsequent economic growth.

Bilal and Känzig have produced an important paper. They are right to force macroeconomists to take climate change seriously, and their focus on global temperature raises valuable questions that local weather studies cannot answer.

But a credible estimate of a historical relationship between a statistical temperature innovation and subsequent GDP is not automatically a credible forecast of how a forward-looking economy will respond to permanent anthropogenic warming. Moving from one to the other requires assumptions about information, physical transportability, extrapolation, innovation, trade, adaptation, and equilibrium adjustment.

Before we treat a social cost of carbon above $1,200 per ton as a policy parameter, we should know how much of that number comes from the historical data and how much comes from the assumptions used to carry a short time series into the twenty-first century.

If the predicted future damages are truly enormous—millions of deaths and a much slower world economy—then it is especially implausible to assume that people, firms, and governments will remain passive.

The new generation of climate econometrics is sophisticated. But the more ambitious its forecasts become, the more seriously it must take the Lucas critique.

IF this argument interests you, take a look at my 2010 Climatopolis book and my 2021 Adapting to Climate Change books!!