Assume we have a time machine and transport ourselves to the year 2100. Would we, with the benefit of decades of extra data, be able to use purely reduced-form econometric methods to estimate the overall economic damage from climate change? I say no. But I also say that we can make progress on understanding climate change impacts by being more syncretic in our methods and by leveraging knowledge from other (non-climate-related) areas of economics.
The argument follows from what’s laid out in the JEL paper on climate damages with Catie and Derek – specifically the connection between the conceptual framework in that paper and the idealized experiment that a causal inference approach would like to approximate and the trilemma that results. Climate change is persistent, anticipatable, and affects the whole world, so the ideal experiment would involve replicating the planet, having one replicate experience a different climate over many years, and then comparing the economic conditions across the two worlds.1
Even with an extra 75 years of data, I don’t see an empirical approach that would allow us to capture all of the important features of this ideal experiment. As we discuss in the paper, cross sectional comparisons of climate and economic output across locations fail to capture the widespread nature of climate change (they are only looking at local effects). Global time series approaches are the mirror image. To gain identification, they need to isolate temperature shocks which are not anticipatable and which have limited persistence.
With more data, one might consider using a global time series analysis that focuses on longer-run variation in temperature. To make things concrete, imagine regressing global output in a given year on global average temperature over the previous 30 years, the 30 years before that, and 30 years before that. A 90 year distributed lag model in long-difference form.
We can notice right away that even with many decades of additional data, this regression will have a tiny number of observations. Ok, so let’s take our time machine hundreds of years further in the future. Does that solve the problem? Again, not really, because we are inherently limited to comparing recent 30 year periods to historical 30 year periods – we don’t have replicate worlds, so we never have a contemporaneous control group. What else has changed between these 30 year periods aside from the climate?
Against epistemic nihilism
Even though I am pessimistic about the ability for purely reduced-form approaches to definitively determine the damages from climate change, that’s not a reason for nihilistically claiming that damages are unknowable. It makes me believe the way forward is to employ a mixture of approaches, taking the strengths from each. Indeed, I would say this is the central message of the JEL paper.
Employing a mixture of approaches starts with all of us being explicit about what our methods assume. For example, a panel fixed effects regression approach to estimating climate damages uses short-run, idiosyncratic variation that implicitly assumes people have a certain form of myopia. Thus, an analysis that looks at forward-looking behavior can help assess what might be missing from these estimates (one of the reasons for my. Similarly, models that incorporate trade can assess the effects of the implicit assumption in these estimates that there are limited spillovers.
In a real switch from the views I had 10 years ago, I’m also keen to see more from IAMs (integrated assessment models). Many IAMs have focused on assessing mitigation policy, with less attention paid to the damage function – especially transitory and durable adaptation including endogenous technical change. IAMs offer the possibility of combining econ knowledge from different (non-climate) settings with estimates of weather and climate damage. If your criticism of climate damage estimates is that they don’t include transition effects or durable effects, then we can use knowledge about the adjustment cost structure of the economy from macro to flesh out our estimates. If your criticism is that the estimates don’t include supply chain spillovers, we can similarly use knowledge from trade econ studies of supply chain effects more broadly. Etc., etc.
With the growing power of dynamic IO and heterogeneous agent macro, I could see a world where careful applied micro studies provide portable statistics for macro models that feature rich characterization of individual (firm and household) and public adaptation (like a supercharged Fried (2022)). The recent spatially granular macro climate models (review) are also in line with this idea. Although the total package of climate change effects is not identifiable, different studies and methodologies get at different slices of the problem – direct effects of local weather, spillovers, effects of past weather on durable adaptation. In principle, if we had a sufficiently rich and well-characterized model, we can use simulation to perform the world replication thought experiment I started the post with. This may feel a little sci fi, but it is less sci fi than a time machine!2
Version history
2026-04-17: First version
2026-04-18: Clarifying credit to Cannon
2026-08-04: Elaboration throughout, in response to conversations with James Rising. More discussion of comparing methods, clarification that IAMs are helpful (for this issue) because they allow us to use existing econ knowledge.
-
Even this might not be enough to capture important dynamics in climate impacts. Much of the economic damages from climate change are likely to be transition costs, which are not well captured by current estimates and which would require a more complicated experimental design subjecting replicate earths to climate change of different speed and duration. ↩
-
Credit to Cannon Cloud for the idea that rich IAMs, in effect, allow us to implement the ideal experiment and thus for helping me realize that my recent bullishness on a new wave of IAMs was indeed about their potential to break out of the climate damages trilemma we lay out in our review. ↩