When a government cuts taxes during a recession and the economy recovers, did the tax cut cause the recovery? When stimulus spending coincides with rising employment, how much of the gain can we credit to the policy? These questions sit at the heart of fiscal policy evaluation, and they are harder to answer than they appear.

The fundamental problem is that we never observe the alternative. We see what happened with the policy, but not what would have happened without it. Economists call this missing observation the counterfactual, and constructing it credibly is the central challenge of empirical policy analysis.

This matters because fiscal decisions involve enormous sums and lasting consequences. Multiplier estimates that differ by a factor of two can justify entirely different policy responses to the same downturn. Understanding how researchers build counterfactuals, and where those constructions are strong or fragile, is essential for anyone interpreting fiscal evidence.

The core empirical challenge in fiscal analysis is endogeneity. Governments rarely change taxes or spending randomly; they respond to economic conditions. This makes simple correlations between fiscal variables and outcomes deeply misleading. If spending rises when growth slows, a naive regression might suggest spending hurts growth, when in fact the causation runs the other way.

Researchers address this through identification strategies designed to isolate fiscal changes that are plausibly exogenous to current economic conditions. The narrative approach, pioneered by Romer and Romer, reads historical documents to classify tax changes by motivation, distinguishing those driven by long-run goals from those responding to short-run conditions.

Other strategies include using instrumental variables, exploiting predetermined components of government spending such as defense procurement, or leveraging cross-sectional variation across regions exposed differently to the same federal policy. Each approach trades off internal validity against external relevance.

No single strategy is definitive. A tax change cleanly identified in one historical period may behave differently under other monetary regimes or debt conditions. Credible fiscal analysis typically triangulates across methods, treating convergent estimates as more trustworthy than any single point estimate.

Takeaway

Correlation in fiscal data almost always reflects bidirectional causation. The work of identification is the work of finding policy changes that the economy did not cause.

Natural experiments occur when fiscal changes happen for reasons unrelated to the outcomes researchers want to study. Wartime spending shocks, court-mandated tax reforms, and sudden shifts in intergovernmental transfers all offer cleaner identification than routine policy adjustments.

The appeal is intuitive. When defense spending surges because of geopolitical events, the increase is not a response to domestic economic weakness. Comparing regions with high and low defense exposure during such episodes provides credible estimates of how government purchases ripple through local economies, an approach developed extensively by Nakamura and Steinsson.

Historical episodes also offer insight into rare conditions. Studies of fiscal consolidations across OECD countries, sovereign debt crises, and the postwar drawdown of U.S. spending have shaped our understanding of how multipliers vary with the state of the economy, the zero lower bound, and debt sustainability concerns.

Yet natural experiments come with caveats. Wartime multipliers may not generalize to peacetime stimulus. Effects estimated at the regional level miss aggregate consequences operating through monetary policy or exchange rates. The cleanest evidence often answers a narrower question than policymakers actually face.

Takeaway

Clean evidence usually comes from unusual circumstances. The price of identification is often relevance, and analysts must judge how far a result travels beyond its original setting.

Empirical estimates tell us what happened in particular episodes, but policy decisions require predictions about settings that may differ from the data. This is where structural models earn their keep. By specifying how households, firms, and governments behave, they allow researchers to simulate counterfactuals that no historical record contains.

A well-constructed dynamic general equilibrium model can explore how the same tax cut would perform under different monetary policy responses, different debt levels, or different degrees of household credit access. It can decompose observed outcomes into channels, separating the effects of expectations, crowding out, and direct demand stimulus.

The credibility of structural conclusions depends entirely on the credibility of the model's assumptions. Different frameworks deliver dramatically different multipliers from the same fiscal change. The discipline lies in matching models to empirical moments and stress-testing conclusions against alternative specifications.

The most useful fiscal analysis treats empirical and structural work as complements. Reduced-form estimates discipline model parameters; models extend reduced-form findings to policy-relevant counterfactuals. Neither alone provides a complete picture, and treating either as definitive invites overconfidence.

Takeaway

Models do not replace data, and data does not replace models. The frameworks supply the logic that lets numbers speak to questions the data never directly addressed.

Counterfactual analysis is less a technique than a discipline. It forces analysts to articulate exactly what they are comparing and to defend the comparison against alternative explanations.

For fiscal policy, where stakes are high and ideological priors are strong, this discipline is what separates credible evidence from motivated reasoning. The goal is not certainty but calibrated uncertainty, knowing what we know and how well we know it.

Policymakers who demand a single number betray the underlying analysis. The honest answer is usually a range, contingent on conditions, and improved by combining multiple approaches that fail in different ways.