When uncertain policy estimates make caution optimal

What the study found

The authors study a linear-quadratic model in which policymakers choose policy instruments to meet one or more targets, potentially over time. The effects of those instruments are subject to statistical uncertainty.

They report that inserting estimated effects directly into standard “sufficient statistics” policy formulas can produce infinite risk. When there is more than one target, they find that maintaining the status quo regardless of the data is the unique minimax-optimal rule. When there is more than one target and more than one instrument, it is also uniquely minimax-regret-optimal.

Despite those results, the authors recommend a dampened version of the plug-in rule. They say it outperforms doing nothing across most of the parameter space and has worst-case risk close to the minimax value in empirically relevant models.

Why it matters

The findings distinguish protection against the worst case from performance across a broader range of possible conditions. The authors also recommend using unbiased but dispersed causal-effect estimates, such as those from a local projection, rather than moderately biased but precise estimates, such as those from a vector autoregression.

Limits and unknowns

These are findings from a theoretical model, not evidence that a particular policy intervention works. The abstract does not specify the dampening formula, the empirically relevant models examined, or how performance varies across individual settings. The work is an arXiv preprint.