A planner operating with a fixed budget of preventive assistance faces a fundamental timing question: whom to help and when. The act of waiting generates valuable information about latent risk, but each period of observation also eliminates future intervention opportunities as higher-risk individuals exit the pool. This work characterizes the identification tradeoff that emerges under repeated noisy signals and risk-dependent attrition.
The canonical model establishes that a observation prefix of length τ identifies exactly the latent risk moments m_j = E[p^j] for j = 0, …, 2τ−1, effectively buying two new moments per additional period. These identified moments shed light on average next-period attrition among survivors, though they do not necessarily identify allocation welfare or attrition among the subset prioritized for treatment. The authors construct observationally equivalent populations that demonstrate unavoidable timing regret at any sample size when observation length is fixed, meaning no estimator can reliably pinpoint the optimal intervention date without additional assumptions.
Exact certificates are derived to determine whether all compatible populations share a unique optimal date. When such a certificate does not hold, the planner faces a fundamental ambiguity: multiple rate distributions consistent with the observed signals support different optimal intervention timing. Matched experiments in the paper separate three practical sources of variation: estimation error, bootstrap averaging, and regret weighting. The experimental results show that bootstrap averaging and early stopping tend to improve allocation decisions, while regret weighting does not reduce pooled mean squared regret. In other words, a better statistical fit does not guarantee better welfare outcomes.
Moreover, the analysis demonstrates that longer observation windows often resolve timing ambiguity, but by the time sufficient data accumulates, the originally optimal intervention date may have already passed. This creates a practical tension: gathering more information improves identification but may render the intervention moot for the highest-risk individuals.
Code and supporting artifacts are released alongside the paper to enable replication and extension of these identification results.
How observation length gates identification
Under the canonical specification, each additional period of observation contributes two newly identified moments from the latent risk distribution. The moments m_j = E[p^j] for j up to 2τ−1 are functionally determined by the τ-length observation prefix. This means that with τ = 1, the first two moments of risk are revealed; with τ = 2, moments up to order three are identified, and so on. The paper proves that this incremental identification structure is both necessary and sufficient for the claims made about average attrition among survivors.
Observationally equivalent constructions and unavoidable regret
The authors construct explicit families of rate distributions that are observationally equivalent across the identified moments yet support different optimal intervention dates. These constructions establish that, for any fixed observation length, there exist compatible populations for which no estimator can guarantee recovery of the uniquely optimal timing. The regret that stems from this ambiguity is inherent: it persists regardless of sample size and is not a consequence of finite-data estimation error.
Matched experimental decomposition
Through a series of matched experiments, the paper isolates the contribution of three operational factors: estimation, bootstrap averaging, and regret weighting. The estimation component captures the error that arises from finite samples. Bootstrap averaging, which averages across resampled datasets, consistently improves decision quality in the experimental setting. Regret weighting, which adjusts the objective function to penalize timing mistakes, adds no measurable improvement in pooled mean squared regret. These results suggest that the dominant barrier to better allocation welfare is not statistical noise but the structural identification limits documented earlier.
Practical trajectory: longer observation versus lost opportunity
The analysis confirms that extending the observation window typically sharpens identification, reducing ambiguity about the optimal intervention date. However, this benefit carries a cost: as more periods elapse, an increasing share of the highest-risk individuals may have already exited the pool, either through the process under study or through attrition unrelated to the intervention. The net effect depends on the relative rates of risk revelation and risk loss, and the paper’s experiments illustrate scenarios where additional observation time shifts the optimal date into a window where the target population has already thinned.
In summary, the work establishes that a planner with a fixed budget and repeated noisy signals faces unavoidable timing regret unless additional structural assumptions are imposed. Bootstrap averaging and early stopping are the practical levers that improve decisions in the identified setting, while regret weighting offers no reliable benefit. The findings have direct implications for resource allocation in settings ranging as from public health intervention scheduling to personnel development timing, where the decision-maker must balance information accumulation against the diminishing pool of eligible recipients.