Systematic review finds COVID behavior models advanced data use but stalled on structure and interdisciplinarity

The review screened 2,143 papers and retained 87 that explicitly modeled feedback between incidence and actions such as distancing or masking. Most relied on mobility or survey data collected after 2020, yet fewer than one-fifth tested alternative behavioral mechanisms against the same dataset. This pattern mirrors earlier compartmental models that treated contact rates as exogenous parameters, a choice that produced over-confident lockdown projections in 2020-2021.

Policy impact is immediate: governments used outputs from these models to set capacity rules, school closures, and workplace mandates. When behavioral response was misspecified, projected peaks diverged from observed waves by weeks, altering economic costs borne by specific sectors. The preprint notes that AI methods could close the gap but only if training data include repeated measures of the same individuals' risk perception and compliance.

Three priorities emerge for the next generation of models: standardized behavioral data pipelines, explicit comparison of utility-based versus norm-based decision rules, and routine involvement of decision-makers during model design. Without these steps, endogenous behavior modules will continue to lag behind the speed at which policies must be revised.

The main limitation is the absence of a quantitative meta-analysis of prediction error across the 87 studies; future work should benchmark against a common set of waves and jurisdictions.