Estimating dynamical systems — in particular identifying theirs parameters — involved in computational biology – for instance in pharmacology, in virology or in epidemiology – is fundamental to put in accordance the model trajectory with the measurements at hand. Unfortunately, when the sampling of data is very scarce or the data are corrupted by noise, parameters mean and variance priors must be chosen very adequately to balance our measurement distrust. Otherwise the identification procedure fails. A circumvention consists in using repeated measurements collected in configurations that share common priors – for instance with multiple population subjects in a clinical study or clusters in an epidemiology investigation. This common information is of benefit and is typically modeled in statistics by nonlinear mixed-effect models. In this paper, we introduce a data assimilation methodology compatible with such mixed-effect strategy without being strangled by the potential resulting curse of dimensionality. We define population-based estimators through maximum likelihood estimation. Then, from filtering theory, we set-up an equivalent robust large population sequential estimator that integrates the data as they are collected. Finally, we limit the computational complexity by defining a reduced-order version of this population Kalman filter clustering subpopulations of common observation background. The resulting algorithm performances are evaluated on classical pharmacokinetics benchmark. The versatility of the proposed method is finally fully challenged in a real-data epidemiology study of COVID spread in regions and departments of France.