I was just reading Online Journal of Public Health Informatics - Nowcasting to Monitor Real-Time Mpox Trends During the 2022 Outbreak in New York City: Evaluation Using Reportable Disease Data Stratified by Race or Ethnicity which uses Nobbs ( Nowcasting by Bayesian Smoothing: A flexible, generalizable model for real-time epidemic tracking ) stratified across different strata and with different time windows.
It’s a nice paper and worth a read. It finds that the stratification helps, but there is no clear time window for fitting that is best and they have some problems with missing data (there is more than this, but that is a high level TLDR).
For people that aren’t aware, Nobbs is a Bayesian nowcasting method that uses a log random walk on cases with a non-parametric fixed delay distribution. As with many methods, it assumes you fit it to a rolling window (i.e D dates in the past for a given nowcast date). This is essentially a subset of the functionality that epinowcast supports.
This task strikes me as a perfect case study for using partial pooling on a time-varying reporting delay (i.e the delay varies based on the data within the D fitting window side stepping the issue of optimal length aside from computational issues) along with partial pooling across strata vs not doing those things (so a bit like Hierarchical nowcasting of age stratified COVID-19 hospitalisations in Germany • epinowcast but with a nice paper to base on).
It could also be a nice testing ground for parametric vs non-parametric. vs hybrid delays i.e. start with a non-parametric one to match Nobbs and then try variations.
Finally, they also note issues with missing onset dates which again could be something else to explore.
As these things are all already in epinowcast as features, it could be quite a straightforward project depending on how many combinations were tested