As we enter yet another period of COVID uncertainity over outcomes (due mainly to human behaviour - what does "baseline/new normal" contacts look like in an European Autumn/Winter) a reminder about models. There are at least 3 different types; explanatory, forecast and scenario.
Explanatory - usually retrospective data to fit an understanding of the world (say infection->hospitalistion/not->death/discharge) for time series. Examples: excess deaths attributable to COVID, vaccine efficacy models and biological properties of variants.
Forecast - fit an up to date time series to understand outcomes in the near future, sometimes just to understand "now" (hence "nowcasting"). Examples: R rate and near time extrapolation; hospitalisation capacity near term management (often not public).
Scenario - provide alternative possible outcomes to explore decision making in the mid to far future. Examples: Scenario planning by SPI-M, often with different decisions and uncertainity about key parameters.
These models inter-relate; importantly parameters /model structure (eg, waning immunity) determined by explanatory models often get folded into scenario and forecast models, but there is *not* one model to rule them all.
(the old maxim; all models are wrong, some models are useful comes out here.)
The different types of models have different utility criteria, and only forecast models have the instinctive "did it predict the future well". Explanatory models are whether the estimated parameters were "right" - at the very least stable and consistent with other data.
Scenario models are the hardest to "score" for utility because its utility is really did the decision makers make good decisions, and these decisions nearly always have more than just the scenario models as input (from other health concerns through to economics and politics).
• • •
Missing some Tweet in this thread? You can try to
force a refresh
