National-level models

Meridian supports national-level models, although we recommend that you use the geo-level model when possible. The national-level model is a special case of the geo-level model with only a single geo. There is no separate model class.

The following parameter restrictions apply, and are automatically enforced, when you use the national-level model:

  • \(\eta_m=0\) and \(\beta_{1,m}=\beta_m \ \ \forall m\)
  • \(\eta^{(rf)}_n=0 \) and \(\beta^{(rf)}_{g,n}=\beta^{(rf)}_n \ \ \forall n\)
  • \(\xi_c=0\) and \(\gamma_{1,c}=\gamma_c \ \ \forall c\)
  • unique_sigma_for_each_geo = False

National-level versus geo-level modeling

Statistical modeling relies on identifying repeatable patterns in data, and this can be done much more effectively with geo-level data, assuming that the patterns are reasonably similar across geos.

The geo-level model pools data across geos to increase the effective sample size. It provides tighter credible intervals, provided the geos are similar in terms of the media impact mechanism as the model assumes. For more information see Geo-level Bayesian Hierarchical Media Mix Modeling.

Geo-level data also improves estimates for time-effects (such as trend and seasonality), due to the fact that there are multiple observations per time period to use for estimation. Geo-level data can support the use of more knots to model the \(\mu_t\) parameter. Often it is reasonable to use one knot per time period for maximum flexibility. However, national-level data has fewer degrees of freedom to spare for time-effects. For example, one knot per time period would completely saturate the model.

The advantages of a geo-level model are so strong, that if only national-level data is available for relatively few channels, we recommend inputting the national-level data across geos.