The case below illustrates the limits of using a simple GLM (Poisson or NegBin) to capture temporal trends :
A better option would be to be able to retrieve growth rates from the asmodee models. This is a lot trickier, and we may want to think of a consistent interface for implementing this in trendbreaker. Some difficulties we will face include:
the best model may not have a 'time' variable
the best model may not be log-linear in time
the time variable could change name from one model to another
the time variable may not represent daily time units
the time variable could be impacted by an interaction term
The case below illustrates the limits of using a simple GLM (Poisson or NegBin) to capture temporal trends :![image](https://user-images.githubusercontent.com/10757154/116215010-92107200-a747-11eb-8f3a-e5d0e04ed358.png)
A better option would be to be able to retrieve growth rates from the asmodee models. This is a lot trickier, and we may want to think of a consistent interface for implementing this in trendbreaker. Some difficulties we will face include: