Hi folks,
my doctoral student investigates dynamic changes of the spread of conspiracy beliefs towards COVID (operationalized by the number of keywords in the series of tweeds on Twitter).
Our goal is to investigate of global events with a certain salience lead to a change of the trend. As we expect the trend being nonlinear, I would like to investigate effects via generalized additive models (GAMs). The problem is: if there is a structural break (a massive change following an event), GAMs are a bit problematic as they tend to "smooth over" the break. But even if the rise in the count of keywords is a bit slower, I have not clue how to investigate the effect of the event (coded as 0 before and 1 after) on the change of the trend.
Does anybody has an idea for can forward some references? Doing an interrupted time series analysis has the problem (AFAIK) that the trends before and after the events are linear.
All the best,
--Holger
Not sure if I could understand the task correctly, but one possible approach (used in predictive maintenance) is as follows:
You need to model your process without the influence of the event and then extrapolate it using the data collected before the event. Then if prediction intervals do not contain the data collected after the event, then there's high probability that the event influenced the process.
The key thing here is to find a good model to obtain adequate prediction intervals.
For this purpose let me recommend the coverage table and coverage chart demonstrated in this work (see slides 29-31):
Conference Paper Data Formats and Visual Tools for Forecast Evaluation in Cyb...
P.S. You can use fancharts such as those presented on slide 21. Forecast origin must be put before the event. Then you can compare PIs with real data obtained after the event. But in order to assess the adequacy of PIs it's better to use rolling-origin evaluation.
Hello, you might take a look at the attached method. I am not entirely sure that a GAM will work in this context. Good luck. David Booth
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series