A rater niche and slightly geeky look at trend lines in general and COVID-19 case trends in particular. I got nerd-sniped a while back by @chadskelton to write this up, and also had discussions with @tom_cardoso and @robroc about this. doodles.mountainmath.ca/blog/2021/01/3…
TL;DR: Trend lines are an important tool to filter noise from signal. Moving averages is one way to do this, but it's not good at removing noise and introduces a data lag. Using methods like STL that can account for the weekly pattern and don't add extra data lag are preferable. 
Moving average trend lines do have the advantage that they are simple to understand, but get misleading when they are artificially shifted to the right to obscure the data lag. This problem is particularly acute when overlaid over case counts as for example in this graph.
For BC data there the an added problem of right-censoring, with subsequent data releases changing the number of cases on previous days. This systematically biases trend lines down. Less of a problem for moving averages because of their data lag, but problem for modelling.
The effect on the tail of trend lines is not huge but clearly visible. The impact on projections from modelling can be quite a bit larger though, especially for models that emphasize the last week of data points for their projections.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
