Well it's Friday morning where I am and it's a bit early for the pub so I'll post this here.
Here's a graph and I'll mumble about it below
Background:
A couple of weeks ago there was weeks worth of hospital admissions figures published in the pub which if you looked at them in isolation might have suggested a plateauing (
blue spots)
I wasn't convinced of this (mostly because in my mind this system seems to have a time constant that means a week's data is predominantly noise) and tried to put some trends to the data to try and help me see what might be going on.
At the time I took the last 30days admissions data and plotted both the linear (
faint red line) and exponential (
faint blue line) best fit curves
and mumbled out loud:
servodude wrote:seeing how the best fit curve changes for a running 30day period of a series could be interesting
why 30 days?
- I think it might be long enough to give scope for averaging out noise (e.g. variations in reporting during the week) and fit a trend to
- and short enough for that trend to be relevant; given the 1-2 week from infection to hospital admission that might be expected and that restrictions/guidelines/behaviour are frequently changing
- and it's easier than adjusting for a rolling month
so we've had a couple of weeks since (
green dots) and I thought I'd have a look at how the trends have changed
I've tried to take it out the "weekly" noise via a 7 day windowed average (
orange and yellow squares)
- it's plotted against the middle day of the window and I've used 30 days worth for the plot
- for this averaged set I've plotted linear (
solid orange line) and exponential (
solid yellow line) best fit curves
- I left a week off the end of this average to see how these curves fitted (or did not fit) the recorded data
Looks like projecting this (10 days ahead) exponentially overshoots a bit and projecting linearly undershoots by a bit more
so what next?
- if I shift the window to end at the last day I can average, the projection of the fitted curves for the next two weeks looks like
I think it takes about 2 weeks for any intervention to show in any figures
- so it will be interesting to see how wrong this is then
time for another coffee
- sd