Awesome project! I'm a big fan of using econometric methods in Quantified Self -type analysis.
A couple of thoughts on your regressions:
Keeping holidays in the data seems counterintuitive. I'm fairly sure just from eyeing the graphs that the increase in your work week might reach statistical significance if you were to use a regression method more resistant to outliers, like Least Trimmed Squares.
Another idea - in microeconometrics it's standard to use some type of Poisson regression four count data, which this could potentially qualify as.
And furthermore, I would have loved to see some sort of distributed lag model of the timeseries (or at least a scatterplot of all activities against each other) to see which ones tend to co-occur.
(Author here.) I'm very pleased to finally receive some feedback on the technical side!
I will have to check out Least Trimmed Squares (and maybe other robust regression techniques). For this dataset, I did try to exclude holidays, but it was difficult to define this exactly, so a more robust technique might indeed work better.
And yes, it would be interesting to see what tags co-occur. Good ideas for a future post. Thanks!
A couple of thoughts on your regressions:
Keeping holidays in the data seems counterintuitive. I'm fairly sure just from eyeing the graphs that the increase in your work week might reach statistical significance if you were to use a regression method more resistant to outliers, like Least Trimmed Squares.
Another idea - in microeconometrics it's standard to use some type of Poisson regression four count data, which this could potentially qualify as.
And furthermore, I would have loved to see some sort of distributed lag model of the timeseries (or at least a scatterplot of all activities against each other) to see which ones tend to co-occur.