Read on Twitter
Good analysis this
on the recent #MRP model - to me, not-a-data-scientist, it looks like it's regressing to the mean, bringing extreme estimates closer to the national average. Also noticeable from the spread of estimates v '19 results, w/ narrower (smaller std dev) in the model https://twitter.com/OwenWntr/status/1346063882847191040
on the recent #MRP model - to me, not-a-data-scientist, it looks like it's regressing to the mean, bringing extreme estimates closer to the national average. Also noticeable from the spread of estimates v '19 results, w/ narrower (smaller std dev) in the model https://twitter.com/OwenWntr/status/1346063882847191040
MRP relies on national averages in part, so this is normal. But the beauty of it is it SHOULD allow for const results to come through more. Otherwise, might as well use uniform national swing. Which I did by quickly dropping the headline VI figs into the @ElectCalculus calculator
Some differences, most notably LibDems not coming out in such a dire position, but otherwise not miles apart. And also similar to a Uniform Regional Swing model (though LibDems propped up by a few v close ones <1%).
Is MRP useful? Yes! If it can tell us something about regional dynamics. For that, you have to first trust it. But in the UK it's only had 2 trials by fire, w/ mixed results. So I think jury is still out. But good that more models are out there that we can evaluate and learn from
In theory you'd want to use it for sub-group analysis because it tends to produce more precise estimates than just weighting your sample and looking at a group. So your 20k sample is in theory better used. Good review here https://medium.com/pew-research-center-decoded/comparing-mrp-to-raking-for-online-opt-in-polls-b05444d9931d
