I just realised something, if I'm reading this right the difference between the greens and NIP here in this sample isn't even as much as 1 person...`1/ https://twitter.com/TomNwainwright/status/1383010343551696897

in fact greens got one more 1st image is unweighted table see greens get 5 & nip gets 4 (still meaningless noise numbers for either), 2nd image is weighted table where both are on 5 because of weighting for likelihood to vote, and that's fine! It's a common methodology but 2/

When we get to the final total weighted w/ undecideds/refuseds removed both still at 5 but Nip has jumped to 2% and green remains 1%. How can that be when they both had 5? Well, remember the 5 is a weighted number, it cld well be not an integer but
the table may have rounded3/

and the 2% is clearly a rounded number as none of the % show decimal places, so my guess is that NIPs weighted total is 5 point something (for example something like 5.4) while greens is also but slightly lower (say 5.2) 4/

5/and that this produces a % of slightly above 1.5% for NIP which is then rounded up to 2% and slightly lower than 1.5% for greens which is rounded down to 1%... so this shock poll? The difference is less than miniscule

6/ it was already clearly too small to read anything from but this kinda really brings home how fucking bad this analysis was from s4l

to be clear my problem isn't that the poll is weighted for voter likelihood, that's common c practice (though everyone has a different opinion on how best to do so) but it shows that the 1% difference is likely mostly a result of rounding

The margin of error on these polls is already way too big to read much from a 1% difference, but this just shows another reason not to read too much from small differences