Sondagens UK: rescaldo final

Está tudo aqui, no UK Polling Report de Anthony Wells, ainda com mais umas sondagens do que aquelas que eu tinha incluído. Com a devida vénia, transcrevo. Para quê inventar?

So, with pretty much everything except Harlow counted, how well did the pollsters do? The bottom line is that everyone got it right - trebles all round! While NOP take the prize, having got the result exactly spot on, not only did all the pollsters get within the 3% margin of error, they all got every party’s share of the vote to within 2%. Basically, it was a triumph for the pollsters.

RESULT - CON 33.2%, LAB 36.2%, LD 22.7%
NOP/Independent - CON 33%(-0.2), LAB 36%(-0.2), LD 23%(+0.3). Av. Error - 0.2%
MORI/Standard - CON 33%(-0.2), LAB 38%(+1.8%), LD 23%(+0.3). Av. Error - 0.8%
Harris - CON 33%(-0.2), LAB 38%(+1.8), LD 22%(-0.7). Av. Error - 0.9%
BES - CON 32.6%(-0.6), LAB 35%(-1.2), LD 23.5%(+0.8%). Av. Error - 0.9%
YouGov/Telegraph - CON 32%(-1.2), LAB 37%(+0.8), LD 24%(+1.3). Av.Error - 1.1%
ICM/Guardian - CON 32%(-1.2), LAB 38%(+1.8), LD 22%(-0.7). Av.Error - 1.2%
Populus/Times - CON 32%(-1.2), LAB 38%(+1.8), LD 21%(-1.7). Av. Error - 1.6%

The other two pollsters, Communicate Research and BPIX, conducted their final polls too early to be counted as proper eve-of-poll predictions, but, for the record, both their final polls were also within the standard 3% margin of error. Their average errors were 0.9% for BPIX and 2.1% for Communicate.

The British Polling Council have a press release out with the same information (although they include the “others” in the average, and use rounded figures for the results, hence the slightly different figures. It doesn’t change the result - everyone was right and NOP did best).

What small errors there were did still tend to favour Labour rather than the Conservatives - of the seven polls above one (NOP) is spot on, one (BES) understated Labour’s lead, and five overstated Labour’s lead. This does suggest there may still be a lingering bias in the polls, but one that is now so small it is hardly worth worrying about. What is interesting is the comparison between the final result and the polls during the campaign - the results from YouGov during the campaign were pretty close to the final result throughout, especially after the first few polls that showed the parties neck and neck. In contrast during the campaign the phone pollsters showed some whopping great Labour leads that disappeared in their final polls - of all the phone polls during the campaign, only one (MORI/Observer, published on the 1st May), did not report a Labour lead larger than the 3% they finally acheived. Of YouGov’s last 10 polls of the campaign, 8 showed a Labour lead of 3 or 4 percent. It doesn’t, of course, necessarily mean that YouGov were right - the “real” Labour lead at that time could have been larger, only to be reduced by a late swing to the Lib Dems - hence the reason why we only compare the eve-of-poll predictions to the final result.

With results that are so close to one another, it’s very difficult to say that technique A worked and technique B didn’t. For what it’s worth, the spiral of silence adjustment made Populus’s final poll less accurate (ICM’s full tables aren’t available yet). At past elections ICM have carried out post election studies, ringing back don’t knows to see how they did vote in the end, so they will hopefully have a far better idea of whether the don’t knows behaved as they expected. If there were Bashful Blairites out there though, they seem to have been equally bashful about going to the polling station. That said, NOP also make a spiral of silence adjustment and it can’t have made their final figures any less accurate.

How did all the polls end up being so close, having been so different during the campaign? Well, I’m afraid there isn’t a simple answer and there’s no evidence of an evil polling conspiracy. The only pollster who I know for sure made methodological changes for their final poll was YouGov, who factored in likelihood to vote for their final prediction (it didn’t have a huge effect, but it did make their prediction more accurate).