The Scottish polls at the end of last week and the weekend were broadly clustered around a small No lead. There are a couple of days campaigning left, but perhaps a more likely route to a YES victory is if the polls are underestimating the level of YES support for some reason. Over the last couple of days I’ve seen several blogs or articles pondering whether the polls could be wrong, could they be underestimating YES or NO?
It would be hubris to suggest the polls couldn’t be wrong. Obviously they can. At most elections there are polls that perform better or worse than their peers, some of that is better methodology. When the polls are close most is probably just normal sample variation. That’s a matter for another time though, here I’m pondering more about the possibly that all the polls are wrong, the potential for a systemic bias with everyone a bit too yes or a bit too no. This is possible too – think of the way all polls overestimated Lib Dem support in 2010, or most famously how all the polls overestimated Labour support in 1992. How likely is that?
The Scottish referendum is a bigger challenge for pollsters than an election would be because it’s a one-off. In designing methodology for voting intention the experience of what worked or didn’t work at previous elections weighs heavy, and most companies’ weighting schemes rely heavily upon the previous election – if not directly through weighting by recalled vote, in using the data from the previous election in designing and testing other weighting targets. For a referendum you can’t take that direct approach, pollsters needed to rely more on modelling what they think is an accurate picture of the Scottish electorate and hoping it reflects the Scottish people well enough that it will also reflect their referendum voting intentions – it’s complicated because Scotland has a complicated electorate. Scottish voters have two Holyrood votes and a Westminster vote, and they use them all in different ways with different political loyalties. Within the space of a year Scotland managed to be a Labour stronghold at Westminster and to produce a SNP landside at Holyrood – using either election alone for weighting gives a rather different picture of what the Scottish electorate are like, even though you are trying to model the same population. Different companies have arrived at different methods of political weighting to deal with the issue – Survation, ICM and TNS weight by Holyrood recalled voted alone, YouGov weight by Holyrood recalled vote with a nod towards 2011 Holyrood voters who backed Labour in 2010, Opinium weight by Holyrood and Westminster recalled vote, Panelbase weight by Holyrood and European recalled vote, Ipsos MORI don’t use political weighting at all. Despite the variance they have all converged to produce the same sort of result, and that gives me some confidence – if there was a particular skew from being online or from using Holyrood recalled vote we would expect to see different results.
Most speculation about whether the polls might be wrong has – rightly in my view – concentrated on two particular issues. Very high turnout and differential response rate.
Polls aren’t very good at predicting an actual percentage for turnout – people overestimate their likelihood to vote, and the actual turnout figures they are compared to are a bit ropey because of inaccuracy and incompleteness of electoral registers – that aside, they are pretty good at predicting relative turnout, and the referendum looks set to have a much higher turnout than any recent election. This poses a problem. Any professionally run opinion poll will make every effort to get as representative a sample as possible, but in practice there are limitations. People on the very fringes of society, people struggling in absolute poverty, those utterly detached from mainstream politics and civic society – people on the extreme edge are probably underrepresented in opinion polls of all sorts. In most voting intention polls this doesn’t matter, as people on the very fringes are also extremely unlikely to vote… but if the Scottish referendum does manage to engage some who were previously totally detached and, crucially, those people vote in a substantially different way to other people of similarly marginal demographics, then it could be a source of error.
The second potential pitfall is differential response. Much of the media discussion around this has called it “shy Noes” – people who want to vote no but are reluctant to admit it to pollsters. That’s possible, but it should be much less of a problem with online polls when people are giving their opinion to an impersonal computer screen. I think there’s more risk from the other side of the same coin – “enthusiastic yesses”. It is very clear from activity online and reported campaigning activity that YES supporters are more enthusiastic, what if that is also reflected in responses to opinion polls? What if the yes supporter, full of zeal and keen to share their view, happily agrees to do the phone interview while the less enthused No supporter doen’t want to interupt their tea? Eagerly clicks on the email when the No voter doesn’t bother? Issues of differential response can be mitigated through careful sampling and political weighting but again, it can only go so far. Pollsters can make sure they aren’t getting too many people who voted SNP in 2011, but there’s not much they can do to be 100% certain they are aren’t, for example, getting too many Yes voting Labour voters and not enough No voting ones.
So, how confident am I about the polling in the Scottish referendum? Well, I suppose I’m fairly confident – if there was anything I thought we were doing critically wrong we’d have corrected it. If I had to put money on the result, I’d certainly back the polls, but the potentials for error are there. We’ll know on Friday if they’ve been avoided.