Almost a month on from the referendum campaign I’ve had chance to sit down and collect my thoughts about how the polls performed. This isn’t necessarily a post about what went wrong since, as I wrote on the weekend after the referendum, for many pollsters nothing at all went wrong. Companies like TNS and Opinium got the referendum resolutely right, and many polls painted a consistently tight race between Remain and Leave. However some did less well and in the context of last year’s polling failure there is plenty we can learn about what methodology approaches adopted by the pollsters did and did not work for the referendum.
The most obvious contrast in the campaign was between telephone and online polls, and this contributed to the surprise over the result. Telephone and online polls told very different stories – if one paid more attention to telephone polls then Remain appeared to have a robust lead (and many in the media did, having bought into a “phone polls are more accurate” narrative that turned out to be wholly wrong). If one had paid more attention to online polls the race would have appeared consistently neck-and-neck. If one made the – perfectly reasonable – assumption that the actual result would be somewhere in between phone and online, one would still have ended up expecting a Remain victory.
While there was a lot of debate about whether phone or online was more likely to be correct during the campaign, there was relatively little to go on. Pat Sturgis and Will Jennings of the BPC inquiry team concluded that the true position was probably in between phone and online, perhaps a little closer to online, by comparing the results of the 2015 BES face-to-face data to the polls conducted at the time. Matt Singh and James Kanagasooriam wrote a paper called Polls Apart that concluded the result was probably closer to the telephone polls because they were closer to the socially liberal results in the BES data (as issue I’ll return to later). A paper by John Curtice could only conclude that the real result was likely somewhere in between online and telephone, given that at the general election the true level of UKIP support was between phone and online polls. During the campaign there was also a NatCen mixed-mode survey based on recontacting respondents to the British Social Attitudes survey, which found a result somewhere in between online and telephone.
In fact the final result was not somewhere in between telephone and online at all. Online was closer to the final result, and far from being in between the actual result was more Leave than all of them.
As ever, the actual picture was not quite as simple as this and there was significant variation within modes. The final online polls from TNS and Opinium had Leave ahead, but Populus’s final poll was conducted online and had a ten point lead for Remain. The final telephone polls from ComRes and MORI showed large leads for Remain, but Survation’s final poll phone poll showed a much smaller Remain lead. ICM’s telephone and online polls had been showing identical leads, but ceased publication several weeks before the result. On average, however, online polls were closer to the result than telephone polls.
The referendum should perhaps also provoke a little caution about probability studies like the face-to-face BES. These are hugely valuable surveys, done to the highest possible standards… but nothing is perfect, and they can be wrong. We cannot tell what a probability poll conducted immediately before the referendum would have shown, but if it had been somewhere between online and phone – as the earlier BES and NatCen data were – then it would also have been wrong.
People who are easy or difficult to reach by phone
Many of the pieces looking of the mode effects in the EU polling looked at the differences between people who responded quickly and slowly to polls. The BPC inquiry into the General Election polls analysed the samples from the post-election BES/BSA face-to-face polls and showed how people who responded to the face-to-face surveys on the first contact were skewed towards Labour voters, only after including those respondents who took two or three attempts to contact did the polls correctly show the Conservatives in the lead. The inquiry team used this as an example of how quota sampling could fail, rather than evidence of actual biases which affected the polls in 2015, but the same approach has become more widely used in analysis of polling failure. Matt Singh and James Kanagasooriam’s paper in particular focused on how slow respondents to the BES were also likely to be more socially liberal and concluded, therefore, that online polls were likely to be have too many socially conservative people.
Taking people who are reached on the first contact attempt in a face-to-face poll seems like a plausible proxy for people who might be contacted by a telephone poll that doesn’t have time to ring back people who it fails to contact on the first attempt. Putting aside the growing importance of sampling mobile phones, landline surveys and face-to-face surveys do both depend on the interviewee being at home at a civilised time and willing to take part. It’s more questionable why it should be a suitable proxy for the sort of person willing to join an online panel and take part in online surveys that can be done on any internet device at any old time.
As the referendum campaign continued there were more studies that broke down people’s EU referendum voting intention by how difficult they were to interview. NatCen’s mixed-mode survey in May to June found the respondents that it took longer to contact tended to be more leave (as well as being less educated, and more likely to say don’t know). BMG’s final poll was conducted by telephone, but used a 6 day fieldwork period to allow multiple attempts to call-back respondents. Their analysis painted a mixed picture – people contacted on the first call were fairly evenly split between Remain and Leave (51% Remain), people on the second call were strongly Remain (57% Remain), but people on later calls were more Leave (49% Remain).
Ultimately, the evidence on hard-to-reach people ended up being far more mixed than initially assumed. While the BES found hard-to-reach people were more pro-EU, the NatCen survey’s hardest to reach people were more pro-Leave, and BMG found a mixed pattern. This also suggests that one suggested solution to make telephone sampling better – taking more time to make more call-backs to those people who don’t answer the first call – is not necessarily a guaranteed solution. ORB and BMG both highlighted their decision to spend longer over their fieldwork in the hope of producing better samples, both taking six days rather than the typical two or three. Neither were obviously more accurate than phone pollsters with shorter fieldwork periods.
During the campaign YouGov wrote a piece raising questions about whether some polls had too many graduates. Level of educational qualifications correlated with how likely people were to support to EU membership (graduates were more pro-EU, people with no qualification more pro-Leave, even after controlling for age) so this did have the potential to skew figures.
The actual proportion of “graduates” in Britain depends on definitions (the common NVQ Level 4+ categorisation in the census includes some people with higher education qualifications below degree-level), but depending on how you define it and whether or not you include HE qualifications below degree level the figure is around 27% to 35%. In the Populus polling produced for Polls Apart 47% of people had university level qualifications, suggesting polls conducted by telephone could be seriously over-representing graduates.
Ipsos MORI identified the same issue with too many graduates in their samples and added education quotas and weights during the campaign (this reduced the Remain lead in their polls by about 3-4 points, so while their final poll still showed a large Remain lead, it would have been more wrong without education weighting). ICM, however, tested education weights on their telephone polls and found it made little difference, while education breaks in ComRes’s final poll suggest they had about the right proportion of graduates in their sample anyway.
This doesn’t entirely put the issue of education to bed. Data on the educational make-up of samples is spotty, and the overall proportion of graduates in the sample is not the end of the story – because there is a strong correlation between education and age, just looking at overall education levels isn’t enough. There need to be enough poorly qualified people in younger age groups, not just among older generations where it is commonplace.
The addition of education weights appears to have helped some pollsters, but it clearly depends on the state of the sample to begin with. MORI controlled for education, but still over-represented Remain. ComRes had about the right proportion of graduates to begin with, but still got it wrong. Getting the correct proportion of graduates does appear to have been an issue for some companies, and dealing with it helped some companies, but alone it cannot explain why some pollsters performed badly.
Another change introduced by some companies during the campaign was weighting by attitudes towards immigration and national identity (whether people considered themselves to be British or English). Like education, both these attitudes were correlated with EU referendum voting intention. Where they differ from education is that there are official statistics on the qualifications held by the British population, but there are no official stats on national identity or attitudes towards immigration. Attitudes may also be more liable to change than qualifications.
Three companies adopted attitudinal weights during the campaign, all of them online. Two of these used the same BES questions on racial equality and national identity from the BES face-to-face survey that were discussed in Polls Apart… but with different levels of success. Opinium, who were the joint most-accurate pollster, weighted people’s attitudes to racial equality and national identity to a point half-way between the BES findings and their own findings (presumably on the assumption that half the difference was sample, half interviewer effect). According to Opinium this increased the relative position of remain by about 5 points when introduced. Populus weighted by the same BES questions on attitudes to race and people’s national identity, but in their case used the actual BES figures – presumably giving them a sample that was significantly more socially liberal than Opinium’s. Populus ended up showing the largest Remain lead.
It’s clear from Opinium and Populus that these social attitudes were correlated with EU referendum vote and including attitudinal weighting variables did make a substantial difference. Exactly what to weight them to is a different question though – Populus and Opinium weighted the same variable to very different targets, and got very different results. Given the sensitivity of questions about racism we cannot be sure whether people answer these questions differently by phone, online or face-to-face, nor whether face-to-face probability samples have their own biases, but choosing what targets to use for any attitudinal weighting is obviously a difficult problem.
While it may have been a success for Opinium, attitudinal weighting is unlikely to have improved matters for other online polls – online polls generally produce social attitudes that are more conservative than suggested by the BES/BSA face-to-face surveys, so weighting them towards the BES/BSA data would probably only have served to push the results further towards Remain and make them even less accurate. On the other hand, for telephone polls there could be potential for attitudinal weighting to make samples less socially liberal.
There was a broad consensus that turnout was going to be a critical factor at the referendum, but pollsters took different approaches towards it. These varied from a traditional approach of basing turnout weights purely on respondent’s self-assessment of their likelihood to vote, models that also incorporated how often people had voted in the past or their interest in the subject, through to a models that were based on the socio-economic characteristics of respondents, modelling people’s likelihood to vote based on their age and social class.
In the case of the EU referendum Leave voters generally said they were more likely to vote than Remain voters, so traditional turnout models were more likely to favour Leave. People who didn’t vote at previous elections leant towards Leave, so models that incorporated past voting behaviour were a little more favourable towards Remain. Demographic based models were more complicated, as older people were more likely to vote and more leave, but middle class and educated people were more likely to vote and more remain. On balance models based on socio-economic factors tended to favour Remain.
The clearest example is Natcen’s mixed mode survey, which explictly modelled the two different approaches. Their raw results without turnout modelling would have been REMAIN 52.3%, LEAVE 47.7%. Modelling turnout based on self-reported likelihood to vote would have made the results slightly more “leave” – REMAIN 51.6%, LEAVE 48.4%. Modelling the results based on socio-demographic factors (which is what NatCen chose to do in the end) resulted in topline figures of REMAIN 53.2%, LEAVE 46.8%.
In the event ComRes & Populus chose to use methods based on socio-economic factors, YouGov & MORI used methods combining self-assessed likelihood and past voting behaviour (and in the case of MORI, interest in the referendum), Survation & ORB a traditional approach based just on self-assessed likelihood to vote. TNS didn’t use any turnout modelling in their final poll.
In almost every case the adjustments for turnout made the polls less accurate, moving the final figures towards Remain. For the four companies who used more sophisticated turnout models, it looks as if a traditional approach of relying on self-reported likelihood to vote would have been more accurate. An unusual case was TNS’s final poll, which did not use a turnout model at all, but did include data on what their figures would have been if they had. Using a model based on people’s own estimate of their likelihood to vote, past vote and age (but not social class) TNS would have shown figures of 54% Leave, 46% Remain – less accurate than their final call poll, but with an error in the opposite direction to most other polls.
In summary, it looks as though attempts to improve turnout modelling since the general election have not improved matters – if anything the opposite was the case. The risk of basing turnout models on past voting behaviour at elections or the demographic patterns of turnout at past elections has always been what would happen if patterns of turnout changed. It’s true middle class people normally vote more than working class people, older people normally vote more than younger people. But how much more, and how much does that vary from election to election? If you build a model that assumes the same levels of differential turnout between demographic groups as the previous election then it risks going horribly wrong if levels of turnout are different… and in the EU ref it looks as if they were. In their post-referendum statement Populus have been pretty robust in rejecting the whole idea – “turnout patterns are so different that a demographically based propensity-to-vote model is unlikely ever to produce an accurate picture of turnout other than by sheer luck.”
That may be a little harsh, it would probably be a wrong turn if pollsters stopped looking for more sophisticated turnout models than just asking people, and past voting behaviour and demographic considerations may be part of that. It may be that turnout models that are based on past behaviour at general elections is more successful in modelling general election turnout than that for referendums. Thus far, however, innovations in turnout modelling don’t appear to have been particularly successful.
Reallocation of don’t knows
During the campaign Steve Fisher and Alan Renwick wrote an interesting piece about how most referendum polls in the past have underestimated support for the status quo, presumably because of late swing or don’t knows breaking for remain. Pollsters were conscious of this and rather than just ignore don’t knows in their final polls, the majority of pollsters attempted to model how don’t knows would vote. This went from simple squeeze questions, which way do don’t knows think they’ll end up voting, are they leaning towards or suchlike (TNS, MORI and YouGov), to projecting how don’t knows will vote based upon their answers to other questions. ComRes had a squeeze question and estimated how don’t knows would vote based on how people thought Brexit would effect the economy, Populus on how risky don’t knows thought Brexit was. ORB just split don’t knows 3 to 1 in favour of Remain.
In every case these adjustments helped remain, and in every case this made things less accurate. Polls that made estimates about how don’t knows would vote ended up more wrong than polls that just asked people how they might end up voting, but this is probably co-incidence, both approaches had a similar sort of effect. This is not to say they were necessarily wrong – it’s possible that don’t knows did break in favour of remain, and that that while the reallocation of don’t knows made polls less accurate, it was because it was adding a swing to data that was already wrong to begin with. Nevertheless, it suggests pollsters should be careful about assuming too much about don’t knows – for general elections at least such decisions can be based more firmly upon how don’t knows have split at past general elections, where hopefully more robust models can be developed.
So what we can learn?
Pollsters don’t get many opportunities to compare polling results against actual election results, so every one is valuable – especially when companies are still attempting to address the 2015 polling failure. On the other hand, we need to be careful about reading too much into a single poll that’s not necessarily comparable to a general election. All those final polls were subject to the ordinary margins of error and there are different challenges to polling a general election and a referendum.
Equally, we shouldn’t automatically assume that anything that would have made the polls a little more Leave is necessarily correct, anything that made polling figures more Remain is necessarily wrong – everything you do to a poll interacts with everything else, and taking each item in isolation can be misleading. The list of things above is by no means exhaustive either – my own view remains that the core problem with polls is that they tend to be done by people who are too interested and aware of politics, and the way to solve polling failure is to find ways of recruiting less political people, quota-ing and weighting by levels of political interest. We found that people with low political interest were more likely to support Brexit, but there is very little other information on political awareness and interest from other polling, so I can’t explore to what extent that was responsible for any errors in the wider polls.
With that said, what can we conclude?
- Phone polls appeared to face substantially greater problems in obtaining a representative sample than online polls. While there was variation within modes, with some online polls doing better than others, some phone polls doing worse than others, on average online outperformed phone. The probability based samples from the BES and the NatCen mixed-mode experiment suggested a position somewhere between online and telephone, so while we cannot tell what they would have shown, we should not assume they would have been any better.
- Longer fieldwork times for telephone polls are not necessarily the solution. The various analyses of how people who took several attempts to contact differed from those who were contacted on the first attempt were not consistent, and the companies who took longer over their fieldwork were no more accurate than those with shorter periods.
- Some polls did contain too many graduates and correcting for that did appear to help, but it was not a problem that affected all companies and would not alone have solved the problem. Some companies weighted by education or had the correct proportion of graduates, but still got it wrong.
- Attitudinal weights had a mixed record. The only company to weight attitudes to the BES figures overstated Remain significantly, but Opinium had more success at weighting them to a halfway point. Weighting by social attitudes faces problems in determining weighting targets and is unlikely to have made other online polls more Leave, but could be a consideration for telephone polls that may have had samples that were too socially liberal.
- Turnout models that were based on the patterns of turnout at the last election and whether people voted at the last election performed badly and consistently made the results less accurate – presumably because of the unexpectedly high turnout, particular among more working class areas. Perhaps there is potential for such models to work in the future and at general elections, but so far they don’t appear successful.