Weekly insight and analysis: 6th June 2024
The fog of war – making sense of the polling situation
There are nineteen different organisations publishing voting polls in this General Election campaign, more than ever before. With roughly three new polls published every day, we are flooded with information about the level of support for each party. To make this torrent of data more digestible, the media defaults in its reporting to showing the average of all the polls. Looked at in that way, very little has changed for at least the last eighteen months, with Labour maintaining an apparently unwavering lead of around 20%.
Looking at the average, however, masks the fact that there are significant differences between different pollsters. Taking the most recent polls over the last week, Conservative support ranges from 20% to 28%, Labour support from 41% to 48%; Labour’s lead is somewhere between 14% and 26%, alternative scenarios that represent hugely different election outcomes.
This variance in poll results stems from important differences in poll methodologies, for instance how they source the people to interview; how precise they are in the demographic profile of the sample they collect; how they try to ensure that their sample is attitudinally as well as demographically representative of the whole country; how they ask the questions; how they estimate who will vote at all; and what they do about people who say ‘Don’t Know’?
Each of these factors can lead to differences in the final numbers, but the biggest factor of all is the ‘Don’t Knows’. Most pollsters simply exclude them. There is a logic to this, but it implicitly presumes that those who can’t or won’t answer the voting intention question will end up splitting between parties in more or less the same proportions as those who do answer the voting question – and the evidence of many past elections is that this assumption is wrong. Some other pollsters therefore ‘squeeze’ the ‘Don’t Knows’ with follow-up questions (e.g. ‘Who would you vote for if it was compulsory?’ or ‘Which party are you leaning more towards voting for?’). A few go further than that, modelling the data to take account of broader views and the demographics of previous elections, inferring the levels of party support rather than relying only on the direct answers to voting questions.
