As polling companies try out different methodologies in the light of recent reviews of their practices, Ailsa Henderson explains how they determine who is likely to vote.
Different polling companies weight their respondents by a range of demographic characteristic to ensure that their quota samples more closely reassemble the electorate as a whole. Most use age, sex and social class. Others add in additional demographic variables like housing tenure, or political variables such as interest or past vote. In addition, when seeking to indicate likely vote intention, they employ a variety of measures to ensure that they are capturing those most likely to vote. The three methods can be summarised as follows:
Filtering out respondents perceived to be unlikely to vote
For the 2015 election, Ipsos employed a 1 to 10 point scale to assess likelihood of voting. Anyone placing themselves at 9 or 10 was included (at equal weight) in the sample to assess partisan preferences. For the 2017 election, Ipsos are including additional variables, tracking previous UKGE vote in England. In Scotland, they ask about voter turnout in the 2014 and 2016 referendums as well as the 2015 UKGE. Anyone who has not voted in all three is also excluded (even if they reported that they were 9/10 likely to vote in 2017).
Panelbase use a 0 to 10 scale and had previous filtered out all but those respondents who selected 8, 9 or 10 on the likelihood to vote question. For 2017 they are selecting only those who select 10 on likelihood to vote.
Weighting respondents by their propensity to vote using self-reported turnout
Survation employs a 0 to 10 likelihood of voting question. Those who select 0 (‘certain not to vote’) are excluded from further analysis when assessing vote intention. All others are weighted proportionately, so that those who select 10 (certain to vote or already voted by post) on the voting likelihood question are weighted at 1, those who select 9 are weighted at .9 and so on.
YouGov do something similar but halve the weights of those who didn’t vote last time for headline vote intention.
Weighting respondents by their propensity to vote, having modelled turnout with external data
For 2015 ComRes relied on likelihood of voting on a 10 point scale. They now employ external data to generate a voter turnout model that is then used to generate weights for different sub-groups. Using aggregate data – for example from the census – they examine the relationship between age, social class, levels of deprivation and turnout to identify demographic groups that they believe are more likely to over report their intention to vote. Such groups are then weighted down examining vote intention.
ICM uses a variation of this method, modelling turnout with individual-level data using the BES (among other sources) to provide likely voter turnout figures for sub-groups on the basis of age and social class (It also excludes those not registered to vote).
The selection and/or weighting of respondents to best capture those most likely to vote is coupled with efforts to include those refusing to answer the vote intention question or those who indicated that they did not yet know who they would vote for. ICM, for example, make adjustments for previous voters who refuse to answer the vote intention question, aligning a portion of them with last party supported. When looking at samples for polls, therefore, the issue is not just quota samples vs random probability samples, or indeed the demographic variables on which respondents are weighted, but how the likely voting population is identified. Those modelling turnout are showing larger leads for Conservatives than those relying on self report measures. There have been some changes after 2015 so it remains to be seen which produce more robust figures.