,
13 tweets,
9 min read
Read on Twitter
The #RevokeArticle50 #RevokeArticle50Petition is the petition that has seen broadest support of any parliamentary petition in UK's political history if I am not wrong. It is inconceivable that among the signatories there are only individuals who voted #Remain in 2016... 
Yet, most opinion polls still suggest only a small swing to #Remain. How can we make sense of this?
A remarkable observation from online polls such as the British Election study @besresearch (using @YouGov) is that among participants in these polls, self-reported turnout in the EU referendum is huge -- 92.5% of BES panel wave participants stated that they voted in the #EURef
This is far away from the 72 percent actual turnout... hence, online polls are very vulnerable and likely affected by a massive #selectionbias, with selection on unobservables featuring large.
People with strong political beliefs/views self-select into participating in these polls. Since pollsters do not observe the nature of the selection bias, the statistical methods usually applied to reweight data to obtain an estimate of @whatukthinks are invalid.
So how do (online) #polls work? Companies like @Survation or @Yougov have a pool of participants or a community to draw on. Most polls work with a sample of 1000 or so drawn from these pools. Individuals get invited and its essentially first come first serve.
Selection bias is main concern for smaller polls, where sample size quotas are quickly met by the set of self-selected politically interested individuals. Larger cross-sectional polls such as the @Survation 's 20k strong poll are less vulnerable as they exhaust their pools
Not surprisingly, the @Survation poll we analyzed in the paper "Who is NOT voting for #Brexit anymore" is less selected. In that sample only 82% stated they voted in the #EUref. Its still massively off, but @chrishanretty 's MRP method can help a lot... goo.gl/CkUkgD
Decision makers, policy makers etc. need to be aware of the many serious limitations that opinion polling has. I would be very happy to work with polling companies to think about how we can tackle selection bias and learn about it -- there are loads of ideas and possibilities,...
all it requires a willingness to cooperate and work with researchers. @Survation works with @chrishanretty and they are doing amazing things. I would love to join forces for the sake of fostering academic knowledge and making sure polls are getting it right.
There are also broader questions. Is there a market failure somewhere? Do Opinion Polling compies that "got it wrong" suffer in the market place? Is this corrective happening? Or is opinion polling, rather than informing people & decision makers mainly a method to influence them?
These are imp. questions, especially given that most of the industry-self regulation through the @britpollingcncl is an artefact of a time, when most opinion polls were still conducted offline (rather than online). Disclosure of polls should extend to releasing micro data!
