In short, it shows the Conservatives losing both seats. The result in Mid Bedfordshire is considerably closer than the one in Uxbridge and South Ruislip, due to the large majority that Dorries had compared to Johnson.
In reality, the Liberal Democrats may be the ones to benefit in Mid Bedfordshire. I think what the model shows is that even a huge 24,000+ majority is vulnerable given the current state of play, but the seat ‘feels’ more like a Liberal Democrat target than a Labour one, despite the model predicting a narrow Labour win.
Labour have had high-profile campaigns in Uxbridge and South Ruislip before, reducing Johnson’s majority to just 5,000 in 2017, and therefore it makes sense that they’d look to capitalise here. The result above, showing a 19pt Labour lead in the seat, implies a 92% chance of Labour winning it. However, there are caveats.
What are the caveats?
There are two main ones. First, the obvious one, is the fieldwork is out of date. This MRP model showed headline voting intention that gave a Labour lead of 20 points; the polls have narrowed significantly since then. We will be running a new MRP model soon, but unfortunately these resignations beat us to it.
The second main one is that the question asked in the poll is how respondents would vote at a General Election were it tomorrow, and not how they’d vote at a by-election. By-elections behave differently, and sometimes strangely, off much lower turnouts, with strong local campaigns that parties can’t afford to replicate in every seat at a General Election. MRP would also fail to assess any incumbency factor, tactical voting, or other local factors, such as a local hospital closure, or planning issues. Our model from December has been fairly close to the actual vote share in recent by-elections, e.g. Stretford and Urmston, West Lancashire, but this may be more luck than judgement.
What is MRP?
Multilevel regression and post-stratification (MRP) is a way of producing estimates of opinion and attitudes for small defined geographic areas. It works by combining information from large national samples with ONS and census data.
The responses given by respondents are modelled on the basis of their demographic characteristics and what we know about their area (its past voting history, how it voted in the EU referendum, and so on). This is the “multilevel regression” part, which examines to what extent a person’s lifestyle, background and life experience affects their electoral behaviour.
In the subsequent “post-stratification” stage, we use census data to calculate how many people of each demographic type live in each area and combine this with additional relevant contextual information to predict how many of these people will vote for each party (or have a certain opinion).
In this way, the estimates, although they are derived from a national sample, end up being representative of the demographic make-up of each constituency.