Brexit in Wales – The Parties

It has been widely reported that the vote to leave the EU was strong in Labour’s heartlands – a worrying split between the party leadership and their voters. By combining the lion’s share of the conservative vote with the non-metropolitan wing of the Labour party’s vote, the Leave campaign caused a political earthquake. As the media generally lacks this  blog’s unabashed cymrucentric bias, however,I’d like to concentrate on how these dynamics played out in Wales.


Awkwardly though, the referendum’s votes were totaled at the Council Authority level, rather than the constituency level. This makes comparisons between votes in the referendum and Assembly/Westminster elections tricky.  I’m sure there are clever ways to estimate the Brexit votes within each constituency, but here I’ve taken a more straightforward approach. I’ve looked at the correspondence between  Leave vote share and the proportion of councilors the different parties have in each Council.

There were a couple of issues with this approach. Firstly the last council elections were in 2012 – so they’re arguably a bit out of date. That’s why I’m not reporting any numbers for UKIP – the last council election pre-dates UKIP’s surge in Wales. Secondly, not being an expert in local government elections, I was quite surprised by how many independent councilors there are. Furthermore there  are some small parties that only exist in a single council – Llais Gwynedd for example. The number of councilors also varies quite a lot across councils. For these reasons I’ve decided to base my main analysis on the proportion of councilors each party has of the councilors belonging to the major parties – so I’m dropping independents and minor parties from the denominator of the proportions. There was no correlation between the number of such councilors and Leave’s share of those vote (.05), so I feel this is a reasonable thing to do.


Onto the results! Below is the proportion of Labour councilors plotted against Leave’s share of the vote. There’s a striking positive relationship – the Spearman’s correlation is .56. Thus the Welsh Brexit vote was unambiguously stronger in areas that elect Labour councilors.


On to Plaid Cymru. This time the relationship was modestly negative – -.26. Councils with lots of Plaid councilors tended to go for Remain. I would note here that two of Plaid’s strongest councils have substantial minor party/independent representation – the aforementioned Llais Gwynedd in Gwynedd and an enormous number of independents on Ynys Mon. Thus the reported proportions exclude a lot of councilors. Putting them back in doesn’t make a big difference though, the correlation becomes -.31.


The Lib Dems are perhaps the most obviously pro-EU party and this is borne out by this analysis. The correlation is -.43, a substantial negative association.


The surprise package, for me at least, was the Conservative party. One might expect that areas returning lots of Conservative councilors would be firmly for out, given the party’s long history  of euroscepticism. Not so. There was actually a small negative relationship between Conservative councilors and Leave vote. The areas with lots of Tory councilors include prosperous Monmouthshire and the Vale of Glamorgan, so my sense is that this is about economic advantage, but I was surprised nonetheless.


To conclude, the Leave vote really does seem to have come from Labour voting areas – there is no other way to slice this. Labour’s electorate appear to be at odds with the party on the defining political issue of the day.


P.S. I’m very grateful to everyone who commented on, shared, and read my previous post Brexit a’r Iaith Gymraeg. I’m somewhat stunned by the response it got. Diolch o galon pawb!


Brexit a’r Iaith Gymraeg


Along with a great deal of shock and dismay, one thing that struck me about the results of the EU referendum was how well the Remain vote held up in Welsh speaking Wales. In England, the remain vote was concentrated in big cities and prosperous regions – Remain voting Gwynedd and Ceredigion don’t really fit that pattern. We all saw those graphics showing demographic factors that predicted voting in the referendum, can we add Welsh-speaking to the list?

Firstly, I got the Leave vote share for each Authority region and census figures on percentage of Welsh speakers and ran a correlation.The two variables were highly negatively correlated: -.40 (Spearman’s correlation). So yes, at first glance the data seem to bear this out.

It seems a bit premature to call it for Yr Iaith Gymraeg yet though, economic disadvantage appears to be a predictor of voting leave too and while Y Fro Gymraeg is not a prosperous area, there are probably fewer pockets of really marked poverty than other areas of Wales. The figure below shows percentage of Welsh speakers on the X-axis and  Leave vote share on the Y-axis. The redder the name of a region, the more areas it has in the most deprived 10% of areas according to the Welsh Index of Multiple Deprivation.


As you can see, the regions with high numbers of Welsh speakers tend towards lower Leave votes, but also tend not to have lots of areas in the most deprived 10%  of regions.

To try and get at this in a more statistical way,  I fitted a regression model to the data predicting Leave vote share using number of areas in the lowest 10% WIMD and saw whether adding in the percentage of Welsh-speakers as a second predictor improved the model – it did. When I did the opposite, seeing if a model that used percentage of Welsh-speakers could be improved by adding in the deprivation data, this didn’t improve the model.

Maybe using the most deprived 10% of area is the wrong measure – it could be that pockets of severe deprivation are less important for explaining voting behaviour than a more diffuse pattern of poverty. I reran the analysis, this time using the proportion of areas that fell into the bottom 50% of areas in the WIMD. This was indeed a better predictor of Leave vote share – .49 as opposed to .35. When I reran the regression models, the model that included both the percentage of Welsh speakers and percentage of areas in the most deprived 50% performed best. Both deprivation and percentage of Welsh speakers are helpful in explaining vote share. Here’s a graph of how vote share (Y-axis) was related to the proportion of areas in the bottom 50%  according to the WIMD (X-axis), this time with the percentage of Welsh speakers reflected by how green the name is. Note that many of the more Welsh-speaking regions are below the line of best fit, suggesting lower levels of support for leave than you would expect based on deprivation data alone.


There remain a few plausible alternative explanations – both Gwynedd and Ceredigion have university towns (although so do Wrexham and Swansea who voted out) and West Wales gets lots of EU funding (although so do the Valleys who voted out). But yes, to conclude, it seems plausible that having a high number of Welsh speakers was associated with lower levels of Leave support, and this wasn’t purely to do with the underlying pattern of poverty in Wales.


P.S. Sorry for the long wait between posts by the way, we had a baby recently!


Enthusiasm gaps


Turnout in May’s Assembly elections was 45.3%. Higher than in 2011 but considerably lower than UK general elections (66.4% in 2015, since you asked). Does this lower level of enthusiasm hit all parties equally though?

Below I’ve made boxplots of the difference between the number of votes each party  got in the Assembly election and the Westminster election in each constituency. Negative scores are lower votes in the Assembly and positive are higher.For those of you unfamiliar with boxplots (a not unusual state of affairs):

  • The black line in the middle is the median – the datapoint exactly in the middle.
  • The coloured bit is the interquartile range – the gap between the 25th percentile and the 75th percentile. So the middle 50% of observations fall here.
  • The ‘whiskers’ go out as far as the outlier 1.5 times the interquartile range.
  • Other outliers appear as the little dots.

Anyway, here it is:


Everyone except Plaid Cymru generally does worse in the Assembly elections – indeed Labour, the Conservatives, and UKIP don’t do as well as well in the Assembly election (in terms of number of votes)  as they do in Westminster in any of the 40 constituencies. The Liberal Democrats generally do a bit worse, but do better in a few seats. This could be related to their pronounced concentration of votes in just four constituencies – locally competitive parties draw tactical voters.

The interesting exception is Plaid Cymru, who do better in the Assembly in 26/40 seats. This could be for a number of reasons:

  1. The Westminster-based parties (yes, you too UKIP) have other fish to fry and this might translate into less organisational focus on Assembly elections.
  2. People who vote for the Westminster parties could have less enthusiasm for the assembly elections than Plaid voters.
  3. Some people might vote for Plaid in the assembly but for another party in the general election.

Plaid’s advantage here, it is important to state, is not just a relative one. They don’t just suffer less than other parties from the lower turnout, they actually outperform their Westminster results. This means that Option 1 above cannot be the whole story.

My last few posts on the predictive power of Westminster results on  Assembly results suggest that it might be possible to build a fairly good predictive model for Assembly elections using Westminster results. The parties’ relative abilities to turn out their vote in the Assembly elections is an important parameter to include in such a model.

Next on Cyfri am Byth: a first draft of a model…


What can Westminster results tell us about Assembly results? 4. UKIP and the Lib Dems

I’ve decided to combine my posts for the last two parties in the Assembly, because the story is pretty straightforward. Here’s UKIP’s plot of vote share in the two elections.


With the exception of Arfon and Aberconwy, where they didn’t stand, UKIP’s performance in the Assembly can be almost perfectly predicted by their Westminster performance. The correlation between the two is .94, excluding Arfon and Aberconwy and .88 including them. The line of best fit on the graph excludes these two constituencies by the way. I’ve not included the plot for change in UKIP vote as UKIP didn’t contest any Assembly constituencies last time and only contested a few Welsh Westminster constituencies back in 2010, so the plot isn’t good for much except playing spot the difference with the above plot.

As for the Lib Dems…


It feels like every plot I make of Lib Dem performance just hammers home how concentrated their vote is in these four constituencies! The correlation is .93, but this is not a reliable number given how weirdly distributed their votes are. Mark Williams, the MP for Ceredigion and new leader of the party in Wales, said that he hoped to improve Lib Dem prospects outside of these four seats. I suppose it’s hard for them to do much worse elsewhere, but that seems like a very long term strategy.


As for change in Lib Dem vote, again this tells us what we might have expected – Kirsty Williams did extremely well in Brecon. The presence of Montgomeryshire above the line, suggesting poorer performance here in the Assembly is perhaps an ominous sign for their recovery.

Interestingly, UKIP actually did slightly better in the Westminster election – 13.8% compared to  12.8% – emphasising that is the list system which gives UKIP its returns in the Assembly, not greater support. The Lib Dems did slightly better in 2016 than 2015 – 7.5% as opposed to 6.5% in vote share and a loss of only 2.9 percentage points rather than the whopping 13.3 in 2015. This could reflect some softening in public anger at the Lib Dems for going into coalition in 2015, the distinctiveness of the Welsh Lib Dems from the national party, or (most likely in my opinion) the low base the Lib Dems were starting from in the Assembly.


What can Westminster results tell us about Assembly results? 3. The Conservatives

Conservative vote share in the Welsh Assembly can be almost perfectly predicted based on Westminster vote share. The correlation is a mighty .97 with the closest thing to an outlier being Kirsty Williams’ heroics in Brecon.


As for change in vote share, no key Conservative seats appear to be outliers from a fairly healthy correlation of .55. Pontypridd stands out because of an enormous improvement in Conservative share of the vote in 2015, which was from a deposit-losingly low base.  Cardiff Central stands out as a relatively good result in the Assembly, perhaps because their vote was squeezed in 2015 – the seat was a high profile Labour-Lib Dem battleground.


The numbers bear out the perception that the Conservatives had a bad Assembly election – they got a mean of 20.2% of the vote, as opposed to 26.2% in 2015. Their direction of travel was also worse – a mean reduction of 3.8 percentage points as opposed to a gain of 1.3%. So the distribution of conservative votes across the country is extremely predictable based on Westminster results, but the overall level was a disappointment for them.



What can Westminster results tell us about Assembly results? 2. Plaid Cymru

How well did Plaid Cymru’s performance in 2015 predict the party’s performance in 2016? Like Labour, there is a very strong relationship between their share of the vote in the two elections. The correlation between the two, for the statistically minded reader, is .91. Again, constituencies above the line of best fit are where performance was relatively better in 2015 and constituencies below the line are where performance was relatively better in 2016PC2015

Interestingly Plaid’s three Westminster seats (Arfon, Dwyfor Meirionydd, and Camarthen East and Dinefwr) are all above the line; perhaps this is explained by regression towards the mean. Seats where Plaid are strikingly stronger in the assembly include Gwent, Aberconwy, and Cardiff West.It’s worth mentioning that UKIP didn’t run in Arfon or Aberconwy. This, as I’ve discussed previously, would have likely had a greater impact on Labour’s vote share than Plaid’s, which would be an issue in Arfon where Labour are Plaid’s rival. In Aberconwy, held by the Conservatives, the evidence is less clear cut that UKIP’s absence would have benefited Plaid – Conservative vote share was higher where UKIP’s was too.

In terms of change in vote share, there is a stronger relationship (the correlation is .44) than seen with Labour, but it’s certainly much weaker than for overall vote share. Interestingly Plaid’s disappointing Llanelli result was arguably better than might be expected based on 2015, while Leanne Wood’s triumphant win in The Rhonda might have been expected to be even bigger – unusual verdicts certainly. Clwyd South, not a target seat for Plaid, stands out as somewhere where their direction of travel was better in 2015, as does Arfon.


Finally, as can be inferred by the difference in the number of constituencies Plaid hold in the Assembly (6) compared to Westminster (3), Plaid are much more successful in the assembly than they are at Westminster. They get a mean of 21% of the vote in assembly constituencies as opposed to 12.9% in the Westminster equivalents. This isn’t due to a sudden surge in Plaid support though, their mean gain in vote share was 1.4% in 2016, compared to 1.2%.

Next time, the Conservatives….


What can Westminster results tell us about Assembly results? 1. Labour


Under the Fixed Term Parliaments Act, it’s possible that Westminster elections will regularly precede elections to the assembly by a year. Given how little polling is conducted for Assembly elections (with the honourable exception of the collaboration between Prof. Scully at Cardiff University,  ITV Wales, and YouGov), this could be a really helpful indicator of the likely outcome of the Assembly elections. But how well did results in 2015 predict assembly results this May? I’ll start with Labour and work my way through the other parties in subsequent posts.

I plotted Westminster vote share (Y-axis) against Assembly vote (X-axis). The colour of text reflects the winner of the assembly seat. There’s clearly a strong relationship here, indeed the correlation is .897, which is very high by the standards of the social sciences. Constituencies under the line of best fit are where Labour did better in the assembly than you might imagine based on Westminster results, those above the line are where they did worse. Gwent and the Rhondda stand out as you’d imagine, but most constituencies are very close to the line, as you’d imagine with such a high correlation.


This is probably unsurprising. You would imagine that places inclined to vote for Labour in Westminster would be fairly likely to vote for them in the assembly too. What about change in vote share? Do shifts from 2010-2015 also predict shifts from 2011-2016? I plotted change in vote share in the two elections against each other in the same way


Here the relationship is much weaker, with a modest correlation of .286. Interestingly the outliers are somewhat different. Gwent was still a surprisingly big loss in Labour vote, but so was Swansea West, where Labour were very much on the front foot in 2015. The defeat in the Rhondda, interestingly was actually quite predictable given the result in 2015 and, based on its position below the line of best fit, could have been much worse for Labour.

Finally, correlations ignore the means of each set of votes. It’s worth pointing out that Labour did much worse in the Assembly elections than in Westminster. Their average vote share fell from 40% to 38.6%. More strikingly, Labour’s mean vote change was a gain of .6% in 2015 but a mean loss of 7.8% in 2016.

Next on Cyfri am Byth, Plaid Cymru…