Constituencies with more migrants are less concerned about immigration

Attention preservation notice: for a full listing of constituencies showing their attitude towards immigration, download the spreadsheet

When journalists were writing their colour pieces about Clacton last week, one commonly occurring theme was concern about immigration. Residents of Clacton seemed concerned about immigration, and their concern seemed unrelated to the actual number of immigrants living in Clacton.

Just as last week we were able to show that Clacton was one of the most Euroskeptic constituencies, this week, using the techniques that we have discussed in previous posts, we’re able to show that Clacton is also one of the constituencies where voters are more likely to say that immigration is having a negative impact on British culture. This makes sense: one of the founding principles of the European Union is freedom of movement; freedom of movement allows larger flows of migration to and from Britain (and in practice has led to larger net inflows), and so any voters who are concerned about migration may be opposed to the European Union for these reasons.

Like last week’s measure, our data is taken from Waves 1 and 2 of the 2015 British Election Study. The question respondents were asked was this:

Do you think that immigration undermines or enriches Britain’s cultural life?

Respondents could give answers on a seven-point scale, where higher numbers indicate that the respondent believes that immigration enriches Britain’s cultural life.

The top and bottom five constituencies by this measure are plotted below. Many of the constituencies which feature in this list also featured in last week’s list. Indeed, the correlation between constituency Euroskepticism and constituency concern about immigration is extremely strong, at 0.94.

top_bottom_immig

It is worth noting that many of the constituencies which are most concerned by immigration are also those with the fewer number of migrants. This holds across constituencies in England and Wales. Using data from Nomisweb (which collates information from the 2011 census, except where it is not yet available… hem, hem, Scotland) we were able to calculate the proportion of residents in each constituency who were born outside of the British Isles.

We plot the percentage of residents from outside of the British Isles against concern about immigration below. Points are coloured according to the party which won the seat in 2010. Note that this can’t (by construction) deal with the number of illegal migrants, or migrants who pass through without being resident for the purposes of the census.

immig_concern_by_immig_freq

As the graph shows, there’s a strong and statistically significant negative correlation between the proportion of migrants and concern about immigration. Constituencies like Kensington are intensely relaxed about immigration — and they have a lot of immigrants. Clacton is, in so so many ways, the polar opposite of this picture.

This is sometimes viewed as paradoxical. If — to take Clacton as our continued example — voters in Clacton have very little experience of migration, how can they be negatively affected by it in a way which would cause them to be concerned?

There isn’t really a paradox here: we can all be concerned by things which don’t personally affect us. Indeed, not only is this relationship not paradoxical, it makes sense if we think about commonly accepted theories of inter-group relations. The contact hypothesis suggests that conflict between potentially antagonistic groups is reduced the more these two groups come into contact. Contact forces people to moderate their stereotypes. Respondents start by thinking “I don’t like [group X], but [member of group X] is a sound bloke”, and then proceed to moderate their views on the group in the light of their evaluations of named individuals.

The contact hypothesis is not universally accepted (see for example, Dan Hopkin’s work) — and a lot of the `nicer’ questions concerning the nature and intensity of the contact have to be tailored to the particular application — but it does make sense of this curious pattern.

For a full listing of constituencies showing their attitude towards immigration, download the spreadsheet

Posted in Uncategorized

Free booklet on constituency opinion

Later today we’ll be presenting some of the results of our project in the House of Commons Library, which were reported on in today’s Times (paywalled).

We’ll be discussing opinion estimates on three issues (Euroscepticism, immigration, and redistribution), and hinting at some of our findings with regard to dyadic representation, or the association between constituency opinion and MP behaviour.

You can find the booklet which accompanies the presentation here. The full set of constituency opinion estimates referred to in the booklet is available here. This data is also mapped in a Times interactive.

The technical details on how opinion was estimated are in this report.

Posted in Uncategorized

Clacton is the most Euroskeptic constituency in the UK

On the 28th August, Douglas Carswell defected to the United Kingdom Independence Party and resigned his seat as Conservative MP for Clacton. At the time of writing, Carswell is the prohibitive favourite to win the upcoming by-election.

Electoral success has many fathers, failure none. Although we could give many reasons for an eventual Carswell win, we wish to identify one reason in particular. Through our research on estimating opinion in parliamentary constituencies, we believe that Clacton is the most Euroskeptic constituency in the UK. Although Douglas Carswell might have personal political reasons for defecting to UKIP, opinion in his constituency is more likely than any other to reward such a move.

Our claim that Clacton is the most Euroskeptic constituency in the UK is based on the techniques we have described in previous posts, applied to the latest data from the British Election Study. Specifically, we use information from Waves 1 and 2 of the British Election Study, which rely on fieldwork carried out as late as June of this year. (We can therefore rule out the possibility that Clacton’s Euroskepticism is a result of Carswell’s defection rather than a potential cause).

By examining responses to a question on vote intention in a future Brexit referendum, and combining this with information on respondents’ and constituencies’ demographics, we can produce estimates of the percentage of respondents in each constituency who would favour exiting the EU. The figure below shows the top and bottom five constituencies in the Great Britain.

eu-topfive

Our best estimate for Clacton is that 75% of voters would vote to leave the European Union. It’s possible that other constituencies are more Euroskeptic than Clacton: our 95% credible interval runs from 68 to 82 percent. But there is clear purple water between Clacton and the least Euroskeptic constituency, Hornsey and Wood Green.

We can make three points about these patterns.

First, the geography of Euroskepticism is similar to the geography of UKIP support identified by Rob Ford and Matt Goodwin in their book Revolt on the Right. Euroskepticism is concentrated in constituencies with a larger percentage of elderly white residents with low levels of education. These constituencies often tend to be located in coastal areas. You can see this in the following map. Darker areas are more euroskeptic.

eumap

Second, Euroskepticism is not the same as the expected probability of UKIP success. Ideological proximity to voters helps electoral success, but success also depends on organisation and the ability to convince voters that their vote will not be a wasted vote. If you are interested in UKIP’s likely electoral success, you can visit electoralforecast.co.uk, a non-ESRC funded project we are also involved in.

Third, the direction of the relationship between constituency positions and MP positions is not clear. We have shown in a conference paper that there is an association between MPs’ votes on Europe and their constituents’ positions — but the direction of causality is not clear. Is Clacton Euroskeptic because Clacton residents were convinced by their Euroskeptic MP, or did their MP become Euroskeptic because his constituents were Euroskeptic?

Certainly, this relationship is stronger for Clacton than it is for Rochester and Strood (no. 250 on our list), where Mark Reckless has defected. But Tory whips (and UKIP strategists) may wish to download our full list of results to find out who might jump next.

Posted in Uncategorized

It’s the BESt thing

There is a special category of things which merit a Wikipedia entry they don’t have. It’s impossible, or at least very difficult, to list members of this category. The very act of listing gives people a strong impetus to include these things in Wikipedia.

I mention this category of things because the British Election Studies are just such a thing. There are lots of other major academic surveys which get some Wikipedia loving: the American National Election Studies; World Values Survey, Eurobarometer — even (with due respect to any Kiwis who may be reading this) the New Zealand Attitudes and Values Study.

Sadly, there is no Wikipedia page for the British Election Studies at the time of writing. That’s a major oversight, given that the BES is (in the phrase usually bandied about), “one of the longest running election studies worldwide”.
That’s not to say there’s nothing on British psephology — there’s a page on the Nuffield Election Studies. But the hard work of putting together a repeated academic survey of mass opinion has not been recognised.

That’s a shame — because so much academic research on electoral behaviour requires dedicated mass surveys. It’s possible to imagine another way of doing things. We could give academics working on elections slightly larger research grants, and get them to buy surveys from polling companies on an ad-hoc basis. But this way of operating would, in addition to penalizing researchers early on in their careers, would make a mockery of the claim to cumulative knowledge.
There is no guarantee that individually-commissioned surveys are going to ask the same questions or use the same operationalizations of key concepts.

So having a study like the British Election Study — which has been running since 1964, and which asks similar questions from election to election — is invaluable. That’s not to say that these election studies have become reified, or otherwise cast in stone: the team conducted the 2015 study has just finished soliciting questions for add-in modules.
That development has been possible thanks to the decreasing cost of commissioning polls. The 1974 study — available at the UK Data Service after registration — had a single poll with a sample of 2,462 respondents. The 2010 study — well, it’s actually quite difficult to say how many respondents were involved in the 2010 study, other than to say, `a lot’.

For our purposes, we generally use the continuous internet panel survey (CIPS) data. You can get it here. We have a Stata file lurking on our shared drive, which features information from a little over 16,000 respondents. For some specialised questions, we’ve looked at the continuous monitoring surveys (CMS), which ran monthly from 2009 until 2012. These have questions which sometimes appear, sometimes disappear. Sadly, there seems to be no way of finding out which questions appear where, save looking through the documentation for each (monthly) survey.

All of our work would not be possible without the BES. So go read up on the project, and the details on the new project team, which is already busy preparing for 2015.

Posted in Uncategorized

South Shields is an island

In an earlier post, we talked about how we can use geographic information about constituencies to improve our estimates of public opinion in those constituencies. That is, we’re trading on the idea that constituencies that are close together are going to be quite similar to each other in many respects, and more similar than constituencies that are far apart.

We neglected to mention how we get data on constituency geographies. The answer lies in something called shapefiles — specifically, an ESRI shapefile. (ESRI is dangerously close to being to computerized map information what Hoover is/was to vacuum cleaners — a trademark used to denote a generic).

Shapefiles are… uh, quite tedious to work with. First, they don’t fit easily with the way most quantitative social science research works. Most social scientists are really happy with information which is stored in a rectangular block: each variable maps on to a column, and each case maps on to a row.

Shapefiles aren’t like that. They’re just collections of points, with some shapes in the file having very few points, and some shapes have hundreds of thousands of points. So statistical packages which work with shapefiles have to handle them using special object types — and that very quickly generates idiosyncracies.

Second, shapefiles are used for lots of purposes where detail is often required. Typically, in our analysis one of the first things we do is throw out half of the points in the shapefile — because these shapefiles are at a level of detail we don’t need. We’re not investigating planning applications for the council: most of the time, we’re just interested in working out whether two constituencies are next to one another or not. (Unfortunately, in the process of throwing out information, more gremlins creep in: in the shapefiles we’re using, we managed to make South Shields an island. We still don’t know why this happened, and we had to manual patch the adjacency matrix we produced).

Third, shapefiles are often not freely distributable. We can’t distribute the shapefiles we use in this project, because they’re only available from the Ordnance Survey (at this link: scroll down to Boundary-Line) under licence. That licence is fairly permissive — but it still means that the results we produce depend on an external file over which we don’t have control — and that limits reproducibility.

Anyway, enough grumbling: here’s the link to the code we use to find out which constituencies are adjacent to each other.

Posted in Uncategorized

All the single ladies, all the single ladies

When we discussed Mr. P before, we talked, rather blithely, about being able to build up a tally of the number of voters of particular types residing in each constituency. Those types depend on the particular model that we use, but for modelling vote choice, we might want to build a tally of all voters of (1) a particular gender, belonging to (2) a particular age-group, in (3) either rented or owned accommodation, working in (4) either the private or the public sector; who are (5) married or single; and who have a (6) particular type of education, and (7) a particular type of social grade (AB, C1, C2, or DE).

That’s quite detailed information. For eminently sensible reasons, the census authorities do not generally release such detailed information at the constituency level. You can get some cross-tabs (ideally from Nomisweb) — you can find out, say, the number of people with a particular educational attainment for each gender and age-group. And you can get oodles and oodles of univariate statistics — raw counts of people according to type of housing tenure; raw counts of people according to type of employment, that kind of thing. But any information which is sufficiently detailed for our post-stratification stage is also going to be sufficiently detailed to risk compromising the anonymity of the census.

So in order to get the detailed breakdowns by constituency that we need, we need to rely on some special census data release — the Sample of Anonymised Records. This is a sample of records from the census, with information on all of the questions that the census asks about. It’s been anonymised in clever ways — some of the data has had some random noise added to it, not enough to ruin anyone’s results, but enough to prevent jigsaw identification.

Using this sample — which, at one hundred thousand records, is a pretty huge sample by anyone’s reckoning — we can get information on the relationship between the different variables which feature in our voter types. We can see that, say, people working in the public sector tend to be slightly more likely to have a university degree than people working in the private sector. So we can create a table for the whole country, with all the information we need about the relationships between these variables.

In order to turn that information into information at the level of the constituency, we do something called `raking’, also known by its Sunday name of `iterative proportional fitting’. We start by assuming that everyone in the constituency has the same probability of landing up in any one of our types, and we gradually chip bits of, and shape this table, until it starts looking more like the relationships we know are there (i.e., the Sample of Anonymised Records). It’s a bit like taking a block of marble and chipping bits off until a beautiful sculpture comes forth.

This raking procedure manages to recover information quite well — we know that it works well at recovering information in the limited cases where the census cross-tabs are available for two or three variables. It means that we can answer quite detailed questions.

If, inspired by the Beyonce song, we wished to identify all the single ladies (all the single ladies) in a particular age-group, in a particular constituency of the UK, we could do (and in fact you’ll see that map plotted below: click to embiggen).

beyonce_plot

Or, conversely, if we wanted to identify the constituencies with the lowest number of males in employment, we could do that to — though perhaps ignoring some obvious problems.

All this is going to be very useful — indeed, essential — for our post-stratification stage.

Posted in Uncategorized

The multiple levels of Mister P

In a previous entry, we began to discuss the Mister P method for estimating constituency opinion.

Specifically, we looked at the `regression and post-stratification’ steps of the Mister P method. To re-cap, in the regression step, national survey data is used to model individual voters’ opinions as a function of their socio-demographic characteristics, yielding a predicted probability of holding a particular opinion for every voter type (where each voter type is a unique combination of socio-demographic characteristics). Then, at the second ‘post-stratification’ stage, opinion for an area is estimated by weighting the predicted probabilities for each voter type by the number of voters of that type living in the constituency.

To keep things simple, last time we left out one important aspect of Mister P that we discuss in more detail here: the use of a particular type of regression — multilevel regression — in the first step of Mister P.

As its name suggests, multilevel regression allows us to explicitly account for the fact that data is often structured at multiple levels. In our case, we can think of two levels in our national survey data. First, we can think of an `individual level’: our political opinion of interest, as well as demographic characteristics, all vary across individual respondents in our data.

Second, we can think of a `constituency level’: our respondents are clustered by the Parliamentary constituencies in which they live, and some variables vary at this constituency level — for example, constituency population density or constituency unemployment levels.

The key thing is that, because we are interested in estimating average political opinion in each constituency, it makes sense to make use of constituency level information by employing multilevel regression.

This involves a tweak to the standard regression that modelled political opinion only as a function of demographic characteristics that vary only at the individual level. To this regression model, we add a `random effect’ for each constituency. These constituency effects are modelled as draws from a common distribution and capture average
differences in opinion between survey respondents from different constituencies after accounting for
demographics.

As a result, they allow for the possibility that geography matters — that voters in some areas may be different from voters in other areas, even though they are similar in terms of age, education, or social grade.

When we incorporate multilevel regression at the first stage of Mister P, the post-stratification step remains very similar to before, except now we also add in the estimated constituency effects at this stage.

One big advantage of the multilevel regression approach is that it is very flexible. In fact, we can begin to model our constituency effects as a function of constituency level variables such as population density. This introduces further useful information about constituencies into our method, which can in turn lead
to further improvements in the accuracy of our eventual opinion estimates.

Posted in Uncategorized

Scree on the slopes

In our last entry, we discussed one way for getting better estimates of constituency opinion using characteristics of the respondents — multilevel regression and post-stratification or Mr. P.

Another way of producing better estimates (where better means `better than the lousy estimates we get from direct estimation’) is to use characteristics of the constituency. In particular, we’re going to use a basic maxim of political geography (well, geography in general) called Tobler’s Law, which states that things which are closer together are more alike than things that are further apart.

Let’s suppose that instead of surveying people about political matters, we’re carrying out some geological surveys for a mining concern — or metal-detecting on a beach, or some such activity. Suppose we want to know the volume of gold found in each square kilometre.

Now, geological soundings are expensive, requiring copious amounts of explosives or lots of spadework. So we want to get a good estimate, but we can’t dig up the entire beach, or blast the entire peninsula.

Suppose that one work team carries out some soundings in a particular area, and finds a better-than-average concentration of gold. We have a hunch that this patch would merit excavation.
But how can we be sure that the levels of gold there are really worth it?

We can firm up with hunch by `borrowing strength’ from observations in neighbouring plots. If our plot, like the button for 5 in your mobile phone’s keypad, has eight neighbours, then we can look at the responses in those eight other plots. If those plots also show higher-than-average levels of gold, we can be more confident than the true figure in the plot we started out with is also higher than average.

In the models that we use, the degree to which we can borrow strength from neighbouring constituencies to `firm up’ our estimates is going to depend on the variance parameter in a conditional autoregressive distribution — but to simplify greatly, we’re going to borrow strength from neighbouring areas only to the extent that doing it that way helps us explain the pattern of responses we get in each individual area.

The more we can explain observations in a particular area by the areas around them, and the less we need to invoke particularities of this or that area, the better this technique works.

Oh, and while other people have labelled this technique `local smoothing’, the name we give to it is (including) spatially correlated random effects, or SCRE (pronounced scree). Hopefully, that both explains the title of this post, and the rather geological example.

In future posts, we’ll look at both how much Tobler’s law holds in our particular case, and how SCRE can be combined with Mr. P.

Posted in Uncategorized

Mister P and the Twenty-Eight Hundred

Statistical methods rarely have cool names.

Tibishrani’s lasso, and ‘bootstrap’ methods, are perhaps the only exceptions.

On the face of it, multilevel regression and post-stratification, or MRP, would seem to be just another acronym.

So we’re going to follow Andrew Gelman’s usage and talk about Mister P instead.

Mister P is a way of estimating opinion in small areas that’s more accurate than directly estimating opinion from tiny samples in the way we discussed a couple of weeks ago.

The Mister P method has a number of features, so we are going to break down our description in to two posts, dealing with the ‘regression and post-stratification’ part in this post, and the ‘multilevel’ part next time.

The `regression and post-stratification’ part of Mister P draws on the two-stage `simulation’ approach used by Pool, Abelson and Popkin to estimate state-level opinion back in 1965.

The first step is to build a statistical model — a regression model — of the opinion you’re interested in, using a national sample.

The precise contours of this model will depend on what opinion you’re interested in. But for our purposes — and bearing in mind what’s coming next — a model of vote choice which uses certain demographic variables as predictors — age, gender, occupation, education, and housing and marital status — will do fine.

Models of vote choice like this can be used to make predictions for particular values of our predictors. Generally, political scientists don’t talk much about predictions — we’re more interested in hypothesis testing; and we rarely have fresh data for which we can generate genuine predictions instead of slightly artificial `retrodictions’. But all of the statistical models we use can make predictions to varying degrees of accuracy. So, based on our model, we can predict the probability of a 55 year old university-educated male working in a managerial role voting Conservative.

We need these predictions for the last step of Mister P: post-stratification. Imagine that we’ve got a really simple model, which just has gender and university education. Using census data, we could draw up, for each constituency, a tally of all the people who satisfy one of the combinations of the variables in this model.

So, we’d tally up university-educated males, university-educated females, non-university educated males, and non-university educated females. If we were thinking visually, we could even put them in a two-by-two table.

And then we could make a prediction for each box in that table. We’d take the predicted probability of each person of that description voting Conservative, and multiply it by the number of people in that box. If we do that for every box, we can predict, for each constituency, how many people vote Conservative there.

Now, our models are a bit more complicated. Instead of two-by-two boxes for two variables, each of which have two categories, we have 2x7x5x2x2x10 tables, which are pretty awkward to work with. That’s 2800 different `types’ of people in each constituency — or 1.82 million different predictions.

This method restricts us to models which can make predictions based on things that are asked about in the census. Other factors — early-years socialization into voting for a particular party, post-materialist values — all that gets swept under the carpet. But as long as there is some association between people’s demographic characteristics and their political opinions, we make some headway.

We will see in the next post how using multilevel regression in the first step can improve our small area opinion estimates by making better use of all the information in our survey data and by allowing us to include extra information on constituency characteristics.

Posted in Uncategorized

Directly estimating opinion in constituencies

Over the next few months, we’re going to be describing some of the special techniques that we use to estimate constituency opinion.

Before we do that, however, it’s useful to explain why we have to use special techniques. Why can’t we just use the same polling techniques that we use to get opinion at the national level?

The answer has to do with the accuracy of our measures, and how this changes with sample size. Almost all reputable newspapers, when they print an opinion poll, will present the margin of error.

The normal margin of error you see reported in newspapers gives us a range of values such that, if we were to carry out this exercise 20 times, the true value would be in that range 19 times out of 20. (It might seem weird to talk about the likelihood of being right in nineteen other polls that we haven’t conducted, rather than the likelihood of being right with this actual poll that we have conducted. If it seems weird to you, you might want to look into Bayesian statistics).

Smaller margins of error are better, and smaller margins of error come through sampling a larger number of people. The returns on sampling more people tend to decrease, so you get less bang for your buck sampling 500 people when you’ve already sampled 10,000, than sampling 500 when you’ve only sampled 1,000. Specifically, the margin of error, in the worst-case scenario where opinion is divided 50:50, can be calculated the following way:

  • divide one by four times your sample size
  • take the square root of this number
  • multiply it by a special constant — we’re going to be using 1.96, which gives us a margin of error where we’re only wrong one time out of twenty

So, for a sample size of 1,000, that means 1 / 4000 = 0.00025, the square root of which is 0.01581, which when multiplied by 1.96 gives us 0.03099, or almost 3%.

What does that mean for estimating opinion in constituencies? It means we either need very big samples in each constituency, which is very expensive, or we need very big margins of error.

Let’s assume that we’re working with a really large sample of 40,000 people, which usually only happens when polling companies aggregate polls conducted over several months. Let’s also assume (and this is an inaccurate assumptions, but it’s the best case scenario for us) that our sample is equally divided between all 650 constituencies in Great Britain and Northern Ireland.

If we’re also happy to assume (and this is another inaccurate assumption) that a nationally representative sample is composed of locally-representative sub-samples, then the sample size in each of our constituencies is 40,000 / 650, or 62 people.

The margin of error on that kind of sample is huge: 1 / 4*62 = 0.004, the square root of which is 0.064, and 1.96 times this number is 12.4%. That’s big. It means that we have to say, `our estimate is 50%, but the true value could plausibly be between 38 and 62%’.

And this is starting with huge, huge national samples. If we shrink our national sample to 10,000, then the situation gets dramatically worse.

Unless we’re happy providing very imprecise estimates, or have lots of money to commission polls with huge samples, we need something better. As you might have guessed, we don’t have lots of money, and we’re not happy with imprecision. So over the next weeks, we’ll be describing some techniques for getting more precise constituency estimates with moderately-sized samples.

Posted in Uncategorized