Delayed shock? How Brexit conditioned campaign effects in British general elections

The impact of campaigns on electoral performance is conditioned in part by contextual effects. The popularity equilibrium model has proven to be an important guide to how the electoral effects of local campaigns vary by a party’s existing level of popularity. Such an equilibrium can, however be disturbed by an electoral shock—a rare event which fundamentally challenges the foundations that underpin predictability in elections. This article analyses the impact of the electoral shock of the UK’s 2016 referendum on EU membership on campaign effects in the subsequent elections of 2017 and 2019. Using a novel theoretical and methodological approach, it shows that while there were observable effects of the Leave vote already present before the referendum, the geography of the effectiveness of Labour and Conservative local campaigns was altered after the referendum. However, it was not until the 2019 election that the shock of the 2016 Leave vote became a particularly important predictor of the electoral efficacy of both parties’ campaigns.


Introduction
There is strong evidence that the electoral consequences of campaigns are conditioned by context (Arceneaux & Nickerson, 2009;Fieldhouse et al, 2014;Fieldhouse et al, 2020a, Hillygus. 2005Niven, 2001). A broad literature across many different countries shows that in general, campaigns deliver electoral payoffs (Andre & Depauw, 2016;Carty & Eagles, 1999;Fisher et al, 2011Gschwend & Zittel, 2015;Sudulich et al, 2013). However, their effects vary depending on electoral context. In a period of relative electoral stability (that is in the absence of electoral shocks) this creates an equilibrium such that the electoral effects of campaigns can be predicted by a party's existing level of popularity -campaigns will be most effective in electoral terms where parties are not unusually unpopular or popular , Fieldhouse et al, 2020a. Fisher, Cutts and Fieldhouse (2011, 2020a refer to this as popularity equilibrium -a relatively stable state of affairs whereby the effectiveness of campaigns shifts in response to existing levels of popularity, but where equilibrium is maintained between that level of popularity and the conditions under which campaigns are more or less electorally effective. But what happens when there is an electoral shock? Popularity equilibrium can adjust to significant changes in parties' electoral fortunes, such as after the British general election of 1997, and indeed through periods of electoral volatility (Fieldhouse et al, 2020a). Electoral volatility in the UK has been driven by both longer-term changes, including a decline in party identification and party system fragmentation (the growth of new, smaller parties) together with short-term electoral shocks (Fieldhouse et al, 2020b). In this article we argue that these electoral shocks may bring about changes in the importance of key contextual variables which influence the effectiveness of campaigns. More specifically we test in the aftermath of such a shock, predicting the likely electoral effectiveness of campaigns is enhanced by accounting for the impact of the shock as well as previous and existing levels of popularity. We test this

Theory
The underlying principle in this paper is that the electoral effectiveness of local campaign efforts will not be entirely within the control of the political party undertaking the campaign.
Effectiveness may also be conditioned by electoral context. Previous work highlights the importance of prior electoral performance and the principle of popularity equilibrium, where the maximum electoral effectiveness of a party's campaigns is likely to occur where a party is neither very unpopular nor very popular Fieldhouse et al, 2020a). The maximum effectiveness of a party's campaigns is driven by popularity. While district (micro) level popularity is correlated with national (macro) levels of popularity, it is also the case that parties' levels of popularity display significant geographical variation. Popularity equilibrium captures the fact that parties' electoral performance across different constituencies is not randomly distributed but is conditioned by that existing level of popularity. Fieldhouse et al, 2020a). Popularity therefore refers to the existing level of support in the district or constituency -measured typically by the level of electoral performance at the previous election. Fieldhouse et al (2020a) show that over time, prior electoral popularity is a strong predictor of campaign effectiveness. Using the principle of popularity equilibrium (for which they devise a model), they show that over the six British general elections between 1992 and 2015, the effectiveness of the party election campaigns is conditioned by the level of popularity (as measured by prior electoral performance). Critically, maximum effectiveness occurs where parties are neither too unpopular nor too popular.
Electoral shocks can, however, fundamentally change electoral dynamics. The principle of electoral shocks is informed by the idea of shocks in public policy and economics. In both, relative stability is the norm. However, when there are shocks such as economic crises, or when rules no longer confer legitimacy (Clift & Fisher, 2004), significant changes in outcomes follow, sometimes leading to permanent and dramatic shifts in policy. In the context of electoral politics, shocks are disturbances which represent a sharp change in the status quo outside the normal course of politics; are highly salient and noticeable over prolonged time periods, and are relevant to party politics (Fieldhouse et al (2020b, 32). Shocks can produce a significant increase in vote-switching because of the strategic response of political parties, (Fieldhouse et al, 2020b:31) Critically, they do not imply a quick return to a former equilibrium. Examples in the British electoral context include the global financial crisis of 2007-8 which significantly affected Labour's reputation for economic competence, the Scottish Independence Referendum of 2014 which resulted in party choice being framed by the constitutional question (Johns and Mitchell, 2016;Fieldhouse and Prosser, 2018), and the referendum on the UK's continuing membership of the European Union in 2016, where party identities subsequently became increasingly tied to a Leave or Remain position.
Electoral shocks, mean that issues most closely associated with the shock become more salient.
Not only do voters become more informed about such an issue, they are also more cognisant of the parties' positions on the matter (Fieldhouse et al, 2020b: 40-1). As a consequence, the issue become more significant in voters' decision-making. Equally, the electoral shock affects parties' image. An electoral shock can re-shape perceptions of parties' positions, priorities and standpoints, and relatedly, their social imagery (which social groups the party is seen to represent). This is likely to affect the efficacy of parties' campaigns as the changes generated by the shock are likely to have a geographically uneven impact leading to a change in the geography of support. If there is an electoral shock, therefore, the popularity equilibrium may be fundamentally disturbed such that prior local electoral performance may be significantly reduced in its capacity to condition the electoral effects of local campaigns.
There are two reasons this may be so. First, popularity still matters, but critically, an electoral shock can alter where a party is now more or less popular. A substantial change in the social basis of party support and a resulting shift in the geography of support will render previous vote share a less good predictor of current popularity. Because campaign effectiveness is normally estimated as the effect of campaign effort after controlling for previous vote or electorate share (Fisher et al, 2011;Fieldhouse et al, 2020a), if the latter becomes a poor predictor of the former then any estimate of campaign effectiveness will become less precise.
The second reason that an electoral shock may affect where local campaigns are most effective is that electoral shocks can abruptly and significantly increase the salience of an associated issue and trigger a strategic response from parties (Fieldhouse et al 2020b). As a consequence, local campaigns have the potential to expose the parties' positions on these issues to the local electorate. Insofar as local voters are receptive to the party's message around the newly salient issue, the local campaign may attract or repulse voters accordingly. Places where a party were previously strong may suddenly become less fertile ground for mobilising supporters if traditional supporters are not sympathetic to the party on the newly salient issue. The geographical distribution of attitudes towards the newly salient issue, therefore, will condition the effectiveness of local campaigns over and above that which is predicted by the popularity equilibrium model.
The vote in the 2016 referendum on the UK's continuing membership of the European Union is an example of such an electoral shock (Fieldhouse et al. 2020b). The result of the referendum, whereby UK citizens voted to leave the EU, represented a fundamental challenge to, and ultimately a shift in, British policy built up over more than 40 years, and arguably since Britain first sought to join the European Economic Community (as was) in the early 1960s. As a consequence of the referendum vote, the process of withdrawal became the principal preoccupation of British politics following the referendum -particularly in the period from 2018-2019 with the various Parliamentary wrangles over the withdrawal agreement (Thompson, 2020  . As Evans and Menon (2017:90) note, the referendum represented a fundamental challenge to party affiliations, with the underlying values behind the choice of referendum vote cutting across party loyalties. Moreover, party cues in the Referendum vote were generally weak -in part because of the mixed or weak messages emanating from both parties (Clarke et al, 2017: 451-454). If cues were weak, there is little surprise that Leave or Remain voting was not strongly tied to existing partisan dispositions.
The response of the parties to the referendum result differed, however. The Conservatives opted for a clear pro-Brexit stance, albeit one where the type of Brexit was not yet defined (Evans & Menon, 2017:93). By 2019, the message was even clearer -Get Brexit Done -following the lengthy difficulties in Parliament over the withdrawal agreement (Thompson, 2020). Labour, however, sought to follow a more ambiguous stance on Brexit -seeking both to appeal to those who wished to Remain, and its own voters who had favoured Leave. These responses were reflected in the parties' manifestos. In the 2017 Conservative manifesto, Brexit was listed as the second of five 'giant challenges' with the word appearing 17 times. By 2019, Brexit was the first section of the manifesto and the word was mentioned 61 times 1 . By way of comparison, 10 pages were devoted to Labour's 124-page 2017 manifesto, with the word appearing 24 times. In 2019, just five pages (of 105) were devoted directly to Brexit, the word appearing 21 times.
The Conservatives' post-referendum position had an important effect on voter perceptions. Fieldhouse et al (2020b: 169) show that while Leave voters saw the party as being very pro-EU integration in 2015, by 2017 they saw the party as being very strongly against. By way of contrast, Labour became, in effect, the default choice for Remainers and those seeking a softer Brexit, in part due to the fact that the Liberal Democrats had not yet recovered from the collapse in electoral viability in 2015, and because Green voters began moving over to Labour after Jeremy Corbyn became Labour leader (Fieldhouse et al, 2020b: 176-86). In effect, the shock of the EU referendum result significantly 'increased the link between attitudes towards the EU and Conservative versus Labour voting' (Fieldhouse et al, 2020b: 184), with record levels of electoral volatility in the 2017 election in terms of voters switching between the Conservative and Labour parties relative to their vote in 2015 (Fieldhouse et al, 2020b: 23-4). Indeed, while partisan affiliation might be expected to condition the effects of electoral shocks, the EU 1 Indeed, the 2019 Conservative manifesto was comfortably the party's most Eurosceptic of the previous 75 years. This is measured using the Comparative Manifesto Project data and calculating a net score of positive and negative indicators in respect of the European Union (Per108 -Per110). See Volkens et al 2020 referendum vote had a stronger impact of the subsequent vote choice of stronger partisans of major parties than amongst weaker partisans or non-identifiers (Fieldhouse et al, 2020b: 45).
As is clear, in keeping with Fieldhouse et al's (2020b) definition of shocks the referendum result represented a sharp change in the status quo outside the normal course of politics; it was highly salient and was relevant to party politics with the potential to reshape perceptions on what parties stood for and who they represented.
The issue also became highly salient for the public. At the time of the 2015 election, the issue of Europe did not even make the top 10 most important issues facing Britain. Just two years later in 2017, Brexit was the second most important issue (behind the NHS), while in 2019, it was seen as the most important issue (Ipsos-MORI). And, public opinion continued to show deep splits on the issue (Evans & Menon, 2017: 115). If the referendum result represented an electoral shock for voters, we might expect, therefore, that it may change where parties' campaigns are most effective. More specifically, an electoral shock could significantly affect the impact of prior levels of popularity on conditioning campaign effectiveness and implies that as the impact of electoral shock played out, Conservative campaigns should have become increasingly effective in pro-Leave areas whilst Labour campaigns should have become more effective in pro-Remain areas.
We can suggest, therefore, that, following an electoral shock, the popularity equilibrium model becomes a less powerful predictor of campaign effectiveness which will instead vary according to the local distribution of opinion on the salient issue. In other words, depending on how attitudes towards the newly salient issue are correlated geographically with previous levels of support, a shock may significantly weaken the relationship between campaign effectiveness and previous electorate share. Moreover, the newly salient issue may become an important factor conditioning where campaigns are most effective. We seek the test the empirical basis for this in the paper by asking the following question: In summary, we should; expect that the geographical distribution of Leave and Remain voters should condition the effectiveness of parties' campaigns after the referendum such that Conservative campaigns are more effective in constituencies with a higher Leave vote, and Labour campaigns in constituencies with a higher Remain vote.
Our hypotheses are therefore as follows:

H1.
Including the geography of the Leave vote will improve the explanatory power of the models assessing the conditioning of campaign effectiveness in 2017 and 2019, but not in 2015.

H2.
There was no relationship between the geographic distribution of Leave voting and the The absolute levels of the indexes can be compared directly across parties within the same year, but not over time as the indexes of campaign intensity reflects campaign developments such that additional variables are entered into the scale in each election to ensure we are capturing as accurate picture of the range of campaign activity as possible.
We can observe the impact of prior electoral performance as represented by the popularity equilibrium model on campaign effectiveness for most parties at each of the elections from 1992 to 2015 (Fieldhouse et al, 2020a; Reflecting the use of constituencies as the unit of analysis, where there may be multiple responses from different parties we use clustered standard errors. Overall, we are seeking to assess whether the incorporation of the level of the Leave vote into the models improves fit and then assess whether the level of the Leave vote conditioned where parties' campaigns were most effective. It is commonplace to use various R 2 statistics and likelihood ratio tests to assess the fit of a model. Yet given that adding terms to a model will always improve fit, it can be difficult to reject more complex models with larger samples as it becomes far more difficult to determine 'actual' from more trivial improvement. As a consequence, there is a tendency to 'overfit' and for models to be less parsimonious than they should be. In seeking to avoid the 'overfitting' problem some researchers will eliminate parameters in a systematic manner thereby exposing the risk that this more simplified model will not explain the data. To offset these problems of model selection, quantifiable measures of the expected value of information are used to compare the relative plausibility of different models and as such determine which best approximates the data. Here we use two common information measures -the AIC (Akaike's Information Criterion) and BIC (Bayesian Information Criterion) -to simultaneously assess overall fit and compare different non-nested or in our case nested models (Long and Freese, 2014;Raftery, 1995;Yang, 2005). Both the AIC and BIC are penalized likelihood criteria as they involve the calculation of a maximum loglikelihood and a penalty term. So as more parameters are added, the fit improves but any extra contribution to model fit must counterbalance the penalty that comes from their inclusion. In other words, there are penalties for those additional variables that do not significantly improve fit. Using the maximum likelihood estimate and the number of parameters, the AIC estimates the information lost in the model. As such it is a useful measure to avert 'overfitting'. The BIC uses a Bayesian set-up to assess overall fit and compare models and differs from the AIC in that it imposes a larger penalty for a higher number of parameters (Schwarz, 1978). The BIC is always higher than the AIC but a lower BIC specifies lower penalty terms while a smaller AIC indicates less information lost. For both measures, the best model is the one that provides the minimum or more negative AIC and BIC value (Burnham and Anderson, 2002;Kass and 5 As a check, we also ran the models including estimates of likelihood of winning the constituency for each party as captured by the British Election Study. This was to test if we observed significance variation in the measure of popularity between Share of the Electorate in the previous election and the estimates of likelihood of winning recorded just before dissolution. The inclusion of the estimates had no statistically significant effect. Raftery, 1995). Here we use both information measures and the magnitude change between models to aid model selection. 6 We test the models for the Conservatives and Labour for each election in the series (2015,2017 and 2019). Of course, the referendum occurred in 2016, after the 2015 election. However, we run the model for 2015 to test whether the referendum marked a shift in the relationship between the constituency Leave vote and campaign effectiveness. In effect, as the referendum occurred after the 2015 election, this provides a placebo test for the Brexit electoral shock. Our expectation is that any effects of the constituency Leave vote will be weak or non-existent in 2015, but will become more evident in the subsequent elections.

The Leave Vote and Model Fit
Tables 1-3 shows the estimates and model fit statistics for each of the models described above The magnitude change in BIC and AIC to assess model selection where model 2 is assessed to determine whether it is a better model than model 1: a) less than 2 -no support for improvement; b) between 2-6 the evidence is more positive; c) between 6-10 the evidence is strong; d) greater than 10 the evidence is very strong. For more details, see Burnham and Anderson (2002) for the BIC measure and Kass and Raftery (1995) for the AIC. Both Yang (2005), Hauser (1995) and Burnham and Anderson (2002) provide deeper discussion about the relative merits of both measures and whether one or the other is preferable. The evidence in 2015 for the difference in model fit between Models 2 and 3 is mixed. The AIC suggests that Model 3 delivers a marginally better fit than Model 2, while the BIC suggests the reverse.

Average Marginal Effects of the Campaigns by the Leave Vote
If our hypotheses are correct we would expect to see a change in the relationship between the efficacy of the campaign and local Brexit preferences after the referendum. We therefore examine the relationship between the Leave vote and the average marginal effect of each party's campaigns using results from our preferred combined model (Model 3). 8 Our expectation was that there would be little or no effect in the Leave vote on the efficacy of parties' campaigns in 2015, prior to the referendum (H2), but that we will observe more tangible effects in 2017 and 2019 such that Labour campaigns should be more effective in Remain-leaning constituencies and Conservative campaigns more effective in Leave-leaning constituencies (H3 and H4).  However, the average marginal effects of the two parties' campaigns are not significantly different from each other at any value of the Leave vote, and overall, it confirms that prior to the referendum, campaign effectiveness was only very weakly related to the Leave vote. There was no relationship in respect of Labour, while in the case of the Conservatives, the very small effect observed was such that the parties' campaigns were slightly more effective in more Remain leaning areas. H2 is therefore confirmed in respect of Labour, and partially in respect of the Conservatives. Thus, despite the fact that Table 1 shows an improvement in overall model fit relative to the popularity equilibrium model by the incorporation of the Leave vote, we do 8 As noted above for 2017 AIC and BIC suggest there is little between Models 2 and 3. For comparability with 2017 and 2019 (where Model 3 delivers the best fit using both measures) we use Model 3. not observe any notable differentiation between the parties in terms of the impact of the Leave vote on campaign effectiveness consistent with an electoral shock. Conservative campaigns were slightly more effective in Remain-leaning constituencies, while the efficacy of Labour's campaigns was not related to the size of the Leave vote. campaigns. By way of contrast, Labour's campaigns were again less successful in seats with higher Leave votes and more effective in Remain areas, with the difference between the two parties' campaigns being statistically significant. However, as Table 5 illustrates, the impact of the Leave Vote on Labour's campaigns was not substantially different from the previous two elections. H3 and H4 are therefore supported for both parties in 2019. Overall, we can observe the following: Conservative campaigns went from being slightly more effective in Remain areas before the referendum to being clearly more effective in Leave areas by 2019. By way of contrast, while Labour's campaigns were more effective in Remain seats in 2019 as predicted, the impact of the Leave vote on Labour's campaign effectiveness did not change significantly over the three elections. Thus, while Table 1 shows that the effects of the Leave vote were already helping to explain variation in overall effectiveness of the parties' campaigns, the substantive effects occurred after the referendum. Critically, however, the substantive effect did not occur until 2019.
To summarize, we therefore find broad support for our hypotheses in respect of electoral shocks. H2 is partially supported. While the constituency Leave vote had no discernible impact on Labour's campaigns in 2015, there was an effect in respect of Conservative campaigns faring less well in areas with a higher constituency Leave vote.  To test the significance of the interactions by year and to generate combined plots shown in Figure 4 and 5 we ran an equivalent model with pooled data from all three elections with interactions between all variables and election year dummies. The results are substantively identical to above (see online supplementary material).

Assessing Overall Campaign Effects
We can further assess the impact of the electoral shock of the 2016 referendum by examining the overall effectiveness of the parties' campaigns. Table 4  effectiveness was not only on the moderating role of prior popularity, but also on the overall level of effectiveness -suggesting further evidence of the impact of the electoral shock. As a further check, we assess whether the apparent exogeneity of the constituency Leave vote on campaign effectiveness is confirmed by examining the distribution of parties' campaign intensity. It is well established that campaign intensity is not equally distributed across all districts or constituencies. The distribution is principally a function of electoral competition, local organisational resource and electoral context (Fisher, 2000;Fisher Denver & Hands, 2006;Fisher et al, 2011Fieldhouse et al, 2020a). In a period of electoral stability, these resource decisions are likely to (and generally do) lead to electoral payoffs, because parties have a good idea of their likely electoral impact, given the importance of prior levels of popularity in the conditioning of electoral effects. As Fieldhouse et al (2020a: 538) show, the electoral impact of campaign intensity is subject to increasing marginal returns. That being so, parties should perform better electorally where they campaign more intensively. In this context, then -parties' decisions about the distribution of resource will be an important aspect influencing the level of their electoral impact.
However, where there is an electoral shock, a distribution of resource which fails to take account of the shock, it is unlikely to produce the same level of electoral payoff as in a period of electoral stability. The basis for the distribution of effort will no longer be optimal for the delivery of payoffs unless it accounts for the issue underlying the electoral shock. What we observe in Figure 3, is that there is a clear relationship between the size of the Leave vote and campaign effectiveness in 2019, with Labour campaigns being more effective in constituencies with a higher Remain vote and Conservative campaigns more effective in constituencies with a higher Leave vote. The question, therefore, is whether such patterns occurred as a result of the parties' decisions to adjust the distribution of campaign effort to account for the electoral shock of the constituency Leave vote.
As  . Table 5 illustrates, therefore that parties were continuing to distribute effort based on the principles of prior popularity. We compare this in Table 6 with the distribution of campaign effort in respect of the level of the constituency Leave vote. In 2017, Labour ran slightly more intense campaigns in Remain seats, the Conservatives similarly ran slightly stronger campaigns in Leave seats. In 2019, both parties engaged in more similar levels of activity in Leave and Remain seats (with slightly more activity in Remain rather than Leave seats). And, in both years, a simple correlation of the level of campaign intensity in the individual district or constituency and the size of the Leave vote reveals no statistically significant relationships.

Discussion & Conclusions
The electoral impact of campaigns is strongly conditioned by context. In periods of relative electoral stability (albeit ones where there is electoral volatility), the most important contextual factor is underlying popularity, and the geography of campaign effectiveness can be described by popularity equilibrium Fieldhouse et al, 2020a). In such circumstances, parties' campaigns are most electorally effective where they are neither too unpopular nor too popular. impacted negatively on campaign effects overall: that is, they were, on average, notably lower for both parties in the two elections following the 2016 referendum, than in the previous six elections. Third, the effect of the constituency Leave vote on campaign effectiveness was exogenous to the parties' own efforts in terms of the distribution of campaign effort. Thus, there is evidence of the impact of the electoral shock on campaign effectiveness in both post referendum elections, but this was not fully realised until the second election in 2019.
The 2016 referendum evidently had a seismic effect on British politics and a very significant one on electoral behaviour and campaign effectiveness. What is unclear at this stage is whether the electoral shock of 2016 will fade in importance over time, or whether it forms the bases of new alignment in politics, embedding the effect long after the issue itself loses relevance. If it does fade, it is not yet apparent how quickly this will occur, particular as we have seen, the impact of 2016 was more fully felt in the election of 2019 rather than the one two years previously.
Equally, the result may produce a new basis of party alignment (Fieldhouse et al 2020c). Just as class became institutionalized into the fabric of politics and survived long after the original