Legislator Responsiveness to Protest

Our central argument is that legislators have an electoral incentive to respond to protests by their constituents. In line with existing work, we assume that the legislators' primary motivation is to achieve re-election, not for its own sake but as a necessary pre-condition for achieving any other political goals (Mayhew Reference Mayhew1974). The main determinant of legislators' re-election chances is usually their party's popularity. However, under electoral systems that allow voters to express a view on individual candidates, MPs can also try to generate a ‘personal vote’: additional support that is attracted to them in particular (André, Depauw, and Martin Reference André, Depauw and Martin2016). In practice, legislators aim to cultivate a personal vote by persuading local voters that they have attractive qualities (‘advertising’), views (‘position-taking’), and achievements (‘credit-claiming’) (Mayhew Reference Mayhew1974).

If legislators wish to echo their constituents' views and priorities, they must first identify them. Existing work suggests that MPs learn about their constituents' views from several sources. These include relatively permanent features of the constituency like the presence of a military base (Soroka, Penner, and Blidook Reference Soroka, Penner and Blidook2009), time-varying conditions such as levels of unemployment (Borghetto, Santana-Pereira, and Freire Reference Borghetto, Santana-Pereira and Freire2020), and more direct signals of public concern like petitions (Blumenau Reference Blumenau2021). In the same vein, protests can influence legislators' behaviour by providing them with information about their constituents' views (Gillion Reference Gillion2012; Lohmann Reference Lohmann1993; Wouters and Walgrave Reference Wouters and Walgrave2017). Collective action involves various costs, ranging from the simple opportunity cost of attending a protest to the serious risk of encountering physical harm or legal difficulties as a consequence (Gause Reference Gause2022, 261). Protests thus highlight to legislators that an issue is sufficiently salient among a group of citizens for them to incur these costs. Of course, even large protests are typically attended by a relatively small percentage of the overall population (Chenoweth and Belgioioso Reference Chenoweth and Belgioioso2019). Even so, protesters are demonstrably politically active and organized, so may be an important portion of voters at future elections. Moreover, the wider public may also take their cue from local protests (Lohmann Reference Lohmann1993), giving them an electoral impact beyond simply the votes of those who attended.

How can MPs signal their responsiveness to local protests? In many parliamentary democracies, individual legislators have relatively limited influence over policy outcomes due to rules that concentrate influence on party groups, and especially the government. As a result, a more feasible strategy is so-called ‘position-taking’ – publicly expressing views and priorities that accord with those of the protesters. MPs have a range of venues for doing this, but two forms of communication are particularly important. First, they can speak about constituents' priorities in parliament itself. This public and publicized arena provides a highly visible platform for signalling responsiveness to constituents' priorities (Fernandes, Debus, and Bäck Reference Fernandes, Debus and Bäck2021). While very few citizens observe parliamentary behaviour directly, they can learn about it indirectly through news coverage and MPs' own publicity, hence the growing comparative evidence that representatives' individual legislative behaviour has at least some impact on their public image and electoral fate (e.g. Ansolabehere and Kuriwaki Reference Ansolabehere and Kuriwaki2022; Hanretty, Mellon, and English Reference Hanretty, Mellon and English2021). Second, MPs can attempt to communicate more directly with voters via social media. Platforms like Facebook and Twitter provide politicians with a low-cost way of posting messages online, which can then be shared widely by their supporters. Consequently, politicians' online communication has become an important avenue for cultivating popularity with constituents (see e.g. Barberá et al. Reference Barberá2019; Enli and Skogerbø Reference Enli and Skogerbø2013).

We therefore expect that politicians should respond to local protests by giving greater emphasis to those protests' subject matter in their offline and online political speech.Footnote ² This produces our first hypothesis.

Online v. Offline Responses

We expect MPs' responsiveness to protests to vary for different kinds of political speech. In particular, we expect greater responsiveness in MPs' online speech than their offline speech. This is for two reasons: the greater constraints on parliamentary speech, and the differing audience of the two domains.

We first expect greater responsiveness online because MPs' online communication is subject to fewer constraints. MPs' offline speech is constrained because parliamentary speaking time is shared, finite, and organized. High demand for speaking time means that MPs may not always be able to speak on a given day or in a particular debate; i.e. speaking time is shared. By contrast, they can post online content whenever they like. Online speech may thus facilitate a more timely reaction to the concerns of local protesters. The limited availability of parliamentary speaking time also means that MPs must choose which issues to prioritize in their speeches; i.e. speaking time is finite. This is much less true of online speech, where MPs can – and do – post many messages in a single day, leaving them freer to address a wider range of subjects.Footnote ³ Finally, MPs' offline political speech is constrained by parliamentary timetabling; i.e. it is organised. At its most prosaic, this constraint might simply mean that the parliamentary arena is unavailable on certain days, such as weekends. But even when parliament is sitting, MPs are still constrained by which issues feature on the agenda. For example, discussion on any given day might be focused on a specific bill or questions to a particular government department. MPs may thus struggle to find a relevant way to raise the issue addressed by recent local protests.

Second, the responsiveness of online and offline speech to local protests may also differ due to their likely audiences. Online communication through social media can be directly observed by voters, who often have selected into following politicians with whom they are ideologically congruent (Barberá Reference Barberá2015). Moreover, users of Twitter, in particular, are more politically attentive and more liberal than the general public, meaning that legislators may tailor their speech accordingly (Barberá et al. Reference Barberá2019; Mellon and Prosser Reference Mellon and Prosser2017). By contrast, the public rarely observes parliamentary speeches directly; their immediate audience is other MPs and the watching media. Offline speech can be – and is – used for indirectly signalling to the public. But this route is highly mediated, depending on news coverage or the publicity efforts of MPs and their parties. Consequently, offline communication has a less guaranteed and less select audience. This means that MPs wishing to signal responsiveness to protesters' concerns may see online communication as a more effective way of doing so.

Online speech thus provides MPs with an immediate and unmediated avenue for signalling their sympathy with protesters' concerns. By contrast, MPs’ offline speech is subject to a range of constraints and only reaches voters indirectly. We thus expect that MPs' online speech should be more responsive to local protests than their offline speech.

Case Details and Selection

To examine the effects of climate protest on political speech, we focus on the FFF campaign in the UK. Following the individual protest sit-in (‘School Strike for Climate’) by Greta Thunberg outside the Swedish parliament in late 2018, the FFF campaign (also known as ‘Youth for Climate,’ ‘Climate Strike’, or ‘Youth Strike for Climate’) emerged as an international grassroots campaign designed to compel elected representatives to act on the impending climate crisis. The stated aim of the FFF campaign was ‘to put moral pressure on policymakers, to make them listen to the scientists, and then to take forceful action to limit global warming’.Footnote ⁴ The movement culminated in thousands of school strikes on named Fridays worldwide throughout the following year. The UK itself saw a considerable number of climate protests organized under the FFF banner over this period (see Fig. 1). The FFF protests explicitly sought to instil urgency into policy debates. Thunberg (Reference Thunberg2019, 57–68) chose to speak of the climate crisis, emergency, and breakdown instead of climate change in her speech at the UK parliament in April 2019. The following week, MPs endorsed Labour leader Jeremy Corbyn's motion to declare a ‘climate emergency’, which he introduced with reference to the ‘unprecedented upsurge of climate activism’.Footnote ⁵

Figure 1. (a) Barplot daily sum of FFF protests; (b) Climate speeches mentioning climate keyword phrases over time. The top panel is daily frequency; the bottom panel plots our full panel dataset, where dark green tiles indicate an MP speech-day where climate keyword phrases are mentioned, and white otherwise; (c) Tweets mentioning climate keyword phrases over time. The top panel is daily frequency; the bottom panel plots our full panel dataset, where dark green tiles indicate an MP tweet-day where climate keyword phrases are mentioned, and white otherwise. (d) Map of the total number of FFF protest events by constituency over the entire observation period; (e) Kernel density plot of climate tweet and speech by two main parties. Points represent daily square root percentages of speeches/tweets mentioning climate; the width of density curve corresponds to the relative frequency at a given point in the distribution. Overlaid boxplot shows the median, first, and third quartiles of distribution.

For the purposes of our analysis, we set an observation period of 8 June 2017 to 12 December 2019. This period's start and end points mark the 2017 and 2019 general election dates and overlap with the peak of FFF protests in 2019. We study the effect of these protests in the UK during this period because of data availability and because it plausibly represents a most-likely case.

The UK parliament records and publishes all political speeches delivered in both Houses in the Hansard record. UK MPs are also highly active on the micro-blogging website Twitter (85 per cent had active Twitter accounts over this period), meaning we are able to draw comparisons between offline speech delivered in the House of Commons and online speech published as tweets. Geolocated FFF protest data is also available, meaning we can assign protests to politically meaningful geographic units (that is, constituencies).

We believe the UK context represents a ‘most likely’ case for establishing a relationship between protests and political speech for two main reasons. First, climate policy has broad support among the public and political elites, with all main parties making prominent policy pledges in this period (see Carter and Pearson Reference Carter and Pearson2020). Climate protests in this context might thus be best viewed as an attempt to push parties to undertake larger and quicker measures, and to raise the salience of existing concerns. Given that the climate is a valence issue, MPs should be more willing to see climate protests as indicative of wider public opinion, and less concerned that responding to these protests would mean diverging from their party compared to a highly polarized context. Second, the UK's single-member plurality electoral system gives MPs relatively strong incentives to appeal to local concerns (André, Depauw, and Martin Reference André, Depauw and Martin2016), further increasing the likelihood of them responding to climate protests in their constituency. We are thus studying a political context that maximizes the chances of finding a relationship between climate protest and individual MPs' political speech. We see this as an advantage since our study is the first to investigate the relationship using large-n speech data from both offline and online contexts.

Data Collection

We use three main data sources to code our two dependent variables and the key independent variable. We collected MPs' online speech from Twitter, using the R package ‘academictwitteR’ (Barrie and Ho Reference Barrie and Ho2021). To determine the full set of elected MPs with active Twitter accounts, we began with a Twitter ‘list’ of UK MP Twitter accounts.Footnote ⁶ We then cross-checked this list against the Hansard record of all 650 MPs elected at the 2017 general election, adding accounts that had been wrongly omitted and removing accounts included in error. This process resulted in a set of 553 UK MPs with active Twitter accounts. With this set of users, we then collected all tweets (including ‘retweets’, ‘quote tweets’, and ‘replies’) by UK MPs using the V2 Academic Research Product Track API endpoint, which provides academic researchers with access to the full archive of all published (and as-yet undeleted) tweets. Our data includes every tweet published by an MP over this period.Footnote ⁷ This resulted in a dataset of ~2.4m tweets over our observation period.

Based on this dataset, we identified all tweets that mentioned climate-related topics. We did this through a two-step dictionary expansion technique. We first trained a word embedding layer across the full corpus of MP tweets using the GloVe algorithm and the R packages quanteda (Benoit et al. Reference Benoit2018) and text2vec (Röder, Both, and Hinneburg Reference Röder, Both and Hinneburg2015). Word embeddings are vector representations of words, which encode semantic meaning in the embeddings space.Footnote ⁸ The distance between vectors thus provides a proxy for semantic meaning: more similar words will be closer to each other in the embeddings space (Rodriguez and Spirling Reference Rodriguez and Spirling2022). This means we can use word embeddings as a discovery technique for expanding dictionaries of relevant terms (Rice and Zorn Reference Rice and Zorn2021). We began with a seed word ‘climate’ and looked up the most (cosine) similar words in our word embeddings. This resulted in an initial set of ten words by which to filter our tweets for relevant terms.Footnote ⁹

These initial terms are useful for reducing the size of relevant data, but are obviously imprecise: ‘climate’ might refer to the ‘political climate’ and ‘nature’ may refer to the ‘nature of debate’. After reducing the data size with this initial filter, we then output the top 4,000 bigrams and hand coded them for relevance to climate-related issues. During this process, we also incorporated relevant unigrams (e.g. unspaced hashtag phrases like #climatechange) and trigrams (e.g. ‘green new deal’). We then sorted the climate-related terms into two lists: a general list of terms that refer to climate in the abstract (for example, ‘climate change’, ‘climate breakdown’, and ‘environmental rights’) and more specific climate-related terms (for example, ‘carbon emissions’, ‘plastic waste’, and ‘green new deal’). We refer to these as our core and expanded lists. The full lists of terms can be found in Supplementary Materials Tables APP-1 and APP-2. We use the core list to create our first dependent variable – the daily number of tweets from each MP containing one or more of the core climate terms during our study period. This results in a total of 15,182 climate-related tweets. In the supplementary materials, we present results for the expanded list of policy terms.

We used the Hansard record to compile data on offline speeches by UK MPs in parliament. Hansard records all instances of parliamentary speech; the speeches and metadata are made available by ‘TheyWorkForYou’. For our analysis, we use only instances of speech classed as ‘Speech’, which are oral questions or oral contributions to debates in the House of Commons.Footnote ¹⁰ In total, we have ~150 k Hansard parliamentary speech records for our set of 553 MPs. Consistent with our measure of online speech, our second dependent variable is a daily count of the number of speeches by each MP containing one or more of the core climate terms during our study period. This results in a total of 1,774 climate-related speeches.

One concern with our technique for capturing climate-related tweets and speech data is that we do not record the speaker's stance toward climate-related issues. That is, we might also be capturing climate-sceptic speech. We checked this by generating two random samples of 500 speeches and 500 tweets and investigated these for climate scepticism. We found no evidence of climate scepticism or denial in either sample. This supports our characterization of climate change as a valence issue in the UK context, and provides reassurance that our measure is well-suited for capturing legislators' responsiveness to climate protests.

Finally, to retrieve data on the incidence of climate protests, we use a crowd-sourced database of FFF protests, available on the dedicated campaign website for the movement.Footnote ¹¹ The data were scraped from a map of geolocated protest actions, alongside corresponding information on the town where the protest occurred, precise location (for example, ‘in front of townhall’), time, date, and link to the Facebook or Instagram event page. We assigned each protest to constituencies using shapefile constituency boundaries for the 2017 general election.Footnote ¹² Our independent variable was then coded as the daily sum of FFF protest events in each constituency.

Descriptive Analysis of Climate Speech

We now present some aggregate descriptive trends in the salience of climate as a topic and the tone in which it was discussed.

Our data are daily protest, tweet, and parliamentary speech data at the constituency- or MP-level for 553 UK MPs during the 2017–2019 parliament. We plot the frequency and location of climate protests and the timing of climate speech in Fig. 1. As protests took off in 2019, so did climate-related speech. MPs sent 5.9 climate tweets per day between 2017 and 2018, but 33.9 per day in 2019. For parliamentary speech, MPs mentioned climate change 1.0 times per day in 2017–2018 and 3.6 times per day in 2019.

We can also observe semantic shifts in the climate speech of individual legislators using recently developed word-embedding techniques. Therefore, we apply a technique called ‘à la Carte on Text’ (ALC) to identify the differences in how MPs discussed climate change during the 2017–2019 parliament.Footnote ¹³ This technique builds from recent contributions that demonstrate the efficiency gains of using pre-trained word-embedding layers to gauge semantic change across document-level covariates or, as in our application, time (Arora et al. Reference Arora2018; Khodak et al. Reference Khodak2018). It relies on the insight that embeddings for a particular (or even very rare) target word may be derived by averaging the vectors of embeddings for words within its (here, six-word) context window from a pre-trained embedding layer.Footnote ¹⁴ In this application, we use the GloVe pre-trained embedding layer (Pennington, Socher, and Manning Reference Pennington, Socher and Manning2014) and follow the recommendation of Rodriguez, Spirling, and Stewart (Reference Rodriguez, Spirling and Stewart2023) by applying a transformation matrix to downweight commonly appearing words.

We can then observe temporal trends by calculating the cosine similarities between our target word ‘climate’ and a set of candidate words over time for both our tweets and speech corpus. Here, as candidate word, we use the word ‘emergency’. To recover the over-time cosine similarities, we first split our observation period into year-week slices and then find the context words around our target word, ‘climate’, for each week. We then estimate a time-period-specific embedding for the word ‘climate’ using the ALC approach. Here, we take the average of the vectors of surrounding context words (from the same pre-trained GloVe embedding layer) and then apply a transformation matrix to downweight commonly appearing words. From this procedure we are able to induce a single period-specific embedding for our target word ‘climate’ over each time period. Once we have recovered these embeddings, we can calculate the (ℓ2-normalized) cosine distance between the vectors for ‘climate’ and ‘emergency’ over time. We display the results of this analysis in Fig. 2.

Figure 2. Weekly cosine similarities between ‘climate’ and ‘emergency’: (a) MP tweets and (b) MP speeches. Lines show loess regression smoothing with bandwidth set to one.

This analysis shows that, in both MPs' tweets and their speeches, the word ‘climate’ became more tightly related to the word ‘emergency’ in our time period, as MPs increasingly discussed climate politics in terms of a climate emergency.Footnote ¹⁵ Notably, this semantic shift arose in 2019 around the time of the FFF protests, which aimed to reframe climate discourse to demand greater urgency. This shows changes in political speech that mirror the demands of the FFF protesters. This analysis lends plausibility to the idea that the FFF protests may have shaped political speech in the UK. Nonetheless, this analysis tells us little about whether this macro-level development is driven by a micro-level relationship between local protests and MPs' attention to climate as a political issue. This is the focus of our main analysis, which we now turn to explaining in more detail.

Estimation

We create a panel dataset for each of the 918 days in our observation period for each of our 553 MPs (when still in position). In addition to our two dependent variables and key independent variable, we incorporate further information for each MP. We use the parlitools R package (Odell Reference Odell2017) to code MPs' party.Footnote ¹⁶ We also include a binary variable indicating any periods when an MP served as a minister or shadow minister, based on Colebrook and Priddy (Reference Colebrook and Priddy2020) and the UK parliament data portal.Footnote ¹⁷ Three MPs in our dataset did not serve their full term in parliament due to resignation, death, or a recall petition. We only include data for these MPs up to the date of the by-election for their replacement. Our analysis excludes the new MPs elected at these by-elections, all Sinn Féin MPs (who do not take up their seats), and any MPs who served as Speaker or Deputy Speaker (who chair debates rather than participating in them).

We appreciate that the underlying propensity to discuss climate change varies across MPs – partisan identity, constituency makeup, and knowledge about climate change necessarily differ across officials. However, these are stable traits that we do not expect to change in our relatively brief observation period. Therefore, we include MP-level intercepts that control for these fixed MP-specific effects. This allows the effect of climate protests to be recovered as a difference in climate speech from the average number of climate speeches for a given MP. Since MPs overlap entirely with constituencies, MP fixed effects also absorb all unobserved constituency-level heterogeneity, including the underlying constituency-level propensity for climate protest.

We must also be sensitive to changes in the incidence of climate speech and climate protests. Both of these key variables increase over time (see Fig. 1), which could confound the effects of protests on legislator behaviour. We mitigate this in two ways. First, given our exceptionally high-frequency (daily-level) and high-resolution (constituency-level) data on both speech and protest, we can precisely locate both variables in space and time. This means that any change in MPs' climate speech can be attributed to changes in climate protest at their constituency level. Second, we include a vector of time dummies that absorb common trends and shocks across all MPs; for example, accounting for a possible rise in climate speeches and protests around the annual United Nations climate negotiations or the visit of Greta Thunberg to the UK in April 2019. Our high-frequency data allows us to consider year-by-month and, later, year-by-week fixed effects. Taken together, the inclusion of MP and time fixed effects allows the effect of climate protest on legislator speech to be recovered through deviations in the incidence of protests from the long-term average in a given constituency, the average political speech behaviour of an individual MP, and common temporal shocks across constituencies. This specification relies on a version of the parallel trends assumption. In the supplementary materials, we consider extensions that include time-varying MP-level covariates, MP-specific time trends, and a placebo test that replaces the treatment with a measure of foreign protests (table APP-4).

Our main estimating equation is then based on the following functional form:

(1)$${\rm Climate}\;{\rm Speec}{\rm h}_{i, t} = \beta {\rm FFF}\;{\rm Protes}{\rm t}_{i, t} + \gamma {\bf X}_{i, t} + \alpha _i + \delta _d + {\rm \epsilon }_{i, t}$$

where i indexes MPs (and necessarily constituencies), t indexes days, d indexes our measures of time controls (time fixed effects), X is a vector of covariates that contains the sum of an MP's speeches or tweets on that given day t, and will later contain a time-varying dummy for an MP's frontbench status. Controlling for the total number of speeches/tweets in this manner means we model the relative emphasis MPs place on the climate, holding constant their daily propensity to make speeches or publish tweets. Note that, in this specification, we are measuring protest and speech at the daily-level t and controlling for temporal trends and shocks at a higher level of aggregation d (months, weeks). We estimate an ordinary least squares (OLS) regression with standard errors clustered on the MP. In the supplementary materials, we present all the results for binary measures of climate speech and the expanded dictionary of policy terms, as well as several robustness tests.