Lessons From the UK's Lockdown: Discourse on Behavioural Science in Times of COVID-19

Sanders, Jet G.; Tosi, Alessia; Obradovic, Sandra; Miligi, Ilaria; Delaney, Liam

doi:10.3389/fpsyg.2021.647348

ORIGINAL RESEARCH article

Front. Psychol., 17 June 2021

Sec. Personality and Social Psychology

Volume 12 - 2021 | https://doi.org/10.3389/fpsyg.2021.647348

This article is part of the Research Topic Contagious Risks: Perception, Behavior, and Management - Lessons from the COVID-19 Pandemic View all 9 articles

Lessons From the UK's Lockdown: Discourse on Behavioural Science in Times of COVID-19

$\nJet G. Sanders$ Jet G. Sanders¹^*

Alessia Tosi^1,2

Sandra Obradovic³

Ilaria Miligi¹

Liam Delaney¹

¹Department of Psychological and Behavioural Sciences, London School of Economics and Political Sciences, London, United Kingdom
²Independent Researcher, London, United Kingdom
³School of Psychology and Counselling, Faculty of Arts and Social Sciences, The Open University, England, United Kingdom

In recent years behavioural science has quickly become embedded in national level governance. As the contributions of behavioural science to the UK's COVID-19 response policies in early 2020 became apparent, a debate emerged in the British media about its involvement. This served as a unique opportunity to capture public discourse and representation of behavioural science in a fast-track, high-stake context. We aimed at identifying elements which foster and detract from trust and credibility in emergent scientific contributions to policy making. With this in mind, in Study 1 we use corpus linguistics and network analysis to map the narrative around the key behavioural science actors and concepts which were discussed in the 647 news articles extracted from the 15 most read British newspapers over the 12-week period surrounding the first hard UK lockdown of 2020. We report and discuss (1) the salience of key concepts and actors as the debate unfolded, (2) quantified changes in the polarity of the sentiment expressed toward them and their policy application contexts, and (3) patterns of co-occurrence via network analyses. To establish public discourse surrounding identified themes, in Study 2 we investigate how salience and sentiment of key themes and relations to policy were discussed in original Twitter chatter (N = 2,187). In Study 3, we complement these findings with a qualitative analysis of the subset of news articles which contained the most extreme sentiments (N = 111), providing an in-depth perspective of sentiments and discourse developed around keywords, as either promoting or undermining their credibility in, and trust toward behaviourally informed policy. We discuss our findings in light of the integration of behavioural science in national policy making under emergency constraints.

Introduction

Public trust in the transparency and reliability of scientific evidence is an important component of effective responses to major challenges and crises (Hendriks et al., 2015; Pittinsky, 2015). Generally, public perceptions of science are positive: science is often held in high esteem with equally high confidence placed in scientists (e.g., Scheufele, 2013; Jonge, 2015; National Science Board, 2016; Lamberts, 2017; Robert Bosch Stiftung, 2017; Lindholm et al., 2018). However, the application of science in policy has variable success (Sanchez-Paramo et al., 2019). Not all science is deemed fit to inform policy (Anvari and Lakens, 2018; Ioannidis, 2018; Cairney, 2020; Stevens, 2020). Determining when a scientific discipline is ready to inform policy is precarious and can be volatile and the criteria for evidence-readiness can vary depending on what is at stake (Ruggeri et al., 2020). In addition, policy choices are trade-offs shaped by many pressures other than those based on evidence. Direct competition from other pressures can shape public perceptions of the evidence and can steer the policy-makers' ability to implement evidence at hand (Cairney, 2020).

In March 2020, the UK was faced with the high stake policy choice of a national lockdown as COVID-19 spread globally (Kreps and Kriner, 2020). As scientific evidence about the virus and its effects was sparse, a broad range of scientists were called onto expert panels to advise governments. In British policy, unlike many other national governments, one prominent perspective was that of the behavioural science (UK Government, 2020).

The integration of behavioural science into UK policy took several forms, but was most notably embodied by (1) the inclusion of Dr. David Halpern, chief executive of the Behavioural Insights Team (BIT) in the government's Scientific Advisory Group for Emergencies (SAGE) and (2) the development of a behavioural advisory group known as the Scientific Pandemic Influenza Group on Behaviours (SPI-B). It is possible that behavioural science was particularly well-represented in the UK because it has been embedded in British policy for longer and more widely than in other national systems. The UK Cabinet Office was amongst the first to embed a dedicated behavioural science unit (often called the “nudge unit”) to that effect (Sanders et al., 2018). Arguably, it is in part due to this unit that the effect of nudges as a novel policy instrument (Lourenco et al., 2016) and methods to test for their effectiveness (Della Vigna and Linos, 2020) were demonstrable on national policy level and embedded elsewhere. We have since seen an increasing popularity for the policy approach, as evidenced by the growing number of behavioural insight units that advise national governments on issues involving citizen choices in the last 10 years (Whitehead et al., 2014; Halpern, 2015).

Despite these successes, in March 2020 the role of behavioural scientists in the UK's COVID-19 response was heavily debated in the media. This left the questions: what caused debate about the role of this emergent science, what were its consequences (if any) and how can we learn from the communication around its scientific contributions to this high-stake policy? While an emerging body of literature exists on support for behavioural interventions (e.g., Reynolds et al., 2019, Sunstein et al., 2019), far less work has been conducted on trust in behavioural scientists more generally, and no work that we are aware of examines public support for the inclusion of behavioural scientists in committees advising government and shaping policy.

To examine this, we were particularly interested in this debate in media discourse, given media's important role in forming public opinion and therefore setting public trust in emerging science (Van Aelst, 2014). In addition, mass media plays an important agenda-setting role: it can direct collective attention and perceived importance (McLeod et al., 1974), shape how severe an issue is perceived to be, and influence how individuals come to perceive their social and political environment (Tyler, 1980; Protess and McCombs, 2016). In other words, mass media play an important “mediating” role in sharing and shaping how scientific and political expertise is understood by the public (Baum and Potter, 2008; Kim et al., 2018).

The rise of social media platforms such as Twitter in recent years (Gil de Zúñiga et al., 2012) provides a further opportunity to consider how public perceptions are then reflected by news media. For example, Chew and Eysenbach (2010) show that during the H1N1 pandemic, individuals used Twitter to share resource-related posts, with news media websites being the most popular sources to share. But social media users do not only reproduce existing information, they are also actively engaged in how key debates are shared and understood, with the potential to impact decision-making in significant ways (Bello-Orgaz et al., 2017). Twitter gives rise to a huge volume of text-based data, with over 500 million “tweets” generated by users each day (Chae, 2015; Mention, 2018). These tweets can be useful for tracking how public opinion develops around key social issues (D'Andrea et al., 2019). Specifically, relevant for the present context, Twitter data has been used to analyze public opinion on vaccination (Bello-Orgaz et al., 2017; D'Andrea et al., 2019), the role of fake news during the pandemic (Gruzd and Mai, 2020) and shifts in public emotions during the pandemic, from fear to anger (Lwin et al., 2020). This emerging research highlights the role of media (both traditional and social) in providing key information, focusing public attention on social issues and shaping public opinion and emotions. We build on this literature in our paper.

This paper provides a key case study on trust and acceptability surrounding the contributions of social and behavioural science in times of crises. Specifically, the paper aims to address the following questions: How was this emergent science debated in the print media, and what can we learn from its perceived credibility in informing policy? To answer these questions, we examine public and media discourse surrounding the high-stake policy decision of the first national lockdown in the UK in March 2020. We draw on a 24-week period to track the discourse as it evolved. Specifically, our case study draws on two independent sources: In Study 1 we use print media to examine the salience, sentiment and co-occurrence of behavioural science keywords in the media. In Study 2, we draw on Twitter data to track how this print media discourse is picked up and appropriated in public discourse. Lastly, in Study 3, we draw on a subset of newspaper articles from Study 1 to provide a more in-depth analysis of how discourses around trust and credibility of behavioural science are constructed and either promoted or undermined.

In bringing together the findings from the three studies, a key objective of this paper is to understand what the consequences (if any) of this discourse were, and how we can learn from it to further trust toward scientific contributions to high-stake policy.

Study 1 Newspaper Discourse Analysis

Top newspapers have been shown to sway common understanding of scientific disciplines and can be used as a proxy to measure understanding of their place in public policy (Bauer et al., 1994; Mutz and Soss, 1997; Schäfer, 2012). As the contributions of behavioural science to the UK's COVID-19 lockdown policies developed, and debate emerged in the British media about its involvement, we reasoned that, in the lead up to, during and after the UK COVID-19 lockdown period in March 2020, public perceptions of behavioural science contributions to this high-stake UK policy decision should be detectable from newspaper articles. With this in mind, we set out to explore (1) the prevalence of behavioural science actors and concepts in relation to national level policy making, (2) the valence associated with such actors and concepts (see Alamoodi et al., 2020 for a review of its use in other COVID-19 policy contexts), and (3) the co-occurrence of key behavioural science concepts and actors over the lockdown period of 2020.

Materials and Methods

Materials

We retrieved news articles from the online database Lexis Nexis for an 8-week window on either side of the hard UK lockdown (27th of January 2020–10th of July 2020). We drew on the 15 UK newspapers with the highest circulation levels (see Supplementary Material 1). We estimate that articles in these newspapers collectively reached almost 8 million people in print and in digital editions (~12% of the British population) on a monthly basis (Mayhew, 2020; Worldometer, 2020).

Using a snowball method, we developed a query to identify articles relevant to the discussion of behavioural science (see Supplementary Material 2A for the various stages and final query). The search produced a corpus of 865 articles. Deduplication and removal of incomplete articles resulted in a sample of 679 articles. These were qualitatively reviewed by three coders for relevance to the topic of behavioural science. This left 647 articles (ranging from 1 to 47 per news outlet; see Supplementary Material 1 for details) for the quantitative analysis (see Figure 1).

FIGURE 1

Figure 1. Flowchart of data selection and cleaning process for Study 1.

Keyword Processing

We defined an initial set of 42 keywords based on the snowball method applied through the search query (see Supplementary Material 3 for a complete set). As one word can be expressed in different ways (e.g., abbreviated, singular/plural form, or by use of synonyms), keywords were grouped to form primary keywords as follows: (1) plurals were standardized into a singular form: e.g., behavioural science and behavioural science as behavioural science; (2) synonyms were unified: e.g., nudge unit and Behavioural Insights Team as Behavioural Insights Team; and (3) we integrated semantically related keywords based on expert knowledge: e.g., nudge, nudging, nudge theory, and nudge strategy were noted as nudge. As exceptions to this rule we kept psychologist, behavioural scientist, and behavioural economist as stand-alone primary keywords. As profession names often preface unique actors (as opposed to their plural counterparts; e.g., Professor of Health Psychology Susan Michie VS Professors at Oxford), they lend themselves as proxies for actors not captured in our keyword base¹.

This resulted in 20 primary keywords: behaviour science, affiliated disciplines (psychology, behaviour economics), behavioural science concepts (nudge, choice architecture, irrational behaviour, behaviour change, behavioural analysis, behavioural insights), commonly named actors in national or international behavioural policy work (SPI-B, Behavioural Insights Team, Michie, Halpern, Chater, Thaler, Sunstein, Kahneman), and unnamed behavioural science actors (behavioural scientist, psychologist, behavioural economist).

Analyses

Salience

To assess the salience of primary keywords over time, we first removed all “parts-of-speech” apart from nouns or keywords. This is based on the assumption that it is nouns that are the part of speech that represent the content of an article (Stuart et al., 2013). A salience score was calculated for each primary keyword for every two-week period. The score was a product of the keyword's normalised corpus frequency (i.e., number of keyword occurrences divided by total word counts per 10,000 words) and the keyword's relative document frequency (i.e., proportion of articles in which the keyword was mentioned). This composite metric allowed us to account both for the centrality of a keyword in the narrative of the articles published in the 2-week period (normalized corpus frequency), and the spread of the keyword in the media in the same period (relative document frequency; Manning et al., 2008).

Sentiment

Targeted sentiment analysis was used to assess perceptions of behavioural science applied in national public policy context. We first identified all sentences (N = 1280) in our corpus where a behavioural science keyword occurred. As a sentence could contain more than one keyword (median = 1, range = 1–5), this resulted in a sample of 1,605 keyword-sentence pairs, termed opinion contexts. Each opinion context was coded manually for sentiment polarity expressed toward the keyword on a 5-point scale: −2 (extremely negative), −1 (moderately negative), 0 (neutral), +1 (moderately positive), +2 (extremely positive). Opinion contexts were also reviewed to contain reference to national-level policy (e.g., mention of government, minister, no. 10, see Supplementary Material 3 for a full list). When this was the case, sentiment polarity toward the policy actor linked to behavioural science was also rated. A subset of 110 opinion contexts were coded by all three coders to produce an inter-rater agreement and solve cases of disagreement, with the remainder coded by single coders (Hallgren, 2012).

To match salience and sentiment scoring, results were presented for two-week intervals over the period of the first national lockdown of 2020 in three sentiment categories: negative (−1; −2), neutral (0) and positive (+1; +2).

Co-occurrence

Finally, we used co-occurrence network analysis to investigate how the conceptual structure of the public narrative around behavioural science evolved over the period of the first national lockdown (Corman et al., 2002; Paranyushkin, 2011). To allow for reasonable variance in co-occurrence, we opted to move from two-week windows to a pre-, during- and post-lockdown window of analysis.

To analyze the narrative around keywords in the relevant context, we calculated co-occurrence at opinion-context level (N = 302 pre-, N = 540 during-, N = 438 post-lockdown) for two keywords or a keyword and any other term appearing in the same sentence. We expressed co-occurrence via the Dice coefficient: the ratio between the co-occurrence of two keywords and the sum of their individual occurrences multiplied by two (Frakes and Baeza-Yates, 1992; see Supplementary Material 6 for details). Simply put, two keywords that never co-occur have a coefficient of 0 and two keywords with identical occurrence have a coefficient of 1².

In the co-occurrence networks, nodes represent terms and an edge between two nodes indicates the terms' co-occurrence, with a weight proportional to the strength of their association (i.e., the Dice coefficient) (Liu et al., 2012; Katsurai and Ono, 2019; Paranyushkin, 2019; Kim et al., 2020; Puerta et al., 2020). By graphically representing patterns of co-occurrence between terms, co-occurrence networks identify the importance of terms and their inter-relatedness (Van Eck et al., 2006; Van Eck and Waltman, 2007). The network analysis was conducted using the Python NetworkX 2.5 library (Hagberg et al., 2008).

To calculate co-occurrence, the opinion contexts were pre-processed as follows to reduce noise interference (cf. Véronis, 2004; Jurgens, 2011; Kim et al., 2020). (i) Common two- and three-word phrases that did not involve any of our behavioural-science keywords, were replaced with the corresponding bigrams/trigrams (e.g., “public health” with “public_health”) based on collocation statistics across all 647 articles, using Python gensim's Phrases model (Rehurek and Sojka, 2010). (ii) We only retained terms (beyond our keywords), whose part-of-speech was either adjective or noun. (iii) We removed terms which appeared less than 20 times across all three time windows (occurrence: median N = 1, Interquartile Range = 2, range = 1–223). Of 3,709 unique terms, 3,635 were eliminated and 74 were retained (appearing in 648 pre-, 1,470 during-, and 1,432 post-lockdown opinion contexts). (iv) The Dice coefficient was calculated between pairs that included at least one behavioural science keyword and with a raw co-occurrence of at least 10 in that time window.

To understand how the relevance of, and narrative around the keywords evolved, we calculated and compared the following network- and node-level metrics (Sudhahar et al., 2015):

(a) Network density: the ratio of the actual number of links between keywords to the maximum possible number of links. On a scale from 0 to 1, higher value indicates a cohesive network.

(b) Network average clustering coefficient: the interconnectedness of nodes in a network on a scale from 0 to 1. If more terms co-occur with each other, the clustering coefficient is high (it is 1 if every node is connected to all other nodes). If terms do not co-occur with each other, the clustering coefficient is low.

(c) Node weighted degree centrality: the sum of the edge weights for edges incident to that keyword. Higher values indicate more frequent direct links to other keywords.

(d) Node weighted between centrality: the degree to which a keyword stands between others. Higher values indicate greater importance in bridging subsets of keywords.

(e) Communities: we used the Louvain algorithm to detect communities of co-occurring words, i.e., “thematic clusters” in our networks (cf. Williams et al., 2016; Lozano et al., 2019). The Louvain algorithm works by maximizing modularity (Blondel et al., 2008). Modularity measures the density of connections within communities compared to the density of connections between communities. It takes on values between−1 and 1, with a higher value representing better community definition (Newman and Girvan, 2004). The Louvain algorithm has been found to be one of the fastest and best performing community-detection algorithms in comparative analyses (Lancichinetti and Fortunato, 2009; Yang et al., 2016).

Finally, the changing trends in keywords' relevance and narrative were identified by comparing the detected communities and ranking of keywords for node centrality metrics (c) and (d) in the three different time periods.

Results

From all analyses we excluded 9 keywords due to extremely low overall frequency (<30 occurrences over the 24-week period of interest; see Supplementary Material 5A for details) as they did not provide enough data points across time to determine trends in our metrics of interest. This left 11 primary keywords: behavioural science, the discipline terms behavioural economics and psychology, four of the eight named actors (Behavioural Insights Team, Halpern, Michie, and SPI-B), two of three unnamed actors (behavioural scientist and psychologist), and two of six concept terms (behaviour change and nudge).

Below, we present trends in salience and sentiment toward behavioural-science keywords over time, followed by reference and sentiment toward public policy application and co-occurrence. As frequently the case in descriptive exploratory studies of linguistic data (e.g., Bian et al., 2016; Kim et al., 2020; Sharma et al., 2020), we contain our results to descriptive findings.

Salience and Sentiment of Keywords Over Time

Primary keyword behavioural science showed two clear surges: the first started one month prior to the UK lockdown (−2) and ended just after lockdown (+1) and the second rather spike-like surge occurred within a two-week period one month after the “hard” UK lockdown measures eased (+6; see Figure 2 and Supplementary Material 4). Simultaneous to the surges, we see an increase in polar sentiments: positive and negative sentiments are greater during these periods compared to other time-periods. This pattern is reminiscent of one commonly reported: “conflict” is deemed of news value and determines the extent to which journalists pay attention to politics (Galtung and Ruge, 1965; Harcup and O'neill, 2001; Van der Pas and Vliegenthart, 2016).

FIGURE 2

Figure 2. Salience of and sentiment toward the keyword “behavioural science” over a 12 two-week time-period surrounding the first British national lockdown of 2020 (gray area) in print media (top 15 UK newspapers). Salience is calculated for a 2-week period as the normalized keyword frequency (per 10,000 words) multiplied by the proportion of articles that mention the keyword. Sentiments are represented in counts of positive (+1 or +2), neutral (0), and negative (−2 or −1) bubbles over time, in green, white and red respectively. The size of the bubble is proportional to the count of sentiments in that polarity class toward the keyword.

What seems to associate with the observed divisiveness? Discipline terms and unnamed actors do not show similar sentimental surges. Psychology (Figure 2I) seems to show a subdued version of behavioural science salience, with notably greater positive than negative sentiment. Behavioural economics is in fact largely absent from the narrative, with minimal salience in newspaper articles, but stable polarity over time.

Similarly, unnamed actors, such as psychologist (Figure 3H) or behavioural scientist (Figure 3D) do not share the surges in sentiment polarity observed for behavioural science. Although unnamed actors show a slight uptick in salience, they show a relatively steady (mostly neutral) sentiment.

FIGURE 3

Figure 3. Salience of and sentiment toward the 10 primary keywords over the 12 two-week time period surrounding the first British national lockdown of 2020 (gray area) in print media (top 15 UK newspapers). (A) behaviour change (concept), (B) behavioural economics (discipline), (C) behavioural Insights Team (named actor), (D) behavioural scientist (unnamed actor), (E) halpern (named actor), (F) michie (named actor), (G) nudge (concept), (H) psychologist (unnamed actor), (I) psychology (discipline), (J) SPI-B (names actor). Salience is calculated per 2-week period as the normalized term frequency (per 10,000 words) multiplied by the proportion of articles that mention the keyword. Sentiments are represented in counts of positive (+1 or +2), neutral (0), and negative (−2 or −1) bubbles over time, in green, white and red respectively. The area of the bubbles is proportional to the count of sentiments toward the keyword.

We reach a different conclusion for named actors and concept terms. Salience for keyword Michie also mimics the behavioural science trend over time in subdued form, but with positive polarity during the first surge (−1). Keywords Halpern and Behavioural Insights Team show a nearly identical rise in salience to behavioural science in the period leading to lockdown, but rather eliciting a negative response. All actors thus seem to associate with the divisiveness we observe, possibly holding opposite perspectives. This narrative finds support in that all three actors only seem to emerge as public figures of behavioural science only around this pre-lockdown time period (−1).

A final pattern of divisiveness is aligned with the keyword nudge. Although nudge was not nearly as salient as other primary keywords, we observe negative sentiment during the first surge. In fact, nudge is the only primary keyword which, throughout the 24-week period, attracted more divisiveness than neutrality. Moreover, nudge, Halpern, and Behavioural Insights Team are the only primary keywords to show greater negative than positive polarity.

What seems to associate with the observed non-divisiness? We are particularly interested in capturing patterns of neutrality as many may deem this to be the category of sentiment best suited to scientific discussion. Here we make three additional observations: (1) keywords Michie and SPI-B (emerging mid-lockdown) showed increasing presence over time but managed to maintain neutrality. Notably, Michie also attracted a small but sustained quantity of positivity over the full period; (2) psychology (with a stable and lowered presence in the media) shows to maintain a neutral presence over time and (3) behaviour change seems to be largely absent from the narrative, although we see a small surge at the point of lockdown (0; one week after the first surge), possibly aligned with an expected moment in time where many needed to change their behaviour. See Supplementary Material 4 for more detail.

Finally, we note that our primary keywords do not provide insight into the second surge in divisiveness in behavioural science (aside from increased salience without sentimental fluctuation for Michie [+6] over this period), which lead to a qualitative inspection of the category of unnamed actors and resulted in identification of an additional key actor: Prof. Stephen Reicher (Supplementary Keyword: Reicher; see Figure 4). Further attention was paid to this in the qualitative analysis (Study 3).

FIGURE 4

Figure 4. Qualitative inference identified Stephen Reicher as an additional actor. Reicher emerged on the topic of behavioural science toward the latter part of the 24-week time period, corresponding with a surge in salience (+6).

Sentiment Toward Keywords in Context of Public Policy Application

We complement our understanding of sentiment expressed toward keywords by separating sentiments by those opinion contexts that refer to the application of behavioural science in public policy and those that do not. We display sentiments in three panels (see Figure 5): keyword sentiment when policy was not mentioned (top), keyword sentiment when policy was mentioned (middle), and sentiment toward policy application in those same opinion contexts (bottom; see data in Supplementary Materials 7, 8).

FIGURE 5

Figure 5. Sentiment toward “Behavioural Science” separated by sentences that do not (top) and do refer to national policy application (middle), and sentiment toward policy contexts of keywords (bottom) over the 12 two-week time period surrounding the first British national lockdown of 2020 from newspaper articles. Sentiments are represented in counts of positive (+1 or +2), neutral (0), and negative (−2 or −1) bubbles over time, in green, white and red respectively. The area of the bubbles is proportional to the count of sentiments toward the keyword. Reference category in bold.

For behavioural science, we observe similar oscillations over time in all three panels, with two noteworthy differences between panels. First, we note higher neutrality and lower negativity toward behavioural science in opinion contexts which did not mention policy application (62% of neutral and 15% of negative sentiments overall) compared to those which did (52% of neutral and 22% of negative sentiments overall). In both contexts, the proportion of neutral sentiments toward behavioural science increased in the lockdown period (from 47 to 65% in contexts that did not mention policy and from 40 to 54% in contexts that did) and remained the highest post-lockdown. When we compare the sentiments toward behavioural science and its related policy actors in contexts in which both were mentioned (middle, bottom): we observe a much higher (57% overall) proportion of negative sentiments toward the policy actors, increasing across the three time windows (37% pre-lockdown, 42% during lockdown, 82% post-lockdown), paired with a decreasing proportion of neutral sentiments (60–40–15% post-lockdown). In opposition, the proportion of negative sentiment toward behavioural science shows a decreasing trend (from 37% to 19% to 18%). This suggests a transference of negative sentiment from the science of behaviour to the actors who are linked to it in this high-stake policy context over time. In other words, we do not only see a greater proportion of negativity toward behavioural science when mentioned in a policy context than when it is not, but we also see that the majority of this negativity is expressed toward the policy actors, and not behavioural science itself.

What may result in the transference of negativity from behavioural science to the policy makers who use it? For sentiments expressed toward keywords in sentences that do not refer to policy application (Figure 6, top row) we recount two observations. First, for negative sentiment expressed toward behavioural science not in reference to policy, the picture is rather simple: prominent negativity is only observed around the concept of nudge (46% negative sentiments overall). This divisive, negative leaning pattern shows a small but consistent presence over the 24-week period, with a negative flare in the lead up to and throughout lockdown (echoed in articles which do mention public policy). Second, most keywords were more likely to appear in contexts that do not mention policy application (range 60–94% of their occurrences). The exceptions (unsurprisingly) Behavioural Insights Team and Halpern, which appeared in relation to policy actors in 69 and 63% of their occurrences respectively. In opinion contexts where policy was not mentioned, all keywords (aside from nudge) were discussed in neutral opinion contexts most often.

FIGURE 6

Figure 6. Sentiment toward the 10 primary keywords separated by sentences that do (top row) and do not refer to public policy application (middle row), and sentiment toward policy actors of keywords (bottom row) over a 12 two-week time period surrounding the first British national lockdown of 2020 in newspaper articles. Sentiments are represented in counts of positive (+1 or +2), neutral (0), and negative (−2 or −1) bubbles over time, in green, white and red respectively. The area of the bubbles is proportional to the count of sentiments toward the keyword. Reference category in bold. Column (A) behaviour change (concept), (B) behavioural economics (discipline), behavioural economics (discipline), (C) behavioural insights team (named actor), (D) behavioural scientist (unnamed actor), (E) halpern (named actor), (F) michie (named actor), (G) nudge (concept), (H) psychologist (unnamed actor), (I) psychology (discipline), (J) SPI-B (names actor).

For opinion contexts that do mention policy application (Figure 6, middle row) and the sentiment toward policy (bottom row), we see a transference of negativity when the keyword is mentioned alongside policy actors for 9 out of the 10 keywords (just as behavioural science). We also observe two patterns: mentions of the common named actors Behavioural Insights Team, Halpern, and concept nudge share approximately equal numbers of negativity with the paired policy actors, suggesting a level of coupling pre-lockdown (−2, −1). This whilst discussion of actor Michie seemed to avoid negativity nearly entirely at cost to their policy co-mentions, suggesting a level of contrasting pre- (−2, −1), and mid- to end- lockdown (2–5). The latter pattern is echoed over the same time periods by a small but noticeable number of unnamed actors (behavioural scientist and psychologist), suggesting that a group of scientists may be “speaking out” against behavioural science application in policy.

The contrasting narrative offers insight into the drivers of a second surge in behavioural science divisiveness (+6). We observe that psychologist, psychology, and SPI-B collectively maintain neutrality, but share in negative sentiment expressed toward the co-mentioned policy application (bottom) in the post-lockdown period.

Co-occurrence Network Analysis

Finally, we look at which keywords, actors, and concepts frequently co-occur in opinion-context with one another, complemented by five metrics: network density, network clustering, node degree centrality, node betweenness centrality and detected communities (“thematic clusters”) (Figure 7; Table 1; Supplementary Material 6).

FIGURE 7

Figure 7. Networks of keyword co-occurrence across the three time periods: pre-lockdown (A), during lockdown (B), and post-strict lockdown (C). Each node represents a keyword. Edgelines represents the strength of the co-occurrence (Dice coefficient) between two keywords. The node size represents the keyword's weighted betweenness centrality (a small constant has been added to all nodes so that all could be visible in the graph). Nodes are colored according to their community, as detected by the Louvain modularity algorithm.

TABLE 1

Table 1. Structural network statistics on the opinion-context co-occurrence networks pre- during- and post-lockdown.

For the network structure (see Table 1). Over the hard lockdown period we observe larger networks associated with lower network density (pre = 0.14; during = 0.09; post = 0.09) and lower clustering coefficients (pre = 0.18; during = 0.11 post = 0.08) than over the other two time periods. This suggests that the narrative structure around keywords was less cohesive and more diverse in opinion contexts in the two later periods. When considering degree centrality, Michie is the most central keyword across all three time periods; Halpern and behavioural Insights Team are central pre-lockdown but move down the structural importance ranks after that. Showing an opposite trend, SPI-B and behaviour change surge to relevance from the start of lockdown onwards. When it comes to connecting clusters of words (betweenness centrality), named actors michie and behavioural Insights Team play a central role pre-lockdown, but behavioural science and SPI-B serve to connect during- and post- lockdown. Notably, psychology is the only other discipline with structural importance across all three networks.

Interestingly, the non-keywords government(s) and people play an important connecting role pre- and during lockdown, probably reflecting the public policy context in which the keywords are discussed in those periods.

Trends in “thematic clusters” (see Figure 7). Pre-lockdown sees three strong associations: keyword Halpern coupled with Behavioural Insights Team, Michie coupled with psychology, and behavioural science coupled with “government,” “people,” and “advice.” Keyword behavioural Insights Team is connected through Michie via “governments” presumably highlighting the influential roles of the two in public policy. Interestingly, behavioural science forms a separate thematic cluster, suggesting behavioural science was discussed independently of its main national actors over these time periods.

During-lockdown, we observe that the narrative around behavioural Insights Team disconnects from that of other keywords in the opinion contexts. Keyword michie remains relevant and now includes relation to behaviour change, and now connects to keywords behavioural science and SPI-B via “government.” These last two keywords form a unique cluster and are discussed in wider relation to policy responses central to pandemic management (e.g., “measures,” “lockdown,” “behaviours,” “public”). We note two interesting facts: psychology forms a separate thematic cluster from behavioural science; “people” surges to a prominent connector role and forms a cluster with keyword nudge and behavioural scientist, presumably highlighting newspaper's practice to introduce behavioural interventions as relevant to individual and group behaviours (i.e., nudging people).

Post-lockdown, we make three novel observations: michie, SPI-B and behavioural science now solidify as the centers of three highly interconnected clusters. We note the emergence of the new actor “(professor) reicher,” and psychology, “SAGE,” and “governments” working as connectors between the michie-behaviour change's and the behavioural science's clusters.

Discussion

Study 1 maps the discourse of behavioural science around the UK lockdown decision through trends in keywords and sentiment toward them. We find that increased salience can be linked to divisiveness in sentiment, associated with a cluster between Behavioural Insights Team and Halpern (and later also nudge) coupled with policy application of behavioural science in the first (pre-lockdown) wave. This coupling may be a reflection of the embedded relationship between application of behavioural science in governance and the work of BIT. Whilst their collaboration has allowed advancement of applying the science of behavioural science in many public policy areas, one possibility is that the tight relationship was deemed less acceptable under the high-stake policy conditions which were faced.

Nudge, independently of whether it was coupled with policy application of behavioural science, also seems to stir divisiveness. This may be a sticking point for trust and credibility in the public eye which seems, to a degree, to be generalizable (Hagman et al., 2015; Treger, 2020), yet simultaneously does not seem to be of any structural importance to the pre- during- or post-lockdown narrative. Other than that, the application of behavioural science in high-stake policy incurred relatively high negativity in media discourse, but this did not reflect necessarily on the science of behaviour, but rather in reference to its policy counterpart. In relation, two other clusters of associations seem to have been impactful. Key actors such as Michie, SPI-B, and the unnamed psychologist and behavioural scientist contrasted positively to behavioural science application in national-level governance. This suggests that one of the factors to have played into the trust and credibility of behavioural science (and its readiness for policy application) emanated from behavioural science actors themselves speaking out against its potential misuse as a policy tool under the high-stakes circumstances, and this seemed of particular influence a few weeks after the lockdown started to ease.

Finally we note that behavioural science is captured in a separable narrative from the frequent actors (SPI-B, Michie, BIT, Halpern). Over the course of the three time periods behavioural science starts to increasingly function to bridge these actors. This seems indicative of its “catch all” terminology: capturing the versatile and heterogeneous perspectives it represents. In addition, behavioural science offers the important bridge to national level policy applications (with terms such as “lockdown,” “measures,” “restrictions,” and “advisory_group”).

Study 2 Social Media Discourse Analysis

Introduction

In Study 1, we looked at patterns of salience and sentiment toward behavioural science in newspaper articles over the 24-week period surrounding the first UK lockdown of 2020. This analysis does not tell us how the public responded to these articles. To identify whether such stories gained traction on social media, we next identified a set of publicly available Twitter data to track the keywords identified in Study 1. Twitter is amongst the most frequently used social media to investigate public's perceptions across a range of topics (Bibo et al., 2014; Arribas-Bel et al., 2015; Bian et al., 2016; Ordun et al., 2020; Sharma et al., 2020). Twitter is popular for capturing public perceptions with over 330 million registered global users who dynamically generate over 500 million messages (also called “tweets”) per day (Chae, 2015; Mention, 2018). We opted for this (as opposed to another) social media platform for: (1) Twitter's informal, colloquially generated and unconstrained opinion data (Fried et al., 2014; Moe and Schweidel, 2017), (2) Twitter's ability to attract individuals focused on information sharing and seeking (Hughes et al., 2012). We reasoned that mapping the salience and sentiments of the identified behavioural science concepts and actors from Study 1 over the same time period in this dataset, would allow us to identify the nature and extent of concordance of public opinion in line with that expressed in print media.