2.1 The impacts of COVID-19 on well-being

An extensive and interdisciplinary literature discusses the negative impact of the COVID-19 pandemic on populations’ well-being. Much of this literature focuses on the direct and indirect consequences of the pandemic—through the social, emotional and economic consequences of lockdowns and health policies—on mental health (Brooks et al. [19], Holmes et al. [20], Blasco-Belled et al. [21], Cooke et al. [22], Rajkumar [23], Xiong et al. [24], Saladino et al. [25], Li and Wang [26], Cao et al. [27]). Salari et al. [28]’s meta-analysis of published studies on mental health in the general population reveal that five studies indicate a prevalence of stress (30%), 17 studies indicate a prevalence of anxiety (31.9%), and 14 studies indicate a prevalence of depression (33.7%) (Please note these rates refer to the first wave of COVID-19 pandemic). Kawohl and Nordt [29] estimated the impact of the expected rise in unemployment during COVID-19 on suicide rates. Using data on expected job losses from the International Labor Organization and a previously developed model about unemployment (from 2000 to 2011) and suicide risk, the authors expect that suicides could increase by about 9570 units per year if the unemployment rate increases from 4.9% to 5.6% (the high scenario). In the low case scenario (unemployment rising to 5.1%), suicides could increase by 2135 units per year. Krendl and Perry [30] found that older adults reported higher depression and greater loneliness following the onset of the pandemic in a sample of American respondents. Sibley et al. [31] documented an increase in anxiety/depression following lockdown and warned about long-term challenges to mental health in New Zealand. Patrick et al. [32] reported marked changes in the mental health of parents and children in the United States. O’Connor and Peroni [33] documented a decline in mental health for nearly a third of residents in Luxembourg. The most important factors associated with the decline in mental health were worsening physical health, income, and a decline in job security.

There are, however, some examples of contrasting results. Sønderskov et al. [34], for instance, documented an increase in the psychological well-being of the Danish population from the first wave (March 31 –April 6, 2020) to the second one (April 22 –April 30, 2020), probably because symptoms of anxiety and depression decreased. Recchi et al. [35] reached a similar conclusion. Using panel data from France (administered at three points in time between the 1^st of April and the 6^th of May 2020), the team found that, in general, self-reported health and well-being improved during the lockdown with respect to the previous year. Although the result hides some heterogeneity within the population (blue-collar workers seem to have suffered more from the crisis), the authors explained their finding, arguing that individuals not affected by the virus judged their situation better than they normally would have. As the authors warn, their findings are based on data from the first six weeks of lockdown in France, and they consider the possibility that the pandemic will affect the population in the long run.

Most studies indicate that well-being decreased in correspondence with the pandemic; however, the mechanisms are much less agreed upon. Rossouw et al. [36] show that lockdown regulations hampered happiness—measured by GNH—in South Africa and argue that the determinants of happiness under lockdown were directly linked to the implemented regulations: lack of access to alcohol (and tobacco), increased social media usage, concerns over future employment, and more time spent at home. Similarly, Greyling et al. [37] found a negative effect of lockdown on GNH in South Africa, New Zealand and Australia. Unobserved factors can confound the findings, however. To address this issue, Foa et al. [38] distinguish the effect of lockdown from that of the pandemic by using weekly data issued from YouGov’s Great Britain Mood Tracker Poll and Google Trends. They found that lockdowns were positively related to subjective well-being and that the main threat to mental health was the severity of the pandemic. The authors also suggested that lockdowns helped relieve the negative impact of the pandemic on well-being by relieving anxiety and stress. Another study, Kivi et al. [39], shows increasing well-being (life satisfaction and reduced loneliness) during the pandemic (among older Swedish residents compared to previous years (2015–2020), but only during the early stage, and also decreasing well-being for those more worried about the pandemic.

With only a few exceptions, previous studies mainly observed well-being at a given point in time. This is because well-being is commonly measured with surveys that take time to administer and elaborate before the data can be published. Only a few studies used Big Data to track the evolution of well-being during the pandemic. As the impact of the pandemic on societies changes over time, it is important to monitor the dynamics of well-being to fully understand their causes and economic, social, and political consequences.

The few available studies on well-being changes during the pandemic reached varied conclusions. Wang et al. [40] found no differences in the Chinese population’s stress, anxiety, and depression levels when comparing two waves (January 31—February 2, to February 28—March 1, 2020). Brülhart and Lalive [41] compared helpline calls to Switzerland’s most popular free helpline during the pandemic to the previous year’s records to infer the impact of COVID-19 on people’s suffering. They concluded that the impact of COVID-19 was negligible as the number of calls has grown in line with the long-run trend. Using GNH data from New Zealand, Rossouw et al. [42] documented a decline in well-being that persisted over time. Sibley et al. [31] reported limited well-being changes during the pandemic’s first stages, possibly due to increased community connectedness offsetting otherwise negative impacts. Brodeur et al. [17] analysed data from Google Trends collected between the 1^st of January 2019 and the 10^th of April 2020 in nine Western European countries and the American States. Although their main aim was to assess the causal link between lockdown and the search intensity for various proxies (of lack) of well-being (terms included boredom, contentment, divorce, impairment, irritability, loneliness, panic, sadness, sleep, stress, suicide, well-being, and worry), they found evidence of mean-reversion in several measures of well-being, but not in all. They concluded that the level of well-being at the beginning of lockdown could be a poor guide to its level later. Cheng and colleagues [43] reached a similar conclusion. The team applied difference-in-difference techniques to monthly data from a longitudinal survey administered on middle-aged and older adults in Singapore. They found large declines in overall life satisfaction and domain satisfaction during the outbreak: the magnitude of the effects is comparable to those of a major health shock or the loss of a beloved person. The figures also indicate that the impact of the COVID-19 pandemic on life satisfaction extended for multiple months but attenuated with time, as life satisfaction trended towards pre-pandemic levels later in the year.

There are various possible explanations for the contrasting results reported above. For instance, previous studies are generally based on data for just one country, often comparing the same country over time or comparing the pandemic to a "normal" year, i.e., 2019. Such analyses hold broad institutional characteristics constant but prevent researchers from comparing these characteristics. Different country contexts, infection rates, and governmental responses are other potential explanations for the heterogeneity of results. This motivates our cross-country investigation over time.

2.2 Use of Big Data to construct happiness indices

Authors from many social sciences have applied Natural Langue Processing to Big Data to address various issues (Eichstaedt et al. [44], Caldarelli et al. [45], Gayo-Avello [46], Bollen et al. [47], Asur and Huberman [48], O’Connor et al. [49]). For instance, tweets have been used to track the rate of influenza in the United Kingdom and the United States (Lampos and Cristianini [50], Culotta [51]). Paul and Dredze [52] found a positive association between public health data and the data issued from sentiment analysis of tweets.

Only a few studies use Big Data to calculate a happiness index. The first measure, The Hedonometer, was created by Dodds and Danforth [53] and their team in 2008. They use the Twitter Decahose Application Programming Interface (API) feed, which is a streaming API feed that continuously sends a sample of roughly 10 per cent of all tweets. This sample allows Dodds and the team to measure happiness levels per day continuously, thus resulting in a time series from 2008 to the present (Dodds et al. [54]). However, the Hedonometer cannot deal with the context in which words are used. Individual words are evaluated, not the overall sentiment of the tweet. For example, a phrase such as "I did not enjoy the holiday" will attract a score of 7.66 for ’enjoy’ and 7.96 for ’holiday’, thus reflecting an overwhelmingly positive sentiment when actually the sentiment is negative. Furthermore, the Hedonometer calculates a happiness index on a scale of 1 (sad) to 9 (happy), but it cannot detect the emotions underpinning the words or the tweets. Thus, it cannot determine if the changes in happiness levels are due to more or less negative emotions such as fear or anger or positive emotions such as joy.

The second known measure was developed in 2012 by Ceron et al. [55]. They used an Integrated Sentiment Analysis (a human-supervised machine learning method) on Big Data extracted from Twitter for both Italy and Japan. They created a composite index of subjective and perceived well-being that captures various aspects and dimensions of individual and collective life (Iacus et al. [56]). Up until 2017, the researchers extracted and classified 240 million tweets over 24 quarters. They applied a new human-supervised sentiment analysis to analyse the sentiment and did not rely on lexicons or special semantic rules.

3.1 Data: Applying natural language processing to Big Data

In the analyses, we use a cross-country panel dataset with high-frequency daily data (see section 3.2). We analyse the time period from 1 January 2020 to 31 December 2020 (366 days since 2020 was a leap year) across ten countries, namely Australia, Belgium, France, Great Britain, Germany, Italy, Luxembourg, New Zealand, South Africa and Spain.

To derive our data, we use the Twitter API (v1) to extract and harvest all original tweets within a geographic bounding box that corresponds with the country in question. After translating all non-English tweets into English, using Microsoft Azure and Google Translate, we performed necessary pre-processing to clean up the punctuation and remove @, #, the letters “https”, control characters, digits, and emojis. We removed emojis because, in further analyses, we use only ASCII characters. We then use Natural Language Processing (NLP) in the form of sentiment and emotion analysis to score each tweet’s underlying sentiment and emotions. Stated briefly, sentiment analysis is an automated process to determine the feelings and attitudes of the author of a written text (or tweet) (Hailong et al. [57]). The time series were then based on the average daily coded values described below.

Applying NLP to Big Data sources has some advantages over traditional tools, as discussed in the Introduction. One of the advantages of Twitter data is that they are abundant, and users are heterogeneous. Twitter accounts include individuals, groups of individuals, organisations, and media outlets, thus providing the moods of a vast blend of users, which is not found in survey data. Another advantage is that they can provide timely and internationally comparable data on individuals’ well-being, sentiments, and behaviours. The data also does not suffer from non-response bias (Callegaro and Yang [58]). Sentiment analysis applied to social media data and location history allows researchers to "listen" and observe what people deem important in their lives. Big data can be used by decision-makers, especially in situations requiring rapid decisions or in the presence of incomplete information. However, using Big Data implies a trade-off between timeliness and accuracy. Big data sources may face large measurement errors and limited representativeness. It is also difficult to establish their convergent validity as there are very few validated data sources to compare.

3.1.1 The dependent variable: The Gross National Happiness index

To measure happiness (the dependent variable), we make use of the GNH, which was launched in April 2019 as part of the Gross National Happiness.today project (Greyling et al. [15]). GNH measures the evaluative mood of a country’s citizens over time. As a measure of mood, the GNH captures the more volatile part of well-being, commonly referred to as happiness (Diener et al. [59]). (Please refer to Rossouw and Greyling [60] for country examples where the GNH accurately captures the mood of a nation). However, the evaluative qualification indicates tweets reflect individuals’ conscious decisions; in other words, they evaluate what they want to say.

The GNH index is constructed using sentiment analysis. After completing the pre-processing of the tweets, each tweet was coded as having either a positive, neutral or negative sentiment primarily using the lexicons Sentiment140 and NRC (National Research Council of Canada Emotion Lexicon developed by Turney and Mohammad [61]). For a full discussion on using sentiment and emotion analyses to derive variables, see Greyling and Rossouw [62]. Then, if a tweet is positive, it is coded as 0, if neutral 2 and if negative 4. The number of positive and negative tweets are used in a sentiment-balance algorithm (a balance score) and scaled to derive a GNH happiness score per hour that ranges from 0 to 10, with higher values indicating higher happiness. To generate daily data, the mean GNH per day is calculated.

3.1.2 Emotion variables

Emotions are derived using emotion analysis operationalized using the NRC lexicon. It distinguishes between eight basic emotions: anger, fear, anticipation, trust, surprise, sadness, joy and disgust (the so-called Plutchik [63] wheel of emotions). NRC codes words with different values, ranging from 0 (low) to 10 (high), to express the intensity of an emotion or sentiment. See Greyling et al. [37] and Greyling and Rossouw [62] for previous examples using these eight emotions.

Table 1 provides three examples of how emotions are extracted. Each tweet is attributed a score according to the presence and intensity of one or more emotions. For example, the tweet “Mask-wearing is really reducing in inner Auckland–I’ve been virtually the only one I’ve seen today. (Lack of) distancing pretty much the same….#COVID19NZ “; resulted in a score of 6 being assigned to the emotion called ‘anger’, a score of 2 to the emotion called ‘anticipation’, a score of 2 to the emotion called ‘surprise’ and a score of 4 to the emotion called ‘disgust’. The daily score of a given emotion corresponds to the average of the scores that emotion received on a given day. As an example, Table 1‘s tweets generate a score of 3.3 for ‘anger’, i.e. (0 + 4 + 6)/3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Examples of coding tweets for emotions.

https://doi.org/10.1371/journal.pone.0275028.t001

To develop our context-specific variables, we extracted three subsets of tweets based on keywords relating to the economic situation, institutions, and loneliness. Emotion analysis using the same process and NRC lexicon was then conducted on these subsets to generate the specific variables:

1. To construct the ’fear in relation to the economic situation’ variable, we first extract all tweets that include the following keywords: jobs, economy, saving, work, wages, income, inflation, stock market, investment, unemployment, unemployed, employment rate, tech start-up, venture capital. We then use the NRC lexicon to return the fear emotion score for each tweet in this subset.
2. To construct the ’sadness in relation to loneliness’ variable, we first extract all tweets that include the following keywords: lonely, loneliness, alone, isolation, abandoned, social distancing, lonesome, by oneself, solitary, outcast, companionless, solitary, homesick. We use the same method to construct the time-series as for economic fear, but now we use the emotion score for ’sadness’ and the loneliness subset.
3. To construct the ‘trust in national institutions’ variable, we first extract all tweets that include the following keywords: parliament, ministry, minister, senator, MPs, legislator, political, politics, prime minister. We use the same method as above, but this time we use the emotion score for ’trust’ and the national institutions subset.

3.1.3 Validity of GNH and the emotion variables

It is vital to assess whether GNH reliably measures the aggregate happiness of a nation. The validity of the GNH can be assessed in several ways. We assume that content validity, i.e., the ability of the GNH to represent happiness correctly, is satisfied. The reason is that the algorithm for sentiment analysis is logically built to measure the affection content of a tweet. Sensitivity to variations in the algorithms used, volumes of tweets, and the sample period are also assessed, with results presented in S1 Appendix. We test the convergent validity of the GNH by checking whether GNH significantly correlates with external data that are known to represent the same or similar concepts.

To assess the external convergent validity, we proceed as follows: firstly, we compare cross-sectional country rankings produced from the GNH to those available from alternative measures of well-being, such as the Eurobarometer’s life satisfaction and the World Happiness Report (Helliwell et al. [64]). Secondly, we analyse the time-series correlation between GNH changes and changes in consumer confidence and other well-being indicators available for the same period from Google Trends. This exercise provides encouraging results: GNH-based country rankings are not significantly different from those based on alternative indicators. Results from time-series correlations are mixed: the correlation of GNH to well-being from survey data for four European countries is uncertain. At the same time, it performs relatively well in relation to the index of negative emotions and consumer confidence. The validation results pertaining to external data are presented in S2 Appendix. Similar validity tests of generalised trust, trust in national institutions, and economic fear are contained in S5 Appendix. Indeed, they show significant relations that support the overall process by which the Twitter-derived variables are generated.

In addition to our tests, previous literature suggests that the GNH index correctly reflects the evaluative mood of a nation (i.e., supporting construct validity). Greyling et al. [37] showed a negative and statistically significant association between the GNH index and ’depression’ and ’anxiety’ in Australia, New Zealand and South Africa. Moreover, data indicate that variations in the GNH index correctly reflect various events, including the COVID-19 pandemic. Data from South Africa show that the GNH dropped well below previous daily averages following the outbreak of COVID-19. Later, when distancing regulations were implemented, the GNH recovered slightly but remained lower than normal (Greyling et al. [37]). In a different field, Lampos and Cristianini [50] and Culotta [51] showed that information extracted from tweets could track health variables.

Additional concern may arise from assigning tweets to specific countries because that limits the number of usable tweets as not all tweets are geo-located. However, a large number of tweets are extracted per country, and the number of users represents significant proportions of the countries’ populations. For example, the number of extracted tweets per day ranges from 257 in Luxembourg to 4,600 in New Zealand and as many as 93,500 tweets in the United Kingdom. The smaller number of tweets in Luxembourg is offset by the smaller population and the analytical techniques used to assess GNH. For instance, we generally focus on longer periods, e.g., weeks, or use rolling averages to smooth daily fluctuations. We also conducted several internal consistency tests (reported in S1 Appendix) to ensure the GNH is not too sensitive to changes in the volume of tweets.

In summary, previous studies and the validation exercises from this study provide initial evidence suggesting that the GNH provides a meaningful measure of happiness. Tests of three emotion variables also support the use of Twitter-derived variables. S5 Appendix gives details on the validity of these measures (Figs S17 and S18 in S1 Appendix). Descriptive statistics of the GNH are presented in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Descriptive statistics by country.

Average values over the year 2020.

https://doi.org/10.1371/journal.pone.0275028.t002

3.2 Additional covariates

We integrate the variables derived from Twitter data with additional information to account for each country’s evolution of the pandemic. These include:

1. Daily new cases: Data on COVID-19 are sourced from Our World in Data (Roser et al. [65]). Among available series, we retain the number of new confirmed cases per day per million in population. We adjust by population to account for countries’ sizes. In much of the analyses, we further transform new cases using an inverse hyperbolic sine transformation, which is roughly equivalent to a log transformation but is identified for zeros. Please note that the European countries’ data are from the ECDC (European Centre for Disease Prevention and Control). Available series also include the number of tests performed, deaths and hospitalisations.
2. Policy responses: The indicator of policy stringency is the Containment and Health Index from the University of Oxford’s COVID-19 Government Response Tracker (Hale et al. [66]). The tracker includes multiple indices summarising 18 indicators of policy response to the COVID-19 pandemic in different dimensions. The Containment Index, also known as the stringency index, is based on the following nine indicators: school closing, workplace closing, cancel events, restrictions of gathering, close public transport, stay at home requirements, restrictions on internal movement, international travel controls, and public information campaigns. Details on the construction of the index and the underlying indicators are available online (www.bsg.ox.ac.uk/covidtracker).). The data and methodology are frequently updated as the pandemic evolves.
3. Behavioural responses (distancing): We use Google Mobility Reports (Google [67]) to measure distancing and account for behavioural responses to the pandemic and government policies. Google Mobility Reports provide daily aggregate mobility/visitation data by geographic location. The data, collected from users’ devices that have opted-in to location history on their Google account, permit the observation of mobility trends across several types of locations: retail and recreation, groceries and pharmacies, parks, transport hubs, workplaces, and residential. The measure of distancing we consider is an index reflecting the time people spend at home. Our choice is motivated by requiring fewer assumptions about people’s movements during the pandemic. The figures are compiled as relative movements (visits’ numbers) compared to the number of visits during the baseline period, 3 January to 6 February. Mobility is also normalised for each day of the week. Seven baseline days are used, corresponding to the median values observed during the five-week baseline period. For this reason, we cannot compare daily movements. We instead use weekly average values or daily data smoothed using a seven-day centered moving average.
4. Economic conditions (unemployment rate): The monthly unemployment rate is made available for the European countries by Eurostat (Eurostat, [68]). We use the raw, not seasonally adjusted series.

Table 2 provides summary statistics for the non-Twitter variables described in points 1–4 above, as well as information on the GNH.

5.1 Descriptive analysis

Before turning to the regression analysis results, we describe our variables’ changes during the year 2020. For brevity, we group countries into two groups: the seven European countries (EU hereafter), and Australia, New Zealand and South Africa (A-NZ-SA hereafter). Please note detailed variables’ evolution for each country is presented in S3 Appendix.

Fig 1 depicts the time series of the GNH (solid line) for the EU countries (panel 1a) and A-NZ-SA (panel 1b) for 2020. Noticeably, in the EU, the GNH dipped at the onset of the first and second waves of the pandemic. The fall in the GNH is also apparent for A-NZ-SA at the beginning of the first wave. The figure also illustrates the dynamics of GNH throughout the year and how measurements based on a single point in time could misrepresent the dynamics (e.g., mid-March or mid-May).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Gross National Happiness, economic fear, trust in national institutions, and loneliness in 2020.

Note: Data are presented using seven-day (centered) moving averages. Figures for Luxembourg are missing due to the scarcity of tweets using the selected keywords. Source: Data are all sourced from the project “Preferences Through Twitter” with the support of FNR, UJ and AUT. a) Average daily data across six European countries. b) Average daily data across Australia, New Zealand and South Africa.

https://doi.org/10.1371/journal.pone.0275028.g001

Fig 1 also shows the time series data for economic fear, loneliness and trust in national institutions. Please note that the number of European countries decreases to six because figures are unavailable for Luxembourg.

One can see from Fig 1 that A-NZ-SA experienced higher levels of economic fear, sadness linked to loneliness, and trust in national institutions than the EU. Trust and sadness exhibit higher volatility in A-NZ-SA than in the EU. We notice a marked decrease in trust in national institutions in correspondence with the first severe coronavirus outbreak (March), followed by a slow recovery in both groups of countries. At the same time, loneliness increased, with a further increase in EU countries during the second wave. Economic fear increased during the first wave. Afterwards, it remained at levels higher than the initial one in the EU countries. At the same time, it decreased slightly but steadily in A-NZ-SA (S15 Fig in S3 Appendix provides detailed trends for each variable for each country). The limited number of tweets produced in Luxembourg did not allow us to compute trust in national institutions, sadness in relation to loneliness, and economic fear.

Fig 2 contrasts the respective evolutions of average GNH (solid line) and COVID-19 infections (dashed line) for the EU (panel 2a) and A-NZ-SA (panel 2b). Please note that infections are the average number of daily new confirmed positive cases per million. Note also that mass testing was not performed during the first pandemic wave.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Gross National Happiness and new COVID-19 cases per day in 2020.

Note: GNH and new cases are presented using seven-day (centered) moving averages. Source: GNH data (Greyling et al. [15]) are sourced from the project “Preferences Through Twitter” with the support of FNR, UJ and AUT. The number of new positive cases is sourced from OurWorldinData.org. A) Average daily data across seven European countries. B) Average daily data across Australia, New Zealand and South Africa.

https://doi.org/10.1371/journal.pone.0275028.g002

Panel 2(a) shows that GNH dipped in correspondence to the two pandemic peaks of March and November 2020. In Europe, GNH dropped suddenly (-8.6%) during the first wave, recovering quickly afterwards (+9.84%). In correspondence to the slow but steady increase in the number of cases during the late European summer-autumn, GNH showed a steady decline culminating with a sharp fall at the beginning of November, when infections reached a second peak.

Panel 2(b) shows a similar pattern for A-NZ-SA’s GNH during the first peak (GNH suddenly dropped by 9.33%). The evolution of GNH declined slightly during the emergence of the second pandemic wave (May-July) and recovered afterwards. We also observe that the number of new positive cases has been substantially lower than those recorded in the EU in this group of countries. Changes in the GNH in 2020 are more volatile in the EU than in A-NZ-SA (S4 Table and S16 Fig in S4 Appendix provide average scores of the GNH by sub-periods, while the changes of GNH for each country are shown in S13 Fig in S3 Appendix). Hence, while the average GNH is about 7, our daily data reveal a varied picture in terms of intensity and duration of the shock and across waves of infection.

However, other variables may influence GNH besides the number of infections. Containment policies, economic conditions, trust in others and institutions may have affected the overall well-being, both directly and through their impact on the pandemic. We account for the joint effect of these variables using regression analysis in section 4.2. In the remainder of this section, we briefly describe changes in the GNH in relation to changes in containment policies and trust in others. These variables are relevant for their direct effects on well-being and because of their effect on the pandemic: containment policies limited the spread of COVID-19 (Fong et al. [71], Chinazzi et al. [72]), thus benefiting well-being. Trust in others promoted cooperation and solidarity with positive spillovers on compliance and well-being (Bargain and Aminjonov [73]).

Fig 3 reports the changes in the GNH (solid line) and those of containment policies (dashed line). We notice the jump in policy stringency and a drop in the GNH, which occurred at the pandemic’s onset. After this, the months of May–July were characterised by a gradual relaxation of containment policies. Then, policies in the two groups of countries took different directions. In A-NZ-SA, increases in stringency during July-September were followed by a marked relaxation of policies. The summer coincided with a relaxation of policies in EU countries, down to a degree of stringency maintained throughout October. Stringency jumped again from about 50 to nearly 80 points during October, accompanied by the sharp fall in the GNH. However, the decline in the GNH had begun earlier. In A-NZ-SA, the increase in stringency saw a dip in GNH, but also, in this case, the latter’s decline had started before.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Gross National Happiness and containment policies.

Average daily data across ten countries. Note: GNH is presented using seven-day (centered) moving averages. Source: GNH data (Greyling et al. [15]) are sourced from the project "Preferences Through Twitter" with the support of FNR, UJ and AUT. The policy index is sourced from Oxford Policy Tracker. a) Average daily data across seven European countries. b) Average daily data across Australia, New Zealand and South Africa.

https://doi.org/10.1371/journal.pone.0275028.g003

It is worth noting that countries’ policy responses to the pandemic have been widely heterogeneous, both within and between the two groups of countries. S14 Fig in S3 Appendix depicts changes in the GNH, the number of new positive cases and the level of containment policies separately for each country. The graphs show that policy response in the EU shares a similar pattern across countries–after the initial shock, stringency in containment policies increased with a rise in new positive cases. In contrast, countries in the Southern hemisphere saw nearly zero new infections, contrasted with heterogeneous governments’ responses. Australia maintained strict containment policies throughout 2020, whereas South Africa gradually relaxed the measures. In New Zealand, the initial suppression strategy was followed by a sharp drop in stringency, interrupted only by a sharp increase in stringency during August-September.

The levels of and changes in trust differ in the two groups of countries (see Fig 4). A-NZ-SA report, on average, higher trust in others than EU countries (respectively, 7.6 and 6.2 throughout 2020). Trust increased in both groups of countries in the first half of the year, markedly in A-NZ-SA (on average, 8.33% in European countries and 13.7% in A-NZ-SA). The upward trend was shortly interrupted in correspondence with the first outbreak of COVID-19. Subsequently, trust started declining (with an initial sudden drop of about 6%), corresponding to a renewed growth in infections (from June onward), with different patterns. In A-NZ-SA, it levelled off before decreasing again in September-October and started increasing again in November-December. In the EU, trust declined (nearly 7.7%) from June to September and exhibited an increasing trend afterwards. S15 Fig in S3 Appendix shows changes in trust for each country in the study. Generalised trust correlates at 81% (statistically significant at 10%) with survey-based measures of trust. See S19 Fig in S5 Appendix and the notes therein for more details.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Gross National Happiness, broad trust and new positive cases of COVID-19 in 2020.

Note: Data are presented using seven-day (centered) moving averages. Source: GNH data (Greyling et al. [15]) and generalised trust are sourced from the project "Preferences Through Twitter" with the support of FNR, UJ and AUT. The number of new positive cases of COVID-19 is sourced from OurWorldinData.org. a) Average daily data across seven European countries. b) Average daily data across Australia, New Zealand and South Africa.

https://doi.org/10.1371/journal.pone.0275028.g004

In summary, the daily evolution of GNH reveals considerable variations in well-being responses during 2020. We found a clear indication that people suffer when the infection worsens and that the recovery takes longer than the decline. A descriptive examination of the changes in the number of new cases, containment policies, and generalised trust indicates a mixed relationship with changes in the GNH. This suggests that multiple factors should be considered jointly to explain the changes in well-being during the pandemic.

The next section explores the joint effects of multiple variables on the GNH using regression analysis.

5.2 Regression results

Following the methodology, as explained in section 4, we present the first set of results in Table 3. Regressors are included step-wise to check their association with GNH before and after controlling for additional variables. Coefficients on dummies for the months of the year are omitted for brevity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Association between the number of positive cases, physical distancing, expected increase and decrease of policy stringency, and GNH.

https://doi.org/10.1371/journal.pone.0275028.t003

The auto-regressive term, the lagged value of GNH, reported in the first row, has a high and significant coefficient, indicating that the current GNH variation depends largely on its previous realisations. Lagged GNH, along with the month and season controls, explains nearly 84% of the overall variance. This indicates that the GNH is highly persistent and not greatly affected by volatile events. Auto-correlation is common in time series. In this case, it also represents a desirable feature in a measure of well-being: "in a world of bread and circuses" (Deaton [74]), the GNH seems to capture the bread more than the circuses.

Dummies for Spring and Fall have positive and significant (10 per cent) coefficients, indicating that the GNH tends to be, on average higher during those seasons. The coefficient for Summer is not statistically different from zero. Models 2 to 4 add, respectively, new cases, the index of residential mobility, and the expected changes in containment policy to the baseline (model 1). We find a negative and significant coefficient for new cases and both expected increase and decrease of containment policies. The index of residential mobility attracts a positive and statistically significant coefficient, indicating that, ceteris paribus, staying at home is positively related to GNH. Notice that the number of observations reduces by 50 when we control for mobility. The reason is that Google Mobility data are available from the beginning of the pandemic (mid-February); thus, the initial five weeks of 2020 are missing. Model 6 shows that the results of the complete model (column 5) do not depend on the smaller sample size due to the inclusion of the control for residential mobility.

Results do not change when we simultaneously include all of the controls (Model 5). The adjusted R-squared indicates that the full model explains 85% of the total variance, slightly improving the baseline. All coefficients maintain their sign, magnitude and statistical significance.

Table 4 presents results for our extended model, which includes the emotions, controls for economic fear, trust in national institutions and loneliness. For ease of comparison, the first column reports the same specification as column 5 in Table 3. Also, in this case, we included variables step-wise. We applied a step-wise selection process for the emotions to minimise the number of controls and preserve degrees of freedom. We kept the emotions with a p-value (after Wild Cluster Bootstrap) smaller than 0.4. After this selection process, only disgust, fear, surprise and trust are retained (column 3). Months and season controls are included in the estimates but omitted from the table.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Regressions of GNH on the complete set of control variables.

Average weekly values by country.

https://doi.org/10.1371/journal.pone.0275028.t004

The full model’s adjusted R squared indicates that 92.7% of the total variance is explained, an improvement over the initial model of nearly seven percentage points, despite the decrease in the number of countries. Luxembourg is excluded from the analysis because tweets about economic conditions, national institutions, and loneliness are scarce.

The auto-correlation term still explains a large part of this variability, but its coefficient decreased from 0.922 to 0.776. The changes in the number of new infections and the expected increase of containment policies maintain their negative sign, magnitude and significance. The expected decrease in the policy stringency index is no longer statistically significant. Ceteris paribus, an increase in the number of people staying at home, remains positively and significantly associated with GNH changes.

GNH also increases when trust and surprise increase and decreases when disgust increases. These results are not surprising: it is well established that trust correlates with well-being both cross-sectionally and over time. Surprise and disgust are emotions that change with mood and correlate with more volatile aspects of well-being, sometimes referred to as affective or momentary subjective well-being. Economic fear, trust in national institutions, and sadness about loneliness are not significant. The reasons for this result are not clear.

In a further specification, we split changes in new positive cases into two variables: increases and decreases in new cases. This allows us to study the symmetry of the effect of contagion on the GNH (results are omitted for reasons of space and can be made available upon request to the authors). We found that an increase in new cases correlates negatively and significantly with the GNH, whereas a decrease does not attract a statistically significant coefficient. This suggests that a worsening pandemic affects the GNH much more than an improvement.

If we compare model 3, which includes Luxembourg, with model 5, we observe that the exclusion of Luxembourg increases our ability to explain the overall variance by 4.8 percentage points (from 88% to 92.8%). This is probably because the time series for Luxembourg are more volatile than those of the other countries. The exclusion of Luxembourg does not qualitatively change the estimated relations.

Model 4 is identical to model 3, except that it excludes the autoregressive term and month and seasonal controls. It is intended to check the robustness of results when increasing the degrees of freedom and removing the strong influence of lagged GNH. As expected, the adjusted R squared decreases considerably, from 88% to 52.7%. The coefficients of the variables of interest maintain their signs, but most of them lose significance: only the change in the number of new positive cases and surprise remain statistically significant.

In sum, our final model seems to explain the GNH changes during 2020 in the studied countries rather well. A large part of the overall variation is explained by the autoregressive term—which is to be expected and signals that the GNH is rather stable throughout the weeks of the year—as well as the month and seasonal controls. GNH decreases when the virus spreads, particularly when new positive cases increase and containment measures become more stringent. Under these circumstances, an increase in people staying at home is associated with an increase in the GNH. This is probably because people feel safe if they stay at home in a difficult or dangerous situation. Disgust and the GNH are negatively associated, whereas we find a positive association with surprise and trust.