key: cord-1041106-8nxfb3zv
authors: Etheridge, Ben; Spantig, Lisa
title: The gender gap in mental well-being at the onset of the Covid-19 pandemic: Evidence from the UK()
date: 2022-04-18
journal: Eur Econ Rev
DOI: 10.1016/j.euroecorev.2022.104114
sha: 0ae7b184f90031308880faf4308cb7db41dc5aa7
doc_id: 1041106
cord_uid: 8nxfb3zv

We assess the decline in mental health after the onset of the Covid-19 pandemic in the UK. This decline was more than twice as large for women as for men. We seek to explain this gender gap by exploring gender differences in: family and caring responsibilities; financial and work situation; social engagement; health situation, and health behaviours, including exercise. We assess their quantitative relevance by applying standard decomposition methods. We find that compositional differences in family and caring responsibilities explain part of the gender gap, but more important are gender differences in social factors, particularly changes in loneliness. We explore this result further by analysing gender differences in personality traits. Even after controlling for all factors there remains a noticeable age-gender gradient, with young females suffering particularly badly.

The Covid-19 pandemic caused large disruption to much of the population across the globe, and along many dimensions. This disruption negatively and substantially affected mental well-being (e.g., Adams-Prassl et al. 2022; Banks and Xu 2020; Davillas and A. M. Jones 2021) . A body of empirical evidence now clearly shows that the effects on well-being were felt unequally, with differential outcomes by several socio-economic characteristics such as age (Banks and Xu 2020; Daly et al. 2020; Davillas and A. M. Jones 2021; Zhou et al. 2020) , gender (Banks and Xu 2020; Davillas and A. M. Jones 2021) and ethnicity (Proto and Quintana-Domeque 2021) . In this paper, we focus on the gender gap in well-being that emerged after the onset of the pandemic. Investigating this gender gap is important not only for understanding the effect of pandemics and associated policies for similar future events, but also for understanding gaps in well-being by gender more broadly. We use rich representative data from the UK to explore potential reasons for this differential impact.

To illustrate the large decline in mental well-being, Figure 1 displays averages over time in the UK by gender, including after the start of the pandemic. It shows a large drop after onset, and, consistently with the existing evidence, a disproportionate effect on women, for whom the impact appears to be over twice as large. 1 In our analysis, we first document how well-being of women and men was related to a variety of factors that have been shown to be affected by Covid, such as economic situation, health outcomes and time use within the household. Building on the literature in psychology (see e.g. Holt-Lunstad et al. 2015 , for a recent review), we additionally consider social circumstances such as friendships and loneliness. We then use a coherent framework to shed light on which particular aspect of disruption affected the gender gap in well-being to the largest degree.

We find that social factors, and particularly increases in loneliness, are the most important factor, indicating a key role of social restrictions.

The UK is a particularly suitable setting for the analysis. During its first wave, the UK was one of the countries most affected by the pandemic. At its peak in mid-April, the 7-day moving average of official daily deaths was 950 (14 per million per day), among the highest rates in the world.

Meanwhile, the data we examine were collected only a little afterwards, when the death rate was around 800 per day. 2 At this time the 'lockdown' was in full force, including strict social distancing measures. 3 Indicators of economic activity were sharply negative. 4 At the same time, the main policy tools relating to the economy, such as the UK Job Retention ('furloughing') Scheme, were already well established. 5 that young women were much more negatively affected in general than older individuals. Most strikingly, we document a strong relationship between declines in well-being and social factors that differ by gender. The declines in well-being are particularly large for those who reported an increase in loneliness since their last pre-Covid interview. These correlations are larger for women, and women were more likely than men to report loneliness deteriorations.

Our main empirical exercise is to quantify these contributions using a standard Oaxaca-Blinder decomposition , that is usually applied to assess e.g. gender gaps in wages (Blinder 1973; Oaxaca 1973 ). Here we focus most on 'compositional' effects, which capture differential exposures across women and men. We find that domestic and time use factors play a significant, but relatively small role in explaining the gender gap across the population. However, as would be expected, this factor plays a substantially larger role when we examine those with children only. Health and medical factors also play a role, with a larger fraction of women reporting changes in their caring arrangements. In line with the discussion above, we find that changes in financial situation narrowed the gap, and that the relatively higher number of women reporting improvements in their financial situation is quantitatively relevant. In terms of explaining the gender gap, the most important role is played by the social factors: these explain around a quarter of the gap overall.

The framework we use allows for both dynamic factors that may have changed substantially at the onset of the pandemic, such as work status, as well as pre-existing exposures, such as the presence of children or being a recipient of care. Of course, the dynamic factors in particular may be endogenous to mental health changes themselves. In keeping with the standard use of decomposition methods, therefore, we refrain from giving our results a causal interpretation. Instead they provide 'bottomline' indicators of the likely quantitative importance of different factors (see, for example, Fortin et al. 2011). However, we pursue our explanations further by seeking additional credible evidence from factors that are highly pre-determined. Specifically we focus on personality traits, which are measured around a decade before the pandemic, which have been shown to predict changes in mental well-being during Covid-19 (Proto and Zhang 2021) and which are consistently shown to differ by gender (Feingold 1994; Weisberg et al. 2011 ).

Our results for personality traits illuminate those from our main model. While quantitatively these explain little of the gender gap, composition effects arising from traits are statistically significant and coincide with the results discussed so far. Specifically, differences in extraversion, which is associated with an orientation towards social engagement (McCrae and Costa 1987) explain a portion of the gender gap: Women score more highly in this factor, and extraverts overall suffered larger well-being declines. In fact, the other personality traits together favoured women during the pandemic, such that the gender gap would have been wider if these traits were evenly distributed, although results for these are less significant.

Throughout our analysis we also discuss 'structural' components from the decomposition. These structural effects capture differential responses to the same exposure. For example, we document 3 J o u r n a l P r e -p r o o f Journal Pre-proof that women, but not men, suffered significantly larger well-being declines when working from home.

Unfortunately, in the context of the full decomposition analysis, these structural components are estimated with too little precision to make strong inference. However, focussing on age, we find that the age-gender gradient in well-being declines remains noticeable across all specifications. This indicates that the factors we examine here provide only a partial explanation of why younger women in particular were so badly affected by the pandemic.

Our main analysis focuses on the immediate onset of the pandemic, in April 2020. This allows us to concentrate on the immediate effect of pandemic conditions. In addition, a priori interesting variables such as changes in receiving formal care and exercise are available in the April 2020 data only. However, for a more comprehensive view, we replicate our analysis as far as possible, and show that our conclusions are unchanged when including data from later in the first wave, collected in May and June 2020. In fact, Figure A .1 shows that the gender gap persisted not only during spring 2020, but also re-appeared, equally strongly, in the UK's second wave, during the winter of 2020-21.

We therefore consider this large gender gap as a highly stable feature of the pandemic period that warrants detailed attention. Our analysis goes beyond existing papers examining the gender gap at the onset of the pandemic (Adams-Prassl et al. 2022; Banks and Xu 2020; Davillas and A. M. Jones 2021) . We do this by using rich data both from before and during the pandemic from the UKHLS and UKTUS, that we analyse in a coherent framework, taking into account findings from both the economics and the psychology literature. This is our primary contribution. This paper also contributes to the growing literature on gender inequality during Covid-19 more broadly (e.g. Alon et al. 2020) . While most studies collect real-time data during the onset of the pandemic, they mostly rely on cross-sectional surveys with limited background characteristics of respondents (e.g. Adams-Prassl et al. 2020 , 2022 Andrew et al. 2020; Sevilla and Smith 2020) . To this literature we provide systematic and rich nationally-representative evidence on changes in the exposures faced by men and women. We then quantify their respective effects on well-being through a formal decomposition. Particularly relevant to this literature, we find limited evidence that the gender gap in well-being at the onset of the pandemic was driven by the increase in time spent on childcare, or differential labour-market outcomes.

Going beyond Covid-19, our work contributes to a longer-standing literature on the 'paradox' of declining female happiness relative to males, which can be contrasted with the increasing success of women across a range of economic and social spheres (Astbury 2001; Seedat et al. 2009; Stevenson and Wolfers 2009) . Although not shown in this paper, we similarly find a persistent gender gap in the level of well-being scores across all waves of UKHLS. We contribute to this literature by gathering detailed evidence from the extra 'shock' of Covid-19. Overall, and in the context of this literature, our work is therefore informative about the role of differences in social needs and social engagement in the production of mental well-being across genders. 4 J o u r n a l P r e -p r o o f Journal Pre-proof Finally, our analysis broadly relates to various strands of literature from economics and psychology that examine different correlates of well-being, both before and during the pandemic. We investigate employment-related and financial factors whose linkages to well-being economists have investigated extensively (e.g. Blanchflower and Oswald 2004; A. Clark and Oswald 1994; L. Winkelmann and R. Winkelmann 1998) and find comparable patterns during a large disruption of normal life. We investigate factors related to time use that have received special attention during the pandemic (e.g. Cheng et al. 2021; Xue and McMunn 2021) . We also investigate social factors, making the distinction between perceived social disconnectedness (i.e., loneliness) and social (dis)connectedness in terms of number of friends. 7 We confirm previous results in psychology finding that perceived connections are more important than actual connections for well-being (e.g. Cornwell and Waite 2009; Coyle and Dugan 2012; Taylor et al. 2018) . 8 Not only do we investigate all these factors separately, but we also quantify their respective importance in a representative sample for the gender gap in well-being at the onset of the pandemic. Our results highlight that social factors, in particular loneliness, are more important than economic factors in explaining the gap. 9 2 Data

Our within-pandemic data come from the first wave of the Covid-19 module from the UK Household Longitudinal Survey (UKHLS) (University of Essex 2019). These data come from interviews conducted in the 7 days from Friday April 24, with 75% of responses completed by Sunday April 26.

For information on the pre-pandemic period, we merge these with the '2019 wave' of the survey, a special release designed for Covid research taken from waves 10 and 11 of the parent UKHLS survey (also known as 'Understanding Society'). For additional background information, we also use data from waves 1-10 of the main UKHLS, which has been administered nationwide yearly from 2009. The UKHLS Covid April wave was conducted entirely over the internet. The underlying sampling frame consists of all those who participated in the UKHLS main survey's waves 8 and 9 (conducted over 2016-2018). To adjust our analysis for non-response, we use the survey weights provided. In addition, to allow for the fact that many respondents are related either through primary residence J o u r n a l P r e -p r o o f Journal Pre-proof or through the extended family, we cluster all regressions at the primary sampling unit level. For a further discussion of the Covid module and underlying UKHLS design see Social and Research (2020) , ISER (2020).

The main variable of interest is mental well-being. Our measure is derived from the Likert index that sums 12 questions from the General Health Questionnaire (GHQ-12). The GHQ battery asks questions regarding, for example, the ability to concentrate, loss of sleep and enjoyment of day-today activities. Importantly, the questionnaire asks participants to evaluate their well-being with respect to 'usual' and thus induces a reference point against which respondents evaluate their current feelings. This feature distinguishes our measure from other measures of mental well-being such as the WHO 5-question module (used e.g. in Adams-Prassl et al. 2022) or the PHQ9 depression questionnaire (adopted e.g. in Fetzer et al. 2020 ) that ask about occurrence of specific feelings or behaviors over the last two weeks. While the latter measures have been shown to reflect the cognitive dimension of well-being, our measure captures affective well-being (see e.g. Diener et al. 1985) . The GHQ-12 from this survey has been widely used, both in psychological (e.g. Bridger Each item of the GHQ is answered on a 4-point Likert scale and can be scaled from 0 (least distressed) to 3 (most distressed). The 'Likert score' is obtained by summing these scores across the 12 items to yield a total score between 0 and 36. We standardize this score across all waves and invert it so that, in our analysis, lower scores indicate lower well-being. To remove seasonal effects in mood, we take account of month effects, adjusting all pre-Covid data to 'April equivalents'.

To remove individual factors in reporting style, we use differences of the Covid-module measures from 2019. We treat all the 2019 data as uniformly 'pre-Covid' and, other than by the seasonal adjustment, do not adjust for differences in interview timing.

A key issue for our analysis is the comparability of GHQ scales across gender. A concern is that gender gaps reflect differences in reporting style rather than genuine differences in mental wellbeing. However, several studies investigating various measures of mental well-being conclude that the measured differences reflect true health states and are not driven by gender-related reporting bias (Drapeau et al. 2010; Galenkamp et al. 2018; Oksuzyan et al. 2019; Spitzer and Weber 2019) .

In the context of the UK, Griffith and K. Jones (2019) use data from UKHLS and find that the GHQ exhibits measurement invariance with respect to gender, and so measures the same concept in the same way for both women and men.

We make use of the extensive background information collected in the Covid April wave, as well as the prior UKHLS surveys. In the Covid module, participants were asked a battery of questions about their current experiences. These include questions on employment, on health, on caring 6 J o u r n a l P r e -p r o o f Journal Pre-proof responsibilities, on time use and childcare, as well as self-assessments of financial situation and feelings of loneliness. We mostly focus on changes in these variables from 2019 to April 2020 to examine dynamic impacts, or use the 2019 variables as lags. Most of our change variables are thus calculated as differences of the same variable measured at two points in time. In this, we differ from cross-sectional Covid studies that either compare a variable that is measured once contemporaneously and once with re-call, or rely on respondents directly reporting the differences.

In addition to the 2019 data, we make use of a specific module conducted in wave 9 (i.e., mostly in 2017 and 2018) on social networks. This module contains detailed self-reports on the quantity, intensity and nature of friendships. We also use wave 3 data (from 2011-12) on personality traits, measured with the 10-item Big-5 inventory. Finally, although our main analysis focuses on data from April 2020, we report results using data from May and June 2020 for completeness. We focus on the earlier data in the main paper to examine the impact of the pandemic at onset, and because some of the factors we examine were not measured in these later waves. Parallel versions of all our main tables with all the available data, and showing similar results, are provided in Appendix B.

To provide a graphical illustration of the main variable of focus, Figure Examining within-individual changes (not shown here), we find that about 52 percent of respondents had worse mental well-being in April 2020 than in 2019. Fifty-seven percent of those were women. 12 percent saw no change (of those, 41 percent are women) and 36 percent had better well-being in 2020 (49 percent women). Within the group reporting a deterioration, women faced on average larger well-being drops: On the non-standardized GHQ-12 scale from 1 to 36, women's well-being decreased an average of 5.6 points, compared to an average 4.5 points for men.

One of the main factors that changed during the pandemic was time spent with children and on childcare. Even early on, childcare was identified as a particularly important facet of gender inequalities (see, e.g., Andrew et al. 2020) . While the UKHLS provides direct information on (2015) for more information.

Specifically, we use the following time use item: 'childcare of own household members'. This item is composed of: 'unspecified childcare', 'physical care and supervision', 'teaching the child', 'reading, playing and talking with child', 'accompanying child', 'other specified childcare'. This definition lines up well with the corresponding item in the UKHLS Covid module, which is elicited from the question 'About how many hours did you spend on childcare or home schooling last week?'.

We use these data by first predicting time spent on childcare using a regression which includes full and flexible interactions of the following: gender; number of children under 4 years of age in the household; number of children between 5 and 10 years of age; the presence of a partner, and economic status (working, inactive, dedicated to care). Because all households complete the survey both for a day from the weekend and from mid-week, we use the survey weights to give an average at the individual level across a typical week. The results of the regression are given in appendix 

We begin our formalization of the relationship between changes in mental well-being and explanatory factors using the following regression model:

where F emale i is a binary dummy capturing the gender of individual i. ∆x t indicates a change

in variable x at time t during the pandemic since the pre-Covid baseline (t = 2019), with ∆ghq it therefore denoting the dependent variable of interest. These time subscripts are included for completeness and for ease of exposition even though in our main analysis we estimate the model using t = April 2020 only.

As shown in model (1) we allow changes in mental health to depend not only on gender but a variety of other factors. We first isolate dynamic factors that may have worsened since before the pandemic, given by the vector Z 1 . Examples of relevant variables here might be the presence of Covid symptoms or labour market status. We also consider, in Z 2 , factors for which pre-existing exposure may be important. These may include, for example, the presence of young children in the household. In various specifications Z 2 also includes purely fixed factors such as year of birth, as well as highly durable factors such as personality traits. We allow vectors Z 1 and Z 2 to overlap: mental health changes might depend on pre-existing and continuing childcare obligations, as well as on changes in childcare brought about by school closures. In this model therefore, when Z 1 and Z 2 include all relevant factors, β captures the gender gap which remains after adjusting for composition differences in relevant exposures.

We explore gender differences in mental health further by examining an extension of model (1) as follows: Models (1) and (2) can be combined to provide the standard Oaxaca-Blinder (OB) decomposition framework, most commonly applied to differences in wages across gender or race. For notational convenience, we further combine Z 1 and Z 2 into Z and drop subscripts i and t. Then, letting

x g denote the population average of variable x for group g, the mean difference in mental health changes across genders can be written as:

where, in line with the discussion in the Introduction, the composition effect captures differences in exposures across genders, and the structural effect captures differences in responses to those exposures. In the language of the OB literature, here we use the 'pooled' version of the decomposition, where we effectively weight (or 'price') differences in exposures at the average mental health change across genders. The composition effect can be estimated from model (1) 

We begin our analysis by presenting correlations of the change in subjective well-being at the onset of the pandemic with a variety of background characteristics/circumstances. We present these results separately by gender, to highlight differential responses to the circumstances that men and women face. As such, the Tables we present here correspond to univariate implementations of model (2) above. As discussed, our main analysis focuses on the immediate onset of Covid-19 and thus uses within-pandemic data from April 2020 only. In Appendix B, we provide results for the entire first

Covid wave in the UK, including data from May and June 2020, yielding similar conclusions.

We start with factors that relate to situation within the household. Alon et al. (2020) discuss that the closure of schools and daycare facilities affected women more than men in the U.S. and that the effects on time use were overall stronger than effects relating to employment. For the UK, Andrew et al. (2020) show that mothers in households with two opposite-gender parents bore a disproportionate share of household responsibilities. In line with the framework presented in Section 3 we therefore examine whether changes in well-being are related to changes in time spent on childcare, as well as pre-existing exposure to childcare duties. We also show the effects of changes in time spent doing housework.

Accordingly, Table 1 shows the change in well-being by gender and when individuals are grouped into coarse bins based on their exposure. Importantly, Table 1 includes all respondents, with and without children. Columns 1 and 4 show the correlation of changes in well-being with experienced changes in childcare duties. All of the three groups faced a decline in mental well-being on average in Spring 2020. Women who faced either no change or an increase in childcare duties were more affected than men in the same categories. However, there is no strong evidence that women with increases in childcare duties suffered substantially more than those whose duties changed little or had no children.

We show the proportions of the sample making up each category in Figure 3 , which we use extensively to discuss differential exposure. Relating to the framework discussed in Section 3, this figure therefore captures Z f and Z m , used in equation (4) and later used to explain the gender gap in well-being in the aggregate. The top-left panel of the figure indicates that although more women faced an increase in childcare than men (17 percent compared to 15 percent), these differences are dwarfed by the approximately equal number facing no changes. This is due to the fact that most adults do not have young children. Therefore, to the extent that women faced some impact from J o u r n a l P r e -p r o o f Journal Pre-proof increasing childcare duties, it seems unlikely that these changes alone contribute much to the gender gap overall. Columns 2 and 5 examine the role of pre-existing exposure to childcare, splitting respondents by weekly childcare hours in 2019. Both women and men who reported childcare duties of 6 hours or more per week before the pandemic faced on average the worst well-being at its onset. 11 While women with higher childcare duties suffered more, we also note a pronounced gender imbalance in exposure, with 12 percent of women in the higher childcare group compared to only 7 percent of men (top center panel in Figure 3 ). On these results, it seems that the more important predictor of well-being declines for women was not so much the increase in childcare at the start of lockdown, but the pre-existing exposure to children who required a high level of attention.

J o u r n a l P r e -p r o o f Journal Pre-proof To examine other aspects of time use, Columns 3 and 6 show the relationship between changes in well-being and changes in time spent on house work. Again, most groups faced on average worse mental health in Spring 2020. In terms of the proportions of men and women that fall into each of the three categories, we in fact find results that contrast with those for childcare, suggesting some substitutions in time use: A higher fraction of men reported increases in housework time (54 percent vs 51 percent for women, top right panel in Figure 3 ), with a higher fraction of women reporting decreases. These results indicate that this factor in particular is unlikely to account for the widening gender gap.

Much of the current literature on consequences of the pandemic has focused on economic impacts such as hours worked or facing financial difficulties. Accordingly, Table 2 shows the relationship of changes in well-being by gender and various indicators of economic position. Columns 1 and 4 show mean group effects for a change in a subjective measure asking how well respondents are getting by. 12 We use this measure as a summary of the complex impacts of loss of earnings and other incomes, as well as changes in expenditure patterns induced by the pandemic. Not surprisingly, we see a stronger average decline for those who report a worse subjective financial situation. 13 This effect is not significantly different for women and men. Only for those who see no change or an improvement in their financial situation are well-being changes for women significantly worse. For the majority of respondents, the financial situation does not change (59 percent of women and 60 percent of men; see middle left graph in Figure 3 ). However, it is noteworthy that more women report an improvement in their financial situation (24 percent vs 20 percent of men), indicating that compositional differences in this factor may, if anything, narrow the gender gap. Again, it is worth remembering, however, that this analysis is broad-brush and does not examine particular subgroups for which gender differences in financial outcomes may be different (see Andrew et al. 2020; Benzeval et al. 2020 ).

In Columns 2 and 5 we turn to furloughing and job loss, the latter of which is usually a strong predictor of subjective well-being (Blanchflower and Oswald 2004; A. Clark and Oswald 1994 ; L.

Winkelmann and R. Winkelmann 1998) . Those who lost their jobs fully saw large declines in wellbeing. 14 However, only less than 0.32 percent of the sample falls into this category, which implies that the explanatory power of job loss for the gender gap is likely to be limited. More usually, for roughly 15 percent of the sample, hours were cut or employees were furloughed. For these, the decline in well-being was not significantly different than for those who continued working as previously.

Examining the fraction of women and men who faced these different working circumstances, we do not see a difference between gender (middle centre graph in Figure 3 ). Table reports grouped means of outcome variable, which is the individual change in standardized, seasonally-adjusted and inverted GHQ Likert score. Standard errors clustered at the primary sampling unit and presented in parentheses. The last column presents p-values testing the difference in female vs male means. Covid survey weights used in all computations. Change in finances is based on self-reports of the present financial situation, measured in 2019 and 2020: variable 'finnow'. Change in employment (no change, reduction in hours/furlough, job loss) comes from Covid wave 1. Changes in work from home are calculated based on self-reported work from home patterns for February 2020 and April 2020 (both measured in Covid wave 1). * p < 0.10, ** p < 0.05, *** p < 0.01

To broadly assess the importance of the type of work and the way it can be done, we examine changes in working from home (WFH) patterns in Columns 3 and 6. All groups saw a decline in their well-being. On the one hand, the effect of increases in WFH on mental well-being declines was significantly stronger for women. On the other hand, there were no gender differences in frequencies of these categories: 22 percent of each gender worked from home more often (middle centre graph in Figure 3 ). This factor therefore provides an illustration of different explanations of the gender gap, aligning with the discussion in Section 3. As the frequencies of WFH were similar across gender, the a higher impact of unemployment on women than men. In fact the literature usually finds a higher impact for men (Blanchflower and Oswald 2004) .

Journal Pre-proof role of compositional differences here are necessarily small. However, WFH was clearly associated with structural differences in well-being, with women who worked from home suffering more than men. We return to this discussion in Section 4.2.

One immediate consequence of the pandemic was social distancing induced by the lockdown policy.

We therefore examine the role of social relationships and loneliness, which have been associated with subjective well-being in a predominantly psychological literature (for a review see e.g. J. Cacioppo et al. 2015) . Table 3 shows social factors and their correlation with changes in well-being. For changes in loneliness, we document strong effects. Those who were less lonely saw an increase in well-being, whereas those who were more lonely faced a large negative decline. 15 Women appear to be worse affected than men. Whereas the majority did not face changes in loneliness (61 percent of women and 72 percent of men), we see that more women than men reported an increase in loneliness (20 percent vs 12 percent; see also bottom left graph in Figure 3 ). This finding indicates that changes in loneliness potentially play an important role in explaining the gender gap.

To examine the role of social connectedness in more detail, we make use of a special module conducted in wave 9 that elicits the number of close friends. A priori, it is not clear how the number of friends would relate to subjective well-being during the pandemic and in particular the lockdown.

On the one hand, one might hypothesise that a strong social network can help coping with such a difficult situation, thus leading to a positive correlation between number of friends and well-being changes. On the other hand, related to the above discussion of loneliness, being more connected might lead to increased feelings of loneliness during physical distance measures and lockdowns, as well as indicating a higher dependence on friendships in general. Columns 2 and 5 suggest that the latter explanation applies: individuals with more close friends faced somewhat larger declines in well-being. 16 This pattern is similar for women and men. Interestingly, and similarly to loneliness, the distributions across categories also display significant gender differences: only 7 percent of women fall in the category for whom outcomes were most favourable, reporting zero or one friend, as opposed to 11 percent of men (see bottom middle graph in Figure 3 ).

Pursuing this analysis further, we examine the role of geographic proximity to friends. Even though meeting friends was prohibited at the time we examine, outdoor exercise was allowed, and therefore social interaction with friends nearby was possible with only a minor infringement of the rules. We thus explore how well-being changes relate to the fraction of friends living close by. Those who have all of their friends living nearby saw smaller well-being declines than those with at least some friends living faraway. 17 In fact, composition effects for this variable are in favour of women, with a larger fraction of women appearing in the most advantaged category with all their friends nearby (16 vs. 13 percent; see bottom right graph in Figure 3 ). However, the differences across genders are small and it is unlikely that proximity of friends is important for the gender gap in aggregate.

15 These correlations are in line with previous psychological studies reporting negative correlations between levels of loneliness and levels of mental health. For example, Coyle and Dugan (2012) find that older adults in the US who reported loneliness had a 17 percent higher change of having a mental health problem and Cornwell and Waite (2009) report a 60 percentage point lower likelihood of reporting very good or excellent mental health for those who feel extremely isolated (as compared to those who do not feel isolated). J. T. Cacioppo et al. (2006) use a withinparticipant design and hypnosis to vary feelings of loneliness in a small sample and find higher levels of anxiety for the high loneliness condition.

16 When pooled across genders and including a gender dummy, the relationship between number of friends and the drop in well-being is significant at the 10% level (not shown). See also studies that find a positive correlation of number of friends and various cross-sectional outcomes such as mental health or life satisfaction. See, for example, Helliwell and Huang (2013) , Ho (2016) , and Lima et al. (2017) . 17 The difference between those with all friends nearby and those with only some nearby is also significant at the 10% level when pooling women and men.

We also investigate additional correlations of well-being changes with changes in medical and health factors, health behaviors and key demographics. These are presented in Appendix tables A.2, A.3 and A.4, respectively. Figure A. 2 shows the proportion of women and men in the different categories.

In sum, we find negative average changes in well-being for almost all groups. While patterns are largely similar across both genders, women in any given category are often worse affected.

Regarding medical (health) factors, we see that those who experienced Covid symptoms (11 percent of the sample) also experienced larger declines in well-being, with women more affected than men (see Table A .2). 18 Those who received more external help experienced larger well-being declines than those without a change and women in both categories saw larger declines than men. Table A .3), patterns differ more clearly between women and men. For example, men who exercised more during the pandemic did not face a change in their wellbeing and did significantly better than those who did not change their exercise level or those who exercised less. For women, changes in exercise show a less systematic relationship with well-being changes.

We present correlations with key demographics in Table A .4. Interestingly, only men appear to be negatively affected by the presence of children. Most notable are results by age (in line with Banks and Xu 2020; Davillas and A. M. Jones 2021). Young women (16-29 years old) faced substantially larger well-being declines than any other groups. This is particularly interesting as overall, the young appear to be comparatively more affected by the policy response to Covid than by Covid itself.

We turn now to the decomposition, formalized in Section 3. The results are presented in Table 4 .

The first column includes a full set of controls grouped into the six relevant factors discussed in Section 4.1: domestic and time-use; financial and work; social; health; health behaviours, and basic demographic. 19

The first row of Table 4 shows the estimated raw gender gap, equaling 17.5 percent of a standard deviation in April 2020 compared to baseline. The top main panel of Table 4 then shows the 18 The presence of symptoms and vulnerability only appear in the Covid data. However, we treat these as dynamic variables and impose their absence across the population in the pre-pandemic period. 19 The allocation of controls to these factors is somewhat arbitrary, and, in some cases overlapping: For example, we include working from home status in 'financial and work' rather than 'domestic and time-use', while 'basic demographic' includes age, which is also strongly correlated with domestic and time use factors. This is why we also examine the most salient factors such as age and changes in childcare and house work separately in Table 4 and Table A .5, respectively. The bottom main panel of Table 4 shows the 'structural' (or 'unexplained') effect, also described

in Section 3. This shows the differential responses to the same exposures. Again as discussed in Section 3, this detailed decomposition can be difficult to interpret, and depends crucially on the omitted categories or bases of the included variables. 21 We use natural bases, which are typically 0 or 'same' on change variables, and 0 on variables that take positive values (such as time spent on childcare). Importantly, for age, we take as base those aged 70 and above. The results in the bottom panel therefore correspond to the increase in the size of the gender gap at the means of the relevant variables compared to these bases, with the precise quantitative results showing a net contribution, weighting for distribution and controlling for the other factors.

In the discussion in Section 4.1, we highlighted some noticeable structural differences between men and women. For example Table 2 showed that women experienced relatively worse outcomes when working from home. Unfortunately, in the context of the full decomposition analysis, the precision on the structural effects is typically lower than on the composition effects, and no factors are found to be significant. For this reason, here we combine several of the groups together. The disaggregrated results are shown in Table A .7. The factor showing the strongest influence is age (p-value of 0.18) 20 The importance of social factors is broadly in line with Armbruster and Klotzbuecher (2020) , who, for a different country and with very different data, discuss that the decline in mental health is not driven by financial worries or fear of the disease, but is due to higher levels of loneliness and anxiety.

21 It is sometimes said that this detailed decomposition suffers from an identification problem (see, for example Oaxaca and Ransom 1999) . In fact, as discussed by Fortin et al. (2011) , the challenge is not one of identification, but of interpretation.

J o u r n a l P r e -p r o o f Journal Pre-proof which accounts for over 80% of the unexplained effect. This estimate reflects the fact that the gender gap is far smaller for the over-70s than for younger groups, and shows the important interaction of age with gender (Banks and Xu 2020).

The second column of Table 4 is similar to the first, except the composition part is evaluated at the female effects (the female 'price' in the language of wage decompositions). The results here are virtually identical to those in the first column, with the composition effects only very slightly larger.

All the components are also very similar when evaluated at the male effects (not shown).

The third column of Table 4 focuses on demographic factors only and removes all other controls.

In this simple specification, the positive contribution of age to the structural effect (the interaction of age and gender) is more precisely estimated and significant at the 5-percent level. The point estimate of 0.148 is around 45% larger than that shown in the first column. This estimate indicates that the controls we include, such as time use and financial factors, explain some of this interaction, but not the majority of it.

The final column of Table 4 uses a largely different set of controls and shows the OB decomposition when we use pre-determined factors: the big-5 personality traits. We continue to control for age in this model because personality traits have been shown to follow age profiles (Roberts et al. 2006) . Importantly in our context, several papers show that personality traits differ by gender (Feingold 1994; Weisberg et al. 2011) . Building on the model shown in the first two columns, we focus in particular on extraversion, which is potentially related to social factors, and which we find to be higher in women in our sample. 22 Among composition factors, this final column shows that differences in extraversion explain around 0.8 percentage points of the gender gap. The small size of the effect itself is to be expected: it is a priori unlikely that large gaps in mental health outcomes would be explained substantially by questionnaires filled in a decade previously. Nevertheless we proceed with using the full apparatus of the decomposition analysis for comparison with our other results. Interestingly, when grouping together the other personality traits, they amount to 2.0 percentage points of the gender gap but in the negative direction. In other words, if women had the same levels of the other traits (agreeableness, conscientiousness, neuroticism and openness) as men, then the gender gap in mental health decline would be noticeably larger. For conscientiousness, for example, women score more highly in this trait, which is associated with better mental health outcomes during the pandemic. The right bars of Figure 4 show the equivalent contributions for personality traits. As discussed above, differences in extraversion explain around 5 percent of the gender gap, while the other traits widen the gender gap by around 13 percent. Of the other traits, the most quantitatively important are conscientiousness and neuroticism, in both of which women score more highly, although no trait widens the gap significantly at the 5 percent level.

We finish our discussion by briefly investigating in additional detail the factors that have received the most attention in the literature: domestic and time-use (Andrew et al. 2020) . As illustrated in Figure 4 , we find that these factors have a modest joint effect on the gender gap for the population as a whole. Table A .5 shows the percentage contribution to the overall gender gap for the three sub-factors factors we consider: pre-existing childcare, change in childcare, and change in housework. The results in the top row come from the same specification and sample as in Column 1 of Table 4 .

Although the detailed estimates are not precise, the point estimates are illuminating: In line with the discussion around Table 1 , levels of childcare from 2019 seem to contribute more to the gender gap (5%), than do childcare changes (2%). Changes in house work also contribute a small amount to the gap at around 1.5%. 25 23 Chandola et al. (2020) report that feelings of loneliness are the most important predictor of overall changes in mental well-being (i.e., for both women and men). While they consider a smaller set of correlates and employ a different analysis, our results help interpret their findings. 24 The importance of feeling lonely relative to the number of friends is in line with studies that find a stronger relationship between perceived isolation and mental health than between actual social disconnectedness and mental health (Cornwell and Waite 2009; Coyle and Dugan 2012; Taylor et al. 2018 , all for older adults in the US). These findings also support the point of Golden et al. (2009) and Hughes et al. (2004) that loneliness and objective social isolation are two distinct concepts. Table 4 , as a percentage of the total gender gap. Areas above the axis are positive contributions to explaining the gap, areas below the axis are negative contributions. See Table 4 for more details on underlying calculations.

The bottom row shows the same decomposition, but restricts the sample to those with children only. As expected, the effects are much larger for this subsample and are in line with Andrew et al. (2020) . 26 As shown in the second row of Table A .5, total time use factors now contribute 24% to the gender gap (p-value of 0.02). Again this contribution seems driven mainly by differential exposure to childcare levels, and less to childcare changes. 27

Overall, Table A .5 illustrates the usefulness of representative samples that allow us to investigate effects for different subgroups as well as for the population as a whole. Not surprisingly, the choice of the sample influences the results. With our analysis, we can put findings of e.g. Andrew et al. (2020) , who explicitly focus on mothers and fathers, into a broader perspective: Given that the majority of the population does not have young children, the gender gap in the overall population cannot be explained much by domestic and time use factors.

Using rich longitudinal data from the UK, we assess the decrease in mental well-being at the onset of the pandemic that is particularly apparent for women. In April 2020, this extra gender gap in well-being amounted to 0.18 standard deviations in our representative sample. We explore possible explanations for this gap, examining a wealth of different factors, ranging from pre-existing domestic situation, to changes in work situation, to changes in loneliness and social factors. In our analysis, we distinguish between different circumstances faced by men and women and differential effects of the same circumstances on mental well-being.

We find that women were more exposed to domestic and time use factors that were associated with worse declines in well-being. For parents, these factors explain a noticeable fraction of the gender gap. However, we show that other factors that are more prevalent across the whole of the population played a larger role overall. Specifically, we document important gender differences in social factors, with women reporting substantially more increases in loneliness. Overall, our results suggest that the early impacts of lockdown on mental well-being of women worked less through its effect on time use or labour market, and more through the perceived loss of social interaction.

In terms of policy, our work has implications for strategies to tackle mental health in the later stages and immediate aftermath of the pandemic. We show that combating loneliness, particularly for women, will be paramount. This is especially true if social distancing continues in the medium term, perhaps mandated by policies necessary to tackle even more transmissible variants. However, even 'light touch' policies might be problematic for gender inequality given that women were found to be more cautious in their behavior at the onset of the pandemic (Galasso et al. 2020), consistent with the notion that women are more risk averse in many domains. This factor might imply that women are more likely to adopt behaviors that result in loneliness, even without the explicit policies such as those in place during the period of our study.

In line with the psychology literature that examines factors related to well-being in general, we emphasize the role of perceived isolation for both well-being declines and the gender gap. This finding poses interesting questions for policy and future research. For example, are there ways to reduce face-to-face interactions without increasing loneliness? Can access to new technologies and social media provide a solution?

Although not the focus of our analysis, we note that the gender gap in well-being grew equally large again in early 2021. Gender gaps in well-being are important not only in themselves, but also because they can exacerbate existing gender inequalities in other domains. For example, lower mental well-being can reduce productivity and hence impact current and future earnings, increasing the gender gap in pay. More generally, the age-gender gradient in well-being that we document is another example where preferences and/or needs of large parts of the population might be addressed more quickly or effectively if politicians reflected the diversity of the entire population.

J o u r n a l P r e -p r o o f Notes: Data from UKHLS main waves and Covid module 1. Table presents summary statistics using survey weights. Changes in finances, work from home, loneliness, help, alcohol, exercise and smoking are coded -1 for less, 0 for same and 1 for more in 2020. Employment takes the value of 1 for no change, 2 for a reduction in hours and 3 for job loss. Friends nearby takes the value 1 for all, 2 for some and 3 for none. Children is an indicator for children below the age of 16 in the household. Table reports grouped Table shows results for a regression of time spent performing childcare on explanatory factors. The model is as follows. A two-part linear function for the number of children aged 4 and under (i.e. linear for one child and above) is interacted with gender, with marital status and a three-valued function for economic status: in work or otherwise occupied (such as being disabled), inactive with free time (unemployed, retired, or 'doing something else') or reporting being a dedicated carer (on maternity leave or 'looking after family or home'). A two-part linear function for the number of children aged 5-10 is interacted with the same variables and also included. The R-squared on this regression is 49% with 14, 688 observations. The columns of the Notes: Data from UKHLS 2019, Covid modules 1-3 and UKHLS wave 9. Table reports grouped means of outcome variable, which is the individual change in standardized, seasonally-adjusted and inverted GHQ Likert score. Standard errors clustered at the primary sampling unit and presented in parentheses. The last column presents p-values testing the difference in female vs male means. Covid survey weights used in all computations. Change in loneliness is based on self-reports of the present frequency of feeling lonely, measured in 2019 and 2020. Number of close friends and fraction of friends living nearby are measured in UKHLS wave 9. * p < 0.10, * * p < 0.05, * * * p < 0.01 J o u r n a l P r e -p r o o f Journal Pre-proof replicates Column 2 but with evaluated at female prices. See Table 4 for further notes. * p < 0.10, * * p < 0.05, * * * p < 0.01 Notes: Data from Covid modules 1-3. Table reports grouped means of outcome variable, which is the individual change in standardized, seasonally-adjusted and inverted GHQ Likert score. Standard errors clustered at the primary sampling unit and presented in parentheses. The last column presents p-values testing the difference in female vs male means. Covid survey weights used in all computations. 'Symptoms' comes from self-reported presence of symptoms since the onset of the pandemic. 'Vulnerable' takes value "yes" either if the individual has received an NHS letter requesting they should stay at home ('shielded') or the individual is pregnant. 'Help' is a self-report of whether the individual has received formal care and is measured twice in the Covid module, once for current help and once for help in 2019. This variable only exists in Covid wave 1. * p < 0.10, * * p < 0.05, * * * p < 0.01

J o u r n a l P r e -p r o o f Journal Pre-proof Notes: Data from UKHLS 2019 and Covid modules 1-3. Table reports grouped means of outcome variable, which is the individual change in standardized, seasonally-adjusted and inverted GHQ Likert score. Standard errors clustered at the primary sampling unit and presented in parentheses. The last column presents p-values testing the difference in female vs male means. Covid survey weights used in all computations. * p < 0.10, * * p < 0.05, * * * p < 0.01

The Impact of the Coronavirus Lockdown on Mental Health: Evidence from the United States

The impact of Coivd-19 on gender equality

How are mothers and fathers balancing work and family under lockdown

Lost in lockdown? Covid-19, social distancing, and mental health in Germany

Gender disparities in mental health". Mental health. Ministerial Round Tables

The mental health effects of the first two months of lockdown and social distancing during the Covid-19 pandemic in the UK

The Idiosyncratic Impact of an Aggregate Shock: The Distributional Consequences of COVID-19

Well-being over time in Britain and the USA"

Wage discrimination: reduced form and structural estimates

People skills and the labor-market outcomes of underrepresented groups

Cognitive ability as a moderator of the association between social disadvantage and psychological distress: Evidence from a population-based sample

COVID-19, lockdowns and well-being: Evidence from Google Trends

Daily Suffering: Helpline Calls during the Covid-19 Crisis

The Neuroendocrinology of Social Isolation

Loneliness within a nomological net: An evolutionary perspective

The mental health impact of COVID-19 and lockdown related stressors among adults in the UK

Working parents, financial insecurity, and childcare: mental health in the time of COVID-19 in the UK

Loneliness, social support networks, mood and wellbeing in communitydwelling elderly

Understanding the population structure of the GHQ-12: Methodological considerations in dimensionally complex measurement outcomes

A global panel database of pandemic policies (Oxford COVID-19 Government Response Tracker

Comparing the Happiness Effects of Real and On-Line Friends

Better health with more friends: The role of social captial in producing health

Loneliness and Social Isolation as Risk Factors for Mortality: A Meta-Analytic Review". Perspectives on

Social relationships and health

A short scale for measuring loneliness in large surveys: Results from two population-based studies

Macroeconomic Conditions and Health in Britain: Aggregation, Dynamics and Local Area Heterogeneity

Mapping population mental health concerns related to COVID-19 and the consequences of physical distancing: a Google trends analysis

All you need is facebook friends? Associations between online and face-to-face friendships and health

Validation of the five-factor model of personality across instruments and observers

Male-female wage differentials in urban labor markets

Identification in detailed wage decompositions

Is the story about sensitive women and stoical men true? Gender differences in health after adjustment for reporting behavior

Mental health before and during the COVID-19 pandemic: a longitudinal probability sample survey of the UK population

Mental health responses to the COVID-19 pandemic: a latent class trajectory analysis using longitudinal UK data

Who Got the Brexit Blues? The Effect of Brexit on Subjective Wellbeing in the UK

COVID-19 and mental health deterioration by ethnicity and gender in the UK

Covid-19 and Mental Health of Individuals with Different Personalities

Patterns of mean-level change in personality traits across the life course: a meta-analysis of longitudinal studies

Cross-national associations between gender and mental disorders in the WHO World Mental Health Surveys

Baby steps: The gender division of childcare during the COVID-19 pandemic

Understanding Society: COVID-19 Study

Reporting biases in self-assessed physical and cognitive health status of older Europeans

The paradox of declining female happiness

Social Isolation, Depression, and Psychological Distress Among Older Adults

Narrative economics

Understanding Society: Waves 1-9

On the reciprocal association between loneliness and subjective well-being

Gender differences in personality across the ten aspects of the Big Five

Why are the unemployed so unhappy? Evidence from panel data

Gender differences in unpaid care work and psychological distress in the UK Covid-19 lockdown

Gender inequalities: Changes in income, time use and well-being before and during the UK COVID-19 lockdown

As discussed in section 2 our measure of mental well-being comes from the Likert scale derived from the 12-question GHQ questionnaire. The GHQ questions are listed below. The Likert scale is obtained by recoding so that the scale for individual variables runs from 0 to 3 instead of 1 to 4, and then summing, giving a scale running from 0 (the least distressed) to 36 (the most distressed).The questionnaire is administered to everyone.In our analysis we standardize this variable across gender and wave to have a mean of zero and a standard deviation of one. We then multiply by −1 to obtain a scale that runs from negative (more distressed) to positive (less distressed). 1. More so than usual 2. Same as usual 3. Less so than usual 4. Much less than usual ghqh [GHQ: ability to face problems]: Have you recently been able to face up to problems?1. More so than usual 2. Same as usual 3. Less able than usual 4. Much less able ghqi [GHQ: unhappy or depressed]: Have you recently been feeling unhappy or depressed?1. Not at all 2. No more than usual 3. Rather more than usual 4. Much more than usual ghqj [GHQ: losing confidence]: Have you recently been losing confidence in yourself?1. Not at all 2. No more than usual 3. Rather more than usual 4. Much more than usual ghqk [GHQ: believe worthless]: Have you recently been thinking of yourself as a worthless person?1. Not at all 2. No more than usual 3. Rather more than usual 4. Much more than usual ghql [GHQ: general happiness]: Have you recently been feeling reasonably happy, all things considered?1. More so than usual 2. About the same as usual 3. Less so than usual 4. Much less than usual