Inequality in Health Status During the COVID-19 in the UK: Does the Impact of the Second Lockdown Policy Matter?

3.1 Sample Design

The data come from the UK Household Longitudinal Study (UKHLS) and the UKHLS COVID-19 survey (University of Essex, ISER 2022, 2021). At the time of writing, the UKHLS is a longitudinal household panel study with 11 waves from 2009 until 2020.

Particularly, for pre-pandemic data, I harmonise waves 9, 10 and 11 of the UKHLS by obtaining only individuals aged 16+, living in the UK and responding in the year 2019. The COVID-19 study includes all UKHLS sample designs, except individuals who refused or were unable to participate mentally or physically, and those with unknown postal addresses or living abroad.

The COVID-19 survey has 8 waves and from April 2020 to July 2020 was a monthly web survey, while from September 2020 to March 2021 the survey became bimonthly and only sample members who had completed at least one partial interview in one of the first four web surveys were invited to participate.

To account for unit non-response to the COVID-19 survey, were selected all individuals who responded to both the year 2019 and at least one of the three waves (i.e. November 2020, January 2021 and March 2021) of the COVID-19 sample with non-missing data for the SAH question.^[4]

For the health inequality analysis, all the periods of the COVID-19 survey were considered, obtaining an unbalanced pooled sample with 37,195 individuals (see Table A.2 in the Appendix A), while only November 2020 was used as pandemic period for the policy intervention analysis reported in the Appendix B (see Table A.3 in the Appendix A).

3.2 Dependent Variable

The outcome of interest is the self-assessed health status, a categorical variable taking values between 1 and 5 (=1 poor, =2 fair, =3 good, =4 very good and =5 excellent) to the question “In general, would you say your health is…”. Data on this health outcome are reported in the waves 6, 7 and 8 of the COVID-19 survey, leading us to cover the literature gap of studies investigating this period of analysis.

To reduce the risk of under- or over-estimation due to a low number of observations, while accounting for the non-basic circumstances listed in Table B.1 (Appendix B), the health variable was transformed into three categories (=1 poor and fair, =2 good, =3 very good and excellent) for the policy intervention analysis reported in Appendix B.

To describe the probabilities of moving from one SAH category to another from 2019 to March 2021, the transition matrices between SAH categories in England and Scotland are shown in Table A.4 in the Appendix A. In the matrix, the rows represent the initial values (year 2019) and the columns reflect the final values (March 2021). Comparing England and Scotland, there is a greater likelihood that individuals will maintain the same health status over time, and this probability is higher in Scotland than in England for individuals reporting poor/fair or very good/excellent health status, without substantial differences between gender in England (see Table A.5 in the Appendix A).

3.3 Control Variables

For the inequalities analysis, relevant circumstances were selected following the choice done by Davillas and Jones (2020a): gender (equal to 0 for males, 1 for females), ethnicity (equal to 0 for whites, 1 for others) and parental occupation when the respondent was 14 years old (one categorical variable for each parent). The latter is a relevant circumstance to better understand the impact of socioeconomic status in childhood and is a relevant source of inequality of opportunity in health in several studies (Davillas and Jones 2020b; Rosa Dias 2009, 2010). For each parent, a categorical variable is constructed whose reference category is the unemployed status and which assumes a value of 1 for administrative and elementary occupation, 2 for corporate and managerial status and 3 for missing data. The skill levels of the occupations used are based on the skill level structure of the Standard Occupational Classification (SOC) 2010.

For the policy intervention analysis in Appendix B, the full list of control variables relevant in the COVID-19 pandemic context is displayed in Table B.1 in the Appendix B, while Table B.2 shows a descriptive statistic of all the variables used in this study.

4.1 Overall Health Inequality: Partial Ordering Models and Indices

In the literature, different studies (e.g. Bangham 2019; Helliwell, Huang, and Wang 2019) have considered subjective wellbeing variables (e.g. happiness and life satisfaction) as cardinal rather than ordinal data. As described by Wagstaff, Paci, and Van Doorslaer (1991), several methods (i.e. range method, Lorenz curve and Gini coefficient) use the mean as a reference point for assessing the spread across socioeconomic groups, but with ordinal data the value of the mean is related to the scale used, hence the ordering of distributions according to their means or standard deviations is not robust to variations in the scale utilised (Madden 2010).

To analyse how health status is distributed among the population and how it changes because of policy interventions, I use ordinal rather than cardinal data. The use of objective data for analysing individual health status is not sufficient and data are often not available, hence self-reported health status data should be used.

To compare the distributions of an ordinal outcome (e.g. life satisfaction, self-assessed health status), a series of partial ordering models and indices have been developed to preserve the ordinal nature of data, also considering that there is not an equivalent definition of the “mean” in an ordinal framework.

Whatever the measurement is considered, first-order stochastic dominance works as a tool to compare distributions. However, this dominance is not consistent with the notion of “inequality reduction”, as described by Pigou-Dalton transfers or Hammond transfers (see Fishburn and Vickson 1978; Gravel, Magdalou, and Moyes 2019; Marshall, Olkin, and Arnold 2011), since distributions can cross after such transfers.

To the best of my knowledge, there are no empirical studies that have applied these partial ordering models and indices developed for ordinal outcome to health outcomes in the COVID-19 era, thus the work aims at contributing to this literature gap. In particular, in the work I apply all these partial orderings and indices to measure the overall health inequality by comparing the distributions of general health status between England and Scotland, before and during the pandemic.

4.2 Inequality of Opportunity in Health Status: the ex-ante Parametric Approach

The approaches presented in the previous section provide a comprehensive view of health inequality, focusing on the overall patterns and distributions over time indicating whether health inequality has increased or decreased without underlying drivers of inequality. To identify the latter, an analysis of inequality of health opportunity is required, shifting from broad inequality measures to inequality driven by unfair factors.

To measure inequality of opportunity, most researchers adopt the model defined by Roemer (1998) that considers circumstances and effort (see Supplementary Material B3). Inequality due to different levels of effort is ethically non-offensive (Checchi and Peragine 2010), indeed it leads to different outcomes whose inequality might be desirable. On the other hand, inequality due to circumstances is ethically offensive because these factors cannot be changed by people through effort but still affect their outcomes. Fleurbaey and Peragine (2013) define two approaches applied in the field of equality of opportunity: the ex-ante and the ex-post approaches. The ex-ante approach is the one most commonly used when circumstances are known and individuals have made no effort. The ex-post approach assumes that effort is observed (e.g. lifestyle). Both approaches are equally valid, but in the work, I use an ex-ante approach because effort cannot be estimated due to a lack of variable information.

There are three methodologies to assess ex-ante inequality of opportunity: non-parametric approaches (Carrieri and Jones 2018; Checchi and Peragine 2010), parametric approaches (Bourguignon, Ferreira, and Menéndez 2007; Chávez-Juárez and Soloaga 2015; De Barros et al. 2009; Juárez and Soloaga 2014) and semi-parametric approaches (Li Donni, Rodríguez, and Rosa Dias 2015). The parametric approach requires estimating the average effect of a certain circumstance on the outcome. In the absence of inequality of opportunity, circumstances should not matter and therefore the regression should have a low fit. Equality of opportunity requires that differences in outcomes due to circumstances, but not to effort, need to be eliminated. One shortcoming of the ex-ante approach is related to only lower-bound estimates of inequality of opportunity because the part of inequality due to unobserved circumstances might be attributed to the effort. In the work, following the parametric approaches proposed by De Barros et al. (2009) and Chávez-Juárez and Soloaga (2015) and used for dichotomous variables and ordered variables, I adopt a probit model to estimate the conditional probability h ̃ i of being in a certain category of health status based on the set of circumstances. The detailed description of the models implemented can be found in Supplementary Material B3.

These two parametric methods differ because the one proposed by De Barros et al. (2009) guarantees scale invariance of the inequality of health opportunity measure, while that proposed by Chávez-Juárez and Soloaga (2015) ensures translation invariance.^[5] These two approaches are the most widely used in recent empirical work for dummy variables and allow the researcher to apply, for instance, a non-parametric approach proposed by Checchi and Peragine (2010) by creating dummies for each type and using them as circumstances.

Chávez-Juárez and Soloaga (2015) highlight an important limitation of the scale invariance approach, namely that it does not permit to compare different countries or the same ones over time due to the impossibility of identifying differences due to changes in the average level of health from those due to variations in the link between outcome and circumstances.^[6]

While some existing studies in the literature (i.e. Fajardo-Gonzalez 2016) transform the ordered variable in a dummy variable before applying the parametric approach, this work takes a novel approach by treating the health outcome as an ordered variable. It selects thresholds and constructs a dummy for each, enabling the estimation of two distinct values for every possible threshold. Furthermore, this study advances the analysis by determining dissimilarity indices at each threshold, providing a deeper understanding of how much of the observed health inequalities, both across countries and over time, can be attributed to circumstances within each health status category.

The decomposition approach that can be adopted for measuring inequality of opportunity with ordered variables is the Shapley decomposition, which decomposes inequality of opportunity in a given country into its sources (Chantreuil and Trannoy 2013; Fajardo-Gonzalez 2016; Juárez and Soloaga 2014; Sastre and Trannoy 2002; Shorrocks 2013). In the work, the decomposition is applied to the dissimilarity indices both before and during the response to the pandemic. A detailed description of the decomposition approach adopted could be found in Supplementary Material B3.

5.1 The Overall Health Inequality Before and After the Pandemic

Existing evidence highlight how health inequalities exacerbated during the pandemic due to pre-existing structural disadvantages (e.g. limited access to healthcare, nutritious food and safe living conditions) and pre-existing chronic conditions of the population (e.g. diabetes, hypertension). Analysing health status distributions can reveal differences over time across the UK regions.

Looking at the distributions of SAH in Figure A.1 in the Appendix A, before the pandemic the median is equal to 3, both in Scotland and England. During the pandemic, the sample median changes in both regions to 4, except in March 2021, where the median is 3 for England and 4 for Scotland. In 2019, the relative frequency of individuals reporting being in “excellent” health status in Scotland is higher than in England (19 % vs. 9 %), while during the pandemic the percentage of individuals reporting to be in “excellent” health status remained stable in both regions. Notably, in November 2020 the relative frequency of individuals reporting being in “very good” health status in Scotland has increased significantly (43 %) compared to Scotland in 2019 (26 %) and England in November 2020 (41 %). The relative frequency of individuals reporting to be in “good” health status increased in January 2021 compared to November 2020 in Scotland. There are no significant differences between the two distributions of England and Scotland in January 2021. Finally, in March 2021, the relative frequency of individuals reported being in “very good” health status in England decreased compared to January 2021.

Policymakers are often concerned with identifying polarization or inequality among distributions and over time to provide insights into the effectiveness of interventions. Inequality and polarization are different concepts: increased polarization often leads to challenges in policymaking, even if overall inequality remains unchanged. Then, post-pandemic recovery policies can benefit from understanding which groups experienced the most significant shifts in outcomes in the UK regions. The results start with the dominance checks for a robustness point of view because it is useful to compare the distributions of the SAH to see if they can be ranked unanimously by all indices in a given family with common features. In addition, comparing pre- and post-pandemic distributions using ordinal-sensitive methods allows to assess how the pandemic exacerbate inequalities and to identify which regions or health status categories experienced the most significant changes. As suggested by Jenkins (2020), authors may disagree about the magnitude of differences emerging from different indices within the family, but it is difficult to disagree on the existence or non-existence of dominance.

These results highlight the suitability of the methods in preserving the ordinal nature of the data. Initially, the data are presented through a graphical representation of dominance tests derived from partial ordering models. Subsequently, the polarization and inequality indices are displayed, ensuring consistency with the dominance results and providing a quantitative measure of the magnitude of differences between groups. The inequality indices are especially useful when the dominance tests have not shown a dominance between the pair of distributions compared.

5.1.1 Partial Ordering Models

5.1.1.1 Cumulative Distribution Function

In Figure 2, focusing on the values where the cumulative population share p = 0.5, we can confirm that 3 is the median value of health status for both regions before the pandemic, while it becomes 4 during the pandemic. Recalling that there is “F-dominance” if country x (Scotland) first-order dominates country y (England), then when F _x (z) ≤ F _y (z) for all z (namely all the possible values of self-reported health statuses), either below or above the median. In addition, given two distributions x (Scotland) and y (England) with the same median, recalling that there is “S-dominance” when distribution y dominates x (ySx) because x exhibits greater dispersion away from the median compared to y. Looking at the cumulative distribution functions (CDF hereafter), for overall inequality, there is neither F-dominance nor S-dominance between Scotland and England in 2019, while Scotland first-order dominates England during the pandemic, especially in November 2020 and March 2021 (Figure A.2 in the Appendix A), because F _x (z) ≤ F _y (z) for all z (namely all the possible values of self-reported health statuses), both below and above the median.

Figure 2:

Cumulative distribution functions (CDFs) for self-assessed health status. Notes: Authors’ computation. Own longitudinal weights are used. p is the cumulative proportion of individuals ordered from lowest to highest SAH. Graphs for each period are in Figure A.2 in the Appendix A.

5.1.1.2 Generalised Lorenz Curve

Removing the common-median requirement and considering the individual’s statues as consistent ordinal measurement, I consider the Generalised Lorenz curve to check for unanimous rankings according to the Cowell and Flachaire (2017) peer-inclusive downward-looking definition of status and J index.

Considering that if the Generalised Lorenz curve of x (England) lies anywhere on or below that of y (Scotland), the health inequality is greater in x than in y for all members of the CF family of inequality indices and J index. In Figure 3, for overall inequality in 2019, the GL _x (z) ≥ GL _y (z) for all z (namely all the possible values of individual statuses), hence England is more equal than Scotland. In contrast, during the pandemic there is no unambiguous ranking because the GL curves cross, thus in this case the inequality and polarisation indices are useful to better understand the orderings between regions.

Figure 3:

Generalised Lorenz curve comparisons of individuals’ statues distributions. Notes: Authors’ computation. Own longitudinal weights are used. p is the cumulative proportion of individuals ordered from lowest to highest SAH. Graphs for each period are in Figure A.3 in the Appendix A.

5.1.1.3 H-Dominance

Following Gravel, Magdalou, and Moyes (2021), there is a dual dominance when H+ and H− curves of one country x are nowhere above the corresponding curves of another country y, and F-dominance implies H+-dominance.

Figure 4 shows that for the overall inequality in 2019 there is neither a dual dominance nor H+-dominance, whereas during the pandemic there is no a dual dominance, but there is H+-dominance, indeed the H+ curve displays that Scotland is on or below England especially in November 2020 (Figure A.4 in the Appendix A).

Figure 4:

Checks for self-assessed health status H+ dominance and H− dominance criteria. Notes: Authors’ computation. Own longitudinal weights are used. Graphs for each period are in Figure A.4 in the Appendix A.

According to Jenkins (2021), I obtained that the rankings results of the GL criteria differ form those of the dual H-dominance criteria, thus also in this case the inequality and polarisation indices are useful to better understand the orderings between regions.

5.1.2 Inequality and Polarisation Indices

Inequality and polarisation indices are essential to confirm the dominance result (if exist) and to define the magnitude of differences when there is not a dominance result. However, different indices reflect varying social perspectives on how to evaluate disparities across different segments of the SAH distribution. Therefore, is essential to check the robustness of rankings through a set of indices.

Figure 5 present the point estimates before and during the pandemic for three polarization indices: ANY(1, 1) where the observations in each category above and below the median are equally weighted; ANY(4, 1) that is sensitive to the spread above the median (top-sensitive); and ANY(1, 4) that is sensitive to the spread below the median (bottom-sensitive). In addition, Figure 5 show the point estimates for inequality indices (I(α) for α = 0, 0.9 and J) that are based on the peer-inclusive, downward-looking status concept. The figures also include the 95 % confidence intervals. The latter are derived using bootstrap standard errors with 500 replications and bootstrap weights.^[7]

Figure 5:

Estimates of overall SAH inequality in England (1) and Scotland (0) before (on the left) and after (on the right) the pandemic. Notes: Authors’ computation. Graphs for each pandemic period are in Figure A.5 in the Appendix A.

Looking at the inequality indices (I(α) and J) estimates, Figure 5 (on the left) shows that the overall health inequality is greater in Scotland than in England in 2019. In particular, the I(0) displays that the difference in the health status inequality between Scotland (the most unequal region) and England (the least unequal region) is around 2 %, while the I(0.9) displays that the difference is around 1.65 % and it is similar to J index. The rankings for the three polarization indices, namely ANY(1, 1), the top-sensitive ANY(4, 1) and the bottom-sensitive ANY(1, 4), are similar to those for I(α) and J, also confirming the results of dominance tests, but the confidence intervals display that this dominance is not statistically significant, except for ANY(1, 1). In particular, the difference in polarisation between Scotland and England for ANY(1, 1) is around 9 %, for ANY(4, 1) it is around 14 % and for ANY(1, 4) it is around 8 %.

However, Figure 5 (on the right) displays that the overall health inequality is greater in England than in Scotland during the pandemic. According to I(0), there is no dominance between England and Scotland, the I(0.9) displays that the difference in the health status inequality between England and Scotland is around 1.5 %, and the J index displays that the difference is around 0.60 %. The rankings of polarization indices ANYs are similar to I(α) but with different magnitudes and the differences are not relevant except for ANY(1, 1). In particular, the difference in polarisation between England and Scotland for ANY(1, 1) is around 9 %, for top-sensitive ANY(4, 1) it is around 6 %, and for bottom-sensitive ANY(1, 4) it is around 4 %. In this case, the GL and dual H-dominance tests displayed no dominance, thus these indices help to better understand the orderings between regions.

Summarising, this section shows that the overall health inequality decreases within UK regions during the pandemic, while that between UK regions is highest in Scotland in 2019 and it is highest in England during the pandemic, except in November 2020 when it is higher in Scotland than in England (Figure A.5 in the Appendix A).

This study advances the filed by implementing ordinal-sensitive methods to ensure robust analysis of health inequality during a pandemic crisis, where traditional cardinal measures often fall short. The use of dominance tests alongside inequality and polarization indices provides a comprehensive understanding of regional disparities, also highlighting the differences between the concepts of inequality and polarization that are insufficiently addressed in the existing literature. The indices display a consistent dominance within and between the UK regions, contrasting the analyses conducted by Kobus and Miloś (2012) where the decomposition measure implemented by population subgroup in seven Switzerland regions shows inconsistent results.

The policy intervention analysis reported in the Appendix B shows how the reduction in inequality within UK regions could be linked to the negative impact of COVID-19 policies that worsened the health status of individuals by reducing the distance between SAH categories. In particular, during the pandemic, people in Scotland were more supportive of public health restrictions, the Scottish Government tended to be more cautious than UK government in its approach to handling health restrictions, and Scottish people who became ill were more certain to receive the best treatment available.^[8]

5.2 Inequality of Health Opportunity: The Dissimilarity Index and Shapley Decomposition

This section reports the inequality of opportunity results obtained by estimating the probit model for each threshold (Juárez and Soloaga 2014). In particular, Table A.6 in the Appendix A display the marginal effect of the probit models for England and Scotland, both before and during the pandemic. For both regions, all circumstances have the expected sign and the marginal effects are higher for “good” and “very good” health status. In addition, in all periods, the parental occupation (especially of the mother) influences the probability of reaching a certain level of health status. In England, for “very good” and “excellent” health status, gender and race also become relevant circumstances during the pandemic. In Scotland, for “very good” health status, gender also becomes a relevant circumstance. Figure A.6 in the Appendix A shows the cumulative distribution function of the estimated probabilities computed with the probit model by region and period. The predicted values computed with the probit regression are used for the inequality measure to provide a point estimate of inequality of opportunity (IOp hereinafter). Table 1 displays the computation of the dissimilarity index (pdb) proposed by De Barros et al. (2009), its modified version (ws) proposed by Chávez-Juárez and Soloaga (2015) and the Shapley decomposition of inequality in health opportunity before and during the pandemic. Table A.5 in the Appendix A displays those results by each pandemic period.

The dissimilarity index measures how unevenly health opportunities are distributed across groups defined by circumstances, quantifying the proportion of individuals who would need to change groups to achieve an equal distribution. Compared to the previous models used for overall health inequality measure, the dissimilarity index provides a direct measure of inequality helping identify specific groups most affected by unequal health opportunities.

In Table 1, for each possible threshold of the ordered variable, two estimates are available. The first line provides the estimate for the IOp in the probability of having at least a fair health status, the second and the third lines are for at least good and very good health status, and the last line is for excellent health status.

Looking at the dissimilarity index proposed by De Barros et al. (2009) it gives the impression that inequality of opportunity increases with the level of health status in both England and Scotland, before and during the pandemic. This suggests a focus on reducing the IOp of the highest level of SAH and illustrates well the strong link of this measure to the average level of health status.

The modified dissimilarity index proposed by Chávez-Juárez and Soloaga (2015) provides a different pattern, in which the highest level of inequality of opportunity is found for the middle categories, suggesting that attention should be paid to the latter.