Introduction

The UK Department of Health defines a long-term condition as “a condition that cannot, at present, be cured but is controlled by medication and/or other treatment/therapies” [1]. Large numbers of people in the UK live with a long-term condition, for example it is estimated that 20.3 million people have arthritis [2] and 1 in 6 people have a mental health condition [3], while 15.9% of the population have asthma [4], and 5.2% have diabetes [5]. Living with a long-term health condition can affect labour market outcomes, including reduced working hours and earnings [6–11]. The COVID-19 pandemic, and reaction to it, caused large-scale upheaval to the labour market [12–15], with substantial increases in remote working [16] and 11.7 million people in the UK being placed on furlough [17]. Given this disruption, it is important to revisit previous findings and examine whether the effect of having a long-term condition on labour market outcomes changed after COVID-19’s onset.

There are many ways in which pandemic-related disruption could have impacted people with long-term health conditions more severely. For example, previous research has shown that employers can be reluctant to employ people with long-term conditions due to real or perceived lower productivity [18,19]. In the pandemic this could potentially have led to a greater likelihood of furlough, or even dismissal, for employees with health conditions. People with long-term conditions may also have felt more at risk from COVID-19 [20,21], and felt obliged to withdraw from work to protect themselves. Another reason people with health conditions may have been disproportionately affected is that new ways of working, rapidly introduced, failed to accommodate their needs. It could also be that health and support services, as well as routine care were harder or impossible to access during the pandemic [22–25].

Examining possibilities like the ones above is a prerequisite for targeting labour market policies to better support people with long-term conditions, and to ensure that interventions are as effective as possible in a post-COVID-19 environment. Knowing who, if anyone, was disproportionately affected by the COVID-19 pandemic, and in what ways, is vital in creating a well-functioning post-COVID-19 labour market.

This article addresses the knowledge gap by quantifying the labour market experiences of UK people with several different long-term health conditions both before and during the COVID-19 pandemic using a large longitudinal dataset [26,27]. We studied whether people employed in January/February 2020 with a given long-term condition left employment at a greater rate than people with similar characteristics but without the condition over the first 18 months of the pandemic. We also studied the intensity of participation by examining hours worked conditional on employment. These results were then contrasted with results for a two-year period before the pandemic. COVID-19 changed the ways many people experienced labour market participation, and so we also looked at whether people with long-term conditions were more likely to be furloughed, thus experiencing employment without actively working, or were more likely to work from home. Finally, we studied whether the returns to labour market participation in the form of earnings conditional on employment were lower for people with a long-term condition, both during COVID-19 and over a similar time-span pre-pandemic.

By studying multiple conditions, we were able to get a broad picture of the scale of the challenges faced by people with health conditions. It also allowed us to compare experiences between different populations, and elicit similarities and differences. Comparing pre- and post-COVID-19 outcomes allowed us to examine to what extent the pandemic changed the employment landscape for people with long-term conditions. It also provided evidence as to what extent COVID-19 had a causal impact on outcomes.

While this is the first paper we are aware of to study the labour market outcomes of many different groups of people with long-term conditions during COVID-19, some research has examined UK labour market outcomes during the pandemic for people with disabilities [22,28,29]. While this is a separate population to the ones we study, there is some overlap, and disabled people and people with long-term conditions may have faced similar issues. Our findings are compared and contrasted with previous and ongoing research on disability and the labour market during COVID-19 in the discussion section.

Methods

All analyses were carried out in R version 4.1.1 [30]. Ethical approval was not necessary for this study, as it used only anonymised data available for download under an end user license [31]. Participant consent was not required for this study, but informed consent to take part in the Understanding Society study was obtained, for details see https://www.understandingsociety.ac.uk/documentation/mainstage/consents/ (accessed 15/3/24).

Mahalanobis distance matching

In the COVID-19 analysis, we included participants who, in the COVID-19 survey, indicated being in employment (including self-employment) at baseline (January/February 2020). For the pre-COVID-19 analysis we included participants who were employed in wave 7 of the main survey. No exclusions were made based on age, as many survey participants who had reached the typical age of retirement were still participating in the labour market: around 12.5% of participants over 65 reported being in employment in at least one survey wave.

We created samples of people with and without given conditions matched on baseline variables using nearest neighbour Mahalanobis distance matching (MDM) without replacement as implemented in the MatchIt package for R [36]. MDM matches treatment/control observations based on measuring the Mahalanobis distance between them, a measure of how similar they are on all variables weighted by the amount of variation in each variable. MDM is similar to the commonly used method of propensity score matching (PSM), which estimates how likely observations are to be in the treatment group and matches treatment/control observations with similar probabilities. However, MDM is more likely to create samples which are balanced on all variables, whereas PSM may result in matched individuals with similar propensity scores, but with wide differences on some variables [37].

Details about what variables were matched on are given in Table 1. Variables for matching were selected based on consulting previous literature [e.g. 38,39], as well as consideration of what characteristics, such as key-worker status, would be relevant during the pandemic. We also matched based on baseline outcome variables. The variables were similar in the COVID-19 and pre-COVID-19 analysis, however, the latter excluded baseline working from home and key-worker status due to these variables not being collected. It also excluded baseline universal credit due to the incomplete rollout of the universal credit programme when data collection took place.

Download:

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

Table 1. Independent and matching variables.

https://doi.org/10.1371/journal.pone.0302746.t001

Where there were few people with a condition, it was possible to find several people without it who made suitable controls, but this became more difficult when more people had the condition. Therefore, the matching ratio depended on the number of people with the condition. For 150 or fewer participants, the ratio was 4:1 (i.e. four controls matched to every participant with the condition), with 150–500 participants, the ratio was 3:1, for 500–1,000 participants, it was 2:1, and for more than 1,000 participants it was 1:1.

The baseline matching procedure required complete information on all baseline variables. Rather than discard all responses from respondents missing one or more baseline variables, missing baseline variables were replaced using random forest multiple imputation, as implemented in the MissForest package [40] with 100 trees and a maximum of 100 iterations.

Modelling

Outcomes were studied separately for each long-term condition and for the COVID-19 and pre-COVID-19 periods. The outcomes were selected to give a broad overview of people’s experience of the labour market during COVID-19: participation, furlough, working patterns, and returns to participation. We did not have any firm expectation of which groups might be affected, or how, given the unprecedented disruption of the pandemic. Nor did we have an expectation that people with long-term conditions would necessarily be worse off, given for example the potential opportunities afforded by increased working from home.

To assess the impact of having a long-term condition on labour market participation, we chose employment status (including self-employed), as a dependent variable, as well as number of hours worked conditional on employment, in line with previous research [38]. Both of these outcomes were examined using random-effects models, with equations of the form (1) where y_it is the outcome for individual i at time t, LTC_i is a dummy variable equal to 1 if i has a given long-term condition and 0 otherwise, X_it and Z_i are vectors of respectively time varying and time invariant control variables, is a normally distributed individual effect, ε_it is an error term and the βs are coefficients to be estimated. The time variable t was defined differently in the COVID-19 and pre-COVID-19 analyses. In the COVID-19 analysis, t was defined as months after April 2020, i.e. t = 0 is the first wave of the Understanding Society COVID-19 survey. In the pre-COVID-19 analysis, t = 0,1,2 corresponds to waves 7, 8 and 9 respectively of the main Understanding Society survey. Control variables were the same as the matching variables in Table 1. These controls were included as an additional way of ensuing balance between people with/without a given long-term condition. As the goal was to create comparable samples of people with/without a given long-term condition to increase the likelihood of identifying causal effects, rather than a representative sample of the UK population, UKHLS sample weights were not used. Logit models were used for employment and Tobit models were used for hours worked, which are standard ways of analysing respectively binary and censored dependent variables [41].

People who were furloughed during the pandemic experienced labour market participation in a fundamentally different way, remaining employed but not actually working. To see if people with a long-term condition were more likely to belong to this group, we used whether participants were furloughed at any point as a dependent variable in cross-sectional logit models of the form

Another way in which experience of labour market participants changed during the pandemic was increase in working from home, so frequency of working from home was also analysed using ordered logit models, a standard approach for categorical outcomes with a natural ordering [42], with the same functional form as Eq (1). (Furlough and working from home were not examined in the pre-COVID-19 analysis due to the former not existing and no data on the latter).

Finally, to see if returns to labour market participation for people with long-term conditions were disproportionately affected earnings conditional on employment was studied using Tobit models with the same functional form as Eq (1).

Full details as to how the dependent variables were defined and how they were analysed are given in Table 2. Coefficients were assessed using t-tests with statistical significance judged at the conventional 5% level.

Download:

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

Table 2. Dependent variables.

https://doi.org/10.1371/journal.pone.0302746.t002

Results

Table 3 provides descriptive baseline statistics for the COVID-19 analysis samples before and after imputing missing baseline variables, and Table 4 provides them for the pre-COVID-19 analysis. A total of 2,867 responses (4.73%) in the pre-COVID-19 sample were obtained by proxy interviews; proxy responses were not possible in the COVID-19 survey. For personal characteristics in the COVID-19 data and all variables in the pre-COVID-19 data there were few missing responses. For job and financial related responses in the COVID-19 data, there were greater fractions of missing observations, for example 18.8% of participants had a missing job type, and 9.4% had missing information about earnings.

Download:

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

Table 3. Summary of COVID-19 data.

https://doi.org/10.1371/journal.pone.0302746.t003

Download:

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

Table 4. Summary of pre-COVID-19 data.

https://doi.org/10.1371/journal.pone.0302746.t004

S2–S10 Tables and S16–S24 Tables give detailed results about the MDM procedure for all conditions for respectively the COVID-19 and pre-COVID-19 analysis samples. In general, control and treatment samples were well matched, with the greatest standardised mean differences (SMDs) seen for number of comorbidities. The SMD for this variable ranged between 0.0614 for cancer and 0.837 for asthma, both in the COVID-19 analysis samples. Table 5 shows how many people were included in the treatment and control samples for each condition. The most prevalent condition was asthma, with over 5,000 respondents reporting the condition, and the least common, with only 119 people, was pulmonary conditions.

Download:

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

Table 5. Number of people in treatment and matched control groups for each condition.

https://doi.org/10.1371/journal.pone.0302746.t005

Table 6 summarises model results, with full results given in S11–S15 Tables and S251–S27 Tables. For time-varying dependent variables, the long-term condition coefficient in the COVID-19 analysis represents the initial impact (if any) on people with long-term conditions’ outcomes compared to a matched control group in the immediate aftermath of the pandemic’s onset in the UK, i.e. April 2020. It thus captures whether COVID-19 initially affected everyone in the same way, or whether people with a given long-term condition were affected differently. In the pre-COVID-19 analysis, the coefficient shows the divergence between people with long-term conditions’ outcomes compared to the matched control group between baseline (main survey wave 7) and time 0 (main survey wave 8). The interaction between the long-term condition variable and time then shows the amount the initial effect changed each month (COVID-19 analysis) or wave (pre-COVID-19 analysis). In the COVID-19 analysis, there were significant and negative initial effects of ENP and liver conditions on probability of employment, and for each the time-trend interactions were small and statistically insignificant. This indicates that the initial reduction in employment rates for people with those conditions persisted throughout the pandemic period. For asthma, arthritis, vascular conditions, and epilepsy there were statistically significant and negative interactions of condition with time trend. This indicates that the gap between the probabilities of people with and without a condition being employed grew during the pandemic. No long-term condition coefficients or time-trend interactions were significant in the pre-COVID-19 models. Hence no significant divergence in employment rates between people with and without long-term conditions over that two-year pre-pandemic period was observed.

Download:

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

Table 6. Results summary.

https://doi.org/10.1371/journal.pone.0302746.t006

For all participants in the COVID-19 analysis there was a reduction in hours worked conditional on employment in April 2020 compared to January/February 2020. Table 6 shows that there was a significantly negative time-trend interaction for arthritis, indicating that people with arthritis were slower than matched controls to increase their working hours following the initial drop. The significant coefficients for diabetes and pulmonary conditions shows that the initial reduction in hours was greater for participants with either condition compared to matched controls. For each, the time-trend interaction was also small and statistically insignificant, indicating they continued to work fewer hours compared to matched controls throughout the pandemic. Table 6 also summarises the pre-COVID-19 analysis for hours worked conditional on employment. There are no significant long-term condition coefficients of time-trend interactions.

Fig 1 shows the differences between how likely people with a long-term condition and matched controls were to be furloughed at some point in the pandemic. The greatest difference is seen for participants with pulmonary conditions. This observation is confirmed by Table 6, where the only significant effect in the analysis of probability of being furloughed is found for that patient group. The odds of being furloughed were 1.7 greater for people with pulmonary conditions compared to matched controls.

Download:

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

Fig 1. Fraction of respondents who were ever furloughed.

https://doi.org/10.1371/journal.pone.0302746.g001

Fig 2 illustrates where participants in the COVID-19 analysis worked, conditional on employment: always at home, hybrid, or never from home. There was a large rise in participants always working from home in April 2020 compared to baseline, and a corresponding fall in people who never worked from home. This pattern was still notably different from baseline in the final survey wave. Table 6 shows that there was no significant initial effect on working from home frequency for any long-term condition. However, for two conditions, arthritis and cancer, there were positive and significant time-trend interactions, showing they were slower to return to in-person working compared to matched controls.

Download:

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

Fig 2. Working from home frequency.

https://doi.org/10.1371/journal.pone.0302746.g002

Table 6 shows that at the onset of COVID-19 there were no significant differences between the earnings conditional on employment of people with any studied conditions and their matched controls. For two conditions (arthritis and cancer), there was a significant and positive interaction with the time trend, though note that in each case the sign of the initial effect is negative. This could represent a recovery following an initial drop in earnings, which the analysis was not sufficiently powered to detect at the chosen significance level. In the pre-COVID-19 analysis, there were no significant effects of having a long-term condition on earnings conditional on employment.

Discussion

The results illustrate substantial disruption to the UK labour market since March 2020. While the broad patterns were the same for people with and without long-term conditions, there were indications that people with long-term conditions were disproportionately affected during the 18 months following the pandemic’s outbreak.

The most common way in which people with long-term conditions were affected was that they were more likely to leave employment. Previous research has shown that many conditions presented barriers to work prior to the pandemic [38,43–46], so it may appear as if the decline in employment rates was unrelated to COVID-19. However, we studied only people with long-term conditions in employment at baseline, and compared their outcomes to people with similar characteristics. We also studied a relatively short space of time in the COVID-19 analysis, so previous research does not necessarily imply that a significant employment gap should have opened up regardless of the pandemic occurring. Indeed, in the pre-COVID-19 analysis, no significant employment gap between people with long-term conditions, and matched controls was seen over a comparable period. We then interpret the decline in the COVID-19 analysis as a causal effect of the pandemic: COVID-19 has exacerbated work barriers for people with long-term conditions.

Previous research has suggested that unemployment can have a causal and negative impact on people’s health, particularly mental health [47–50]. For people with long-term conditions who left employment during COVID-19, there is hence a risk of further ill health. This in turn may further reduce their chances of re-entering the labour market.

For some conditions, the intensity of labour market participation was also affected, with those in employment working fewer hours than matched controls during the pandemic. Again, no analogous results were seen in the pre-COVID-19 analysis was seen, suggesting a causal effect of the pandemic. The largest effect was seen for participants with pulmonary conditions, and those participants were also more likely to be furloughed. Perhaps this is because people with pulmonary conditions were regarded as more at risk from COVID-19 by employers, even compared to people with other long-term conditions. Then where there was flexibility in which employees were furloughed, employees with pulmonary conditions were prioritised for selection.

It could also be that people with pulmonary conditions were among the most cautious during the pandemic, and therefore the most reluctant to return to in-person work. That latter hypothesis is supported by the fact that the pulmonary group likely included the largest proportion of individuals advised to shield (i.e. minimise all physical interactions). All people with severe respiratory conditions were included in this guidance [51]. On the other hand, only people with specific cancers undergoing certain treatments were advised to shield, so the cancer group probably included many people who fell outside the guidance. Likewise, for the vascular group, only women with heart disease who were pregnant were recommended to shield. Other people included in the guidance were those with rare metabolic disorders, on immunosuppression therapies, or who had previously received a whole organ transplant. It seems unlikely these are a large group in our sample. The finding that for all long-term conditions but one there was no increased probability of furlough is comparable to previous findings that people with disabilities were no more likely to be furloughed when controlling for personal and job characteristics [52].

Some people with long-term conditions may have benefitted from increased working from home, although they could also have been negatively impacted. For example, commuting to work might be challenging, both physically and mentally, for individuals with some conditions [53,54]. On the other hand, people with other conditions may have greater need of a specially adapted workspace not available at home. We found little evidence that people with long-term conditions initially opted for different working conditions than their matched counterparts, either by working from home for a greater or lesser extent. This is probably due to initial changes being legally mandated by job type: If a job could be done from home, it had to be, and non-essential jobs which couldn’t be done from home were not done at all. As restrictions were relaxed, the general trend was for increased in-person working, but people with arthritis and cancer did so at a slower rate than matched controls. For the former group, it could be due to them feeling a benefit from avoiding commuting. For both groups, it could be that some people were shielding due to taking immunosuppressant medication, and thus were avoiding in-person work. If that were the case, then an even larger impact would be expected for pulmonary conditions. The time-trend interaction for pulmonary conditions was statistically insignificant, though this may be due to the analysis being underpowered as the coefficient’s magnitude was the largest out of all conditions. Going forward there are both challenges and opportunities in finding new ways of working, and it is necessary to account for the needs and preferences of people with long-term conditions in doing this.

For people with long-term conditions in employment, there was little evidence that an earnings gap developed between them and matched controls, either in the COVID-19 or pre-COVID-19 analysis. Indeed, for some conditions, participants’ earnings increased significantly faster during the pandemic than matched controls, although it is notable that in each case there was a sizeable, albeit statistically insignificant initial drop, so that the higher trend likely represents recouping initial losses. However, the COVID-19 data spanned only 18 months, whereas the effects of the pandemic on promotions, early retirement, etc., may not be seen for some time.

Rising labour market inactivity, especially in the older population is an ongoing policy discussion in the UK [55], with crucial implications for social care costs, growth and prosperity [56–59]. Previous research has shown that ill-health is a contributory factor in increased levels of labour market inactivity [60], via mechanisms of more people being in ill-health [61], and via long COVID-19 affecting employment [62]. Our study complements these findings by identifying another mechanism: people already living with many long-term conditions at the pandemic’s onset have transitioned to labour market inactivity at a greater rate than people without the condition. This study’s contribution to the debate is to highlight the disproportionate impact of the COVID-19 pandemic on employment for people with existing long-term conditions, and the economic benefit that reversing the trend could bring.

There are similarities between the labour market outcomes during COVID-19 for several of the studied groups. However, there are also many differences, for example reduced working hours were seen for participants with only three out of nine studied conditions, and only those with pulmonary conditions were more likely to be furloughed. It is therefore unlikely that a “one fits all” approach to labour market interventions would be particularly effective. Successful interventions should be targeted at the different needs and challenges faced by individuals with different long-term health conditions in order to maximise the individual and economic benefit.

This study is the first of which we are aware to study the impact of long-term conditions, rather than disability, on labour market outcomes during COVID-19. Having a long-term condition is not the same as having a disability, although many long-term conditions create a risk of disability, especially when combined with comorbidities [63,64]. The current study’s findings complement previous research on disability: it includes many people whose long-term condition was successfully managed to the extent of not considering themselves disabled. However, COVID-19 may have had a disproportionate impact on their labour market outcomes compared to people without a long-term condition by disrupting the management of their condition, for example due to changing working conditions, restricted access to support, or the impact of the virus itself. An advantage of studying conditions separately is the potential variation both in the extent to which the condition was manageable prior to the pandemic and the extent to which the pandemic disrupted management strategies.

Comparing our findings for people with long-term conditions with related research on people with disabilities, Emerson et al. [28] found that people with a disability were more likely to work fewer hours and more likely to experience financial stress using the first three waves of UKHLS COVID-19 survey data. Byrne, Barber & Lim [22] found a post-pandemic onset fall in employment rates for people with a mental health-related disability, and here a similar fall was observed for people with an emotional, nervous or psychiatric problem. However, in that paper, the gap between mental health-disabled and non-disabled people had begun to close by the end of 2021, whereas no comparable trend was observed here. Jones [29] found that people with a disability were more likely to be away from work, including being on furlough. Here, only people with pulmonary conditions were significantly more likely to be furloughed, possibly indicating that people who would consider themselves disabled made up a relatively small fraction of other studied conditions. Jones found that the disability pay gap did not increase during the pandemic, just as we found that people with long-term conditions did not experience lower earnings than matched controls. However, the aforementioned paper also reported no increase in the disability employment gap, whereas we found lower probability of employment for people with several long-term conditions. The contrasts above demonstrate that care should be taken in future research to study each group separately, as well as clearly distinguishing between long-term conditions and disability.

Some previous studies using the Understanding Society COVID-19 survey have used a more complex time structure than the linear trend used here, with one coefficient per survey wave (e.g. [39,65]). While that may have led to a more detailed picture given varying restrictions and disease prevalence over 2020–21, it would also have implied many more model coefficients, which may have caused problems with long-term conditions where the sample size was small. In addition, simplifying models’ time structure into an initial impact, then a broad trend over the next 17 months facilitated comparing patterns of results over several outcomes for many long-term conditions. It would be useful for future research, when examining the labour market outcomes during COVID-19 for a specific condition to model time in a more flexible way, and could also link outcomes to fluctuating restrictions and case levels.

It would have been possible to match participants on a greater number of observed variables, for example on details of what industry people worked in, rather than three categories of job type. However, we wished to avoid overfitting [66], and so relied on for example three-category job type combined with key-worker status (thereby effectively giving six categories) to capture the major differences in job situations. As we also wished to include the matching variables as model covariates as an extra step to ensure balance between treatment/control, having fewer matching variables had the advantage of fewer model coefficients in cases where sample sizes were small.

Conclusion

The main policy conclusion from these results is that people with many disparate long-term health conditions require even greater support in accessing work in a post-COVID-19 world. Interventions could take many forms, for example supported employment, in which people are placed into a job with support systems provided on an ongoing basis [68]. Other options include eHealth interventions based on mindfulness or cognitive behavioural therapy [11], or ergonomic training [69]. The variety of available interventions, and of people with long-term conditions’ experiences and needs, means interventions will require targeting to be effective, and this should be a topic for future research. Such support would have been of benefit to many people even prior to the pandemic [6–11], but their need is even more pressing now. Disruption in labour market outcomes due to a shock can persist for many years [70–72], creating an ongoing need for support.

Future research should investigate further the causes behind the observed differences in outcomes, particularly whether they reflect the preferences of people with long-term conditions or not. In turn, this could inform research about how best to target labour market interventions. Different people will have different reasons why their long-term conditions affects their labour market participation. Exploring these reasons is important in knowing what support they need to improve their wellbeing.

The studied time period was until September 2021. Yet the impact of COVID-19 on labour market outcomes is still evolving, both on the demand and supply side, for example in the development of new ways of working according to employer and employee priorities [16,73–75]. Future research could assess the longer-term impact of health conditions, including long COVID-19, on labour market outcomes. A further possibility is to investigate how partners of people with long-term conditions, especially those providing informal care, were affected [76].

Supporting information

Acknowledgments

This work was undertaken on ARC4, part of the High Performance Computing facilities at the University of Leeds, UK.