Data source

The paper uses data from wave 12 (collected between 2020 and 2022) and wave 13 (collected between 2021 and 2023) of the UKHLS, a longitudinal survey of approximately 40,000 households in the UK which began in 2009 (University of Essex, 2023). This dataset is ideal for our analysis, as it includes an Ethnic Minority Boost Sample of at least 1,000 members of five ethnic groups: Indian, Pakistani, Bangladeshi, Caribbean and African. The analytical sample for each part of the analysis differs depending on the valid responses to the variables of interest. Ethical permission for the secondary data analysis was obtained from the University of Southampton. Table S1 shows that the UKHLS sample closely mirrors the ethnic composition of the UK Census 2021.

Outcome measures of pension membership

A workplace pension is a way of saving for retirement arranged by the employer (Department for Work and Pensions, 2025c). Workplace pension membership (No = 0, Yes = 1) was assessed using four survey questions, which are nested: being in paid work (No = 0; Yes = 1); among those in paid work, being an employee (No = 0; Yes = 1); among employees, working for employer with a workplace pension scheme; (No = 0; Yes = 1); and among those working for such an employer, being a member of such scheme (No = 0; Yes = 1) (see Figure S1). As these survey responses are self-reported, we are not able to assess whether all respondents asked such questions were eligible for work-based pensions, and this limitation is discussed at the end of the paper. The information about workplace pension membership was asked in wave 12, which is used for this part of the analysis.

Personal pensions are pensions that individuals manage for themselves, and which are separate from the state pension (Department for Work and Pensions, 2025a). Personal pension membership is derived from the following question in wave 13: ‘Other than your main employer pension scheme, are you currently a member of any personal pension scheme or do you currently contribute to any personal pension scheme?’ (No = 0; Yes = 1).

The outcome variable ‘savings or investments (other than pension)’ is derived from the following question in waves 2–11: ‘Do you expect to use savings or investments to provide for your retirement?’ (No = 0; Yes = 1), and the following question in wave 13: ‘Apart from pensions, are you likely to use savings or investments (other than pension) to provide money in later life?’ (No = 0; Yes = 1).

Predictor variables

Ethnicity and gender are key variables in the analysis. Gender is an important predictor of pension membership, both individually and when it interacts with ethnicity (Vlachantoni et al., Reference Vlachantoni, Feng, Evandrou and Falkingham2015 and references therein). Ethnicity was recorded according to the respondents’ self-report of their ethnic group into White British (includes British, English, Scottish, Welsh, Northern Irish), Other White (includes Irish, Gypsy or Irish traveller, any other White background), Indian, Pakistani, Bangladeshi, Chinese, Other Asian, Black (includes African/Caribbean/any other Black background), Mixed (includes White and Black Caribbean/White and Black African/White and Asian/any other Mixed background) and Arab & Other ethnic groups (see Table S1).

Socio-economic characteristics

Individuals’ socio-economic characteristics can directly relate to their pensions and other financial plans. Housing tenure has an impact on individual health, social security and welfare (Stevenson et al., Reference Stevenson, Aalabaf, Babad, Billig, Goldman, Grenham and Gutterman2019), while high house prices have a crowding-out effect on pension saving for middle-earning employees (Crawford & Simpson, Reference Crawford and Simpson2020). Housing tenure was recoded as: own outright; own with mortgage; social rent; rent privately; others (Waine, Reference Waine2006). Education has an evidenced impact on individuals’ type of employment in the UK (Paterson, Reference Paterson2022). We used the respondents’ highest educational qualification, which was recoded as: degree or higher; A-level; General Certificate of Secondary Education (GCSE); other qualification; no qualification. Drawing on existing research, which has used income quintiles to represent the income distribution (Cribb et al., Reference Cribb, Emmerson, Johnson and Karjalainen2024), we derived an income quintiles variable from the total household net income, divided by the size of the household, and distinguished between households in the bottom two and the top three quintiles of the income distribution. The use of quintiles, and the comparison between the top and bottom of the distribution, allows us to explore the socio-economic characteristics associated with membership of different types of pension schemes. The derived variable Job type was recoded as: educational/sport; public administration; health services; banking; transports/ trade; agriculture/construction/ industrial; did not work last week (but has a job); and others. Finally, we used income support receipt (No/Yes) as an indicator of lower socio-economic status.

Demographic characteristics

Age was recoded into two groups (forty-five to fifty-four and fifty-five to sixty-four-year-old adults) to capture the effect of an individual’s proximity to retirement on their pension and other financial arrangements. We acknowledge that there is variation in the retirement age within this study population, and that the state pension age is currently increasing; however, the distinction between individuals in mid-life (forty-five to fifty-four) and those closer to retirement (fifty-five to sixty-four) allows us to examine patterns of pension membership for two age groups that are central to our analysis. Legal marital status (married/ cohabiting; widowed; divorced/ separated; single or never married) is included as research has shown that married men and women are more likely to participate in pension schemes than their unmarried counterparts (Dushi & Iams, Reference Dushi and Iams2013). At the same time, living with one’s spouse or civil partner was used to incorporate the relevance of one’s living arrangements on their pension and other financial arrangements (Joseph & Rowlingson, Reference Joseph and Rowlingson2012). Finally, previous research has documented a strong link between households with three-plus children and poverty (Resolution Foundation, 2024), as such, we have included the number of children in the analysis as: No children, 1–2 children, 3+ children.

Individuals’ migration history, for example the timing of their arrival in the UK, directly affects the amount of time they have been living in the UK, and by extension, their employment patterns and workplace pension eligibility (Allmark, Reference Allmark2010; Vlachantoni et al., Reference Vlachantoni, Feng, Evandrou and Falkingham2015). The migration history variable allows us to examine the relevance of the migration timing into the UK by those born abroad but who migrated between 1930 and 1969; 1970–1999, and 2000–2020, compared to those who were born in the UK (non-migrants). The place of residence (urban or rural) variables allow us to take into account nation-level and regional differences in pension membership (Zhang et al., Reference Zhang, Wang, Li and Wang2023).

Health characteristics

Individuals’ self-reported health is a commonly used indicator in social, clinical and epidemiological research and a reliable health indicator among older adults (Akhtar et al., Reference Akhtar, Saikia and Muhammad2023). Self-reported general health (SRH) was recoded into two categories: good (excellent, very good, good) and poor (fair, poor). The variable indicating the report of any long-standing illness/disability was recoded as: No/Yes. We also included indicators of health satisfaction (satisfied; neutral; dissatisfied) and general mental health in our analysis, to obtain a broader picture of individuals’ wellbeing. The General Health Questionnaire (GHQ) screens for general psychiatric illness and is a widely used self-report measure of non-psychotic mental illness (Donath, Reference Donath2001). The GHQ-12-caseness scale (score ranging from 0 to 12) was recoded into two categories: score < 3 (no risk of mental illness); and score > 3 (risk of mental illness).

Subjective financial wellbeing indicators

Indicators of individuals’ subjective wellbeing, measured through a range of dimensions, have been shown to accurately reflect the quality of individuals’ lives in different domains (Diener et al., Reference Diener, Inglehart and Tay2013). Previous research on measuring subjective financial wellbeing has pointed to a range of indicators (Fan & Henager, Reference Fan and Henager2022), of which we used two in our analysis: income satisfaction coded as: satisfied; neutral; dissatisfied; and the respondents’ subjective financial situation in the present (comfortable; just about getting by; difficult) in the present.

Statistical analysis

In the first part of the analysis of trends between 2009 and 2023 we calculate the percentage distribution by ethnic group and gender of: those being in paid work; among those in paid work, being an employee; among employees, working for an employer who offers a workplace pension; among those working for such an employer, being a member of such scheme; being a member of a personal pension scheme; and expecting to use non-pension savings/ investments in retirement. Chi-square values were applied to each outcome variable to examine significant associations. We conducted a variance inflation factor (VIF) analysis to investigate multicollinearity, which did not detect any issues with the models (see Tables S5–S10). In the second part, the multivariate analysis focusing on wave 13 used a binary logistic regression approach, which is widely used in this topic area (Akhtar et al., Reference Akhtar, Saikia and Muhammad2023; Vlachantoni et al., Reference Vlachantoni, Evandrou, Falkingham and Gomez-Leon2020, Reference Vlachantoni, Falkingham, Evandrou and Qin2023). For each regression, the odds ratio indicates the change in the odds of having the outcome (e.g. personal pension membership) relative to the reference group while holding the control variables constant. The analyses are performed in STATA (version 18.5) (StataCorp, 2023), and figures are generated in R software (version 4.3.1) using the ggplot2 package (Wickham & Chang, Reference Wickham and Chang2011).

The equation of the final model for personal pension membership is:

• •

  $$\eqalign{ & {{Y}} = {{Logit}}\,\left( {{{P}} = {{personal}}\;{{pension}}\;{{membership}}} \right) \cr & = {{log}}{{{P}} \over {1 - {{P}}}} = {{a}} + {{{b}}_1}{{*\,Ethnicity}} + {{{b}}_2}*{{Housing}}\;{{tenure}}\,\cr & {{ + }}\;{{{b}}_{{3}}}*{{Educational}}\;{{qualifications + }}{{{b}}_{{4}}}*{{Income}}\;{{quintiles}} \cr & \;{{ + }}\;{{{b}}_{{5}}}*{{Job}}\;{{type}}\;{{ + }}\;{{{b}}_{{6}}}\,*{{Income}}\;{{support}}\;{{ + }}\;{{{b}}_{{7}}}\,{{*\,Any}}\;{{savings}}\;{{/}}\;{{investments}} \cr & \;{{ + }}\;{{{b}}_{{8}}}*{{Gender}}\;{{ + }}{{{b}}_{{9}}}*{{Age}}\;{{group \,+\, }}{{{b}}_{{{10}}}}*{{Migration}}\;{{history}} \cr & \;{{ + }}\;{{{b}}_{{{11}}}}*{{Number}}\;{{of}}\;{{children}}\;{{ + }}\,{{{b}}_{{{12}}}}*{{SRH}}\;{{ + }}\,{{{b}}_{{{13}}}}*{{SFS}}\;\left( {{{present}}} \right) \cr}$$

The equation of the final model for expecting to use non-pension savings/ investments in retirement is:

• •

  $$\eqalign{ & {{Z}} = {{Logit}}\,\left( {{{S}} = {{Any}}\;{{savings}}/{{investments}}} \right) \cr & = {{log}}{{{S}} \over {1 - {{S}}}} = {{a}} + {{{b}}_1}{{*\,Ethnicity}} + {{{b}}_2}*{{Any}}\;{{personal}}\;{{pension}}\;{{membership}} \cr & \;{{ + }}\,{{{b}}_{{3}}}*{{Housing}}\;{{tenure}}\;{{ + }}{{{b}}_{{4}}}*{{Educational}}\;{{qualifications}} \cr & \;{{ + }}\,{{{b}}_{{5}}}*{{Incomequintiles}}\;{{ + }}{{{b}}_{{6}}}*{{Job}}\;{{type}}\;{{ + }}\;{{{b}}_{{7}}}*{{Income}}\;{{support}} \cr & \;{{ + }}\;{{{b}}_{{8}}}*{{Gender}}\;{{ + }}\;{{{b}}_{{9}}}*{{Marital}}\;{{status}}\;{{ + }}\;{{{b}}_{{{10}}}}*{{SRH}} \cr & \;{{ + }}\,{{{b}}_{{{11}}}}*{{Income}}\;{{satisfaction}}{\mkern 1mu} {{ + }}\;{{{b}}_{{{12}}}}*{{SFS}}\;\left( {{{present}}} \right) \cr & {{+ }}\,{{{b}}_{{{13}}}}*{{Ethnicity}}*{{Gender}} \cr}$$

where ‘ ${{P}}$ ’ is the probability of the occurrence of event ‘ ${{Y}}$ ’, ‘ ${{S}}$ ’ is the probability of the occurrence of event ‘ ${{Z}}$ ’, ‘a’ is the constant coefficient and ‘b’ is the coefficient associated with the independent variable, where the odds change with unit increases in the independent variable. If ‘b’ is positive, the odds ratio (OR) will increase, as this factor will be >1, whereas if ‘b’ is negative, the OR will decrease, as this factor will be <1. When ‘b’ is zero, the factor exponential of ‘b’ is one and the odds remain unchanged (Ali et al., Reference Ali, Akhtar, Chauhan, Malik and Singh2023).