Study participants and setting

This study analysed data from participants in the MRC 1946 National Survey of Health and Development (NSHD) [29]. The NSHD includes a socially stratified and nationally representative sample of 5,362 participants born in 1 week in March 1946 in England, Scotland, and Wales and recruited at birth. Extensive social, anthropometric, medical, lifestyle, and biological data were collected either during home visits or at clinical research centres in later years by professional interviewers and research nurses. More information on the study is available from the study’s website: https://www.nshd.mrc.ac.uk/. Each sweep of data collection has the required ethics approvals, and study members provide their consent for each assessment sweep.

The analysis sample for this prospective longitudinal study included participants who attended the age 36 years assessment in 1982 and any one of the follow-up assessments at ages 43 in 1989, 53 in 1999, 63 in 2009, and 69 years in 2015 (total N = 3,723, including 1,843 women [49.5%]). Participants lost to follow-up (due to death or emigration) between ages 36 and 63 were excluded from the study sample.

Attrition: Of the original 5,362 cohort participants, 3,723 participants were included in this analysis. By age 63 years, 718 (13.4%) had died, 594 (11.1%) had previously withdrawn from the study, 567 (10.6%) lived abroad, and 320 (5.9%) had been untraceable for more than 10 years.

Conditions

Data for 18 selected common chronic conditions on which consistent and reliable information was collected across the adulthood sweeps are the focus of this study [30]. These included diabetes, dyslipidaemia, hypertension, obesity, coronary heart disease, stroke, cancer, anaemia, respiratory disorders, kidney disorders, gastrointestinal disorders, skin disorders, osteoarthritis, rheumatoid arthritis, Parkinson disease, epilepsy, depression, and psychosis. Including conditions with consistently collected data across the 5 sweeps would aid in estimating total multimorbidity score at each time point. Data sources, sweeps at which information is available and definitions for each of the 18 conditions, are listed in Table A in S1 File.

Chronicity

We used the Delphi technique [31]—a structured communication tool—to give an informed consensus on which conditions would be considered chronic (i.e., once a participant was diagnosed with a condition then the participant would be considered to have the condition throughout follow-up). The Delphi team consisted of 10 clinicians (of ranging specialities including general practice, cardiovascular, geriatrics, and respiratory medicine) who received the same form that listed the 18 conditions, how they were measured, and a column for them to indicate which ones they would consider chronic (i.e., to be carried forward; Table B in S1 File). Each clinician was requested to indicate their decision along with a brief explanation supporting their decision. All responses were collated and examined in an Excel sheet. We chose the majority decision for each condition on whether it should be considered chronic or not.

Multimorbidity

All 18 conditions were coded as binary variables indicating whether a participant had a condition or not at each sweep. The multimorbidity score in this paper is defined as the unweighted sum of the number of conditions (out of 18 conditions) accumulated by an individual [2,14].

For those conditions considered chronic (like diabetes or hypertension), a participant was coded to have the condition throughout follow-up once diagnosed (Table A in S1 File). For all participants, we calculated a multimorbidity score at each age (i.e., across the 5 age sweeps), which indicated the number of conditions at each time point, and this multimorbidity score was the main outcome.

Socioeconomic position (SEP)

Childhood SEP was based on the father’s occupational social class, reported at age 11. We used 2 measures of SEP in adulthood: educational level attained at age 33 and occupational social class ascertained at age 43. Occupational social class was coded into 4 categories: (1) Professional and intermediate (reference group in analysis); (2) Nonmanual; (3) Manual; and (4) Partly skilled/unskilled. Educational levels were ascertained by the Burnham classification. The original 9 categories were collapsed into 3 groups for analysis: (1) None or vocational courses; (2) School leaving certificate (GCE “O” and “A” levels); and (3) Degree level and higher (university education, reference group).

Statistical analysis

We used frequencies, percentages (prevalence of each condition at all age sweeps for the full study sample), cross-tabulations (mean differences in multimorbidity score across the 5 age sweeps by socioeconomic indicators), and graphical display for descriptive analysis. To visualise prevalence and associations between the conditions that constituted the multimorbidity score at each age, networks were constructed using the R package “igraph” (conditions represented by nodes with sizes proportional to prevalence). These networks effectively illustrate the changing makeup of the total multimorbidity score at each age. Partial correlations between each of the 18 conditions are represented as “edges” in networks (networks only included those edges with partial correlations with absolute values >0.05 and p-values <0.05). The width of an edge was proportional to the strength of the partial correlation between 2 conditions (with positive partial correlations >0.05 represented by grey edges and negative correlations <−0.05 shown in red).

Socioeconomic patterning in multimorbidity at each of the 5 age sweeps was analysed using multivariable linear regression. We first ran models to test individual associations between each of the 3 indicators and the multimorbidity score. The fourth model was mutually adjusted for all 3 socioeconomic indicators in order to assess each of their associations independent of other indicators. All models were adjusted for sex.

Longitudinal change in multimorbidity trajectories were analysed using linear spline mixed-effects modelling. This accounts for repeated measurements (or clustering) of multimorbidity scores within individuals over time and individual-level deviations from an overall average trajectory, with a nonlinear specification of the average trajectory using linear splines [32]. Given the data structure—cohort members are of the same age at each time point—knots were assigned at ages 43, 53, 63, and 69 years giving 4 splines corresponding to ages 36 to 43, 43 to 53, 53 to 63, and 63 to 69 years. The model with a random intercept plus random slopes for the individual spline coefficients had a superior fit and was used in all subsequent modelling. This breaks up the average change in multimorbidity score into separate slopes for each of the 4 spline variables representing an age band (each slope coefficient indicating rate of change in multimorbidity score per year in each period). We used linear spline mixed-effects modelling to first estimate the population-average multimorbidity trajectory for the full sample and by sex. The model for multimorbidity trajectories by sex included interaction terms between sex and the spline variables for ages 36 to 43, 43 to 53, and 53 to 63 years. We then ran 4 models, each one having a different socioeconomic exposure of interest plus sex: Model 1: childhood social class; Model 2: adulthood social class; Model 3: adulthood educational level; and Model 4: mutual adjustment for all 3 socioeconomic indicators.

To examine whether the slope of multimorbidity trajectories varied by social class categories over time within each of the 4 spline variables (i.e., whether inequalities in the rate of multimorbidity accumulation changed over time), we tested for interactions between childhood or adulthood social class and the individual spline variables. Only those interactions terms that were significant (p < 0.05) were subsequently included in the model. The model for adulthood social class included interactions with all 4 spline variables (as all 4 interactions were significant, p < 0.05), whereas the one for childhood social class included interactions with the ages 53 to 63 years and 63 to 69-years’ variables only. These models were mutually adjusted for other socioeconomic indicators. Interactions were significant, and multimorbidity trajectories were plotted at the group level (childhood and adulthood social class separately) for visualisation based on model 4 above. Interactions between spline variables and education were not significant and are not shown.

Missing data across the 5 age sweeps were addressed using multiple imputation assuming missing at random (MAR) using 25 imputed datasets. The MAR mechanism implies that systematic differences between the missing values and the observed values can be explained by observed data [33], which is a plausible assumption in the British birth cohorts given the rich data available from birth [34]. Additionally, the main purpose was to impute missing data in health conditions across the 5 sweeps. The imputed data are more likely to be closer to the true values (i.e., predicted with greater precision), as many of the health conditions are correlated longitudinally further strengthening the imputation model. Information on the 3 socioeconomic indicators was available for more >94% of study participants. Frequencies and distributions of nonimputed and imputed variables were largely similar for most conditions (Table C in S1 File). All analyses including the descriptive analysis were run on the imputed sample (N = 3,723). All analyses were conducted using Stata version 15 (StataCorp LP, College Station, Texas). A completed STROBE checklist is provided as a separate supplementary file (S1 STROBE Checklist). Analyses were prespecified, but no written analysis plan is available.

Strengths and limitations

This study utilises data on individuals followed up from ages 36 to 69 years enabling us to investigate the development of and increasing inequalities in multimorbidity trajectories using robust longitudinal modelling. Additionally, this study uses both periodical cross-sectional and longitudinal analyses to investigate socioeconomic patterning in multimorbidity. The NSHD was a nationally representative sample of all births in 1 week in the UK, with a >95% response rate. As with most longitudinal studies, the NSHD suffers from attrition. Nonetheless, this study included 70% of the original birth cohort (N = 5,362) and is reasonably representative of the British born population of this generation (and with a socioeconomic distribution and mortality rate similar to that observed nationally for this generation) [38,39]. Missing data were addressed using multiple imputation, which is a routinely used robust method. While imputation relies on the MAR assumption, which is not empirically verifiable, we increased the plausibility of the MAR assumption by including socioeconomic variables (which are known predict nonresponse or missingness) in the imputation model. This should help predict missing data with greater precision and minimising nonrandom variation in the values [40]. The NSHD is composed of only members of white European heritage, and findings may not be generalizable to the non-white British population of the same generation. Some conditions like coronary heart disease, stroke, and kidney disorders were defined by only self-report of a doctor’s diagnosis. However, this should not affect estimation of multimorbidity, as self-reporting of doctor’s diagnoses was not found to bias findings in other studies [41,42], and in NSHD, >90% agreement for diagnosis of diabetes when comparing self-report with primary care diagnosis was shown [43]. The population-based nature of the study resulted in a comprehensive set of 18 conditions over time; however, studies based on electronic health records usually consist of a larger number of conditions [10,21]. The inclusion of a limited range of common conditions and excluding more rare diseases in mid-adulthood is likely to have led to some underestimation of the extent of multimorbidity. Information on certain conditions (for instance, stroke, diabetes, and arthritis) that mostly occur in older ages was not collected in early adulthood ages, for instance, at age 36 years. However, it is possible that a small number of participants could have experienced early onset of these conditions. The definitions and criteria for diagnoses of various diseases have changed over time, making it difficult to study some conditions in greater detail, for example, all respiratory conditions had to be grouped together due to changing measurement methods and definitions over the last few decades. While there are several definitions of multimorbidity, we chose to use disease counts, which is the most commonly used definition of multimorbidity enabling comparison to other studies, and appropriate as multimorbidity is the outcome of interest (more complex definitions like weighted measures and indices are better suited to studying multimorbidity in relation to outcomes like mortality and healthcare usage) [44]. There are other important dimensions of SEP shown to be related to multimorbidity in older ages such as income and health literacy, which we were unable to take into consideration [24,45]. Lastly, we cannot exclude the possibility of residual confounding by other factors (for instance, family history of disease), which we were unable to control for, which limits causal inferences we can draw about lifetime SEP and multimorbidity trajectories.

Clinical and policy implications

The increase in multimorbidity in older ages is both a natural consequence of longer life expectancies and significant change in lifestyle seen in recent decades. By age 53 years, 52% of our study population experienced multimorbidity, and we observe substantial inequalities in multimorbidity development. The largest increases in conditions over time were observed for those part of the cardiometabolic syndrome (obesity, hypertension, dyslipidaemia, and diabetes), and these conditions were more strongly interconnected over time as demonstrated by the network analysis. Primary care physicians are the main point of contact and ongoing care for patients with multimorbidity, with additional input from geriatricians with expertise in dealing with multimorbidity for older patients, and organ-specific clinical specialists for patients with specific conditions on a single disease basis. While there are limited clinical guidelines for addressing multimorbidity in the UK, these as yet do not include any guidance on targeted care for disadvantaged individuals at higher risk for greater multimorbidity as is clearly shown in this study [46].

The independent and long-lasting shadow of childhood SEP on both the experience and the rate of accumulation of multimorbidities in later life highlights the need for interventions early in life to minimise inequalities in later life health. Importantly, the ongoing and additional inequities observed in relation to adulthood socioeconomic factors stress the need for intervention and prevention efforts starting in childhood, but also their continued provision through the lifecourse to reduce these age-widening inequities in health and their corresponding consequences for healthy ageing, years lived with disability, and reduced life expectancy across all of society.

By 2035, the proportion of the UK population living with 4 or more conditions is expected to increase to 17% from 9.8% in 2015 [47]. Hypertension, chronic pain, obesity, depression, and anxiety are expected to be the leading contributors to multimorbidity, indicating the heterogeneity of diseases likely to co-occur and the strong interlinking between mental and physical health conditions [3]. Given that conditions like obesity, depression, and diabetes are socioeconomically patterned and more prevalent in younger generations, and combined with the projected increase in older age groups due to increasing life expectancies (those >75 years is expected to double to 13% over the next 20 years), the issue of multimorbidity, and its inequalities, will be a heavy burden that poses serious challenges for healthcare systems. Multimorbidity is linked to increased healthcare utilisation, which is largely attributable to the number of interventions required to deal with multiple chronic health conditions in individuals [48]. In addition to placing a substantial economic burden on the healthcare system, it will also require significantly more healthcare professionals and potential restructuring of healthcare provision in the near future to effectively address the issue.

The synergistic clustering of diseases at both the population and individual level is one of the mechanisms by which health inequalities develop and perpetuate in sections of the population [15,49,50]. Hence, any model developed to tackle multimorbidity and to reduce inequalities associated with socioeconomic circumstances should incorporate societal and individual-level factors. Additionally, public health policies that aim to reduce multimorbidity need to start earlier in the lifecourse, must be universal but proportional with increased targeting towards the more socioeconomically disadvantaged population. Individuals with multimorbidity living in deprived areas receive poorer healthcare (for example, shorter consultation times and lower perceived GP empathy) compared to their counterparts in affluent areas [51]. This reflects the inverse care law, which states that the availability of good medical care varies inversely with the need for it in the population served [52]. Hence, there also needs to be a focus on improving healthcare access and services at the ecological level as there is a greater concentration of individuals with multimorbidity among the socioeconomically disadvantaged.

Final summary

In conclusion, this study shows marked and independent effects of childhood and adulthood socioeconomic circumstances on onset, rate of accumulation, and burden of multimorbidity. These inequalities widen with time and the effects of socioeconomic disadvantage in childhood on multimorbidity emerges in late adulthood and early old age. Strikingly, the most socioeconomically disadvantaged by each measure of SEP experienced an additional 0.39 conditions (childhood social class), 0.83 (adult social class), and 1.08 conditions (adult education) at age 69 years. Multimorbidity is perhaps inevitable in ageing societies with increasing life expectancies, but the inequalities as demonstrated by our study are stark. This calls not only for better access and delivery of healthcare for the more vulnerable, but also for population-based interventions in early life and through the lifecourse that reduce the impact of childhood inequalities if we are truly to address the societal impact of multimorbidity.