Effect of body mass on future long-term care use

Long-term care support for adults with chronic health conditions and disabilities usually comprises nursing care, personal care and assistance with domestic tasks [22]. Care can be provided either formally through professional services paid by individuals or local authorities or informally by family members, friends and neighbours [23]. Formal LTC, or more generally, social care as it is known in England, is provided by voluntary organisations, local councils, health authorities and private agencies and includes a range of services such as institutional care in care (nursing) homes, home care, day care, meals on wheels and others. England has a means-tested formal social care system whereby a publicly-funded ‘safety-net’ is in place for those with eligible financial circumstances and with greater need [24]. For our purposes we define ‘informal’ care as being provided by unpaid carers (e.g. often family members). Approximately 85% of all older people with a functional disability living in private households in England receive some informal care [22]. The number of informal care providers has increased over the years (by 11% between 2001 and 2011) and informal care itself has become more intensive, and, according to some estimates, reaching the total annual value of £55 billion [23].

Data is taken from the English Longitudinal Study of Ageing (ELSA), which is a longitudinal, biennial survey of individuals aged 50 and over. It was originally sampled from the pool of respondents to the Health Survey of England (1998, 1999, 2001) and collects data on individual and family circumstances and quality of life. We pooled data from waves 1 to 5.

In our analysis, the dependent variable is a categorical indicator measuring either any care or types of care, and non-response outcomes, including death. Care use was indicated where the survey respondent answered that they had help with activities affected by functional limitations from the associated question or were living in a care institution. Answers to further questions were used to determine the type of care used - see details in Table A1, Additional file 1: Appendix A. To avoid inconsistencies with the correspondence of categories across waves and to ensure a reasonable share of cases per category, we aggregated to broader care categories.

In line with the existing literature, our main indicator for obesity is derived from the body mass index (BMI) calculated from height and weight clinically measured by the nurses to the nearest millimetre and the nearest 0.1 kg respectively. Compromised measurements have been set to missing values [25]. We classify respondents into four groups according to the World Health Organisation’s (WHO) definition: underweight (BMI less than 18.5), normal weight (BMI 18.5 to 24.99), overweight (BMI 25 to 29.99) and obese (BMI of 30+). BMI was calculated directly for waves 2 and 4, and imputed for wave 1 (using wave 0 data). This was used as a risk factor for outcomes for waves 2, 3 and 5 respectively. Excluding the data from waves 0/1 does not change the main results, however it does prevent us from analysing heterogeneous effects due to the small sample size.

We use four sets of control factors [26‐29]:

The analysis was performed using Stata 15. We started with the descriptive analysis and using locally weighted regression (LOWESS) to gain insights into the functional relationship between care use and BMI.

The main models were estimated to account simultaneously for a range of outcomes, including the various types of long-term care support, no support, non-response and death:

For person i at time (wave of the ELSA survey) t, p_itj = prob(y_itj|X_it − 1, W_it − 1) is the probability that the individual experiences outcome j. In the baseline specification, j includes (i) any type of care, (ii) non-respondent and (iii) dead. In the extended specification, j includes (i) informal care, (ii) informal and privately-paid care and (iii) formal care (care home and LA social care), (iv) non-respondent and (v) dead. The base category in both models is no use of care. We exploited the longitudinal nature of the data (ELSA has waves 2 years apart) in an attempt to mitigate any contemporaneous bias from unobserved confounding factors that have a short-term effect (for example, a person’s current level of self-confidence, which is unobserved, might affect both the need for LTC and obesity in the current period, but is less likely to be correlated with past obesity). As such in this analysis, the person’s care use in the current period t is specified to depend on their obesity status, W_it − 1, and the other controls, X_it − 1, as measured in the previous period (wave). Additional file 1: Appendix B provides details on setting up the econometric model and strategy of dealing with bias in more detail.

These models were estimated using multinomial logit. Standard errors were clustered at the individual person level. As a robustness check, to account for the unobserved time-invariant individual effects, we also estimated an alternative specification with a quadratic function in BMI using the unobserved effect logit model (xtmelogit in Stata).

Table 1 presents summary statistics for the main sample used in the analysis, as a whole and by type of care. Overall, in the whole sample we used, 30% of respondents receive some type of care in the future, of which 1% are in care homes, 27% receive informal care, 2% receive formal care and 4% paid for care privately. BMI measures for people not receiving care of any type are lower than in the overall sample, as is the obesity prevalence. The obesity prevalence is highest for those receiving informal care (36% compared to 27% in the whole sample; p < 0.01). However, the prevalence of being overweight is highest for those not receiving any type of care. Summary statistics for control variables are available in Additional file 1: Appendix C, Table C1.

Figure 2 reports the (LOWESS) analysis of the relationship between care use and BMI. It shows that individuals with higher BMIs are far more likely to use care in the future, except for being in the care home.

Table 2 reports the main results for the any-care estimation specified in eq. (1). The coefficients in the table are relative risk ratios, with (clustered) standard errors in parenthesis (full estimation results are in Additional file 1: Appendix C, Tables C2-C3). We estimate various specifications to explore the impact of the inclusion of additional controls on the magnitude of obesity’s effects on future care use. Panel A features the results for the full sample of people aged 65 and up, with respondents who do not use any care being the base category. Panel B restricts the sample to the individuals who receive no care of any type at the start of the period. As reported in column (1) in Panel A (the whole sample), obese people, compared with people of normal weight, are 75% more likely (p < 0.01) to use some care in the future (controlling for death and non-response). If we add controls for health behaviours such as physical activity, smoking and drinking (column [2]), the effect’s magnitude decreases somewhat, but still remains significant at 65% (p < 0.01).  Column 3 reports the results as we add demographic and socioeconomic controls, as well as ADL, iADL and mobility limitation-counts in the third specification, the effect of obesity decreases further, but still remains statistically: obese individuals are 28% more likely (p < 0.01) to use care in the future compared to individuals at a normal weight.

Column (4) of Table 2 presents the specification that includes a full set of health risk factors, such as high blood pressure, diabetes, cancer, lung and heart problems, stroke, psychiatric problems and arthritis. As can be seen, the effect has decreased further, while still remaining statistically significant: an obese person is around 25% more likely (p < 0.01) than a person at a normal weight to be using some type of care in the future.

Comparing the results (in Table 2) for the full sample (Panel A) and those for the restricted sample of individuals starting with no care (Panel B), we found that the estimates for the variable of interest become slightly larger in the specification with full set of controls, but are still statistically significant at the 5% level.

Rather than outcomes categorised as any care (or not), plus non-response and death, the analysis was also conducted using an extended set of outcomes – for various types of care. Columns (5)–(7) in Table 2 (panel A) show results in which categories are defined according to types of care: (i) only informal care (IC); (ii) informal and privately-paid care (IC + PC); and (iii) formal care (both care homes and social care provided by Local Authorities) (FC). Respondents who receive the latter type of care are grouped in this category regardless of their use of informal or privately-paid care.

The overall impact of obesity on any care-use appears primarily due to the effect on informal care, while the effect on privately-paid care or formal care is smaller (16% (p > 0.05) compared with 26% (p < 0.01)) and not statistically significant. However, the latter may be due to the relatively low number of cases in this category (see Table 1 for descriptive statistics).

Potentially, respondents’ current care status may be driving the effect on the future care use. To test this, we ran all the specifications on the sample restricted to those who did not use any care initially (see panel B in Table 2). We found almost no qualitative difference in the results between the two samples. If anything, the effect was larger in magnitude for the sample with no initial care use.

We also assessed whether the effect sizes regarding obesity differ by gender. When estimating models with interaction terms on these variables, we found no statistically significant difference between genders with regards to obesity effects (results are available upon request).

To assess robustness to different model specifications, we estimated a range of alternatives (see Table 3). For easy reference, column (1) repeats the results from the main specification with full set of controls (column (4) in Table 2).

First, we investigated the use of the BMI-based obesity measures vs. an abdominal obesity (AO) measure. If AO better captures the risk for future care use, we expect to see a further decline in the magnitude of the coefficient on general obesity indicator. The AO indicator is calculated based on the waist-hip ratio (WHR),¹ which was available for a sub-sample of the data.² We found that controlling for AO did not reduce the magnitude and significance of the main coefficient of interest – on the contrary, it became larger in magnitude with obese individuals being 34% more likely (p = 0.01) to use some type of care in the future (column (2) in Table 3). Moreover, the results suggest that regardless of the BMI-based obesity status, having abdominal obesity means being a further 15% more likely (p < 0.05) to use care in the future. Moreover, this effect is preserved when we drop the BMI-based obesity measures from the specification (results are available upon request), suggesting an independent AO effect on future use of care, which merits further investigation.

Second, we considered pre-diabetes as an explanation for the obesity effect we found. ELSA contains data on blood sugar levels for around a quarter of the sample,³ from which we calculated a ‘pre-diabetes’ indicator using fasting blood glucose levels.⁴ As column (3) in Table 3 reports, we found that pre-diabetes increases the probability of using care in the future, but it is not statistically significant. At the same time, while controlling for it, obese individuals are now 70% more likely (p < 0.1) to use some type of care in the future. However, these results should be treated with caution given a significant drop in the number of observations in the blood sample.

Third, we explored subjective health and depression as further explanations for the obesity effect, where in the main analysis we focussed mostly on the functional limitations and health conditions diagnosed by a doctor as major determinants of care. In this way, we examined the effect of having good or better self-rated health and also of depression – using the Centre for Epidemiologic Studies Depression scale – as proxies for other yet-to-be-diagnosed health conditions. As reported in columns (4)–(5) in Table 3, when various combinations of these control factors were specified in the main model, we found no difference from the main result regarding the effects of obesity, while the effect of having good or better self-rated health reduces use of care in the future and depression has no effect at all.

Finally, column (6) reports the estimates from a regression in which the concurrent counts of ADLs, iADLs and functional limitations were used, i.e., not lagged with respect to the outcome measure. Their inclusion reduced the significance and the magnitude of the obesity’s effect. We might expect the current need for care to be highly correlated with current impairment rates (essentially by definition). Indeed, (lagged) obesity does not appear to affect care need beyond its effect on impairment rates. We also explored the consequences of using different estimators as noted above (unobserved effect logit model, using a quadratic function in BMI). Figure 3 reports the results – we find no statistically significant difference in the predictions from the two models.