The association between obesity, health service use, and work productivity in Australia: a cross-sectional quantile regression analysis

The burden of disease attributable to obesity is rapidly increasing and becoming a public health challenge globally. Using a nationally representative sample in Australia, this study aims to examine the association of obesity with healthcare service use and work productivity across outcome distributions. We used Household, Income and Labour Dynamics Australia (HILDA) Wave 17 (2017–2018), including 11,211 participants aged between 20 and 65 years. Two-part models using multivariable logistic regressions and quantile regressions were employed to understand variations in the association between obesity levels and the outcomes. The prevalence of overweight and obesity was 35.0% and 27.6%, respectively. After adjusting for socio-demographic factors, low socioeconomic status was associated with a higher probability of overweight and obesity (Obese III: OR = 3.79; 95% CI 2.53–5.68) while high education group was associated with a lower likelihood of being high level of obesity (Obese III OR = 0.42, 95% CI 0.29–0.59). Higher levels of obesity were associated with higher probability of health service use (GP visit Obese, III: OR = 1.42 95% CI 1.04–1.93,) and work productivity loss (number of paid sick leave days, Obese III: OR = 2.40 95% CI 1.94–2.96), compared with normal weight. The impacts of obesity on health service use and work productivity were larger for those with higher percentiles compared to lower percentiles. Overweight and obesity are associated with greater healthcare utilisation, and loss in work productivity in Australia. Australia’s healthcare system should prioritise interventions to prevent overweight and obesity to reduce the cost on individuals and improve labour market outcomes.


Methods
Sample and data. We used the nationally representative cross-sectional data from the Household, Income and Labour Dynamics in Australia (HILDA) wave 17, which was conducted between 2017 and 2018. HILDA is a nationally representative longitudinal study of Australian residents aged 15 years and above. The baseline survey of HILDA commenced in 2001, collecting information annually on various aspects of life in Australia, including demographics, socio-economic status, healthcare utilisation, work productivity, and the Health-Related Quality of Life (HRQoL). Multi-stage sampling approach was used to select sample households. First, a stratified random sample of 488 Census Collection District containing 200 to 250 households in each district was selected 20 . A sample of 22 to 34 dwellings were then randomly selected, and within each dwelling, up to three random households were selected 20 . Interviews and self-completed questionnaire following the University of Melbourne's ethical guidelines were used for data collection. Further detailed description of the survey objectives and data collection method can be found elsewhere 20 . HILDA Wave 17 includes a total of sample including 17,570 respondents with a response rate of 96.4%. In this analysis, we included working-age respondents between 20 and 65 years old with Body Mass Index (BMI) ≥ 18.5 kg/m 2 and excluded those who had a missing value on outcomes and any of the covariates. The final sample consists of 11,211 respondents.
Outcomes. In this study, health service use was measured using the number of General Practitioner (GP) visits and nights spent at hospital, as well as the number of prescribed medications for regular use in the previous 12 months. Work productivity was measured through the number of paid sick leave days taken in the last 12 months for the respondents who were employed at the time of the survey.
Covariates. Covariates included sex, age group (20-29, 30-39, 40-49, 50-59, 60-65), education level (low, middle, high), household income (1-4 quantiles of sample), Indigenous status (Non-Indigenous Australian, Indigenous Australian), country of birth (Australia, main English-speaking countries, other), marital status (married/de facto, others), Australian state (New South Wales, Victoria, South Australia, Queensland, Western Australia, Tasmania, Northern Territory, Australian Capital Territory), residential area (urban, rural), alcohol intake (no alcohol intake, drink alcohol less than 3 days per week, drink alcohol 3 and more days per week), smoking (non-smoker, ex-smoker, current smoker). www.nature.com/scientificreports/ Statistical analysis. We examined the determinants of overweight and obesity using multivariable logistic regression models. Two-part models were used to investigate the association of overweight and obesity with health service use and work productivity loss. First, we applied a multivariable logistic regression model to examine the probability of observing positive binary outcomes for any health service use (yes/no) and any sick leave taken (yes/no) against obesity levels. Second, for those with positive binary outcomes for these variables, we performed multivariable quantile regression to estimate the different impacts of overweight and obesity on health service use and loss of work productivity across their distributions (10th, 25th, 50th, 75th, and 90th percentiles). The quantile regression is estimating the quantiles for the outcome variable associated with a set of independent variables and covariates without assuming normality or homoscedasticity of the underlying distribution 22,23 . Advantage of the quantile regression is the robustness to outliers as it allows for assessing the full distribution of the outcome variable and is suitable for modelling outcomes that are not normally distributed or are highly skewed 22,23 .
In this study, the coefficients at the lower percentiles of the health service use distribution (e.g. the 10th, 20th, and 30th percentiles) indicate the association between obesity and low level of health service use, while the coefficient at the higher percentiles of the health service use distributions (e.g. 70th, 80th, and 90th percentiles) reflect the association between overweight and obesity and high level of health service use. In addition, we performed Ordinary Least Square models (OLS) and presented estimated mean change of the outcomes by overweight and obesity levels to compare with the estimated change at different points of outcome distributions obtained from multivariable quantile regression models. All multivariable regression models were adjusted for covariates listed above and p-values < 0.05 were considered statistically significant. All analyses were performed using Stata 15 (Stata Corp., College Station, Texas).

Results
Sample characteristics. The study sample consisted of 11,211 respondents. The median age of respondents was 42 (IQR 30-53) and BMI was 26.4 (IQR 23.4-30.5). The majority of respondents were married/de facto (70.6%), living in an urban area (87.4%), with a middle or high education level (84.1%). The prevalence of each obesity level was 37.5% (normal weight), 35.0% (overweight), 16.8% (obese I), 6.69% (obese II), and 4.08% (obese III). (Table 1). Table 2 presents the determinants of each obesity level. The odds of being obese were higher for those in lower socioeconomic group (SEIFA 5: Obese II OR = 2.46, 95% CI 1.85-3.28, Obese III OR = 3.79, 95% CI 2.53-5.68) compared to higher socioeconomic group. Respondents who are not in the labour force have 65% higher likelihood of being highest level of obese (95% CI 1.29-2.11). Middle age group (40-49 years old and 50-59 years old) showed higher likelihood of being overweight and obese compared to other age groups. On the other hand, higher education level and higher alcohol intake showed adverse relationship with obesity level. High education group had 53% (95% CI 0.36-0.61) and 58% (95% CI 0.29-0.59) lower likelihood of being obese II and obese III, respectively. The odds of being severe obese were significantly lower for those who were born in non-English speaking countries (Obese II OR = 0.29, 95% CI 0.21-0.41, Obese III: OR = 0.26, 95% CI = 0.17-0.41). Drinking alcohol more than 3 days per week had 70% lower likelihood of being obese III (95% CI = 0.22-0.41). In addition, we found that Indigenous status has a positive association with lower obesity level (Obese II OR = 1.68, 95% CI 1.09-2.60) but the association was not significant for Obese III (OR = 0.9, 95% CI 0.50-1.60).

Sociodemographic correlates of overweight and obesity.
Association between obesity, health service use, and work productivity. Health service use across percentiles of outcome variables. Table 3 describes the association between obesity levels and health service use. A higher level of obesity was associated with higher probability of using any GP (Obese III: OR = 1.42 95% CI 1.04-1.93), hospitalisation (Obese III: OR = 1.60 95% CI 1.22-2.09), as well as use of prescribed medication (Obese III: OR = 2.40, 95% CI 1.94-2.96).
Amongst those with normal weight and had any health service use, the estimated mean number of GP visit, nights at hospital, and prescribed medications were higher than their median (GP visit: 5.32 vs. 3, nights in hospital: 5.68 vs. 3, prescribed medication: 1.90 vs. 1), illustrating the positively skewed nature of the distributions. Table 3a shows the relationships between obesity level and GP visits by health service use quantile. An increase in BMI (from normal weight to overweight, obese I, obese II, and obese III) was associated with a higher likelihood of GP visits and a higher number of GP visits in every quantile group. The effect of obesity on GP visits was found to be greater in the upper tail of health service use distributions than in the lower tail. For example, the coefficient for obese II was 0 (95% CI − 0.15-0.15) for the 10th percentile, and 3.58 (95% CI 2.21-4.95) for the 90th percentile). The coefficient for obese III was 1.0 (95% CI 0.81-1.19) for the 10th percentile and 6.54 (95% CI 4.83, 8.25) for the 90th percentile.
Table 3b presents the relationship between obesity level and number of nights at a hospital. Number of nights in hospital was positively associated with obesity II and III at the 75th and 90th percentiles of the distribution while it was negatively associated with a lower level of obesity. However, these associations were not statistically significant.
The association between obesity level and number of prescribed medications is presented in Table 3c. Higher obesity level was associated with higher likelihood of an increasing number of prescribed medications compared to normal weight at the upper tail of the distribution. The effect of obesity was greater at the 75th and 90th percentiles of the distribution compared to the lower tail. For example, the coefficient for obese II was 0.76 (95% CI 0. 43

Discussion
Principal findings. Overall, the prevalence of overweight and obesity was 35.0% and 27.6%, respectively, among Australian residents aged 15 years and above during the period of 2017-2018. After adjusting for sociodemographic factors, low socioeconomic status was associated with overweight and high level of obesity while     www.nature.com/scientificreports/ high education group was associated with a lower likelihood of being obese. Generally, a higher level of obesity was associated with higher probability of health service use, and work productivity loss. Regarding quantile analyses, using quintile regression methods, this study showed that the high level of obesity had significantly increased number of GP visits and prescribed medications across the outcome distributions, In terms of work productivity loss, the impact of obesity levels on number of sick leave was larger at upper tiles of distributions. These findings suggested that the impacts of overweight and obesity on health service use could be higher for obese individuals mainly because these individuals may have more healthcare needs.
Comparison with existing literature. We found that the prevalence of overweight and obesity was 35.0% and 27.6%, which is consistent with population data from Australia 1,24 . Our findings that lower income and educational attainment were associated with higher probability and levels of obesity is also in line with literature in Australia and other high-income countries 25,26 .
Higher levels of obesity were associated with increased number of GP visits and number of prescribed medications, which has been found in studies from other high-income countries 9,11,12,27,28 . The quantile regression analysis show that such association was more profound among those at upper tiles of distribution (i.e. among those use more GP service and prescribe more medications). This is in line with a study used quantile regression to examine association between BMI categories with health costs 29 . The finding of increased health service use is indicative of the direct costs of overweight and obesity in Australia, with annual estimates of between $2 Billion and $3.8 Billion 24,30,31 .
Overweight and obesity were associated with increased likelihood and number of sick leave days taken, and the effect was larger at upper tiles of distributions. Though focused on a sample from defence force in Australia, the author also found that productivity losses from full days off work were higher in the obese subgroup compared with the normal cohort 17 . Our finding is also in line with a border literature on the indirect costs of obesity on productivity loss [32][33][34] , with obesity believed to incur productivity costs of almost $5 Billion annually in Australia 31,35,36 . Strengths and limitations. This is the first study to explore healthcare use and productivity loss associated with obesity using nationally representative data in Australia. Besides, this is a novel study to estimate the effect of obesity applying quantile regression which reflects the variations in the impact of obesity across the outcome distributions.
However, there are some limitations. First, we use self-reported data for BMI which may be biased due to measurement error. An existing study indicates that obese people are likely to underestimate weight 37 . In addition, self-reporting of health service use and productivity is likely to cause recall error as the respondents are required to remember and report information from the last 12 months. Second, the outcomes of the impact of obesity on health service use and productivity in this study might be underestimated, as the survey data was not collected from patients with severe illness or those admitted to hospital. Third, healthcare utilisation and productivity loss were not specifically due to illness or disability associated with obesity, and could have included other unrelated reasons and conditions such as acute illnesses and physical injury. Fourth, we used alcohol consumption based on weekly frequency but not on intensity, thus, it may not reflect the actual amount of alcohol consumption. Further, our analysis on work productivity only looked at number of sick leave days as an indicator due to the availability of question. Thus, the results might not capture the full picture of the burden of obesity. Other indicators of work productivity such as wage loss or presenteeism should be taken into account for future studies. Another limitation of our study is that our estimation may be biased as a result of omitting confounding factors that may be associated with outcomes, such as individuals' local environment and genetic factors, which are not captured in the HILDA survey. Lastly, the interpretation of causal effect might be limited because of the use of cross-sectional study design. Future studies need to consider using longitudinal data to observe a long-term effect of obesity on outcomes.
Policy implications. These findings highlight the significant variance in the demands that differing severities of overweight and obesity can place on health care systems, as well as its implication for employment productivity on both a personal and population level. This raises the question as to the appropriateness of direct and indirect cost estimates for obesity based on linear regression models. While this may capture mean estimated costs, they do not reflect the implications at the different quantiles of the distribution of outcomes.
Our results of the quantile regression analyses demonstrate the associations between obesity or overweight largely varied at different quantiles. The impact of obesity is larger for those with a higher degree of obesity and with greater use of health service and lower work productivity. A likely account for this result was that obese individuals suffer double economic burdens of increasing use of health services and reduced work productivity.
It is evident from our findings that policy approaches to address obesity and its impacts must be two-fold; firstly, addressing the prevalence of overweight and obesity in Australia, and secondly in reducing its association with adverse health service and productivity outcomes. A range of policy approaches are needed in order to confer both benefits, targeting obesity on both an individual and a population or workforce level. There are a number of broad measures to reduce the prevalence of excess weight which are supported by the literature; taxes on low nutrition foods and sugar sweetened beverages, while politically difficult have been modelled to reduce obesity and improve productivity 35,38,39 , as have investments in urban design to facilitate greater physical activity 40,41 . On an individual level, there is evidence for utilisation of bariatric surgery to both reduce excess weight and its impact on employees wellbeing, as well as reducing barriers to employment 42,43 .
It is also necessary that interventions and programs target those who already suffer from overweight and obesity, in order to improve their wellbeing and occupational performance. There is encouraging evidence that www.nature.com/scientificreports/ workplace-based health interventions can see improvements in health and productivity before a noticeable decline in overweight and obesity [44][45][46] . Evidence supports targeted interventions by Primary care practitioners 35,47 , however a lack of implementation of established guidelines for obesity management in Australia has hampered the effectiveness of this 48,49 . Failure to focus on both the management and prevention of obesity will see poor returns in the form of ongoing costs for both individuals and wider societal costs (Supplementary Table 1).

Data availability
Data is publicly available from Household, Income and Labour Dynamics in Australia Survey website: https:// melbo urnei nstit ute. unime lb. edu. au/ hilda.