Introduction

socioeconomic status (SES) reflects a person’s absolute and relative position in society. It captures a combination of access to material and social resources, as well as relative status, i.e., prestige- or rank-related characteristics. SES is reflected in a broad spectrum of variables, including financial, occupational, and educational influences. Although interrelated, each of these domains reflects somewhat different individual and societal forces associated with health and disease risks [22, 24].

SES has generally been found as a strong and consistent predictor of a person’s health risk, including morbidity and mortality risks (e.g., [5, 10, 18]). Across most diseases with few exceptions, SES has a strong association with disease risk factors and the association has a tendency to persist throughout an individual’s life [2, 15, 19].

Although individuals with SES disadvantage may have greater health needs due to generally poorer health than non-disadvantaged individuals, their greater health needs do not necessarily translate into greater use of health care. Insurance, funding, co-payments and other financial and non-financial constraints are important considerations, as these are integral parts of the social system under which health care is accessed. As such, the evidence on the link between SES and health care utilization is mixed.

In comparing health care utilization from 12 European countries, Terraneo [26] reports a clear positive association between education levels and specialist and dentist visits, no evidence of educational disparities for GP use, and mixed results regarding hospitalizations. Lusardi et al. [17] compare health care use in five advanced countries and finds that socially disadvantaged persons (the unemployed and financially fragile) are likely to forgo health care during an economic crisis, and the extent appears to align with differences in out-of-pocket costs across countries.

Elsewhere variable and contrasting findings have also been reported, e.g., Lostao et al. [14] found modest effects of income on health care use in Germany and Spain, while Löfvendahl et al. [16] found a bell-shaped relationship for Sweden, with patients in the second and third income quintiles having the highest mean annualized cost. For the English NHS, Morris et al. [21] find that socioeconomic and ethnic variables influence utilization even after controlling for a large set of health variables. The review by Dixon et al. [7] suggests that in the UK the disadvantaged tend to use less relative to need for certain health services; see also Goddard and Smith [9], Van Doorslaer et al. [27], among others. However, a more recent review by Cookson et al. [6] find that health care use tends to be generally pro-poor in the UK.

Related to this literature is the vast literature on measuring inequity in health care use [8, 28]. An important consideration there is the notions of horizontal and vertical equity. Most empirical studies tend to focus on measuring horizontal equity. A typical approach is to measure the association between SES and some measures of utilization (e.g., hospital admissions) while allowing for ‘appropriate’ or ‘fair’ differences in individual needs [6, 21]. However, to allow for appropriate differences in utilization invariably requires some normative notion of vertical equity [25]. Another difficulty with most studies is the failure to account for disease progression. Patients in an advanced stage of a chronic disease obviously tend to incur greater health care use than those at the initial stage.

This study quantifies the effect of SES on hospital utilization and adverse in-hospital events of chronic disease patients in Australia. The investigation focuses on chronic disease patients for whom an initial disease spell can be identified, and utilization and adverse events are measured over the identified spell. Further analysis examines the extent to which differences in the incidence of adverse outcomes across SES groups explain the observed difference in utilization across SES groups.

This paper contributes to the literature by proposing an approach of identifying chronic disease spells and measuring utilization and adverse outcomes over the entire period of the spell. Since patients are placed on similar disease progression time-line, by the principle of horizontal equity, they should have similar utilization. With sufficient data, the approach can be extended to map out health care use in successive spells by different SES groups, thereby providing useful insight into the long-term utilization trajectory of chronic disease patients. This study also produces new insight into how SES directly and indirectly affects costs of hospital utilization, where the indirect effect works through the incidence of adverse in-hospital events.

Methods and data

Background

This study makes use of hospital administrative data from the state of Victoria, Australia. With about 6 million residents, Victoria is the second most populous state in Australia. The state has about 300 hospitals, about 120 are public hospitals and the remaining are private hospitals. The latter are typically small, and about half are day-surgery hospitals catering to patients undergoing same-day procedures.

Australia has a tax-funded universal health insurance system, Medicare, which provides coverage for public hospital admissions for all Australians. Equity of access to health care is a primary goal of the health care system. Hospital care is provided free of co-payments to all Australians admitted to public hospitals. In addition to Medicare, private health insurance can be purchased to provide coverage for private patient admissions. About 46% of the population has private health insurance, which allows the insured to be admitted as private patients in either public or private hospitals. Using private health insurance allows the insured to bypass the waiting list for elective surgery at public hospitals, in addition to providing a choice of doctors and generally better room amenities than public patients. However, using private health insurance generally entails out-of-pocket payments which can be substantial in some cases.

About 60–70% of all in-patient admissions take place in public hospitals. Health care in public hospitals in Victoria is mostly funded via casemix funding administered by the state government. Private hospitals, in contrast, obtain their funding through several channels, from health insurance funds (by contractual agreements), the federal government (through Medicare rebates) and from patients in the form of co-payments. In recent years the state government has been investigating various funding models to better support the broad range of services delivered by Victoria’s public health services. This includes a pilot program known as the HealthLinks model, which allows providers greater flexibility in managing patients with chronic disease and complex health needs. An important policy issue that arises is the extent to which SES should be accounted for in setting individual health service budgets. It is in this context that the current study is undertaken.

Data

The main data source for this research is the Victorian Admitted Episodes Dataset (VAED), a hospital administrative data set maintained by DHHS. The data set contains all admission episodes occurring in all hospitals, public and private, in Victoria. The data allow patients to be followed across admission episodes via a linked patient identifier. This study makes use of a special version of the data that has been linked by staff at DHHS to the usage records of 12 different health and human services administered by DHHS. Some of these services are targeted specifically at SES disadvantaged groups via means testing criteria.

The data set covers a 3-year period from 2013/14 to 2015/16 and contains about 7.5 million admission episodes. Of these, approximately 1.7 million episodes were from patients who were also users of one or more of the 12 health and human services. After restricting to the list of chronic diseases and applying our exclusion criteria (described below), we arrive at a sample of 237,743 admission episodes.

The utilization of specific health and human services administered by DHHS forms the basis of our SES measure. DHHS administers a range of services for the disadvantaged segment of the population in the state. These services include family assistance (e.g., family violence, child protection), low-income assistance (e.g., concessions for energy and water bills, finance hardship support), housing assistance (e.g., low-income housing support, crisis accommodation), etc. Whether a service is considered to indicate low, moderate or high SES disadvantage is determined by DHHS staff using an in-house algorithm. For privacy reasons, the precise nature and type of services included, and the algorithm used in constructing the SES measure are not revealed to the researchers. However, it is reasonable to speculate that services that are subjected to means testing are key elements in determining SES disadvantage.

Patients in the sample are classified into SES Groups by DHHS staff as follows.

  • SES Group 0: No SES disadvantage—non-user of human services.

  • SES Group 1: Low SES disadvantage—user of services not linked to SES or SES specific, not explicitly targeted at low SES users, or not extremely rationed, but still more likely to be used by disadvantaged individuals.

  • SES Group 2: Moderate SES disadvantage—user of services that are universal and not SES specific, but is explicitly targeted at SES disadvantaged individuals, or is rationed to the extent that high SES individuals were unlikely to use.

  • SES Group 3: High SES disadvantage—user of services designed for a prima facie SES issue, or is restricted exclusively to disadvantaged individuals.

Methods

We construct the sample by identifying disease spells. We have in mind a stylized disease progression time-line as depicted in Fig. 1a, where typically utilization is low during the initial period but begins to rise as the disease progresses. The figure shows the utilization pattern across the complete chronic disease spell, \(t_0\) to T, which can last many years and is seldom captured in its entirety in most data sets. Instead, what we intend to capture is utilization and adverse outcomes during the first year of the spell, \(t_0\) to \(t_1\), as shown in Fig. 1b. The key is therefore in identifying the start of a spell.

Fig. 1
figure 1

Health care utilization and chronic disease progression

Full size image

A patient is said to begin with a spell of a chronic disease if: (i) the patient has an admission episode with a principal diagnosis code in one of the chronic diseases listed in “Appendix 1”; (ii) during the previous 360 days (or longer), the patient was not admitted for that chronic disease (i.e., the principal diagnosis code not indicating the particular chronic disease). The assumption is that a fresh spell is deemed to form if there are no previous admission episodes for at least a year. Patients who have other complex conditions such as HIV, transplants, trauma, etc. are excluded, so were those who died during the sample period, were too young, or too old. The complete list of exclusion criteria can be found in “Appendix 1”.

For each disease spell, six outcomes are measured: (i) total hospital costs, (ii) total length of stay (LOS), (iii) number of admission episodes (iv) number of readmissions within 28 days, (v) number of admission episodes with hospital-acquired complications as measured by Classification of Hospital Acquired Diagnoses (CHADx), and (vi) number of potentially preventable hospitalization (PPH) episodes as measured by Ambulatory Care Sensitive Conditions (ACSC). All measures are taken over the entire spell, which can contain multiple admission episodes. The first three measures capture utilization while the last three measures are attempts to measure adverse in-hospital outcomes. Total hospital costs are obtained from cost information reported by hospitals; they include all fixed and variable costs attributed to an admission episode. These costs are routinely collected by DHHS for the purpose of administering the casemix funding system. Since costing information is not collected for private hospital admissions episodes, these episodes are assigned a cost of zero. We conduct a sensitivity analysis (discussed below) in which we drop spells that contain private hospital admissions and find that results are broadly similar.

We estimate linear regression models for all outcome measures, allowing for hospital fixed effects using dummy variables, which account for unobservable differences across hospitals. Note that the hospital dummy refers to the hospital where the first admission episode occurred. Additional covariates, listed in “Appendix 1” together with summary statistics, include personal characteristics such as age, gender, marital status, location of residence, as well as clinical characteristics such as total ICU hours, number of diagnoses at the beginning of the spell, and so on.

For the regression models, the first three outcome measures are logarithmically transformed to accommodate the skewness of the distribution; the remaining three measures are in untransformed raw scale. Given that a spell can comprise multiple admission episodes, some covariates, e.g., marital status and hospital campus, may change during the spell. We determine the value of these time-varying covariates at the beginning of the spell, i.e., from the first admission episode of each spell. We also check the robustness of our results by estimating nonlinear models; we use generalized linear models (GLM) for measures of utilization and count models for measures of adverse in-hospital outcomes. The distribution families under GLM models are determined using the modified Park test (see, e.g., [13]).

We further extend the analysis to examine the direct and indirect effects of SES. It is plausible to assume that hospital costs are affected by, among other factors, adverse in-hospital outcomes, which can drive up the hospitalization costs of a spell. As such, SES can affect hospital costs indirectly through its effects on these adverse outcomes. The total effects of SES can thus be decomposed into direct and indirect effects, with the latter estimated using standard techniques of mediation analysis [1, 23]. The analysis involves two steps. First, we estimate the effect of adverse outcomes on total hospital costs through a linear regression of hospital costs on adverse outcomes while controlling for other covariates.Footnote 1 Next, we obtain the effect of SES on each adverse outcome through a series of linear regressions of adverse outcomes on SES (again controlling for other covariates).Footnote 2 The indirect effect of SES through an adverse outcome is the product of the effects obtained from the first and second regressions. The standard errors of the estimated effects are obtained by bootstrapping [23].

Results

The sample consists of 237,743 disease spells in 12 chronic diseases. The frequency distribution by disease group is shown in Table 1. The largest disease group is digestive system disorders, accounting for nearly half the sample, followed by heart disease at 14.5%. The smallest disease group is rheumatoid arthritis, accounting for only 0.7%.

Table 2 presents the sample mean and standard deviation of the six outcome measures by SES group. The mean values show that greater SES disadvantage appears to be associated with higher utilization and worse adverse outcomes. It is worth noting the standard deviation values are generally two to three times larger than the mean values, suggesting a high degree of variations across SES groups.

Table 1 Chronic disease patient spells, frequency distribution
Full size table
Table 2 Outcome Measures by SES Group, Mean and standard deviations
Full size table

In relative terms, the differences in mean values between no SES disadvantage (SES Group 0) and moderate (SES Group 2) or high SES disadvantage (SES Group 3) groups are large. Relative to patients with no disadvantage, moderate and high disadvantage patients incurred respectively about 55% and 35% higher total costs, 129% and 81% longer LOS, 60% and 41% higher number of admission episodes, more than twice the number episodes with 28-day readmissions, complications and incidences of PPH. Note however these figures are unadjusted for risk factors, which are expected to account for some of the differences.

We employ a number of covariates in the regression analysis to allow for differences in personal characteristics and case complexity of patients. A list of all covariates with summary statistics and by SES group is given in “Appendix 1”. Generally, patients in moderate and high disadvantage groups are older, more of them are widowed or divorced, living in major cities (where social services are more widely available than in regional and remote areas), and with more complex health conditions.

The regression coefficients on SES Groups are presented in Table 3. The full list of all coefficient estimates (except hospital dummies) for all models can be found in Online Appendix B. Since a logarithmic transformation is applied to hospital costs, LOS, and number of admission episodes, the coefficient estimates on SES Groups can be interpreted as a percentage change over the reference category.

Table 3 Estimated SES effects
Full size table

The estimates suggest that, compared to patients with no SES disadvantage, low, moderate and high disadvantage patients incurred about 4.5%, 18.8% and 16.8% higher costs on average during the spell. In dollar terms, these figures translate to additional costs of about $475, $1983 and $1772 on average. The coefficient estimates on LOS and number of admission episodes suggest that low, moderate and high disadvantage patients tended to stay respectively about 3.7%, 20.5% and 16.0% longer and had 6.6%, 16.1% and 14.8% more admission episodes on average than patients with no SES disadvantage. All estimates reported above are statistically significant (\(p<\) 0.001).

For adverse outcomes, since the measures are in raw scale, the SES estimates can be interpreted in the original measurement unit. For the number of 28-day readmissions, admission episodes with complications and PPH, patients with low SES disadvantage were found to be statistically no different from those with no SES disadvantage. In contrast, the effects are highly statistically significant (\(p<\) 0.001) for moderate and high SES disadvantage patients. Compared with patients with no disadvantage, patients with moderate and high disadvantage were found to have respectively 0.07 and 0.06 higher number of 28-day readmission episodes. These same group of patients were found to have, respectively, 0.07 and 0.05 higher number of episodes with complications, and about 0.1 and 0.08 more episodes with PPH than patients with no disadvantage. The magnitude of the estimated effects is large relative to the respective sample mean as reported in Table 3. In relative terms, these additional adverse events range from about 27% to 99% of the mean values of the group with no SES disadvantage.

Direct and indirect effects of SES on hospital costs

We next examine how SES contributes, directly and indirectly, to total hospital costs. The indirect effects are constructed from two sets of regression models: (i) Regressing hospital costs on adverse outcomes; and (ii) regressing each adverse outcome on SES. All regression coefficient estimates are listed in “Appendix 1”. Table 4 presents the indirect effect estimates with standard errors obtained by cluster bootstrapping with 500 replications.

Table 4 Estimated indirect effects of SES
Full size table

As expected, the estimates are highly statistically significant for patients with moderate and high SES disadvantage, but not statistically significant for those in the low disadvantage group. Among the three adverse outcomes, complication appears to have the largest indirect effects, followed by 28-day readmission, with PPH having the lowest indirect effects. The sum of all indirect effects amounts to 0.0498 and 0.0398 for, respectively, the moderate and high disadvantage groups, and are highly statistically significant. These indirect effects represent approximately a quarter of the total effects shown in Table 3.

Sensitivity analyses

Several alternative models are considered for sensitivity analyses with results collected in Online Appendix B. In estimating the base models, a potentially complicating factor is the presence of private patient admission episodes in some spells. This is especially problematic for calculating total hospital costs because no costing information is collected for private patient admission episodes. This omission could systematically bias the results since SES disadvantaged patients are less likely to be admitted as private patients. As a sensitivity check, we re-estimate the models by removing spells with private patient admissions.

A second sensitivity check is to use an alternative measure of SES in the form of the SEIFA index in place of our SES measure. We also note that patients with digestive system disorders account for about half of the sample. In view of the dominant share, we perform a sensitivity analysis by dividing the sample into two halves, one consisting of only patients with digestive system disorders and the other half consisting of all other patients. The results are consistent across the two halves of the sample. A further sensitivity check is to estimate hospital random-effects models instead of using fixed-effects estimation. We find almost identical results whether random effects or fixed effects are used.

Another sensitivity check is to include additional observations removed by exclusion restrictions. Specifically, in this variation we include observations due to HIV and cancer patients, mental health patients, and patients with multiple chronic diseases. We find the resulting estimated SES effects to be similar and in most cases slightly higher in magnitude than before.

The linear model specification is ill-equipped to deal with count variables and measures that are highly skewed. As alternative specifications, we estimate non-linear models using generalised linear models (GLM) with different specifications of distribution families and link functions. Total costs and LOS are estimated using a Poisson distribution with log link, number of admission episodes is estimated using inverse Gaussian distribution with log link, and the remaining three outcome variables are estimated as count models using a negative binomial distribution (with the identity link function). These specifications are not new and have been extensively discussed in the literature, see, e.g., Jones et al. [12], Hill and Miller [11], and Jones [13]. From the marginal effects of SES, we find similar results to those from our linear model specifications.

Discussion

This study finds that patients with moderate and high SES disadvantage, in comparison to those without SES disadvantage, tend to incur additional hospital costs and longer LOS of about 20% more, and higher number of admission episodes by about 16%. These patients also have more adverse outcomes in hospitals—ranging from 27 to 99% more incidents than other patients. These effects are not only highly statistically significant but are also large in magnitude. We also note the gradient effect in our estimates does not extend beyond moderate SES disadvantage—the estimated SES effects do not appear to be greater for patients with high SES disadvantage compared to patients with moderate advantage. For all outcomes, the estimated effects are smaller for patients with high SES disadvantage, although in most cases the differences are not statistically significant at 5% level. We believe this pattern is related to access barriers facing patients with high SES disadvantage. Although under Australia’s universal health insurance system, hospital care is provided free of co-payments to all Australians admitted to public hospitals, there can still be indirect monetary and non-monetary access barriers such as travel costs, waiting times, and information and search costs in relation to the availability of care services. Some of these costs can present significant barriers for patients with high SES disadvantage such as the homeless and those with severe mental health issues.

Our results have implications on the funding of hospital care for chronic disease patients. Hospitals serving disproportionately large number of disadvantaged patients are likely to face financial shortfalls under the current casemix funding system, unless additional payments are allocated to hospitals serving SES disadvantaged patients. Recent policy developments in Australia indicate that capitation funding models are under consideration for funding chronic diseases. An example is the Health Care Homes program, which aims to provide coordinated care for patients with chronic and complex conditions. Our results suggest that allowance for SES disadvantage is even more important since chronic disease patients with SES disadvantage attract significantly higher costs and utilization, and suffer from a higher incidence of adverse outcomes than patients with no disadvantage. Given that capitation funding is typically set for a specified period of time, usually one year, not allowing for SES disadvantage would put hospitals serving disproportionately high number of disadvantaged patients under severe financial handicaps. In extreme cases, hospitals may have incentives to turn away or transfer disadvantaged patients elsewhere, i.e., cream skimming [4].

That SES disadvantaged patients have higher utilization than non-disadvantaged patients is partially explained by the former’s higher incidence of adverse in-hospital outcomes. In extending the analysis to estimate the mediating effects of adverse outcomes on hospital costs, we find that the indirect effects of SES through complications are the largest, followed by unplanned readmissions and PPH. This result suggests that removing SES inequality in the occurrence of complications would have the largest impact on reducing the discrepancy on hospital costs. Together the three adverse outcomes account for about a quarter of the total effects of SES on hospital costs. Hence, although these adverse outcomes are material and significant, tackling SES inequality in the occurrences of these adverse events would at best remove about a quarter of the estimated SES effects on hospital costs. The other three-quarters would have to be found in other factors.

Besides adverse in-hospital outcomes, other factors that can potentially cause the discrepancy in hospital utilization between SES groups include differences in education [26], extent of comorbidities [3], housing and environmental exposure [20]. Other possibilities include the availability of support services at home and in the community, and the quality of care received by different SES groups. Whether disadvantaged patients receive a lower quality of care than non-disadvantaged patients is a question that warrants further investigation, especially if hospital funding is to be adjusted to reflect the higher costs of caring for disadvantaged patients. The higher level of funding may provide a perverse incentive for hospitals to provide an ever lower quality of care for disadvantaged patients. We note, however, such incentives may be countered by other institutional features such as quality audits and the incentive structure inherent in capitation funding schemes.

Our approach of measuring SES relies on human services usage records and differs from other studies, which often use income-, education- or location-based SES measures. For studies relying on location-based measures such as SEIFA, a key limitation is all residents in a location are classified in the same SES group regardless of circumstances. This means that a patient can be classified into a different SES group simply by moving to a new location, even though the patient’s circumstances may not have changed at all. In contrast, our measure has the advantage of being individual specific. It is also more reliable in measuring SES disadvantage since many of the services are means tested and only accessed by the disadvantaged segment of society. Our measure also has advantages over income-based measures which can be problematic for measuring the SES of older individuals, e.g., retirees with low income but are well off in terms of assets.

Conclusion

This study investigates the extent to which SES disadvantage affects hospital utilization and adverse outcomes of chronic disease patients. It contributes to the literature by proposing an approach of identifying disease spells based on the disease progression time-line. It finds that SES disadvantage has large and statistically significant effects—disadvantaged patients are found to incur higher utilization and have more incidents of adverse outcomes. Further analysis finds that the estimated effect of SES on hospital costs can be attributed, partially and by about 25%, to the effects of SES on in-hospital adverse outcomes.

This study is exploratory in nature and has several limitations. With only 3 years of data available, identifying chronic disease spells is less than precise—essentially the identification can only rely on data for the first year. Ideally for chronic diseases one should be analysing patients with complete disease spells, i.e., from the beginning of the disease diagnosis to the end of treatment (e.g., complete recovery or death). However, this is seldom feasible for most data sets. Data limitations also prevent us from investigating utilization beyond hospital care. For example, whether hospital care is substituted for primary care by SES disadvantaged patients, e.g., Terraneo [26]. If this holds, our estimates would be biased against non-disadvantaged patients since their greater reliance on primary care is not accounted for in our data.

Another potential source of bias is the possible endogeneity of SES, in that unobserved shocks to individuals’ health may affect health care utilization and SES. We note that although the use of health and social services may to some extent reflect individual health conditions, in cases such as housing assistance, family violence, disability supports, etc., it is likely unrelated to hospitalization due to chronic diseases. Moreover, a large number of these services are subjected to means testing, which is unrelated to the severity or complexity of patients’ health conditions.

Finally, while the estimation allows for differences in health care needs through the covariates as far as the data permit, no allowance is given for differences in preferences. We assume, as in previous studies, that any observed differences are not the result of free choice (e.g., [7]).