Study setting and health care delivery system
Australia has a publicly funded universal health care scheme known as Medicare. Residents are entitled to subsidised treatment from private medical practitioners, and for some services from nursing and allied health professionals. Australians can obtain free treatment in public hospitals, and private health insurance is available for patients preferring private services in hospital. Under the broad umbrella of Medicare there is also a Pharmaceutical Benefits Scheme (PBS), which provides subsidised prescription drugs. An estimated 80% of all prescription medicines dispensed in Australia receive subsidy via the PBS [28]. There is a concessional and a general co-payment rate, and also a safety net so that when a patient reaches the threshold their PBS patient contribution is reduced or removed [29].
Participants
A validated questionnaire was mailed to a representative cross-section of the membership (n = 10,000) of National Seniors Australia during mid-2009. National Seniors Australia is a nation-wide organisation with 285,000 members aged over 50 years. A stratified sampling procedure by age, rurality and state of residence was applied.
Respondents were asked ‘Has a doctor ever told you that you had any of the following illnesses?’ This was followed by the list of 11 conditions and allowed for other conditions to be reported under ‘other chronic condition’. While information was collected on all conditions that lasted more than six months, the listed conditions were cancer, heart disease, high blood pressure (HBP), diabetes, arthritis, osteoporosis, asthma, bronchitis, Parkinson’s disease, depression and anxiety. As Parkinson’s disease had a very low prevalence (<2.0%), it was excluded from clustering.
The data collection method is described in detail in McRae et al. [17]. The survey and study were approved by the Australian National University Human Research Ethics Committee (no. 2009/309).
Out-of-pocket expenditure
OOPE was defined as the total amount of own money respondents spent on both medical expenses and nonmedical expenses (e.g. transport, home care) related to care processes pertinent to healthcare. In Australia, home care includes domestic assistance, personal care and respite care and depending on individual needs may include services such as meals, transport, shopping and home maintenance [30]. Respondents were asked to report their personal OOPE during the previous three months under the main categories of health-related services, including medication, medical services, transport, medical equipment, home care and other expenses. Health insurance premiums were not included, because the focus of this study was to measure the financial burden that is directly related to out-of-pocket costs for medical care. Respondents reporting ‘do not know’ to any category (15%) were omitted from calculations of total costs. As extreme expenses such as those for housing modifications (one observation over $20,000) and very expensive hearing aids had the potential to significantly influence estimates, observations with quarterly costs of $5,000 or over were excluded when estimating total costs (removing 26 observations or ~1% of observations reporting total expenditure).
Comorbid groups and prevalent pairs
Cluster analysis was used to establish the “natural groups” of chronic conditions – specifically, a partitional cluster analysis was undertaken using k-medoids and Yule’s Q similarity measure [27]. For modelling purposes, participants with none of the ten conditions were classified as the reference cluster.
From a disease management perspective in clinical practice [31] our clinical and content expert identified a set of clusters that, unlike partitional clusters, are not mutually exclusive. Thus a participant may belong to more than one clinically suggested cluster depending on the type of chronic conditions. Here again participants with no diseases become the reference group.
As an alternative means of addressing co-morbidity, following Schoenberg et al. [16]’s approach we created a categorical variable labeled ‘multiple morbidity’. Among those individuals with only one chronic illness HBP and arthritis were the most frequently occurring, so we categorised individuals with one condition into three sub-groups: those with HBP only, those with arthritis only, and those with only one condition but not arthritis or HBP. Participants with only two conditions were divided into five sub-groups: HBP + arthritis, HBP + diabetes, HBP + heart disease, arthritis + asthma, all other combinations of only two conditions. We constructed three sub-groups with the participants with only three conditions: HBP + arthritis + cancer, HBP + arthritis + diabetes, and all other combinations of three conditions. Although we could have used more sub-groups with two or three condition combinations, we did not go any further as the prevalence of such combinations became increasingly very small. This categorical variable ‘multiple morbidity’ contained all the above groups together with a category for those with four conditions, and a category for those with more than four conditions. The group with HBP only was selected as the reference group as it was the most common condition that tends to be associated with other conditions.
For many conditions the cost faced by a single patient with two conditions may be different from the sum of having the same two conditions separately [32], a phenomenon known as interaction, which then modifies the outcomes. The modified effect could be greater (positive interaction, synergism) or less (negative interaction, antagonism) than simple addition of the two effects [33]. To see whether there is any interaction due to having two particular conditions we used the most prevalent pairs (because there are 45 combinations of ten conditions, many of which are rare, we use conditions pairs with observed prevalence of ≥ 5%) along with the individual conditions in the regression models to examine the effect modification of these combinations.
Other variables
A number of covariates were considered during analyses selected from analytical domains that previous studies have shown to be associated with OOPE [17]. These included socio-demographic variables such as age, sex, income, physical and mental health status reflected by SF-12 [34], region and number of ‘other chronic conditions’. The SF-12 is a widely used 12-item measure of health-related quality of life. Items are summarized into two weighted scales representing perceived impairment in role functioning associated with physical and mental health problems, with lower scores indicating greater impairment [34]. We included SF-12 measures to address participants’ physical and mental health status – as a proxy for disease severity. Income was converted to ‘household equivalent income’ using the modified Organisation for Economic Co-operation and Development (OECD) equivalence scales which apply a scale of 1 to the first adult in a household, 0.5 to the second and later adults, and 0.3 to children [35].
Financial burden
As well as analysing OOPE, it is important to understand which groups of people face the greatest financial burdens due to their healthcare costs. For the purposes of this study, a heavy financial burden was defined as expending over 10% of equivalised household income on OOPE. Although this percentage is necessarily somewhat arbitrary, it has been used by a number of previous studies [3, 7, 17, 36].
Approach to modelling
The distribution of the OOPE variable contains an abundance of zeros (30% of those who responded) and a highly skewed distribution of nonzero values. There are in practice two processes occurring – one which establishes a requirement to expend any OOPE on health matters, and second process which establishes the size of the OOPE conditional on it being non-zero. One approach to handling this data is to undertake two sets of regression: firstly a logistic regression exploring the probability of incurring OOPE, and secondly a linear regression of how much is spent with the subset who reported more than zero expenditure, after logarithmic transformation of that subset. These two regressions are then interpreted separately.
Alternatively, the two processes can be estimated jointly using an approach based on a parametric mixture distribution [37], which addresses both the abundance of zeros and the skewed distribution of non-zero values in the same model. Both of these approaches are known as two-part models, and the latter is also known as two-part joint regression model [38]. For the joint regression model we used STATA tpm command [37]. In order to show the nature of the two distinct processes as well as the overall process, in this article we used both of these approaches as shown in Figure 1. This means we report on four groups of models – (i) whether a respondent has any OOPE, (ii) the amount of OOPE for those with any OOPE, (iii) the joint modelling of the two previous models, and (iv) models based on whether the respondents faced a heavy cost burden from their conditions.
Models estimated
For each of these four groups (i-iv mentioned above) a set of five models were estimated: model 1 estimated the association between the relevant measure of OOPE and aggregated number of chronic conditions; model 2 estimated the association between OOPE and specific chronic conditions; model 3 estimated the association between OOPE and ‘multiple morbidity’; model 4 estimated the association between OOPE and natural clusters; and model 5 estimated the association between OOPE and clinically relevant clusters.
To see the effect modification of one condition on another we also assessed the effect of prevalent pairs on both of the dependent variables. However, for the sake of parsimony we have chosen to report these results briefly in the text rather than in the tables.
Multicolinearity was assessed using variance inflation factors (VIF) and not found to be a problem. Models were compared using Akaike information criterion (AIC) and Bayesian information criterion (BIC). Data were analysed using STATA (version 12). To make the coefficients (β) of the models in group-ii regressions easily interpretable they have been exponentially transformed and reported as a value B. The interpretation is that a one unit (e.g. from zero to one) increase of independent variable would result in (B-1)*100 percentage change in OOPE.
The Mantel-Haenszel chi-square test for linear trend (χ2
trend) was used to assess whether the proportion of participants who spent over 10% of income as OOPE showed a trend in relation to their level of household income.