Determinants and health outcomes of trajectories of social mobility in Australia

Objectives To investigate trajectories in socio-economic position (SEP) and the onset of a range of physical and mental health outcomes and commencement of treatment. Methods The Household Income and Labour Dynamics Australia (HILDA) study, a nationally representative prospective cohort study over the period 2001 to 2020 was used to define trajectories of SEP. Trajectories of low, low-middle, upper-middle and high SEP and decreasing (low-middle to upper-middle SEP) or increasing (upper-middle to lower-middle SEP) SEP were identified using k-longitudinal means. Cox-regression was used to assess SEP trajectories and physical (arthritis or osteoporosis, any cancer, asthma, chronic bronchitis or emphysema, Type 1 diabetes, Type 2 diabetes, hypertension or high blood pressure, and coronary heart disease), and mental health (depression or anxiety) outcomes, and treatment commencement. Predictors of SEP trajectories were also investigated using multinomial logistic regression and random forests. Results Decreasing SEP had a higher relative risk of new onset illness than increasing SEP for all health outcomes. Increasing SEP had relative risk estimates that were more consistent with upper-middle income groups and decreasing SEP had a relative risk consistent with lower-middle income groups. In contrast, there was no socio-economic gradient in treatment commencement for physical health outcomes, or depression or anxiety, with the exception of arthritis or osteoporosis. Conclusion Decreasing SEP was associated with poor health outcomes, and increasing SEP with better health outcomes. A range of socio-demographic and psychosocial determinants of SEP trajectories were identified to inform policy responses that could modify trajectories of health inequalities in the Australian context.


Introduction
Socio-economic position (SEP) is a major determinant of poor health and social outcomes. SEP is determined by socio-economic, cultural and political context, and affects health via material, psychosocial and biobehavioural factors (World Health, 2010), and can have effects over the life course through latent effects, specific pathways or trajectories, accumulation of effects, or via social mobility (Ben-Shlomo & Kuh, 2002;Lynch & Smith, 2005;Niedzwiedz et al., 2012;Pollitt et al., 2005). These outcomes often also follow a social gradient, where there are poorer health outcomes among those of lower (compared to higher) SEP, and these improve as affluence increases There is clear and consistent evidence of a social gradient of association between SEP and a range of outcomes including physical health (Marmot, 2005), mental health and substance use (de Oliveira et al., 2022), crime (Ljujic et al., 2017), COVID-19 hospitalisations and mortality (Eva, 2022) and educational attainment. Reliable statistical evidence for the effect of SEP on educational achievement, dates as far back as 1967 (Sewell VPS, 1967).
Social mobility relates to inter-and intra-generational changes in SEP, based on measures of income, occupational status, and/or educational achievement (Patricia & Rachel, 2020). Recent evidence shows that downward social mobility is associated with a range of health and behavioural outcomes, including depression and other mental health outcomes (Euteneuer & Schafer, 2018;Melchior et al., 2018),hypertension (Lopes et al., 2021) and, acute coronary syndrome. (Burazeri et al., 2008), lower leisure time physical activity (Elhakeem et al., 2016), and increased tobacco consumption (Motta et al., 2015). Improving social mobility for discriminated groups in a community can lead to lower cause-specific cancer mortality for the whole community (Poulson et al., 2022).There are also differential impacts of social mobility on health among migrant groups (associated with migration to lower relative SEP settings and under-employment) (Das-Munshi et al., 2012) and different points across the life course (Fu et al., 2022;Niedzwiedz et al., 2012) Much of this evidence arises from population cohort studies including the French nationally representative CONSTANCES study of 67,057 individuals (Melchior et al., 2018) or the European Social Survey of 52,773 individuals (Gugushvili et al., 2019) or the ELSA-Brasil study of 8754 participants (Lopes et al., 2021). These are high quality prospective cohorts with follow-up over many years.
In the Australian context, there has been extensive previous research investigating health inequalities based on socio-economic position (Mech et al., 2016;Shepherd et al., 2012;Williams et al., 2013), including for a range of chronic disease outcomes, health service access, and trends in socio-economic differentials over time. However, there has been limited investigation of how trajectories in social mobility (that is, changes in SEP) over the life-course affects health outcomes and health service access, and whether these trajectories are specific to particular health outcomes or are consistent across health outcomes.
Accordingly, this research aims to investigate the association between SEP and the impact of upward and downward social mobility on the onset of a range of (i) physical health outcomes (obesity, cardiovascular disease, diabetes, respiratory disease, and cancer), (ii) mental health outcomes (psychological distress, depression or anxiety) and (iii) associated treatment commencement, based on a large representative prospective cohort study of Australia (Roger Wilkins and BothaSarah, 2021). An additional aim was to identify socio-demographic and psychosocial determinants associated with changes in SEP (social mobility). Potential determinants included a wide range of time-varying variables, for example educational achievement, unemployment, marital status, alcohol consumption, body-mass index (BMI), Kessler 10 distress score, among others. This research can inform policy responses in both preventing downward social mobility as a distal factor associated with poor health, and in ameliorating the impact of falling or low SEP on new onset serious chronic diseases.

Study setting
This study uses a prospective cohort study design based on the nation-wide Household, Income, and Labour Dynamics Australia (HILDA) survey designed by the Melbourne Institute of Applied Economic and Social Research at The University of Melbourne and funded by the Australian government through the Department Of Social Services. The HILDA cohort is a representative longitudinal cohort of people residing in Australia for the period 2001 to 2020 (N = 34,573 people ever followed up and 22,932 people at last follow-up). The HILDA study data consists of self-reported questionnaires on domains including family background, health, education, income and wealth, housing, relationships, and employment over 20 waves of follow-up. A detailed description of this cohort is presented elsewhere (Roger Wilkins and BothaSarah, 2021). The data dictionary, summary frequencies and response rates for each variable and across each wave from wave 1 to 20, is also documented elsewhere (Melbourne Institute, 2022). Participants enter and leave the cohort, with an average age of 48.5 years.

SEP and social mobility
The main exposure of interest was SEP, and how SEP changed over the course of follow-up in the HILDA cohort. Categories of social mobility were based on household income (adjusted for inflation over time), collected for seventeen waves of follow-up, and derived empirically using the K-means longitudinal clustering approach (Genolini et al., 2015). This approach grouped similar trajectories of people in the cohort into clusters based on the principles of K-means. Six clusters of similar proportions of the cohort (12%-22%) were generated automatically (not a priori) based on un-supervised learning from patterns of trajectories in household income using the R-package kml. For each cluster 60%-80% of participants had house-hold income recorded for the full 17 years of follow-up (Genolini et al., 2015). Five or six clusters were recommended to optimise various information criteria such as the Calinkski-Harbatz coefficient, AIC, BIC, and Ray-Turi index. Six clusters were identified as interpretable clusters of SEP in this study (Fig. 1). The K-means longitudinal method identified 4 clearly delineated clusters across a range of stable SEPs, and 2 clusters of rising or falling SEP, over time (Tables 1 and 2, Fig. 1). Kml requires that there is no missing data for anyone in the analysis, so household incomes were carried forward for a complete case analysis without extrapolation. Imputation of data from the HILDA cohort has already been carried out by the data custodians at the Melbourne Institute (Breunig, 2022). There were four stable groups (that is, where participants did not change their SEP) who were defined as 'High Income' ([18%, N = 2780]), 'Upper Middle Income'([22%,N = 3593]), 'Lower Middle Income'([15%, N = 2451]) and 'Low Income' ([16%, N = 2830]). There were also two groups where participants' SEP changed over time, defined as changes from higher to lower SEP (that is, 'decreasing' social mobility, N = 2003, or 12% of the cohort), and from lower to higher SEP (that is, 'increasing' social mobility, N = 2530 or 17% of the cohort).
For the derived social mobility groups, decreasing SEP commenced at the upper middle SEP category and decreased to the lower middle SEP category. In contrast the increasing SEP group started at the lower middle SEP category and increased to the upper middle SEP category ( Fig. 1).

Outcome variables
All of the mental health, physical health and treatment initiation outcomes were based on questions like: "Have you ever been diagnosed with XXXX serious illness?"

Onset of mental health outcomes
Mental health outcomes included self-reported diagnosis of depression or anxiety, and psychological distress. Psychological distress was measured as a time-varying variable using the Kessler 10 (K10) questionnaire score. The K10 score is measured every alternative wave from Wave 7 and was carried forward between waves. K10 is higher with rising distress and is a marker of allostatic stress over time. The K10 contains ten questions on symptoms of anxiety and depression. Note however, that K10 is not diagnostic of depression or anxiety. Onset of depression or anxiety was based on responses to questions in wave 13 and wave 17 of follow-up among participants who previously did not report the outcome (as described below).

Onset of physical health outcomes
Physical health outcomes included self-reported past clinical diagnoses of arthritis or osteoporosis, any type of cancer, asthma, chronic bronchitis or emphysema, Type 1 diabetes, Type 2 diabetes, hypertension or high blood pressure, and coronary heart disease. BMI was recorded annually from Wave 6. Onset of physical health outcomes was based on responses to questions in wave 13 and wave 17 of follow-up among participants who previously did not report the outcome at Wave 9 (as described below).

Onset of treatments
Additional outcome variables included initiation of treatments associated with the physical and mental health outcomes described above, each as separate variables. Onset of initiation of treatment variables were recorded only in wave 9, wave 13 and wave 17 of follow-up as either 'yes' or 'no'.

Other study factors
A range of other socio-demographic, health and wellbeing factors that were likely associated with the identified social clusters were also identified and included as potential confounders in analyses (a crosssection of the study at wave 9 is presented in Tables 1 and 2).
Descriptive data for wave 9 are presented as this was the point at which new onset physical or mental illness at treatment association rates were defined for survival analyses (as described below). Socio-demographic factors included age at when the participant entered the study, gender ('male', 'female' -gender at entry to the cohort), indigenous status ('Aboriginal, Torres-Strait Islander', 'both Aboriginal and Torres Strait Islander', 'neither Aboriginal or Torres Strait Islander' -measured annually since Wave 1), highest level of educational achievement ('Masters or PhD', 'Graduate Diploma or Grad Certificate', 'Bachelor or Honours', 'Advanced Diploma', 'Cert III or IV', 'Year 12', 'Year 11 and Below', 'Undetermined' -annual wave 1), marital status ('married', 'separated but not divorced', 'divorced', 'widowed', 'never married but not in a relationship', 'never married and not with someone' -annually since Wave 2), and the percent of time in the previous year spent unemployed (annually since Wave 1). Dwelling type ('Free Standing House', 'Non-private dwelling -nursing homes', 'Non-private dwellingothers', 'Separate house with attached shop, office, etc', 'Semi-detached house with one storey', 'Semi-detached house with two or more storeys', 'Semi-detached house attached to a shop, office etc', 'Flat/unit/apartment in one-storey block', 'Flat/unit/apartment in two-storey block', 'Flat/unit/apartment in three-storey block', 'Flat/unit/apartment in four to nine-storey block', 'Flat/unit/apartment in ten or more storey block', 'Flat/unit/apartment attached to a house', 'Flat/unit/apartment attached to a shop, office etc', 'Caravan/Tent/Cabin/Houseboat'annually from Wave 2), and the Accessibility/Remoteness Index of Australia (ARIA) ('Live in a City', 'Inner Regional Australia'. 'Outer Regional Australia', 'Remote Australia', 'Very Remote Australia'annually from Wave 1), Speaks Language Other Than English at home -annually from Wave 2, was also included. Additional health and wellbeing factors included participant responses to questions relating to selfreported physical and emotional problems affecting social functioning (measured on a 5-point Likert scale from 1 to 5 -annually from wave 1), general health and wellbeing (measured on 5-point Likert scale from 1 to 5 -annually from Wave 1), weekly alcohol consumption ('I have never drunk alcohol', 'I no longer drink', 'I drink alcohol everyday', 'I drink alcohol 5 or 6 days per week', 'I drink alcohol 3 or 4 days per week', 'I drink alcohol 1 or 2 days per week', 'I drink alcohol 2 or 3 days per week', 'I drink only rarely' -annually from Wave 2), and significant lifeevents in the past year (including birth of a child, victim of physical violence, or incarceration in the householdall annually from Wave 2). Detailed descriptions of these variables can be found in the HILDA data dictionary (Melbourne Institute, 2022).

Data analysis 2.5.1. Health and treatment outcomes
Survival analysis using Cox-Proportional Hazard regression was used to assess the association between SEP trajectories and new onset of each physical health, mental health, and treatment outcome as described above. Participants were selected at Wave 9 if they did not report the described health outcome or treatment. The specific illness or treatment status was measured again at Wave 13 or Wave 17 and these waves indicated the survival outcomes, allowing for new onset health outcomes to be identified.
Confounders were classified as fixed at baseline, or time varying covariates if they changed over time, for each of the outcome variables of interest. Fixed confounders were age at entry to the study, gender, and whether the person spoke a language other than English. Time-varying confounders included percent of time unemployed in the last financial year, urban-rural residence, marital status, highest level of education, indigenous status and alcohol consumption.
Random effects analyses (where the individual person identifier was the random intercept) was used to assess the temporal association between SEP category and BMI (continuous random effects regression) and psychological distress as measured by the discrete K10 score (random effects negative binomial regression) with the same approach to fixed and time-varying confounder adjustment as described above.  Tables 1 and 2 corresponds to year 6 (Wave 9).

Factors associated with SEP and social mobility
Additional analyses also investigated those socio-demographic, health and wellbeing factors associated with SEP and social mobility. A multinomial logistic regression model was developed where SEP clusters were specified as the outcome and age, gender, sociodemographic (dwelling type, highest education qualification, marital status, percent of time unemployed, remoteness of residence (ARIA), English as the language spoken at home, First Nation's Status, life events in the last year, (member of the family jailed, victim of physical violence, birth/adoption of a new child), general health and wellbeing question (1-5), social functioning given health status (1-5), distress score (K10), body mass index were specified as exposures, with individuals defined within each cluster. Decreasing SEP, increasing SEP, high, low-middle, low, were compared to upper-middle SEP as the reference category. Upper-middle SEP was chosen as this income trajectory was the origin of decreasing SEP and destination of the increasing SEP.
An explanation of the equations is given for multinomial logistic regression: In is the probability of a participant being in the uppermiddle SEP, which is the Kth or sixth cluster. Each other cluster is assigned a value between 1 and K-1 or one to five. The variables or columns in the dataset are represented by x i . The corresponding β K-1 vector or vector of coefficients exists for each of the other five (K-1) clusters where the reference group is the upper-middle SEP cluster. These β ′ s are exponentiated to give a relative risk. In this study this model was extended where each individual's information across all the waves were used as clusters in a multinomial logistic regression random effects model.
Finally, an atheoretical prediction model was also constructed to classify people into categories of SEP or social mobility. The aim of this analysis was to predict those who experienced a decreasing SEP. The cohort was divided into a 'test' dataset (20% of the cohort) and a 'training' dataset (80% of the cohort). Seven techniques were screened to identify the technique with the lowest cross-over error rate, including random forest (55.3%), multinomial regression (68.2%), regularised multinomial regression (68.2%), k-nearest neighbours (68.9%), neural nets (62.9%), naïve Bayes (78%) and supported vector machines. The random forest technique had the lowest cross-over error rate and is presented in the current study. Also, the random forest technique is most adaptive to models where there is a combination of categorical and continuous predictors and multi-class outcomes, especially in large datasets (Fu et al., 2022). The random forest technique combines a range of decision trees with continuous variables classified at different cut-points. The combination of these decision trees is conducted randomly while optimising desirable model features such as error, strength of individual trees in the random forest and correlation between trees in the random forest (Breiman, 2001).
The random forest technique was optimised in R. Based on the data, an optimal mtry value of 8 was calculated and used for the random forest model. The mtry defines the number of variables randomly selected as candidates at each split. The variables used were ranked in order of the mean decrease in the Gini co-efficient and mean decrease in accuracy when the corresponding variable was removed from the model. The random forest model was evaluated using the measures of sensitivity, specificity, positive predictive value, negative predictive value, prevalence, detection rate, detection prevalence and balanced accuracy.

Missing data
Information on health outcomes were first measured from wave 9 of the HILDA survey. In the HILDA cohort there were 17,632 people enrolled at Wave 9. This system of classification of household income level categories using k-longitudinal means and a minimum of 7 years follow-up classified 16,187 people out of 17,632 (92%).

SEP and social mobility and new onset chronic disease
An increasing socio-economic gradient (from highest to lowest SEP) was evident for each of the included health outcomes (Fig. 2) For all health outcomes, those participants with decreasing social mobility (from higher to low SEP) generally had higher relative risk of illness than those with increasing social mobility, as compared to the high SEP referent category, with the exception of asthma and chronic bronchitis of emphysema (Fig. 2). Participants with decreasing social  mobility (from higher to lower SEP) had relative risk estimates that were more consistent with low-middle SEP groups. Participants with increasing SEP (from low-middle to upper-middle SEP) had relative risk estimates that were more consistent with upper-middle income groups (Fig. 2). In contrast, there was no evidence of a socio-economic gradient in self-reported treatment commencement for all health outcomes (Fig. 3), with no difference between lower SEP, or between increasing or decreasing SEP groups, compared to the high SEP referent category. The exception was for arthritis or osteoporosis treatment commencement, where those of lower SEP and either decreasing or increasing SEP were more likely to have commenced treatment (Fig. 3). Additionally, those with lower SEP and those of decreasing SEP, were less likely to visit a mental health clinician, compared to the high SEP group (Fig. 3).

Determinants of falling SEP
The results of the full multinomial regression model are provided in Supplementary Materials Table 1. Of particular interest was the comparison between upper-middle SEP and decreasing SEP groups (Table 3). Factors associated with decreasing SEP included older age; being female, general health and wellbeing, social functioning given health, highest level of educational qualifications, not being in a marital relationship, period of unemployment in the last year, remoteness of residence, being a victim of physical violence and not living in a free-standing house. Variables used in the full multinomial regression model were not highly correlated with the highest correlation coefficient at r = − 0.61 for K10 (distress) and social functioning given health (not sown).
More generally, the random forest model found that the seven most important variables that classified participants into one of the SEP or social mobility groups were BMI, age, marital status, psychological distress score, alcohol consumption, education level, and remoteness of residence gradient (Fig. 4). The balanced accuracy ranged from 89% for low SEP to 77% for decreasing SEP. The specificity (96%) and negative predictive values were very high (94%), but sensitivity (57%) and positive predictive values (69%) were modest. (Table 4) This implies that people not satisfying the criteria for decreasing SEP are unlikely to experience decreasing SEP, whereas there is a high level of uncertainty for people satisfying the criteria for decreasing SEP as to whether they actually experience decreasing SEP.

Discussion
This study investigated the association between SEP and social mobility (that is, change in SEP), and a range of physical health, mental health and treatment outcomes, in a population-based prospective cohort study in Australia. Overall, there was an increased risk of new onset chronic disease among lower SEP groups, evident for all reported conditions. Those experiencing rising or falling SEP had a risk of new onset chronic disease similar to their final SEP, rather than their initial SEP. That is, the risk of new onset chronic disease outcomes for the decreasing SEP group was similar to the lower middle-income SEP group, and not the upper middle SEP group. Similarly, those experiencing increasing SEP had a similar risk of new onset chronic disease outcomes to the upper middle SEP group, rather than the lower middle income. This pattern was evident for both physical health outcomes (with the exception of Type 1 diabetes and any cancer diagnosis) and mental health outcomes. No previous studies in the Australian context or globally, have shown the consistent effects of social mobility over wide range of physical and mental health outcomes over multiple decades in this detail.
Additional analyses suggested that older age, poorer social relationships, psychological distress, lower educational achievement and periods of unemployment, and type and area of residence were factors associated with a decreasing trajectory of SEP. The transition period of decreasing SEP was also found to be important in predicting the onset of the disease outcome, baseline 'protective' factors such as educational qualifications, diet and lifestyle, before the decline in SEP may not be preventive of subsequent illness. This suggests that it is current SEP which is associated with health outcomes rather than past SEP.
In contrast, there was limited evidence of a socio-economic gradient in self-reported initiation of treatment relating to both physical health and mental health outcomes. Overall, there was no substantial difference in treatment initiation between the lower SEP group, or between increasing or decreasing SEP groups, compared to the high SEP referent category, with the exception of treatments for arthritis or osteoporosis, and treatment for poor mental health. For arthritis or osteoporosis treatment commencement, those of lower SEP, and either decreasing or increasing SEP, were more likely to have commenced treatment earlier; whereas those with lower SEP and those of decreasing SEP, were more likely to visit a mental health clinician later. The less prominent social gradient for the initiation of treatment may reflect a health system that is nominally universal (under the Australian Medicare system), and which is broadly accessible across socio-economic position. However, this finding is not consistent with other Australian studies that show socioeconomic differences in health service access by measures of other markers of socio-economic position such as socio-economic area of residence (Robards et al., 2019) and remoteness (Carey et al., 2013). The role of transitions in educational achievement, employment status and occupation and access to health care in the Australia context is less clear and is an area for future research.
There are several methodological considerations when interpreting the findings from the current study. Firstly, establishing the extent to which SEP is a putative 'cause' of a given health outcome is difficult to ascertain, and there are clearly bi-directional associations between SEP and health as has been previously noted (Hoven & Siegrist, 2013), and evidence of both social causation and social selection of health outcomes. However, a strength of the current study was the longitudinal study design, based on a representative sample of the Australian population. Analyses could clearly delineate temporal associations between a given SEP, or level of social mobility, and incident health outcomes allowing a determination of how SEP, or trajectory of SEP, was associated with the subsequent onset of selected disease outcomes. The HILDA cohort also contains a large set of variables to document the breadth and depth of social variables with some focus on health variables. This allowed for the incorporation of a range of confounders associated with social determinants of incident health outcomes.
Secondly, the definition of SEP and social mobility in the current study was limited to household income level. Income was selected as the main exposure measure given that it was asked of study participants at every follow-up for the last seventeen waves, allowing the construction of detailed trajectories over time. A limitation of k-longitudinal means is that only one continuous variable can be us used to identify trajectories, so annual household income was employed.
There is also potential measurement bias in both the main exposure and outcome variables, as data on SEP and health outcomes were based on self-report. For income level it may be participants under-or overestimated their household income resulting in biased SEP-health outcome associations. Similarly, health outcomes were based on a selfreported diagnosis of selected outcomes (for example for diabetes, respiratory disease, and cardiovascular disease), and are potentially affected by recall bias. However, findings for mental health outcomes were consistent when comparing a self-reported diagnosis of depression or anxiety with a validated screening measure of psychological distress measure (the Kessler 10).
Findings are similar to previous research which has found improvements in improved cause-specific cancer mortality with improving SEP (Poulson et al., 2022), and studies showing the cumulative effects of low SEP on hypertension (Lopes et al., 2021). Similarly, previous studies have also shown healthy lifestyle factors (healthy selection bias) associated with rises in SEP, affecting subsequent health outcomes (Missinne et al., 2015), and rising SEP improved depressive symptoms and falling SEP increased depressive symptoms, mainly among men (Gugushvili et al., 2019). Less consistent findings were evident in specific sub-populations, for example for cardiovascular outcomes among African Americans where a rise in SEP increased risk of diabetic outcomes (Chen & Miller, 2022), among refugees who experienced more severe depressive symptoms post-migration with falling SEP (Euteneuer & Schafer, 2018), and increased anti-depressant use among those with falling SEP (Melchior et al., 2018).
Findings of the current study suggest that modifying SEP can have impacts on subsequent health outcomes and health inequalities. While disentangling the individual determinants of SEP is problematic, there are an array of potentially modifiable risk factors and sub-populations of relevance for specific policy responses. For example, education, unemployment, and housing, alcohol consumption, and specific subpopulations such as older age groups, those living alone, rural and remote populations, and Aboriginal and Torres Strait Islanders. Policy interventions that acknowledge the syndemics of health disparities and address the main determinants of social mobility could potentially prevent decreases in SEP, reducing socio-economic inequality and preventing poor outcomes across a range of physical and mental health conditions. Similarly, interventions to facilitate upward social mobility, such as early childhood education resources (Garon-Carrier et al., 2022), housing affordability (Wetzstein, 2017), education and training (Baum et al., 2011), structured employment support and income re-distribution (Knotz, 2018), and health service access (McMaughan et al., 2020), also have the potential to facilitate higher SEP and reduced socio-economic inequalities.

Conclusion
This study investigated the association between changes in socioeconomic position (SEP) trajectories and the onset of physical health outcomes, mental health outcomes, and associated treatment commencement, and found strong associations between decreasing SEP and poor health outcomes and improving health outcomes with increasing SEP. This study identifies a range of associated sociodemographic and psychosocial determinants for these SEP trajectories. Preventing people from falling in SEP and assisting those with low SEP across the social determinants of health can reduce risk of new onset chronic disease, and these findings can be used to inform interventions and policy responses to prevent the emergence of health inequalities in the Australian context.

Ethics approval
The authors assert that all procedures contributing to this work

Author contributions
MD led the conception and design of the study, conducted the data analyses, and drafted the manuscript. AP contributed to the conception and design of the study, data interpretation, and critical revision of the manuscript. SS contributed to the study design, oversight of the statistical analysis, and provided critical revision of the manuscript. GU contributed to data interpretation and provided critical revision of the manuscript. CHS contributed to data interpretation, provided perspectives on policy implications of findings, and provided critical revisions of the manuscript. JE contributed to conception of the study and provision of critical revision of the manuscript.

Funding
This work was supported by the Centre for Research Excellence for Integrated Health and Social Care (CREHSCI), funded by the National Health and Medical Research Council [APP1198477]. Fig. 4. Importance of variables for random forest Model as per the Mean Decrease in Accuracy and Mean Decrease in Gini statistic upon the variable's exclusion from the model.* Caption: Variables with increasing importance are listed from top to bottom with their corresponding mean decrease in accuracy and mean decrease in Gini. Both are performance measures of importance. *Note: hgage1 is "Age at first follow-up", mrcms is "Martial Status over time", bmi "BMI over time", lsdrkf "Alcohol Consumption over time", hhra "Remoteness of Residence over time", edhigh1 "Highest Educational Qualification over time", pdk10s "K10 score over time", dodtyp "Dwelling type over time", hgsex1 "Gender at first followup", gh1 "General Health and wellbeing over time (1-5)", anlote "Speaks Language other than English over time", capune "Percent of time spent unemployed in the last financial year", anatsi "Aboriginal and Torres Strait Islander status over time", gh10 "Social functioning given health status (1-5) over time", lejif "Member of family jailed in the last year", levi "Victim of Physical Violence in the last year", lebth "Birth/adoption of a new child".