2.1 Data

The Australian Bureau of statistics (ABS) census data was used to extract socio-economic indicators for areas (SEIFA) from 2013 to 2021, at local government area (LGA) level across rural and remote Australia [28]. SEIFA scores rank areas by socio-economic advantage and disadvantage through variables measured in census data [29]. Age distributions for rural and remote LGAs were also sourced from the ABS, using data-explorer [30]. The proportion of residents aged over 85 years old and aged under 5 years old in each LGA were calculated. The following data were collected from the Health Workforce Data Tool [31], provided by the Australian Government Department of Health and Aged Care for the same time periods (2013–2021) and by LGA:

• Population size.
• Number of all full time equivalent (FTE) medical practitioners that are registered with the Australian Health Practitioner Regulation Agency (AHPRA) and working.
• Number of all FTE nurses and midwives (as a singular total) working and registered with AHPRA.
• Number of FTE allied health professionals registered with AHPRA and working. As is consistent with the Australian Institute of Health and Welfare [4], this includes Aboriginal and Torres Strait Islander health practitioners, chiropractors, Chinese medicine practitioners, medical radiation practitioners, occupational therapists, optometrists, osteopaths, pharmacists, physiotherapists, podiatrists, paramedics and psychologists.
• Remoteness area, measured by the Australian statistical geography standard remoteness structure.

2.2 Study variables

The outcome variable is binary, with each LGA being coded as either being in deficit or not being in deficit of the categories of health care worker of interest, i.e., “medical practitioners”, “nurses and midwives”, and “allied health professionals”. Thresholds to define a “deficit” were extracted from WHO’s 2016 report on health workforce requirements. This report concluded that 4.45 medical practitioners, and 4.45 nurses and midwives are required per 1,000 people to achieve health related sustainable development goals [32]. Therefore, under 4.45 FTE workers per 1,000 people was defined as a deficit for medical practitioners and nurses and midwives.

The WHO health workforce requirements report recognised the importance of allied health professionals but did not define a minimum workforce requirement for this profession due to difficulty centralising data [32]. Due to a lack of other literature surrounding the required quantity of allied health professionals, the 2020 Australian national average of allied health professionals per 1,000 population was used to define a deficit in this category, which is 5.94 FTE allied health professionals per 1,000 people [4].

Index of relative socioeconomic advantage and disadvantage (IRSAD) was the SEIFA score chosen to represent SES in this analysis, as it considers elements of both advantage and disadvantage, and encompasses the broadest range of variables [29]. IRSAD will be measured in quintiles, as is consistent with national [18,33], and international literature [22,23] examining the socioeconomic status of areas in a similar context. A higher IRSAD quintile represents higher SES.

The following variables were included as covariates:

• ◦Remoteness area (RA), measured by the Australian statistical geography standard remoteness structure.
• ◦Age distribution variables:
  • ▪proportion aged under 5 per 1,000 population.
  • ▪proportion aged over 85 per 1,000 population.
• ◦Time in years from 2013 to 2021.

These covariates were included based on the assumptions illustrated in Fig 1, which were guided by current literature. National statistics demonstrate that the remoteness of an area is independently associated with both SES [1,2] and health workforce availability [4], making RA an important covariate to adjust for to avoid possible confounding. Year was also included as a covariate, as time may influence both health workforce availability and IRSAD scores through events such as the covid-19 pandemic. Age distributions from two extreme tails of the LGA population, i.e., less than 5 and over 85 years old were also included as covariates, due to possible relationships between age and local area-level SES [26], which may influence health workforce availability through workforce retention factors such as availability of childcare services [16] and community engagement [14].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Diagram depicting possible relationships between the included covariables and the outcome variable (health workforce deficit), and the exposure variable (index of relative socioeconomic advantage and disadvantage quintile).

https://doi.org/10.1371/journal.pone.0321198.g001

3.3 Statistical methods

Data were examined using summary statistics, and repeated observations were deleted. Examination of diagnostic plots showed a non-linear relationship between the continuous variables and the binary outcome variable, which prompted the development of a non-linear additive structure in our model for the continuous variables. More specifically, we used Generalised Additive Models (GAMs), which allow for inclusion of non-linear continuous variables by use of cubic regression splines [34]. GAMs were created for each healthcare profession.

Due to the clustering effects in the data, generalised estimating equation (GEE) estimation was also implemented. The data were considered to be clustered on the basis that separate observations from a single LGA are likely to be more correlated than observations from different LGAs [35]. This clustering violates the independence assumption of GAMs, making GEEs a necessary addition to address this violation [36]. This decision was supported by the large intra-cluster correlation values within each GEE model. GEEs address clustering by assuming a correlation structure within the dataset, which allows for calculation of robust standard errors [36]. GEEs are robust to misspecification of correlation structures [37], and were selected as the most appropriate method to account for clustering as they allow for population averaged estimations.

To create the GEE-GAM models, the splines of the continuous variables were extracted from the GAMs and applied to GEE-GLMs. Similar techniques of developing GEE-GAMs have been shown to be successful in literature [38]. The QIC (quasi-likelihood under the independence model criterion) was used to identify the best and final GEE-GAM for all workforce categories. Age distribution variables were considered for removal from the final GEE-GAM models if they did not cause an increase in the QIC value or a significant change in the IRSAD quintile coefficients after removal. RA and year variables were not considered for removal due to their strong contextual relevance to the model.

The model selection process was completed with both exchangeable and autoregressive correlation structures for each GEE-GAM, and the correlation structure for each final model was selected based on the lowest QIC value. Effect modification between year and age distribution variables was also tested in the final model and considered for inclusion if interaction terms were significant. QIC values were compared between the initial GEE-GLMs and the final GEE-GAMs to ensure the best model was selected.

We used R programming language to implement the models [36,39], and used “glm”, “mgcv” [34,40] and “geepack” [41] packages to get model-based results. Table B and C in S1 File includes model-based output for crude and adjusted GLMs for the health professional categories considered in this paper.

4.1 Descriptive statistics

A total of 3,294 observations were collected from 368 rural and remote LGAs across Australia, over nine years. There were no missing data. IRSAD quintile 5 (most advantages quintile) contained the smallest proportion (8.4%) of observations, while quintiles one to four each contained between 20% to 24% of the remaining observations. The highest proportion of observations were from inner regional Australia (41.6%), and the lowest proportion were from remote Australia (11.3%).

There was an overall 15.5% relative decrease in the proportion of LGA’s in deficit of allied health professionals between 2013 and 2021 (Table 1). The proportion of LGAs in deficit of medical practitioners in 2013 had decreased by 3% in 2021, which is the smallest relative change amongst all health profession categories. Nurses and midwives had a 9.6% relative decrease in the proportion of LGAs in deficit between 2013 and 2021.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The relative change in proportion of deficit for three categories of healthcare professions, between the years 2013 and 2021 across 368 rural and remote Australian local government areas, by remoteness area and index of relative socioeconomic advantage and disadvantage (IRSAD).

https://doi.org/10.1371/journal.pone.0321198.t001

4.2 Model-based results

All GEE-GAMs had lower QIC values than their corresponding GEE-GLMs, and were therefore selected as final models. The final GEE-GAM for allied health professionals included all the covariates, as listed in Table 2. All continuous covariates were observed to have a non-linear relationship with the logit of the outcome variable, and were therefore included as smoothed variables, which are plotted in Fig 2. Age distribution variables remained in the final model as their exclusion increased the QIC score. There was no evidence of effect modification between year and age distribution variables. The C-statistic for this model is 0.77, which is greater than the null model (0.5). The QIC value for this model (1652) is lower than that of the null model (1824). Table A in S1 File provides the model-based results.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Results from the final Generalised Additive Model with Generalised Estimating Equations for each healthcare profession.

https://doi.org/10.1371/journal.pone.0321198.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Smooth functions for continuous covariates included in the GEE-GAM (generalised estimating equations applied to a generalised additive model).

https://doi.org/10.1371/journal.pone.0321198.g002

The estimated odds ratios calculated for each IRSAD quintile (Table 2) have large robust standard errors, and there is no statistical evidence (p>0.1) of association between IRSAD quintile and allied health professional deficit. However, there is strong evidence (p<0.01) that LGAs in very remote areas have on average 70% lower odds of being in deficit of allied health professionals compared to LGAs in inner regional areas, after adjusting for the covariates listed in Table 2.

The GEE-GAM model for medical practitioners included all covariables, and all continuous variables were included as smoothed functions (Fig 2) due to non-linearity. The age distribution variables increased the QIC value when removed from the model and were therefore included. There was no evidence of effect modification between the age distribution variables and year. The C-statistic for this model was 0.73 which is higher than that of the null model (0.5). The QIC for the final model (2016) was also lower than that of the null model (2140). The model showed no statistical evidence of association between IRSAD quintile and medical practitioner deficit, due to large robust standard errors. There was also no statistical association between area remoteness and medical practitioner deficit.

The GEE-GAM model for nurses and midwives also included all variables, with both age distribution variables being included as smoothed terms (Fig 2). Year was included as a linear term in the model after graphical inspection. The age distribution variables were not removed from the model, as their presence lowered the model’s QIC value. There was no evidence of effect modification between either of the age distribution variables and year.

This model showed a significant association between nurse and midwife deficit and IRSAD quintile. Strong statistical evidence (p<0.01) demonstrated that on average, LGAs in quintile 5 had 2.54 times the odds being in deficit of nurses and midwives compared to LGAs in quintile 1. There was also evidence (p<0.05) to show that LGAs in IRSAD quintile 4 had on average 1.65 times the odds of being in deficit of nurses and midwives compared to LGAs in IRSAD quintile 1. Additionally, there is weak evidence (p<0.1) to show that on average LGAs in IRSAD quintile 2 and 3 had respectively 1.34 and 1.43 times the odds of being in deficit of nurses and midwives when compared to LGAs in IRSAD quintile 1. All conclusions were after adjustment for the covariates listed in Table 2, and after accounting for clustering.

Additionally, strong evidence shows that LGAs in remote areas have on average 88% lower odds of being in deficit of nurses and midwives than LGAs in inner regional areas, and that LGAs in very remote areas have 72% lower odds of being in deficit of nurses and midwives than LGAs in inner regional areas, after adjusting for the included covariates.

5.1 Limitations

The use of IRSAD as a proxy for SES is a limitation of the models created. While IRSAD is a widely used measure of SES that undergoes validation by the ABS [28], socioeconomic status remains a difficult concept to quantify. IRSAD ranks relative socioeconomic advantage and disadvantage, meaning LGAs in lower IRSAD quintiles are not necessarily of low socioeconomic status, but are of lower SES compared to LGAs in higher quintiles [29]. Hence, the conclusions of this study cannot be equated to actual SES, but relative SES.

Another key limitation of the study is the lack of detailed data on the breakdown of health workforce professions. For example, nursing and midwifery data are combined by AHPRA, despite their distinct scopes of practice. These professions often have separate roles and are frequently the first point of contact for providing essential services in health and primary care, health promotion, and disease prevention. Therefore, analysing these two professions separately would provide a more meaningful perspective for policy implications.

Analysing health workforce deficit as a binary outcome rather than as a continuous measure also introduces model limitations. The use of a binary outcome variable means that the magnitude to which an observation is above or below the deficit threshold is not accounted for, and more detailed associations cannot be uncovered. The efficacy of the final models is also dependent on the accuracy of the deficit thresholds defined. The lack of evidence surrounding the minimum requirements for allied health professionals means deficit in this category was based on the national average of FTE allied health workers in 2020, which may be an inaccurate representation of deficit. The minimum health workforce requirements estimated by the World Health Organisation are not specific to the rural and remote Australian setting, and therefore may decrease accuracy of the thresholds defined for medical practitioners and nurses and midwives.

Additionally, areas of greater remoteness with more sparsely distributed populations may require an increased number of health professionals to account for the potentially larger distances of travel required. This means a single population to health practitioner ratio may not be appropriate for all levels of remoteness, which may also decrease the accuracy of the models created. However, to our knowledge more accurate minimum thresholds are not available, and the estimations supplied by the World Health Organisation are widely used.

Another limitation of this study was the inability to adjust for an LGA’s level of need. Areas of lower SES are likely to have increased burden of disease [25], and therefore may require more healthcare workers per head. Hence LGAs in lower IRSAD quintiles may be in deficit of healthcare workers due to increased demand, but in this analysis would not be classified as such if they are above the minimum threshold. This may bias the results towards the null, as was evident in findings from Fisher et al [23]. Further research into the association between SES and health workforce availability with adjustment for need may be valuable, and uncover associations that this paper was unable to.

It must also be noted that this analysis examines healthcare worker availability, and does not account for quality, acceptability, or accessibility of care. The number of FTE healthcare staff available is only a precondition to adequate care [32], with aspects of quality, acceptability and accessibility being equally important to the delivery of essential health care services [22,32].