Traumatic brain injury (TBI) and spinal cord injury (SCI) are both potentially catastrophic injuries resulting in lifelong disability. Combined, they contribute significantly to the global burden of injury. In 2016, there were an estimated 27 million new cases of TBI and 0.93 million new cases of SCI globally, with age-standardised incidence rates of 369 per 100 000 and 13 per 100 000, respectively.1

Both are major contributors to permanent disability and burden of disease in Australia. The 2016 incidence rates in Australia were estimated at 275 per 100 000 for TBI, which is lower than the global rates. Australian estimates for SCI were 23 per 100 000, which is higher than the estimated global rates.1 The incidence of TBI in Australia specifically is not well defined and varies based on the study methodology and scope of data. In New South Wales, Australia’s most populous state, the annual incidence of hospitalised TBI was estimated at 99.1 per 100 000 based on 2007 data.2 The annual incidence of hospitalised traumatic SCI was estimated between 2.1 and 3.2 per 100 000 based on 2006 hospital data.3

These injuries can result in long-term physical, neurological and psychological impairment. This results in ongoing medical, rehabilitation and care and support costs. Many of these injuries occur at a younger age, resulting in a long lifetime of care and support and specialised accommodation.4 The economic cost of these injuries in Australia was estimated at $8.6 billion annually for TBI and $2 billion annually for SCI for new incidence in 2008.5

The long-term care pathways and living arrangements for people with TBI and SCI are varied. Some people can live in the community with an informal support network. Some people live in specialist-supported accommodation with funded supports provided by an insurance scheme (if their injury was compensable)6 or government-funded disability support. Some people will be living as young people in residential aged care due to a lack of alternative care and living arrangement.7

The introduction of the National Disability Insurance Scheme (NDIS) in Australia, beginning in 2013, reflected the view that the previous disability support system was ‘underfunded, unfair, fragmented and inefficient’8 (page 2), and had the aim of providing ‘reasonable long-term supports for people acquiring a significant disability’8 (page 12) aged under 65 years regardless of the origin of their disability.

The NDIS has seen increasing costs associated with SCI and brain injury since its inception. At the end of 2020, the NDIS had approximately 14 000 participants with a brain injury with an approved plan of supports. Annual committed supports were in the order of $1.6 billion, although only just more than two-thirds used. In 2017/2018, the average annualised committed supports for scheme participants with brain injury was $99 000 per annum, compared with a scheme average of approximately $58 000 per annum. In 2020, the average annualised committed supports for participants with brain injury had risen to $148 000 per annum (a 50% increase), compared with a scheme average of $71 000 (a 22% increase). SCI shows a similar, though less pronounced trend. There were approximately 4900 participants with an SCI and average annualised committed supports have increased from approximately $120 000 in 2017/2018 to $166 000 in 2020 (a 38% increase). This shows not only a significant increasing trend but also a higher rate than for average scheme participants.9

The NDIS has piloted independent assessment tools in an attempt to ensure funding is applied consistently in line with need,10 but these independent assessment tools have been temporarily discontinued in the face of opposition from the disability sector.11 The use of an independent assessment tool has the potential to reduce inflation of support costs associated with ‘bracket creep’12 (page 17)—the incentive to inflate need for support to receive additional funding.

The Australasian Rehabilitation Outcomes Centre (AROC) collects data from rehabilitation units across Australia and New Zealand.13 Data collected on inpatients include the Functional Independence Measure (FIM) score. The FIM score provides a clinically assessed, standardised measure of functional impairment and is recorded at both admission and discharge from the rehabilitation episode. Assessors are credentialed in the use of the tool to ensure consistency of results.14 Research has shown a relationship between FIM score and ongoing support needs for people with SCI in particular.15

By linking AROC data, including FIM score at discharge, to NDIS participant data, including annualised planned support budgets, we will create a dataset that includes both an independent measure of functional independence and the resources allocated by the NDIS for care and support, with the following primary aims:

To quantify the relationship between the FIM score and the annualised funded supports allocated by the NDIS.

To analyse whether this relationship has changed over time or varies between cohorts.

To identify the other factors that predict the annualised funded supports allocated by the NDIS.

To estimate the prevalence and incidence of serious brain injury and SCI entering the NDIS each year.

The research will provide an important first step in understanding the link between functional impairment and disability support needs for TBI and SCI, as well as a foundation for ongoing data linkage between rehabilitation hospitals and the NDIS, assisting in the successful roll-out of the NDIS.

Methodology

This study is a retrospective, national, population-based cohort study of people with TBI or SCI who have transitioned between an inpatient rehabilitation setting and the NDIS. The protocol outlines the approach to data linkage between these two data sources.

Datasets to be linked

The AROC inpatient dataset records information on all inpatients treated in AROC member rehabilitation units in Australia and New Zealand; most public and private rehabilitation units are members of AROC.16 Data are collected from rehabilitation units in a deidentified form. Names of patients are not provided, replaced by a statistical linkage key, SLK-581,17 which comprises letters of name, date of birth and sex and has been used in Australia for linking health and disability records of diverse types.18 In addition to SLK-581, AROC linkage data include postcode, injury onset date, age at admission, episode end date and flags for TBI and SCI.

The NDIS participant dataset contains information on scheme access requests, plan approvals, committed plan supports and disability severity measures for all people who have requested access to the NDIS with a brain injury or an SCI as their primary disabling condition. It is extracted from the NDIS participant management system and contains full names. The linkage file includes an SLK-581 that is generated using the name, date of birth and sex information, as well as modified versions of SLK-581 that take into account known variations in the name (eg, Joe and Joseph would result in different SLK-581s). The file also includes postcode, scheme access request decision date, scheme phasing date and flags for TBI and SCI.

The scheme phasing date refers to the date the NDIS was made available to an individual based on their age and location. The roll-out of the NDIS between 2013 and 2019 was complex and based on geographical location and age groups.19

Although AROC data do not have identifiers, the SLK-581 makes it potentially reidentifiable in a linkage process. For this reason, linkage was done using a secure, separated process. The AROC and NDIS linkage files were securely transferred to NDIS linkage staff who undertook the linkage without access to analysis data (ie, FIM scores or funded supports). These linkage staff produced a project link file which includes a unique identifier for the AROC record, a unique identifier for NDIS record and a unique project record identifier. Each data custodian was provided only with their respective unique record identifier and the project record identifier so that analysis variables could be extracted and sent to the researcher with the project record identifier. This enables the researcher to join the two analysis datasets without having to see the linkage variables.

AROC data was available for all inpatient admissions between 2008 and 2019. NDIS data was available for all access requests since scheme inception on 1 July 2013 until June 2022, noting that national coverage of the scheme was not achieved until 2019.

Linkage approach

Data linkage creates a new and improved data resource by combining different administrative data sources that relate to the same person. Data linkages can be subject to error in the form of false matches and missed matches. As a result, the linkage methodology seeks to find a trade-off between sensitivity (the proportion of all records that have a true match in both files that are correctly accepted as linked) and specificity (the proportion of all records in one file that do not have a corresponding match in the other file that are correctly not accepted as linked).20 Errors in linkage can lead to bias in the resulting dataset and subsequent analysis.21 This bias can be hard to identify due to the different agencies involved in the distinct stages of the linkage pathway. The secure, separated linkage approach can increase the risk of bias.22 It is intended that the linked analysis dataset be investigated for potential bias.

The challenge of this linkage is the limitation of SLK-581. The SLK-581 is a concatenation of 14 characters comprising the second, third and fifth letters of the family name; the second and third letters of the given name; the date of birth in the form DDMMYYYY and a single character for sex (1 for male, 2 for female and 9 for unknown).17

The SLK-581 is designed to maximise the probability of being unique, though it is not a perfect tool for linkage. The Australian Institute of Health and Welfare (AIHW) has shown that across all 440 000 residential aged care residents in Australia, only 0.6% had a common SLK-581 with one or more people,23 which suggests a high sensitivity. A recent study on Australian data indicated that the use of SLK-581 and other non-identifiable data was able to identify 97.5% of the links that a gold standard, fully identifiable approach could identify.24 The AIHW estimated that compared with name-based linkage, an SLK-581 linkage had a sensitivity of 98.5%.25 International studies of statistical linkage keys show comparable results.26 Name changes (eg, after marriage) reduce the sensitivity of SLK-58127 as well as the specificity. Specificity (the rate of true non-matches as a proportion of all non-matches) is reduced by missing, incorrect or out-of-date data in the SLK.28 However, specificity is less informative where the number of non-matches far outweighs the number of matches.21 It has been shown that the use of probabilistic data matching techniques can improve the linkage results.21 28 This is the motivation for the stepwise deterministic approach taken in this protocol.

The linkage algorithm undertaken here has been designed to address these shortcomings by identifying other plausible and valid links. It operates in two stages: a linkage step with a hierarchical set of linkage passes using components of SLK-581 with broadly decreasing levels of information and a validation step using additional linkage data items to assess the validity of the match. The validation step relies on additional variables such as key clinical or administrative dates and geographical location (ie, postcode or state). Previous studies have shown the effectiveness of using variables such as these in probabilistic data linkages29 and they are employed here to optimise the linkage.

The AROC data represented episodes of care (n=20 886) for SCI (n=3937) and TBI (n=16 583) and both (n=366). A single individual may have multiple episodes of care, and therefore, the AROC data is not unique for individuals. The NDIS data represented individuals who have applied for scheme access (n=21 828), including those with SCI (n=5602) and brain injury (n=16 226) as their primary disabling condition. The NDIS data should be unique for individuals.

This project is the first time these two datasets have been linked and so there is no prior data to train a purely probabilistic approach. The linkage uses a stepwise deterministic approach with probabilistic or ‘fuzzy matching’ elements.30 Probabilistic approaches can produce more sensitive matching because they allow for variation in identifiers by calculating a measure of similarity, or match weighting, between values. This is used to decide whether a pair is accepted or rejected, usually after some clerical review.31–33 The stepwise deterministic approach for SLK-581—using the most discriminating match keys first and combined with additional variables—has been shown to increase linkage sensitivity compared with a simple deterministic approach.34 It has also been shown to be attractive when sufficient data are not available for the clerical review a researcher requires for the probabilistic approach.34 Further, it has been shown that detailed deterministic algorithms can be as successful, or even more successful, in identifying links compared with probabilistic algorithms.35

Figure 1 visually represents the 15 linkage passes used in broadly decreasing order of SLK-581 information used. For example, pass 1 is a match on the full SLK-581, pass 5 is a match on date of birth and letters of first name only and pass 14 is a match on date of birth only.

Figure 1

Linkage passes with broadly decreasing information through inclusion of SLK-581 components. SLK, statistical linkage key.

Not every possible combination of the components of SLK is used because not all combinations were deemed useful linkages based on the prior experience of the AROC data custodian.

Higher linkage passes correspond to lower degrees of confidence in the accuracy of the matches. This is reflected by the categorisations on the left-hand side of figure 1. Passes have been broadly grouped into four categories (A–D). These categories will be maintained on the linked analysis dataset so that analysis can be performed taking the different levels of confidence into account. Evaluating the sensitivity of future analysis results to the changes in linkage pass will help identify potential sources of bias in analysis.31

After each linkage pass, matches were validated using the validation steps. Successful matches were excluded from future passes. The validation steps address three main pieces of information: disability type, key dates of onset, NDIS availability and NDIS access and geographical location. The use of disability type is effectively a blocking technique used in probabilistic data linkage that limits the records to be matched to only those that have the same disability type.30

The flow chart in figure 2 outlines the validation steps undertaken.

Figure 2

Linkage validation process.

The first step checks to see if both records have the same disability type in each dataset. The AROC data created a flag for both TBI and SCI, so it is possible for an individual to have both TBI and SCI. The NDIS data contain only a primary disabling condition. In order to pass validation the NDIS primary disabling condition must also be flagged in the AROC dataset. It is possible that this approach may miss some true matches that have incorrectly coded disability types. However, for pass 1, the rejection rate due to disability was <1%, which is in line with the proportion of false matches we might expect from SLK-581.23

The second step checks to ensure the date of onset of disability in the AROC data is before the NDIS access decision date. If the date of onset of disability is missing in the AROC data, then the rehabilitation start date is used as a proxy. This validation step assumes that the TBI or SCI that led to the rehabilitation episode also led to the disability which led to NDIS eligibility. It is possible that a person could be eligible for the NDIS on the basis of a different disability (eg, autism, intellectual disability, etc) and has sustained the TBI or SCI subsequent to scheme access. As a result, this validation step is not applied to pass 1. This means that for cases where the date of TBI or SCI onset is after the NDIS access decision date, a match on full SLK-581 is required for inclusion as a match. Therefore, AROC inpatients with a disability other than TBI/SCI as the cause for NDIS eligibility will only be included if they match on Pass 1.

The third step considers the order in which the key dates occur for the linked records. Table 1 shows the six possible orders of the three key dates: rehabilitation end date (R), scheme phasing (the availability of the scheme to an individual) (P) and the NDIS access decision date (A).

Table 1

Potential key date order scenarios

The use of dates reflects the expectation that the normal pathway for TBI and SCI would progress from injury to rehabilitation, and then to NDIS eligibility. If a matched record follows this pattern, it is more likely to be a true match than a record that does not follow this pattern. Logically, one would expect the NDIS access decision date to follow the scheme phasing date. That is, the individual would not be granted access to the scheme until it was available to them based on their age and geography. Note that this validation step was not applied to pass 1 for the same reason outlined in the second step above. A match on full SLK-581 is deemed sufficient to imply a non-TBI/SCI disability as the cause for NDIS eligibility.

Scheme phasing is based on a complicated series of rules based on geographical location and date of birth. This is further complicated by the fact that during the roll-out of the scheme, it was widespread practice for potential participants to have the eligibility assessed prior to the scheme being available to them in order to expedite the planning process. It also could be that the scheme phasing date has changed due to a change of address of a participant after entering the scheme. However, since both the scheme phasing date and the NDIS access decision date were provided in the NDIS data, it is assumed that any ordering of these dates is acceptable. As a result, date orders 1, 2 and 3 are all considered valid.

It is possible for a scheme access decision to be made while a person is still in the rehabilitation setting (as per date orders 4, 5 and 6). However, we would expect there to be a relatively short time period between access and the end of the rehabilitation episode. To investigate this time lag, records matched on pass 1 (full SLK-581) were considered. These records are considered the most reliable matches and are not subject to this validation step. This analysis showed that for date orders 4, 5 and 6, more than two-thirds of access decisions were made within 6 months (183 days) of the rehabilitation episode end date. There was a sharp drop-off in decisions in the seventh month prior to rehabilitation episode end date, with the majority of the remaining decisions made 12 months or more prior to the episode end date. Therefore, it was assumed that pass 1 links with a >6-month time period between access decision and rehabilitation episode end date were likely to be individuals with TBI or SCI as a secondary disability. Therefore, matches from passes 2 and later with these date orders were only accepted if the time between access and rehabilitation episode end date was <6 months (183 days).

The final step is to use geographical location data from addresses to validate matches. Location information available on both datasets included postcode and state/territory. Postcode can be mapped to each NDIS district (a region smaller than a state/territory). A match with corresponding postcode, district or state/territory information is more likely to be a true match than one without these validations. However, it is possible that NDIS participants will move location, even interstate, particularly where there is a time lag between rehabilitation episode and NDIS access.

For Passes 1, 2, 3 and 4, no location validation was conducted, recognising the higher quality of these matches. For all other passes, records needed to be in the same state or territory to be accepted. Where this resulted in a unique match, the matched record was accepted.

Duplicates

Some of the lower linkage passes (passes 5–15) resulted in a high number of duplicate matches, where a single AROC record could be matched to multiple NDIS records. These duplicates were subjected to a further validation step to see if unique valid matches could be identified. If restricting the validation to include only records that matched on postcode (rather than just state) resulted in unique matches, then these records were accepted.

Validation results—unique matches

In total, the linkage passes resulted in 10 449 unique potential matches. Of these, 3143 were recommended for acceptance and 7306 were rejected on the basis of this validation.

Table 2 summarises the matches accepted and rejected by linkage pass. This shows that only 13 of 1758 (0.7%) pass 1 matches were rejected on validation. This is the order of magnitude of false-positive matches to be expected from previous studies.23

Table 2

Summary of matches recommended for acceptance and rejection (by reason)

Table 2 also shows the reason for rejection by linkage pass. Each of the reasons for rejection corresponds to one of the steps in the flow chart in figure 2: disability type, date of onset, date order and geography.

Validation results — duplicate matches

In addition to the 3143 unique matches identified above, a further 558 duplicate matches were able to be refined to unique matches by only accepting records that match on postcode across both datasets. This stricter validation increases the likelihood of a true match. A summary of these matches by linkage pass is included in figure 3.

Figure 3

Accepted linked records by linkage pass with included components of SLK-581. SLK, statistical linkage key.

As a result, there are a total 3701 matches validated for acceptance. These are outlined in figure 3. This summarises the linked records by linkage pass and the corresponding confidence level (A–D) based on the amount of identifying information used for linkage.

Records linked with lower levels of confidence will be subject to further analysis once the analysis data is made available. Particular attention will be paid to any potential biases, and some records may be subsequently excluded from further analysis. This analysis will form the basis of future investigations into linkage approaches and will be reported on in future publications.

Analysis

It is anticipated that this linkage might be the first of many linkages between these two datasets, with this protocol providing the basis for future linkages as well as being refined over time. Analysis data, to be provided by data custodians, will be reviewed for any indications of false matches and analysed for bias between different linkage passes to provide feedback on the linkage protocol.

The linked dataset will be analysed to address the research questions.

Is there a correlation between the FIM score and the annualised funded supports allocated by the NDIS?

The FIM scores for individuals who are eligible for the NDIS will be compared with the FIM scores of people found not eligible for the NDIS on the basis of disability (if any such people are found in the linked data). This will give an indication of the relationship between scheme eligibility criteria and FIM score. Generalised linear modelling will be performed on the initial annualised funded supports for NDIS participants, to determine the extent to which the FIM score or the individual elements of the FIM score are predictive of annualised funded supports. This analysis can be further broken down by the types of funded supports (core, capacity building and capital) and the used supports (payments).

Has this relationship changed over time or between cohorts?

The above analysis will incorporate cohort elements such as year of injury, year of scheme entry and year of annualised planned support costs. This will identify any trends in scheme eligibility or resource allocation. Modelling can determine whether these elements are independent predictors of scheme resource allocation.

What other factors predict the annualised funded supports allocated by the NDIS?

In addition to FIM score and time, other factors need to be included in the modelling to control for potentially confounding effects and to further understand the drivers of scheme costs. These may include age, comorbidities and length of stay in rehabilitation setting.

What is the prevalence and incidence of serious brain injury and spinal cord injury entering the NDIS each year?

This project forms a part of a broader research project that seeks to better understand the incidence, prevalence and economic cost of SCI and particularly TBI in Australia. Although the data limitations mean this study is unlikely to capture all new incidents, it will provide an insight into the quantum of people entering the NDIS from rehabilitation settings each year.

Patient and public involvement

People with disability, NDIS scheme participants and rehabilitation and disability professionals will be involved in the future of this project as one or more of the following: consultants, reference group or steering committee members or coauthors. This will provide a greater understanding of the data and context of the supports provided. This can guide future data analysis.

Ethics and dissemination

Ethics approval has been obtained from the Macquarie University Human Research Ethics Committee in July 2020. The requirement for obtaining consent for the use of personal data was waived on the basis that the data were pre-existing administrative and research datasets, the secure, separated data linkage process created minimal risk to privacy and the potential benefits of the research are significant. It is anticipated that dissemination of research findings to key stakeholders will be through peer-reviewed publication in scientific journals, conference proceedings and directly with NDIS and AROC. This study has the potential to establish an ongoing linkage between rehabilitation hospitals and the NDIS. Relevant linked information such as FIM score could be used as an input to the NDIS planning process. Functional independence is only one element of the whole person that needs to be considered in the planning process. The planning process considers all aspects of a participant’s life. However, having this information linked prior to the planning process may serve to smooth the administration of transition of participants with severe injury into the NDIS.