Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Submitter: An AHPRA-registered psychologist and PhD candidate. Name-withheld status requested.

Disclosures: The author has no financial or contractual ties to the National Disability Insurance Agency (NDIA), the Centre for Disability Studies (CDS), or the i-CAN’s developers. As a provider of services to some of the schemes participants, the author could be indirectly affected by provisions affecting private-sector providers. AI tools assisted in preparing this submission; all statistics, citations and quotations were manually verified against original sources, and the author is responsible for its accuracy.

Lay Summary

This submission examines the scientific evidence for the i-CAN, the assessment tool that the Bill will use to set National Disability Insurance Scheme (NDIS) funding for more than 700,000 Australians. A scientific analysis of the published research shows a number of serious methodological and procedural problems.

 - There is no published reliability or validity evidence for the versions of the tool that will be deployed.

 - The current available evidence shows that i-CAN scores accounted for less than 4% of the differences in
    the funding people receive; over 96% of funding was determined by factors other than i-CAN results.

 - The current available evidence shows that the i-CAN’s ability to reliably measure participant needs was
    poor, uncertain, and feasibly overestimated.

 - The current available evidence shows that the i-CANs precision to accurately classify participant needs

cannot rule out failing to meet the developers cited success criteria for low-stakes observational research.

 - The current available evidence has significant design limitations, including too few participants to
    adequately power the studies to have appropriate levels of precision given the human stakes. The
    submission body sets these issues out in detail, while Appendix A calculates the number of participants
     that would be required to meet both the minimum contemporary threshold (80% power) and a more
     suitable threshold for high stakes assessment (95% power) for each estimate

Together, these findings do not support the NDIA’s characterisation of the i-CAN as a gold-standard tool in disability needs assessment (NDIA, 2025), and do not support its use for high-stakes assessment contexts. Furthermore, there does not appear to be any published evidence to show that the NDIA’s method for translating i-CAN assessment results into funding for participant plans is itself reliable or valid. The Bill also does not require the government to publish how i-CAN results will be translated into a funding amount.

The submission examines these issues from a scientific perspective, and identifies additional risks associated with the proposed assessment system, including changes to the assessor workforce that sidelines the involvement of federally registered clinicians with domain-relevant expertise and their associated regulation systems, reductions to the rights of appeal available to participants, a need for stronger regulatory oversight (incl. potential involvement by the Therapeutics Goods Administration) regarding the use of an invisible software to circumvent clinical determinations of participant impairment and support needs while autonomously allocating funds to their NDIS plans, and omissions in the Bill’s Statement of Compatibility with Human Rights. It also raises a possible tension with the Australian Constitution that may leave service providers vulnerable to civil conscription.

This submission asks the committee to delay the relevant Schedules until both the deployed instrument and the funding allocation method are properly validated for the use contemplated by the Bill. Until this occurs, deployment carries is a real risk of material harm to Australians with disabilities and their families.

While the submission is quite critical of the i-CAN in the context of this intended use-case, I would like to affirm that the i-CAN is a promising instrument that warrants continued development, and wish to praise the CDS and its staff for its commitment to Australians facing disability, and for its work towards innovation in the assessment of its impacts. It is my hope that this detailed engagement would be interpreted as a sign of respect by the i-CANs authors and serve a constructive aim by assisting CDS in its future endeavours.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Executive summary

This Bill (Parliament of Australia, 2026b) commences an algorithmic system to set NDIS funding for more than 700,000 Australians by amending the National Disability Insurance Scheme Act (Parliament of Australia, 2013). The validation evidence for that system fails the developers’ own benchmark for system-level deployment; it cannot reliably predict the funding allocations the system will be used to make. The system’s central instrument has no public validation evidence for the version to be deployed, and the algorithmic budget translation method has never been disclosed. The closest published evidence base fails to meet reasonable standards even for low- stakes observational research, let alone high-stakes use. Meanwhile, the mechanisms for detecting and correcting errors in participant plans are being narrowed in the same legislation.

The submission’s appraisal of the published validation literature identifies additional concerns. The i-CAN’s conceptual approach is in tension with the goals of NDIS needs assessment. The tool’s test-retest reliability in core domains is poor. The i-CAN’s classification accuracy did not clearly meet the minimum standards the developers themselves cite for observational research, and, in the only study where i-CAN scores were compared with actual funding allocations, the scores predicted less than 4 per cent of the variation.

Monte-Carlo simulation analyses set out in Appendix A demonstrate that across these studies, the statistical power to detect the relationships actually observed was, with few exceptions, well below the 80 per cent minimum that contemporary scientific journals expect of low-stakes observational research, with low precision for many of the observed coefficients. None of the published evidence covers the version of i-CAN that will be deployed, and there are no peer-reviewed validation studies for i-CAN v6 or the NDIA’s customised variant.

These concerns arise in a system already facing structural difficulties. The Administrative Review Tribunal (2025) reports a 73 per cent change rate for challenged NDIS decisions in its most recent reporting period.

The submission sets out the empirical case in detail, drawing on novel precision and power analyses (Appendix A). It also examines and identifies serious concerns with the Bill’s workforce provisions, identifies a subtle legal- inguistic change that, functionally, hands over clinical duties to automated software, and argues that this, in turn, means the assessment-to-funding pathway should be considered against the Therapeutic Goods Act 1989 as a software-based medical device. Finally, it points out that, given the lack of evidence for the assessment, that the Statement of Compatibility with Human Rights does not appear to engage with the Bill’s central provisions, and further identifies a possible tension with the Australian Constitution that leaves service providers vulnerable to civil conscription. It recommends nine amendments to allow the substance of the reform to proceed once its scientific and ethical foundations are adequately and independently demonstrated:

• Recommendation 1. [CRITICAL] Defer commencement of Schedule 4 and the dependent provisions of Schedule 1 until independent scientific validation of the deployed instrument, across the full NDIS population demographics, has been published in peer-reviewed literature and assessed by an expert panel that has no ties to the instrument, its developers, or the NDIA.

• Recommendation 2. [CRITICAL] Amend Schedule 4 to require, in primary legislation, publication of the budget method’s key methodological details, including validation metrics for the assessment-to- funding translation algorithm in a form open to reasonable scientific scrutiny and/or independent regulatory authority (e.g., the Therapeutic Goods Administration (TGA), per recommendation 9), before any participant is assessed under it.

• Recommendation 3. Amend Schedule 4 to require that all support needs assessors be registered health practitioners under the Australian Health Practitioner Regulation Agency with a relevant scope of practice in accordance with existing legislation and codes of conduct.

• Recommendation 4. Amend Schedule 1 Part 6 to restore the obligation to consider independent medical and allied health evidence in planning decisions, and to clarify that the new evidence hierarchy operates as a guide rather than a precondition.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

• Recommendation 5. Amend Schedule 4, Item 8 (new section 32L(4)(b) to (c)) so that any NDIS rule prescribing information an assessor must or must not have regard to is consulted on with the relevant National Boards of the Australian Health Practitioner Regulation Agency, and so that the rule must not, in principle, direct assessors to disregard information the assessor reasonably believes to be relevant to the assessment.

  • In addition, amend Schedule 4 so that NDIS rules made under section 32L(4A)(c) – which prescribe persons additional to Agency staff and engaged consultants who may undertake the assessment – are reclassified from category D to category A, so that there is parity regarding the information-constraint rules under section 32L(4)(b) and (c) (Item 13).

• Recommendation 6. [CRITICAL] Amend Schedule 1 to preserve the Administrative Review Tribunal’s full remedial powers, including the power to substitute or vary a plan decision.

• Recommendation 7. Amend Schedule 1 Part 4 to make ministerial determinations subject to sunsetting and parliamentary disallowance.

• Recommendation 8. [CRITICAL] Insert a provision requiring independent biennial public review of the assessment framework’s measurement performance, decision-change rates, and population-level outcomes, conducted by a body without commercial or contractual ties to the instrument or its developers, and without political interests in the perceived success of the framework.

• Recommendation 9. Require regulatory oversight under the Therapeutic Goods Act 1989 of any automated or algorithmic tool used to determine support needs in assessments or funding allocations under the NDIS. Software that takes private health information related to a person’s impairment and produces outputs that determine the supports required to compensate for that impairment feasibly meets the definition of a software-based medical device under the Act, in which case it would be regulated as Software as a Medical Device in any other clinical or commercial setting.

  • The committee may wish to recommend that the Department of Health, Disability and Ageing and the TGA jointly issue guidance on the application of this framework to the assessment-to-funding pathway empowered by Schedule 4 and Schedule 3 Part 2, and that commencement of those provisions be conditional on appropriate regulatory classification and oversight.

The Royal Commission into the Robodebt Scheme (2023) established that fiscal objectives do not justify systems that are opaque, inaccurate, and inadequately overseen. The same principle applies here. The founding architect of the NDIS, Professor Bruce Bonyhady, has described a similar needs assessment proposal as ‘robo-planning’ (Bonyhady, as cited in ABC News, 2021), and the Minister, introducing the Bill on 14 May 2026, said: ‘This government will not repeat those mistakes.’ The recommendations above set out conditions under which that commitment can be honoured. Without amendments along these lines, particularly Recommendations 1, 2 and 8, those prior mistakes are at high risk of being repeated.

Table 1 below summarises the provisions which this Bill (Parliament of Australia, 2026b) engages with.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Table 1. Bill provisions engaged by this submission

Bill provision	What it does
s 9B (Schedule 1 Part 1)	Statutory definition of functional capacity, requiring assessment ‘in a context that excludes, as far as possible, the impact of the person’s environmental and personal circumstances’, which appears to conflict with commitment to the social model of disability.
s 32K(1A) (Schedule 4, Item 3); s 32K(3B), (3C) (Schedule 4, Item 4)	Provides for a budget method, specified in NDIS rules, that translates the needs assessment report into a total funding amount, and authorises funding amounts to be more than, equal to, or less than the actual cost of supports
9, 10
s 32L(6)(a) (Schedule 4, Items 9 and 10)	Replaces ‘identify’ with ‘include information about’ in the assessment-report obligation (Item 9), and requires relevant support needs to arise ‘directly’ from the participant’s impairment (Item 10). Functionally circumvents need for clinical input and restricts supports only to whatever partition
of a persons impairments are presumed attributable to recognised disability.	section 3.6, 8,
9.1
s 32L(4A) (Schedule 4, Item 8);
NDIS rules under s 32L(4A)(c)
rules (EM p. 147)	Specifies who may undertake an assessment; permits the Minister to prescribe additional assessors with limited parliamentary scrutiny.
s 32L(4)(b) to (c) (Schedule 4, Item 8)	Specifies information an assessor must or must not have regard to, in NDIS rules (EM p. 146).
Schedule 4, Item 7 incorporation-by-reference note (EM pp. 145 to 146)	Permits NDIS rules to apply, adopt or incorporate Agency-prepared material ‘as in force or existing from time to time’.
Schedule 1 Part 6	Establishes a hierarchy under which peer-reviewed research outranks clinical evidence and may, on its own, defeat a support; the structure separates what good clinical practice combines (the integration of scientific literature with evidence about the person), and does not engage with the substantial gaps in the Australian disability literature.
s 34A;
s 34A(3), (5) (Schedule 1 Part 4)	Authorises the Minister, by legislative instrument, to reduce funding for specified groups of supports in old framework plans during the transition period, with effect ‘even if … less than the total cost’ of a reasonable and necessary support,
subject to the requirement to have regard to participant safety.	section 9, 10
New Division 5 of Part 1
of Chapter 4;
ss 59B
to 59E (Schedule 3 Part 2)	Authorises computer programs to take administrative action, including ‘evaluative determinations’ about what’s needed to reasonably compensate for a disability related impairment,
functionally replacing clinical decision makers.	section 8, 9
Schedule 1 Part 2	Limits unscheduled reassessment by raising the threshold to a significant and ongoing change in functional capacity that substantially reduces daily activities; implementation details left
to transitional ministerial rules.	section 6
Schedule 5	Transitional ministerial rules of limited duration and constrained parliamentary disallowance.

4

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

1. About this submission

I am an AHPRA-registered psychologist and a PhD candidate with interests in psychometrics and the philosophy of science. Over the past year, I have examined the published evidence base for the Instrument for the Classification and Assessment of Support Needs (i-CAN), which the National Disability Insurance Agency has selected as the basis for the NDIS Support Needs Assessment and described as ‘the gold standard of available, validated needs assessment tools’ (NDIA, 2025).

As part of that examination, I conducted precision and power analyses of the published validation studies for the i-CAN, to clarify the level of confidence the evidence supports for the use cases the Bill contemplates. Appendix A reports the analyses; Section 3 summarises them alongside additional critique of the study designs. The analyses are also intended to be of constructive use to the CDS and the NDIA in planning future studies sample sizes and controls.

This submission represents my best efforts to accurately compile my conclusions and articulate my arguments in the Senate Inquiries submission timeframe. It is made in a personal capacity within a highly limited daily schedule, and does not represent any employer, university, or professional body. For personal reasons, I have requested name-withheld status. The submission’s intended contribution is to place on the public record what the published scientific evidence does and does not currently support, and to identify methodological conditions under which the Bill’s central provisions could be made fit for the use they contemplate.

I invite the committee to consider the evidence I present and ask that it satisfy itself that the empirical and related ethical foundations of the Bill’s central provisions are sound before assenting to commence them.

2. The Bill’s central pathway

Based on my reading, the Bill empowers an algorithmic assessment-to-funding pathway as the basis for funding NDIS participant supports. It does so through four interlocking components.

• Functional capacity (new section 9B). Schedule 1 Part 1 introduces a statutory definition of functional capacity that excludes ‘as far as possible, the impact of the person’s environmental and personal circumstances’.

• Assessment report and budget method (new sections 32L(6)(a) and 32K). Schedule 4, Item 9 revises the needs assessment report so that it ‘includes information about’ disability support needs rather than ‘identifies’ them. Schedule 4, Item 3 provides for a budget method, specified in NDIS rules, that translates the report into a total funding amount.

• Below-cost funding (new sections 32K(3C) and 34A). Schedule 4, Item 4 inserts new section 32K(3C), which authorises any funding amount specified in the budget method to be ‘more than, equal to or less than the actual cost of providing or acquiring the support’. This is the central below-cost provision for new framework plans. Schedule 1 Part 4 also inserts new section 34A, authorising the Minister, by legislative instrument, to reduce funding for a specified group of supports in old framework plans during the transition period, with effect ‘even if the result is… the funding provided under a participant’s plan for a reasonable and necessary support is less than the total cost of the support’. The Ministrermust,undersection34A(3),haveregardtothesafetyofparticipantsinmakingthe determination;itisnotclearthatthisentailsanythingbeyondasimpleacknowledgment.

• Automation (new sections 59B to 59E). Schedule 3 Part 2 inserts a new Division 5 of Part 1 of Chapter 4, authorising computer programs to take administrative action, including ‘evaluative determinations’. Schedule 5 leaves implementation details to transitional ministerial rules, subject to limited duration and constrained parliamentary disallowance.

The pathway has two technical components:

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

1. An assessment instrument that produces a descriptive profile about a participant’s support needs; and

2. A budget allocation method that translates that profile into a funding amount for participant plans.

3. The empirical foundations of the assessment instrument

Both the Australian and International scientific communities accept that, for the kinds of measurement tools represented by the i-CAN, the ways in which we are to identify ‘Validity’ – the notion that an instrument assesses what it is intended to assess – is governed by several international and Australian standards. On this basis, three principles are foundational in making accurate statements about the validity status of any such test.

• First, validity is to be recognised as a property of a) a specific version of a test that is b) used for a specific purpose with c) a specific population; it is not a property of any test name in the abstract, and any change in any one of a), b) or c) materially impacts on what can be said about a tests validity (American Educational Research Association, American Psychological Association & National Council on Measurement in Education 2014, Chapter 1).

• Second, when an instrument is revised in any way, validity evidence does not transfer automatically from the prior version; though the validity evidence of prior versions might help inform that of later versions, it is an expectation that each version of an instrument obtain its own evidence for the various kinds of validity before respective statements about that version of test can be empirically supported (AERA et al., 2014, Chapter 4, Cluster 4).

• Third, validity for use with a population must be established with that specific population, and validity for a particular purpose must be established for that particular purpose (AERA et al., 2014, Chapter 1; Australian Psychological Society 2014, section 3.4; International Test Commission 2013). With Psychologists having specialised training in developing, evaluating, and administering these kinds of instruments in applied settings, the Psychology Board of Australia’s Code of Conduct (Psycbology Board of Australia, 2024; [Effective December 1, 2025]) requires that any psychological services, including in the selection of assessment tools, be grounded in evidence consistent with these standards, which are informed by the respective scientific consensus. The observations that follow are evaluated against these standards.

It is my view that the concerns set out in 3.1 to 3.5 below would substantially resolve if independent scientific validation of v6 and the NDIA customisation, across the full NDIS population, were published and assessed by an expert panel without ties to the instrument or the NDIA. The position taken in this submission is contingent on the evidence made public to date.

Table 2 summarises the precision and power findings most relevant to the Bill, drawn from the published validation studies (full results in Appendix A). This table presents two key metrics needed for an informed interpretation.

• A confidence interval is the range of values consistent with what the study actually observed: any value within the interval is consistent with the study’s data, meaning that the data alone cannot distinguish between which of the values inside the interval is the true coefficient. Confidence intervals of the coefficients reported here are based on Monte Carlo simulations of the available data, and are described further in Appendix A. How narrow an interval should be depends on the stakes of the decisions it will inform. For high-stakes individual allocation, this submission proposes that classification-agreement (kappa) coefficients be estimated to within 0.05, and funding-prediction (rho) correlations to within 0.10. No published i-CAN study meets either target (see Appendix A for the rationale and supporting analyses).
• Statistical power is the chance that a study has, given its size, to detect the relationships it observed if they were real. The conventional minimum power that contemporary scientific journals expect for exploratory, low-stakes academic research is 80 per cent (Cohen, 1988); a threshold appropriate for high- stakes decisions about disability funding allocation would be much higher, 95 per cent, for example.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Table 2. What the published i-CAN validation studies establish

Construct	Published result	Power	Bottom line
Does the same person score the same on a second occasion?	Core Health and Well-Being domains r ≈ −0.22 to 0.25; Activities and Participation domains r = 0.46 to 0.94	16–20%	Unreliable in the core Health and Well-Being domains; stronger in Activities and Participation, though measured on a sample unlike the norming population.
Does the algorithm assign the support level an expert clinician would?	κ = 0.94 (best-rated 49 of 186); 0.65 (all 186); 0.66 (independent 41)	30% (full sample)	The headline figure rests on a curated quarter of the sample; full-sample agreement may sit below the developers’ own 0.60 minimum.
Do i-CAN scores predict the funding a participant receives?	ρ = 0.13 to 0.20 (Total Raw 0.19)	38–72%	Explains under 4 per cent of funding variation; no domain reaches the 80 per cent power floor.

Note: r and ρ are correlations (0 = unrelated, 1 = perfect); κ is Cohen’s kappa (0 = chance, 1 = perfect). Power is the chance, at the published sample size, of detecting the reported relationship. Intervals are from Monte Carlo simulation (10,000 replications); full figures and methods in Tables 2a to 2c and Appendix A.

3.1 All published validation data are for older versions of the instrument

Discontinuity in validity evidence across versions. The published peer-reviewed literature on the i-CAN concerns versions 2 through 4.2, in studies conducted between 2003 and 2015. The version that the NDIA’s deployment is based on is v6 (released 2022). The instrument the NDIA will actually use is a further customised, non-public derivative of v6 (Centre for Disability Studies [CDS], 2025b). The Centre’s own publication list confirms that there have been no peer-reviewed validation publications for any version of the i-CAN since 2016 (CDS, 2025a). Under the standards set out above, the validity evidence accumulated for versions 2 through 4.2 cannot be expected to transfer to v6 or its NDIA customisation, particularly given that substantive changes were reported to have taken place between versions. Each version, therefore, requires its own validation data.

The NDIA proposes a power to change their assessment process on whim. The Bill provides the statutory mechanism that authorises this gap. The Explanatory Memorandum confirms (Schedule 4, Item 7; Parliament of Australia, 2026a, pp. 145 to 146) that NDIS rules specifying the support needs assessment tool ‘may apply, adopt or incorporate any matter contained in a document prepared by the Agency (such as a questionnaire or manual) as in force or existing from time to time’. The Memorandum states the purpose of this drafting as ensuring that updates to the inputs and content of the assessment ‘can be made in real time’. This appears to be the mechanism by which a customised, non-public version of the i-CAN v6 can be deployed, and by which the Agency can modify the deployed instrument without the modification appearing on the face of the NDIS rules or being subject to disallowance as a rule. Allowing the NDIA the power to modify assessment processes without careful analysis of the impacts of any changes introduces risks of unreliability and invalidity to the process. This also appears to replace existing provisions that require assessments to regard any other information that relates to the participant,

Given the above, my analysis should be considered a best-case analysis, as it asks whether the published evidence, taken at face value and granting every benefit of the doubt, would support the NDIA’s proposed use cases. From a scientific point of view, the unpublished current versions cannot improve on this position without additional validation data, as there is currently no publicly available evidence for them.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Existing evidence is split across different versions of the instrument. The published validity evidence itself is quite fragmented, with inter-rater and test-retest reliability data reported only for versions 2 and 3 (Riches et al., 2009b). Meanwhile, criterion validity data are available only for versions 4.0 and 4.2 (Arnold et al., 2014). Funding-prediction data are further derived from a different version – a research-only short form, the i-CAN-Brief (Arnold et al., 2015). The Mental Emotional Health domain, central to assessment of psychosocial disability, was substantially revised between v2 and v3, in changes the authors acknowledge but do not detail, with other modified versions appearing in later publications regarding the ability to predict appropriate funding amounts.

Given this, from a scientific perspective, the published evidence cannot provide sufficient validation of any single version of the i-CAN instrument for applied use. The deployed instrument and its NDIA customised variant are at least two design generations beyond the latest published evidence and therefore necessarily currently lack any evidence of reliability or validity.

3.2 The older versions’ evidence is too thin to support high-stakes use

By Reliability, I refer to the property of an instrument by which it produces consistent results when the same person is assessed on different occasions, in the absence of any real change. Low reliability in the funding context means a person’s funding can shift substantially between assessments, even when their support needs have not changed. Table 2a below presents the test-retest reliability figures reported in the published studies. This section explains how to read those figures, why the worst rows matter most, and why the better rows do not save the instrument for its proposed use.

Two distinct reliability questions. For an assessment instrument, ‘reliability’ covers two distinct questions which can be easily confused. The first is whether two raters observing the same assessment session agree on the scoring – inter-rater reliability. The second is whether two independent administrations of the assessment, with different teams interviewing the same participant on separate occasions, yield similar results – test-retest reliability. Inter-rater reliability is a necessary but not sufficient condition for assessment reliability in the deployment sense. If two raters watching the same session cannot agree, the scoring rubric is broken; if they can agree, the rubric is workable, but that tells us nothing about whether the assessment process as a whole, run by different teams on different occasions, produces stable results. For the use the Bill contemplates, the primary question is whether an assessment will yield consistent funding-relevant outputs across administrations, which deals with test-retest reliability.

What the published literature reports on each. Riches et al. (2009b) report a high overall inter-rater reliability of r = 0.99 (14 cases, 3,228 items), based on a design in which one trained researcher independently scored the same assessment sessions observed by the facilitator-led team. This confirms the scoring rubric is workable, but it does not establish that two independent administrations of the assessment would produce similar scores. Arnold et al. (2014) also report inter-rater reliability (rs = 0.89, n = 25), but for the clinical judges who served as the criterion comparator in their classification study – a separate question from the i-CAN’s own scoring behaviour (discussed in 3.3 below). The test-retest figures, which speak to the question the Bill turns on, are reported by Riches et al. (2009b) and are the subject of the rest of this section.

• The polling analogy. Imagine a polling organisation surveys 20 people and reports that 65 per cent support a particular policy. If the same poll were run again with a different 20 people, the result could easily be 45 per cent, or 80 per cent. Each figure for those twenty people is real, but with only 20 people, it cannot be relied on as a reliable indicator, nor a valid representative of the average Australians support for that policy.

  • The published i-CAN studies do report confidence intervals on some statistics, but they do not report statistical power, and they do not provide simulation-based precision analyses calibrated against use-case-appropriate thresholds.

  • Appendix A provides those analyses for the first time, clarifying how precise each study’s results are and how likely each study was to detect the effects it reported.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

• Confidence intervals and statistical power.

  • The confidence interval is a margin of error. Any value within the interval is 95% consistent with what the study observed, meaning the data alone cannot distinguish among the values within the interval. A wide interval indicates that the study’s data are imprecise and consistent with very different conclusions about the scientific property of the i-CAN.

  • Statistical power is a separate measure: it is the probability that a study with that number of participants would detect a real relationship if one were present. The conventional minimum for exploratory research is 80 per cent (Cohen, 1988); the threshold appropriate to high-stakes funding decisions would likely be closer to 95 per cent. Studies that fall short of 80 per cent power cannot reliably support exploratory conclusions, let alone high-stakes ones. This standard is now reflected in many scientific journals, which do not consider publishing a study unless it demonstrates at least 80% statistical power for its reported results. Low power has a second, less obvious consequence that bears on how the i-CAN evidence base should be read. Research has found that underpowered studies also include a higher proportion of larger, false positives than adequately powered ones (see Appendix A on Statistical Power). Because of this. where the statistical power is low for a published i-CAN coefficient, it may be best to read it as a plausible upper bound on what an independent replication would find, rather than as a best estimate of the true relationship.

Table 2a. Test-retest reliability (Riches et al., 2009b)

Domain	Retest interval	n	r	95% CI	Power	What the data can establish
Mental Health	2 years	20	−0.22	[−0.62, 0.24]	16%	Effectively unrelated; the data cannot establish any reliability.
Emotional Health	2 years	20	0.25	[−0.21, 0.64]	20%	Spans slight disagreement to strong agreement; reliability not demonstrable.
Physical Health	6–12 months	23	0.21	[−0.21, 0.59]	16%	Cannot distinguish no agreement from moderate agreement.

Note. Riches et al. (2009b, Table 2); see Appendix A, Tables A10 and A11. The three weakest and most psychosocially relevant domains are shown; the Activities and Participation domains performed better ($r = $ 0.46 to 0.94) but were measured on a sample unlike the i-CAN’s norming population.

• The headline finding. The most consequential row in Table 2a is the first. When the same 20 participants were reassessed two years after their original interview on the Mental Emotional Health domain of the i-CAN, the correlation between the two sets of scores was r = −0.22. A correlation of 1.0 would indicate perfect agreement; a correlation of 0 would indicate that the two assessments are completely unrelated.

  • The published −0.22 sits slightly below zero: slightly worse than unrelated. The confidence interval runs from r = −0.62 to r = +0.24, which means the data are consistent with anything from strong disagreement to moderate agreement.

  • In terms of statistical power, the study had a 16 per cent chance of detecting whether this domain produces consistent scores at all. Reaching even 80 per cent power would require about 200 participants. The study used 20.

• The pattern across Health and Well-Being. Behaviour at two years and Physical Health at six to twelve months both produce confidence intervals that span zero relationship, with statistical power of 20 per cent and 16 per cent respectively. Across the Health and Well-Being domains, these analyses

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

show that the published data cannot rule out the possibility that the instrument yields no statistically reliable or scientifically valid information about the person being assessed in those domains.

• The developers’ explanation. The developers attribute this instability to real change in participants’ support needs between assessments, not to measurement error. While this explanation is possible, it cannot be tested from the published data. This is because the validation study did not report whether participants who experienced documented health changes showed score changes consistent with those changes, or whether participants without documented changes showed stable scores. Without reporting this information, it’s not possible to determine whether the instability is attributable to real change.

• The Activities and Participation domains. The Activities and Participation domains fare considerably better, with reliability correlations in the 0.6 to 0.9 range; however these scales face a separate problem.

  • The i-CAN was developed on a group of 1,012 people, 84 per cent of whom had intellectual disability as their primary diagnosis. The smaller groups that were used to check reliability had a markedly different profile: 60 per cent of participants had physical disability , sensory disability or acquired brain injury.

  • This means that the reliability figures were not obtained from the population on which the instrument was normed. A reliability figure measured on one population does not automatically apply to another.

• The composition of the normative sample itself. Of the 1,012 individuals in the original development group, 84 per cent had intellectual disability, 91 per cent lived in congregate residential settings (group homes and similar shared-care arrangements), and 2.3 per cent were aged 65 or older.

  • No published validation data exist for autistic individuals, people with psychosocial or physical or sensory disabilities, people with acquired brain injury, First Nations peoples, culturally and linguistically diverse communities, or older Australians (Riches et al., 2009a).

  • Despite this, the Bill, as it currently stands, appears to empower the NDIA to deploy the i-CAN instrument precisely to these populations.

3.3 The criterion-validity headline figure depends on selection from a small sample

Table 2b. Classification accuracy (Arnold et al., 2014)

Subsample	n	κ	95% CI	Power	Benchmark	What the data can establish
Best-rated reports (of 186)	49	0.94	[0.85, 1.00]	77%	κ > 0.80	even on the developers’ best-rated quarter of the sample.
Full validation sample	186	0.65	[0.56, 0.74]	30%	κ > 0.60	about one-in-three chance of clearing it.
Independent sample	41	0.66	[0.46, 0.83]	13%	κ > 0.60	Interval too wide to place; spans inadequate to high agreement.

Note: Arnold et al., (2014); see Appendix A, Tables A1, A4 and A5. Benchmarks differ by row: the best-rated subsample is tested against the high-stakes floor ($\kappa$ > $0.80$; McHugh, 2012), the representative samples against the developers’ own observational-research minimum ($\kappa$ > $0.60$; Suen & Ary, 1989). Including ‘average’-rated reports lowers $\kappa$ to $0.75$ ($n = 114$).

When the i-CAN is administered, an algorithm sorts the assessed person into one of five categories of support need: no formal support, mild, moderate, substantial, and pervasive. Classification accuracy refers to how often the algorithm places someone in the same category that an expert clinician would, if they were to assess the same person. The statistic used to measure this agreement is Cohen’s kappa, which ranges from 0 (agreement no better

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

than chance) to 1 (perfect agreement). In their papers, the developers provide citations suggesting that a kappa of 0.6 is the minimum acceptable value for low-stakes observational research (Suen & Ary, 1989). Other widely cited benchmarks include 0.75 for ‘excellent’ agreement (Cicchetti, 1994) and 0.80 for ‘strong agreement’ (McHugh, 2012); the latter is a reasonable floor for individual classification with high-stakes consequences.

The most widely cited figure for the i-CAN’s classification accuracy is kappa = 0.94 (Arnold et al., 2014). Two features of the derivation of that figure warrant the committee’s attention, as seen in Table 2b.

Selection.

The 0.94 figure is drawn from a carefully curated subset of 49 from the total 186 assessment reports in the validation study, those that the developers’ own team rated as ‘good’ or ‘very good’ quality.

• When all 186 reports were included, the kappa fell to 0.65, with a 95 per cent confidence interval of [0.56, 0.74]; the lower bound of this range sits below the 0.60 threshold the developers themselves cite as ‘acceptable for observational research’.

• The independent validation sample of 41 reports yielded a kappa of 0.66 with a confidence interval of [0.46, 0.83], a range too wide to distinguish between an instrument with adequate scoring consistency and one with marginal consistency.

• At N = 186, the study had only 30 per cent statistical power to demonstrate that kappa exceeds the 0.60 threshold. This means that their study had only a 1-in-3 chance of correctly identifying whether the instrument’s true scoring consistency met the lowest acceptable bar for low-stakes, observational research.

• Even at the developers’ favoured n = 49 subsample where kappa = 0.94, statistical power against a high- stakes threshold of 0.80 reached only 77 per cent, which is still below the 80 per cent minimum for exploratory research, and well short of a value (e.g., 95 per cent) that high-stakes funding determination would call for.

Criterion contamination.

The i-CAN’s classifications were not tested against an independent comparison. To test whether an instrument produces accurate classifications about functional need, the test requires an independent source of the right answer: for example, an experienced clinician who assesses the person directly, without seeing the instrument’s output.

• In the i-CAN study that attempted to assess this quality (i.e., criterion validity), the comparison was made by assessors who, although making their judgement independent of the i-CAN process itself, drew on the i-CAN’s assessment reports as the basis for their judgement about the person’s level of need (Arnold et al., 2014).

• The instrument and the comparison were therefore working from the same information, which is a form of criterion contamination. In other words, the authors tested the i-CANs accuracy using data generated by the i-CAN.

As a result, this cannot count as a test verifying whether i-CAN’s scoring accurately classifies functional needs. It is a test of whether two summaries of the same i-CAN data agree. While that is useful information, it does not establish that the i-CAN places people in the right category, only that two readings of the same instrument’s output reach similar conclusions about that same output.

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

By the developers’ own criteria, the instrument does not predict funding

Table 2c. Funding prediction (Arnold et al., 2015)

Score / domain	n	ρ	95% CI	Power	What the data can establish
					Explains under 4 per cent of funding variation;
i-CAN Total Raw Score	163	0.19	[0.04, 0.33]	69%	below the 80 per cent power floor.
					Interval includes zero; no link to funding
Health and Well-Being	163	0.13	[-0.03, 0.28]	38%	demonstrable.
					Near-zero to small; still short of the 80 per
Activities and	163	0.20	[0.05, 0.34]	72%
Participation					cent power floor.

Note. Arnold et al. (2015, Table 4, standard scoring); see Appendix A, Tables A14 and A17. These figures show whether scores track funding already allocated, not whether they predict the funding a new participant would require, which the published literature has not evaluated.

On the benchmark cited by the developers, the published data fall well short.

A question of intended use. As mentioned earlier, an instrument validated for one purpose cannot be assumed to perform adequately for another without additional validity testing (AERA et al., 2014, Chapter 1; International Test Commission 2013). The i-CAN has been developed to identify needs and plan clinical supports, whereas its application to algorithmic funding determination is a materially different use that, according to accepted scientific standards, will require its own evidence of validity.

Under the scoring method on which the assessor workforce is to be trained, the relationships between i-CAN scores and the actual funding allocations participants (Arnold et al., 2015, Table 4) can be observed in Table 2c.

Implications. A correlation of 0.19 between i-CAN total score and funding allocation means that less than 4 per cent of the variation in funding decisions is accounted for by the i-CAN score. To put it another way, this evidence demonstrated that knowing someone’s i-CAN total score tells us almost nothing about what funding they will receive. More than 96 percent of the difference between participants’ allocations is determined by something other than their i-CAN scores. This is far lower than would be expected if i-CAN scores were useful predictors of the funding a participant requires to meet their support needs.

Granted, it would not be reasonable to expect i-CAN scores to strongly predict participant funding; in the context of the author’s study, the i-CAN was only one of several inputs that determined participants’ funding allocations. In this case, it might be reasonable to expect a modest correlation between i-CAN scores and participants’ funding allocations; however, a much weaker correlation than this was observed.

Furthermore, and of particular warrant for the committee’s attention, the proposed Bill, unlike this study, would make i-CAN scores the central input for determining participant funding. The relevant question for the committee’s assent is, therefore, in my view, what predictive performance the i-CAN exhibits as the basis of the allocation.

If the data from this prior version of the i-CAN are any indication, i-CAN scores will not reliably or validly predict the amount of funding required to meet participants’ support needs, and will likely struggle to predict any meaningful amount of funding without integrating far more information than what i-CAN scores can provide.

Imprecise estimates. Further still, these small correlations were themselves estimated imprecisely. Based on simulations from the data reported in the study, The Total Raw correlation of 0.19 has a 95 per cent confidence interval of [0.04, 0.33] and 69 per cent statistical power, which is again below the 80 per cent minimum expected of exploratory research. The Health and Well-Being correlation of 0.13 has a confidence interval of [-0.03, 0.28] and only 38 per cent power, which means that the data cannot reliably establish whether the Health and Well- Being component of the i-CAN predicts funding at all. The Activities and Participation correlation of 0.20 has a confidence interval of [0.05, 0.34] and 72 per cent power, which is still below the exploratory-research threshold.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Model Instability. The predictive models in Arnold et al. (2015) sit below the conventional ratio of 15 observations per predictor recommended for stable prediction (Harrell, 2015), with the standard i-CAN + demographics model achieving only 8.6 observations per predictor. The stepwise model reaches the recommended ratio only because variable selection reduced the predictor count from 19 to 4 within the same sample.

Recognition due to Authors. Although they did not report simulation-based precision and power analyses themselves, the developers nonetheless appreciated the significant limitations in this data in their own interpretations; they are on record expressing disappointment in these early funding-prediction attempts (Riches et al., 2009b) and appropriately noted that a larger sample would be necessary in order to develop a sufficiently precise allocation algorithm (Arnold et al., 2015).

Post-hoc prediction. One other feature of this study warrants the committee’s attention. In this study, the developers reported that one variable showed a substantially stronger funding prediction ($R^2 = 0.87$). What needs to be appreciated, however, is that this was constructed post-hoc based on ‘funded items’ data; here, items were re-coded according to whether the support each item described was funded through Disability Services. Three features limit what this result establishes, and prevent it from establishing the predictive value of i-CAN scores that the NDIA is interested in.

• First, the coding was applied after the fact, not as part of standard administration; assessor workforces would not be receiving training on it.

• Second, the criterion (total funding) and the predictor (sum of items coded as funded) both index the same allocation decisions at different levels of detail, so the correlation partly reflects the internal consistency of funding that had already been allocated, which is substantially different to the instrument’s capacity to determine appropriate allocations for new participants.

• Third, the model used stepwise regression with one outlier removed; the standard error of the resulting predictions was approximately $14,000, across allocations ranging from near-zero to over $90,000. The developers themselves noted that the findings require independent verification (Arnold et al., 2015). To date, such verification has not happened.

Put plainly, altogether this shows that, in their study that attempted to use an older version of the i-CAN to predict funding, the developers’ best result showed that the instrument could recreate funding decisions that had already been made, though even this best result demonstrated an average prediction error of around $ 14,000 per participant. That data, however, did not show that the instrument could determine appropriate funding for new participants.

The latter question is the one that matters for an instrument empowered by this Bill (Parliament of Australia, 2026b) to determine funding for new participants on a national scale, and it was not answered in these studies. Depending on how the current versions of the instrument have been changed, it would not be unreasonable to hypothesise from this data that i-CAN scores will not accurately predict participants’ funding needs, at least not by themselves.

3.5 No independent replications of the measurement tools have been completed

Conditions for Scientific Credibility. For scientific data to be deemed credible within the community, it is expected to be replicable by independent parties without conflicts of interest. Affecting independence are that:

• All published validation studies on the i-CAN have been conducted by its development team.

• The lead author holds a royalty agreement with the CDS regarding use of the instrument (Arnold et al., 2015), while CDS owns the intellectual property and is a consortium partner in the NDIA’s implementation.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

• In 20 years, no independent replication of the instrument’s scientific properties has been published.

Authors acknowledged the need for more rigorous study. During one of their evaluations of i-CAN validity, Riches et al., (2009b) themselves anticipated the need for independent replication and more appropriate controls in their study designs, remarking how at the time that funding was not conditional on the outcome, and that if the results did determine funding that much more rigour would be required.

• The NDIS deployment represents a material increase in conflicting interests and is precisely the kind of scenario that the developers appear to have alluded to, one that requires more rigour before the measure is ready to be deployed for this purpose.

• To date, there has been no published research using any of the controls the authors alluded to as necessary before the i-CAN can be reasonably deployed for the purposes the Bill is intended to empower it for.

3.6 The instrument produces different scores for the same underlying needs

The i-CAN, based on its developers’ design philosophy, assesses support needs in the context of the person’s specific environment, counting only items relevant to that environment (Stancliffe et al., 2016). From a theoretical point of view, this appears to be both a novel and genuinely meritorious proposal with a philosophical sophistication that distinguishes the i-CAN from alternative measures of functional need. Other measures tend to assume that functional needs can be estimated as a stable, trait-like state that operates largely independently of environmental context, whereas appreciation of this context appears to be a key feature of i-CAN. As a property of a clinical support-planning instrument, the i-CAN therefore has a well-defensible underlying logic that warrants further scientific study.

However, this theoretical aspiration appears to conflict with the Bill’s intended use of the i-CAN. As a property of an instrument used to determine consistent funding for participants in the national scheme, this commitment to support needs arising from a person’s specific environment will introduce a structural problem: two individuals with the same enduring support needs but different environmental circumstances may end up with dramatically different total scores on the i-CAN and, under the method being proposed to budget funding allocations, receive completely different amounts of funding. The Bill’s text moves in the opposite direction: Item 10 of Schedule 4 amends paragraph 32L(6)(a) to require that relevant support needs arise ‘directly’ from the participant’s impairment(s), leaving the deployed instrument’s environment-sensitive design in tension with the very text authorising its use.

A supplementary environmental-circumstances measure could be added to standardise i-CAN outputs across environments. However, any such module would itself require published validation evidence, both for its measurement properties and for its use in combination with the i-CAN to predict funding.

4. The budget translation method has not been disclosed

The Bill enables a budget method, specified in NDIS rules, that translates assessment outputs into funding allocations (Schedule 4, Items 3 and 4). New section 32K(3B) authorises that method to identify levels of need for supports and to specify funding amounts for each level, while new section 32K(3C) authorises those amounts to be ‘more than, equal to or less than the actual cost’ of the support. The methodology by which the levels of need will be derived, the basis on which the funding amounts will be set against those levels, and the validation evidence supporting either, have not been published or made publicly available. Analysis of documents obtained under Freedom of Information legislation indicates that the NDIA already employs machine learning models to generate draft budgets from participant profile data (van Toorn & Carney, 2024). The model that will link assessment data to funding outcomes under the Bill has not been made available for independent review, preventing independent scrutiny of its scientific properties.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Invisibility prevents public scrutiny. The methodology behind this half of the pathway is therefore currently invisible to the public and to the scientific community. In principle, this prediction model can be examined and challenged by a trained practitioner when it yields an invalid outcome. However, the translation into funding appears to occur within a system that, by design, cannot be examined independently by appropriately trained experts.

In practice, this will likely mean that:

• A participant who receives a lower funding allocation than expected or needed has no way to know whether the shortfall lies in the assessment, in the budget method, or in their interaction.

• Without meaningful human oversight, determining the source of a shortfall would be difficult, if not impossible for the NDIA itself. It is not possible to meaningfully challenge what cannot be seen.

The need for ‘test security’. One legitimate reason for not publishing the details of this algorithm may be ‘test security’. If too much information about the budget allocation method were publicly available, it would be at risk of being gamed by malicious actors. However, if the NDIA does not publish anything about this method, it leaves the public and the scientific community without any reasonable assurance that the method is fit for purpose, and not unreasonably disadvantaging certain populations. This may seem to leave the NDIA in a bind.

• Without public assurance that the allocation model is fit for purpose, disabled Australians are left vulnerable to a system that, if not fit for purpose, would be structurally unable to be challenged.

• On the other hand, standard mechanisms of public assurance (e.g., detailed technical publications) themselves risk undermining the scheme’s effectiveness by making it vulnerable to malingering.

Possible solutions to the test security problem. Several options are available and worth considering:

• One is that the NDIA may simply choose not to disclose any information about the budget allocation method, though this would substantially undermine public confidence in the scheme and the presiding government.

• Alternatively, the NDIA might choose to publish summary statistics on validation evidence regarding its predictive accuracy on out-of-sample data; this would help mitigate the opacity concern, though this might entail non-trivial risks to the scheme’s security.

• Another option, proposed for other reasons in Section 8 of this submission, might involve an independent agency serving as a quality-assurance regulator, which can critically evaluate the system while keeping technical details secure – for example, the TGA.

Irrespective of what the NDIA chooses to do about this issue, this submission’s position is that establishing adequate validity not just for the i-CAN, but for the funding translation mechanism, and doing so in accordance with the kinds of scientific standards outlined earlier, should be considered a necessary precondition for assenting to commencement of the relevant Schedules, and not simply a desirable accompaniment that can be left for the scheme to achieve later, after it has begun materially impacting participants.

5. Participant needs will be evaluated against a standard of evidence that the assessment process itself does not meet

Schedule 1 Part 6 establishes a hierarchy of evidence types the CEO must consider in determining whether a support is reasonable and necessary, with published, peer-reviewed research at the top. As a strong adherent of evidence-based practice, scientific standards, and the scientist-practitioner model, I consider applying peer-reviewed standards to the allocation of public resources to be a legitimate aim. The way the provision has been drafted, however, does not achieve that aim. The following issues should be noted.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

• Schedule 1 Part 6 Item 73 inserts new paragraph 34(1E), which requires the CEO, when considering two or more of the listed evidence types, to rank them in this order:

  • peer-reviewed and generalisable research;
  • evidence of the support’s effectiveness given the participant’s circumstances;
  • evidence of outcomes for the participant from their previous plan;
  • other matters the CEO considers appropriate. New paragraph 34(1F) further authorises the CEO to decide that the satisfaction requirement is not met on the basis of limited or absent category (a) evidence, even where category (b) or (c) evidence is available.

• The hierarchy treats peer-reviewed research and individual clinical evidence as if they were separable when, in well-conducted practice, they are not. Under the scientist-practitioner model, the central task of clinical practice is the critical engagement of the peer-reviewed literature and – taking into account its gaps and limitations – its translation into decisions appropriate for the specific person receiving services. Basing practice and clinical reasoning on the peer-reviewed literature is a professional obligation of AHPRA-registered health practitioners, explicitly set out in their respective Codes of Conduct. The hierarchy structurally separates two activities that, in good clinical practice, must always occur together, even if not explicitly obvious to consumers of clinical services.

• The framing presumes a completeness that the Australian disability literature does not have. The disability literature has substantial gaps, is underfunded relative to other clinical domains, and provides very limited evidence on the needs and effective services for many participant cohorts. Clinical reasoning is the means by which defensible use of the sparse literature in individual cases is possible, and is one of the competencies that AHPRA-regulated training establishes. Any reading of the actuarial-methods literature that disregards the role of clinician decision-making is a misreading of that evidence; algorithmic tools can guide and structure decision-making, but there is no reasonable evidence that they can safely or accurately perform this duty autonomously, particularly at scale.

• Where category (a) evidence is sparse or absent – which is the situation across many disability supports – paragraph 34(1F) permits the CEO to refuse the support on that basis alone, however strong the category (b) or (c) evidence is. The effect on participants and their treating practitioners will be similar whether the legal text describes this as ranking or as exclusion. Practitioners whose clinical evidence sits in category (b) will know that, where category (a) is absent, their evidence may carry little or no decision-weight regardless of the strength of the individual case for the support.

• The Bill imposes a standard of evidence on participants that the assessment instrument it empowers does not itself meet. As Section 3 sets out, the i-CAN has no published peer-reviewed validation for the version that will be deployed, and what is published does not consistently meet even the standards the developers themselves cite for low-stakes observational research. Participants will be required to provide peer-reviewed evidence to access supports, while being assessed by an instrument that does not satisfy that same standard.

In effect, this provision will place an intolerable burden on participants that most will never be able to meet. The Bill stands to impose on participants a standard of evidence that its own central instrument and funding mechanisms will not. It represents a serious ethical and procedural issue that will place many disabled Australians at material risk of serious harm.

6. The error-correction mechanisms are being narrowed when the system is already underperforming

The Bill commences these provisions at a time when the Administrative Review Tribunal has reported a 73 per cent change rate on finalised NDIS decisions for the period 14 October 2024 to 30 June 2025 (Administrative Review Tribunal, 2025). For plan-related decisions specifically, the figure was 74 per cent. This submission takes for granted that these percentages don’t just reflect errors in original decision-making, but also reflect cases where new evidence has become available since that decision.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

The figures are nevertheless material to the present argument, because the Bill’s central provisions act on both components: section 3 bears on the rate at which original decisions are sound, and the provisions discussed below bear on the visibility of any errors of either kind.

The Tribunal in Ray and NDIA [2020] AATA 3452 held that a single-occasion assessment by an unfamiliar practitioner should be considered less reliable than longitudinal evidence from treating professionals, which is consistent with established scientific methods for evaluating individuals. The Bill empowers the former assessment model.

One provision of the Bill, taken together with two further changes indicated in internal NDIA briefings tabled at Senate Estimates, will functionally reduce the visibility of any errors the new system will produce.

• Reassessment threshold raised (Schedule 1 Part 2). Schedule 1 Part 2 limits unscheduled reassessment, raising the threshold to a ‘significant and ongoing change in functional capacity that substantially reduces’ daily activities. Implementation details are left to transitional ministerial rules of limited duration and subject to constrained parliamentary disallowance (Schedule 5).

• Tribunal remedial powers restricted (reported). Internal NDIA briefings tabled in Senate Estimates by Senator Steele-John (Senate Community Affairs Legislation Committee, 2025) indicate that the Administrative Review Tribunal’s remedial powers will be restricted to ordering reassessment by the same system that produced the contested result. The briefings entered the public record through Estimates tabling, where the Agency was represented and had the opportunity to correct the record.

• Removal of the obligation to consider independent medical evidence (reported). Those briefings further indicate that NDIA staff will no longer be required to consider independent medical evidence in planning. In the same briefing, an NDIA senior manager stated there would be ‘no requirement for staff to consider independent medical evidence provided by NDIS participants in the development of plans’ (Lyons, 2025). A staff member present raised the foreseeable consequence directly, asking how participant wellbeing could be protected, ‘if there are needs not captured in that assessment and we’re under no obligation to consider independent evidence, then you can see how there might be shortfalls’ (Lyons, 2025). The concern raised in this submission is the same concern NDIA staff have raised internally.

A future reduction in the decision-change rate would not, by itself, indicate that the new system produces better initial decisions. It could equally indicate that the primary mechanism for detecting and correcting poor outcomes has been weakened. The workforce concerns examined in Section 7 and the visibility-narrowing set out here are features of the same legislative package and, together, are likely to have structural effects on the scheme and its participants, whether intended or not.

7. The workforce that will administer the assessment

The Bill (new section 32L(4A), inserted by Schedule 4, Item 8) provides for three categories of person who may undertake the assessment: members of the Agency staff under section 169 of the Act, consultants engaged by the Agency under section 171, and persons prescribed by NDIS rules. Across all three categories, the Bill does not require the assessor to hold registration under the Australian Health Practitioner Regulation Agency with the relevant scope of practice. Internal NDIA briefings indicate that, for the Agency staff workforce, an allied health background will be ‘desirable’ but ‘not mandatory’ (Lyons, 2025, reporting on the NDIA Queensland staff briefing of 13 November 2025), while the category of ‘persons prescribed by NDIS rules’ is left open by the Bill.

The remainder of this section examines two problems that fall back on participants: the position of an AHPRA-registered practitioner who is asked to administer the assessment under the Bill’s information-handling rules, and the position of any assessor who is not AHPRA-registered.

The AHPRA-registered practitioner.

Item 8 of Schedule 4 rewrites section 32L(4)(b) to (c) so that NDIS rules will ‘prescribe specific information an assessor must or must not have regard to’ (Parliament of Australia, 2026a, p. 146), without providing any examples. For an AHPRA-registered practitioner, a statutory direction about what information they must not consider sits in tension with the professional judgment obligations under their

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

respective boards’ Codes of Conduct, which they are required to adhere to as part of their registration to practice.

• These Codes require AHPRA practitioners to consider information they reasonably believe to be relevant to the case of the person they are providing services to.

• A subordinate instrument that constrains that judgement would appear to create conditions in which the AHPRA practitioner is forced towards either non-compliance with their AHPRA Code of Conduct, or non-compliance with the NDIS rules. An AHPRA-registered practitioner who is participating in the Needs Assessment is then placed in a position that the Code does not contemplate, and the Bill does not resolve.

• The lack of examples in the provision, taken together with Schedule 4, Items 7 and 8, suggests that what an assessor must or must not have regard to becomes dependent on NDIS rules that could be changed ‘in real time’, and replaces existing provisions that requires assessments to regard any other information that relates to the participants. Specific content regarding this (Item 13) does not appear to exist yet.

• These changes could direct AHPRA registered practitioners not to consider information the participant provides to them that they, on the basis of their expert judgements, reasonably believe is relevant to their needs, which they are likely required to consider in order to comply with their conditions of registration.

The non-registered assessor. The assessment involves substantive judgements about a person’s impairment, heir functional capacity, and the supports that would compensate for their disability. These are the standard subject matter of registered health practitioners, for which there is a considerable public infrastructure to ensure minimum standards of competence and public safety.

• In the current proposal, a workforce that is not required to hold either mandatory clinical qualifications or relevant tertiary backgrounds will make determinations about a domain whose content falls within professional clinical scope and for which competence and safety regulations are of substantial public interest.

• The category of ‘persons prescribed by NDIS rules’ appears to be left open by the Bill, and Item 14 of Schedule 4 classifies the rules that prescribe additional assessors as category D, which I understand to be the lightest legislative-instrument category, on the ground that prescribing additional assessors ‘will not have any significant policy or financial implications’ (Parliament of Australia, 2026a, p. 147).

• By contrast, Item 13 of the same Schedule places the rules that prescribe information an assessor must or must not have regard to (under section 32L(4)(b) and (c)) at category A, a more rigorous classification. The rule-making power that determines who can administer the assessment is therefore subject to less scrutiny than the rule-making power that determines what information an assessor can consider.

• Whether the prescription of additional assessors carries no significant policy implications is precisely one of the questions this submission raises. For an assessment regime that determines the funding of more than 700,000 Australians, the policy weight of deciding who is qualified to administer it is not a small matter.

• Any decision to prioritise a non-clinical workforce carries implications for the existing workforce of regulated health practitioners and its productivity.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

The participant will have unequal recourse. A participant assessed by an AHPRA-registered practitioner can lodge a notification with AHPRA if concerns arise about the conduct or competence of their assessment, and the practitioner is subject to professional standards enforced by a regulatory body on the basis of that practitioner’s individual knowledge and competence. A participant assessed by a non-registered assessor has no equivalent regulatory pathway; the individual assessor is not held to the same standards.

Under this proposal, participants undergoing identical assessments with similar consequences would have materially unequal recourse depending on the type of assessor they have. For the participant, the protection available against assessment error has traditionally stood with the assessor they draw. The choice to utilise a public-servant workforce to carry out what are, in virtually all other settings, clinical health practices, removes this protection.

These concerns are not novel, either. As mentioned earlier in Ray and NDIA [2020] AATA 3452, the Tribunal held that a single-occasion assessment by an assessor unfamiliar with the participant should be given less weight than longitudinal evidence from treating health practitioners. The Bill empowers the assessment model the Tribunal found less reliable: a single-occurrence assessment, by an assessor who in many cases will not be a registered health practitioner, with the obligation to consider independent treating-practitioner evidence, and participant rights to appeal narrowed in the same legislation.

The direction of design. Read together with the wording in new sections 32L(6)(a) and 32K where the process is empowered to generate and act on information about a participant’s needs instead of identifying them, the workforce design features set out above can reasonably be read to imply something stronger than simply the absence of a registration requirement for needs assessors: that the NDIA and its presiding ministers do not consider AHPRA-regulated training, expertise and safety regulations necessary for the conduct of these assessments or the assessors carrying them out, and does not intend such practitioners to be central to the workforce that administers them.

• This, despite the fact that the subject matter – impairment and functional capacity and the supports that compensate for disability – sits squarely within those same training pathways, professional identities, core skills, and competencies of AHPRA-regulated professions.

• Those professions also operate within a regulatory architecture established in the public interest: registration with statutory standards of conduct, defined scopes of practice, mandatory continuing professional development, and complaints and disciplinary mechanisms enforced by the National Boards and AHPRA.

• An Australian Public Service-based workforce, in contrast, sits entirely outside that architecture.

• Disabled Australians have a direct interest in the accuracy of the assessments that will determine their funding, and in regulatory protections that apply when some of those assessments inevitably go wrong.

The committee may wish to seek the NDIA’s operational rationale for designing the assessment regime to place the clinical professions most aligned with the subject matter at the periphery rather than the centre, and to clarify what considerations were the primary drivers of this direction.

The structural fix. Both problems resolve under the same two legislative changes: that the assessor workforce be required by statute to hold AHPRA registration with the relevant scope of practice (Recommendation 3), and that any NDIS rule constraining what information assessors must or must not consider be consulted on with the relevant National Boards of AHPRA before it is made (Recommendation 5). The committee may also wish to consider whether the rules made under section 32L(4A)(c) – which determine who beyond Agency staff and engaged consultants can administer the assessment – should be reclassified from category D to category A, on a parity argument with the information-constraint rules under section 32L(4)(b) and (c).

Together, these changes would ensure that individual practitioners remain accountable to public-facing institutions, and that every participant continues to be assessed by an appropriately trained and regulated practitioner whose information-gathering judgement and professional norms are preserved. They would also reorient the regime to be built around the clinical professions most closely aligned with its subject matter, rather than bolt them onto its periphery or sideline them entirely.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

8. An unregulated clinical instrument

The provisions in combination.

The provisions discussed above are not independent of each other. Schedule 3 Part 2 enables automation of administrative action. Schedule 1, Part 6, ranks individual clinical evidence below peer-reviewed research when determining which supports are reasonable and necessary. Internal NDIA briefings tabled in Senate Estimates (Lyons, 2025) indicate that staff will no longer be required to consider independent medical evidence in planning, and that the Administrative Review Tribunal’s remedial powers will be restricted to ordering reassessment by the same system that produced the contested result.

Subtly redefining assessment as a non-clinical act.

The wording shift in new paragraph 32L(6)(a), from ‘identify’ to ‘include information about’ participant needs, is subtle but appears to have significant consequences. The practical work the assessor performs is extremely similar in each case, with the primary difference being the legal and conceptual status of the work.

• When an assessor’s job is to ‘identify’ needs, their report is considered a clinical determination of needs.

• When the same evaluation protocols are used merely to ‘include information about’, the report is a data input that new section 32K(1A) then applies, through a method specified in NDIS rules, to derive a funding amount, potentially without the participant’s needs ever actually needing to be legally identified.

• This re-framing would seem to act as a linguistic tool that allows the workforce design examined in section 7 to engage in what is functionally a clinical assessment process while plausibly denying that it constitutes a clinical act. Granting the questionable idea that assessors are merely gathering data for determinations made by other processes raises several other questions. Two issues are whether the automated budget allocation method should determine reasonable and necessary supports for a participant whose disability/impairment-related needs have not actually been legally identified, and whether software that functionally absorbs clinical reasoning duties should be placed under stricter regulations.

• Note that the fact the i-CAN authors sought to study its ability to match clinical expert judgement in correctly identifying the level of a participants support needs (i.e., its Classification Accuracy) is itself a tacit admission that the tool has an inherently clinical-evaluative purpose, as is the case for every similar tool.

The transfer of clinical judgment.

Taken together, these provisions create the conditions for the substantive judgment involved in identifying a person’s impairment, determining the supports their impairment requires, and assessing whether those supports are reasonable and necessary (until now the responsibility of regulated health practitioners) to be transferred from health practitioners to an undisclosed software-based decision-making algorithm that determines the same things using information about the participant rather than a clinical determination by the practitioner.

• As section 7 sets out, the human element of the assessment will be carried out by a workforce that does not require a mandatory clinical qualification and will, in practice, primarily collect quantitative data to be entered into an automated algorithmic process designed to absorb, via the wording shift discussed above, decisions that currently rely on the clinical reasoning of health practitioners.

• By design, then, the Bill empowers automated processes within a software program to perform what is currently handled by the clinical judgments of federally registered health practitioners regarding impairment, function, and support, and to do so on the basis of non-clinical inputs from a workforce that will not be required to have a background that facilitates critical engagement with the assessment instruments, and will not be subject to a regulatory framework like AHPRA.

• There is currently no way to know whether the budget allocation software will be capable of making sound judgements about the peer-reviewed literature in individual cases, but this is part of the function it appears to be designed to perform.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

The TGA definition. Software that performs the functions just described in virtually any other setting would feasibly meet the Australian regulatory definition of a software-based medical device under the Therapeutic Goods Act (Parliament of Australia, 1989). The Act and its associated regulations s 41BD(1)(a)(ii) appear to be inclusive of ‘any instrument, apparatus, software, or other article intended for “diagnosis, monitoring, treatment, alleviation of or compensation for an injury or disability”. The wording shift in paragraph 32L(6)(a), discussed above, is what I believe will place the software within this definition when all relevant factors are considered.

• Under the existing Act, a human clinician identifies the needs; downstream software handles calculation.

• Under the Bill, the assessor includes information, and new section 32K(1A) directs that the software apply that information to derive a funding amount. The intended purpose of the budget allocation method is, on the face of the legislation, to determine what supports are reasonable and necessary for a person on the basis of information about their impairment, and allocate funding which is intended to be used to compensate or treat the impairment the information is about.

The Software as a Medical Device framework classifications scale to the risk and significance of the clinical decisions that software drives. An automated software whose outputs determine the level of disability support available to a person, on the basis of what has traditionally been considered private health information via privileged clinical judgements, and whose use the Bill empowers in the absence of independent clinical evidence and meaningful merits review, would appear to fall squarely within that definition.

On standard application of the classification criteria, given the population, the stakes, and the absence of routine human override, the budget allocation system would (in my view) plausibly qualify for the higher-risk classes (IIb or III).

Why existing automation safeguards are not equivalent. Schedule 3 Part 2 contains three safeguards: section 59B(7) lets the CEO substitute the automated action where it is “not correct or preferable”; section 59C(3) requires the Minister to be “satisfied that it is appropriate to do so” before specifying provisions involving evaluative determinations; and section 59D requires a Standard Operating Procedure instrument specifying “sufficiently objective” circumstances for those determinations.

As safeguards, these provisions are substantially weaker than both of the regulatory frameworks Parliament has established for clinical decision-making in Australia: AHPRA’s regulation of human practitioners and the TGA’s regulation of software-based medical devices.

• Against AHPRA protections, the Bill’s safeguards lack the framework Section 7 describes: individual registration, statutory standards of conduct, defined scopes of practice, mandatory continuing professional development, and the notification-and-disciplinary pathway available to participants. The substitute-action power under section 59B(7) sits with the CEO, not the participant, and depends on the CEO having sufficient domain-relevant clinical knowledge pertinent to the case, to somehow exercise an ability to recognise when the automated output of a black-box algorithm is incorrect, and the competence to exercise the clinical judgements that any substitution would require.

• Against TGA SaMD requirements, the Bill’s safeguards lack the standard mechanisms by which the TGA regulates software whose outputs drive clinical decisions: pre-market validation, post-market surveillance, and adverse-event reporting. The ministerial threshold under section 59C(3) is internal to the executive; the TGA’s classification of a SaMD is an external, technical determination by a body whose function is the safety of clinical decision-support. The “sufficiently objective” requirement under section 59D(3) reduces but does not eliminate evaluative judgement, since section 59B(4) explicitly authorises automation of “a discretion being exercised, an evaluative judgement being made, or a state of mind being formed”, much of which has been part of the duties of registered health practitioners.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

The regulatory gap. No regulatory pathway of this kind has been described for the i-CAN customisation, for the budget method that operates downstream of it, or for the integrated system the Bill empowers. The committee may wish to consider whether commencing an automated system that executes what has, until now, been reserved as a duty for health practitioners at this scale, in this regulatory vacuum, is consistent with the safety architecture that Parliament has established for every other clinical software application in Australia.

The fix. If the NDIA sought TGA classification for the integrated assessment-to-funding pathway and demonstrated compliance with the Software as a Medical Device framework, the regulatory anomaly identified in this section would, in my opinion, be substantially resolved. Involving TGA oversight for quality assurance of the budget allocation mechanism, as a software-based medical device, would also increase public confidence in the scheme by alleviating the concern outlined in section 4 and providing an independent mechanism to ensure that the needs assessment processes meet reasonable scientific standards.

9. The Statement of Compatibility with Human Rights

The Bill’s Statement of Compatibility finds the measures compatible with the Convention on the Rights of Persons with Disabilities (CRPD; United Nations 1966a; United Nations, 1966b; United Nations, 2006), the International Covenant on Economic, Social and Cultural Rights (ICESCR), and the International Covenant on Civil and Political Rights (ICCPR), on the ground that they ‘protect and preserve the capacity of the NDIS to deliver supports to people with permanent and significant disability’. That framing is plausible at the level of intent. At the level of mechanism, four features of the Bill warrant closer engagement than the Statement currently provides. Each connects directly to findings set out earlier in this submission.

9.1 The new definition of ‘functional capacity’ conflicts with how the i-CAN is designed

Section 3.6 of this submission noted that the i-CAN was designed to assess support needs in the context of a person’s specific environment, and that this environment-sensitive logic is one of the instrument’s most defensible theoretical features. New section 9B of the Bill moves in the opposite direction: it requires functional capacity to be assessed ‘in a context that excludes, as far as possible, the impact of the person’s environmental and personal circumstances’. Item 10 of Schedule 4 appears to reinforce this narrowing from a second direction by amending paragraph 32L(6)(a) so that the relevant support needs must arise ‘directly’ from the participant’s impairment(s). Where new section 9B excludes environmental and personal circumstances from the assessment context, Item 10 excludes needs that arise indirectly through those same circumstances. The Statement of Compatibility does not engage either narrowing, despite their combined effect on how the social model of disability translates into the assessment regime.

This creates a structural tension. The social model of disability that is affirmed in the Preamble to the CRPD and in Articles 1 and 3 locates disability in the interaction between a person’s impairment and the environmental, atitudinal and structural barriers around them. A statutory definition that requires those barriers to be excluded from the assessment of functional capacity is in tension with that principle, and is also in tension with the design philosophy of the very instrument the Bill empowers. The Statement does not address how environmental factors would be factored into the assessment, if at all, nor how the deployed instrument is expected to handle a contradiction between its design and the statutory framework under which it must operate.

9.2 The reductions fall heaviest on the cohorts the instrument has not been validated for

Section 3.2 of this submission noted that the published validation evidence for the i-CAN rests on a sample that was 84 per cent people with intellectual disability and 91 per cent in congregate residential settings, with no published validation data for people with psychosocial, visual or sensory disabilities. The Explanatory Memorandum elsewhere acknowledges (Parliament of Australia, 2026a, p. 229) that participants with psychosocial disability, visual impairment, Down syndrome and intellectual disability ‘may be more affected’ by the proposed reductions to social, civic and community participation funding, and that 30 per cent of plan funding for participants whose primary disability is psychosocial is currently committed to those supports.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Article 19 of the CRPD requires States Parties to take ‘effective and appropriate measures’ to facilitate full inclusion in the community. The committee may wish to consider what it means, under Article 19, to reduce community-participation supports for the same cohorts for which no version of the assessment instrument has ever been validated to assess. The Statement does not address the distributional impact, nor whether alternatives less restrictive of Article 19 rights were considered.

• Both the CRPD (Article 4(2) ) and the ICESCR (Article 2(1)) require States Parties to progressively realise the relevant rights to the maximum of available resources. As interpreted by the Committee on Economic, Social and Cultural Rights (General Comment No. 3) and applied by the Committee on the Rights of Persons with Disabilities, this obligation carries a strong presumption against measures that go backwards, which can be justified only after careful consideration of alternatives and only by reference to the totality of rights protected by the treaty.

New section 34A authorises the Minister to reduce funding for a specified group of supports by legislative instrument, with effect ‘even if the result is … the funding provided under a participant’s plan for a reasonable and necessary support is less than the total cost of the support’. Read together with Section 6 of this submission, which examines the narrowing of error-correction mechanisms in the same legislative package, this raises a question that the Statement does not engage with: how is enabling funding for an acknowledged reasonable and necessary support to fall below its cost consistent with the obligation of progressive realisation? This is particularly important to consider for participants whose plan arrangements make it a breach of the rules for providers to charge gap fees, which would likely result in providers withdrawing services, or risking civil conscription (see section 10).

The combination of validation gaps and automation engages Articles 5 and 12

Article 5 requires States Parties to guarantee equal and effective legal protection against discrimination because of disability. Where the instrument that determines a person’s funding has been validated against a population that is 84 per cent intellectual disability in 91 per cent congregate settings (Riches et al., 2009a)), and has not been validated for the deployed v6 or its NDIA customisation at all (CDS, 2025b), the foreseeable misclassification of under-represented disability cohorts is a substantive Article 5 question.

Article 12 requires States Parties to recognise that persons with disabilities enjoy legal capacity on an equal basis with others, and to provide support for the exercise of that capacity. Section 8 of this submission sets out the case for treating the assessment-to-funding pathway as a software-based medical device, on the grounds that it performs determinations traditionally made by regulated health practitioners.

• New section 59B authorises computer programs to make ‘evaluative determinations’, defined as ‘a discretion being exercised, an evaluative judgement being made, or a state of mind being formed’.

• Determining a person’s reasonable and necessary supports is the exercise of legal capacity over a matter of direct consequence to their life. Whether the safeguards in sections 59B(7), 59C(3), and 59D are sufficient to discharge the Article 12 obligation is, ultimately, a legal question.

• As Section 8 sets out, however, those safeguards are structurally weaker than both of the regulatory frameworks Parliament has otherwise established for clinical decision-making: AHPRA’s regulation of individual practitioners, and the TGA’s regulation of software-based medical devices. The safeguards do not provide a participant with the notification-and-disciplinary pathway available against an individual practitioner, nor the pre-market validation, post-market surveillance, and adverse-event reporting that the TGA imposes on equivalent software.

That the Statement of Compatibility does not engage Article 12 in connection with the automation provisions, given this combination, is, on any reading, an omission.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

The four points set out in 9.1 to 9.4 are observations on which the Statement of Compatibility’s findings appear to diverge from features of the Bill that this submission examines on methodological grounds. The committee may wish to seek targeted advice from the Parliamentary Joint Committee on Human Rights, the Australian Human Rights Commission, or independent CRPD legal specialists on whether the Statement’s compatibility findings stand once the scientific evidence set out in Sections 3 to 8 has been taken into account.

10. A separate constitutional consideration

The provisions examined above also raise a constitutional question that lies outside the scope of this submission but warrants the committee’s attention. The civil conscription limitation in section 51(xxiiiA) of the Constitution, as developed in British Medical Association v Commonwealth (1949), General Practitioners Society v Commonwealth (1980), and Wong v Commonwealth (2009), restricts the Commonwealth’s power to legislate in ways that compel medical or allied health practitioners to deliver services on terms set by the Commonwealth.

New section 32K(3C) authorises funding amounts in the budget method itself to be below the actual cost of a reasonable and necessary support, as a default rule of the new framework. New section 34A authorises a parallel reduction by ministerial determination for old framework plans during the transition period. Read together with NDIS rules that prohibit providers from charging gap fees to bridge the difference, this combination may engage the civil conscription limitation. Because section 32K(3C) operates by default rather than by ministerial discretion, the engagement of the limitation under it is arguably more direct than under section 34A.

The committee may wish to seek targeted constitutional advice on whether new sections 32K(3C) and 34A, read with the relevant gap-fee rules, are consistent with section 51(xxiiiA), and to consider whether commencement of those provisions should be conditional on that advice.

11. The ‘gold standard’ characterisation

The NDIA has described the i-CAN as ‘the gold standard of available, validated needs assessment tools’ (NDIA, 2025). As discussed above, the published evidence does not support that characterisation, for any prior version; and as the version that will be deployed has no published validation evidence at all, it necessarily follows that the notion it approximates a gold standard in validated needs assessment is not supported by the available evidence.

In stating this, I wish to be clear that this submission does not oppose the i-CAN in principle, nor does it object to actuarial methods in resource allocation or clinical decision-making; substantial scientific literature (e.g., following research by Paul Meehl) demonstrates how this can be done well when executed to assist structure and guide clinical reasoning, not replace it.

There is also much to admire about the CDS, such as its 20-year commitment to developing the i-CAN instrument in a geographical context where measurement innovations of this kind have largely stagnated. As stated earlier, I consider the i-CAN framework to have genuine conceptual strengths. This submission’s concerns about the i-CAN are narrow, concerning the facts that a promising instrument, still very much under development, is being linked to an automated decision-making algorithm with undisclosed properties that performs traditional clinical decision-making duties, deployed by a workforce whose domain-relevant expertise and professional obligations are not assured, against populations and use-cases for which it has not been validated, all under a legislative framework that does not require disclosure of its translation logic. On the grounds of scientific method and ethical obligations, the premature deployment of this promising instrument for a questionable purpose is the subject of this submission’s scrutiny, not the i-CAN or its authors per se.

This submission also does not oppose the idea that the scheme requires serious fiscal and regulatory reforms. What it contests is the possibility that fiscal pressure justifies commencing the specific assessment-to-funding architecture that the Bill is set to empower, based on the evidence currently in the public domain. In my view, those two questions are separable. That reform to the NDIS is needed does not establish that this particular reform proposal, on these foundations, is the one that should be commenced.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

The position this submission takes is also contingent on the current state of the evidece. If the conditions identified in the recommendations were met (independent validation, methodological rigour, an AHPRA-registered workforce, and appropriate oversight), the system the Bill empowers would become both much more defensible and more useful to the public. However, the current evidence is not consistent with any of these conditions being met satisfactorily.

12. Recommendations

Based on my analysis, I provide the following recommendations.

• Recommendation 1. [CRITICAL] Defer commencement of Schedule 4 and the dependent provisions of Schedule 1 until independent scientific validation of the deployed instrument, across the full NDIS population demographics, has been published in peer-reviewed literature and assessed by an expert panel that has no ties to the instrument, its developers, or the NDIA.

• Recommendation 2. [CRITICAL] Amend Schedule 4 to require, in primary legislation, publication of the budget method’s key methodological details, including validation metrics for the assessment-to-funding translation algorithm in a form open to reasonable scientific scrutiny and/or independent regulatory authority (e.g., the TGA, per recommendation 9), before any participant is assessed under it.

• Recommendation 3. Amend Schedule 4 to require that all support needs assessors be registered health practitioners under the Australian Health Practitioner Regulation Agency with a relevant scope of practice in accordance with existing legislation and codes of conduct.

• Recommendation 4. Amend Schedule 1 Part 6 to restore the obligation to consider independent medical and allied health evidence in planning decisions, and to clarify that the new evidence hierarchy operates as a guide rather than a precondition.

• Recommendation 5. Amend Schedule 4, Item 8 (new section 32L(4)(b) to (c)) so that any NDIS rule prescribing information an assessor must or must not have regard to is consulted on with the relevant National Boards of the Australian Health Practitioner Regulation Agency, and so that the rule must not, in principle, direct assessors to disregard information the assessor reasonably believes to be relevant to the assessment. – In addition, amend Schedule 4 so that NDIS rules made under section 32L(4A)(c) – which prescribe persons additional to Agency staff and engaged consultants who may undertake the assessment – are reclassified from category D to category A, so that there is parity regarding the information-constraint rules under section 32L(4)(b) and (c) (Item 13).

• Recommendation 6. [CRITICAL] Amend Schedule 1 to preserve the Administrative Review Tribunal’s full remedial powers, including the power to substitute or vary a plan decision.

• Recommendation 7. Amend Schedule 1 Part 4 to make ministerial determinations subject to sunsetting and parliamentary disallowance.

• Recommendation 8. [CRITICAL] Insert a provision requiring independent biennial public review of the assessment framework’s measurement performance, decision-change rates, and population-level outcomes, conducted by a body without commercial or contractual ties to the instrument or its developers, and without political interests in the perceived success of the framework.

• Recommendation 9. Require regulatory oversight under the Therapeutic Goods Act 1989 of any automated or algorithmic tool used to determine support needs assessments or funding allocations under the NDIS. Software that takes private health information related to a person’s impairment and produces outputs that determine the supports required to compensate for that impairment feasibly

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

• meets the definition of a software-based medical device under the Act, in which case it would be regulated as Software as a Medical Device in any other clinical or commercial setting.

• The committee may wish to recommend that the Department of Health, Disability and Ageing and the TGA jointly issue guidance on the application of this framework to the assessment-to- funding pathway empowered by Schedule 4 and Schedule 3 Part 2, and that commencement of those provisions be conditional on appropriate regulatory classification and oversight.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

References

ABC News. (2021, April 23). Former NDI chairman slams changes to support scheme. https://www.abc.net.au/news/2021-04-23/former-ndis-chairman -slams-changes-to-support- scheme/100091348 Administrative Review Tribunal. (2025). ART caseload report: For the period 14 October 2024 to 30 June 2025. Australian Government. https://www.art.gov.au/help-and-resources/statistics American Educational Research Association, American Psychological Association, & National Council on Measurement in Education. (2014). Standards for educational and psychological testing. Arnold, S. R. C., Riches, V. C., & Stancliffe, R. J. (2014). I-CAN: The classification and prediction of support needs. Journal of Applied Research in Intellectual Disabilities, 27(2), 97–111. https://doi.org/10.1111/jar.12055 Arnold, S. R. C., Riches, V. C., & Stancliffe, R. J. (2015). Does a Measure of Support Needs Predict Funding Need Better Than a Measure of Adaptive and Maladaptive Behavior? American Journal on Intellectual and Developmental Disabilities, 120(5), 375–394. https://doi.org/10.1352/1944-7558-120.5.375 Australian Psychological Society. (2014). Ethical guidelines for psychological assessment and the use of psychological tests. British Medical Association v Commonwealth [1949] HCA 44; (1949) 79 CLR 201. Centre for Disability Studies, University of Sydney. (2025a, November). List of I-CAN research publications. https://cds.org.au/wp-content/uploads/2025/11/List-of-I-CAN-Research-Publications.pdf Centre for Disability Studies, University of Sydney. (2025b, December). Frequently asked questions: I-CAN v6 and National Disability Insurance Scheme (NDIS) Support Needs Assessment. https://cds.org.au/wp- content/uploads/2025/12/I-CAN-v6-Support-Needs-Assessments-FAQs.pdf Cicchetti, D. V. (1994). Guidelines, criteria, and rules of thumb for evaluating normed and standardized assessment instruments in psychology. Psychological Assessment, 6(4), 284–290. https://doi.org/10.1037/1040-3590.6.4.284 Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Lawrence Erlbaum Associates. General Practitioners Society v Commonwealth [1980] HCA 30; (1980) 145 CLR 532. Harrell, F. E., Jr. (2015). Regression modeling strategies: With applications to linear models, logistic and ordinal regression, and survival analysis (2nd ed.). Springer. https://doi.org/10.1007/978-3-319-19425-7 International Test Commission. (2013). ITC guidelines on test use (Version 1.2). https://www.intestcom.org/files/guideline_test_use.pdf Lyons, K. (2025, December 3). NDIS plans will be computer-generated, with human involvement dramatically cut under sweeping overhaul. The Guardian. https://www.theguardian.com.australia- nnews/2025/dec/03/ndis-plans-computer-generated McHugh, M. L. (2012). Interrater reliability: The kappa statistic. Biochemia Medica, 22(3), 276–282. https://doi.org/10.11613/BM.2012.031 National Disability Insurance Agency. (2025). Developing a new support needs assessment (News item 10926). Australian Government. https://www.ndis.gov.au/news/10926-developing-new-support-needs- assessment Parliament of Australia. (1989). Therapeutic Goods Act 1989 (Cth). https://www.legislation.gov.au/C2004A03952 Parliament of Australia. (2013). National Disability Insurance Scheme Act 2013 (Cth). https://www.legislation.gov.au/C2013A00020 Parliament of Australia. (2026a). Explanatory Memorandum to the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026 (Cth). Parliament of Australia. (2026b). National Disability Insurance Scheme Amendment (Securing the NDI for Future Generations) Bill 2026 (Cth). Psychology Board of Australia. (2024). Code of conduct for psychologists [Effective December 1, 2025]. Australian Health Practitioner Regulation Agency. https://www.psychologyboard.gov.au/standards-and- guidelines/code-of-conduct.aspx Ray and National Disability Insurance Agency [2020] AATA 3452. Riches, V. C., Parmenter, T. R., Llewellyn, G., Hindmarsh, G., & Chan, J. (2009a). I-CAN: A new instrument to classify support needs for people with disability: Part I. Journal of Applied Research in Intellectual Disabilities, 22(4), 326–339. https://doi.org/10.1111/j.1468-3148.2008.00466.x

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Riches, V. C., Parmenter, T. R., Llewellyn, G., Hindmarsh, G., & Chan, J. (2009b). The reliability, validity and practical utility of measuring supports using the I-CAN instrument: Part II. Journal of Applied Research in Intellectual Disabilities, 22(4), 340–353. https://doi.org/10.1111/j.1468-3148.2008.00467.x

Royal Commission into the Robodebt Scheme. (2023). Report. Commonwealth of Australia. https://robodebt.royalcommission.gov.au/publications/report

Senate Community Affairs Legislation Committee. (2025). Supplementary Estimates 2025–26: Document tabled by Senator Steele-John regarding NDIS New Framework Planning [K. Lyons, The Guardian, December 3, 2025]. Parliament of Australia. https://www.aph.gov.au/-/media/Estimates/ca/supp2526/Health_Disability_and_Ageing/15_TabledDoc_SenatorSteeleJohn.pdf

Stancliffe, R. J., Arnold, S. R. C., & Riches, V. C. (2016). The supports paradigm. In R. L. Schalock & K. D. Keith (Eds.), Cross-cultural quality of life: Enhancing the lives of people with intellectual disability (2nd ed., pp. 133–142). American Association on Intellectual and Developmental Disabilities.

Suen, H. K., & Ary, D. (1989). Analyzing quantitative behavioral observation data. Lawrence Erlbaum Associates.

United Nations. (1966a). International Covenant on Civil and Political Rights. United Nations Treaty Series, 999, 171. https://treaties.un.org/Pages/ViewDetails.aspx?src=TREATY&mtdsg_no=IV-4&chapter=4

United Nations. (1966b). International Covenant on Economic, Social and Cultural Rights. United Nations Treaty Series, 993, 3. https://treaties.un.org/Pages/ViewDetails.aspx?src=TREATY&mtdsg_no=IV-3&chapter=4

United Nations. (2006). Convention on the Rights of Persons with Disabilities. United Nations Treaty Series,2515, 3. https://treaties.un.org/Pages/ViewDetails.aspx?src=TREATY&mtdsg_no=IV-15&chapter=4

van Toorn, G., & Carney, T. (2024). Decoding the algorithmic operations of Australia’s National Disability Insurance Scheme.Australian Journal of Social Issues. Advance online publication. https://doi.org/10.1002/ajs4.342

Wong v Commonwealth [2009] HCA 3; (2009) 236 CLR 573.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Appendix A - Simulation-Based Precision and Power Analysis: i-CAN Criterion Validity, Test-Retest Reliability, and Predictive Validity

Overview

This appendix contains the output of the precision and power analyses from section 3 of the submission. The analyses are based on Monte Carlo simulations using 10,000 replications per condition with fixed seeds. Monte Carlo simulation is a statistical approach that involves repeatedly simulating a study to assess the robustness of its results. The simulations treat the published i-CAN sample statistics as though they are the true values for the whole population. The point of treating them this way is to analyse whether the sample sizes used in the published validation literature are sufficient to estimate the coefficients with the precision required for high-stakes assessment. As described below and in the submission, the analyses find that this is not the case and that substantial work remains to achieve this. If the true values are lower than the estimates reported by the studies, their precision is worse, not better. Three sets of analyses are reported.

---

How to Read This Appendix

For whatever we are studying, the sample of a study gives us a brief snapshot of varying resolution. Depending on that resolution, the agreement statistics and correlations a study produces are estimates that would differ if the same study were repeated with a different sample of similar size. A 95% confidence interval (CI) describes how much each attempt at the same study would differ in the results it observes: if the published value is taken as the truth, then 95 times out of 100, the effects/discovered relationships in a fresh study at the same sample size would land somewhere inside the interval. The wider the interval, the less an individual study can say about how strong the true relationship actually is.

Width (CI high − CI low) is the precision measure used throughout this Appendix. A narrow interval means a study can support a confident statement about the underlying value. A wide interval means the result is consistent with several different conclusions and that the data cannot, on its own, support any of them over any other.

Statistical Power is the probability that a study will detect a real effect if one exists. A study with 80% power has a one-in-five chance of missing a true effect; at 95% power, one in twenty. However, low power not only reduces the chances of a true discovery; it also increases the chances of a false discovery. Relationships discovered in studies with low power tend to appear much larger than they actually are, with only larger-than-average effects clearing the threshold for statistical significance; as such, the chance of a false positive increases the less statistical power there is (Ioannidis, 2005; Button et al., 2013; Gelman & Carlin, 2014). This is sometimes called the ‘winner’s curse’. These issues sit at the centre of what is widely known as the ‘replication crisis’, and they are part of why independent replication, as raised in section 3.5 of the submission, is so important.

How precise an estimate needs to be depends on the decisions it will inform. Recent work on sample size planning (Maxwell, Kelley, & Rausch, 2008) treats this as the first design question: what width would the CI need to be for the result to be useful? Two width targets are used in this Appendix, both proposed in the body of the submission, calibrated to the decisions an NDIS Support Needs Assessment will drive. The principle in each case is the same: the interval should be narrow enough that the data can distinguish between the practically meaningful categories of the metric.

• For κ (the criterion-validity statistic): width ≤ .05. The standard kappa benchmarks define agreement bands roughly 0.10 wide (e.g., McHugh’s, 2012, ‘strong agreement’ band, .80–.89). A width of 0.05 is half that, keeping the CI inside a single band: at κ = .85, the CI runs from roughly .82 to .88, clearly inside ‘strong agreement’. At the widths the published studies actually produce (.15 to .37), a single point estimate spans two or three categories, and the data cannot show which one the true value belongs in.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

• For ρ (the predictive-validity correlation): width ≤ .10. Cohen’s (1988) standard effect-size categories for correlations (small ≈ .10, medium ≈ .30, large ≈ .50) sit roughly 0.20 apart. A width of 0.10 is half that, keeping the CI inside a single category: at ρ = .19 with this width, the CI runs from about .14 to .24, a modest but real relationship. At the width the published N = 163 actually produces (.30), the CI runs from .04 to .34, straddling ‘effectively no relationship’ and ‘a meaningful relationship’. A single study at this width cannot support a decision about whether the instrument predicts funding need.

Benchmarks Invoked in This Appendix

The following thresholds are cited throughout. Values labelled ‘proposed in this submission’ are the high-stakes values argued for in the body of the submission; the lower thresholds are conventional values from the cited sources.

• κ ≥ .60: acceptable for observational research (Suen & Ary, 1989).
• κ ≥ .75: ‘excellent’ threshold (Cicchetti, 1994).
• κ ≥ .80: ‘strong agreement’ (McHugh, 2012); proposed in this submission as the minimum for high-stakes individual NDIS allocation.
• Power ≥ 80%: conventional minimum for exploratory research (Cohen, 1988).
• Power ≥ 95%: proposed in this submission as the minimum for high-stakes applied use.
• CI width ≤ .05 (for κ) and ≤ .10 (for ρ): proposed in this submission, with the rationale set out in ‘How to Read This Appendix’ above.
• $N : k$ ≥ 15: recommended for prediction-model stability (Harrell, 2015).

Abbreviations

A&P = Activities and Participation (composite of seven sub-domains); AKGT = Applying Knowledge and General Tasks; CSCS = Community, Social and Civic Life (original-paper spelling preserved); H&WB = Health and Well-Being; ICAP = Inventory for Client and Agency Planning (Bruininks et al., 1986); iir = Interpersonal Interaction and Relationships; MEH = Mental Emotional Health; sCDL = Self-Care and Domestic Life.

A Note on Interval Construction

The intervals reported in this Appendix are based on running each study 10,000 times in simulation, with the published value held fixed as if it were the true value for the whole population. The middle 95% of the resulting estimates define the 95% interval. In plain terms: it a study could be repeated many times under the assumption that the published value is correct, this is the range its results would fall in 95 times out of 100. One implication follows directly: the intervals reported here are an optimistic measure of precision. If the true effect is in fact weaker than the published study reports, The actual interval would be wider still.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Part I. Criterion Validity (Cohen’s κ)

Part I addresses criterion validity (Cohen’s κ) from Arnold, Riches, and Stancliffe (2014). The headline value of κ = .94 is derived from 49 reports (26 per cent of the development team’s 186-report sample) that the developers themselves rated as ‘good’ or ‘very good’ quality. Including reports rated ‘average’ reduces κ to .75 (n = 114); on all 186 reports, κ = .65; the independent validation sample (n = 41) produced κ = .66. Part I quantifies what each of these values can and cannot establish at the published sample sizes. Simulations based on published values from Arnold, Riches, and Stancliffe (2014).

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Table A3

Power to Demonstrate $\kappa$ = .94 Against the Conventional Threshold ($\kappa \leq$.75)

n	Power $\geq$ 80%	Power $\geq$ 95%
30	76.0%	No
40	89.6%	No
49	94.4%	No
60	97.6%	Yes
80	99.3%	Yes
100	99.9%	Yes
150	100.0%	Yes

Note. Power derived from Monte Carlo simulation (10,000 replications); one-sided α = .05 against H₀: κ ≤ .75 (Cicchetti, 1994, ‘excellent’ threshold). The 95th per centile of the simulated null distribution serves as the critical value; power is the proportion of simulations under κ = .94 exceeding this value. Benchmarks: 80% power conventional for exploratory research (Cohen, 1988); 95% proposed in this submission for high-stakes applied use. At the published n = 49, power against the conventional threshold is 94.4%, just below the 95% high-stakes minimum. Achieving 95% power requires $\geq$ 60.

Table A4

Power to Demonstrate $\kappa$ = .94 Against the High-Stakes Threshold ($\kappa \leq$.80)

n	Power $\geq$ 80%	Power $\geq$ 95%
30	57.0%	No
40	70.3%	No
49	77.4%	No
60	86.9%	No
80	94.5%	No
100	97.5%	Yes
120	99.1%	Yes
150	99.8%	Yes
200	100.0%	Yes

Note. Power derived as in Table A3, against H₀: κ ≤ .80. The .80 threshold corresponds to McHugh’s (2012) ‘strong agreement’ category and is proposed in this submission as the minimum for high-stakes individual NDIS allocation. At the published n = 49, the power against this threshold is 77.4%, below the conventional 80% minimum. Achieving 80% power requires $\geq$ 60; reaching the 95% high-stakes power minimum requires $\geq$ 100.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Table A5

Power to Demonstrate Full-Sample κ Exceeds the Suen and Ary (1989) Threshold

n	Power ≥ 80%	Power ≥ 95%
41	12.8%	No
49	12.4%	No
100	21.0%	No
114	22.2%	No
186	30.1%	No
300	40.1%	No
500	56.1%	No
1,000	82.5%	Yes
2,000	97.7%	Yes
3,000	99.8%	Yes

Note. Power derived as in Table A3, against H₀: κ ≤ .60 (Suen & Ary, 1989). True κ = .65 (the full-sample value from Arnold et al., 2014, N = 186). At the published N = 186, power to demonstrate the full-sample κ exceeds .60 is 30.1%; at the validation sample (n = 41), 12.8%. Reaching 80% power requires N ≥ 1,000; 95% power requires N ≥ 2,000.

Table A6

Power to Detect Departures from κ = .94 (n = 49)

True κ	.	Power
.60		99.9%
.65		99.3%
.70		97.0%
.75		90.5%
.80		74.6%
.85		47.1%
.90		15.7%

Note. Power derived from Monte Carlo simulation (10,000 replications). For each candidate true κ, the table reports how often a study at n = 49 would correctly conclude that the true value is below .94. The calculation is technically two-sided at α = .05, but only the lower-direction rejection matters in practice (the probability of falsely concluding the true value is above .94 is essentially zero). At n = 49, the study has 74.6% power to detect a true κ as low as .80, falling to 47.1% for .85 and 15.7% for .90. The published sample cannot reliably distinguish κ = .94 from values plausibly within McHugh’s (2012) ‘strong agreement’ range.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Table A7

Power to Detect Departures from κ = .65 (N = 186)

True κ	Power
.40	99.9%
.45	98.8%
.50	89.8%
.55	59.1%
.60	20.1%

Note. Power derived as in Table A6, with H₀ centred on the published full-sample value (κ = .65, N = 186). At N = 186, the study has 99.9% power to detect a true κ as low as .40, but only 20.1% power to detect a true κ of .60. Small downward departures from the published value cannot be reliably distinguished.

Table A8

Sensitivity of κ Precision to Marginal Configuration

Marginals	Mean	CI low	CI high	Width
Observed	.938	.847	1.000	.153
Equal	.939	.846	1.000	.154

Note. Intervals derived from Monte Carlo simulation (10,000 replications) at κ = .94, n = 49. The ‘Observed’ row uses the marginal distribution from Arnold et al. (2014, Table 10); the ‘Equal’ row imposes 20% per category. Sensitivity of the precision estimate to the marginal configuration is negligible (widths .153 versus .154), supporting the use of observed marginals as the primary configuration for Part I.

Part II. Test-Retest Reliability (Spearman’s ρ)

Part II addresses test-retest reliability (Spearman’s ρ) as reported by Riches et al. (2009b). Sample A is the to-12-month retest (n = 23); Sample B is the two-year retest (n = 20). Sample B carries a version confound: participants were assessed on Version 2 and reassessed on Version 3, with major revisions to the Mental Emotional Health domain between versions, and a 33 percent attrition rate (30 intended, 20 completed). Part II quantifies the precision of each domain’s reliability estimate at the achieved sample sizes. Simulations based on published values from Riches et al. (2009b).

Table A9

Latent-Threshold Configurations for the Ordinal Simulation

Configuration	Cumulative proportions (categories 0–4)
Equal-width	.17, .33, .50, .67, .83
Floor-heavy	.40, .55, .70, .82, .92

Note. The two configurations apply to the latent variable underlying the discretisation of bivariate normal draws into ordinal values 0–5. The equal-width configuration assigns approximately 17 per cent of the distribution to each category. The floor-heavy configuration assigns approximately 40 per cent to the lowest category, with the remainder distributed across categories 1–5. Equal-width is used as the primary configuration; floor-heavy is included for sensitivity analysis (Table A13).

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Table A10

Precision of Spearman’s ρ for the 6-to-12 Month Retest (n = 23)

Domain	ρ	CI low	CI high	Width	Power	Spans 0
Mobility (A&P highest)	.93	.834	.975	.142	100.0%	no
SCDL	.86	.689	.946	.257	100.0%	no
A&P Total	.84	.653	.938	.286	100.0%	no
AKGT	.84	.646	.936	.290	100.0%	no
IIR	.78	.548	.913	.365	99.6%	no
Communication	.74	.483	.898	.415	99.0%	no
Behaviour	.66	.351	.861	.511	94.2%	no
CSCS (A&P lowest)	.60	.260	.832	.572	86.5%	no
Mental Emotional Health	.30	-.113	.654	.768	27.6%	YES
Physical Health	.21	-.208	.590	.798	15.6%	YES

Note. Intervals derived from Monte Carlo simulation (10,000 replications) using a latent-threshold ordinal model with equal-width thresholds and bivariate normal latent variables; the latent correlation is calibrated so that the simulated ordinal Spearman’s ρ matches the published value (Riches et al., 2009b, Table 2). Width = CI high − CI low; Power tests ρ = 0 at two-sided α = .05; Spans 0 indicates whether the CI includes zero. Abbreviations defined in the abbreviations block. At n = 23, Mental Emotional Health and Physical Health produce CIs that span zero: a true ρ of zero cannot be ruled out for these domains at the published sample size.

Table A11

Precision of Spearman’s ρ for the Two-Year Retest (n = 20)

Domain	ρ	CI low	CI high	Width	Power	Spans 0
Mobility (A&P highest)	.94	.843	.983	.140	100.0%	no
SCDL	.76	.481	.910	.429	98.6%	no
AKGT	.74	.449	.901	.451	97.8%	no
A&P Total	.73	.430	.899	.469	97.2%	no
IIR	.66	.318	.867	.550	91.4%	no
Communication	.62	.241	.851	.610	85.5%	no
CSCS (A&P lowest)	.46	.027	.766	.739	54.8%	no
Physical Health	.51	.094	.799	.705	65.5%	no
Behaviour	.25	-.212	.637	.849	19.9%	YES
Mental Emotional Health	-.22	-.619	.238	.857	16.0%	YES

Note. Intervals derived as in Table A10, at n = 20 (Riches et al., 2009b, Table 2). The two-year retest carries a version confound: participants were assessed on Version 2 and reassessed on Version 3, with substantial revisions to the Mental Emotional Health domain between versions (Riches et al., 2009b, p. 350), and a 33 per cent attrition rate (30 intended, 20 completed). At n = 20, Mental Emotional Health andBehaviour produce CIs that span zero. The negative MEH coefficient is consistent with the domain having been substantively redefined between assessment occasions.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Table A12

Power to Distinguish $\rho$ from Zero by Sample Size (Health and Well-Being Domains)

Domain (interval)	$\rho$	n = 20	n = 50	n = 100	n = 150	n = 200	n = 300	n = 500
Mental Emotional Health (2 yr)	-.22	16.0%	32.8%	57.0%	77.7%	86.4%	96.3%	99.9%
Behaviour (2 yr)	.25	19.0%	42.0%	71.3%	86.9%	94.8%	99.1%	100.0%
Physical Health (6–12 m)	.21	14.5%	31.2%	54.5%	72.7%	85.6%	95.7%	99.7%
Mental Emotional Health (6–12	m)	.30	26.4%	56.0%	86.3%	96.3%	99.3%	100.0%
Physical Health (2 yr)	.51	65.3%	97.0%	100.0%	100.0%	100.0%	100.0%	100.0%

Note. Power derived from Monte Carlo simulation (10,000 replications) using the latent-threshold ordinal model with equal-width thresholds; two-sided $\alpha$ = .05. The five domains shown are those most relevant to psychosocial support funding and where the i-CAN’s evidence base is weakest. Reaching the 95% high-stakes power minimum proposed in this submission requires n ≈ 290 for $\rho$ = .21, n ≈ 270 for $\rho$ = -.22, and n ≈ 200 for $\rho$ = .25. The published n = 20 and n = 23 fall far below these requirements.

Table A13

Sensitivity of Reliability Estimates to Threshold Configuration

Domain	$\rho$	Equal-width CI low	Equal-width CI high	Floor-heavy CI low	Floor-heavy CI high
A&P Total (2 yr)	.73	.429	.897	.410	.915
Mental Emotional Health	- .22	- .620	.239	- .621	.237
Behaviour (2 yr)	.25	- .207	.639	- .207	.647

Note. Intervals derived from Monte Carlo simulation (10,000 replications) at n = 20, comparing the equal-width and floor-heavy threshold configurations defined in Table A9. Sensitivity of the precision estimate to the threshold configuration is negligible across the three domains tested (widths within ±.02), supporting the use of equal-width thresholds as the primary configuration for Part II.

Part III. Predictive Validity (Spearman’s $\rho$ with Funding Allocation)

Part III addresses predictive validity (Spearman’s $\rho$) for funding allocation, as reported by Arnold, Riches, and Stancliffe (2015). The published Spearman correlations between i-CAN scores and individualised funding allocations under standard scoring are $\rho$ = .19 (Total Raw), $\rho$ = .13 (Health and Well-Being), and $\rho$ = .20 (Activities and Participation), with N = 163. Part III quantifies the precision of these estimates and the sample sizes required to detect them reliably.

The variables are i-CAN domain scores (aggregate sums of ordinal item ratings) and annual individualised funding allocation (Australian dollars). Two scoring methods appear in the published Table 4: the standard administration protocol (which is the basis for the NDIS deployment) and a post-hoc ‘funded items’ coding (which is exploratory, was not part of standard administration, and is not deployed). The ICAP comparator is included as a benchmark. The sample is N = 163, mixed disability, with person- centred funding allocations already in place. The instrument used in the published study is the i-CAN-Brief Research version, not the full instrument and not v6.

Part III approximates the underlying distribution as bivariate normal, calibrated so that the simulated Spearman $\rho$ matches the published value. Under heavier-tailed or more skewed distributions (likely for dollar-valued funding allocations), the sampling variance of $\rho$ would be no smaller and probably slightly larger, so the precision estimates reported here are conservative: actual published-study CI widths are no narrower than what is shown.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Table A14

Precision of Predictive Correlations Under Standard i-CAN Scoring (N = 163)

Score	ρ	CI low	CI high	Width	Power	Spans 0
Total Raw	.19	.035	.333	.298	68.6%	no
Health and Well-Being	.13	-.026	.281	.307	38.1%	YES
Activities and Participation	.20	.049	.340	.292	72.3%	no

Note. Intervals derived from Monte Carlo simulation (10,000 replications) using bivariate normal draws with population ρ from Arnold et al. (2015, Table 4); standard scoring is the published protocol and the basis for the NDIS deployment. Width = CI high − CI low; Power tests ρ = 0 at two-sided α = .05; Spans 0 indicates whether the CI includes zero. The Health and Well-Being CI spans zero at the published N = 163: a true ρ of zero cannot be ruled out for this domain.

Table A15

Precision of Predictive Correlations for the ICAP Comparator (N = 163)

Score	ρ	CI low	CI high	Width	Power	Spans 0
Service Score	-.19	-.337	-.038	.299	68.1%	no
Broad Independence	-.25	-.390	-.100	.290	89.4%	no

Note. Intervals derived as in Table A14. The ICAP (Bruininks et al., 1986) is included as a comparator from the same Arnold et al. (2015) study. Correlations are negative because higher ICAP scores indicate greater independence, hence lower support need.

Table A16

Precision of Predictive Correlations Under Post-Hoc Funded Items Scoring (N = 163)

Score	ρ	CI low	CI high	Width	Power	Spans 0
Total	.63	.519	.720	.201	100.0%	no
Health and Well-Being	.41	.272	.535	.263	100.0%	no
Activities and Participation	.65	.544	.736	.192	100.0%	no

Note. Intervals derived as in Table A14. Funded Items scoring is a post-hoc exploratory recoding that was not part of standard administration and is not the basis for the deployed NDIS protocol. The substantially higher correlations under this scoring (in contrast to the standard scoring in Table A14) reflect the recoding rather than improved measurement of underlying support need.

Table A17

Power to Detect a Non-Zero Correlation Under Standard i-CAN Scoring

Score	ρ	n = 50	n = 100	n = 163	n = 200	n = 300	n = 500	n = 800	n = 1,000
Total Raw	.19	23.8%	49.2%	69.3%	78.0%	91.0%	99.0%	99.9%	100.0%
Health and Well-Being	.13	13.2%	25.9%	38.1%	45.9%	59.8%	84.2%	95.6%	98.7%
Activities and Participation	.20	26.8%	52.2%	72.7%	81.9%	93.0%	99.7%	100.0%	100.0%

Note. Power derived from Monte Carlo simulation (10,000 replications); bivariate normal draws; two-sided α = .05. ρ values from Arnold et al. (2015, Table 4). For i-CAN Total Raw ($\rho$ = .19), the sample of N = 163 used in the published study achieves 69.3% power against a two-sided null. Reaching 80% power (conventional minimum; Cohen, 1988) requires approximately N = 200; reaching the 95% high-stakes minimum proposed in this submission requires approximately N = 500 for Total Raw and Activities and Participation, and N = 800 for Health and Well-Being.

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

Table A18

Confidence Interval Width by Sample Size at ρ = .19

n	CI low	CI high	Width	Width ≤ .10
50	-.095	.451	.546	No
100	-.007	.374	.381	No
163	.039	.338	.299	No
200	.052	.323	.270	No
300	.080	.296	.217	No
500	.103	.274	.171	No
1,000	.129	.248	.119	No
1,500	.141	.239	.098	Yes
2,000	.147	.233	.086	Yes

Note. Intervals derived from Monte Carlo simulation (10,000 replications); bivariate normal draws; ρ = .19 from Arnold et al. (2015, i-CAN Total Raw, standard scoring). The .10 width target, set out in ‘How to Read This Appendix’ above, distinguishes a near-zero from a meaningful relationship at ρ = .19. The target is met at n ≥ 1 500; the published N = 163 produces a width of .299, meaning the true ρ could plausibly lie anywhere from .04 to .34.

Table A19

Analytical R² 95% Confidence Intervals for Predictive Models

Model	R²	k	N	CI low	CI high	Width	N : k
ICAP + Demog	.22	12	163	.059	.273	.214	13.6
i-CAN + Demog	.35	19	163	.142	.382	.240	8.6
Funded Items + Demog	.60	19	163	.440	.630	.190	8.6
Stepwise (Funded Items)ᵃ	.87	4	162	n/a	n/a	n/a	40.5

Note. Intervals derived analytically via the non-central F distribution (Smithson, 2003; Steiger, 2004). The intervals are asymmetric: the upper CI bound usually sits closer to the point estimate than the lower bound. Two factors drive this. Sample R² tends to be biased upward, so the population value is on average smaller than the observed value. R² also cannot exceed 1.0, which compresses the upper end of the distribution. Demog = demographic covariates. N : k ratios below 15 do not meet the Harrell (2015) recommendation for stable prediction; only the stepwise model row clears this, and only because stepwise selection reduced the predictor count from 19 to 4 within the same N.

a Stepwise selection precludes a standard confidence interval: the procedure inflates R² by capitalising on chance, regression through the origin further inflates R² because the total sum of squares is not computed against the intercept ( Arnold et al., 2015, Table 9; one outlier removed).

Submission 1567

Submission to the Senate Community Affairs Legislation Committee

Inquiry into the National Disability Insurance Scheme Amendment (Securing the NDIS for Future Generations) Bill 2026

References Cited in This Appendix

Arnold, S. R. C., Riches, V. C., & Stancliffe, R. J. (2014). I-CAN: The classification and prediction of support needs. Journal of Applied Research in Intellectual Disabilities, 27(2), 97–111. https://doi.org/10.1111/jar.12055

Arnold, S. R. C., Riches, V. C., & Stancliffe, R. J. (2015). Does a measure of support needs predict funding need better than a measure of adaptive and maladaptive behavior? American Journal on Intellectual and Developmental Disabilities, 120(5), 375–394. https://doi.org/10.1352/1944-7558-120.5.375

Button, K. S., Ioannidis, J. P. A., Mokrysz, C., Nosek, B. A., Flint, J., Robinson, E. S. J., & Munafò, M. R.(2013). Power failure: Why small sample size undermines the reliability of neuroscience. Nature Reviews Neuroscience, 14(5), 365–376. https://doi.org/10.1038/nrn3475

Bruininks, R. H., Hill, B. K., Weatherman, R. F., & Woodcock, R. W. (1986). Inventory for client and agency planning (ICAP). DLM Teaching Resources.

Cicchetti, D. V. (1994). Guidelines, criteria, and rules of thumb for evaluating normed and standardized assessment instruments in psychology. Psychological Assessment, 6(4), 284–290. hhttps://doi.org/10.1037/1040-3590.6.4.284

Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Lawrence Erlbaum Associates.

Gelman, A., & Carlin, J. (2014). Beyond power calculations: Assessing Type S (sign) and Type M (magnitude) errors. Perspectives on Psychological Science, 9(6), 641–651. https://doi.org/10.1177/1745691614551642

Harrell, F. E., Jr. (2015). Regression modeling strategies: With applications to linear models, logistic and ordinal regression, and survival analysis (2nd ed.). Springer. https://doi.org/10.1007/978-3-319-19425-7

Ioannidis, J. P. A. (2005). Why most published research findings are false. PLoS Medicine, 2(8), e124. hhttps://doi.org/10.1371/journal.pmed.0020124

Maxwell, S. E., Kelley, K., & Rausch, J. R. (2008). Sample size planning for statistical power and accuracy in parameter estimation. Annual Review of Psychology, 59(1), 537–563. https://doi.org/10.1146/annurev.psych.59.103006.093735

McHugh, M. L. (2012). Interrater reliability: The kappa statistic. Biochemia Medica, 22(3), 276–282. https://doi.org/10.11613/BM.2012.031

Riches, V. C., Parmenter, T. R., Llewellyn, G., Hindmarsh, G., & Chan, J. (2009b). The reliability, validity and practical utility of measuring supports using the I-CAN instrument: Part II. Journal of Applied Research in Intellectual Disabilities, 22(4), 340–353. https://doi.org/10.1111/j.1468-3148.2008.00467.x

Smithson, M. (2003). Confidence intervals. Sage Publications.

Steiger, J. H. (2004). Beyond the F test: Effect size confidence intervals and tests of close fit in the analysis of variance and contrast analysis. Psychological Methods, 9(2), 164–182. hhttps://doi.org/10.1037/1082-989X.9.2.164

Suen, H. K., & Ary, D. (1989). Analyzing quantitative behavioral observation data. Lawrence Erlbaum Associates.