Do Mini–Mental State Examination and Montreal Cognitive Assessment predict high‐cost health care users? A competing risks analysis in The Irish Longitudinal Study on Ageing

Abstract Objectives Policymakers want to better identify in advance the 10% of people who account for approximately 75% of health care costs. We evaluated how well Mini–Mental State Examination (MMSE) and Montreal Cognitive Assessment (MoCA) predicted high costs in Ireland. Methods/Design We used five waves from The Irish Longitudinal Study on Ageing, a biennial population‐representative survey of people aged 50+ (2010–2018). We used competing risks analysis where our outcome of interest was “high costs” (top 10% at any wave) and the competing outcome was dying or loss to follow‐up without first having the high‐cost outcome. Our binary predictors of interest were a ‘low score’ (bottom 10% in the sample) in MMSE (≤25 pts) and MoCA (≤19 pts) at baseline, and we calculated sub‐hazard ratios after controlling for sociodemographic, clinical and functional factors. Results Of 5856 participants, 1427 (24%) had the ‘high cost’ outcome; 1463 (25%) had a competing outcome; and 2966 (51%) completed eight years of follow‐up without either outcome. In multivariable regressions a low MoCA score was associated with high costs (SHR: 1.38 (95% CI: 1.2–1.6) but a low MMSE score was not. Low MoCA score at baseline had a higher true positive rate (40%) than did low MMSE score (35%). The scores had similar association with exit from the study. Conclusions MoCA had superior predictive accuracy for high costs than MMSE but the two scores identify somewhat different types of high‐cost user. Combining the approaches may improve efforts to identify in advance high‐cost users.

CAPI and SCQ follow-up occurs at each Wave; health assessments were conducted again at Wave 3. In the event of a participant death, TILDA approaches a family member or close friend to complete a voluntary interview on end-of-life experience; full details including ethical procedures have been provided previously. 3 This study uses data up to and including Wave 5 (2018).

| Setting
Ireland is an island in north-western Europe comprising two jurisdictions: ROI, which is an independent country, and Northern Ireland, which is a part of the United Kingdom. The Irish Longitudinal Study on Ageing is conducted within the ROI (ROI) only. ROI has a population of approximately 4.9 million people, with a relatively young age distribution for a high-income country. 15 The healthcare system has mixed public and private provision. A means-tested medical card confers free primary care and hospital visits, and also caps co-payments for prescriptions. 16 Those without a medical card pay full primary care costs out of pocket as well as capped co-payments for hospital care and prescriptions. Over half of people aged 50+ have voluntary health insurance to access private hospital care in the context of lengthy waiting lists for planned care. 17 Compared to similar countries the ROI health system has unusually high acute hospital bed occupancy and relatively low primary and community care provision. 18

| Predictors
Our primary predictors of interest were MMSE and MoCA at Wave 1.
An MMSE cut-point of 24 points has been suggested previously to identify normal function but in practice a range of points are used. 19 We created a binary variable in MMSE where the participant had a value of 1 with 25 MMSE points or fewer. Previous research has found that this has good sensitivity and specificity. 20 The developers of MoCA advocated a cut point of 26 or lower to indicate mild cognitive impairment (MCI), 21 but this has been shown to have low specificity. 22,23 A range of alternative cut-offs on MoCA have been used depending on the cohort.
The two tests have differing purposes-MMSE was derived with respect to detecting dementia, MoCA to detect MCI-and as such the distributions of these scores differ within a given sample. 24 To ensure that observed differences in comparing MMSE and MoCA were not due to sample size, we defined a binary MoCA variable at the same percentile as we created the binary MMSE variable.
To identify additional predictors we drew on Andersen's model of predisposing, enabling, need characteristics and prior utilisation 16 ; and prior cost modelling using TILDA data. 3 The full list of predictors is provided in Table 1 25,26 and standardised to 2018, the final year of data collection, using the consumer price index. 27 If an end-of-life interview was conducted for a participant who died in a given wave, we calculated costs for that person in that wave from the end-of-life interview. We categorised the outcome as binary because we were interested specifically in predicting the high-cost class, and not in the association between coefficients and the full outcome distribution.
Our secondary outcome of interest was mortality. All registered deaths in the ROI are recorded with the General Register Office (GRO). The Irish Longitudinal Study on Ageing data between Wave 1 and Wave 5 are linked to the GRO data to March 2018, in a procedure detailed elsewhere. 28 In addition to this, TILDA may become aware of participant deaths after being notified by a family member or after the TILDA team approach for an interview. To account for this, we have a mortality file providing full coverage of participant death dates within Ireland during the time period (Wave 1 to Wave 5), via the GRO, and additional non-comprehensive coverage on deaths outside the State (from family members).

| Statistical methods
In descriptive analyses we examined the overall sample on baseline predictors, and we stratified the sample by MMSE and MoCA scores.
In presenting our outcome data we reported the distribution of costs in the sample at each wave, and we calculated the proportion of total sample costs that are accounted for by the top 10% of the distribution.
In our main analyses we used regression to analyse the association between our predictors, low MMSE and low MoCA scores, and our outcome, high costs at any wave. We conducted unadjusted bivariate regressions and multivariable regressions adjusting for factors in Table 1. At each wave, TILDA participants may die or drop out of the study. Since mortality and dropping out are not independent of our predictors, simply treating these outcomes as missing data increases the risk of bias. [29][30][31] We treated these outcomes as competing risks; that is, events that potentially prevent occurrence of the primary outcome of interest but which should not be treated as missing in analysis. 32 At wave 2, if a participant was in the top 10% of costs then they were deemed to have the outcome of interest, if the participant had died or did not participate they were deemed to have the competing risk, and if they had neither then they were retained to examine outcomes at wave 3, and so on. Thus we estimate associations between predictors and outcome, after taking account of any participants who had died or dropped out and could not achieve the outcome. As such our analyses adjust for participants' mortality rather than allocating the deceased zero costs or dropping them from the analysis. 33

| Bias
While TILDA was representative of the community-dwelling population aged 50+ at baseline, MMSE and MoCA data were collected only for the sub-sample who attended the health assessment. While all participants were invited to attend the health assessment, approximately 30% did not do so. This group is therefore missing from the analysis and not at random. In the Appendix we present summary statistics for our sample alongside the Wave 1 participants who did not complete MoCA and MMSE. Missingness on baseline

| Outcome data
The health care costs across five waves are presented in Table 2

| Main results
The regression results are presented in Table 3

| Key results
In our main analyses (Table 3)

| Limitations
All variables in Table 1 and Table 2 are self-reported, which increases the risk of biases related to response accuracy. Initial recruitment of community-dwelling people means that ADRD diagnosis and declining cognitive function are more unusual in TILDA than in the general population of older people. A corollary of this sampling strategy, and a strength of our analyses, is that our cost data do not reflect high residential care expenditures, which may not be substantively avoidable, but rather hospital admissions, which may be avoidable or shortened with appropriate supportive care. Within the TILDA sample, health assessment participants who comprise our analytic sample were younger, more healthy and more socioeconomically advantaged than those who did not do the health assessment and are therefore excluded (Appendix). Non-negligible measurement error has been noted previously on both cognitive scores. 34 The cut-points on MMSE and MoCA at the 10 th percentile were based on reason but ultimately pragmatic to ensure that any   Table 1. comparison was consistent on sample sizes; our results were substantively similar in sensitivity analyses with different cut-points to the cognitive scores.

| Interpretation
Prior literature on economics of cognition has noted that while the health care costs associated with ADRD are well established, much less is known about the costs of MCI. [35][36][37] We analysed a sample of people who at baseline were over 50 years old and living in the community with <1% prevalence of self-reported ADRD diagnosis (Table 1). An estimated 10%-20% of people aged 65+ will develop MCI, which is a significant risk factor for ADRD. 38 The MoCA and MMSE indices were developed to identify differing populations: respectively, those living with MCI, who have lower costs than those with ADRD but are a much larger group 39  There's estimated to be more than 64,000 people are living in Ireland with dementia currently, 40 approximately 1.3% of the population. Therefore by definition cognitive function scores will never be a powerful tool for prospectively identifying large proportions of the 10% of high-cost users. However, as efforts to address the "denominator challenge" become more sophisticated, in recognition that high-cost users are drawn from multiple potentially latent subgroups, people with lower cognitive function and at risk of cognitive decline will form one such important sub-group. Moreover, as Ireland's young population ages, this sub-group will increase in both total size and in mean per-capita costs.
Our results highlight the scope for well-known cognitive tests,

| CONCLUSION
Longitudinal follow-up over eight years in a sample of people aged 50 + in Ireland found that MoCA was a more reliable predictor than MMSE of people subsequently becoming high-cost health care users.
Montreal Cognitive Assessment also had superior predictive accuracy to MMSE but the two scores identify somewhat different types of high-cost user and so combining the approaches may improve predictive accuracy.