Introduction
Difficulties in performing cognitively complex everyday activities, so-called “instrumental activities of daily living” (IADL), are a diagnostic criterion for dementia. One way to measure IADL functioning is by means of the Amsterdam Instrumental Activities of Daily Living Questionnaire (A-IADL-Q), an extensively validated study partner-reported instrument (see e.g. [Dubbelman et al., Reference Dubbelman, Verrijp, Facal, Sánchez‐Benavides, Brown, der Flier, Jokinen, Lee, Leroi, Lojo‐Seoane, Milošević, Molinuevo, Pereiro Rozas, Ritchie, Salloway, Stringer, Zygouris, Dubois, Epelbaum, Scheltens and Sikkes2020; Jutten et al., Reference Jutten, Peeters, Leijdesdorff, Visser, Maier, Terwee, Scheltens and Sikkes2017; Koster et al., Reference Koster, Knol, Uitdehaag, Scheltens and Sikkes2015; Sikkes et al., Reference Sikkes, Knol, Pijnenburg, de Lange-de Klerk, Uitdehaag and Scheltens2013; Sikkes et al., Reference Sikkes, Pijnenburg, Knol, de Lange-de Klerk, Scheltens and Uitdehaag2013]). To interpret A-IADL-Q scores on an individual level and to successfully implement IADL in clinical practice, normative data are imperative.
Norm scores typically reflect the standardized difference between an individual’s observed and expected outcome measure, as based on a cognitively healthy reference population. Hence they are imperative for results interpretation, which is particularly useful in clinical practice. Nevertheless, researchers can also benefit from the availability of norm scores, as they allow additional research opportunities, like splitting clinical cohorts based on individual deviations from normal functioning, which encourages personalized treatments (Marquand et al., Reference Marquand, Rezek, Buitelaar and Beckmann2016). Clinical use of the A-IADL-Q can be further facilitated by a norm-based diagnostic cutoff value, which aids in the diagnostic process and subsequent communication of results.
Traditional norming has the important limitation that it treats all demographic variables as discrete values (Voncken et al., Reference Voncken, Albers and Timmerman2019). As a result, this approach has been widely replaced by continuous norming, including standard polynomial regression and generalized additive models for location, scale, and shape (GAMLSS). The first allows for varying means, while the latter additionally allows for varying scale (σ, approximate coefficient of variation), skewness (ν, transformation to symmetry), and kurtosis (τ, power exponential parameter) (Voncken et al., Reference Voncken, Albers and Timmerman2019). Due to this flexibility, and because it is efficient, regression-based norming using GAMLSS has become a highly popular approach to estimate norm scores for psychological tests (Timmerman et al., Reference Timmerman, Voncken and Albers2021).
Here, we aimed to establish demographically adjusted norm scores for the A-IADL-Q using GAMLSS. Several combinations of age, sex and/or education were fitted to find the best normative model. Additionally, we aimed to provide the optimal cutoff value to distinguish cognitively normal (CN) from dementia, along with its diagnostic accuracies for other diagnostic contrasts, including subjective cognitive decline (SCD) and mild cognitive impairment (MCI) diagnoses.
Methods
Ethical considerations and consent
This study made use of European data from multiple studies, all of which were approved by ethical review boards local to each study site. All participants provided written informed consent to store, share, and use their data for research purposes. The research was completed in accordance with Helsinki Declaration.
Participant selection criteria
See Supplementary Methods for information. Figure S1 shows a flowchart of our exclusion criteria.
Sample characteristics
Cognitively normal sample
Cross-sectional data from Dutch Brain Research Registry (Zwan et al., Reference Zwan, van der Flier, Cleutjens, Schouten, Vermunt, Jutten, van Maurik, Sikkes, Flenniken, Howell, Weiner, Scheltens and Prins2021), European Medical Information Framework – Alzheimer’s Disease (EMIF-AD) 90 + study (Legdeur et al., Reference Legdeur, Badissi, Carter, de Crom, van de Kreeke, Vreeswijk, Trappenburg, Oudega, Koek, van Campen, Keijsers, Amadi, Hinz, Gordon, Novak, Podhorna, Serné, Verbraak, Yaqub, Hillebrand, Griffa, Pendleton, Kramer, Teunissen, Lammertsma, Barkhof, van Berckel, Scheltens, Muller, Maier, Herholz and Visser2018), and European Prevention of Alzheimer’s Dementia Longitudinal Cohort Study (EPAD-LCS) (Ritchie et al., Reference Ritchie, Molinuevo, Truyen, Satlin, Van der Geyten and Lovestone2016) were combined into one large, European normative sample of 1,374 participants with a wide age range (mean age = 63.3 ± 11.9years [min = 18, max = 97years], 69.2% female, 69.8% highly educated). See Supplementary Methods for details per dataset.
Memory clinic sample
In order to compute the optimal cutoff value to distinguish normal functioning from dementia, the CN sample was combined with a memory clinic sample from the Amsterdam Dementia Cohort (ADC) (van der Flier & Scheltens, Reference van der Flier and Scheltens2018), comprising 2,155 individuals recruited from the Alzheimer Center Amsterdam between 2013 and 2022 (mean age = 64 ± 8years [min = 32, max = 86 years], 44.4% female, 42.0% highly educated).
Measures
A-IADL-Q
The A-IADL-Q is a well-validated, study partner-reported questionnaire, measuring impairments in everyday functioning in the context of dementia (Dubbelman et al., Reference Dubbelman, Verrijp, Facal, Sánchez‐Benavides, Brown, der Flier, Jokinen, Lee, Leroi, Lojo‐Seoane, Milošević, Molinuevo, Pereiro Rozas, Ritchie, Salloway, Stringer, Zygouris, Dubois, Epelbaum, Scheltens and Sikkes2020; Jutten et al., Reference Jutten, Peeters, Leijdesdorff, Visser, Maier, Terwee, Scheltens and Sikkes2017; Koster et al., Reference Koster, Knol, Uitdehaag, Scheltens and Sikkes2015; Sikkes et al., Reference Sikkes, Knol, Pijnenburg, de Lange-de Klerk, Uitdehaag and Scheltens2013; Sikkes et al., Reference Sikkes, Pijnenburg, Knol, de Lange-de Klerk, Scheltens and Uitdehaag2013). The shortened version counts 30 items, each of which is scored according to a 5-point Likert scale (ranging from no difficulty to unable to perform a certain task). Total scores were calculated using item response theory (IRT), resulting in T-scores that are normally distributed in a memory clinic population around a mean of 50 with a standard deviation of 10, where higher scores indicate better IADL functioning (Jutten et al., Reference Jutten, Peeters, Leijdesdorff, Visser, Maier, Terwee, Scheltens and Sikkes2017). For the purpose of fitting a GAMLSS with a beta distribution, the T-scores were divided by 100, such that the values fell within the beta interval [0,1].
Other measures
Education of participants was either measured in years or classified according to the Dutch ‘Verhage’ system, and dichotomized to low/medium (i.e., < 13 years of education, or Verhage scales 1–5; up to high school education) and high education (i.e., ≥13 years of education, or Verhage scales 6-7; college level degree).
Statistical analyses
All analyses were performed in R version 4.2.1. The R packages gamlss (Rigby & Stasinopoulos, Reference Rigby and Stasinopoulos2005) and cutpointr (Thiele & Hirschfeld, Reference Thiele and Hirschfeld2021) were used for fitting GAMLSS and establishing the optimal diagnostic cutoff value, respectively.
The normative model
To determine the appropriate distribution family for our normative sample, a histogram plot of the A-IADL-Q T-scores was visually inspected. The beta distribution was found appropriate (see results), and accordingly used for fitting GAMLSS with various combinations of age (years), sex (0 = male, 1 = female) and educational attainment (0 = low, 1 = medium/high), including interaction effects and a cubic smoothing spline function for age. Stepwise model selection using a Generalized Akaike Information Criterion (GAIC) was used to determine the optimal model, i.e. the model with the lowest GAIC value. A Worm plot (i.e., de-trended Q-Q plot) of the optimal model was visually inspected to determine model fit, and term plots were created to display the cubic smoothing spline functions that were fitted.
Generalizability of the normative model was tested using k-folds cross validation, and norm scores (i.e., Z-scores, or normalized quantile residuals) were obtained for each individual in the joint study sample (see Supplementary Methods for details).
The optimal diagnostic cutoff value and corresponding diagnostic accuracies
The optimal cutoff value to distinguish CN individuals from those with dementia was derived from a bootstrap average estimate on the Youden Index, i.e., a metric of accuracy, whereby the highest value represents the optimal balance between sensitivity and specificity. Bootstraps were run with 500 repeats, to optimize this metric. The diagnostic accuracy was determined by means of Area Under Received Operating Curves (AUC-ROC).
The diagnostic accuracy of the optimal cutoff value for (1) CN vs dementia (CNvsDem) was also calculated for other diagnostic groups: (2) CN vs MCI (CNvsMCI), (3) SCD vs MCI (SCDvsMCI), (4) SCD vs dementia (SCDvsDem), and (5) MCI vs dementia (MCIvsDem).
Results
Sample characteristics
Our CN sample comprised 1,374 individuals, and the memory clinic sample comprised 1,725 individuals. Table 1 displays descriptive information about the two study samples, as well as the total study sample. Individuals from the memory clinic sample were older (t(2179) = −2.47, p = 0.01), lower educated (χ 2(1) = 247.35, p < 0.001), more often male (χ 2(1) = 227.85, p < 0.001) and performed significantly worse on the A-IADL-Q (t(3271) = 59.39, p < 0.001) compared to the CN sample. Table S1 shows the participant count per cohort across different age categories.
The normative model
Visual inspection of the histogram plot of the A-IADL-Q T-scores in the normative sample showed a left skew of the data (Figure S2), resembling a beta distribution with alpha = 2 and beta = 8 parameters. Hence, the beta distribution was used for fitting GAMLSS.
We found that a GAMLSS with an interaction effect between education and a cubic spline for age was the best model for our normative sample (GAIC = −4946.624, log(σ) = −2.38) (Table S2). Figure S3 shows the relationships between IADL functioning, age, and educational attainment. Our results show that A-IADL-Q performance generally decreases with age (albeit in a cubic-fashion). The interaction effect with educational attainment indicates that the effect of age on A-IADL-Q is dependent on it. Indeed, older participants with low educational attainment had lower A-IADL-Q scores compared to older participants with medium/high educational attainment. Furthermore, educational attainment is positively associated with A-IADL-Q, meaning that participants with medium/high educational attainment perform better on the A-IADL-Q than those with low educational attainment (Table S2, Figure S3).
The smooth functions used for fitting the cubic spline for age are displayed in the term plots (Figure S4). Figure S5 shows the Worm plot of the residuals, which indicates that the beta distribution does not adequately fit the data.
Cross-validation
The generalizability of the normative model was tested by comparing its global deviance (i.e., a goodness of fit measure, equal to −2*LogLikelihood) with the global deviance after k = 10-fold cross-validation. The difference was only 0.48% (global deviance = −4960.623, cross-validated global deviance = −4936.66), indicating that our normative model is highly generalizable to the general population.
The optimal cutoff value for CN versus dementia
For the demographically adjusted A-IADL-Q norm scores, we estimated the optimal cutoff value to distinguish CN from dementia, based on the highest Youden index. We found this cutoff to be -1.85 SD from the population mean (AUC = 0.97; 95%CI [0.96, 0.98]; sensitivity = 90.43%; specificity = 93.89%) (Table S3, Figure S7).
Figure S6 demonstrates the density distribution of the demographically adjusted norm scores in the CN reference population, as well as in the SCD, MCI and dementia groups. The high accuracy of the optimal cutoff to distinguish between CN and dementia is supported by the fact that 94.25% of CN individuals had demographically adjusted norm scores above this cutoff (indicating good performance), compared to only 7.59% of individuals with dementia. The percentage of MCI individuals scoring above the optimal cutoff was 30.55%, and for SCD this was 50.70%.
Diagnostic accuracies corresponding to the optimal cutoff value
We also studied the diagnostic accuracies of the CNvsDem cutoff (which is −1.85SD from population mean, see above) for other diagnostic contrasts (Table S3). This cutoff appeared just as sensitive in distinguishing CN from dementia as in distinguishing SCD and MCI from dementia (sensitivity = 90.43% for all three diagnostic contrasts), although with lower specificity: 49.16% for SCDvsDem, and 31.03% for MCIvsDem, compared to 93.81% for CNvsDem. Moreover, using the −1.85SD cutoff to distinguish CN and SCD individuals from those with MCI resulted in a lower sensitivity (68.97% for both), as well as lower specificity for the SCDvsMCI contrast (specificity = 49.16%). For CNvsMCI, however, the specificity remained the same (specificity = 93.89%). Figure S7 shows the corresponding ROC curves.
Normative data
Figure 1 illustrates our normative model. Individual deviations from the norm are reflected by different percentile categories (indicated with different colors), which can be deduced from reading the figure. Moreover, the cutoff for low vs medium/high IADL functioning (indicated with different shapes) can be read.
Discussion
The current study aimed at establishing demographically adjusted norm scores for the A-IADL-Q and providing a corresponding cutoff for dementia, thereby easing results interpretation and accordingly clinical use of the A-IADL-Q.
To our knowledge, this is the first study to establish norm scores for an IADL instrument. For other IADL measures, like the Functional Activities Questionnaire (FAQ) (Pfeffer et al., Reference Pfeffer, Kurosaki, Harrah, Chance and Filos1982), Everyday Cognition (ECog) (Farias et al., Reference Farias, Mungas, Reed, Cahn-Weiner, Jagust, Baynes and Decarli2008), and the Alzheimer’s Disease Cooperative Study ADL scale (ADCS-ADL) (Fish, Reference Fish, Kreutzer, DeLuca and Caplan2011), normative values, as well as a norm-based diagnostic cutoffs, have not yet been computed.
While cutoffs for different diagnostic contrasts can be of interest, we only present a single cutoff value here, to avoid possible confusion due to multiple cutoff values. Moreover, our cutoff for dementia showed sufficient diagnostic accuracies for the other diagnostic contrasts examined (all AUC > 0.6), such that one single cutoff suffices.
Notably, a prior A-IADL-Q study has calculated the optimal cutoff for dementia based on the raw A-IADL-Q T-scores (Sikkes et al., Reference Sikkes, Pijnenburg, Knol, de Lange-de Klerk, Scheltens and Uitdehaag2013). This value ( = 51.4), however, had 16% lower sensitivity and 30% lower specificity compared to our norm-based cutoff, and did not take into account age, sex or educational attainment. Clinicians are therefore recommended to use the norm-based cutoff we present here.
We used GAMLSS – a regression-based parametric norming method – for the computation of norm scores, which has two benefits over other types of continuous norming (i.e., semi-parametric and inferential norming): (1) it has readily available statistical criteria for model selection and assessment and (2) it does not require discretization of continuous norm predictors, which possibly introduces imprecision (Timmerman et al., Reference Timmerman, Voncken and Albers2021). A known downside of regression-based norming, however, is that the shape of the data must be modeled with any of the available distribution functions, which might not optimally model the shape of the data. This is particularly problematic in the case of large floor or ceiling effects, which cannot be captured well by parametric distributions (Lenhard et al., Reference Lenhard, Lenhard, Gary and Christensen2019). A floor or ceiling effect exists when, respectively, less or more than 20% of the participants have the maximum score on an instrument (Garin, Reference Garin and Michalos2014). While our data showed a left skew, only 17% of individuals had a T-score >= 69.5, supporting the absence of a ceiling effect, and advocating for regression-based norming.
The best normative model comprised a cubic spline for age and its interaction with educational attainment. Sex appeared not to be a determinant of IADL functioning, in line with a prior study on differential item functioning of the A-IADL-Q across eight countries (Dubbelman et al., Reference Dubbelman, Verrijp, Facal, Sánchez‐Benavides, Brown, der Flier, Jokinen, Lee, Leroi, Lojo‐Seoane, Milošević, Molinuevo, Pereiro Rozas, Ritchie, Salloway, Stringer, Zygouris, Dubois, Epelbaum, Scheltens and Sikkes2020). Nevertheless, sex differences in IADL have been reported in Europe (Scheel-Hincke et al., Reference Scheel-Hincke, Möller, Lindahl-Jacobsen, Jeune and Ahrenfeldt2020). Those differences were significantly larger in Southern compared to Northern Europe (Scheel-Hincke et al., Reference Scheel-Hincke, Möller, Lindahl-Jacobsen, Jeune and Ahrenfeldt2020), which may suggest that our normative sample more closely represents Northern Europe. Indeed, most of our normative sample is from the Netherlands, so some caution is warranted when the model is applied to non-Northern Europeans.
To make our norm scores applicable to a broader, European population, we have included European data from the EPAD-LCS cohort, which is an important strength of our study, because it increases generalizability of our results. Furthermore, we included a cohort of cognitively healthy nonagenarians, which further adds to the generalizability of results, by increasing applicability of our normative sample to older individuals. Other strengths of this study are the broad age range of our normative sample (18y–97y) and its large sample size(n > 1,000).
One limitation of our study is that the Worm plot indicated bad model fit. Be that as it may, these normative values are computed using the highly flexible GAMLSS framework, which is in many ways superior to the traditional(mean-based) standard polynomial regression approach. Thus, even though our data did not fit the beta distribution well, it is the best fit we could have obtained without subjecting our data to complex transformations (e.g., a square- or cube root), which come with their own limitations. Another limitation is that our normative data are not directly accessible. Hence, we are currently exploring the implementation of an A-IADL-Q norm score calculator in existing tools for research and clinical use, such as ADappt (van Maurik et al., Reference van Maurik, Visser, Pel-Littel, van Buchem, Zwan, Kunneman, Pelkmans, Bouwman, Minkman, Schoonenboom, Scheltens, Smets and van der Flier2019). Finally, our normative data are limited to a European population. Future studies should therefore include participants from different continents, cultures, and socioeconomic backgrounds, in order to facilitate world-wide use of the A-IADL-Q norm data in the clinic.
Conclusion
To conclude, we present norm scores for the A-IADL-Q and provide an optimal cutoff for dementia. Our data are key to improving clinical assessment of IADL performance, as they allow for easier interpretation of scores.
Figure 1. Illustration of the normative model. This figure shows the A-IADL-Q T-score distribution in the normative sample across age, split by educational attainment. Left: < 13 years of education, or Dutch Verhage scales 1-5. Right: ≥13 years of education, or Dutch Verhage scales 6–7. The different colors indicate the different percentile categories individuals can fall into, with lower percentile scores (yellow) reflecting better IADL functioning, as compared to the normal sample. The different shapes indicate the expected level of IADL functioning (i.e., low [triangle], or medium/high [circle]), based on the optimal cutoff for dementia. Reading example: The top right shows a reading example for a 50-year individual with medium/high education and an A-IADL-Q T-score of 68. The place where the two solid lines cross each other indicates the place where this person falls within the normal sample. That is, within the third percentile category, meaning that 50%–75% of cognitively normal individuals score the same or better on the A-IADL-Q than this person. The cutoff for this person (triangle) suggests medium/high IADL functioning.
Table 1. Participant characteristics
CN = cognitively normal, SCD = subjective cognitive decline, MCI = mild cognitive impairment, AD = Alzheimer’s disease, EMIF-AD 90+ = European Medial Information Framework-Alzheimer’s Disease 90+ study, EPAD-LCS = European Prevention of Alzheimer’s Dementia Longitudinal Cohort Study, ADC = Amsterdam Dementia Cohort.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S1355617724000031.
Acknowledgments
The authors would like to thank all study participants, as well as everyone involved in data collection and data sharing.
Funding statement
The Amsterdam IADL Questionnaire is free for use in all public health and not-for-profit agencies and can be obtained via https://www.alzheimercentrum.nl/professionals/amsterdam-iadl. The development of the Amsterdam IADL Questionnaire is supported by grants from Stichting VUmc Fonds and Innovatiefonds Zorgverzekeraars.
This work was supported by public–private funding of which SAMS was the recipient. She received funding from Health-Holland, Topsector Life Sciences & Health (PPPallowance; LSHM20084-SGF, project DEFEAT-AD), and the National Institutes of Health, as well as license fees from Green Valley, VtV Therapeutics, Alzheon, Vivoryon, and Roche, and honoraria from Boehringer, Toyama and Biogen. All funding is paid to her institution. In addition, the work was supported by license fees from Brain Research Center Holding BV obtained from the for-profit commercial use licensing of the Amsterdam IADL Questionnaire to third-party clients.
Research of Alzheimer Center Amsterdam is part of the neurodegeneration research program of Amsterdam Neuroscience. Alzheimer Center Amsterdam is supported by Stichting Alzheimer Nederland and Stichting Steun Alzheimercentrum Amsterdam. The chair of Wiesje van der Flier is supported by the Pasman stichting. The clinical database structure was developed with funding from Stichting Dioraphte.
The Dutch Brain Research Registry (Hersenonderzoek.nl) is supported by ZonMw-Memorabel (project no 73305095003), Alzheimer Nederland, Amsterdam Neuroscience, and Hersenstichting (Dutch Brain Foundation).
The EMIF-AD 90+ Study was funded by the EU/EFPIA Innovative Medicines Initiative Joint Undertaking EMIF grant agreement no. 115372.
The European Prevention of Alzheimer’s Dementia project was funded by the Innovative Medicines Initiative grant agreement 115736.
Competing interests
MCP, MAD, JC, CR, MV, LV, PV, and MDZ report no disclosures relevant to the manuscript.
WMF has received further funding from NWO, EU-FP7, EU-JPND, Alzheimer Nederland, CardioVasculair Onderzoek Nederland, Stichting Dioraphte, Gieskes-Strijbis fund, Stichting Equilibrio, Pasman Stichting, Biogen MA Inc., Boehringer Ingelheim, Life-MI, AVID, Roche BV, Fujifilm, and Combinostics. W.M. van der Flier holds the Pasman chair, has performed contract research for Biogen MA Inc. and Boehringer Ingelheim, is a consultant to Oxford Health Policy Forum CIC, Roche and Biogen MA Inc., and has been an invited speaker at Boehringer Ingelheim, Biogen MA Inc., Danone, Eisai, WebMD Neurology (Medscape). All funding is paid to her institution.
SAMS co-developed the A-IADL-Q, which is freely available for use by academic users and not-for-profit organizations. License fees from for-profit organizations are paid to her institution. SAMS was supported by grants from JPND and Zon-MW and has provided consultancy services in the past 2 years for Nutricia and Takeda. All funds were paid to her institution.
Open access
