Gerontology 2004;50:39–43

DOI: 10.1159/000074388

MMSE Cross-Domain Variability

Predicts Cognitive Decline in
Centenarians
Matthias Kliegel a Martin Sliwinski b
a Institute of Psychology, Department of Gerontopsychology, University of Zurich, Zurich, Switzerland;
b Department of Psychology, Syracuse University, Syracuse, N.Y., USA

Key Words The development of methods for the early detection of

Oldest old W Mini-Mental State Examination W dementia has been a central aim of applied gerontological
Cross-domain variability W Cognitive decline W Dementia research for many years. The need for accurate methods
of early detection has become even more pressing with the
aging of the population and the availability of effective
Abstract treatments for certain types of dementias that may be
Background: Early detection of dementia is one of the most efficacious during early disease stages. However,
key issues in cognitive gerontology. However, so far the despite development of highly sensitive psychometric
detection of early stages in cognitive decline has been measures [2, 3], our ability to identify individuals in the
rather unreliable. One central limitation of current as- very early stages of dementia (i.e. years before they reach
sessment strategies is that they rely on information clinical threshold) has been hampered by inherent limita-
about a person’s level of performance obtained from a tions of current assessment strategies that rely on infor-
single assessment. Objective: In the first part of the mation about a person’s level of performance obtained
present paper, we propose three strategies for overcom- from a single assessment. A problem with this approach is
ing this limitation by using information from several the difficulty in distinguishing between poor performance
measurement occasions to improve diagnostic reliabili- due to low premorbid intellectual ability, preclinical de-
ty. Methods: In the second part, we present one exem- mentia, or state-based fluctuations in cognition [4]. Re-
plary empirical analysis of one of these approaches cent research has demonstrated that only half the individ-
reporting data from the Heidelberg Centenarian Study. uals classified as high-risk for dementia are consistently
Results: In this sample of very old individuals tested at classified on each of several repeated assessments [5]. The
baseline and at a follow-up 18 months later, the intra- purpose of this paper is to consider alternative assessment
individual cross-domain variability across MMSE sub- strategies that rely on the measurement of variability in
scales [1] at baseline predicted cognitive decline (i.e. cognitive performance.
change in an external cognitive functioning rating from The reliability of a measurement of cognitive perfor-
baseline to follow-up) much better than baseline mean mance reflects the joint characteristics of the specific test
MMSE scores. Conclusions: In conclusion, we discuss and the specific individual being tested. For example, a
the advantage of performance varibility as a predictor for test of free recall might be very reliable if administered to
cognitive decline. a sample of cognitively intact older adults, but far less reli-
Copyright © 2004 S. Karger AG, Basel able in a sample of neurologically impaired elders because

© 2004 S. Karger AG, Basel Matthias Kliegel

ABC 0304–324X/04/0501–0039$21.00/0 Institute of Psychology, Department of Gerontopsychology
Fax + 41 61 306 12 34 University of Zurich, Schaffhauserstrasse 15
198.143.47.65 - 8/13/2015 4:46:38 PM

E-Mail [email protected] Accessible online at: CH–8006 Zürich (Switzerland)

www.karger.com www.karger.com/ger Tel. +41 1 635 3431, Fax +41 1 635 3421, E-Mail [email protected]
Emory University
Downloaded by:
the latter’s behavior may be intrinsically more variable. measures will increase the reliability by reducing the error
Recent research has demonstrated increased cognitive variance by a factor of N, where N is the number of mea-
variability in subgroups of individuals with degraded neu- surements. This holds in much the same way that the SD
rological function [6, 7]. Particularly relevant is the study of the sampling distribution of the mean (i.e., standard
by Hultsch et al. [6] that demonstrated that individuals error of the mean) is smaller than the standard deviation
with mild dementia exhibited twice as much trial-to-trial of single scores in the population by a factor of [(SD single
variability on simple reaction time tasks than non- score)/√ N]. As an example, a common strategy for identi-
demented healthy and arthritic controls. fying preclinical dementia is to classify individuals scoring
Efforts to detect individuals in the very early stages of more than 1.5 SDs below the mean of age-matched con-
dementia are thus complicated by the likelihood that trols as high risk (i.e. MCI [8]). Long-term longitudinal
behavior is most variable (and hence, cognitive assess- studies of incident dementia [12] suggests that 5 years
ments are most unreliable) in precisely those individuals prior to diagnosis, preclinical individuals score on average
at greatest risk for dementia. As a consequence, findings 1.5 SDs below the mean of controls. Assuming equal stan-
that demonstrate that many individual are not consistent- dard deviations in preclinical and nondementia groups,
ly classified as Mild Cognitive Impairment (MCI [8]) this yields a sensitivity of 50% and a specificity of 93%. If,
across multiple testing sessions [5, 9] may reflect two for example, the average of 6 scores were used instead of a
sources of variability: classification error and amplified single score and the same –1.5 SD cut score were applied,
intraindividual variability. Hence, variable cognitive per- the sensitivity would increase to approximately 98% by
formance impairments are likely to precede consistent reducing the effect of intraindividual variability.
performance impairments in individuals in the very early A third approach accounts for cross-domain variability
stages of dementia. We now consider three possible ap- across several cognitive measures or cognitive sub-do-
proaches to addressing these difficulties. mains at one measurement occasion. While the first two
The first is to obtain bursts [10] of repeated measure- approaches are mainly methodologically driven, the
ments in order to estimate performance variability and cross-domain approach directly utilizes the theoretical
then to use this estimate in conjunction with information assumption that a decrease in neurological integrity not
about performance level to augment predictive validity. In only leads to increasing variability of performance in one
the first step of this approach, one would estimate the specific domain but also to increasing variability across
magnitude of variability for each individual (i.e. obtain- several (cognitive) sub-domains. Recent evidence for this
ing a measure of within-person variability like an intrain- proposal comes from findings reporting greater inconsis-
dividual standard deviation (SD)). In the second step, one tency in cognitive and non-cognitive domains in groups
would then add this estimate to information about the characterized by central nervous system dysfunction or
level of performance for each person in a model that pre- poorer cognitive test performance, respectively [13]. In
dicts (early stages of) dementia. This approach is support- addition, sporadic evidence has been presented that drop-
ed by recent work from the Victoria Longitudinal Aging ping performance in specific subscales of the Mini-Mental
study that demonstrates mildly demented individuals State Examination (MMSE [1]) rather than in the overall
show increased variability compared to nondemented sum score can predict cognitive changes in preclinical
and arthritic controls even after statistically controlling Alzheimer’s disease [14]. Taken together, it seems plausi-
for mean differences between groups [6]. This result ble to assume that the variability of performance across
implies that using an index of performance variability several cognitive sub-domains is a better predictor of
would have incremental validity above and beyond a sin- beginning cognitive decline than consistently high, me-
gle index of performance level in discriminating mildly dium, or low mean level performance as it is most fre-
demented individuals from their nondemented peers. quently used in (clinical) cognitive assessment.
The second approach is to combine information across While all three approaches have to be tested (more)
multiple assessments to maximize the reliability of perfor- extensively for their ‘added value’ in early diagnostic cog-
mance level estimates for individuals. In this approach, nitive aging research, in the following part of the paper we
which could be easily implemented in clinical assessment will present an example of the third of these analytic
procedures, one would measure people more than once approaches to demonstrate that it can optimize the detec-
and then take the average of the multiple measures. While tion of cognitive decline. We, therefore, analyze a sample
typically in clinical memory assessments only one mea- of very old adults utilizing the occurrence of very rapid
surement of each indicator is obtained [8, 9, 11], multiple cognitive changes in this population.

40 Gerontology 2004;50:39–43 Kliegel/Sliwinski

198.143.47.65 - 8/13/2015 4:46:38 PM
Emory University
Downloaded by:
 Method able cognitive stages, ranging from no cognitive impairment to very
severe cognitive impairment. Content validity of the GDS was dem-
Participants onstrated by Overall, Scott, Rhodes, and Lesser [19], and inter-rater
The Heidelberg Centenarian Study [15] is a population-based and retest reliability has been demonstrated regularly at over 0.9
study. All inhabitants of a defined geographical region approximately [20].
60 km around Heidelberg, Germany (comprising 172 communities
and 2.6 million inhabitants), who were aged 100, were potential par-
ticipants. According to this dynamic cohort design [16], there were Results
no other inclusion or exclusion criteria. For recruitment, resident
registration offices were asked to provide information on inhabitants
born in 1901 or earlier. Overall, 156 centenarians or a close person In the following analyses, we will utilize the shortened
(proxy) were contacted. 91 of them took part in the study. The main MMSE-scale(s) as predictor variable(s) assessing a range
reasons for refusal were health-related problems (50%) or no interest of four cognitive abilities: orientation, immediate recall,
in participation (22%). Except in one case, the interviewer visited number, and delayed recall. In addition, the Global Dete-
each centenarian personally in his or her domestic environment. One
rioration Scale (GDS) serves as the external criterion
(confused and aggressive) woman, although encouraged by a relative,
could not be persuaded to meet the interviewer. Thus, at baseline, reflecting overall cognitive status and its consequences in
participation rate in this study was 58.3%, indicating that the sample daily life.
represented 58.3% of the entire defined population. The average age
was mean = 100.2 years (SD = 0.41). Detailed analyses of variables Descriptives
describing the sample at baseline are summarized in Rott et al. [15].
MMSE: At baseline, the mean MMSE score was 10.01
In sum, participants were mainly female, had mostly graduated from
primary and secondary school, were for the most part widowed and (SD: 7.32, range: 0–21). Distinguishing the four sub-
in half the cases lived in residential care. Every second participant’s domains, the mean score for orientation was 5.03 (SD:
health state was reported to be good. Compared to the entire defined 3.85, range: 0–10), for immediate recall 1.88 (SD: 1.31,
population, a positive selection seems only to be likely with regard to range: 0–3), for number 2.42 (SD: 2.23, range: 0–5), and
the participants’ health state [15]. After 18 months, we conducted a
for delayed recall 0.68 (SD: 1.05, range: 0–3).
follow-up on the sample. By that time, 49 of the 91 participants had
deceased, 6 of the remaining 42 refused further participation. This GDS: At baseline, the mean GDS score was 3.97 (SD:
left 36 centenarians in the follow-up sample, all of whom were exam- 2.09, range: 1–7). 22% of the participants were rated as
ined on the same measures that had been used at baseline. showing no cognitive impairment, 7% very minor, 12%
minor, 10% moderate, 21% substantial, 16% severe, and
Instruments and Procedure
12% very severe cognitive impairment. At follow-up,
The design and methods of the Heidelberg Centenarian Study
have been thoroughly detailed elsewhere [15]. Here, we will merely these values changed to M = 4.4 (SD: 2.21, range: 1–7).
report the instruments that were relevant to the present purpose. 19% of the remaining participants were now rated as
Cognitive status, i.e., the amount of cognitive impairment, was showing no cognitive impairment, 6% very minor, 11%
examined with two tests. The first one was the MMSE [1]. We minor, 6% moderate, 6% substantial, 38% severe, and
applied a shortened version with which we intended to avoid disad-
14% very severe cognitive impairment. Analyzing intra-
vantages due to sensorimotor impairments when examining people
aged 100 years or above. According to initial systematic reports on individual change (GDS T2 – GDS T1), 55% of the par-
the problems that can arise in psychodiagnostics in very old people, it ticipants did not change in their rating, 6% of the centen-
is conceivable that contradictory data on very old adults’ cognitive arians improved their rating by 1 stage, 19% declined by 1
status may in part be explained by how cognitive impairment is stage, 14% of the participants by 2 stages, 3% by 3 and 3%
assessed. For example, Holtsberg et al. [17] demonstrated that on
of the participants declined by 6 stages.
MMSE-items requiring reading and writing skills, sensorimotor im-
paired centenarians were systematically disadvantaged. Hence, we
attempted to minimize biases by using only items that had found to Prediction of Cognitive Decline
be empirically unproblematic regarding such impairments. We short- The core question of these analyses was: Can we pre-
ened the original test version, leaving out items requiring reading or dict the relatively short-term changes in oldest old adults’
writing skills and thus reducing the maximum score to 21. The items
cognitive functioning as determined by changes in GDS
we used were: orientation (10 points max.), immediate free recall of
three items (3 points max.), counting backwards (5 points max.), and ratings using information on very rudimentary cognitive
delayed free recall of three items (3 points max.). This procedure measures obtained at baseline? In a first step, we analyzed
apparently has no effect on test reliability or validity [17]. the correlations between the interindividual variability in
As a second, broader measure of overall cognitive functioning MMSE sumscore and the four distinct MMSE subdo-
and its consequences to everyday functioning, we included the Glob-
mains and the intraindividual change scores in GDS rat-
al Deterioration Scale (GDS) [18], a rating scale to be filled out by a
trained interviewer. The GDS is a 7-step rating scale made up of ings. However, there were no significant correlations. In a
detailed clinical descriptions of seven major clinically distinguish- second step we computed a score of intraindividual cross-

Cross-Domain Variability Gerontology 2004;50:39–43 41

198.143.47.65 - 8/13/2015 4:46:38 PM
Emory University
Downloaded by:
Table 1. Hierarchical regression predicting cognitive change changes in oldest old’ cognitive functioning using infor-
mation on very rudimentary basic cognitive measures
Predictors Cognitive change (¢GDS)
obtained at baseline?’, the answer was ‘yes’, but only when
ß R2 ¢ R2 considering the intraindividual variability in perfor-
mance across the four sub-domains of the assessed MMSE
Step 1: mean level performance 0.05 0.05
version. T1 mean level performance in both the sum score
MMSE Sum Score T1 0.12
as well as the distinct subscales did not correlate with the
Step 2: variability 0.20* 0.15* observed changes in overall cognitive functioning.
MMSE intraindividual
cross-domain variability T1 0.39*
There are a number of possible interpretations for this
finding. First, performance variability across multiple do-
* p ! 0.05. ß values are taken from the model that includes both mains may be a signature of cognitive impairment in old-
predictors. er adults and hence provide a more sensitive indicator of
impairment than performance level in any single domain
or all the domains combined. Second, the variability score
we used may not reflect variability per se but rather differ-
ential impairment in some or a single domain, and rela-
domain variability. For this purpose we z-standardized all tive intactness in the remaining domains. Although possi-
four sub-domain scales and then computed an intra-indi- ble, this interpretation seems unlikely, since the correla-
vidual standard deviation across all four subdomains. tions of the individual MMSE scores with decline were
This measure of intra-individual cross-domain variability small and nonsignificant. Third, on a more general level,
correlated significantly with change in cognitive perfor- the obtained measure of intraindividual variation refers
mance (r = 0.51; p ! 0.01). to the heterogeneity of cognitive performance within a
Finally, we computed a hierarchical regression analysis person and so contains information about the multivar-
predicting cognitive change in GDS ratings, including the iate status of the individual. Thus, considering this infor-
mean MMSE sum score in the first step and the intraindi- mation may be useful for both early detection of cognitive
vidual cross-domain variability measure in the second decline and more precise (age-related) assessment of cog-
step. This allows assessing the predictive power of the nitive performance. Although the results from the present
intra-individual cross-domain variability measure after study are limited with respect to the covered age range,
controlling for the influence of mean level performance. breadth of cognitive assessment, and size of sample, it
As summarized in table 1, while MMSE sum mean level seems worthwhile to pursue further studies that assess the
performance at baseline does not explain a significant utility of using intraindividual variability as a marker of
amount of variance in cognitive change (R2 = 0.05; p = cognitive decline as well as an indicator of the multivar-
0.2), the inclusion of intra-individual cross-domain vari- iate status of a person and investigating the oldest old as a
ability increases the amount of explained variance signifi- population of rapid changes.
cantly (R2 = 0.15; p ! 0.05). Considering both variables
simultaneously, only the intra-individual cross-domain
variability at baseline remains as a significant predictor of
cognitive change in the investigated population of the
oldest old.

Discussion

In the present paper we propose three possible ap-

proaches of how to utilize intraindividual variability as a
predictor for cognitive decline in old age. In addition, we
present one exemplary empirical analysis of one of these
approaches analyzing data from extremely old adults who
display rapid change in cognitive functioning. Following
the research question of ‘Can we predict short-term

42 Gerontology 2004;50:39–43 Kliegel/Sliwinski

198.143.47.65 - 8/13/2015 4:46:38 PM
Emory University
Downloaded by:
 References

1 Folstein MF, Folstein SE, McHugh PR: Mini- 8 Petersen RC, Smith GE, Waring SC, Ivnik RJ, 14 Small BJ, Fratiglioni L, Viitanen M, Winblad
Mental-State: A practical method for grading Tangalos EG, Kokmen E: Mild cognitive im- B, Bäckman L: The course of cognitive impair-
the cognitive state of patients for the clinician. pairment: Clinical characterization and out- ment in preclinical Alzheimer disease: Three-
J Psychiatr Res 1975;12:189–198. come. Arch Neurol 1999;56:303–308. and 6-year follow-up of a population-based
2 Buschke H, Sliwinski M, Kuslansky G, Lipton 9 Bischkopf J, Busse A, Angermeyer MC: Mild sample. Arch Neurol 2000;57:839–844.
R: Diagnosis of early dementia by the Double cognitive impairment: A review of prevalence, 15 Rott C, d’Heureuse V, Kliegel M, Schönemann
Memory Test: Encoding Specificity improves incidence and outcome according to current P, Becker G: The Heidelberg Centenarian
diagnostic sensitivity and specificity. Neurolo- approaches. Acta Psychiatr Scand 2002;106: Study: Theoretical and methodological founda-
gy 1997;48:989–997. 403–414. tions of psychosocial research in the oldest old.
3 Buschke H, Kuslansky G, Katz M, Stewart WF, 10 Nesselroade JR: The warp and woof of the Z Gerontol Geriatr 2001;34:356–364.
Sliwinski M, Eckholdt HM, Lipton RB: Screen- developmental fabric; in Downs R, Liben L, 16 Andersen-Ranberg K, Vasegaard L, Jeune B:
ing for dementia with the memory impairment Palermo DS (eds): Visions of Aesthetics, the Dementia is not inevitable: A population-
screen (MIS). Neurology 1999;50:231–238. Environment, and Development: The Legacy based study of Danish centenarians. J Gerontol
4 Darby D, Maruff P, Collie A, McStephen M: of Joachim F. Wohwill. Hillsdale, Erlbaum, Psychol Sci 2001;56B:152–159.
Mild cognitive impairment can be detected by 1991, pp 213–240. 17 Holtsberg PA, Poon LW, Noble CA, Martin P:
multiple assessments in a single day. Neurology 11 Burns A, Zaudig M: Mild cognitive impair- Mini-mental state exam status of community-
2002;59:1042–1046. ment in older adults. Lancet 2002;360:1963– dwelling cognitively intact centenarians. Int
5 Collie A, Maruff P, Currie J: Behavioral char- 1965. Psychogeriatr 1995;7:417–427.
acterization of mild cognitive impairment. J 12 Sliwinski M, Hofer S, Hall CB, Buschke H, Lip- 18 Reisberg B, Ferris SH, deLeon MJ, Crook T:
Clin Exp Neuropsychol 2002;24:720–733. ton R: Modeling memory decline in older The Global Deterioration Scale (GDS) for as-
6 Hultsch DF, MacDonald SWS, Hunter MA, adults: The importance of preclinical demen- sessment of primary degenerative dementia.
Levy-Bencheton J, Strauss E: Intraindividual tia, attrition and chronological age. Psych Am J Psychiatry 1982;139:1136–1139.
variability in cognitive performance in older Aging, in press. 19 Overall J, Scott J, Rhodes H, Lesser J: Empiri-
adults: Comparison of adults with mild demen- 13 Strauss E, MacDonald, SWS, Hunter M, Moll cal scaling of the stages of cognitive decline in
tia, adults with arthritis, and healthy adults. A, Hultsch DF: Intraindividual variability in senile dementia. J Geriatric Psychiatry Neurol
Neuropsychology 2000;14:588–598. cognitive performance in three age groups of 1990;3:212–220.
7 Knotek PC, Bayles KA, Kaszniak AW: Re- older adults: Cross-domain links to physical 20 Reisberg B, et al: Overview of methodologic
sponse consistency on a semantic memory task status and self-perceived affect and beliefs. J issues for pharmacologic trials in mild, moder-
in persons with dementia of the Alzheimer Int Neuropsychol Soc 2002;8:893–906. ate and severe Alzheimer’s disease. Int Psycho-
type. Brain Language 1990;38:465–475. geriatr 1996;8:159–193.

Cross-Domain Variability Gerontology 2004;50:39–43 43

198.143.47.65 - 8/13/2015 4:46:38 PM
Emory University
Downloaded by: