Mini Mental State Examination

Abbreviation: MMSE

This text describes the MMSE as a single instrument, as opposed to its use as part of the CERAD Test Battery. Here, a norming based on purely random samples is described, including in old age. In contrast, the norming of the MMSE as part of the CERAD Test Battery, like the norming of the MoCA, is also based on random samples, but on samples from which age-related cognitive disorders were excluded. For the testing of patients with suspected age-associated cognitive impairment, only these latter normings are meaningful. The present text on the MMSE as a single test will soon be revised so that the difference between the two kinds of norming becomes clearer.

This text describes the MMSE as a standalone test. In a second text, it is described as part of the CERAD test battery. The norming procedures described in the two texts differ fundamentally. This text describes a norming based on purely random samples. In such (rarely found) samples, the proportion of participants with identifiable cognitive impairments increases with age. Nevertheless, these participants were included in the samples. Purely random samples thus serve to describe actual test performance measured across an entire population.

However, the Mini-Mental State Examination (MMSE) is most commonly used in the diagnosis of individuals with cognitive impairments. For this purpose, normative data, which include an increasing number of individuals with cognitive impairments as age increases, are not suitable. If one wishes to diagnose individuals with cognitive impairments in older adults, a normative group is needed that, while ideally also derived from a random sample, does not include any individuals with cognitive impairments. The standardization of the MMSE as part of the CERAD test battery—as well as the standardization of the MoCA—is based on such samples. For testing patients with suspected age-related cognitive impairment, only the latter standardizations are appropriate.

The Mini Mental State Examination was presented by Folstein et al. in 1975 in a journal article.¹⁾ The aim of the publication was to present a simple and practical procedure for estimating the level of mental performance of cognitively impaired persons of advanced age. The measurement range of the instrument covers a very broad range between “extremely impaired” (0-5 points) and “normal” (28-30 points). The scores are thus suited to the intended goal of staging mental performance over a very wide range.

The MMSE is most often used with demented patients. Here its measurement range extends from “normal” to severe forms of dementia. In the first five of the seven stages comprised by the frequently used Global Deterioration Scale of Reisberg et al.,²⁾ the MMSE can measure, namely from stage 1 (no impairment of memory in the clinical interview) to stage 5 (forgets important data of personal life, e.g., names of children, is disoriented in time or place). Stages 6 (needs help with activities of daily living) and 7 (loss of all verbal abilities, incontinence) can no longer be captured by measurements of cognitive abilities.

The tasks were already described so well in the original publication³⁾ that anyone could produce the little test material required themselves.

In the German-speaking region many different versions are in use, some of them with modified tasks. Well comparable with the original is the German version published long ago by Kessler et al. by the Beltz publishing house⁴⁾, which is still available as part of the Demenz-Test. A German-language MMSE is also available as a component of the CERAD Test Battery. Professionals can order an authorized version of the German CERAD as a paper version or download it as a pdf after registering on the website of the Basel Memory Clinic.

The MMSE was first published in 1975 in an article in the Journal of Psychiatric Research by Folstein and colleagues as a simple method for examining mental functions in psychiatric patients⁵⁾. The appendix of this article contained a ready-to-fill-in answer sheet, which also included the questions and instructions. Between 1975 and 2000, nothing was publicly heard or read about copyright questions concerning the MMSE. The test was reprinted or copied by clinicians and researchers in many versions, variants, and languages. Publishers that published foreign-language versions of the MMSE, and pharmaceutical companies that used the test in studies, presumably settled the copyright questions beforehand, however.

From 2000 on, the situation changed. It also became publicly known that the copyright, which had been held by the authors since 1975, was transferred to a joint company of the authors and from there, a year later, to Psychological Assessment Resources, a large test publisher in Florida. Since then the copyright has been enforced by the publisher - above all in the USA and against large organizations - and violations have been sanctioned. There are various statements on this by authors and organizations; an overview is given by the Wikipedia article already cited above. If one searches the internet for MMSE copyright, one finds many further statements.

Users of the German MMSE are probably on the safe side with the purchase of the test materials mentioned above. The use (but not the copying) of original forms that originate from the pharmaceutical industry should also be possible without a copyright violation, since one may assume that the industry paid license fees for the reprint.

The administration is already explained in detail in the original publication. The working memory task was formulated there in two alternatives: either subtract 7 from 100 five times in succession or spell “World” backwards. Both tasks should serve their purpose; in the German-speaking region the word “Preis” is frequently used instead of “World”.

The number of correct answers is the test raw score. It ranges from 0 to 30.

The MMSE is not meant to measure a normally distributed cognitive ability in a population, but grades a loss of cognitive abilities from the “normal state” down to very low scores. The upper end of the measurement range is arbitrarily fixed at a “healthy, unimpaired” level, the lower end at the loss of all cognitive abilities. One must therefore expect the raw score distributions to differ greatly depending on which persons are examined: healthy persons by definition all crowd at the upper edge of the measurement range and yield strongly negatively skewed distributions, whereas for samples of demented persons a rather symmetrical distribution is to be expected.

In contrast to other instruments used in dementia testing, the MMSE has been used in random samples of young adults. Crum et al.⁶⁾ examined 18056 healthy adults over the age of 17 in 1993 in connection with an epidemiological study of the NIMH and processed the results in such a way that at least rough raw score distributions of healthy adults can be presented. Figure 1 shows the data excerpted from Figure 2 of the article and back-calculated into distributions. (The data were taken from Figure 2 and not from the table because the table reports only the 25th, 50th, and 75th percentiles).

Figure 1: Back-calculated distribution shape of the MMSE raw scores in the epidemiological study by Crum et al., 1993 for four selected age groups

In the figure, which for clarity shows only four of 14 age groups, it is easy to see that in most age groups very narrow, negatively skewed distributions result. Only among the over-80-year-olds does the raw score distribution approach a normal distribution, but it is still negatively skewed. All of this is of course a consequence of the deliberate capping of the upper limit of the measurement range in the average or just-average range of cognitive abilities.

In a further study⁷⁾ on a very large random sample of persons over 65 years (total N: 13009), percentile distributions were calculated and plotted for three age groups (Figure 1 on page 325 of the paper), which are resolved much more finely than those of Crum et al. (1993). Figure 2 shows the primary distributions of the MMSE raw scores back-calculated from these data for the three age groups 65-69, 80-84, and over-90-year-olds.

Figure 2: Back-calculated distribution shape of the MMSE raw scores in the MRC study 1998 for three selected age groups

Owing to the large samples and the detailed presentation of the percentiles, the distribution shape can be seen down to very low MMSE scores.

With the two large studies mentioned above, a linearization on the basis of percentile-to-normal transformations is readily possible. The data compiled in Figure 3 help us with this.

Figure 3: Percentile-to-normal transformations of percentile values into IQ-scaled standard scores for selected samples (see text)

The first two curves show the IQ-scaled standard scores as they would result from the simple percentile-to-normal transformation of the percentiles for the group of 80-84-year-olds from the two large studies (Crum et al., 1993⁸⁾ and MRC, 1998⁹⁾). The next two curves show the same for the group of 65-69-year-olds. It can be seen that the two studies yield quite similar norms for these two age groups, which gives the data even more weight than they already have by virtue of their sheer size. The fifth curve from the top consists of two parts: a darker part showing the percentile-to-normal transformation of the percentiles for the 25-29-year-olds from the study by Crum et al. (1993), and a lighter part showing the extrapolation of this curve. The extrapolation roughly follows the actual course among the 65-69-year-olds. Together, the two parts of the curve make up the performance score transformation for the MMSE as an individual test in TDB2Online.

Figure 3 shows yet another curve (drawn in green). It represents the values that result from the percentile data in the publication by Welsh et al. (1994)¹⁰⁾. This is probably the most frequently used percentile curve for the complete CERAD Test Battery, which contains the MMSE as a subtest. Until August 2019 it was also used in TDB2 and TDB2Online. Welsh et al. (1994) examined a sample of 413 healthy examinees at the time, mostly relatives of patients of a memory clinic or other volunteers in the age range between 50 and 89 years (mean 68.6 years). It is remarkable that this sample, with an average age of about 69 years, performs considerably better than the 65-69-year-olds in the two randomized studies, even somewhat better than the young adults of Crum et al. (1993). We will go into this in more detail in the next section.

There are many norming studies for the MMSE. The collection of norms for frequently used dementia screening instruments compiled by Busch and Chapin in 2008¹¹⁾ also contains, for example, 13 publications for the MMSE, although not all of them refer to healthy examinees and not all are broken down by age. After a literature analysis we supplemented the list of studies cited there with 5 further studies. Only studies of examinees in which defined age groups were examined were counted as norming studies. In total, 9 such studies were found; they are compiled with their most important characteristics in Table 1.

The first two studies^{12) 13)} were already discussed in the preceding sections; in both, huge unselected random samples were examined. In the third study (Dufouil et al., 2000¹⁴⁾), too, the examination was unselected, namely all patients of four GP practices in Cambridge. The ADAMS study¹⁵⁾ is distinguished by particularly careful random sampling. The first four studies thus have in common that no selection of patients on the basis of cognitively relevant variables took place, at least none apart from the ability to take part in such an examination. In the next two studies (Bravo & Hébert, 1997¹⁶⁾ and Tombaugh et al., 1996¹⁷⁾), too, random samples were examined first. Only in a second step were persons who suffered from dementia or another acquired cognitive disorder on the basis of an extensive clinical examination removed from the sample.

Table 1: Overview of the norming studies combined meta-analytically

In the remaining four studies, by contrast, there was no intention from the outset to draw a random sample. In these samples, deliberately only healthy persons were examined as far as possible. The aim was thus not random samples but ideal samples. In two studies (Swiss CERAD norming study, 2002¹⁸⁾ and Welsh et al., 1994¹⁹⁾), predominantly the healthy relatives of patients of a memory clinic were examined, in one²⁰⁾ study participants were recruited via advertisements, via talks in senior centers, and through further word of mouth, and another²¹⁾ gives no information on the mode, except that the participants were healthy.

As usual, we find different statistics for describing central tendency and dispersion in the studies, some only in graphs. Remarkably often, percentile values are given, which made the derivation of performance scores possible (see above). In almost all studies, means and standard deviations of the MMSE raw scores per age group were reported; the exceptions are the studies by Dufouil et al. (2000) and Bleecker et al. (1988). In both, however, medians and interquartile ranges are given. The question was to what extent these can stand in for means and standard deviations given the distributions at hand. In the present case this could be checked empirically. For four studies, including the two large ones, both means and standard deviations as well as medians and interquartile ranges were available. In the following figures the corresponding statistics are compared side by side, in Figure 4 means and medians, in Figure 5 standard deviations and the interquartile range divided by 1.35 (75th percentile minus 25th percentile divided by 1.35). This is the fraction of the interquartile range that corresponds to the standard deviation in a normal distribution.

Figure 4: Course of means and medians of the MMSE raw scores in four studies

It is quickly seen that the statistics correspond astonishingly well. In none of the studies are there larger deviations between means and medians, which was not necessarily to be expected given the strong skew of the distribution. (In the MRC study, the medians are available for only three age groups, which is why no line is drawn through the three open triangles there.)

Figure 5: Course of standard deviations and corresponding interquartile range proportions of the MMSE raw scores in four studies

For the dispersion measures, too, the agreement is greater than one would have expected. Only in the study by Crum et al. (1993) are there larger deviations in the older age groups: here the medians are higher than the corresponding standard deviations. The result is probably attributable to the broadened middle range that the percentile curve of the Crum et al. data (Figure 1) shows compared with the MRC data (Figure 2) (without a reason for this being found). Overall, however, the correspondence here too is good enough that, for the meta-analysis of the data, we can with a clear conscience include the medians and IQRs of the two studies that do not report means and standard deviations.

The next two figures now show the age trajectories of the central tendencies and dispersions from all 9 studies, Figure 6 the means (or the medians for two of the studies) and Figure 7 the standard deviations (or the interquartile range divided by 1.35 for two of the studies).

Figure 6: Age trajectories of the mean MMSE raw scores in nine norming studies

In Figure 6, studies with random samples are drawn in shades of green and blue, those with selected healthy persons in shades of red and yellow. It can be seen at first glance that the random samples yield considerably lower raw scores than the selected samples. This holds across all age groups, but the effect becomes much stronger in the higher age groups. The average raw score difference between the random and the selected samples is about 1 raw score point among the younger examinees, 2 to 3 raw score points among the 70-80-year-olds, and 4 to 7 raw score points among the very old examinees. Even the two studies in which the examinees with clinically abnormal signs of dementia were eliminated from the random sample (Bravo et al., 1997 and Tombaugh et al. 1996) still yield lower raw scores than the four samples with examinees who were meant to be as healthy as possible from the outset.

Figure 7: Age trajectories of the dispersion values of the MMSE raw scores in nine norming studies

The dispersion trajectories in Figure 7 likewise show the influence of sampling, although not as clearly. The dispersions of the random samples widen with increasing age and thus yield a plausible trajectory. The dispersions of the selected samples, by contrast, are all lower and remain at a low level even in old age.

The MMSE norming studies impressively show the effect of sampling on the norms of psychological test instruments. Convenience samples of examinees who volunteer for a cognitive examination, or of relatives who bring their family members to a university dementia outpatient clinic, are not representative of the general population. At the same time, however, another effect also comes into play in these graphs: the difference between the two sample types becomes larger the older the examinees are. This has to do with the increasing proportion of demented examinees in the higher age groups of the random samples. In the discussion we will use a further study to examine this matter and the different strategies behind it in more detail.

It was already explained above that the transformation of the raw scores into performance scores is based primarily on the young adults (25-29 years) of the study by Crum et al. (1993)²²⁾. For the age norms in TDB2Online, the MRC study of 1998²³⁾ was added in order to broaden the basis for the age norms in the relevant age range from 65 years. The studies with selected samples of healthy persons were not taken into account; for the smaller studies with representative samples, the effort of the meta-analytic calculation was not worthwhile, because the two large studies determine the means anyway. The percentile values of the MRC study are available only for the age groups 65-69, 80-84, and >90, which is why the age groups in between were linearly interpolated at the performance score level. They were then averaged, weighted, with the corresponding data of the Crum et al. study. As a last step, the age norm limits were smoothed across the age groups with a fifth-order moving average.

These calculations result in the ability and age norming of the MMSE as an individual test for TDB2Online shown in Figure 8. As usual, the vertical lines in this figure mark the performance scores from 40 to 145. The raw scores are plotted on the black horizontal line. The alternating red and blue lines contain the norm limits for all age groups in the overview. The five markers on each line stand for the percentile ranks 2.5, 16, 50, 84, and 97.5. One such line, namely the one that corresponds to the examinee's age group, is drawn (in a different form) in the TDB2Online profile sheet to help with the individual interpretation of the test results.

Figure 8: MMSE performance scores with raw scores and age norm limits (see text)

In the young adult cohorts, the median of the raw scores is about 29 raw score points. The measurement range of the instrument (30 raw score points) ends at a performance score of 110. In the upper measurement range (between 27 and 30 raw score points), the MMSE measures only very coarsely; one raw score point more or less corresponds to about 10 (IQ-scaled) performance scores. Below 26 raw score points the measurement precision is markedly finer; here one raw score corresponds to only about 3 performance scores.

As expected, mean performance on the MMSE varies strongly with age, with the medians declining progressively with increasing age. The mean performance of 80-year-olds is about one standard deviation below the mean performance of 60-year-olds.

One must bear in mind that these data are all based on purely random samples. In the higher age groups these contain an increasingly large proportion of persons who suffer from pathological disorders of cognition. In the present norming these persons are not excluded. The norming simply reflects the result of a random sample. The MMSE is included in TDB2Online a second time as a subtest of the CERAD with a different norming, in which the demented persons were excluded from the normative samples.

In an earlier version of TDB2 and TDB2Online, which was in use until August 29, 2018, the performance score transformation of the MMSE was based on the study by Welsh (1994) and the age norm limits (essentially) on the data of the Basel norming study. This led to the age norm limits shown in Figure 9.

Figure 9: Old version of the MMSE performance scores with raw scores and age norm limits

It can be seen that there are no large differences in the performance score transformation itself: for good performance (28 to 30 raw score points) the performance scores are almost the same in both versions, and toward the lower end they differ increasingly. 15 raw score points corresponded to a performance score of 41.1 in the old version, but to 48.5 in the new version. One must bear in mind, however, that the performance score transformation in the old version was based on the data of older adults (from convenience samples).

For the age norms the differences are larger. For 80-92-year-olds (there was no finer age resolution in the old version), a median performance of about 27.5 raw score points was expected, whereas in the new version only 22 (90-94-year-olds) to 25 (80-84-year-olds) are expected. Low test scores of old and very old persons were therefore rated relatively quickly as “abnormally low” according to the old norms. This was a direct consequence of the normative data used, which did not come from random samples.

TDB2Online does not store test scores, TDB2 only raw scores. Reprinting a test profile in TDB2 is automatically done with the new norms and can, in an individual case, be compared with old printouts at hand.

Why do the available norming studies on the MMSE diverge more from one another than those on other test procedures? Why do the sampling strategies diverge above all for instruments used in old age?

The classical strategy for norming psychological tests is studies with samples representative of the population. It is standard practice here not to exclude from the sample even persons with identifiable causes of a cognitive disorder, as long as they can carry out the test according to the instructions. It is normally even accepted as standard (although by no means always done) that persons with intellectual disability, who are not captured by the usual sampling procedures (children, for example, in schools), must be sought out specifically so that the corresponding quotas can be met ^{24) 25)}. Such a strategy accepts that the sample will contain some persons who, at the time of testing, are not in full possession of their cognitive abilities for external reasons (primarily disabilities, physical illness, mental disorders). In general, however, there will not be many of them, and disregarding this guideline will not be very noticeable.

When persons of advanced age are examined, this no longer holds. Illnesses that directly affect cognition increase in old age to an extent that can no longer be neglected. The older the group examined, the stronger the effect. Purely random samples of persons of advanced age inevitably contain many persons whose performance is lowered for identifiable reasons, in particular because of pathological disorders of cognitive abilities.

Some of the studies that have supplied normative data for the MMSE were not primarily planned as studies representative of the population, but were meant to examine healthy examinees suitable as a control group for patient groups with cognitive disorders. Thinking in a medical model, the adequate control group for patients with dementia is a healthy group without dementia. Tests for the detection of dementia (the MMSE is primarily used as such a test) are meant to separate affected from unaffected persons as well as possible. A narrow selection of the unaffected and their selection by further exclusion criteria favors statistical separation. Here, then, the focus is not on the adequate description of ability but on the most effective possible separation of groups. There are a number of studies that have dealt with how samples can be made even more “robust”. In the studies by Sliwinski et al. (1996)²⁶⁾ and De Santi et al. (2008)²⁷⁾, as is usual in such studies, initially only persons with dementia or mild cognitive impairment (MCI) at the baseline examination were excluded from the sample. In addition, the persons in the sample were followed longitudinally for several years. If one then retrospectively eliminated from the baseline samples persons who had not developed MCI or dementia at the baseline measurement but did so in later follow-up examinations (four or more years), the cognitive performance of this baseline sample, thus made more “robust,” increased. In the study by Hassenstab et al. (2016)²⁸⁾, persons were removed from the baseline sample who reached abnormal values in biomarkers for dementia of the Alzheimer type (PET examination for amyloid deposits, cerebrospinal fluid biomarkers, and hippocampal density on MRI) without being abnormal in behavioral measures. Here too the cleaned sample proved to be higher-performing and more homogeneous, especially in the higher age groups. The focus of these studies is the cleaning of the control sample of persons who have cognitive disorders not only because of a manifest but also because of a developing disease.

A sample selected in this way, as healthy as possible, is useful as a control sample for patients with dementia for certain investigative purposes. However, such samples cannot be used to pursue the goal of developing instruments for measuring cognitive abilities representative of the population. That is possible only with random samples, even if healthy and ill persons are mixed there, especially in old age.

The effect that selections by diagnoses, or by later diagnoses in a longitudinal design, have can be demonstrated by way of example with the data of a US longitudinal study. The Institute for Social Research of the University of Michigan has been conducting a Health and Retirement Study (HRS) since 1992, which now represents the entire US population over 50 years with over 30,000 persons (of whom 5000 died after entering the study). In connection with the HRS, the Aging, Demographics, and Memory Study (ADAMS) was carried out, a population-based dementia study. An overview of the study is given by the paper by Langa et al. (2005)²⁹⁾. A sample of 856 persons aged at least 70 years was examined at home between 2001 and 2003 (Wave A) by professionals using, among others, the MMSE, CERAD, and TMT-A and B, and further dementia ratings were added. In the following years, further waves of examination (B after two years, C after 4 years, D after 6 years) were conducted, which allow a longitudinal view, also retrospectively. The HRS also provides many further data with which the results can be specified for particular groups. The data are accessible to researchers on request via the HRS website.

To show what effects targeted selections have in originally population-representative samples, four groups of persons were formed with the ADAMS data and their MMSE raw scores were calculated (the ADAMS data set supplies weighting factors, including ones for longitudinal data):

• Group 1: the complete population-representative (US population 2001) baseline sample at Wave A
• Group 2: baseline sample without those currently with dementia
• Group 3: baseline sample without those currently with dementia and without those who developed dementia over the next 4 years
• Group 4: only those who were neurologically and psychiatrically completely healthy at time A

The mean MMSE raw scores of these four groups were plotted in the familiar figure with the samples examined so far, and the result is shown in Figure 10.

Figure 10: Age trajectories of the mean MMSE raw scores including the four ADAMS groups (see text)

In addition to the data that were already visible in Figure 6, Figure 10 first shows the mean course of the MMSE raw scores from the meta-analysis of the random samples as a solid black line. Added to this are the dashed trajectories of the four ADAMS groups. It can be seen that the ADAMS random sample (labeled ADAMS_M1 in the figure, black dashed) corresponds well to the meta-analytically calculated TDB2 norms (black line). Every further selection of this sample toward the healthiest possible persons leads to samples with better performance. The red dashed curve results when persons with current dementia are removed, the green dashed one when persons with future-onset dementia are also excluded, and the blue dashed one when only persons who were completely healthy at Wave A are included.

Something else is noticeable, however: even when the ADAMS random sample is selected so strongly that only the roughly two thirds who are healthy remain, their performance is still worse than the performance of the selected samples drawn in shades of red and yellow (Basel, Welsh, Bleecker, and Beeri in the figure). This can no longer be explained by the exclusion of persons with dementia in these samples, but only by the fact that intellectually higher-performing persons entered the sample from the outset.