CERAD Test Battery

Abbreviation: CERAD

The Mini Mental State Examination (MMSE) is a simple procedure, in use since 1975, for estimating the level of mental performance of cognitively impaired older persons. One of its special features is that the measurement range of the procedure is broad, from “extremely impaired” (0-5 points) to “normal” (28-30 points).

In TDB2Online the MMSE is included twice with different norm sources and normative data: once as a subtest of the CERAD (this chapter) and once as an independent test. There you will also find a detailed description of the MMSE with regard to history, test material, administration, and copyright.

From the ADAMS study, the distributions of the raw scores of the MMSE subtest can be calculated for all age groups between 70 and 94. The distribution shape (Figure 2) is extremely skewed in all age groups. In the younger age groups, more than half of all persons achieve at least 28 points, and in the oldest group still at least 27.

Figure 2: Distribution of the CERAD MMSE raw scores in the ADAMS study

Because of the skewness of the distribution, the data from Welsh-Bohmer, which are available only in coarse percentiles, cannot be used meaningfully.

If, on the basis of the ADAMS data in Figure 2, the raw scores are transformed into standard scores by means of an percentile-to-normal transformation, here IQ standard scores with a mean of 100 and a standard deviation of 15, the relationships shown in Figure 3 result. With raw scores of 27 or 28, IQ standard scores of 100 are reached in all age groups. Above a standard score of 100, hardly any measurement is possible, because the raw score range almost ends here. Toward the lower end, however, the measurement range is very differentiated, which is also the purpose of the procedure.

Figure 3: Transformation of the ADAMS CERAD MMSE raw scores into IQ standard scores as an percentile-to-normal transformation, values partly extrapolated

Figure 4 shows the transformation curves after the values were interpolated within the available age groups (by moving averages over three age groups, with the oldest and youngest left unchanged) and extrapolated downward.

Figure 4: Interpolation and extrapolation of the standard scores

The percentile-to-normal transformation of the MMSE raw scores achieved a good linearization; however, the measurement range at the upper end of the scale is inadequate.

To evaluate the norm sources across the age course, the data of all samples that are listed in Table 2 and provide data for the MMSE were summarized in Figure 5. Data from random samples are shown in reddish colors, those from convenience samples in blue and green.

Figure 5: MMSE means across the age course

Many studies provided the data only in coarser age groups. These studies can be recognized in the figure by the fact that the means remain the same in successive age groups. The ADAMS data were calculated in five-year groups; the Welsh-Bohmer data were linearly interpolated. Both samples were linearly extrapolated downward by one (W-B) and two (ADAMS) age groups, respectively. Given the fairly linear age trajectories, this seemed justified to us.

It can be seen quite clearly that test performance in the studies based on random samples is lower than in those with convenience samples. This is particularly evident in the upper age groups. The performance of the examinees in the Swiss norming (green line) is, in most age groups, one or two raw scores above the performance of the examinees in the ADAMS or Welsh-Bohmer study.

Because of the clear mean differences between random and convenience samples, we did not use the latter for the norming. The two random samples show similar trajectories. However, since the data of the ADAMS study are available in much finer gradations, we base the performance scores and the age norms on the data of this study only. We calculate the performance scores themselves on the youngest age group, the 60-64-year-olds.

Figure 6 shows the conversion of raw scores into performance scores and, in addition, the age norm limits for all age groups.

Figure 6: CERAD MMSE Performance Scores with raw scores and age norm limits

In this figure (and all similar ones that follow), the vertical lines mark the performance scores from 40 to 145. On the black horizontal line, selected raw scores are plotted as diamonds in their correct positions. The alternately red and blue lines contain the norm limits for all age groups at a glance. The five markers on each line stand for the percentile ranks 2.5, 16, 50, 84, and 97.5 (in the age group 60-64, the last two coincide at the end of the measurement range). 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. The MMSE's scoring range ends at a performance score of 115.

As already emphasized, the ability of the MMSE to differentiate in the upper half of the distribution is insufficient. But the test is not intended for that. In the performance score range from 40 to 100 (corresponding to raw scores between 15 and 28), the test differentiates sufficiently well; here a performance difference of one standard deviation corresponds to about three to four raw scores. The raw score range below 15 cannot be made visible within the norming, because this is where the range of severe dementia begins and examinees with manifest dementia were not included in the normative sample.

There are three random samples for which the distribution of the raw scores of the Word List Learning subtest can be calculated or back-calculated, albeit with varying quality. For the ADAMS sample the original data are available. For Welsh-Bohmer there are the percentile values 5, 10, 25, 50, 75, 90, and 95, and for Luck there are the percentiles for all raw scores from 0 to 30, although separately for three educational levels and two sexes.

The raw score distributions of the age groups from 70 to 94 calculated from the ADAMS study (Figure 7) are obviously psychometrically favorable and - apart from the marginal ranges, where random effects play a role even with this sample size - correspond largely to a normal distribution.

Figure 7: Distribution of the CERAD Word List Learning raw scores in the ADAMS study

If the raw scores are transformed into standard scores by means of an percentile-to-normal transformation, here IQ standard scores with a mean of 100 and a standard deviation of 15, a largely linear relationship results. Figure 8 shows the curve resulting from the percentile-to-normal transformation, together with a superimposed linear trend, for three age groups.

Figure 8: Transformation of the ADAMS CERAD Word List Learning raw scores into IQ standard scores as an percentile-to-normal transformation (solid line) and as a linear transformation (dashed lines)

As can be seen, the linear approximation is good, certainly better at the ends of the distribution than the percentile-to-normal transformation, which here is subject to random fluctuations due to small numbers even with relatively large samples. We check this against the two other random samples. Figures 9 and 10 show the corresponding data for the norming study by Welsh-Bohmer. There, only the raw scores for seven percentiles are available, which is why the distribution data are coarse. Nevertheless, the linear approximation is similarly good as in the ADAMS study.

Figure 9: Distribution of the CERAD Word List Learning raw scores in the Welsh-Bohmer study

Figure 10: Transformation of the Welsh-Bohmer CERAD Word List Learning raw scores into IQ standard scores as an percentile-to-normal transformation (solid line) and as a linear transformation (dashed lines)

Unfortunately, the age groups in the Welsh-Bohmer study are broader. In Figure 10 they are labeled with the empirical mean age of the respective groups. Since they run linearly and in parallel, it is possible to interpolate the usual 5-year age groups and to extrapolate a group of 60-64-year-olds.

The data of Luck et al. (Figure 11) are unfortunately available for only two age groups, one narrowly defined (75-79) and one of about the same size but very broadly defined (80+). Since the mean age of the total sample was 80.2 years, however, most of those in the 80+ group are likely to have been in the range 80-84.

Figure 11: Transformation of the CERAD Word List Learning raw scores (Luck et al.) into IQ standard scores as an percentile-to-normal transformation (solid line) and as a linear transformation (dashed lines)

These data, too, confirm the normal distribution and good linearity of the raw scores.

For the Word List Learning subtest, the raw scores already yield a very good distribution, which permits a simple linear conversion of the raw scores into performance scores.

To evaluate the norm sources across the age course, the data of all samples that are listed in Table 2 and provide data for Word List Learning were summarized in Figure 12. Data from random samples are shown in reddish colors, those from convenience samples in blue and green.

Figure 12: Mean CERAD Word List Learning performance across the age course

Many studies provided only coarser divisions of the age groups, so that 10 or even 20 years were pooled. This can be recognized by the step-shaped trajectories. The ADAMS data were calculated directly in five-year groups; the Welsh-Bohmer data were linearly interpolated. Both samples were linearly extrapolated downward by one (W-B) and two (ADAMS) age groups, respectively. Given the linear age trajectories, this seemed justified to us. The linear age trajectory in memory performance in this age range can also be seen in the data on the AVLT.

It can be seen that almost all studies with convenience samples yield higher test performance than the studies with random samples. An exception is the study by Murphy (2012), which hardly differs from the random samples. With 99 examinees, however, it was also very small. The Swiss CERAD norming (green line, “Monsch”) shows a similar age trajectory to the random samples; however, the means are about two raw scores above the weighted mean performance of the examinees in the ADAMS and Welsh-Bohmer studies (red line with circles). They are thus about half a standard deviation higher.

Looking at the available norm sources, it makes sense to use only the two random samples that span a wide age range for the norming of the Word List Learning subtest. The study by Luck et al. does have a very large sample but only a very restricted age range. In terms of content, the examinees in this study are only marginally higher-performing than those in the other two studies; nevertheless, including them would distort the age trajectory. We calculate the performance scores themselves on the youngest age group, the 60-64-year-olds.

Figure 13 shows the conversion of raw scores into performance scores and, in addition, the age norm limits for all age groups. Its details are explained in Figure 6.

Figure 13: CERAD Word List Learning Performance scores with raw scores and age norm limits

For the Word List Recall subtest as well, there are three samples from which distributions can be calculated or back-calculated: the ADAMS sample with the original data, and Welsh-Bohmer with the percentile values 5, 10, 25, 50, 75, 90, and 95. For Luck, percentiles are available for all raw scores from 0 to 10, although separately for three educational levels and two sexes. Because of the restricted age range, we dispensed with the distribution analysis here.

The raw score distributions of the age groups from 70 to 94 calculated from the ADAMS study (Figure 14) come - considering the restricted raw score range - still fairly close to a normal distribution. There are slight ceiling effects in the 70-74-year-olds and slight floor effects in the oldest groups.

Figure 14: Distribution of the CERAD Word List Recall raw scores in the ADAMS study

If the raw scores are transformed into standard scores by means of an percentile-to-normal transformation, here IQ standard scores with a mean of 100 and a standard deviation of 15, a still largely linear relationship results, as expected. Figure 15 shows the curves resulting from the percentile-to-normal transformations, together with superimposed linear trend lines, for three age groups.

Figure 15: Transformation of the ADAMS CERAD Word List Recall raw scores into IQ standard scores as an percentile-to-normal transformation (solid line) and as a linear transformation (dashed lines)

As can be seen, the linear approximation is a good approximation of the relationship between raw scores and standard scores, certainly better at the ends of the distribution and in the higher age groups than the pure percentile-to-normal transformation, which here is subject to random fluctuations due to small numbers.

For the norming study by Welsh-Bohmer, only the raw scores for seven percentiles are available, which is why the distribution data (Figure 16) are coarse. Nevertheless, the linear approximation (Figure 17) is similarly good as in the ADAMS study.

Figure 16: Distribution of the CERAD Word List Recall raw scores in the Welsh-Bohmer study

Figure 17: Transformation of the Welsh-Bohmer CERAD Word List Recall raw scores into IQ standard scores as an percentile-to-normal transformation (solid line) and as a linear transformation (dashed lines)

The broadly defined age groups of the Welsh-Bohmer study are labeled in Figure 17 with their empirical mean age. Since a parallel approximation also fits the real data well, it is possible to interpolate the usual 5-year age groups and to extrapolate a group of 60-64-year-olds.

For this subtest, too, the raw scores provide a good approximation to the normal distribution, which permits a simple linear conversion of the raw scores into performance scores.

To evaluate the norm sources across the age course, the data of all samples that are listed in Table 2 and provide data for Word List Recall were summarized in Figure 18. Data from random samples are shown in reddish colors, those from convenience samples in blue and green.

Figure 18: Mean CERAD Word List Recall performance across the age course

Many studies provided coarser divisions of the age groups; this can be recognized by the step-shaped trajectories. The ADAMS data were calculated directly in five-year groups; the Welsh-Bohmer data were linearly interpolated. Both samples were linearly extrapolated downward by one (W-B) and two (ADAMS) age groups, respectively. Given the linear age trajectories, this seemed justified to us. The linear age trajectory in memory performance in this age range can also be seen in the data on the AVLT.

Here, too, it can be seen that the convenience samples generally show higher test performance than the random samples. However, the difference is not as large as for Word List Learning. Again, the study by Murphy (2012) hardly differs from the random samples. In the age range up to 79 years, the performance of the examinees in the Swiss norming (green line) is about one raw score above the weighted mean performance of the examinees in the ADAMS and Welsh-Bohmer studies (red line with circles); in the older cohorts they are comparable. A difference of one raw score corresponds to about half a standard deviation in this subtest.

Just as for the Word List Learning subtest, for Word List Recall we also use only the two random samples that span a wide age range for the norming. We calculate the performance scores themselves on the youngest age group, the 60-64-year-olds.

Figure 19 shows the conversion of raw scores into performance scores and, in addition, the age norm limits for all age groups. To say it again, the details of this figure are explained in Figure 6.

Figure 19: CERAD Word List Recall Performance scores with raw scores and age norm limits

For younger healthy examinees, a memory test with only 10 words has too low a test ceiling. For the 70-74-year-olds this was already evident in the raw score distribution. For the extrapolated data of the 60-64-year-olds it applies even more. However, the CERAD was not developed to examine healthy older persons but those with cognitive deficits. The measurement precision in the lower ranges is not impaired by the ceiling effect. Even in the 90-94-year-olds, test performance at the level of two standard deviations below the age mean (ANS 70 or just under 1 raw score) can still be measured well.

In the Word List Recognition subtest, various definitions of the raw scores are in circulation. In many studies, the number of correctly recognized previously presented words is calculated as raw score 1 and the number of correctly rejected non-presented words as raw score 2. It makes more sense to combine the two scores into one. In TDB2Online we use the sum of these two raw scores minus 10. If this calculation yields a value below zero (very rare, but theoretically possible), it counts as zero. This calculation corresponds to the raw score originally provided by the CERAD Consortium²⁵⁾.

In the Basel version of the CERAD, recognition is evaluated as a discriminability score. The sum of the two raw scores is divided by 20 and multiplied by 100. A discriminability index calculated in this way of 100 (percent) corresponds to a raw score of 10 in TDB2Online, and one of 50 corresponds to a TDB2Online raw score of 0. Discriminability indices below 50 percent indicate results that are below the guessing probability.

Word List Recognition is a CERAD subtest that in some cases permits reasonable clinical conclusions but that is hardly tractable psychometrically. The biggest problem is the low test ceiling. Almost all healthy persons, even those of advanced age, achieve between 8 and 10 raw score points in this test. Differentiated measurement of performance is naturally difficult with only three raw score values. We try it nonetheless.

From the ADAMS study, the raw score distributions can be calculated for all age groups between 70 and 94 years (Figure 20). The extreme skewness of the distribution can be seen, even in the highest age group.

Figure 20: Distribution of the CERAD Word List Recognition raw scores in the ADAMS study

On close inspection, it is noticeable that the 85-89-year-olds in this subtest are better in the lower performance range (roughly between 2 and 8 raw scores) than the 80-84-year-olds. This can also be seen in Figure 21, which shows the percentile-to-normal transformation of the raw scores into IQ standard scores. With the limited sample size, such a break in the age trajectories can certainly occur. It is corrected by moving averages across the age groups.

Figure 21: Transformation of the ADAMS CERAD Word List Recognition raw scores into IQ standard scores by means of percentile-to-normal transformations

In Figure 21 it can also be seen clearly that, because of the different distribution shapes of the raw scores, the transformation curves take different courses from age group to age group.

Figure 22 shows the transformation curves after the values were interpolated within the available age groups (by moving averages over three age groups, with the oldest and youngest left unchanged) and extrapolated downward.

Figure 22: Interpolation and extrapolation of the standard scores

For the norming study by Welsh-Bohmer, the reported percentile norms are far too coarse in view of the skewness of the distribution. We must therefore restrict ourselves to the ADAMS norms.

One of the good properties of percentile-to-normal transformations is that, even with such skewed distribution shapes that in addition differ from age group to age group, they always lead to linear standard scores. The transformation is based merely on the assumption that the measured construct is normally distributed in the population (and, with better quantification of the test, would also be measured as such, one could add).

The poor quantification of test performance in this subtest, however, has the consequence that average and very good performance can practically no longer be distinguished. The highest attainable age-normed score lies between 100 and 110 (see Figure 22).

Clinically, this subtest is of particular importance when a very poor result in Word List Recall meets an average result in Word List Recognition. One can then conclude that attention deficits and/or language deficits — and sometimes a lack of willingness to make an effort — are more likely to have led to poor recall performance, rather than a genuine memory deficit.

Figure 23 shows the means of all relevant possible norm sources for the Word List Recognition subtest across the age course. Again, the data from random samples are shown in reddish colors, those from convenience samples in blue and green colors.

Figure 23: Mean CERAD Word List Recognition performance across the age course

In the younger age groups, the comparison of the studies is unproductive because all means lie between 9 and 10 raw scores. Only from about age 75 do the individual studies differentiate in the manner familiar from Word List Learning and Word List Recall: here the random samples then yield the lower means.

Two further methodological remarks in passing:

Means are unsuitable for reflecting the central tendency of skewed distributions. For most of the studies listed here, however, there are no medians, which would be better suited.

The percentile data in the literature, too, are often not well suited for further calculations, because in their calculation it is often not taken into account that test raw scores yield discrete distributions and that the percentile ranks do not apply to the actual raw score x but to the raw score x +0.5 (because, if the raw scores were made continuous, x would extend from x-0.5 to x+0.5). The smaller the number of raw score values, the greater the possible error due to the incorrect quantification of the percentile ranks.

For the Word List Recognition subtest we use only the ADAMS study for the norming. We calculate the performance scores themselves on the youngest (extrapolated) age group, the 60-64-year-olds.

Figure 24 shows the conversion of raw scores into performance scores and, in addition, the age norm limits for all age groups.

Figure 24: CERAD Word List Recognition Performance scores with raw scores and age norm limits

For healthy examinees, a recognition test for only 10 words has a far too low test ceiling. The highest attainable raw score, 10, thus leads to a performance score of only 102. The test cannot measure better performance. Toward the lower end, however, about three standard deviations can be captured, although not with high precision, as can be seen from the large spacing of the raw scores in the drawing.

Word fluency tasks in which as many animals as possible must be named within one minute are among the most frequently used procedures. This test also occurs as a subtest of the Regensburg Word Fluency Test. In the discussion we will compare the norms of the procedures.

The distribution of the raw scores of the Semantic Word Fluency (SWF) subtest can be calculated or back-calculated both from the ADAMS sample and from the work of Welsh-Bohmer. For the ADAMS sample the original data are available. For Welsh-Bohmer there are the percentile values 5, 10, 25, 50, 75, 90, and 95.

The raw score distributions of the age groups between 70 and 94 years calculated from the ADAMS study (Figure 25) apparently correspond quite well to a normal distribution.

Figure 25: Distribution of the CERAD semantic word fluency raw scores in the ADAMS study

An percentile-to-normal transformation of the raw scores into IQ standard scores yields - apart from the marginal ranges - a largely linear relationship. Figure 26 shows, for three age groups, the curve resulting from the percentile-to-normal transformation, together with the linear trend lines, which were calculated only from the well-populated middle performance ranges.

Figure 26: Transformation of the ADAMS CERAD semantic word fluency raw scores into IQ standard scores as an percentile-to-normal transformation (solid line) and as a linear transformation (dashed lines)

As can be seen, the linear approximation is good, certainly better at the ends of the distribution than the percentile-to-normal transformation, which here is subject to random fluctuations due to small numbers even with relatively large samples. However, a certain relationship between age group and the slope of the curve can be observed: in older persons the slope becomes steeper, that is, the distribution of values becomes narrower. This is probably a consequence of a frequently observed correlation between means and standard deviations in raw score ranges that are bounded below and unbounded above: the better the average performance of a group, the higher the standard deviation as well.

Figure 27 shows the distribution data for the norming study by Welsh-Bohmer. Because only the raw scores for seven percentiles are available there, the distribution shape can be outlined only roughly.

Figure 27: Distribution of the CERAD semantic word fluency raw scores in the Welsh-Bohmer study

An percentile-to-normal transformation of the values into IQ standard scores yields values that can be described well with linear trends (Figure 28)

Figure 28: Transformation of the Welsh-Bohmer CERAD semantic word fluency raw scores into IQ standard scores as an percentile-to-normal transformation (solid line) and as a linear transformation (dashed lines)

Here the trends are largely linear across the age groups, which is why we will also describe them by parallelized linear trends. For the ADAMS samples we generalize the observed correlation of the means with the standard deviations. For the norming, the weighted mean of both samples is used.

For the calculation of the performance scores for the Semantic Word Fluency subtest, we draw on both samples, ADAMS and Welsh-Bohmer, and use the weighted mean of the linear trends of both samples for the conversion.

To evaluate the norm sources across the age course, the data of all samples that are listed in Table 2 and provide data for semantic word fluency were summarized in Figure 29. Data from random samples are shown in reddish colors, those from convenience samples in blue and green.

Figure 29: Mean CERAD semantic word fluency performance across the age course

In many studies the age trajectories were summarized only in coarse 10- or 20-year groups. In the figure they can be recognized by the step-shaped course. The ADAMS data were calculated directly in five-year groups; the Welsh-Bohmer data were linearly interpolated. Both samples were linearly extrapolated downward by one (W-B) and two (ADAMS) age groups, respectively. Given the linear age trajectories, this seemed justified to us.

For semantic word fluency, the studies based on random samples do not stand out as clearly from those based on convenience samples, at least not in the younger cohorts. The Swiss CERAD norms, however, lie clearly above the values of ADAMS and W-B (and above all other studies) across the entire age course. The difference is a good half standard deviation. The patients in the Luck study are also relatively high, which was not the case for the memory tests.

For the norming of the SWF subtest, only the two random samples that span a wide age range were used. We calculate the performance scores themselves on the youngest age group, the 60-64-year-olds.

Figure 30 shows the conversion of raw scores into performance scores and, in addition, the age norm limits for all age groups.

Figure 30: CERAD semantic word fluency performance scores with raw scores and age norm limits

Naming animals for one minute is probably the most thoroughly studied word fluency task. In the discussion of the Regensburg Word Fluency Test, 22 studies on this version of the word fluency task were summarized meta-analytically in the section on the categorical-semantic subtests. The vast majority of these studies examined convenience samples, and the age ranges overlap only partially, so that the means in different age groups are based on different studies. In Figure 31, the mean of the meta-analytically pooled studies, smoothed across the age groups, was added as a black curve labeled M(MA) to the other curves of Figure 28.

Figure 31: Mean CERAD semantic word fluency performance across the age course, results of a meta-analytic norm comparison added

It can be seen that the meta-analytic mean of the SWF raw scores in the age range between 65 and 89 fits our mean from the ADAMS and W-B data quite well. That the age-associated decline in performance is less pronounced overall is to be expected: most studies with convenience samples do not systematically examine all persons in the high age range but rather the fitter ones. One would expect that there. What is unexpected is that our extrapolated means for the two youngest age groups lie above the meta-analytically derived means. However, we found no indication in the data that the age trajectory in this range is not linear.

The Naming subtest is a very easy test. Healthy persons solve all tasks more or less without error up to the highest age. The mode of correct answers lies between 13 and 14 in all age groups (with 15 items).

From the ADAMS study, the raw score distributions can be calculated for all age groups between 70 and 94 years (Figure 32). The distribution is extremely skewed; many persons of every age reach the test ceiling of 15 raw score points.

Figure 32: Distribution of the CERAD Naming raw scores in the ADAMS study

As a consequence, the test can no longer differentiate in the upper performance range. This can be seen in Figure 33, which shows the percentile-to-normal transformation of the raw scores into IQ standard scores. Values below an IQ of 100 are captured well, those above 100 only poorly.

Figure 33: Transformation of the ADAMS CERAD Naming raw scores into IQ standard scores by means of percentile-to-normal transformations

Figure 34 shows the transformation curves after the values were interpolated within the available age groups (by moving averages over three age groups, with the oldest and youngest left unchanged) and extrapolated downward.

Figure 34: Interpolation and extrapolation of the standard scores

For the norming study by Welsh-Bohmer, the reported percentile norms are far too coarse in view of the skewness of the distribution. We therefore restrict ourselves to the ADAMS norms.

The linearization of the measurement scale is achieved by means of the percentile-to-normal transformation.

Figure 35 shows the means of all relevant possible norm sources for the Naming subtest across the age course. Again, the data from random samples are shown in reddish colors, those from convenience samples in blue and green colors.

Figure 35: Mean CERAD Naming performance across the age course

For this subtest we find similar means in most studies, also across the age course. Lower means are found in the study by Murphy, without any apparent reason.

Here, however, the methodological remark from the section on Word List Recognition also applies: means are not good parameters for skewed distributions.

For the Naming subtest we use only the ADAMS study for the norming. We calculate the performance scores themselves on the youngest (extrapolated) age group, the 60-64-year-olds.

Figure 36 shows the conversion of raw scores into performance scores and, in addition, the age norm limits for all age groups.

Figure 36: CERAD Naming performance scores with raw scores and age norm limits

For healthy examinees, the Naming subtest has too low a test ceiling. The highest attainable raw score of 15 leads only to a performance score of 103. The test cannot measure better performance. Toward the lower end, about three standard deviations can be captured well in terms of measurement. That is, after all, the purpose of this subtest.

Constructional Praxis is a CERAD subtest that is psychometrically rather difficult: a limited range of values is combined with a low test ceiling.

From the ADAMS study, the raw score distributions can be calculated for all age groups between 70 and 94 years (Figure 37). It can be seen that this distribution, too, is very skewed. With a raw score range of 0 to 11, the mode lies in all age groups in the range between 9 and 10. At the same time, practically no raw scores below 4 occur.

Figure 37: Distribution of the CERAD Constructional Praxis raw scores in the ADAMS study

Figure 38 shows the percentile-to-normal transformation of the raw scores into IQ standard scores. Certain discontinuities in the course are noticeable here, which could already be seen in the distribution. They arise from the fact that the 11 raw score points do not stem from independent items but from ratings of four figures for which between 2 and 4 points can be obtained.

Figure 38: Transformation of the ADAMS CERAD Constructional Praxis raw scores into IQ standard scores by means of percentile-to-normal transformations

Figure 39 shows the transformation curves after the values were interpolated within the available age groups (by moving averages over three age groups, with the oldest and youngest left unchanged) and extrapolated downward. The discontinuities in the course were deliberately not smoothed, because they did not arise by chance but are due to peculiarities of the construction of the subtest.

Figure 39: Interpolation and extrapolation of the standard scores

For the norming study by Welsh-Bohmer, the reported percentile norms are far too coarse in view of the skewness of the distribution. We therefore restrict ourselves to the ADAMS norms.

The linearization of the measurement scale was again achieved by means of the percentile-to-normal transformation.

Figure 40 shows the means of all relevant possible norm sources for the Constructional Praxis subtest across the age course. Again, the data from random samples are shown in reddish colors, those from convenience samples in blue and green colors.

Figure 40: Mean CERAD Constructional Praxis performance across the age course

In view of the weak measurement properties of this subtest, it is not surprising that most samples show similar means across the age course. For the two outliers, Collie and Beeri, it is again unclear why they show lower means. Furthermore, the caveats mentioned above regarding the informative value of means for such distributions apply. For Constructional Praxis there is the additional point that the scoring rules were certainly not applied uniformly from study to study.

For the Constructional Praxis subtest we use only the ADAMS study for the norming. We calculate the performance scores themselves on the youngest (extrapolated) age group, the 60-64-year-olds.

Figure 41 shows the conversion of raw scores into performance scores and, in addition, the age norm limits for all age groups.

Figure 41: CERAD Constructional Praxis performance scores with raw scores and age norm limits

For healthy examinees, this subtest has a far too low test ceiling. The highest attainable raw score of 11 thus leads only to a performance score of 105. The test cannot measure better performance. Toward the lower end, about three standard deviations can be captured in terms of measurement, although not with high precision, as can be seen from the large spacing of the raw scores in the drawing. The entire relevant measurement range lies between 6 and 11 raw score points; with six measurement points, fine measurement is simply not possible.

The Constructional Praxis Recall subtest was added to the CERAD only in the mid-1990s and was often included only hesitantly by users. Of the studies listed in Table 2, only ADAMS, Welsh-Bohmer, Monsch, and Fillenbaum provide data, the latter, however, from three different studies at once. The ADAMS data are best suited for deriving performance scores. In the section Evaluation of the norm sources we will compare them with the other available studies and explain these in more detail.

From the ADAMS study, the raw score distributions can be calculated for all age groups between 70 and 94 years (Figure 42). The Constructional Praxis Recall subtest has an almost symmetrical distribution and is thus, despite the small range of values, psychometrically easier to handle than Constructional Praxis with its low test ceiling.

Figure 42: Distribution of the CERAD Constructional Praxis Recall raw scores in the ADAMS study

As expected on the basis of the distribution shape, the percentile-to-normal transformation of the raw scores into IQ standard scores shows a largely linear course that can be captured well with linear trends (Figure 43).

Figure 43: Transformation of the ADAMS CERAD Constructional Praxis Recall raw scores into IQ standard scores as percentile-to-normal transformations (solid lines) and as linear transformations (dashed lines)

The distribution of the raw scores corresponds approximately to a normal distribution, which permits a simple linear conversion of the raw scores into performance scores.

In addition to the ADAMS and Monsch data, the data of three samples are available that are summarized in the work of Fillenbaum et al. (2013)²⁶⁾:

1. a random sample of Americans of African descent from Indianapolis (Indiana) without pathological cognitive impairment (CDR score of 0), examined between 1997 and 2004 (N=369)
2. a random sample of Americans of almost exclusively European descent from Cache County (Utah) without pathological cognitive impairment (clinical consensus diagnosis, N=484). This sample is the same as in Welsh-Bohmer et al. (2009)²⁷⁾, but reprocessed in 5-year groups, which is why we prefer it here.
3. a convenience sample of Americans of European descent from the original CERAD normative sample²⁸⁾ without pathological cognitive impairment (N=177).

None of these samples is thus a study representative of the entire USA, but only a random sample from different and limited catchment areas. The schooling of the Indianapolis sample (57% less than 12 years) is clearly below that of the two other studies (17 and 10% less than 12 years, respectively).

Figure 44 shows the means of the five norm sources for the Constructional Praxis Recall subtest across the age course. The norm source favored by us, the ADAMS sample, lies between the two regional random samples, which represent a very highly educated (Cache County) and a very poorly educated (Indianapolis) population. The Swiss sample is found in the upper performance range.

Figure 44: Mean CERAD Constructional Praxis Recall performance across the age course

For the Constructional Praxis Recall subtest we use only the ADAMS study for the norming. We calculate the performance scores themselves on the youngest (extrapolated) age group, the 60-64-year-olds.

Figure 45 shows the conversion of raw scores into performance scores and, in addition, the age norm limits for all age groups.

Figure 45: CERAD Constructional Praxis Recall performance scores with raw scores and age norm limits

For young and middle-aged examinees, this subtest, too, is too easy. The highest attainable raw score of 11 leads to a performance score of 107, standardized on the performance of the 60-64-year-olds. Nevertheless, the subtest is well suited for its actual task, namely assessing the visual memory deficits of older people. Here the measurable performance range covers a full four and a half standard deviations across 11 measurement values.

The Basel CERAD-Plus contains one further subtest (lexical or phonemic word fluency, LWF) in which as many words as possible beginning with S must be named within one minute. This subtest is not included in the American CERAD. For the norming it would not help in this case either: for this subtest, norms can be used only if they come from the German-speaking area. Word production for a letter depends on the frequency of occurrence of the letter in everyday language, and this differs across languages. This aspect is discussed in somewhat more detail in the documentation on the Regensburg Word Fluency Test (RWT).

To our knowledge, German norms for lexical word fluency exist only in the Swiss CERAD norming and in the RWT. While the Swiss norms at least extend over three age groups between about 60 and about 85 years, the RWT norms contain only one age group in the age range of interest here, namely “over 65”. The norming of the LWF subtest is therefore based exclusively on the Swiss CERAD norms.

In general, word fluency tests have quite reasonable raw score distributions. This is so for the CERAD version of semantic word fluency (above in this documentation) and it is also so for the Regensburg Word Fluency Test (RWT). Empirical data on this are available only from the RWT, in the form of percentile data. Figure 46 shows the raw score distribution of the “S words” subtest in the one-minute version of the RWT for five age groups.

Figure 46: Distribution of the CERAD LWF raw scores in the one-minute version of the S-words subtest of the RWT

An percentile-to-normal transformation of the raw scores into IQ standard scores yields - apart from the marginal ranges - a largely linear relationship. Figure 47 shows, for three age groups, the curve resulting from the percentile-to-normal transformation, together with the linear trend lines, which were calculated only from the well-populated middle performance ranges.

Figure 47: Transformation of the CERAD LWF raw scores in the one-minute version of the S-words subtest of the RWT into IQ standard scores as an percentile-to-normal transformation (solid line) and as a linear transformation (dashed lines)

As can be seen, the linear approximation is good, certainly better at the ends of the distribution than the percentile-to-normal transformation, which here is subject to random fluctuations due to small numbers even with relatively large samples. Assuming that the conditions in the Swiss CERAD norming are similar, we can average the means and standard deviations reported there across the various groups in a weighted manner and take this as the basis for a simple linear norming.

For this subtest there are only two sources, Monsch and RWT, both convenience samples, which for other subtests have shown relatively high means compared with random samples. In addition, the age norming of the RWT is insufficiently differentiated in the upper age range. Thus only the Monsch norms remain.

For the norming of the LWF subtest, only the Swiss CERAD norming (“Monsch” in Table 2) was used. We calculated weighted means of the means and standard deviations of all subsamples per age group and interpolated the values for 5-year age groups. The performance scores were here, too, normed on the performance of the 60-64-year-olds. Figure 48 shows the conversion of raw scores into performance scores and, in addition, the age norm limits for all age groups.

Figure 48: CERAD LWF performance scores with raw scores and age norm limits

In Figures 49 and 50 we compare the Basel CERAD means and standard deviations with those of the RWT. In both figures, the age groups are plotted at the age midpoint of the respective group. Except for the outer groups, this should roughly coincide with the actual mean age of the group.

Figure 49: CERAD LWF means of the two norm sources across the age course

Figure 50: CERAD LWF standard deviations of the two norm sources across the age course

With regard to the means, the Basel norms and the RWT norms are not too far apart; presumably both are too high in comparison with random samples. The standard deviations are relatively low in the Monsch samples, especially in the oldest group. As a consequence, the age norm limits are closer together in the oldest groups than in the younger groups. On the basis of the RWT data, a correlation between means and standard deviations across the age course should not necessarily be assumed, although this would also be quite conceivable in terms of content. With the available data this cannot be assessed further.

Just like the MMSE, the two Trail Making Tests are also present twice in TDB2Online: once as independent procedures with the usual definition of the performance scores based on the performance of young adults, and once as subtests of the CERAD with norming based on 60-64-year-olds (this section). Of the studies listed in Table 2, only ADAMS, Welsh-Bohmer, Monsch, and Beeri provide data. Only the ADAMS data are suited for deriving performance scores. In the section Evaluation of the norm sources we will compare them with the other available studies and explain these in more detail.

For the two Trail Making Tests, high raw scores (= long times) mean poor test performance.

From the ADAMS study, the raw score distributions can be calculated for all age groups between 70 and 94 years (Figure 51). The TMT-A subtest has an extremely positively skewed distribution. Hardly any examinee over 70 completes the test in less than 25 seconds, the modes lie roughly between 45 and 65 seconds, and many examinees need two minutes or longer, which leads to the far-reaching right tail of the distribution.

Figure 51: Distribution of the CERAD TMT-A raw scores in the ADAMS study

For the conversion of the raw scores into standard scores, only an percentile-to-normal transformation is therefore feasible. Figure 52 shows the corresponding relationship, also for the values extrapolated to the age range 60-69.

Figure 52: Transformation of the ADAMS CERAD TMT-A raw scores into IQ standard scores as percentile-to-normal transformations

The linearization of the raw scores was achieved by means of an percentile-to-normal transformation based on the ADAMS data.

Of the norming studies named in Table 2, four also collected data for TMT-A and TMT-B: ADAMS, Welsh-Bohmer, Monsch, and Beeri.

Figure 53 shows the means of these four norm sources for the TMT-A subtest across the age course. Welsh-Bohmer, Beeri, and Monsch provide very similar data here. The trajectory curve of the ADAMS sample lies clearly above them in all age groups. One reason for the higher time raw scores is probably that, in the ADAMS study, examinees were given comparatively much time to complete the form. For the TMT-A, testing is otherwise often discontinued after 90 seconds. In ADAMS, times of 6 or 8 minutes also occur. These very high values bias the means upward. This can easily be seen when comparing the means with the medians of the ADAMS study, which are plotted as brown dashed lines: the medians lie clearly below the means. With such skewed distributions, means and medians usually yield different values.

The black dashed curve, which can also be seen in the figure, represents the mean of the many studies that entered into the norming of the TMT-A as an individual test. It can be seen that, at least at the level of the means, there are no large deviations from the norm sources included here.

Figure 53: Mean TMT-A raw scores across the age course

For the TMT-A we use only the ADAMS study for the norming. We calculate the performance scores themselves on the youngest (extrapolated) age group, the 60-64-year-olds.

Figure 54 shows the conversion of raw scores into performance scores and, in addition, the age norm limits for all age groups. On the raw score scale in the diagram one can once again see very nicely the distortion of the time scale: the difference between 20 and 30 seconds is, at the level of the performance scores, about as large as the difference between 60 and 120 seconds, namely about one standard deviation in each case. However, it can also be seen that in the higher age groups the test probably reaches its limits. Beyond a test duration of 90 seconds, many motivational factors certainly come into play, also with regard to the interaction with the test administrators.

Figure 54: Performance scores with raw scores and age norm limits

The description is kept brief because most of what was said in the last section about the TMT-A also applies here.

Figure 55 shows the raw score distributions for all age groups between 70 and 94 years. The distributions of the TMT-B are less skewed than those of the TMT-A, especially in the older cohorts. However, many examinees here need times of over 5 minutes, which in the practice of a dementia outpatient clinic probably often leads to the test being discontinued.

Figure 55: Distribution of the CERAD TMT-B raw scores in the ADAMS study

The raw scores were first converted into standard scores by means of an percentile-to-normal transformation. Subsequently, the curves, which partly overlapped because of the uneven trajectories of the distributions, were homogenized by trend calculations across the age groups. Figure 56 shows the corresponding relationships, also for the values extrapolated to the age range 60-69.

Figure 56: Transformation of the ADAMS CERAD TMT-B raw scores into IQ standard scores as an percentile-to-normal transformation

The linearization of the raw scores was achieved by means of an percentile-to-normal transformation based on the ADAMS data.

Of the norming studies named in Table 2, four also collected data for TMT-A and TMT-B: ADAMS, Welsh-Bohmer, Monsch, and Beeri.

Figure 53 shows the means of these four norm sources for the TMT-B subtest across the age course. All four studies provide very similar data here. Also, the difference between the means and the medians in the ADAMS study is not as large as for the TMT-A. This is related to the less skewed distribution of the TMT-B.

The black dashed curve, which can also be seen in the figure, represents the mean of the many studies that entered into the norming of the TMT-B as an individual test. As with the TMT-A, there are no large deviations here either from the norm sources collected in the course of the CERAD norming.

Figure 57: Mean TMT-B raw scores across the age course

For the TMT-B, too, we use only the ADAMS study for the norming. We calculate the performance scores themselves on the youngest (extrapolated) age group, the 60-64-year-olds.

Figure 58 shows the conversion of raw scores into performance scores and, in addition, the age norm limits for all age groups. For the higher-performing half of each age group, the percentile-to-normal transformation works in the sense of a linearization of the scale. For the lower-performing, this is evidently no longer the case. At the lower end of the scale (that is, with long and very long completion times), the TMT-B reaches its limits, at least in the higher age groups.

Figure 58: Performance scores with raw scores and age norm limits