Journal of Gerontology: PSYCHOLOGICAL SCIENCES

2008, Vol. 63B, No. 5, P271P278

Copyright 2008 by The Gerontological Society of America

Death, Dropout, and Longitudinal Measurements of

Cognitive Change in Old Age
Patrick Rabbitt,1 Mary Lunn,2 and Danny Wong2
1

Department of Experimental Psychology, University of Oxford, England and University of W. Australia.

2
Department of Statistics, University of Oxford, England.

During a 20-year longitudinal study of cognitive change in old age 2,342 of 5,842 participants died and 3,204
dropped out. To study cognitive change as death approaches, we grouped participants by survival, death, dropout,
or dropout followed by death. Linear mixed-effects pattern-mixture models compared rates of cognitive change
before death and dropout from four quadrennial administrations of tests of fluid intelligence, vocabulary, and
verbal learning. After we took into account the significant effects of age, gender, demographics, and recruitment
cohorts, we found that approach to death and dropout caused strikingly similar reductions in mean test scores
and amounts of practice gains between successive quadrennial testing sessions. Participants who neither dropped
out nor died showed significant but slight cognitive declines. These analyses illustrate how neglect of dropout
miscalculates effects of death, of worsening health, and of all other factors affecting rates of cognitive change.
Key Words: Cognitive changeDeathDropout.

ANY studies of changes preceding death in old age have

found that individuals who are between 18 months and
11 years from death score less well on tests of mental abilities
than survivors do (e.g., Johansson & Berg, 1989; Lieberman,
1965; Palmore & Cleveland, 1976; Rabbitt, Watson, Donlan
et al. 2002; Reimanis & Green, 1971; Riegel & Riegel, 1972;
Riegel, Riegel, & Meyer, 1967; Siegler & Botwinick, 1979;
Small, Fratiglioni, von Strauss, & Backman, 2002). The typical
methodology has been to assess participants only once and to
compare the scores of survivors and decedents at a later,
arbitrary, census date. This underestimates the effects of death,
because younger decedents are compared against younger
survivors who will survive longer beyond the census date
whereas elderly decedents are compared against elderly survivors who are likely to die soon after census (Rabbitt, Lunn, &
Wong, 2005; Rabbitt, Watson, Donlan, Bent, & McInnes,
1994). This explains the otherwise paradoxical findings that
differences in ability between deceased persons and survivors
are larger in younger than in older samples (e.g., Riegel &
Riegel; Riegel et al.).
Longitudinal studies in which participants are repeatedly
assessed over many years avoid these problems but encounter
other methodological difficulties. Participants improve with
practice as a result of repeated testing (Rabbitt, Diggle,
Holland, McInnes, Bent, et al., 2004); recruitment cohorts
may markedly differ in ability; and participants typically drop
out of studies because of deteriorating health and survivors
become increasingly elite and able (Lachman, Lachman, &
Taylor, 1982; Rabbitt, Watson, Donlan, Bent, & McInnes,
1994; Schaie, Labouvie, & Barrett, 1973). Thus, if scores of
decedents are compared against those of all survivors, including
less able dropouts, then the effects of approaching death are
underestimated (Rabbitt et al., 2005). Demographic factors
must be taken into consideration. Women, the socioeconomically advantaged, and the most able live longer, so deaths
selectively alter sample composition (Hart et al., 2003). More
and less advantaged individuals tend to die from different

causes and so experience different patterns and rates of terminal

declines in health and cognition (Nagi & Stockwell, 1973;
Pincus, Callahan, & Birkhauser, 1987; Snowden, Ostwald,
Kane, & Keenan, 1989). Differences in age must be considered
because, independent of approach to death, age accelerates
the rate of change in performance over time; the effects of
pathologies, and so of approaching death, may also alter as age
advances (Rabbitt, Diggle, Holland, & McInnes, 2004).
Data from the University of Manchester Longitudinal
Study, described in detail elsewhere (Rabbitt, Diggle, Holland,
McInnes, Bent, et al., 2004), allowed us to make analyses of
rates of change preceding death and dropout after the effects
of practice, age, cohort effects, and demographics had also been
considered.

METHODS

Participants and Procedure

A panel of 5,842 volunteers, that is, 2,615 residents of Greater
Manchester and 3,227 residents of Newcastle-upon-Tyne,
United Kingdom, were all sufficiently healthy and motivated to
travel independently to the University of Newcastle-upon-Tyne
or the University of Manchester, where they were given
cognitive tests in groups of 10 to 20. There were 1,711 men
aged between 49 and 93 years (M 65.6, SD 7.7) and 4,131
women aged between 49 and 92 years (M 64.4, SD 7.8).
They were each paid 5 (UK) per session to cover expenses. A
search by Her Majestys Registry Office UK obtained exact
dates and proximate causes for all 2,342 deaths between 1983,
when the study began, and the census date, July 2004. Of 3,204
dropouts, 1,208 also died before the census date. Most dropouts
only appeared as failures to answer invitations for testing
and so could only be dated from the last session attended.
Consequently, to compare rates of changes preceding dropout
and death, we had to use the same dating. We could not
compare changes preceding dropouts from different causes

P271

P272

RABBITT ET AL.

Figure 1. Profiles of percentage scores on the Heim AH4-1 test

(AH41) of fluid intelligence across testing sessions for survivor (C),
death (D), and dropout (W) groups.

because many participants did not reveal them. However, a

previous survey found that, although most respondents before
1993 cited poor health as their reason for dropping out, many
cited positive reasons such as taking up new employment
(Rabbitt et al., 1994). Thus, data pooled over all dropouts
underestimates the effects of illness and increasing frailty. The
remaining 1,996 participants did not drop out before July 2003
and also survived the July 2004 census of deaths. During the
course of the study from 1983 to 2004, 7 participants were
identified as suffering from dementias, either by death certificates or information from relatives, and we excluded these.
Since 2004, all survivors have been screened every 6 months
on the Mini-Mental State Examination and other assessment
protocols for dementias, and so far nine cases have been identified. We have also retrospectively excluded these. Although it
is impossible to be certain that no individuals contributing data
did not also suffer from dementia, the latter figure suggests that
the incidence was sufficiently low as to have hardly any effect
on the comparisons described.

Materials
Some of us have given details of the entire study elsewhere
(Rabbitt, Diggle, Holland, McInnes, Bent, et al., 2004). The
data analyzed here are from tests of general fluid intelligence,
that is, the Heim (1970) AH4-1 intelligence test; of vocabulary,
that is, the Raven (1965) Mill Hill B Vocabulary Test; and
cumulative verbal learning, that is, the Cumulative Verbal
Learning task (CVL task) administered on four successive
occasions at 4-year intervals between 19831985 and 2003.
The AH4-1 test consists of 64 logic, verbal comparisons, and
arithmetic problems. Scores are the percentages of correct
answers given within 10 minutes. The Mill Hill B Vocabulary
Test requires correct definitions for each of 34 words with no
time limit. For the CVL task, 15 three-syllable words matched
for frequency (1/100,000) and concreteness are projected, one
at a time, on a screen at a rate of 1/1.5 s. Participants recall as
many words as possible and the words are then shown three
times in different random orders and recalled without sight of
previous attempts. Scores are percentages of correct answers.

Figure 2. Profiles of percentage scores on the Cumulative Verbal

Learning task across testing sessions for survivor (C), death (D), and
dropout (W) groups.

Levels of socioeconomic advantage (SEA) are indexed by

the Office of Population Censuses and Surveys (1980) in their
classification of occupational categories. These categories are
as follows: C1, made up of professionals such as doctors,
lawyers, senior managers, and academics; C2, which consists of
other professionals such as schoolteachers, junior managers,
and pharmacists; C3N, made up of skilled nonmanual workers
such as secretaries and clerical workers; C3M, consisting of
skilled manual workers such as plumbers, craftsmen, joiners,
fitters, and machinists; C4, made up of nonskilled, nonmanual
workers such as security guards; and C5, which consists of
nonskilled manual workers such as cleaners. We categorized
those participants who did not reveal occupations as nonresponders (labeled NR). We include SEAs, city of residence,
gender, and recruitment cohort in all analyses.
We divided participants into 11 groups according to their
histories of survival, dropout, and death, logged with respect to
the four quadrennial assessment time points: T1, T2, T3, and
T4. The groups, labeled as D (for death), W (for withdrawal), or
WD (for withdrawal followed by death), are as follows: D1
completed T1 but died before T2; D2 completed T1 and T2 but
died before T3; D3 completed T1, T2, and T3 but died before
T4; D4 completed T1, T2, T3, and T4 but died before the close
of the census of deaths in 2004; WD1 completed T1 but
withdrew before T2 and also died before the 2004 census; WD2
completed T1 and T2 but withdrew and then died before the
2004 census; WD3 completed T1, T2, and T3 but withdrew
and then died before the 2004 census; W1 completed T1 and
withdrew before T2 but survived the 2004 census; W2 completed T1 and T2, withdrew before T3, but survived the 2004
census; W3 completed T1, T2, and T3, withdrew before T4, but
survived the 2004 census. Group C was a control group whose
members completed all four assessments and survived the 2004
census.
The average raw scores for each of these death and dropout
groups at each of the four testing occasions, that is, T1, T2, T3,
and T4, are plotted for the AH4-1 test in Figure 1, for the CVL
task in Figure 2, and for the Mill Hill B Vocabulary Test in

DEATH, DROPOUT, AND COGNITIVE CHANGE

P273

Table 1. Means and Standard Deviations of Scores on

Intelligence, Verbal Learning, and Vocabulary Tests for
Age and Demographic Groups
M(SD) AH4-1

Subset

Figure 3. Profiles of percentage scores on the Mill Hill B

Vocabulary Test (MHB) across testing sessions for survivor (C), death
(D), and dropout (W) groups.

Figure 3. Note that we have not taken effects of differences in

age, recruitment cohort, and demographics into consideration.
Although these means illustrate the logic of grouping individuals in this way, they only provide very approximate indications of specific comparisons between data points. Results of
comparisons between these means, after age, cohort, and
demographics have been considered, are described in the
following paragraphs.
Figures 1, 2, and 3 show that, for all tests, and most clearly
for the AH4-1, there is a clear vertical separation between most
groups from Group C through Group WD3. We exploit this in
the pattern-mixture mixed-effects model by introducing an
indicator for the group to which the individual belongs.

RESULTS
Table 1 shows details of the demographic and age categories,
and death and withdrawal categories, and mean scores of these
subgroups on the AH4-1 test, the Mill Hill B Vocabulary Test,
and the CVL task.

Analyses
Methodology. Our analyses were based on the models
described by some of us elsewhere (Rabbitt, Diggle, Smith,
Holland, & McInnes, 2001; Rabbitt, Diggle, Holland, McInnes,
Bent, et al., 2004; Rabbitt et al., 2005). We consider age,
gender, socioeconomic status, and whether participants are
taking the test for the first, second, third, or fourth time (the
practice effect). We also include cities of residence and years of
recruitment of cohorts to adjust for unidentified confounding
factors. The model can be considered in two parts: a model for
the average response over time for a subject with given values
of all explanatory variables, and a model for the random
variation about the main response.
The aim is to determine the effect of imminent death on
a participants cognitive performance. However, some participants who withdraw survive beyond the census date whereas
others do not, giving three mechanisms by which they fail to

Manchester (2,615)
Newcastle (3,227)
Men (1,711)
Women (4,131)
Oclass C1 (261)
Oclass C2 (1,849)
Oclass C3(NM)(2,061)
Oclass C3M (771)
Oclass C4 (433)
Oclass C5 (47)
Oclass Na (420)
Age 4960 (1,347)
Age 6170 (2,849)
Age 7180 (1,477)
Age 80 (169)

M(SD) CVL

M(SD) MHB

49.6
47.5
50.9
47.4
61.1
56.2
48.4
39.2
34.4
31.0
40.5
55.4
49.5
41.8
33.7

(17.7)
(17.3)
(17.7)
(17.4)
(14.0)
(16.6)
(14.9)
(15.8)
(14.0)
(16.6)
(18.3)
(17.1)
(16.6)
(16.5)
(15.5)

71.6
67.5
66.2
70.4
72.2
73.1
70.1
63.0
61.7
55.5
62.6
75.5
69.7
63.5
57.5

(13.9)
(14.8)
(15.1)
(14.2)
(13.4)
(13.4)
(13.6)
(14.4)
(14.7)
(17.4)
(18.0)
(12.9)
(13.8)
(14.7)
(16.7)

54.8
45.2
52.1
48.5
63.3
58.3
48.1
40.6
35.3
29.9
42.7
51.7
49.4
48.1
47.8

(17.9)
(17.9)
(18.4)
(18.5)
(15.8)
(16.6)
(16.2)
(16.7)
(16.3)
(16.7)
(20.1)
(18.0)
(18.2)
(19.4)
(20.0)

56.3
43.0
46.7
49.9
53.6
39.4
42.9
48.7
45.2
49.8
51.4

(16.2)
(17.4)
(16.6)
(14.8)
(16.3)
(17.5)
(15.4)
(16.0)
(17.1)
(16.8)
(15.6)

74.9
63.4
65.3
70.1
70.2
62.5
63.2
67.9
69.6
71.7
71.4

(12.1)
(16.8)
(14.5)
(12.7)
(12.9)
(15.9)
(14.8)
(14.1)
(14.3)
(13.2)
(13.3)

53.0
48.3
50.8
52.8
50.4
45.2
47.0
50.5
45.6
50.4
52.9

(17.4)
(19.1)
(19.2)
(18.7)
(18.8)
(19.5)
(19.4)
(19.5)
(18.2)
(17.8)
(16.6)

Pattern
C (1,504)
D1 (365)
D2 (408)
D3 (246)
D4 (115)
WD1 (745)
WD2 (354)
WD3 (109)
W1 (1,013)
W2 (595)
W3 (388)

Note: AH4-1 Heim AH4-1 test of fluid intelligence; CVL Cumulative

Verbal Learning task; MHB Mill Hill B Vocabulary Test; Oclass occupational class (see text for details on the C designations). For patterns, C
control, D death, and W withdrawal (see text for particular designations).

complete the tests: death, withdrawal, and withdrawal followed

by death. The statistical method determines whether or not
these three mechanisms have similar effects on the test scores.
There are at least two possible approaches to handling
missing data. One uses a selection model, and the other uses
a pattern-mixture model. We are interested in a retrospective
analysis of the effects of death and dropout, and so a patternmixture model is an ideal tool. The probability density function
is factored as f(y, d) f(yjd)f(d), where y represents the response
and d represents the dropout data. We will be interested in the
first factor on the right-hand side of the equation, which
represents the response conditional on the dropout pattern. In
this data set, dropouts occur between test sessions or between
the last session and the end of the study, but they are also of
three types: death, dropout followed by death, and dropout with
survival after census. We can thus classify volunteers into
groups by dropout type or pattern. We model the conditional
density as we did in Rabbitt and colleagues (2001), that is,
Yij lij Ai Bi xij1 Eij ;
where the mean value of Yij is given by
lij b0 b1 xij1 b2 xij 2 . . . bp xijp c2 gi 2 . . . cr gir :
Index i is the identifier of the volunteer, index j gives the
occasion of the test (j 1, . . ., 4), and xijk denotes the values of
the explanatory variables, where k runs from 1 to p and gim is an

RABBITT ET AL.

P274

Table 2. Model for Analysis of Scores on the AH4-1

Intelligence Test
CE

SE

df

53.63
0.77
0.014
2.25
1.42
9.58
6.14
8.56
11.96
14.85
5.42
3.71
5.54
7.22

0.75
0.04
0.00
0.44
0.66
0.96
0.46
0.60
0.74
2.07
0.79
0.18
0.29
0.52

6732
6732
6732
5814
5814
5814
5814
5814
5814
5814
5814
6732
6732
6732

70.68
19.16
5.78
5.09
2.13
9.91
13.26
14.09
16.06
7.14
6.78
20.43
19.08
13.78

,.0001
,.0001
,.0001
,.0001
.03
,.0001
,.0001
,.0001
,.0001
,.0001
,.0001
,.0001
,.0001
,.0001

2.49
1.86
1.16
0.27
1.59
0.55
9.28
1.86
3.08

1.73
0.89
0.85
0.68
1.10
0.74
2.79
0.86
0.82

5814
5814
5814
5814
5814
5814
5814
5814
5814

0.28
2.09
1.36
0.40
1.43
0.74
3.32
2.16
3.72

.7737
.0365
.1713
.6884
.1502
.4552
.0009
.0305
.0002

7.22
5.03
3.71
0.41
9.93
7.24
3.43
8.15
4.71
2.46

0.87
0.79
0.96
1.34
0.69
0.85
1.38
0.60
0.68
0.80

5814
5814
5814
5814
5814
5814
5814
5814
5814
5814

8.21
6.31
3.86
0.31
14.28
8.47
2.48
13.45
6.85
3.07

,.0001
,.0001
.0001
.7569
,.0001
,.0001
.0129
,.0001
,.0001
.0021

0.08
0.04
0.12
0.20

0.03
0.02
0.04
0.07

6732
6732
6732
6732

2.66
1.82
2.62
2.61

.0076
.0686
.0086
.0089

Item
(Intercept)
Age
Age2
Gender
City
Oclass C1
Oclass C2
Oclass C3M
Oclass C4
Oclass C5
Oclass Na
T2
T3
T4

Entry year
1982
1983
1984
1986
1987
1988
1990
1991
1992
Pattern
D1
D2
D3
D4
WD1
WD2
WD3
W1
W2
W3
Interactions
Age
Age
Age
Age

3
3
3
3

Gender
T2
T3
T4

Residual standard deviations: Levels 1 and 2

Level
2
2
1

Intercept
Age
Residual

SD

correlation with intercept

13.086
0.389
5.467

0.05

Note: CE coefficient; T1T4 Times 14; Oclass occupational class

(see text for details on the C designations); Oclass Na occupational class
not specified. For patterns, D death and W withdrawal (see text for particular designations).

indicator, a value 1 denoting to which of r type or pattern

groups the ith volunteer belongs. Only one of these indicators is
nonzero. The first two variables xij1 and xij2 are related to age
and age squared, with a value of zero corresponding to age 49,
which is close to the mean age.
As in Rabbitt and colleagues (2001), the random effects
(Level 2) are given by Ai Bixij1 , where (Ai, Bi) are bivariate
normal with a mean of zero and covariance matrix

rA2
rAB

rAB
rB2

and are independent of the Level 1 error, Eij, which follows a univariate normal distribution with a mean of zero and
variance rE2.
This allows us to quantify the mean effect of death or
withdrawal, including also the test occasion by which this has
occurred. It also allows us to make a direct comparison of the
effects of death and of withdrawal, as we shall see.
We have also considered interactions between the various
explanatory variables and these are included in the final models.
Analysis of AH4-1 intelligence test scores. To illustrate
general trends common to all tasks, we give analyses of AH4-1
scores in detail. Scores for vocabulary and cumulative learning
are jointly discussed in the subsequent text. Table 2 shows
results from a linear mixed-effects pattern-mixture model comparing percentages of correct unadjusted AH4-1 scores for the
death and dropout groups after the effects of age, gender,
occupational category, city of residence, recruitment cohort,
and practice are considered. Age was centered at 49, the lowest
recorded in the study. Variance between individuals has
been modeled with random effects because these are longitudinal data and measurements from the same individual are
correlated.
Significant linear and quadratic terms show an accelerating
decline in mean scores with increasing age. Because there is no
Age 3 Death or Age 3 Dropout Group interaction, there is no
evidence that the cognitive effects of death or dropout differ
with the ages at which they occur. There is no interaction
between group and any of the variables of age, gender, city,
occupational class, or cohort (entry year), but there is a
significant constant effect of group at each session time, as seen
in the plot of the raw data. (It is not possible to look for an
interaction with session time because these regression coefficients would be unidentifiable.)
There are, however, some significant interactions of age with
practice gains. The interaction between age and the difference
between scores at T1 and T2 is not significant, but the
interactions between age and the T3 versus T1 difference (p
.0086) and between age and the T4 versus T1 difference (p
.0089) are significant. In survivors, deceased individuals, and
dropouts alike, the effect of age on test scores increases with
the interval over which it is measured. Overall, men (M 50.9,
SD 17.7) scored higher than women did (M 47.4, SD
17.4). The significant Age 3 Sex interaction shows that even
when longevity has been taken into account, women decline
less as they age. Mancunians or persons from Greater
Manchester (M 49.6, SD 17.7) score higher than
Novocastrians or persons from Newcastle (M 47.5, SD
17.3). Scores range from 61.1, SD 14.0, for C1 (the most
advantaged SEA group), to 31.0, SD 16.6, for SEA C5. This
is reflected in the main model, shown in Table 2. There, we can
see that the mean score of C5 is over 14% lower than that of
C3 (both C3N and C3M), the baseline class, and that of C1 is
9.5% higher than C3 (C3N and C3M). There are significant
differences between recruitment waves with performance for
entry years 1990 and 1992 that are greater than those for 1985
(baseline). There is a highly significant practice gain between

DEATH, DROPOUT, AND COGNITIVE CHANGE

Table 3. Specific Comparisons of Mean AH4-1 Intelligence Test

Scores Between Death and Dropout Groups
AH4-1
C
D1
D2
D3
D4
W1
W2
W3
WD1

D1

D2

D3

D4

.0001 .0001 .0001

W1

W2

W3

.0001 .0001
.0001
.0002

.0001

.0001 .0001

Table 4. Model for Analysis of Scores on the Cumulative

Verbal Learning Task

WD1 WD2 WD3

Item

.0001 .0001
.001

(Intercept)
Age
Age2
Gender
City
Oclass C1
Oclass C2
Oclass C3M
Oclass C4
Oclass C5
Oclass Na
T2
T3
T4

.0009

.0001

Note: Values of t statistics are derived from the main model. After correction, p , .001 is taken as the threshold for significance. C control group;
D death and W withdrawal (see text for particular designations).

T1 and T2 followed by rather less substantial gains (less than

2%) between T2 and T3 and between T3 and T4.
After we consider these effects and interactions, then we compare the death, withdrawal, or death followed by withdrawal
groups against those for the control group of survivors,
Group C. Table 3 compares average scores throughout the study
between Group C survivors and all others by using t values
computed from the main model. Group C survivors score
higher than any of the members of the death and withdrawal
groups except Group D4, whose members also completed all
four assessments but then died within the 12 months before the
death census in July 2004. Table 3 shows comparisons by twotailed t tests between all other death and withdrawal groups.
After applying a Bonferroni correction for multiple testing, we
take the threshold level for significance for these multiple
comparisons as p .001.
Comparing the effects of death and dropout. Specific
comparisons using t tests (calculated from the main model)
compared differences in the amounts of practice gains between
survivors and death and withdrawal groups. After correction the
threshold for significance is p .001. Persons in Group D2
completed assessments at T1 and T2 but then died before T3.
From the raw data their nonsignificant gain in average scores
between T1 and T2 was 0.03 points. Again from the raw data,
members of Group C, who completed all four assessments and
survived the census date, shows a gain of 1.34 points between
test sessions T1 and T2. From the main model, over all test
sessions, we see that Group D2 and Group C have significantly
different mean scores (5%), p , .0001. Similarly, Group D1
and Group C have significantly different scores (7%), p ,
.0001, with a rather larger loss of score than Group D2, as
would be expected from their more imminent time of death.
Group D3 again has a significant loss as compared with Group
C but with a smaller mean difference than those of D1 and D2.
Similar results can be seen for comparisons between Group
C and (dropout and survivor) Groups W1 through W3. Rather
more pronounced results of a similar nature are also seen with
Groups WD1 through WD3 (participants who dropped out and
subsequently died).
Both from the raw data (Figure 1) and from the main model
(Table 2), we see that death, dropout, and dropout with subsequent death persons do indeed model in much the same way
as do overall survivors who did not withdraw.

P275

CE

SE

df

68.96
0.82
0.01
4.18
0.53
3.14
2.93
4.59
6.09
11.44
3.17
0.56
4.79
0.21

0.70
0.04
0.01
0.46
0.76
0.90
0.41
0.55
0.68
1.95
0.93
0.24
0.35
0.60

5603
5603
5603
4774
4774
4774
4774
4774
4774
4774
4774
5603
5603
5603

97.93
19.06
3.95
10.28
0.70
3.48
7.12
8.23
8.87
5.85
3.38
2.32
13.57
0.35

,.0001
,.0001
.0001
,.0001
.4824
.0005
,.0001
,.0001
,.0001
,.0001
.0007
.0202
,.0001
.7237

1.77
5.68
1.23
1.41
0.59
0.48
0.62
0.75

1.53
0.87
0.85
0.59
0.98
0.71
0.97
1.49

4774
4774
4774
4774
4774
4774
4774
4774

1.15
6.52
1.45
2.37
0.60
0.68
0.63
0.50

.2490
,.0001
.1468
.0176
.5479
.4957
.5237
.6138

5.25
4.56
0.78
0.17
7.05
6.06
2.40
4.78
2.56
1.58

0.93
0.77
0.92
1.18
0.68
0.75
1.16
0.58
0.60
0.65

4774
4774
4774
4774
4774
4774
4774
4774
4774
4774

5.64
5.91
0.84
0.14
10.34
8.06
2.06
8.22
4.24
2.41

,.0001
,.0001
.4004
.8811
,.0001
,.0001
.0391
,.0001
,.0001
.0157

0.36
0.08
0.18
0.36

0.03
0.03
0.05
0.09

5603
5603
5603
5603

10.41
2.46
3.25
3.92

,.0001
.0137
.0012
.0001

Entry year
1982
1983
1984
1986
1987
1988
1991
1992
Pattern
D1
D2
D3
D4
WD1
WD2
WD3
W1
W2
W3
Interactions
Age
Age
Age
Age

3
3
3
3

City
T2
T3
T4

Residual standard deviations: Levels 1 and 2

Level
2
2
1

Intercept
Age
Residual

SD

correlation with intercept

10.607
0.444
7.433

0.533

Note: CE coefficient; T1T4 Times 14; Oclass occupational class

(see text for details on the C designations); Oclass Na occupational class
not specified. For patterns, D death and W withdrawal (see text for particular designations).

Analyses for cumulative verbal learning and vocabulary.

Tables 4 and 5 and Tables 6 and 7 show the same analyses for
correct percentage scores on the CVL task and Mill Hill B
Vocabulary Test, respectively.
For the CVL task, significant linear and quadratic age terms
show that mean decline accelerates as calendar age increases
over all survival, dropout, and death groups. In the Mill Hill B
Vocabulary Test, the mean effects of age and of age squared are
not significant. In both tests, because age does not significantly
interact with death or withdrawal group membership, there is

RABBITT ET AL.

P276

Table 5. Specific Comparisons of Mean CVL Task Scores

Between Death and Dropout Groups
CVL

D1

D2

D3

D4

W1

W2

W3

WD1 WD2 WD3

C
.0001 .0001
.0001 .0001
.0001 .0001
D1
.0002 .0004
.0066 .0003
D2
.0005 .0009
.0007
D3
D4
W1
.0007 .0001
W2
W3
WD1
.0002
WD2
Note: Values of t statistics are derived from the main model. After correction, p , .001 is taken as the threshold for significance. CVL task Cumulative Verbal Learning task; C control group; D death and W withdrawal
(see text for particular designations).

Table 6. Model for Analysis of Scores on the Mill Hill B

Vocabulary Test
Item
(Intercept)
Age
Age2
Gender
City
Oclass C1
Oclass C2
Oclass C3M
Oclass C4
Oclass C5
Oclass Na
T2
T3
T4

CE

SE

df

57.88
0.01
0.01
1.02
6.08
12.67
8.91
7.36
11.69
16.88
3.57
0.13
1.40
3.24

0.82
0.04
0.01
0.47
0.72
1.03
0.49
0.65
0.80
2.26
0.86
0.26
0.37
0.66

6797
6797
6797
5837
5837
5837
5837
5837
5837
5837
5837
6797
6797
6797

70.51
0.02
0.87
2.12
8.48
12.24
18.06
11.21
14.46
7.46
4.12
0.51
3.70
4.89

,.0001
0.97
.3789
.0339
,.0001
,.0001
,.0001
,.0001
,.0001
,.0001
,.0001
.6039
.0002
,.0001

1.82
0.07
2.36
1.81
3.75
0.12
1.05
3.34
3.06

1.83
0.95
0.90
0.74
1.18
0.80
2.94
0.93
0.90

5837
5837
5837
5837
5837
5837
5837
5837
5837

0.99
0.08
2.59
2.45
3.18
0.15
0.35
3.56
3.38

.3210
.9331
.0094
.0141
.0015
.8751
.7209
.0004
.0007

5.41
3.57
2.31
1.02
7.90
6.09
1.78
7.26
4.67
1.73

0.98
0.85
1.01
0.73
0.76
0.92
1.47
0.65
0.72
0.84

5837
5837
5837
5837
5837
5837
5837
5837
5837
5837

5.51
4.20
2.27
1.39
10.29
6.61
1.21
11.09
6.48
2.05

,.0001
,.0001
.0227
.4652
,.0001
,.0001
.2239
,.0001
,.0001
.0402

0.01
0.04
0.31

0.03
0.06
0.09

6797
6797
6797

0.41
0.72
3.25

.6755
.4687
.0011

Entry year

no evidence that approaching death causes more rapid decline

in older than in younger participants. Women score higher than
men on the CVL task, but a similar result on the Mill Hill B
Vocabulary Test has only a 5% significance level. On both
tasks there are no significant Sex 3 Age interactions and so
there is no evidence that men and women decline at different
rates. Mancunians score significantly higher than Novocastrians
on the Mill Hill B Vocabulary Test, but persons from the
different cities do not differ on the CVL task. Greater SEA
(occupational class) is associated with better overall performance but there is no Occupational Class 3 Test Sessions
interaction and so no evidence that occupational class affects
rates of decline. On both tasks, members of Group C, who
continued and survived the census date, have a mean score that
is higher than that of all other groups. Differences in mean
gains again reflect the time to death or withdrawal.
Tables 5 and 7 show the results when t values from the main
models for the CVL task and the Mill Hill B Vocabulary Test
are used to compare overall scores between all death or withdrawal groups throughout the study. The corrected threshold for
significance is p .001.
For the CVL task, Tables 4 and 5 show that overall average
scores for the D3 and D4 groups are significantly higher than
for the D1 and D2 groups and that the D2 group scores higher
than the D1 group, although this last item is not significant at
the corrected level of p .001. In other words, mean declines in
verbal learning scores accelerate as death approaches. There are
similar graded effects of dropout, with scores for the W3 group
being higher than those for the W2 or W1 group and scores for
the W2 group being higher than those for W1. The D1 and the
W1 groups do not differ, suggesting that over this brief interval
the effects of impending withdrawal are as severe as those of
impending death. For the Mill Hill B Vocabulary Test, Table 6
shows that the mean differences follow similar patterns. Even
scores on a test of crystallized intelligence, production vocabulary, fall as death approaches. There are no differences between
those groups who drop out and then die shortly thereafter and
those who die without first dropping out.

GENERAL DISCUSSION
These analyses replicate the main findings of Rabbitt and
colleagues (2005) for AH4-1 intelligence test scores on a much

1982
1983
1984
1986
1987
1988
1990
1991
1992
Pattern
D1
D2
D3
D4
WD1
WD2
WD3
W1
W2
W3
Interaction
Age 3 T2
Age 3 T3
Age 3 T4

Residual standard deviations: Levels 1 and 2

Level
2
2
1

Intercept
Age
Residual

SD

correlation with intercept

13.513
0.097
9.163

0.845

Note: CE coefficient; T1T4 Times 14; Oclass occupational class

(see text for details on the C designations); Oclass Na occupational class
not specified. For patterns, D death and W withdrawal (see text for particular designations).

larger sample, extend and compare these to the CVL task and
production vocabulary (Mill Hill B Vocabulary Test), and
examine the time courses of changes in greater detail, concluding that mean scores on all analyzed cognitive tests decline
according to nearness of death or dropout.

Time Courses of Effects of Death and Dropout

There are significant declines up to 8 years preceding death,
but participants who survived for more than 8 years after their

DEATH, DROPOUT, AND COGNITIVE CHANGE

first assessment (T1) performed almost as well as those who

survived the final testing session. Because there is no Age 3
Death or Age 3 Dropout interaction, there is no evidence, in
this study of atypically healthy participants, of change in the
sizes or the time courses of the cognitive effects of approaching
death between the ages of 49 and 93 years.
A new finding is that amounts of decline preceding dropout
closely resemble those before death in all tests. Because only
some participants gave reasons for dropout, we could not
compare the decline of individuals who dropped out for different reasons. However, some of us (Rabbitt et al., 1994) found
that the group of dropouts who gave reasons for withdrawal
included relatively healthy and able people who withdrew for
reasons such as employment. Because these analyses include
such robust dropouts, the actual similarity of declines preceding dropout caused by illness or frailty to those preceding
death must be even closer than these analyses suggest. The
finding that amounts of declines preceding impending death
and dropout are so strikingly similar suggests that they are both
caused by declining health. On this interpretation, death and
dropout are empirically useful, but rough and indirect, markers
for worsening health. The effects of approach to death, on their
own, tell us little about the functional causes of cognitive
decline. To learn more we must study precisely how particular
terminal illnesses affect our brains and central nervous systems.

Are Different Cognitive Abilities Differentially

Affected by Death and Dropout?
Declines with age for production vocabulary (Mill Hill B
Vocabulary Test) are much less (and nonsignificant) than those
for intelligence and cumulative verbal learning. Vocabulary,
a skill crystallized because it is acquired early in life and
maintained by continual practice into old age (Horn, 1987;
Horn, Donaldson, & Engstrom, 1981), is not only relatively
resistant to normal aging but also to the pathologies that
accompany aging and that may terminate in death. The effect of
death or withdrawal group is also weaker in Mill Hill B
Vocabulary Test.
This result differs from previous findings that vocabulary test
scores may be especially sensitive to approaching death (Berg,
1987; Birren, 1965; Siegler, McCarty, & Logue, 1982). It is
noteworthy that these previous studies were cross-sectional and
used relatively elderly samples and brief census periods of
24 months or less. It is therefore possible that participants in
these studies were quite near to their deaths but that this could
not be observed because the census periods were so short. It is
therefore plausible that declines in vocabulary only become
marked very shortly before death and so were less pronounced
in the present study, in which the times of measurement were
4 years or longer. It seems likely that individuals who are, as
yet, only experiencing declines in so-called fluid abilities are
relatively further from death than those whose terminal pathologies have become severe enough to affect even their crystallized abilities. Accepting this context declines in vocabulary
are particularly sensitive markers of approaching death.

Effects of Gender and Demographic Variables

Women perform better than men on the CVL task and on the
Mill Hill B Vocabulary Test (the latter at a marginal significance level), and men perform better than women on the AH4-1

P277

Table 7. Specific Comparisons of MHB Scores Between

Death and Dropout Groups
MHB
C
D1
D2
D3
D4
W1
W2
W3
WD1
WD2
WD3

D1

D2

D3 D4

.0001 .0001

W1

W2

W3

.0001 .0001

WD1

WD2

WD3

.0001 .0001
.001

.0002 .0001
.001

.0009

.0004

Note: Values of t statistics are derived from the main model. After correction, p , .001 is taken as the threshold for significance. MHB Mill Hill B
Vocabulary Test; C control group; D death and W withdrawal (see text
for particular designations).

intelligence test. On the AH4-1 task, although women score

lower on average than men do, they also decline more slowly as
they grow older. This is informative because the effects of
differences in death and dropout have been taken into consideration, as the slower declines of women of advancing age
cannot be attributed simply to their greater longevity or lower
dropout rate. This raises a fruitful topic for further investigation.
It is possible that examining the differences in the nature of
terminal pathologies in men and women would be informative.
A different point is that significant gender advantages in
baseline scores are opposite on different cognitive tests. This
means that if we do not take into account the typical progressive, age-related increases in the proportion of women to men
during longitudinal studies, then our speculations as to which
cognitive abilities are most and least sensitive to age and
pathology will be insecure.
Recruitment cohorts differ markedly in cognitive test scores
and, as exploratory analyses show, also in average levels of
SEA. Because mortality is strongly associated with level of
general intellectual ability (Hart et al., 2003) and with SEA
(e.g., Nagi & Stockwell, 1973), it is essential that researchers
check for differences between recruitment cohorts when
analyzing the effects of impending death on cognitive
performance.

Methodological Issues
Most studies have found that particular pathologies that
become common in later life, such as diabetes (e.g., Bent,
Rabbitt, & Metcalf, 2000), hypertension and other cardiovascular problems (Fahlander et al., 2000; Hertzog, Schaie, &
Gribbin, 1978; Lopez et al., 2003), respiratory problems
(Holland & Rabbitt, 1991), and undifferentiated health
problems (McInnes & Rabbitt, 1997), have significant but
surprisingly small effects on cognitive performance. Most
studies have compared patients and healthy controls only at
a single time point, and the few longitudinal studies have
ignored death and dropout. The current analyses suggest that if
individuals in so-called patient groups who die or withdraw
are not included in comparisons against healthy controls, then
the true effects of pathologies must be severely underestimated.
If dropouts and deaths are excluded from analyses then the

P278

RABBITT ET AL.

comparisons will involve only patients whose conditions are, as

yet, relatively mild. The effects of terminal declines and of early
stages of pathologies can only be compared if all deaths and
dropouts are logged and taken into account.
CORRESPONDENCE
Address correspondence to P. M. A. Rabbitt, Department of Experimental Psychology, University of Oxford, England. E-mail: Patrick.
Rabbitt@psy.ox.ac.uk
REFERENCES
Bent, N., Rabbitt, P. M. A., & Metcalf, D. (2000). Diabetes mellitus and the
rate of cognitive change. British Journal of Clinical Psychology, 39,
349362.
Berg, S. (1987). Intelligence and terminal decline. In G. Maddox and E. W.
Busse (Eds.), Aging: The universal human experience (pp. 234259).
New York: Springer.
Birren, J. E. (1965). Age changes in speed of behavior: Its critical nature
and physiological correlates. In A. T. Welford and J. E. Birren (Eds.),
Behavior, aging and the nervous system (pp. 171189). Springfield, IL:
Charles C Thomas.
Fahlander, K., Wahlin, A., Grut, M., Forsell, Y., Hill, R. D., Winblad, B.,
et al. (2000). The relationship between signs of cardiovascular
deficiency and cognitive performance in old age. Journal of
Gerontology: Psychological Sciences, 55B, P259P265.
Hart, C. L., Taylor, M. D., Davey-Smith, G., Whalley, L. J., Starr, J. M.,
Hole, D. J., et al. (2003). Childhood IQ, social class, deprivation and
their relationships with mortality risk in later life: Prospective
observational study linking the Scottish Mental Survey 1932 and the
Midspan Studies. Psychosomatic Medicine, 65, 877883.
Hertzog, C., Schaie, K. W., & Gribbin, K. (1978). Cardiovascular disease
and changes in intellectual functioning from middle to old age.
Gerontology, 33, 872883.
Holland, C. M., & Rabbitt, P. M. A. (1991). The course and causes of
cognitive change with advancing age. Reviews in Clinical Gerontology,
1, 8196.
Horn, J. L. (1987). A context for understanding information processing
studies of human abilities. In P. A. Vernon (Ed.), Speed of information
processing and intelligence (pp. 201238). Norwood, NJ: Ablex.
Horn, J. L., Donaldson, G., & Engstrom, R. (1981). Application, memory
and fluid intelligence decline in adulthood. Research on Aging, 3, 3384.
Houx, P. J. (1991). Rigorous health screening reduces age effects on
a memory scanning task. Brain and Cognition, 15, 246260.
Johansson, B., & Berg, S. (1989). The robustness of the terminal decline
phenomenon: Longitudinal data from the digit-span memory test.
Journal of Gerontology: Psychological Sciences, 44B, P184P186.
Kleemeier, R. W. (1962). Intellectual changes in the senium. Proceedings
of the Social Statistics Section of the American Statistical Association,
23, 290295.
Lachman, R., Lachman, J. L., & Taylor, D. W. (1982). Reallocation of
mental resources over the productive lifespan: Assumptions and task
analyses. In F. I. M. Craik & S. Trehub, (Eds.), Aging and the cognitive
process (pp 304350). New York: Plenum Press.
Lieberman, M. A. (1965). Psychological correlates of impending death:
Some preliminary observations. Gerontology, 20, 181190.
Lopez, O. L., Jagust, W. J., Dulberg, C., Becker, J. T., De Kosky, S. T.,
Fitzpatrick, A., et al. (2003). Risk factors for mild cognitive impairment
in the Cardiovascular Health Study cognition study. Archives of
Neurology, 60, 13941399.
McInnes, L., & Rabbitt, P. M. (1997). The relationship between functional
ability and cognitive ability among elderly people. In Facts and
research in gerontology (pp. 3445). Paris: Sardi.
Nagi, M. H., & Stockwell, E. G. (1973). Socioeconomic differentials in
mortality by cause of death. Health Services Reports, 88, 449456.

Office of Population Censuses and Surveys. (1980). Classification of

occupations 1980. London, England: Her Majestys Stationery Office.
Palmore, E., & Cleveland, W. (1976). Aging, terminal decline and terminal
drop. Gerontology, 31, 7681.
Pincus, T., Callaghan, L. F., & Birkhauser, R. V. (1987). Most chronic
diseases are reported most frequently by individuals with fewer than 12
years formal education in the age 1864 US population. Journal of
Chronic Disorders, 40, 865874.
Rabbitt, P. (2002). Cognitive gerontology. In H. Pashler & J. Wixted (Eds.),
Stevens handbook of experimental psychology (Vol. 4., pp. 203254).
New York: Wiley.
Rabbitt, P., Diggle, P., Holland, F., & McInnes, L. (2004). Practice and
drop-out effects during a 17-year longitudinal study of cognitive aging.
Journal of Gerontology: Psychological Sciences, 59B, P84P97.
Rabbitt, P., Diggle, P., Smith, D., Holland, F., & McInnes, L. (2001).
Identifying and separating the effects of practice and of cognitive
ageing during a large longitudinal study of elderly community residents.
Neuropsychologia, 39, 532543.
Rabbitt, P., Lunn, M., & Wong, D. (2005). Neglect of dropout
underestimates effects of death in longitudinal studies. Journal of
Gerontology: Psychological Sciences, 60B, P106P109.
Rabbitt, P., Watson, P., Donlan, C., Bent, N., & McInnes, L. (1994).
Subject attrition in a longitudinal study of cognitive performance in
community resident elderly people. In B. J. Vellas, J. L. Albarende, &
P. J. Garry (Eds.), Facts and research in gerontology: Epidemiology
and aging (pp. 2934). Paris: Serdi.
Rabbitt, P., Watson, P., Donlan, C., McInnes, L., Horan, M., Pendleton, N.,
et al. (2002). Effects of death within 11 years on cognitive performance
in old age. Psychology and Aging, 17, 114.
Rabbitt, P. M. A., Diggle, P., Holland, F., McInnes, L., Bent, N., Abson,
V., et al. (2004). The University of Manchester longitudinal study of
cognition in normal healthy old age, 1983 through 2003. Aging,
Neuropsychology and Cognition, 11, 245279.
Raven, J. C. (1965). The Mill Hill Vocabulary Scale. London: Lewis.
Reimanis, G., & Green, R. (1971). Immanence of death and intellectual
decrement in the aging. Developmental Psychology, 5, 270272.
Riegel, K. F., & Riegel, R. M. (1972). Development, drop and death.
Developmental Psychology, 6, 306319.
Riegel, K. F., Riegel, R. M., & Meyer, G. (1967). A study of the drop-out
rates in longitudinal research on aging and the prediction of death.
Journal of Personality and Social Psychology, 5, 342348.
Roth, M. (1986). The association of clinical and neuropsychological
findings and its bearing on the classification and etiology of
Alzheimers disease. British Medical Bulletin, 42, 4250.
Schaie, K. W., Labouvie, G. V. & Barrett, T. J. (1973). Selective attrition
effects in a fourteen-year study of adult intelligence. Gerontology, 28,
328334.
Siegler, I. C., & Botwinick, J. (1979). A long-term longitudinal study of
intellectual ability of older adults: The matter of selective subject
attrition. Gerontology, 34, 242245.
Siegler, I. C., McCarty, S. M., & Logue, P. E. (1982). Wechsler Memory
Scale scores: Selective attrition and distance from death. Gerontology,
37, 176181.
Small, B. J., Fratiglioni, L., von Strauss, E., & Backman, L. (2003).
Terminal decline and cognitive performance in very old age.
Psychology and Aging, 18, 193202.
Snowden, D. A., Ostwald, S. K., Kane, R. E. L., & Keenan, N. L. (1989).
Years of life with good and poor mental and physical functions in the
elderly. Journal of Clinical Epidemiology, 42, 10551066.
Van Boxtel, M. P. J., Buntink, F., Houx, P. J., Metsemakers, J. F. M.,
Knottmerus, A., & Jolles, J. (1998). The relation between morbidity
and cognitive performance in a normal aging population. Journal of
Gerontology: Medical Sciences, 55A, M147M154.
Received June 28, 2005
Accepted January 30, 2008
Decision Editor: Thomas M. Hess, PhD