Home
HOME ARCHIVE SEARCH TABLE OF CONTENTS
This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Services
Right arrow Download to citation manager
Citing Articles
Right arrow Citing Articles via HighWire
PubMed
Right arrow PubMed Citation
The Journals of Gerontology Series B: Psychological Sciences and Social Sciences 60:P106-P109 (2005)
© 2005 The Gerontological Society of America

RESEARCH ARTICLE

Neglect of Dropout Underestimates Effects of Death in Longitudinal Studies

Patrick Rabbitt1,, Mary Lunn2 and Danny Wong2

1 Age and Cognitive Performance Research Centre, University of Manchester, England.
2 Department of Statistics, University of Oxford, England.

Address correspondence to Patrick Rabbitt, University of Manchester, Age and Cognitive Performance Research Centre, Zochonis Building, Brook St., Manchester, M13, 9PL, United Kingdom. E-mail: rabbitt{at}psy.man.ac.uk


   Abstract
TOP
 Abstract
Methods
 Results
 Discussion
 References
 
Investigations of terminal declines in mental abilities have assessed cognitive performance at a single time point and retrospectively compared survivors and decedents at a single later census date. Neglect of outcomes other than death, such as dropout, causes a loss of information on the relative frailty of survivors and deceased persons before the census date and on incidence of mortality and frailty among survivors after the census date. This discards information on differences in health status between younger and older survivors. The Heim AH4-1 intelligence test was given to 4,228 people between the ages of 42 and 92 years, and both deaths and dropouts were logged during three successive census periods during the subsequent 11 years. Within and across census periods, effects of impending death and dropout did not differ, decreasing with time from initial assessment. Thus the effects of terminal decline, or indeed of any other variable affecting cognitive performance, are miscalculated if dropout is ignored.

Nearly all studies of cognitive changes preceding death have assessed participants only once and then recorded deaths, but no other events, before a single census date, and retrospectively compared performance of survivors and decedents. Typical findings are that individuals who die within 1 to 11 years of assessment perform less well than survivors (e.g., Rabbitt et al., 2002Go; Small, Fratiglioni., von Strauss, & Backman, 2003Go). This methodology can lead to misleading conclusions, because it neglects useful information about subgroups of survivors between whom the effects of mortality are compared. For example, Riegel and Riegel (1972)Go reported that differences in ability between survivors and decedents were significantly larger for young adults than for elderly adults because "death strikes more randomly and psychological differences between survivors and non-survivors are less marked" (p. 309). A simpler explanation is that, because young survivors are typically healthy and survive the census date for much longer, differences between young decedents and survivors are correspondingly larger than those between old decedents and survivors (see discussion in Rabbitt et al., 2002Go).

To examine how use of a single census point and neglect of events other than death affect estimates of amounts and rates of terminal declines, we find that it is necessary to use several successive census periods and to log not only deaths but also other markers of frailty and of biological changes. Dropout is a strong marker for pathology and frailty because most individuals who withdraw from longitudinal studies do so because of poor health and mobility (see, e.g., Rabbitt, Watson, Donlan, Bent, & McInnes, 1994Go). To illustrate the extent to which estimates of the cognitive status of decedents are affected by the incidence of frailty and consequent dropout, we compared subsets of individuals who died or dropped out before, after, and during three successive periods between assessment and census.


   METHODS
TOP
 Abstract
 Methods
Results
 Discussion
 References
 
Participants
Between 1982 and 1988, 2,190 healthy, independent citizens of Greater Manchester and 2,038 citizens of Newcastle on Tyne independently traveled to university laboratories to complete the Heim (1970)Go AH4-1 test of general fluid intelligence in quiet, well-lit rooms in groups of 5 to 15, supervised by two experimenters. For complete details, see Rabbitt, McInnes, and colleagues (2004)Go. A computer search of death certificates by Her Majesty's Registrar General's office ensured that census of deaths was complete and that exact dates of 576 deaths between assessment and February 4, 1994 were logged.

Design and Procedure
AH4-1 scores are the numbers out of 64 problems involving mental arithmetic, number series, and verbal comparisons that were correctly answered within 10 min. Table 1 shows means and SDs of AH4-1 scores by age band, gender, city of residence (Manchester or Newcastle), and level of socioeconomic advantage recorded in terms of the UK National Register of Occupational Categories (Office of Population Censuses and Surveys, 1980Go). These categories are good proxies for years and level of education and of income. Table 1 shows that, on February 1, 1994, there were 576 participants (13.6 %) who had died, 1,629 (38.5 %) who had dropped out, and 2,023 (47.9 %) who had survived and continued.


View this table:
[in this window]
[in a new window]
  		Table 1. Selected Demographics of Sample With Mean AH4-1 Scores.

 

   RESULTS
TOP
 Abstract
 Methods
 Results
Discussion
 References
 
Categorization of Death and Dropout Groups
We compared AH4-1 scores at initial assessment (T1) between participants who survived beyond and those who died or dropped out within each of three successive subsequent time periods before the February 1, 1994 census point. We identified dropouts by failures to respond to invitations to further testing sessions at approximately 4 years (T2), 8 years (T3), and 11 years (T4, which was 1994) after initial assessment. We did not count individuals who missed a particular testing session (e.g., between T1 and T2) but returned to take a later session (e.g., at T3 or T4) as dropouts; instead, we categorized them in terms of their history after the last session that they attended. We also censored deaths within each of these three successive periods. We compared AH4-1 scores at T1 between subgroups who fell into each of nine different possible temporal patterns of death (D) and dropout (W) patterns as follows:
C, Did not drop out and survived beyond 1994; n = 2,023 (baseline group)
D1, Died between T1 and T2; n = 234
D2, Died between T2 and T3; n = 127
D3, Died after T3 but before 1994; n =16
WD1, Recorded as dropout before T2 and died before 1994; n =133
WD2, Recorded as dropout between T2 and T3 and died before 1994; n = 66
WD3, Recorded as dropout after T3 and then died before 1994; n = 0
W1, Recorded as dropout before T2 and survived beyond 1994; n = 1,133
W2, Recorded as dropout between T2 and T3 and survived beyond 1994; n = 457
W3, Recorded as dropout after T3 and survived beyond 1994; n = 39

Patterns of subsequent death and dropout markedly affect AH4-1 scores. On a simple comparison, the range of 17.44 points between groups C and WD1 is much larger than that between cities (1.96 points) or genders (2.74 points) and approaches the range for age (23.44 points) and for socioeconomic advantage (28.39 points).

Regression Analysis
Table 2 shows the results of a multivariate regression analysis in which we compare the effects of patterns of death and dropout after we have considered effects associated with age, city, gender, and socioeconomic advantage. Main effects of age, socioeconomic advantage, and gender are all significant, with a single, unexplained, significant interaction between city and gender (p <.05). Consequently, we retain city in the model. The quadratic term for age is significant, confirming findings by Rabbitt, Diggle, Holland, and McInnes (2004)Go that, within this same sample, declines in mean AH4-1 scores accelerate after the individual reaches the age of 70. There are no significant interactions involving death–dropout pattern groups. Table 3 shows specific comparisons of AH4-1 scores between career-pattern groups made with an analysis of variance (F test for linear combination).


View this table:
[in this window]
[in a new window]
  		Table 2. Effects of Demographic Variables and of Differences Between Temporal Patterns of Death and Dropout on AH4-1 Intelligence Test Scores.

 

View this table:
[in this window]
[in a new window]
  		Table 3. Specific Comparisons Between AH4-1 Scores for Death and Dropout Groups.

 
The effects of subsequent death and dropout on scores at initial assessment are statistically indistinguishable. Group D1, who died before T2, did not differ either from Group WD1, who dropped out before T2 and then died before the 1994 census, or from Group W1, who dropped out before T2 but survived beyond census. Similarly, Group D2, who did not drop out but died between T2 and T3, did not differ from group WD2, who dropped out before T3 and died before census, or from Group W2, who dropped out between T2 and T3 but survived beyond census. Further, Group D3, who died after T3 but before T4, did not differ from Group W3, who dropped out between T3 and census but survived thereafter. However, note that numbers in these latter groups are small. Assessment scores are lower if either death or dropout occurs within 8 years. Within this period, the effects of death and dropout reduce with their temporal distance from assessment. Because scores for groups D3 and W3 equal those for the baseline group of census survivors, C, there is no evidence that assessment scores are affected by death or dropout occurring more than 8 years later. It follows that both the size and the time course of the effects of impending death and dropout are statistically indistinguishable. The analysis also shows that the effect sizes of both deaths and dropouts will be miscalculated if incidence of mortality and frailty among census survivors is neglected. This is illustrated by the finding that mean scores for group D1, who died between T1 and T2, are significantly lower than for group C, who survive and continue beyond 1994, but are not significantly lower than for any other groups who die after T2 or withdraw before 1994. Similarly group D2, who died between T2 and T3, do not differ from groups who survived T3 but thereafter dropped out or died. However, group D2 did score lower than group W1, who withdrew between T1 and T2 and survived, or than group WD1, who withdrew between T1 and T2 and later died before 1994.


   DISCUSSION
TOP
 Abstract
 Methods
 Results
 Discussion
References
 
Although this analysis examines performance on only a single cognitive test, the results are general because tests of fluid intelligence, and specifically the AH4-1, robustly predict performance on most other tests of mental abilities (Deary, 2000Go; Rabbitt, Diggle, et al., 2004Go).

Declines associated with dropout have been widely discussed (e.g., Little, Lindenberger, & Maier, 2000Go), but to our knowledge the time course of changes preceding dropout have not previously been reported. Our new findings are that the effects of impending dropout and of death are statistically indistinguishable in size and time scale and that performance at initial assessment is affected by both death and dropout up to 8 years, but possibly not between 8 and 11 years later. Note that these estimations of the sizes and time scales of dropout effects must be conservative, because Rabbitt and colleagues (1994)Go found that a subset of dropouts are robust individuals who withdraw for positive reasons such as taking up employment, marriage, or changes of residence. These individuals, on average, have AH4-scores higher than the general sample mean. Because in this study it was not possible to identify all such "positive dropouts," their inclusion means that the true effects of dropout caused by illness or increasing frailty are greater than these analyses indicate. The indistinguishable effects of impending dropout and mortality raise the question whether it makes theoretical sense to regard death as a uniquely informative outcome in longitudinal studies of age-related cognitive change. Death is certainly a strong and unambiguous marker for the presence of pathologies, but the equivalence of death and dropout effects suggests that the severity, the time course, and no doubt the particular nature of the pathology resulting in death are of greater interest than the date of its conclusion. Future resources may be more usefully expended in studying the cumulative effects, the time course, and the nature of pathology and frailty as age advances (e.g., Hertzog, Schaie, & Gribbin, 1978Go, Houx, 1991Go; Holland and Rabbitt, 1991Go; Van Boxtel et al., 1998Go).

These findings show that, if comparisons of decedents and survivors use only a single census point and neglect other outcomes such as dropout, the true extents and rates of terminal declines are underestimated. Within samples of frail elderly individuals, absolute differences in survival duration between those who die before and those who die soon after an arbitrary census date are likely to be so small that their effects on scores at earlier assessment are imperceptible. This methodological artefact explains why many studies based on samples of very elderly and hospitalized individuals have found no evidence for ter-minal declines in cognition. It also explains findings such as those by Riegel and Riegel (1972)Go that, when death is the only outcome considered, "terminal decline" effects appear larger in younger than in older samples. Finally, this analysis extends earlier discussions of the effects of informative dropout by Rabbitt, Diggle, and colleagues (2004)Go with direct evidence that, if dropout is ignored, then the true effects on rates of cognitive change of mortality, pathology, sociodemographic factors, and indeed of so-called normal or usual aging are miscalculated.


   Footnotes
 
Decision Editor: Thomas M. Hess, PhD

Received for publication March 18, 2004. Accepted for publication September 23, 2004.


   References
TOP
 Abstract
 Methods
 Results
 Discussion
 References
 




This article has been cited by other articles:


Home page
 Age AgeingHome page
R. E. Hubbard, M. S. O'Mahony, and K. W. Woodhouse
Characterising frailty in the clinical setting--a comparison of different approaches
Age Ageing, January 1, 2009; 38(1): 115 - 119.
[Full Text] [PDF]

Home page
 Journals of Gerontology Series B: Psychological Sciences and Social ScienceHome page
P. Rabbitt, M. Lunn, and D. Wong
Death, Dropout, and Longitudinal Measurements of Cognitive Change in Old Age
J. Gerontol. B. Psychol. Sci. Soc. Sci., September 1, 2008; 63(5): P271 - P278.
[Abstract] [Full Text] [PDF]

Home page
 Journals of Gerontology Series B: Psychological Sciences and Social ScienceHome page
P. Rabbitt, M. Lunn, D. Wong, and M. Cobain
Sudden Declines in Intelligence in Old Age Predict Death and Dropout From Longitudinal Studies
J. Gerontol. B. Psychol. Sci. Soc. Sci., July 1, 2008; 63(4): P205 - P211.
[Abstract] [Full Text] [PDF]

Home page
 Journals of Gerontology Series B: Psychological Sciences and Social ScienceHome page
P. Rabbitt, M. Lunn, D. Wong, and M. Cobain
Age and Ability Affect Practice Gains in Longitudinal Studies of Cognitive Change
J. Gerontol. B. Psychol. Sci. Soc. Sci., July 1, 2008; 63(4): P235 - P240.
[Abstract] [Full Text] [PDF]

Home page
 Age AgeingHome page
K. Rockwood, D. Jones, Y. Wang, D. Carver, and A. Mitnitski
Failure to complete performance-based measures is associated with poor health status and an increased risk of death
Age Ageing, March 1, 2007; 36(2): 225 - 228.
[Full Text] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF)
Services
Right arrow Download to citation manager
Citing Articles
Right arrow Citing Articles via HighWire
PubMed
Right arrow PubMed Citation

HOME ARCHIVE SEARCH TABLE OF CONTENTS