| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
| ||||||||||||||||||||||||
RESEARCH ARTICLE |
a University of Toulouse-le-Mirail, France
Jean Claude Marquié, Laboratoire Travail \|[amp ]\| Cognition, UMR 5551 du CNRS, MDR, University of Toulouse-le-Mirail, 5 all\|[eacute]\|es A. Machado, 31058 Toulouse Cedex, France E-mail: marquie{at}univ-tlse2.fr.
Toni C. Antonucci, PhD
 |
Abstract |
|---|
 TOP Abstract MethodsResultsDiscussionAppendix ENDIXReferences |
|---|
IN line with the popular belief that with increased age people are less confident and more cautious, it has long been contended that older people show a more conservative response bias than younger ones. In perceptual tasks, for instance, even at the psychophysical assessment level, several early studies have suggested that part of the observed age-related change goes beyond purely sensory considerations and can be ascribed to changes in decision criteria (see Welford 1958
). Increased cautiousness with age has generally been inferred from the observation that, compared with their younger counterparts, older participants take more time than necessary to respond to difficult signals in visual discrimination tasks (Botwinick, Brinley, and Robbin 1958) or are more reluctant to accept items as perfect in simulated industrial inspection tasks (Belbin and Shimmin 1964). In other tasks, greater cautiousness in older participants has been deduced from their preference, when in doubt, not to respond at all rather than to give an incorrect answer (Korchin and Basowitz 1957), their choice of the least risky alternative from those presented on a list (Wallach and Kogan 1961), or their shift in emphasis from speed to accuracy (Botwinick 1978; Hertzog and Vernon 1993; Rabbitt 1979; Welford 1958, Welford 1977).
For years the so-called choice dilemma instrument (Wallach and Kogan 1961)—in which participants choose between alternatives involving various degrees of risk—was the sole methodology used specifically to address age differences in risk-taking behavior. But later criticisms of this measuring device led most authors to stop using it (see Okun 1976). More recently, methods derived from signal detection theory (SDT) have been considered to offer powerful tools for investigating this issue. SDT, first developed by Swets, Tanner, and Birdsall 1961 and D. M. Green and Swets 1966 for detection tasks, and then extended to other tasks (Swets 1996), provides a performance assessment framework in which performance components related to the sensitivity of a psychological function can be separated from those related to decision-making processes, the two components being theoretically independent. The decision criterion, which reflects the participant's readiness or reluctance to give a certain response under uncertainty, is based on the relative proportion of two kinds of errors: omissions and false alarms (commissions).
Several researchers have used the SDT approach to study age differences in decision criteria on a variety of tasks, including auditory detection (Baron & Belongia Le Breck, 1987; Craik 1969; Marshall 1991; Potash and Jones 1977; Rees and Botwinick 1980); visual discrimination (Baracat and Marquie 1992); visual inspection (Craik 1969, who reanalyzed data from Belbin and Shimmin 1964); visuo-spatial perception (Hutman and Sekuler 1980; Morrison and Reilly 1986); pain perception (Clark and Mehl 1971; Harkins and Chapman 1976, Harkins and Chapman 1977); weight discrimination (Baron & Belongia Le Breck, 1987; Danziger and Botwinick 1980; Watson, Turpenoff, Kelly, and Botwinick 1979); vigilance (Neal and Pearson 1966; Tune 1966); and recognition of words, letter-number combinations, faces, prose, or speech (Baron & Belongia Le Breck, 1987; Baron and Surdy 1990; Ferris, Crook, Clark, McCarthy, and Rae 1980; Gordon and Clark 1974a, Gordon and Clark 1974b; Gordon-Salant 1986; Harkins, Chapman, and Eisdorfer 1979; Poon and Fozard 1980; Yanz and Anderson 1984).
Even with this more specific or purer method, however, these studies have failed to obtain consistent results concerning age effects on response bias. Of the 26 studies involving decision criterion assessment we reviewed, only 8 revealed a more conservative response bias in older participants: the 3 pain perception studies, 3 of the 5 auditory perception studies (Craik 1969; Potash and Jones 1977; Rees and Botwinick 1980), the Belbin and Shimmin study (1964, data reanalyzed by Craik 1969), and the Poon and Fozard 1980 word recognition study. In the others, what appeared was sometimes the opposite age effect (Baracat and Marquie 1992; Gordon-Salant 1986; Neal and Pearson 1966; Tune 1966); mitigated results depending on the trial, type of response measure, or way of estimating the decision criterion (Danziger and Botwinick 1980, Gordon and Clark 1974b; Harkins et al. 1979); or most often no age differences at all. This makes it currently quite difficult to draw any definite conclusions about whether there is a response-bias difference between older and younger participants. Our first goal in the present study was to obtain further information about whether increasing age is associated with changes in response bias and, especially, with more conservative decision criteria. We examined this issue in a large sample by means of a memory recognition task.
One possible reason for the inconsistencies mentioned previously is the small size of the samples used (usually less than 50 participants, all experimental groups taken together) and a lack of adequate control in those samples of variables likely to play a role in such processes. For instance, some studies have worked on either men or women only, whereas others have included both men and women in their samples but in unequal proportions. Given that earlier studies have provided some evidence of sex differences in verbal and spatial abilities (Halpern 1986; Willis and Schaie 1988) and, as Botwinick (1973) pointed out, possible relationships between cautiousness and verbal ability, sex effects on decision processes cannot be ruled out. Moreover, Wallach and Kogan 1961 and Gordon and Clark 1974b found direct evidence of sex differences in decision criteria. Likewise, in most studies on response bias, the participant's educational level was not specified or its range was limited. Because education has often been shown to influence cognitive processes strongly, and given that a vast majority of studies on cognitive aging are cross-sectional, one can assume that age differences in decision making might reflect cohort differences in schooling or other related experiences rather than the aging process per se (e.g., R. F. Green 1969; Schaie and Hertzog 1983). Our view is that sex and education should be controlled, at least until substantial evidence has been provided of their lack of effect on response bias. Thus, we used a large sample made up of members of both sexes with the widest possible range of educational levels, in order to examine the extent to which these variables interact with the effect of age on response bias. By using an index derived from SDT, we expected that if age effects on response bias have a large degree of generality, then they will be apparent in the present recognition task. An alternative hypothesis would be either that age differences in decision criteria are specific to certain types of tasks or that previously observed age differences are due to sampling artifacts, especially the underlying effects of educational level or sex.
|
Methods |
|---|
|
TOPAbstractMethodsResultsDiscussionAppendix ENDIXReferences |
|---|
The breakdown of the sample into age groups was 27.8% for the 32-year-olds, 30.1% for the 42-year-olds, 27.6% for the 52-year-olds, and 14.5% for the 62-year-olds. In the oldest age group, 82.4% had already retired. In the sample as a whole, there were equal proportions of men (51.6%) and women (48.4%). Both sexes were also well represented in each age group, although their proportions were not strictly equal (48.2% and 51.8%, 50.7% and 49.3%, 51.9% and 48.1%, and 59.7% and 40.3% for men and women, for the 32-, 42, 52-, and 62-year-olds, respectively; 
2 significant at p < .01).
Individuals from all educational levels were found in the sample, but they were grouped for the present study into five classes so that the age and sex variables could be represented in equal proportions to the greatest extent possible. The educational classes, defined by the number of years of schooling, were EL1 (
19.6%), EL2 (8 and 9 years, 25.2%), EL3 (10 and 11 years, 14.5%), EL4 (12 and 13 years, 15.5%), and EL5 (
, 25.1%).
Current health status was self-assessed on a 10-point Likert-type rating scale ranging from 0 ("very poor") to 10 ("excellent"). Mean scores of 7.45 (
), 7.12 (
), 6.91 (
), and 6.96 (
) were obtained for the 32-, 42-, 52-, and 62-year-olds, respectively, indicating small but significant differences between age groups, 
, p < .0001. The administration of the Digit-Symbol subtest of the Wechsler Adult Intelligence Scale gave rise to the typical age-related result pattern: the younger the participant, the higher the Digit-Symbol scores (32-year-olds, M = 57.33, SD = 12.96; 42-year-olds, M = 53.96, SD = 12.71; 52-year-olds, M = 46.81, 
; 62-year-olds, 
; 
, p < .0001).
Material
For the learning phase, the test material was made up of three sets (A, B, and C) of 16 frequent, two-syllable, phonetically unambiguous common nouns, each set combining an equal number of low- and high-imagery words (Hogenraad and Orianne 1981). For the recognition phase, the material consisted of three corresponding sets of 48 words made up of the 16 target words mixed in with 32 new words (or distractors) with the same characteristics as the targets. In each target and corresponding distractor set, words belonging to various categories (body parts, clothes, animals, fruit, etc.) were balanced.
Procedure
Participants first underwent a learning phase in which the words were loudly and distinctly pronounced at a rate of one per second. They were given three consecutive trials followed by immediate free recall. Then, after a 20-min period during which they completed the Digit-Symbol test and carried out two other cognitive tasks (completion tasks), they performed the memory recognition task. They were instructed to check as many of the 16 previously learned words as possible from a list on which learned words were mixed in with 32 new words, but to check previously learned words only. Because the study was longitudinal and we planned to repeat the measures twice within the next few years, each participant was assigned one of the three sets of words (A, B, or C). The task was self-paced.
|
Results |
|---|
|
TOPAbstractMethodsResultsDiscussionAppendix ENDIXReferences |
|---|
). Response bias was assessed with the B'H index (Grier 1971; Hodos 1970). Some of the advantages of this index are that it does not require the assumption of normality of the underlying "noise" and "signal" distributions and that it has a good validity level even with low false-alarm rates (see also Snodgrass and Corwin 1988). B'H values range from -1 to +1, with negative values indicating a greater relative proportion of false alarms over omissions, that is, a participant's greater tendency to consider a new item as belonging to the previously learned word list (risky criterion), and positive values indicating the opposite trend, that is, a greater relative proportion of omissions over false alarms (conservative criterion). The formulas for calculating A' and B'H are given in the Appendix.
Before examining differences in sensitivity and response bias, we checked to see whether the three word sets were equivalent in terms of memory demands and whether the participants in each set had the same characteristics on the variables of interest to this study. Participants were assigned to Sets A, B, and C in the following proportions: 51.1%, 26.7%, and 22.2%, respectively. Two one-way ANOVAs were conducted with Set as a factor (3 levels), one with A' and the other with B'H as the dependent variable. Neither revealed any significant differences between Sets A, B, and C on A', 
, p < .54, or on B'H, 
, p < .81. Moreover, age, education, and sex, which are the relevant variables in this study, turned out to be equally represented within each set (all 2s nonsignificant).
Sensitivity
We performed sensitivity analyses on the transformed data
 |
, p < .0001; education, 
, p < .0001; and sex, 
, p < .0001. The magnitude of the age (32 vs 62), education (EL1 vs EL5), and sex (M vs F) effects was greater for education (effect size, 
SD) than for age (
) and sex (
). The power (for 
) was maximal for age, education, and sex (
). No interaction between any factors could be found. In addition to the mean number of hits and false alarms, the resulting mean A' values are given in Table 1 , by age, education, and sex. The analyses indicated overall high memory sensitivity but slightly less sensitivity with increased age in discriminating between old and new words. The Bonferroni t test (at p < .05) revealed significant differences for all age-group pairs except for the 32–42 pair and the 52–62 pair. The significant effects of educational level and sex reflect higher sensitivity in more educated subjects and higher sensitivity in women than in men. The Bonferroni t test revealed that each educational class differed significantly from all others.
|
, but significant differences across the educational levels, 
, and sexes, 
. The last two results mean that the higher the level of education, the lower the B'H value, and that men exhibited stricter decision criteria than women.
|
. Increasing education was also linearly associated with decreasing bias, 
. The magnitude of the education (EL1 vs EL5) and sex (M vs F) differences was greater for education (
) than for sex (
). For age (32 vs 62), the effect size was d = .32. The power (for ) was maximal for both education and sex (). For age it was more moderate (
), suggesting some caution in concluding that there is no age effect. No interaction was found between age and sex or between education and sex, thus indicating that the sex difference was constant across age groups and levels of education. By contrast, a significant although rather weak Age x Education interaction was found, 
, indicating greater age differences in response bias in the more educated individuals (see Fig. 1).
|
, p < .001, whereas in the other less educated groups (EL1 to EL4) the effect of age was nonsignificant (all ps between .17 and .99). The age effect in the EL5 group showed a significant linear trend, 
, 
, whereas in the other EL groups no significant trend was found. For EL5, the Bonferroni t test indicated that the 32-year-olds differed significantly from both the 52-year-olds and the 62-year-olds. By looking at the Age x Education interaction in another way, that is, by examining the effect of education within each age group, other aspects of that interaction can be highlighted. Indeed, although the main effect of education was significant in every age group (all ps < .05), the effect sizes (EL1 vs EL5) were strong in the youngest group and then diminished with age, with 
, .62, .56, and .30 for the 32-, 42-, 52-, and 62-year-old groups, respectively. Bonferroni t tests revealed the following significant differences: for the 32-year-olds, between each of the first three EL groups and EL4, and between each of the first four EL groups and EL5; for the 42-year-olds, between each of the first three EL groups and both EL4 and EL5; for the 52-year-olds, between each of the first three EL groups and EL5; and for the 62-year-olds, between EL1 and EL4. Considering the data in this fashion emphasizes the fact that decision criteria were influenced less by the level of education in the older groups than in the younger ones and that education is a better predictor of individual differences than age.
Although the sensitivity and bias indexes are theoretically independent, previous studies (Danziger 1980; Danziger and Botwinick 1980) have drawn attention to a problem that can arise when participants who differ in sensitivity are compared on decision criterion measures. They have shown that confidence in interpretation and generalization of bias differences can be maximal only when similar levels of sensitivity are obtained. In the present study, differences in sensitivity were found as a function of age, education, and sex, and the correlation between A' and B'H was -.33 (). The correlations were higher in younger groups than in older ones (
, -.38, -.29, and -.24 for ages 32, 42, 52, and 62, respectively, all 
), with the following age-group pairs showing significant differences at 
: age 32–62, age 42–52, and age 42–62 (Fishers' r' test). Therefore, to be certain that the response bias results reported previously are independent of sensitivity differences, we matched participants' decision criteria for equivalent A' values. Because the slopes lacked homogeneity, as reflected by the age-related differences in the correlation coefficients between A' and B'H, an ANCOVA was not appropriate. So we created five classes from the A' distribution to allow for comparison of bias differences for several A' levels. A' values below .875 were excluded because they corresponded to too few subjects (10.1%) to make additional homogeneous classes. The five classes represented steps of 5% of the total A' range, or 0.025 units of the index. An ANOVA with Age (4), Education (5), and Sex (2) as factors and B'H as the dependent variable was computed in each A' class. The analyses confirmed the lack of a main age effect, because this effect was never observed no matter what the A' value was (all ps > .05). An exception was an interacting effect of age with education and sex for the highest A' class, 
. This interaction reflected stricter decision criteria with age for the most educated men but not women, for participants whose sensitivity was maximal. Education had an effect in only two of the five A' classes: the middle, 
, and highest, 
, ones. Sex effects were only found for the highest A' level, 
. Thus, after factoring out sensitivity, this confirmed the lack of clear age differences in response bias, as well as the effect of educational level, and to a lesser extent sex, even if the last two effects were not found for all A' levels.
|
Discussion |
|---|
|
TOPAbstractMethodsResultsDiscussionAppendix ENDIXReferences |
|---|
First, analysis of the differences in sensitivity revealed that a moderate but significant decrease in the ability to recognize old items was observed with advancing age. This result is consistent with previous research regarding age effects on the ability to discriminate between old and new items in recognition memory tasks, where techniques derived from signal detection theory were used. In the nine earlier papers we reviewed on this issue, seven reported lower sensitivity in older adults than in younger ones (Baron & Belongia Le Breck, 1987; Baron and Surdy 1990; Ferris et al. 1980; Gordon and Clark 1974a, Gordon and Clark 1974b; Harkins et al. 1979; Poon and Fozard 1980). This was true whatever type of material was used, which in most cases included words as in this study, and it was even true in wider age range samples than ours. Thus, measures that assess this memory ability independently of the participant's decision criteria, as well as more classical measures of the recognition rate (e.g., McCarthy, Ferris, Clark, and Crook 1981; Schonfield and Robertson 1966; Smith 1975; see also Craik and Jennings 1992), support the view that older people have poorer recognition memory than younger ones. We also found that sensitivity was higher in women than in men, but the largest differences were obtained between lower and higher levels of education, the latter being associated with greater sensitivity. As the results obtained by Poon and Fozard 1980 suggest, the higher verbal aptitude usually observed in participants with more schooling might partly explain this education-linked effect.
With regard to decision criteria, both education and sex had significant effects, with more risky criteria being found in more educated individuals compared with less educated ones, and in women compared with men. However, contrary to what has often been suggested, we found that older people did not differ reliably from younger ones in spite of a steady mean increase with age toward stricter criteria. The only reliable age effect was observed for the most educated group: the younger the participants, the more risky they were, with the B'H values of the youngest and most educated subjects reaching lower levels than in any other age and educational group. Even after factoring out sensitivity, we found no overall age effect on response bias. Moreover, our data showed that in participants whose sensitivity was maximal, the stricter decision criteria observed with age at the highest educational level applied to men but not to women.
Because of the restricted age range in the present study, and the fact that the power to detect the small age effect size was only moderate, the lack of an age effect on response bias must be taken cautiously. However, this finding is consistent with 11 of the 26 earlier findings reviewed. The other ones revealed either stricter criteria, more conservative criteria, or mitigated effects associated with age. For instance, in an earlier study on visual discrimination involving participants with more than 13 years of education, Baracat and Marquie 1992 found the opposite age-related pattern to the one observed here in the corresponding education group. It may be that age differences depend on the type of task. Some results suggest this (e.g., stricter criteria in pain perception, more risky in vigilance), especially those obtained on different tasks with the same participants (e.g., Baron & Belongia Le Breck, 1987). But the scarcity of studies on each type of task and a number of within-task inconsistencies in the age effects observed, even in sensory tasks (auditory and weight perception), make it difficult to draw a definite conclusion on the issue at this time. The lack of a consistent age effect is especially true in recognition tasks. Indeed, no age differences on verbal or face recognition tasks were found by Baron and Belongia Le Breck (1987), Baron and Surdy 1990, Ferris and colleagues 1980, Gordon and Clark 1974a, or Yanz and Anderson 1984. But different results were obtained in other experiments. Gordon-Salant 1986, using a speech recognition task, found more risky criteria in the older participants. Gordon and Clark 1974b, using a nonsense syllable recognition task, found more risky criteria on the first trial only and stricter criteria for words on the first trial, but no age difference for either type of material on the second trial. Harkins and colleagues 1979 also used a recognition task but with a different procedure in which participants rated each stimulus on a 5-point confidence scale. They found mitigated results, with elderly participants exhibiting a criterion that was less stringent than the young on the first category, similar on the three intermediate categories, and stricter on the fifth category, mainly indicating that older participants restricted the range of their criteria more than young ones.
Perhaps more interesting from a methodological point of view is our finding of an age effect in only the most educated subjects. It confirms our hypothesis that a possible source of discrepancy among previous findings regarding age effects on response bias could be due to differences in the characteristics of the samples. As stressed previously, the level of education was not always reported in previous studies (one out of two), and when it was, it was often described in a fairly sketchy manner (e.g., college level). As far as we can tell, the only studies that allow for a clear-cut comparison with our most educated group (
) and have the same characteristics in terms of type of task and equivalent proportions of men and women are the studies by Gordon and Clark 1974a, Gordon and Clark 1974b. However, their results are only partially similar to ours, because they found no age difference on the prose recognition task (Gordon and Clark 1974a) or on the second trial of the second study (Gordon and Clark 1974b), as indicated previously. They found stricter and riskier criteria for words and nonsense syllables, respectively, on the first trial. As with education, and for the same reasons, it is difficult to compare the present sex effects with those of previous studies. Some exceptions, however, are the lack of a clear overall sex-related trend across the various analyses performed in the weight perception study by Danziger and Botwinick 1980 and no sex difference in the pain perception study by Clark and Mehl 1971. By contrast, the Gordon and Clark 1974b findings were similar to ours, with stricter response bias in men than in women.
The issue of how to interpret age differences, when they exist, is still unresolved. In fact, the goal of most of the work done so far has been to determine the extent to which age-related differences in sensory and cognitive tasks result from differences in target ability or response strategy. Some interpretations have been suggested, but they are mostly sketchy and post hoc. A case in point is the hypothesis of Watson and colleagues 1979 that elderly people are more cautious only in tasks where there is a perceived deficit, which could explain between-task differences in age effects. Other examples are Clark and Greenberg 1971, Clark and Mehl 1971, and Gordon and Clark 1974b, who assumed that the greater anxiety of elderly people over having their cognitive abilities tested in experimental situations accounts for age differences in response bias. Another possibility is close to the hypothesis we examined: that some results in aging studies may be due to the underlying effects of education. As shown here, education is a fairly powerful predictor of response bias. And it has been suggested that education is closely associated to verbal skills, which in turn may influence decision criteria (Gordon-Salant 1986; Yanz and Anderson 1984; see also Botwinick 1978). Further research is needed to examine these or other hypotheses.
The study of age-related changes in response bias is useful for obtaining a clearer view of whether cautiousness is a general age-dependent trait, but it also serves a methodological and applied purpose. Given that decision processes are omnipresent in a wide range of laboratory and everyday tasks, identifying age-specific bias trends can help avoid misinterpretations of many perceptual and cognitive research findings involving young and older adults. From an applied standpoint, better knowledge of the bias and sensitivity components of the behavior of elderly people would help in the design of appropriate remedial and training programs that could help make their behavior better suited to environmental demands and to their own goals, as illustrated by Belbin and Shimmin 1964. We successfully demonstrated that in the word recognition task studied here, and in the age range considered, age effects are not general but are a function of both educational level and sex. It may be that our finding of a stronger and more robust educational effect on decision criteria than the age effect cannot be extended downward to sensory domains or upward to more complex social-cognitive domains. Likewise, it is possible that with a wider age range than we used age effects might be more marked. The present findings only suggest that, because education and sex clearly affect decision processes in this type of task, these variables should be more systematically controlled in future work on age effects.
|
Acknowledgments |
|---|
Received for publication September 30, 1998. Accepted for publication December 22, 1999.
|
Appendix ENDIX |
|---|
|
TOPAbstractMethodsResultsDiscussionAppendix ENDIXReferences |
|---|
H = Hits
F = False Alarms
A' = 0.5 + (H - F) (1 + H - F) / [4H(1 - F)], when H 
F.
A' = 0.5 + (F - H) (1 + F - H) / [4F(1 - H)], when H 
F.
B'H = 1 - F(1 - F) / [H(1 - H)], when H 1 - F.
B'H = H(1 - H) / [F(1 - F)] - 1, when H 1 - F.
|
References |
|---|
|
TOPAbstractMethodsResultsDiscussionAppendix ENDIXReferences |
|---|
| ||||||||||||||||||||||||
| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
).