| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
| ||||||||||||||||||||||||||||||||
RESEARCH ARTICLE |
1 Hebrew Rehabilitation Center for Aged Research and Training Institute, Boston, Massachusetts.
2 Department of Family Practice and Community Medicine, University of Pennsylvania, Philadelphia.
Address correspondence to Richard N. Jones, ScD, HRCA Research and Training Institute, 1200 Centre Street, Boston, MA 02131. E-mail: jones{at}mail.hrca.harvard.edu
 |
Abstract |
|---|
 TOP Abstract Subgroup Differences in Measured...Methodological Approaches to...IRTDetecting Disturbances in...MIMIC Model Approach to...HypothesesMethodsResultsDiscussionReferences |
|---|
The correlation of education with tests of cognitive functioning may be the most robust finding in gerontologic public mental health (Bassett & Folstein, 1991; Boone, Ghaffarian, Lesser, Hill-Gutierrez, & Berman, 1993; Brayne & Calloway, 1990; Crum, Anthony, Bassett, & Folstein, 1993; Evans et al., 1993; Fillenbaum, Hughes, Heyman, George, & Blazer, 1988; Ganguli et al., 1991; Jorm, Scott, Henderson, & Kay, 1988; Magaziner, Bassett, & Hebel, 1987; Murden, McRae, Kaner, & Bucknam, 1991; O'Connor, Pollitt, Treasure, Brook, & Reiss, 1989; Wiederholt et al., 1993), and has led some to argue that education may be an important risk factor for dementia (Mortimer & Graves, 1993). The epidemiology of cognitive impairment has been facilitated by the availability of brief cognitive status assessment instruments such as the MMSE (Folstein, Folstein, & McHugh, 1975). The MMSE is a short assessment instrument that assesses orientation to time and place, attention, memory, and ability to follow commands. This article is concerned with the validity of the MMSE as a measure of cognition and with threats to its validity according to education and sex. In the context of Messick's (1989) articulation of the unitary concept of validity, this investigation can be seen as an attempt to demonstrate the extent to which the validity of the MMSE is compromised by test-irrelevant variance. Our goal is to determine the degree to which item-level performance is influenced by level of education after controlling for differences in underlying cognitive ability.
|
Subgroup Differences in Measured Cognitive Impairment |
|---|
|
TOPAbstractSubgroup Differences in Measured...Methodological Approaches to...IRTDetecting Disturbances in...MIMIC Model Approach to...HypothesesMethodsResultsDiscussionReferences |
|---|
|
Methodological Approaches to Investigations of Subgroup Differences |
|---|
|
TOPAbstractSubgroup Differences in Measured...Methodological Approaches to...IRTDetecting Disturbances in...MIMIC Model Approach to...HypothesesMethodsResultsDiscussionReferences |
|---|
|
IRT |
|---|
|
TOPAbstractSubgroup Differences in Measured...Methodological Approaches to...IRTDetecting Disturbances in...MIMIC Model Approach to...HypothesesMethodsResultsDiscussionReferences |
|---|
|
Detecting Disturbances in Measurement With IRT |
|---|
|
TOPAbstractSubgroup Differences in Measured...Methodological Approaches to...IRTDetecting Disturbances in...MIMIC Model Approach to...HypothesesMethodsResultsDiscussionReferences |
|---|
IRT-based methods for detecting DIF are superior to methods based on comparing proportion correct or mean score comparisons across groups and are reviewed by Teresi, Kleinman, and Ocepek-Welikson (2000). Briefly, a test item is considered to be free from measurement disturbance or DIF when item response functions or ICCs for two groups are equivalent. That is, individuals matched on ability from different groups should have the same probability of responding correctly to a given item (Angoff, 1993; Camilli & Shepard, 1994; Osterlind, 1983). The degree to which ICCs differ across groups is summarized with DIF statistics: One convenient summary of DIF is the area between ICCs plotted for two groups (Raju, 1988).
The IRT approach to DIF attempts to disentangle measurement noninvariance from population heterogeneity, and it represents an important development for the evaluation of measurement devices. One statistical approach to this is the multiple indicators–multiple causes, or MIMIC, model (Jöreskog & Goldberger, 1975) for dichotomous variables (Gallo et al., 1994, 1998; Muthén, 1987, 1989).
|
MIMIC Model Approach to DIF |
|---|
|
TOPAbstractSubgroup Differences in Measured...Methodological Approaches to...IRTDetecting Disturbances in...MIMIC Model Approach to...HypothesesMethodsResultsDiscussionReferences |
|---|
|
Hypotheses |
|---|
|
TOPAbstractSubgroup Differences in Measured...Methodological Approaches to...IRTDetecting Disturbances in...MIMIC Model Approach to...HypothesesMethodsResultsDiscussionReferences |
|---|
|
Methods |
|---|
|
TOPAbstractSubgroup Differences in Measured...Methodological Approaches to...IRTDetecting Disturbances in...MIMIC Model Approach to...HypothesesMethodsResultsDiscussionReferences |
|---|
The total number of respondents interviewed in the five-site ECA program was 20,861. Excluded participants were those whose age at initial interview was younger than 50 years (n = 10,594), missing data for year of birth (n = 28), not living at home (n = 1,294), did not at least start the MMSE (n = 326), or missing data for educational attainment (n = 63). Characteristics for the final sample of 8,556 are summarized in Table 1.
|
Age, sex, and ethnicity
Three age groups were included in our MIMIC models. Adults in the 50–64 age group served as the reference group for the elderly (age 65–74) and very old (age 75 and older) age groups. Sex was recorded as observed.
Self-reported ethnicity
Information on self-reported ethnicity was obtained by asking "Would you please look at this card and give me the letter of the group that best describes your racial background?" Respondents then selected from this list: American Indian, Alaska Native, Asian, Pacific Islander, Black–Not Hispanic, Hispanic, White–Not Hispanic. We considered three ethnic groups based on this self-report—White, Black or African American, and all other racial/ethnic groups—in our examination of the distribution of education.
Education
Education was based on participants' responses to the question "What is the highest grade in school or year of college that you completed?" Responses were assigned values ranging from 0 (indicating no formal education) to 17 (representing graduate school). It should be noted that this method of ascertaining educational attainment is not wholly consistent with methods used by the National Center for Health Statistics National Health Interview Survey (Aday, 1989) and may result in overestimates of years of completed education. We are unaware of evidence that this method of ascertaining educational attainment performs differentially according to birth cohort, sex, or ethnicity.
A number of alternatives to express education as an analytic variable were considered, motivated by discomfort with the assumption that estimated regression effects are constant across grade levels (implicit when treating education as a continuous variable). Splitting participants into groups on the basis of the sample's mean years of completed education, or on the basis of educational milestones, relies on the dubious assumption that each year of completed education or attainment of milestones implies the same education for all age, sex, and ethnic subgroups in the sample.
We used a new approach to describe educational attainment: a relativistic education indicator. This indicator was created by stratifying the sample by age at initial interview (grouped by ages 50–54, 55–59, 60–64, 65–69, 70–74, 75–79, 80–84, 85+), sex, and self-described ethnicity (White, Black or African American, and other). Within each of the resulting 48 strata, participants with fewer than the median years of completed education were classified as having low education. Five-year intervals were chosen, as finer age strata would result in sparsely populated strata and unstable ranks. This indicator removes the influence of cohort factors (age, sex, ethnicity) in grouping respondents into high- and low-education groups.
A comparison of the performance of the relativistic educational attainment indicator to an indicator split at the overall sample mean (Grade 10) is reported elsewhere (Jones, 1997; Jones & Gallo, 2001). Briefly, the relativistic indicator is not correlated with age (Spearman's 
= -0.01, p = .64), but a mean-split indicator is ( = 0.22, p < .001). The relativistic indicator is not associated with being Black or African American ( = -0.01, p = .46) but the mean-split indicator is ( = 0.19, p < .001). In addition, the relativistic educational attainment indicator and the mean-split educational attainment indicator are comparably associated with cognitive performance after partialing out the effects of age, sex, and ethnicity (standardized regression coefficients, ßs = -0.30 and -0.32, respectively).
Analytic Approach: The MIMIC Model
Consistent with the modern SEM approach to dichotomous dependent variables (Muthén, 1978; Muthén & Lehman, 1985), our model specified a threshold model for the observed dichotomous MMSE items (y). The likelihood of responding incorrectly to a MMSE item was modeled using multivariate probit regression. Probability of an incorrect response is viewed as a function of a latent ability or a latent trait and background variables. Background variables may influence items directly or indirectly as mediated by the underlying trait. This latent trait (
) fulfills the same role as the latent ability (
) in the IRT model. Slopes relating the underlying trait to the item responses are probit regression coefficients (
) and are analogous to factor loadings in confirmatory factor analysis or item-discrimination parameters in IRT. The underlying trait (in this case, cognitive impairment) is assumed to be continuously distributed and normal, conditional on the included covariates.
The latent trait is regressed on background variables (covariates, x). The corresponding linear regression parameters (
) are referred to as indirect effects. This terminology reflects the fact that these parameters capture the increase in the likelihood of making an error on the item associated with the covariate mediated by the underlying trait. When the covariate is a dichotomous dummy indicator, these regressions can also be conceptualized as expressing mean differences in the underlying trait in line with traditional analysis of covariance models. Our model included background variables for participants older than 75 years (1 if age at interview was 75 years or more, 0 otherwise) and for participants aged between 65 and 74 years at interview (1 if true, 0 otherwise). Participants aged 50–64 formed the reference group. Our model also included indicator variables for Blacks or African Americans, leaving Whites and all other ethnicity groups in the reference group. The model also included indicators for relative educational attainment (1 = low education, 0 otherwise) and male sex. With the exception of education, all background variables were centered at their respective means in the analysis. This modeling strategy results in equivalent model fit and regression parameter estimates, with the exception of mean and threshold parameters, that can be interpreted for participants balanced at the overall sample mean on the other background variables.
DIF is detected by including regressions of the individual items on the background variables. These probit regression parameters (
) describe differences in item difficulty associated with the background variable. These regressions are termed direct effects and describe differential item difficulty over and above that expected due to the effect of the covariate on the underlying trait. Finally, it is important to note that all of these regressions are estimated simultaneously and thus reflect independent effects conditional on the other parameters in the model.
Assessing Model Fit
Model fit was assessed with the root mean square error of approximation (RMSEA; Browne & Cudeck, 1993; Muthén & Muthén, 1998) and the comparative fit index (CFI; Bentler, 1990; Muthén & Muthén, 1998). The RMSEA provides a measure of discrepancy per model degree of freedom. The RMSEA will approach 0 as model fit improves. Browne and Cudeck recommended rejecting models with RMSEA values greater than 0.1; Hu and Benter (1998) suggested values close to 0.06 or less represent adequately fitting models. The CFI is based on the model chi-square: Values range between 0 and 1 and values greater than 0.95 are generally accepted for adequately fitting models.
MIMIC Model Building and Estimation
The procedure for building a structural probit or MIMIC model to detect DIF has been outlined by Muthén (1988). The first model freely estimated all indirect effects, with all direct effects fixed to zero. Disturbance in measurement by background variables was identified by examining the matrix of first-order derivatives of the fit function corresponding to direct effects. Large values (absolute value) indicate that if the corresponding parameter were freely estimated, significant improvement in model fit would result. The model was then reestimated with the associated parameter freely estimated. This procedure is similar to that used in general SEM, using the modification index to inform model building (Jöreskog & Sörbom, 1996). If the resulting improvement in model fit was significant (p < .05) as judged by the chi-square difference test, the parameter was retained and the process repeated. Iterations were stopped when improvement in model fit was not significant. Models were estimated by using the Mplus program's weighted least squares estimator (Muthén & Muthén, 1998).
|
Results |
|---|
|
TOPAbstractSubgroup Differences in Measured...Methodological Approaches to...IRTDetecting Disturbances in...MIMIC Model Approach to...HypothesesMethodsResultsDiscussionReferences |
|---|
|
|
) reveal that all items were strongly related to the underlying latent trait, a reflection of the essentially unidimensional structure of the MMSE (Jones & Gallo, 2000). The least discriminating MMSE item was the orientation to place item "what county are we in?" The metric of the latent trait was set with the attention item repeat "apple" ( = 1). Those with low education had a significantly greater mean level of cognitive dysfunction (indirect effect, = 0.55, SE = 0.22). Male and female ECA respondents did not differ in level of cognitive dysfunction (indirect effect, = 0.00, SE = 0.03).
Evidence of DIF
There was evidence of a lower likelihood of error on the remember "penny" item for those with low education. Those with low education were more likely to err on the first serial subtraction but not on subsequent subtractions. Although those with low education were more likely to err on naming items (Table 2), this excess error rate is lower than would be expected, due to the estimated mean level of cognitive disability for this group (Table 3). In other words, the MIMIC-model–implied mean level of cognitive dysfunction overestimates the likelihood of error on these items for the low-education group. This was also true for the fold paper in half task. Those with low education were more likely to err on the write a sentence and copy design tasks.
Men were less likely to err on the county orientation item ( = -.29, SE = 0.05). This direct effect provided the shift in the difficulty (or difference in the probability of making an error in the probit scale) for men relative to women, holding constant the effects of education, age, and ethnicity on the construct and item levels. Men were more likely to make an error on two of the three attention items, two delayed recall items, and on the repeat phrase item. Relative to women, men were more likely to err on the spelling backwards item but much less likely to err on the serial subtraction items. Men were more likely to err on the write a sentence item, but less likely to err on the copy design task. Men were also much more likely to err on the read and follow command tasks.
Although not summarized, the effects listed in Table 3 include control for DIF attributable to old (age 65–74) and very old (age 75 years and older) age groups (relative to those aged 50–64) and for those self-identifying as Black or African American relative to all other racial/ethnic groups. Briefly, detected direct effects suggest DIF for the orientation to season, delayed recall tasks, naming, repeat phrase, three-step command, write a sentence, and copy polygon items for the older age groups. The model detected DIF relative to race/ethnicity for the registration, delayed recall, naming, repeat phrase, read and obey command, three-step command, write a sentence, and copy polygons tasks. The full matrixes of direct and indirect effects, as well as residual variances, are available from Richard N. Jones on request.
DIF and Estimates of Underlying Ability
To assess the influence of DIF on estimates of the level of underlying ability, we estimated a purposefully misspecified model fixing all direct effects for education to zero (ignoring DIF). The standardized regression coefficients of latent cognitive impairment on low education can be compared between the final model and the purposefully misspecified model to obtain an estimate of the magnitude of bias that is due to educational attainment. The standardized mean latent cognitive dysfunction level for the referent group is presumed to be zero with unit variance, and the standardized regression coefficient for group membership provides the standardized difference in latent cognitive impairment for members of the indicated group. Considering first the educational attainment group differences in latent cognitive dysfunction, the standardized coefficient was .498 for the final model and .490 for the purposefully misspecified model (ignoring DIF). Thus, in both models, the low-education group has about a half a standard deviation greater level of cognitive impairment relative to the high-education group, but only about 1.6% of the difference between estimated latent cognitive dysfunction for the high- and low-education groups is due to DIF. By way of comparison, the standardized coefficient for the regression of latent cognitive impairment on male sex is -.004, implying no significant difference in mean cognitive dysfunction by sex (estimate/standard error [z] = -0.16, p = .870). In a purposefully misspecified model holding all sex direct effects to zero, the standardized regression coefficient was -.081 (z = -4.51, p < .001), implying significantly less cognitive dysfunction for men, but of very small magnitude. Thus, ignoring DIF (the misspecified model) overestimates the level of cognitive dysfunction for women by about 95%. Nearly all of the very small difference in underlying cognitive dysfunction attributable to sex is due to items that perform differently by sex.
|
Discussion |
|---|
|
TOPAbstractSubgroup Differences in Measured...Methodological Approaches to...IRTDetecting Disturbances in...MIMIC Model Approach to...HypothesesMethodsResultsDiscussionReferences |
|---|
Limitations
Before putting the results in the context of previous research, it is important to discuss the limitations of the current analysis. The relativistic education indicator addresses some of the difficulties in treating years of completed education as an analytic variable. As it is unknown how early life exposure to education influences cognition in late life, the relativistic indicator may mask a true education effect. On the other hand, raw years of education have been known to capture residual correlation of age-related constructs via the strong association of years of completed education and birth cohort (Cobb, Wolf, Au, White, & D'Agostino, 1995).
Another limitation is the problem of constant bias. Detecting DIF is straightforward when most items on a test do not show DIF. IRT approaches are not able to detect constant bias (Camilli & Shepard, 1994). Constant bias refers to the condition where many or all items on a test are equivalently biased. This situation is conceptually, but not statistically, distinct from a situation where all items are, in fact, not biased, but persons from contrasted groups differ in underlying ability. When a large number of items on a test are similarly biased, DIF analyses will overestimate group differences in underlying ability and reveal a small number of items apparently biased in favor of the minority or disadvantaged group. Interpretation of results is further complicated if group differences in underlying ability exist and a large number of items are biased. If many of the MMSE items are biased in the same way by education level, we would expect exactly the findings observed: Little evidence of item bias or small bias apparently favoring the minority or disadvantaged group and large group differences in underlying ability (Camilli & Shepard, 1994).
Previous Research on Item Bias
One previously published study examined MMSE responses for DIF according to education by using an IRT approach. Teresi and colleagues (1995) found several items in a cognitive battery (that included items from the MMSE) biased by education among participants in a dementia case register study (n = 550). Teresi and colleagues highlighted the orientation to state, repeat phrase, naming, read and follow instruction, three-step command, and write a sentence items as biased by level of education: All but the orientation to state item were more difficult for respondents with low education. Our results agreed for the write a sentence and repeat phrase items. Unlike Teresi and colleagues, our results suggest DIF for naming and the three-step command item fold paper in half that favors respondents with low education. Our results do not replicate evidence of DIF for the state orientation item and detected DIF for the first serial subtraction, spelling, orientation to season, and copy polygon items. Although there are many reasons why results of Teresi and colleagues and our own differ, the commonalities deserve mention. The write a sentence task was flagged as possibly biased in both studies, and it is plausible to suspect this item of education bias given the manifest content. This finding suggests DIF analyses are capable of identifying items that may need to be dropped, revised, or scored differently for individuals with varying levels of education. Given the results of our investigation of the global impact of measurement bias using purposefully misspecified MIMIC models, however, a recommendation to drop this item is not supported.
There are important differences between the work of Teresi and colleagues (1995) and the analyses we present that could result in variant findings. Firstly, Teresi and colleagues analyzed a set of cognitive assessment items that included tasks from the MMSE, the Blessed (Blessed, Tomlinson, & Roth, 1968), the Short Portable Mental Status Questionnaire (Pfeiffer, 1975), the Kahn-Goldfarb Mental Status Questionnaire (Kahn, Goldfarb, Pollack, & Peck, 1960), the Caregiver Assessment and Referral Evaluation Mental Status Questionnaire (Golden, Teresi, & Gurland, 1984), and the Caregiver Assessment and Referral Evaluation diagnostic scale (Golden, Teresi, & Gurland, 1983). This design feature is an important advantage of the Teresi and colleagues study. If the array of non-MMSE items were not biased, the analysis with the larger item set would have more power to detect bias in MMSE items. However, Teresi and colleagues found that many of the non-MMSE items demonstrated DIF attributable to education. Therefore, the longer item set may also suffer from the problem of constant bias to the same extent that the MMSE suffers from constant bias. Second, Teresi and colleagues fit a three-parameter IRT model (guessing, discrimination, and difficulty parameters) using the LOGIST software package. In contrast, we fit a two-parameter IRT model (discrimination, difficulty) with item-discrimination parameters assumed to be equal across groups. Teresi and colleagues found most guessing parameters to be close to zero, so this does not seem to be an important difference between our studies. However, Teresi and colleagues allowed discrimination and difficulty parameters to differ across groups, whereas we allowed only difficulty parameters to vary across groups. An adequate discussion of the strengths and limitations of these two parameterizations of item-bias detection models is beyond the scope of this report (but see Hambleton, Wright, Crocker, Masters, & van der Linden, 1992). This difference in parameterization could lead to differences in results. Finally, our study concerned a relatively large (N = 8,550), geographically diverse, and representative community sample of older adults. In contrast, Teresi and colleagues conducted their analyses with a much smaller sample (N = 550) of older adults, about a third of whom had been referred to a memory disorder clinic for neurological diagnosis, and with approximately equal numbers of White, Black or African American, and Hispanic participants. Thus, Teresi and colleagues' sample is not representative of community-dwelling older adults, and their low-education group (defined as completion of 8 or fewer years of formal education) was disproportionately represented by minorities. Thus, DIF detected for low education in Teresi and colleagues' work may be due to ethnicity (or language, culture) rather than education. These differences highlight the advantage of the MIMIC model (multiple background variables can be included) and the relativistic educational attainment indicator (reduce confounding by age or ethnicity).
Perhaps the most intriguing finding of our analyses—a finding that cannot be elicited from previous IRT-based research on the MMSE that derives from our use of the MIMIC model—is the finding that there are at best minor measurement disturbances in MMSE by level of education. Detected DIF is not sufficient to account for education group differences in overall level of estimated cognitive ability. Furthermore, the magnitude of detected DIF by level of education is small relative to the magnitude of effects by sex.
Sex differences detected in how the MMSE measures cognition were strong and to a large extent predicted by neuroendocrine and neuropsychological research. The detection of specific sex differences is important: It highlights that the MMSE is sensitive to group differences expected by theory and that these item-level differences can be detected with the MIMIC model approach. The specific items for which potential item bias may be at work support a theory that the tasks draw on activities learned in early schooling such as spelling, writing, and familiarity with paper-and-pencil tasks. A very important feature separates the importance of DIF attributable to sex and DIF attributable to education. This is that researchers and clinicians often adjust total scores in an attempt to attenuate or remove education-level differences in scores derived from the MMSE (and other instruments) under the assumption that the test is biased. Such adjustments or modifications according to sex are unknown to us. However, researchers often use the serial subtraction or spell world backwards item interchangeably or one or the other exclusively. Using spell world backwards instead of serial subtractions would systematically overestimate the level of cognitive impairment for men. If the research question includes demonstrating sex differences in cognitive function or in the prevalence or incidence of dementia and the MMSE is used as a screening device, researchers should give some consideration to item choice and should perhaps include both the serial subtraction and the spell world backwards items.
Conclusion
Compelling hypotheses or explanations for the association of education and cognitive function in late life include (a) the brain reserve hypothesis (higher education confers richer neuronal or dendritic density, providing resistance to loss associated with aging; Katzman, 1993), (b) the accelerated (cognitive) aging hypothesis (individuals from socioeconomically disadvantaged groups age at a faster pace), (c) reverse causation (i.e., low intelligence causes lower educational attainment and poor cognitive performance in late life), (d) that overall differences are due to item (Anthony et al., 1982) or test bias (O'Connor, Pollitt, & Treasure, 1991; group differences reflect differences in skills or capacities conferred by education), or (e) that education is either a surrogate for unmeasured lifestyle risk or neuroprotective factors (Orell & Sahakian, 1995; Pitt, 1993) or those with high premorbid intelligence (as evidenced by high academic attainment) experience cognitive decline that is not detected by the available crude mental status assessment instruments (Slater & Roth, 1969).
Our analyses evaluate the test or item bias hypothesis. Our findings reveal that although item bias or DIF attributable to education level was detected, it accounted for a very small fraction of the overall group difference in assessed cognition. Our results set limits on the extent to which item bias influences overall education group differences on assessed cognition. After the presence of DIF was adjusted for, large group differences in underlying cognitive impairment across educational attainment group remained. Furthermore, not only was the magnitude DIF attributable to education not as great as that observed for sex, but the sex DIF was responsible for most of the difference between the genders on assessed cognitive functioning.
As for the other hypotheses or explanations for the association of education and assessed cognitive function in late life, our conclusion that item bias accounts for little of the observed association encourages additional research into the other explanations. Thus, our findings have implications for researchers investigating the role of education in cognitive impairment, cognitive decline, and risk of neurological disease. After DIF was adjusted for DIF, considerable education group differences in underlying cognitive functioning remained. Thus, other mechanism(s) must account for education group differences in assessed cognitive function. Researchers concerned with these hypotheses may appreciate our evidence that measurement bias accounts for only a small portion of observed group differences in assessed cognitive function. Furthermore, our results argue that ad hoc MMSE total score adjustment methods that completely remove education group differences in MMSE scores (e.g., Kittner et al., 1986; Mungas, Marshall, Weldon, Haan, & Reed, 1996) overadjust and replace bias in assessed cognition favoring those with high educational attainment with bias favoring those with lower educational attainment. Also, efforts to remove the influence of education by dropping items (cf. Bazargan, Baker, & Bazargan, 2001) were not supported by the results of this psychometric analysis. The failure of DIF and possible item bias to account for much of the difference in assessed cognitive functioning may be one reason why such total score adjustments fail to substantially improve neuropsychiatric case identification activities (see Belle et al., 1996, for an applied example and Kraemer, Moritz, & Yesavage, 1998, for a review). The failure of detected DIF to account for more than a small fraction of the difference in assessed cognition between high- and low-education groups may simply be evidence that the construct measured bears only a weak link to late-life cognitive functioning. The test may more strongly measure academic achievement and familiarity with the testing situation (cf. Flynn, 1987). If this is true, then no item-level, ad hoc, or item-omission modification will improve the validity of the mental status assessment device. Research using neuropsychological batteries and assessment devices relevant to the daily cognitive functioning of older adults in large, representative, and longitudinal samples are needed to further our understanding of how socioeconomic status and early life experiences influence cognitive functioning in late life.
|
Acknowledgments |
|---|
Received for publication May 12, 2000. Accepted for publication February 7, 2002.
|
References |
|---|
|
TOPAbstractSubgroup Differences in Measured...Methodological Approaches to...IRTDetecting Disturbances in...MIMIC Model Approach to...HypothesesMethodsResultsDiscussionReferences |
|---|
This article has been cited by other articles:
 |
|
 |
|
  F. M. Yang and R. N. Jones Measurement Differences in Depression: Chronic Health-Related and Sociodemographic Effects in Older Americans Psychosom Med, November 1, 2008; 70(9): 993 - 1004. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||
| HOME | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|---|
