Use of Spoken and Written Japanese Did Not Protect Japanese-American Men From Cognitive Decline in Late Life

Abstract

Objectives.

Spoken bilingualism may be associated with cognitive reserve. Mastering a complicated written language may be associated with additional reserve. We sought to determine if midlife use of spoken and written Japanese was associated with lower rates of late life cognitive decline.

Methods.

Participants were second-generation Japanese-American men from the Hawaiian island of Oahu, born 1900–1919, free of dementia in 1991, and categorized based on midlife self-reported use of spoken and written Japanese (total n included in primary analysis = 2,520). Cognitive functioning was measured with the Cognitive Abilities Screening Instrument scored using item response theory. We used mixed effects models, controlling for age, income, education, smoking status, apolipoprotein E e4 alleles, and number of study visits.

Results.

Rates of cognitive decline were not related to use of spoken or written Japanese. This finding was consistent across numerous sensitivity analyses.

Discussion.

We did not find evidence to support the hypothesis that multilingualism is associated with cognitive reserve.

Multilingualism and Cognitive Reserve

The movement of people from country to country has fascinated social scientists for decades. Immigrants embody dialogue between cultures, and the social, psychological, and medical consequences of migration have been extensively investigated. One specific area of interest is the cognitive effect of multilingualism. Individuals raised speaking one language who migrate to a place with a different predominant language face challenges and opportunities different from those faced by nonmigrants, many of whom will need to master only a single language to function in their daily lives.

An important area of interest is cognitive reserve. One of the prevailing explanations of why the burden of neuritic plaques and neurofibrillary tangles has an imperfect correlation with clinical Alzheimer’s disease (AD) during life is referred to as reserve (Le Carret et al., 2005; Scarmeas & Stern, 2004; Stern, 2002, 2003; Whalley, Deary, Appleton, and Starr, 2004). The reserve hypothesis is that “brain” reserve (total brain volume, dendritic/synaptic density, etc.—“hardware”) and/or “cognitive” reserve (higher educational attainment, cognitively or physically demanding occupations or behaviors, etc.—“software”) buffer the effects of neurodegenerative processes, leading to relative protection. These concepts have been referred to as “neurological brain reserve” and “behavioral brain reserve” (Valenzuela & Sachdev, 2006b).

In this article, we examine the possibility that multilingualism may increase cognitive reserve. Phonological language processing in bilingual individuals occurs in parallel with both shared and separate brain structures (Marian, Spivey, and Hirsch, 2003). One study found lower rates of cognitive decline for a period of 2 years in Japanese-American immigrants who were less assimilated to Western culture than those who were more assimilated. One of the reasons suggested for this finding was that facility with written Japanese language led to increased cognitive reserve (Graves et al., 1999). Bialystok, Craik, and Freedman (2007) state that “The speculative conclusion … is that bilingualism does not affect the accumulation of pathological factors associated with dementia, but rather enables the brain to better tolerate the accumulated pathologies,” p. 463.

There are some cross-sectional data, however, that suggest that some cognitive difficulties may be related to multilingualism as reviewed in Kave, Eyal, Shorek, and Cohen-Mansfield (2008), see p. 71. Specifically, bilinguals may have more tip-of-the-tongue retrieval failures and fewer correct responses on confrontation naming tests and generate fewer items on verbal fluency tasks. These findings may be confounded by difficulties with assessing cognitive function in multilingual individuals using monolingual (often secondary language) test instruments. There have been few longitudinal studies addressing the effects of multilingualism on cognitive trajectories. As Salthouse (2006) argues, longitudinal data are necessary to address questions of cognitive aging.

Spoken and Written Japanese and Reserve

There are two distinct sets of characters in written Japanese, the ideographic Kanji script and the syllabic Katakana script. A third distinct set of characters, the syllabic Hiragana script, is used for non-Japanese words. Japanese schooling involves learning more than 1,000 Kanji characters in elementary and junior high school. Kanji is perceptually more complex and demanding than English script because many more than 26 basic perceptual elements are used (Matsuoka, Uno, Kasai, Koyama, and Kim, 2006). If fluency in written Japanese language were protective for dementia and cognitive decline, we would predict that mastery of written Kanji would be more protective than Katakana or Hiragana.

We thus predict an ordinal relationship between Japanese language use and rates of cognitive decline. The least protection would be provided by neither speaking nor reading Japanese (monolingual English), followed by speaking Japanese but not being able to write it (spoken bilingual), followed by speaking and reading and writing Japanese (spoken and written bilingual).

In a previous article, we tested this hypothesis using AD and dementia outcomes from the Honolulu-Asia Aging Study (HAAS) (Crane et al., 2009). We found no protection associated with self-reported use of spoken or written Japanese. Here, we address the question of whether spoken and/or written Japanese language use might be associated with protection from cognitive decline.

Second-Generation Immigrants As a Particularly Valuable Population to Study

It would be impossible to perform a controlled study in which individuals are randomized to speak only a single language or to speak two or more languages and then follow those individuals for decades to observe trajectories of cognitive functioning in late life. Attempts to understand the implications of multilingualism on late life cognition must be limited to observational data.

An important consideration in observational studies of multilingualism is that immigrants are often very different from individuals who do not migrate; those differences are by no means limited to language use. Many families who choose to migrate perceive improved economic opportunities after migration, whereas those who do not migrate may perceive sufficient economic opportunities where they are. Childhood experiences and environmental exposures may be very different for migrants compared with nonmigrants. It can be very difficult if not impossible in an observational study to parse out differences related specifically to multilingualism from differences due to the very different life experiences of migrants and nonmigrants. As an example of this difficulty, in an article on multilingualism by Bialystock and colleagues (2007), almost all (81 of 93 or 87%) of the bilingual individuals were immigrants and almost none (13 of 91 or 14%) of the monolinguals were immigrants. In a second article by the same group, 20 of the 24 older bilinguals (83%) were immigrants, whereas the number of immigrants among older monolinguals was not reported (Bialystok, Craik, and Luk, 2008). Various statistical approaches may be applied to try to account for differences not related to multilingualism itself, but it would be difficult if not impossible to identify and control for all the factors that are likely to confound the relationship between multilingualism and cognitive reserve among first-generation immigrants. This is particularly challenging if none or very few monolingual study participants are immigrants (Bialystok et al., 2007, 2008).

An alternative approach would be to study first-generation immigrants and limit analyses to those who were able to learn the new host language (Kave et al., 2008). By requiring all included participants to be able to speak at least two languages, Kave and colleagues address the additional cognitive reserve added by moving beyond bilingualism to three or more languages. Such a study cannot address the effects of mono- versus bilingualism.

In this article, we address these questions with a complementary approach. We analyzed a sample of second-generation immigrants whose parents moved from Japan to Hawaii. Although there are differences in environment across any two families, the childhood environments of these second-generation immigrants are much more similar to each other than those of first-generation immigrants and nonmigrants.

These participants were raised in Hawaii where they were exposed to English language schooling. They grew up in a society in which they were exposed to some English, and all had at least functional literacy in English in midlife. Indeed, the Honolulu Heart Program (HHP) that collected this sample in midlife did not need to employ Japanese translators; all interviewing was done in English. Many of these individuals also were able to speak and understand at least some Japanese. They would have been exposed to Japanese to varying degrees in their families of origin (all of which were first-generation immigrant families) and in their neighborhoods and towns. Only a minority of these individuals learned to read and write in Japanese, topics not typically covered in either public or private schools on Oahu when these individuals were of school age. Reading and writing Japanese were learned in special Japanese language schools that some families sought for their sons.

Cognitive Reserve, Brain Reserve, and Genetic Reserve

Our exposure of interest is Japanese language use. Because we are interested in cognitive reserve (software), we controlled analyses for head circumference and for apolipoprotein E (APOE) genotype. Head circumference is an indirect but robust measure of brain reserve (Borenstein Graves et al., 2001) and is related to nutritional status in childhood. Interestingly, the mean head circumference of first-generation immigrants from Japan is smaller than that of second-generation immigrants in the HAAS, likely reflecting early life nutritional differences. There is still some variability within second-generation Japanese immigrants in head circumference, and we control our analyses for this variable in an effort to isolate the effect of interest, which is the relationship between Japanese language use and rate of cognitive decline. Likewise, we control for APOE genotype. APOE genotype is the most robust genetic risk factor for AD identified to date and was found to be associated with risk of AD in HAAS (Havlik et al., 2000). It is included in analyses here to account for its likely effects on the aging brain, in an effort to isolate the effect of interest.

Methodological Issues in Studying Cognitive Decline

It seems intuitive that analyses of cognitive “decline” require longitudinal data (Salthouse, 2006). A perfect study of the effect of interest here would include an extensive neuropsychological battery administered on multiple occasions in late life. Unfortunately, the present data set does not include such a neuropsychological battery. Instead, we have item-level data from a single global measure of cognition, the Cognitive Abilities Screening Instrument (CASI) (Teng et al., 1994). The CASI represented a feasible compromise between incredibly brief cognitive tests such as the Short Portable Mental Status Questionnaire or the Mini-Mental State Examination (MMSE) and a more burdensome multicomponent neuropsychological battery. Indeed, in their systematic review of studies of brain reserve and cognitive decline, Valenzuela and Sachdev (2006a) found that nearly half of the studies used outcome measures that are shorter than the CASI.

The National Institutes of Health formed a trans-NIH Project on Cognitive and Emotional Health, whose Critical Evaluation Study Committee identified 36 longitudinal studies of cognitive functioning with at least 500 subjects (Hendrie et al., 2006). In a previous article, we noted that 33 of these studies with longitudinal data from nearly 150,000 individuals included one of the four global cognitive tests similar to the CASI (Crane et al., 2008). The inferential power of such studies will not lie in detailed understandings of relationships among cognitive domains because these tests are too short to obtain reliable data on any particular cognitive domain. Nevertheless, these studies can identify factors associated with decline in global cognition. Indeed, previous studies of the relationship between multilingualism and cognitive decline published in the psychology literature have employed similar tests (Kave et al., 2008).

There are important statistical challenges in using data from studies that employ longitudinal cognitive tests. We have reviewed some of these challenges (Crane et al., 2008). Tests such as the CASI and MMSE include few easy items, many medium-difficulty items, and few hard items. The sensitivity of such tests to detect change over time is greatest for people with moderate levels of cognitive difficulties. There are few hard items, so people with intact cognition have to decline quite a bit to be detected by the medium-difficulty items that populate the majority of the test. These global cognitive tests have curvilinear (non-interval) scaling properties—that is, the meaning of a difference in the underlying global cognition measured by the test implied by a given number of standard score points varies dramatically. A 5-point decline in standard scores over any time period for people with high test scores implies a great deal of cognitive decline because there are few hard items. A 5-point decline for people with medium test scores implies a much smaller amount of cognitive decline because there are many medium-difficulty items. Standard scoring (total up all the items correct) does not account for variable difficulty of test items.

Curvilinear (non-interval) scaling properties such as those found in these tests pose important concerns for any longitudinal analysis using standard scores. This problem has been discussed by several investigators (Crane et al., 2008; Mungas & Reed, 2000; Proust, Jacqmin-Gadda, Taylor, Ganiayre, and Commenges, 2006). In our own investigation of this issue, we simulated cognitive decline over time using simulated item-level data and compared the fidelity of several scoring techniques for recovering the known rate of decline that generated the data. Use of standard scores led to bias in estimating the rate of decline (Crane et al., 2008).

There are several potential solutions to this common problem of obtaining valid inference from longitudinal studies that employ a global cognitive test with curvilinear scaling properties. The one we use here is to treat the item as the fungible unit rather than the total test score. This approach is known as item response theory (IRT). A nice introduction to IRT can be found in the textbook Item Response Theory for Psychologists (Embretson & Reise, 2000), and a recent review on the use of IRT in the psychology literature is also a good resource (Reise & Waller, 2009). Here, we emphasize that IRT results in interval scaling and is thus a technique to overcome biased estimates of the rate of cognitive decline caused by curvilinear scaling properties of global cognitive tests (Crane et al., 2008).

Another important methodological consideration is the regression modeling strategy. We use a mixed-effects model, an excellent choice when data are incomplete, as is common in studies of elderly individuals (Laird & Ware, 1982). For a nice review of the implications of regression modeling choices for longitudinal psychological or psychiatric data, see Gibbons and colleagues (1993).

METHODS

Overview

We categorized subjects into three groups on the basis of midlife self-reported use of spoken and written Japanese for our primary analysis and on the basis of more specific questions assessed in late life for a secondary analysis. We censored participants at the time of death, dropout, or diagnosis of dementia.

For our primary analyses, we limited our study sample to second-generation Japanese-Americans (Nisei). It was not uncommon to send Nisei children to Japan for their education; these individuals are called Kibei. The number of years of childhood spent in Japan was previously found to be associated with poorer baseline cognitive functioning in HAAS (Yano et al., 2000). Not surprisingly, a high proportion of the Kibei participants spoke, read, and wrote Japanese. We excluded Kibei participants from our primary analyses because we could not differentiate between Japanese language use and other environmental exposures in Japan.

Study Population

The HHP cohort included Japanese-American men born between 1900 and 1919 who were living on the island of Oahu, Hawaii, in 1965 (Syme, Marmot, Kagan, Kato, and Rhoads, 1975). Three midlife Examinations conducted as part of the HHP in 1965–1968, 1968–1970, and 1971–1974 included collection of clinical and demographic information. In the fourth Examination in 1991–1993, when the cohort was 71–93 years old, HAAS was started as an extension of the HHP; 3,734 members of the cohort (80% of the survivors) participated and had their cognitive functioning measured (see subsequently). As shown in Figure 1, we limited our primary analysis to the 2,520 Nisei without dementia at enrollment in 1991–1993. Our analytic sample was limited to those for whom complete cognitive score, exposure, and covariate data were available for at least one examination.

Figure 1.

Analytic sample for Models 1 and 2. CASI = Cognitive Abilities Screening Instrument; HAAS = Honolulu-Asia Aging Study.

HAAS was approved by Institutional Review Boards of the Kuakini Medical Center (Honolulu, Hawaii) and the Honolulu Department of Veterans Affairs. All participants gave written informed consent. The funding sources had no role in the decision to conduct the study or to publish results.

Assessment of Cognitive Functioning

Cognitive functioning was measured with the CASI (Teng et al., 1994) at HAAS baseline (ages 71–93 years) and three additional waves: HAAS Examinations 5 (1994–1996, ages 74–95 years), 6 (1997–1999, ages 77–98 years), and 7 (1999–2000, ages 79–100 years). Trained interviewers administered the CASI and noted problems with vision or hearing that might interfere with the validity of the resulting score. We limited our analyses to valid assessments for nondemented participants.

The CASI is a 25-item, 100-point global cognitive functioning instrument designed for use in studies of Japanese and Japanese-American older adults. It combines elements from a commonly used cognitive test in Japan, the Hasegawa Dementia Screening Scale (Hasegawa, 1983), the most commonly used instrument in the United States, the MMSE (Folstein, Folstein, and McHugh, 1975), and the Modified MMSE (3MS) (Teng & Chui, 1987). The CASI was characterized by minimal differential item functioning impact with respect to a wide variety of covariates in the HAAS study (Gibbons et al., 2009), which suggests that any findings of the present investigation are unlikely to be due to test bias. CASI items include identification of birth place; birth year; date of birth; age; number of minutes in an hour; direction of the sunset; instant, short-term delayed, and medium-term delayed recall of three items; digits backward; serial subtractions; identification of the year, month, date, day of the week, and season; spatial orientation; number of animals named in 30 seconds; similes; judgment; repeating phrases; reading and following a written command; writing a specific sentence; copying a figure of interlocking pentagons; following a three-step verbal command; naming body parts; naming objects; and immediate recall of the objects when covered. The full scale has been published (Teng et al., 1994).

Standard CASI scores were used to identify participants for further evaluation to detect cases of dementia at Examinations 4–7. At Examination 4 (1991–1993), a stratified sampling scheme was used. A subgroup of participants were identified for full dementia evaluation based on results of CASI scores, age, education, and scores from the Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE) (Jorm & Jacomb, 1989). At Examination 5 (1994–1996), those who scored less than an education-adjusted CASI cut-point (77 for participants with low education and 79 for those with high education) or who had an absolute drop of ≥9 points in CASI score were selected for full dementia evaluation. At Examinations 6 (1997–1999) and 7 (1999–2000), those who scored <70 points on the CASI were selected for full dementia evaluation. At all HAAS Examinations, the dementia evaluation included the Consortium to Establish a Registry for AD neuropsychological battery (Morris et al., 1989) and a neurological examination. A consensus diagnosis was reached by the study neurologist and at least two other physicians with expertise in geriatrics and dementia. The panel used the Diagnostic and Statistical Manual of Mental Disorders, 3rd edition, revised criteria (Association, 1987) to diagnose dementia.

We employed IRT scoring (Embretson & Reise, 2000; Hambleton, Swaminathan, and Rogers, 1991; McDonald, 1999) in these analyses. We used Parscale (Muraki & Bock, 2003) to obtain IRT CASI scores. Many CASI items are polytomous (two or more response categories). We used the graded response model (Samejima, 1969, 1997), a polytomous extension of the two-parameter logistic model, to calibrate CASI items from HAAS baseline for the nondemented population. We applied these item parameters to CASI item responses at each Examination to determine IRT CASI scores. We rescaled all IRT scores to have a mean of 100 and a standard deviation of 15 (Mungas, Reed, and Kramer, 2003).

Assessment of Spoken and Written Japanese Language Use

We considered two different categorizations of Japanese language use for the two main analyses in this article.

Group Categorization for Model 1.—

At HHP Examination 1 in 1965–1968 (ages 45–68 years), participants were asked two questions regarding Japanese language use. They were asked “Do you speak Japanese?” and “Do you read or write Japanese?” with response options for both questions of no, very little, fair, well, and unknown. We divided participants into three groups based on responses to these questions (dichotomized as no/very little vs fair/well). Group 1 was participants who indicated that they spoke at most very little Japanese (“neither spoke nor read Japanese”; all except one indicated that they read or wrote at most very little Japanese, and the lone exception was excluded from these analyses). Group 2 was participants who indicated that they spoke Japanese at least fairly well but who read or wrote at most very little Japanese (“spoke but did not read Japanese”). Group 3 was participants who indicated that they both spoke and read or wrote Japanese at least fairly well (“both spoke and read Japanese”). Group 2 was our reference group for analyses.

Group Categorization for Model 2.—

A limitation of the HHP Examination 1 exposure questions is that specific forms of written Japanese (Hiragana, Katakana, and Kanji) were not specified. At Examination 5 in 1994–1996 (ages 74–95 years), additional questions were asked regarding language(s) spoken and how well participants could read and write in Japanese and English in the middle years of their lives (ages 40–50 years). One question addressed how well the participant could read Japanese. We dichotomized responses to this question to identify participants who indicated that they “could read most Japanese magazines, books, and newspapers” versus participants who at most “could read simple Japanese magazines, books, and newspapers.” A second question asked participants how well they could write Japanese. We dichotomized responses to this question to identify participants who indicated that they “could write well in Japanese in any form” versus participants who at most “could write in Hiragana or Katakana but not in Kanji.” We then combined responses to these items to identify participants who indicated that they could read Japanese at this level, write Japanese at this level, or both (“highly literate in Japanese”) versus participants who could neither read nor write Japanese at this level (“not highly literate in Japanese”). We combined this categorization with responses to a question about languages spoken in midlife to identify three groups: (a) participants who did not speak Japanese and were not highly literate in Japanese; (b) participants who spoke at least some Japanese but who were not highly literate in Japanese (reference category); and (c) participants who were highly literate in Japanese.

Assessment and Handling of Covariates

Years of education and income category were collected from participants at Examination 4. A blood sample was collected at Examination 4 and analyzed for APOE genotype (Havlik et al., 2000), categorized as presence versus absence of any APOE ϵ4. Smoking status information was collected at Examination 4 and coded as current versus prior versus never smoking. Head circumference is an indirect but robust measure of brain reserve (Borenstein Graves et al., 2001); it was measured at Examination 5. Age was centered by subtracting 70 years and treated in decades by dividing by 10.

Statistical Methods

We analyzed repeated IRT CASI scores using linear mixed-effects models, which combine population (“fixed”) and individual-specific (“random”) effects (Laird & Ware, 1982; Verbeke & Mollenberghs, 2000). All models included fixed effects of age and age squared with their corresponding random effects and specified an unstructured covariance structure among the random effects. Random intercepts and random slopes for age accounted for heterogeneity at baseline and in the rate of decline of cognitive functioning. We censored data from participants diagnosed with dementia; censoring began with the study visit that prompted a positive dementia evaluation.

For Model 1, the quantities of interest were the interactions between age and Japanese language use group and age squared and Japanese language use group. Our primary analysis model included terms for years of education and years of education squared, income, age at HAAS Examination 4, and smoking; these covariates were selected based on prior analyses (Yano et al., 2000). We also included APOE genotype and the number of times the participant had taken the CASI (1–4) to account for learning or familiarity effects.

For Model 2, we reestimated the same model with exposures determined from HAAS Examination 5 questions.

We performed a number of secondary analyses that included substitution of the standard CASI score as the outcome (Model 3), the inclusion of Kibei in the analytic sample and Kibei status as a covariate (Model 4), and the inclusion of head circumference as an additional covariate (Model 5). Next, we performed model selection on a more complex model that included interaction terms between all covariates in the primary model and the age and age-squared terms. We undertook model selection to identify parsimonious structures that sufficiently describe the data while allowing us to answer questions of interest to determine any differences from the primary model estimates and for possible use in future research. Model selection was performed using Akaike’s (1973) information criterion, the Bayesian Information criterion (Schwarz, 1978), and likelihood ratio tests. In the first selected model, we required that interactions between Japanese language use group and both age and age squared be retained. In the second selected model, we required only that the interaction between Japanese language use group and age be retained.

Finally, we assessed the potential influence of missing data using a multiple imputation framework (Little & Rubin, 1989). We imputed 10 data sets using multiple imputation by chained equations (Van Buuren & Oudshoorn, 2000). With this approach, we were able to define the imputation method for each variable independently by specifying a separate conditional density. We employed predictive mean matching to impute missing continuous data, logistic regression to impute missing binary data, and polytomous logistic regression to impute missing categorical data.

All statistical analyses were performed using R (R Development Core Team, 2005).

RESULTS

There were 2,520 participants with 6,899 valid CASI assessments (mean 2.7 per participant) in our primary analysis. Of these participants, 465 (18%) neither spoke nor read Japanese, 1,495 (59%) spoke Japanese only, and 560 (22%) both spoke and read Japanese. Demographic characteristics, smoking status, and APOE genotype are summarized in Table 1. Participants who both spoke and read Japanese were older than those in the other groups, and higher proportions were found in the extremes of education (both more poorly educated and more well-educated individuals) and income. Lower baseline CASI scores were more common among participants who spoke and read Japanese. Although current smoking status at baseline was similar across groups, participants who both spoke and read Japanese were less likely to have ever smoked.

Table 1.

Demographic Characteristics, Smoking Status, and Presence of APOE ϵ4 Alleles for Model 1 Sample by Exposure Status.

	Neither spoke nor read Japanese (n = 465)		Spoke Japanese but did not read Japanese (n = 1495)		Both spoke and read Japanese (n = 560)		Total (n = 2520)
Characteristic	M or N	SD or %	M or N	SD or %	M or N	SD or %	M or N	SD or %
Age at HAAS Examination 4
M (SD)	76.1	(3.5)	76.6	(3.6)	77.5	(4.4)	76.7	(3.8)
71–75	247	53%	709	47%	235	42%	1191	47%
76–80	162	35%	583	39%	195	35%	940	37%
81–85	52	11%	176	12%	98	18%	326	13%
86–93	4	1%	27	2%	32	6%	63	3%
Education
M (SD)	11.1	(2.9)	10.8	2.9	11.5	(3.6)	11.0	(3.1)
1–6	19	4%	61	4%	41	7%	121	5%
7–8	84	18%	322	22%	96	17%	502	20%
9–11	105	23%	352	24%	96	17%	553	22%
12	164	35%	499	33%	148	26%	811	32%
13+	93	20%	261	17%	179	32%	533	21%
Standard CASI score at HAAS Examination 4 (0–100 points)
M (SD)	89.0	(7.2)	87.4	(7.1)	87.0	(8.3)	87.6	(7.4)
<74	12	3%	47	3%	29	5%	88	3%
74–81	44	9%	207	14%	81	14%	332	13%
≥82	402	86%	1196	80%	426	76%	2024	80%
Missing	7	2%	45	3%	24	4%	76	3%
IRT CASI score at HAAS Examination 4
M (SD)	104.3	(13.8)	101.5	(13.7)	100.9	(14.8)	101.9	(14.0)
Annual income at HAAS Examination 4
<15,000	87	19%	313	21%	132	24%	532	21%
15,000–19,999	79	17%	330	22%	100	18%	509	20%
20,000–29,999	132	28%	383	26%	108	19%	623	25%
≥30,000	167	36%	469	31%	220	39%	856	34%
Smoking status at HAAS Examination 4
Never	155	33%	535	36%	244	44%	934	37%
Past	280	60%	853	57%	284	51%	1417	56%
Current	30	6%	107	7%	32	6%	169	7%
APOE ϵ4 alleles
None	381	82%	1228	82%	459	82%	2068	82%
1 or 2	84	18%	267	18%	101	18%	452	18%

Note: APOE = apolipoprotein E; CASI = Cognitive Abilities Screening Instrument; HAAS = Honolulu-Asia Aging Study; IRT = item response theory.

In our primary analysis (Model 1), there was no significant association between Japanese language use groups and the rate of cognitive decline; neither the interaction terms for exposure group with age or with age squared were statistically different from 0 (Table 2). Figure 2 graphically presents selected coefficient estimates and confidence intervals (CIs). Dots indicate point estimates for each parameter from the regression model, and lines represent 95% CIs around each parameter estimate. In Figure 2a, there is a slight quadratic effect for age and a clear linear effect for age, adjusting for all other covariates in Model 1. The effects of the other variables are all close to null, and all CIs overlap with 0, indicating no significant effect at the 95% confidence level. The overall effect of change over time in the group that both spoke and read Japanese was not significant (multivariate Wald test for the interactions with age and age squared, p = .82). Likewise, the overall effect of change over time in the monolingual English group was not significant (multivariate Wald test for the interaction with age and age squared, p = .74).

Table 2.

Linear Mixed Regression Models Predicting IRT-Based CASI scores^a

	Analysis 1 (Examination 1 language)	Analysis 2 (Examination 5 language)
	2,520 participants	2,029 participants
Covariate	Coefficient (95% CI)	Coefficient (95% CI)
Examination 1 or 5  language use
Spoke but did not read Japanese	0 (reference)	0 (reference)
Neither spoke nor read Japanese	1.94 (−0.95 to 4.84)	−2.31 (−4.90 to 0.28)
Both spoke and read Japanese	0.02 (−2.83 to 2.88)	−2.15 (−6.73 to 2.43)
Age	−12.24 (−17.53 to −6.95)	−12.66 (−18.19 to −7.12)
Age × Examination 1 or 5 language
Spoke but did not read Japanese	0 (reference)	0 (reference)
Neither spoke nor read Japanese	−1.54 (−7.76 to 4.69)	3.17 (−2.27 to 8.62)
Both spoke and read Japanese	−1.24 (−6.93 to 4.45)	1.97 (−6.72 to 10.67)
Age2	−1.96 (−3.60 to −0.33)	−1.11 (−30.41 to 4.26)
Age2 × Examination 1 or 5 language
Spoke but did not read Japanese	0 (reference)	0 (reference)
Neither spoke nor read Japanese	1.07 (−2.18 to 4.33)	−1.74 (−4.53 to 1.05)
Both spoke and read Japanese	0.77 (−1.99 to 3.54)	−0.59 (−4.67 to 3.49)

Note: CASI = Cognitive Abilities Screening Instrument; CI = confidence interval; IRT = item response theory.

^aModels were adjusted by including random effects for age and age squared and fixed effects for exposure category, age at Honolulu-Asia Aging Study (HAAS) Examination 4, education, education squared, income category, smoking status, apolipoprotein E ϵ4, and the number of tests taken. CASI scores are IRT based. Age effects are for 10 years, centered at age 70. IRT scores were calibrated such that the mean of nondemented HAAS participants at baseline was 100 and the standard deviation was 15, similar to IQ scores. Over a decade, a man in the reference group is predicted to decline 12.24 points (almost 1 SD). The decline for men in the other two groups was not statistically different than this value (CIs for both the linear and quadratic interaction terms with exposure status include the null value of 0).

Figure 2.

(a) Estimates and 95% confidence intervals (CIs) for selected coefficients for predicting item response theory (IRT)-based Cognitive Abilities Screening Instrument (CASI) scores in Model 1. This figure plots the results shown in Table 2, where language use was defined by questions asked at Honolulu Heart Program Examination 1. The top two elements show that the intercept (the baseline cognitive function level) for those who neither spoke nor read Japanese was somewhat higher than that for the reference group (those who spoke but did not read Japanese, not shown in the figure) and for the group that both spoke and read Japanese. Both these intervals cross the null value of 0, indicating no statistically significant effect. The third element shows that over a decade, the cohort was expected to decline about 12 points, almost 1 SD. CIs for the interactions between language groups and age both cross the null value of 0. The age-squared term indicates that on average, older individuals at baseline were expected to have somewhat faster trajectories of cognitive decline. The last two elements show the interactions between language groups and age squared. CIs for both these interaction terms cross the null value of 0. (b) Estimates and 95% CIs for selected coefficients for predicting IRT-based CASI scores in Model 2. This figure plots the results shown in Table 3, where language use was defined by questions asked at Honolulu-Asia Aging Study Examination 5. The top two elements show that the intercept (the baseline cognitive function level) for those who spoke but did not read Japanese (the reference group, not shown here) was modestly higher than that for either the group who neither spoke nor read Japanese (the top element) or the group who both spoke and read Japanese. The 95% CIs around these parameter estimates include the null value of 0, indicating that neither of these differences reached statistical significance. The third element shows that over a decade, the cohort was expected to decline about 12 points, almost 1 SD. CIs for the interactions between language groups and age both cross the null value of 0. The age-squared term indicates that on average, older individuals at baseline were expected to have modestly faster trajectories of cognitive decline, although in this case the 95% CI for this parameter included the null value of 0. The last two elements show the interactions between language groups and age squared. CIs for both these interaction terms cross the null value of 0.

To frame this result in more interpretable terms, we compared the expected declines among typical participants in different Japanese language use groups. Compared with the reference group of those who spoke but did not read Japanese, those who both spoke and read Japanese are expected to decline 0.3 points less for a period of 10 years from age 75 to age 85 (95% CI = 1.4 points more to 2.1 points less). The IRT scores were scaled to have a standard deviation of 15 points, so the expected difference of 0.3 points between groups for a period of 10 years from age 75 to age 85 corresponds to 2% of a standard deviation. Compared with the reference group, the monolingual English group is expected to decline 0.6 points less for a period of 10 years from age 75 to age 85 (1.4 points more to 2.6 points less). This expected difference of 0.6 points between groups for a period of 10 years from age 75 to age 85 corresponds to 4% of a standard deviation.

Additionally, we fit a similar model on a subset of HAAS participants who had more specific Japanese language exposure data from Examination 5 and at least one subsequent follow-up visit. There were 2,029 participants with 6,365 valid CASI assessments (mean 3.1 per participant) in this secondary analysis. Of these participants, 603 (30%) did not speak Japanese and were not highly literate in Japanese, 1,258 (62%) spoke some Japanese but were not highly literate in Japanese, and 168 (8%) were highly literate in Japanese, which meant that they could read the hundreds of ideographs required for Kanji. Model results were similar when we used this different definition of exposure, also shown in Table 2 and Figure 2b. Compared with the reference group of those who spoke some Japanese but were not highly literate in Japanese, those who were highly literate in Japanese are expected to decline a statistically and clinically insignificant 0.8 points less in the decade from 75 to 85 years of age (95% CI = 1.9 points more to 3.5 points less), which corresponds to 5% of a standard deviation. Compared with the reference group, monolingual English speakers are expected to decline a statistically and clinically insignificant 0.3 points more (2.0 points more to 1.4 points less), which corresponds to 2% of a standard deviation.

We performed several sensitivity analyses, including using standard CASI scores rather than IRT scores, including the Kibei, and including head circumference as a covariate. All these analyses produced similar results and did not alter our conclusion that there was no significant difference in the rate of cognitive decline associated with speaking Japanese or speaking and reading Japanese (see Table 3). Our model selection process also resulted in similar estimates and similar models. The selected models retained the required interactions terms, the covariates in the primary model, and an interaction term between age and APOE ϵ4 (see Table 4). To further illustrate our null findings, we used the left column of Table 4 to calculate scores at baseline and 10 years later, with baseline ages 70 and 80 years, for representative men in each of the three language groups (see Figure 3). At baseline, the reference group of speaking but not reading Japanese has a moderately higher predicted cognitive functioning score than either of the other language groups. The rate of decline in each group appears entirely comparable. Certainly, it is difficult to argue that both speaking and reading Japanese (the blue lines) is associated with much reduced rates of cognitive decline than neither speaking nor reading Japanese (the red lines).

Table 3.

Secondary Analyses: Linear Mixed Regression Models for CASI Scores^a

	Model 3: standard CASI scores	Model 4: include Kibei participants	Model 5: include head circumference
	2,520 participants	2,763 participants	2,207 participants
Covariate	Coefficient Est. (95% CI)	Coefficient Est. (95% CI)	Coefficient Est. (95% CI)
Language use
Spoke but did not read Japanese	0 (reference)	0 (reference)	0 (reference)
Neither spoke nor read Japanese	0.78 (−0.79 to 2.36)	2.04 (−0.86 to 4.92)	2.34 (−0.67 to 5.35)
Both spoke and read Japanese	0.06 (−1.48 to 1.61)	−0.80 (−3.43 to 1.83)	−0.73 (−3.51 to 2.06)
Age	−6.23 (−9.63 to −2.83)	−13.21 (−18.22 to −8.20)	−11.35 (−16.68 to −6.02)
Age × Language use
Spoke but did not read Japanese	0 (reference)	0 (reference)	0 (reference)
Neither spoke nor read Japanese	−0.06 (−3.98 to 3.86)	−1.67 (−7.89 to 4.55)	−1.85 (−8.20 to 4.51)
Both spoke and read Japanese	−0.85 (−4.44 to 2.73)	0.10 (−5.10 to 5.30)	−0.53 (−5.92 to 4.87)
Age2	−3.11 (−4.25 to −1.96)	−1.72 (−3.34 to −0.10)	−1.8 (−3.45 to −0.14)
Age2 × Language use
Spoke but did not read Japanese	0 (reference)	0 (reference)	0 (reference)
Neither spoke nor read Japanese	0.19 (−2.07 to 2.46)	1.14 (−2.10 to 4.38)	1.24 (−2.03 to 4.52)
Both spoke and read Japanese	0.67 (−1.28 to 2.63)	0.44 (−2.08 to 2.96)	0.98 (−1.61 to 3.57)

Note: CASI = Cognitive Abilities Screening Instrument; CI = confidence interval; Est. = estimate.

^aThe group that spoke but did not read Japanese served as the reference group for each model. CASI scores were item response theory based for Models 4 and 5. Age effects are for 10 years, centered at age 70. See explanatory note to Table 2.

Table 4.

Linear Mixed Regression Model Selection for IRT-Based CASI Scores^a

	Selected model requiring interactions of language use group, age, and age2	Selected model requiring interactions of language use group and age
Covariate	Coefficient (95% CI)	Coefficient (95% CI)
(Intercept)	78.58 (73.70 to 83.46)	78.86 (74.04 to 83.70)
Examination 1 or 5 language use
Spoke but did not read Japanese	0 (reference)	0 (reference)
Neither spoke nor read Japanese	1.92 (−0.96 to 4.80)	1.20 (−0.72 to 3.12)
Both spoke and read Japanese	0.09 (−2.75 to 2.94)	−0.56 (−2.47 to 1.35)
Age at Examination 4	9.41 (5.12 to 13.70)	9.41 (5.12 to 13.70)
Education (y)	2.90 (2.14 to 3.67)	2.91 (2.14 to 3.67)
Education2 (y)	−0.07 (−0.10 to −0.040)	−0.07 (−0.10 to −0.040)
Annual income at Examination 4 ($)
<15,000	0 (reference)	0 (reference)
15,000–19,999	2.16 (0.84 to 3.49)	2.17 (0.84 to 3.49)
20,000–29,999	3.67 (2.40 to 4.93)	3.67 (2.40 to 4.94)
≥30,000	4.60 (3.37 to 5.83)	4.60 (3.37 to 5.83)
Smoking status at Examination 4
Never	0 (reference)	0 (reference)
Past	1.03 (0.14 to 1.92)	1.03 (0.13 to 1.92)
Current	0.91 (−0.87 to 2.68)	0.90 (−0.87 to 2.67)
0 APOE ϵ4 alleles	0 (reference)	0 (reference)
1 or 2 APOE ϵ4 alleles	0.62 (−1.31 to 2.54)	0.60 (−1.32 to 2.53)
No. of test taken
1st	0 (reference)	0 (reference)
2nd	−1.51 (−2.8 to −0.18)	−1.50 (−2.83 to −0.18)
3rd	−0.84 (−3.42 to 1.74)	−0.84 (−3.42 to 1.74)
4th	0.47 (−2.79 to 3.74)	0.46 (−2.80 to 3.73)
Age	−11.81 (−17.10 to −6.52)	−12.61 (−17.50 to −7.71)
Age2	−2.02 (−3.64 to −0.40)	−1.59 (−2.82 to −0.37)
Age × Examination 1 or 5 language
Spoke but did not read Japanese	0 (reference)	0 (reference)
Neither spoke nor read Japanese	−1.52 (−7.712 to 4.67)	0.44 (−1.44 to 2.32)
Both spoke and read Japanese	−1.40 (−7.06 to 4.26)	0.29 (−1.45 to 2.03)
Age × 1 or 2 APOE ϵ4 alleles present	−1.96 (−3.80 to −0.12)	−1.94 (−3.78 to −0.10)
Age2 × Examination 1 or 5 language
Spoke but did not read Japanese	0 (reference)
Neither spoke nor read Japanese	1.07 (−2.17 to 4.31)
Both spoke and read Japanese	0.87 (−1.88 to 3.61)

Note: APOE = apolipoprotein E; CASI = Cognitive Abilities Screening Instrument; CI = confidence interval; IRT = item response theory.

^aAge effects are for 10 years, centered at age 70.

Figure 3.

Model-predicted trajectories of cognitive function for representative Japanese-American men from the Honolulu-Asia Aging Study. We used the regression results shown from the selection model (left column of Table 4) to determine item response theory (IRT)-based CASI scores for men entering the study at age 70 (top three solid lines) and at age 80 (bottom three dashed lines). All six groups had 10 years of education, were nonsmokers, had an income from $20,000 to 29,999, and had no apolipoprotein E ϵ4 alleles. We used these values in Excel to obtain model-predicted IRT CASI scores at baseline and then 10 years later (when men in the first group would be 80 and men in the second group would be 90). Men who neither spoke nor read Japanese (depicted in black) had a modestly higher cognitive functioning level at baseline compared with both the group that spoke but did not read Japanese (depicted in red) and the group that both spoke and read Japanese (depicted in blue). This “advantage” of neither speaking nor reading Japanese was maintained consistently across a decade. *CASI = Cognitive Abilities Screening Instrument. The IRT scores were scaled such that the mean score for nondemented individuals at baseline in the Honolulu-Asia Aging Study had a mean of 100 and a standard deviation of 15. The representative men depicted here are predicted to decline roughly a full standard deviation over a decade. This anticipated decline is not affected by whether the men spoke or read Japanese.

Finally, we used multiple imputation to evaluate sensitivity of our primary analysis (Model 1) results to missing data. Missing data were a particular concern in this analysis as a potential analytic sample of 9,462 observations for 2,986 cognitively normal participants was reduced to 6,899 observations for 2,520 participants due to missing cognitive test scores, exposure information, and covariate data. Results of the multiple imputation analyses suggest that those who both spoke and read Japanese are expected to decline 0.3 points more (2.0 points more to 1.4 points less) in the decade from 75 to 85 years of age on average than those who spoke but did not read Japanese. The multiple imputation results further suggest that monolingual English speakers are expected to decline 0.6 points less (1.2 points more to 2.5 points less) than the reference group over the same time period (see Table 5). (At the request of an anonymous reviewer, we performed additional analyses that strengthened our conclusion that our results were unlikely to be due to inflated error due to retaining insignificant terms in our models. Specifically, we reran analyses excluding random effects for the age and age-squared terms, and results were essentially unchanged. We also reran analyses excluding the age-squared direct effects and interactions with the age-squared term, and results were essentially unchanged.)

Table 5.

Results From Multiple Imputation for Model 1, Predicting IRT-Based CASI Scores^a

Covariate	Coefficient (95% CI)
(Intercept)	79.62 (75.21 to 84.04)
Examination 1 language use
Spoke but did not read Japanese	0 (reference)
Neither spoke nor read Japanese	2.01 (−0.93 to 4.96)
Both spoke and read Japanese	−0.52 (−3.48 to 2.44)
Age at Examination 4	−10.94 (−16.34 to −5.55)
Education (y)	2.94 (2.26 to 3.62)
Education2 (y)	−0.073 (−0.10 to −0.04)
Annual income at Examination 4 ($)
<15,000	0 (reference)
15,000–19,999	2.43 (1.00 to 3.85)
20,000–29,999	3.70 (2.25 to 5.16)
≥30,000	5.06 (3.72 to 6.41)
Smoking status at Examination 4
Never	0 (reference)
Past	0.52 (−0.27 to 1.31)
Current	−0.60 (−2.18 to 0.98)
0 APOE ϵ4 alleles	0 (reference)
1 or 2 APOE ϵ4 alleles	−1.27 (−2.27 to −0.28)
No. of test taken
1st	0 (reference)
2nd	−6.92 (−8.5 to −5.34)
3rd	−12.14 (−15.39 to −8.90)
4th	−13.92 (−18.14 to −9.69)
Age	4.78 (−1.77 to 11.32)
Age2	−0.89 (−2.60 to 0.81)
Age × Examination 1 language use
Spoke but did not read Japanese	0 (reference)
Neither spoke nor read Japanese	−2.32 (−8.52 to 3.89)
Both spoke and read Japanese	−0.46 (−6.32 to 5.40)
Age2 × Examination 1 language use
Spoke but did not read Japanese	0 (reference)
Neither spoke nor read Japanese	1.48 (−1.65 to 4.60)
Both spoke and read Japanese	0.079 (−2.62 to 2.78)

Note: APOE = apolipoprotein E; CASI = Cognitive Abilities Screening Instrument; CI = confidence interval; IRT = item response theory.

^aAge effects are for 10 years, centered at age 70.

DISCUSSION

We did not find a relationship between use of spoken or written Japanese in midlife and protection from cognitive decline in late life. These findings were consistent using two methods of defining exposure—based on less specific questions that were available for the entire cohort and on more specific questions that were available for a subset that specifically addressed the ability to read the hundreds of ideographs of the Kanji form—and when including or excluding participants who had some education in Japan, including or excluding head circumference from our regression models, and including only linear or both linear and quadratic relationships with age. We also determined that our findings were not likely to be due to missing data. These results were not consistent with what would have been predicted based on the reserve hypothesis.

We previously found that there was no relationship between spoken or written Japanese use and protection from dementia outcomes (any dementia, AD, or vascular dementia) in HAAS (Crane et al., 2009). The present study complements those findings by examining whether cognitive reserve associated with use of spoken or written Japanese may protect from cognitive decline.

We limited these analyses to second-generation sons of Japanese immigrants (Nisei), and we excluded Kibei participants who had some of their education in Japan to minimize potential confounding. Results when including and excluding the Kibei were essentially unchanged, suggesting that there was minimal effect of mastery of Japanese language, whether that mastery was obtained in Japan or in Hawaii.

Families who chose to enroll their sons in special Japanese language schools may have had different priorities than families that did not make that choice. These data provide us with the opportunity to determine whether the choice to study Japanese language among children of Japanese immigrants might have a relationship with rates of cognitive decline decades later. We found no evidence that use of spoken or written Japanese in midlife had any impact on rates of cognitive decline. There are obviously many reasons one might choose to learn and use the language of one’s ancestors or any second language. These data suggest that prevention of cognitive decline should not be part of the rationale for encouraging second language acquisition.

The multilingualism construct evaluated here is related to the mental exercise (Salthouse, 2006) and adult enrichment (Hertzog, Kramer, Wilson, and Lindenberger, 2009) constructs discussed in detail elsewhere. Salthouse (2007) notes that longitudinal data are needed to address questions related to cognitive aging, and no study will have a perfect design. Regarding the relationship between multilingualism with Japanese in general, and with written Japanese in particular, the present study—especially in conjunction with our prior study of dementia outcomes (Crane et al., 2009)—suggests that there may be limited benefits in terms of the rate of global cognitive decline or delay in dementia onset anticipated in late life. Unlike previous studies (Bialystok et al., 2007; Kave et al., 2008), the present investigation includes a prospective cohort design, followed participants longitudinally prior to developing dementia, and used appropriate statistical methods. This does not eliminate the possibility of a protective effect on some specific cognitive domain or domains that the CASI may be too insensitive to detect. Nevertheless, the present study suggests that any such specific cognitive effect on longitudinal decline either does not exist or is small enough to be overwhelmed by noise in other cognitive domains measured by the CASI. Furthermore, our prior study suggests that any such specific cognitive effect is insufficient to meaningfully delay the onset of AD or dementia (Crane et al., 2009).

Several limitations should be considered for these analyses. Japanese language use was assessed only through self-report. We did not have additional acculturation, vocabulary, or reading-level data (Manly, Jacobs, Touradji, Small, and Stern, 2002). The CASI measures several cognitive domains but is too brief to address changes in any specific cognitive domain. Although we used two complementary exposure definitions, a more robust measure of exposure—such as measured written and spoken Japanese skill in midlife—would have strengthened our confidence in our findings. We do not have measures of frequency of exposure to written Japanese. This study was limited to men. For these reasons, these analyses should be considered exploratory.

Another limitation was the amount of missing data, which could lead to biased estimates, with the amount and direction of the bias impossible to predict. Results from sensitivity analyses using multiple imputation did not alter our conclusions (see Table 5). However, given the magnitude of missing data, it would be wise to approach the results with caution.

In summary, we did not find evidence to support the hypothesis that multilingualism is associated with cognitive reserve. Specifically, we found that neither use of spoken Japanese nor use of written Japanese was associated with reduced rates of cognitive decline in late life.

FUNDING

Supported by an Alzheimer's Association Investigator Initiated Research Grant (to P.K.C., PI). Additional support was provided from grants from the National Institute on Aging P50 AG 05136 (to L.E.G.) (M. Raskind, PI) and R01 AG 029672 (to P.K.C., J.G., E.E., and L.E.G.) (P.K.C., PI). Data collection was supported by the National Institute on Aging through Contract N01-AG-4-2149, grant U01-AG-1-9349-01, grant R01 AG 17155, and grant RO1 AG 19349 and by the National Heart, Lung, and Blood Institute through Contract N01-HC-05102.