Age of Migration and Cognitive Function Among Older Latinos in the United States

Data and study population

This study uses 5 waves of biennial data (2006-2014) from the Health and Retirement Study (HRS), a nationally representative sample of adults 50 years and older in the US which initially began in 1992. The HRS sampling has a multi-stage area probability design, including geographical stratification and clustering, with an oversample of Black and Latino households. Collection of interviews occurs through a multi-modal design, inclusive of primary respondent interviews and proxy interviews when necessary. Response rates across waves of HRS have remained relatively high (∼81% –89%). More detailed information regarding the HRS sample is published elsewhere [42, 44]. Our choice of study period is guided by two factors: 1) we wanted to reduce the potential for loss to follow up as a result of death for individuals who meet research criteria for dementia diagnosis during the study period. While estimates vary depending on type of dementia, severity stages, age at diagnosis, sex, and potentially race/ethnic background, a systematic review suggests a range of between 1 and 7 years [44]. As such, four biennial waves (8-years; 2006–2014) was a reasonable upper limit for longitudinal analyses. As importantly, 2) we wanted to validate our primary results using a different set of waves that stretched an equivalent and non-overlapping time period. To do so, in sensitivity analyses, we re-estimated our models using HRS data from 1998 to 2006 (4-waves). The results were consistent with the reported main findings in this study (see below, and sensitivity).

We take advantage of several HRS resources (https://hrs.isr.umich.edu/data-products) to construct our analytic sample for the current investigation; specifically, the HRS tracker file and the RAND HRS Longitudinal files [45]. The baseline (2006) sample consisted of unweighted n (u.n) = 17,106 unique respondents 50-years and older that were derived from the 2006 wave of the HRS. Given the design of the study, we focus on individuals who self-reported race/ethnic background as non-Latino Whites or Latinos and exclude foreign-born Whites and individuals that reported race/ethnic background other than non-Latino White or Latino (u.n = 13,817). Furthermore, we excluded 4% of participants (u.n = 551) due to missingness on baseline covariates. The variables with the highest prevalence of missing values were migration status (1.36%) and body mass index (BMI; 1.48%). Participants with missing values on any of the baseline covariates were more likely to be female (65.7% versus 53.5%), slightly older (66.1 versus 63.8 years), and had higher reported levels of education (16.3 years versus 13 years).

A detailed scheme showing the exclusion criteria and the analytic sample inclusion is provided in Supplementary Figure 1. Since most attrition in the HRS is a result of death, we used the HRS tracker file to construct an indicator measure capturing whether respondents were 1) alive and responding by end of each wave respectively; 2) attrited from the sample as a result of death; and/or 3) were lost to follow-up but not due to death. A table detailing the rates and causes of attrition in the HRS in the study period is provided as a supplement (Supplementary Table 1). We include death as an outcome category in analyses of longitudinal cognitive classification to minimize bias arising from attrition. All analyses accounted for the complex design of the HRS, including probability weighting, stratification, and clustering to allow appropriate generalization and correct inferences relative to the target population [42]. Given the scope of this study we used the baseline weights reflective of the target population of individuals who were not institutionalized or incarcerated at baseline. Data included in analyses are publicly available data, which received institutional review board (IRB) approval from the University of Michigan [HUM00061128]; therefore, the corresponding authors’ IRB determined our research was exempt from IRB consideration. Below we describe our analytic sample in more detail.

Outcome(s)

Our primary baseline and longitudinal analyses focus on cognitive functioning based on an abbreviated version of the Telephone Interview of Cognitive Status (TICS-M) [46–49]. The HRS TICS-M includes 10 word immediate and delayed recall tests of memory, a serial 7s subtraction test for working memory, and counting backwards to assess attention and processing speed. We exclude the object naming items used to assess language, and the probes for date and president/vice-president used to assess orientation as these were only administered to respondents 65 years and older. The composite score values were normally distributed (see Supplementary Figure 2) and ranged between 0 and 27. In secondary analyses we independently model the 1) memory (sum of immediate and delayed recall; range = 0–20) and 2) processing speed, working memory, and attention subcomponents (total serial 7s and backward counting from 20; range = 0–7) to assess potential education induced bias on cognitive performance. Finally, we consider baseline and longitudinal (i.e., 8-years later) cognitive status classification based on the Langa-Weir (LW) algorithm. The cut points for the LW classification are derived through extrapolations from more comprehensive adjudication processes using data from the Aging, Demographics, and Memory Study (a subsample of the HRS) and allow distinction between a) normal cognition; b) cognitive impairment no dementia (CIND); and c) dementia. LW classifications have been used extensively [9, 19]; details and rationale for the underlying algorithm are published elsewhere [9, 50–53]. For the longitudinal classification outcome, we include an additional category to group individuals that died between the baseline wave and 2014 (1 = normal cognition; 2 = CIND; 3 = dementia; and 4 = deceased). The outcome is set to missing if respondent status indicates loss to follow-up [35, 54].

Exposure

The primary exposure is a measure constructed to distinguish US-born NLWs, USBLs, and FBLs and capture life course migration patterns among FBLs. This was coded as follows: 0 = NLWs; 1 = USBLs; 2 = FBLs who migrated to the US prior to age 18 (FBLs < 18); 3 = FBLs who migrated to the US between 18 and 34 years (FBLs18–34); and 4 = FBLs who migrated to the US after age 35 (FBLs35+).

Baseline covariates

Our conceptual framework is guided by two theoretical underpinnings. First, social determinants of health, and education in particular, are critical for explaining differences between groups. Second, we are also guided by a lifecourse perspective which links early life factors to late-life health of immigrants and underserved groups in general. The variables that we chose to control for, given availability in the HRS, are reflective of these conceptual frameworks. In line with previous work [12, 23], we account for a series of measures that capture potential differences in sociodemographic and economic characteristics as well as comorbid health conditions, and health behaviors known to affect cognitive functioning. Sociodemographic and economic measures included: (a) participant sex (0 = male; 1 = female); (b) age in years at baseline interview; (c) educational attainment, based on the numbers of years of schooling completed; (d) marital status using a binary measure for whether the respondent was married or partnered versus not (married but spouse absent, separated, divorced, widowed, and never married); (e) and health insurance status using a 4-category indicator that differentiates participants with 1) government supplied insurance only; 2) government-supplied and private insurance; 3) private insurance only; and 4) no health insurance. We also account for geographical residence using a 4-category census region indicator: a) Northeast; b) Midwest; c) South; and d) West. For chronic comorbid health conditions, we used the count of eight self-reported conditions based on history of diagnoses including high blood pressure, diabetes, cancer, lung disease, arthritis, stroke, heart, and psychiatric problems. These conditions were included since they account for the largest share of health, social, and economic burden in the US [55]. Additionally, we included BMI (kg/cm) as another health condition as it has been shown to influence cognitive aging processes [56, 57]. Health behavior measures known to influence cognitive aging processes were also included; namely, smoking status, and alcohol consumption [58]. BMI was mean centered and entered into the models as a continuous linear measure. In supplementary models (available upon request) we accounted for BMI as a categorical variable using four groups (underweight, typical, overweight, and obese), as well as using quadratic and cubic forms. The results were unchanged. Smoking status was a binary measure indicating whether the respondent had a history of smoking. Alcohol consumption was ascertained using a binary self-reported measure of ever drinking alcoholic beverages. Finally, we also controlled for a count of the number of waves participated by an individual. We did not adjust for language of interview (Spanish versus English) as this measure is completely confounded with the NLW referent group.

Analytic approach

Our analyses were executed in six steps. First, we generated descriptive statistics to characterize the target population at baseline year by race/ethnicity, nativity and age of migration. We used survey adjusted chi-squared tests to assess group differences over categorical indicators (e.g., marital status) and survey-adjusted two-sided t-tests to assess differences in continuously measured indicators (e.g., age, education). Estimates can be found in Table 1.

Second, we fit linear models to examine baseline global and domain-specific (e.g., memory, processing speed and working memory) performance as a function of the primary exposure and sequentially adjusting for model covariates. This modeling approach takes advantage of our outcomes’ composite summed score and includes a pooled cross-sectional orientation of the data. Specifically, we estimate (1) crude, (2) age and sex adjusted, (3) additional adjustment for education, and (4) full covariates adjusted models. Beta coefficients, standard errors, and p-values for the estimated parameters are presented in Table 2A. To facilitate interpretation of the results we estimate and plot post-hoc average marginal effects (AMEs) derived from the sequentially adjusted models and their corresponding 95% confidence intervals (see Fig. 1).

OPEN IN VIEWER

Fig. 1.

Estimated (a) average marginal means of baseline global and domain specific cognitive function and (b) marginal probabilities of baseline Langa-Weir cognitive status classification (Normal, CIND, and Dementia) by race/ethnicity, nativity, and age of migration. Note: NLW refers to US-Born (USB) non-Latino Whites; USBL refers to USB Latinos; <18L refers to Foreign Born (FB) Latinos who migrated prior to age 18; 18–34L refers to FB Latinos who migrated between the ages of 18–34 years; 35+L refers to FB Latinos who migrated at age 35 and older Full Adjustment includes age, sex, education, marital status, insurance, census region, alcohol consumption, smoking, BMI, count of number of chronic conditions, and number of participated waves.

Third, we fit survey multinomial logit regression models to examine the association between baseline cognitive classification, using the normal cognitive group as the reference, and the exposure and sequentially adjusting for covariates of interest as described in step 2 above. This modeling approach takes advantage of the Langa-Weir classification cut-points indicating: a) normal cognition (reference), b) CIND, c) dementia, and d) deceased. Additionally, this modeling allowed us to assess model performance and identify any trends using the full analytic sample (2006–2014), by applying a pooled cross-sectional orientation to the data. The estimated relative risk ratios derived from these models and corresponding 95% confidence intervals are presented in Table 2B. We estimate and plot post-hoc average marginal probabilities of classification in each of the cognitive status groups derived from the sequentially adjusted models and corresponding 95% confidence intervals (see Fig. 1).

Fourth, we used latent growth curve models (LGCM) to test longitudinal associations between race/ethnicity, nativity and age of migration (NLWs, referent; USBLs, FBLs < 18; FBLs18–34; and FBLs35+) and the trajectory of global cognition (TICS-M) (Table 3A). Although the cross-sectional models described above are useful for testing associations between key measures of interests and the baseline outcome measures, these models are insufficient to clarify whether and to what extent nativity and age of migration influence cognitive change and prospective cognitive status. As with the above modeling approach, we sequentially adjusted for baseline measured covariates. LGCMs are conceptually and, in most cases, mathematically equivalent to mixed effects models of longitudinal data [59]. Detailed theoretical and applied treatments of latent growth models are provided elsewhere [60–63]. Briefly, under LGCM the repeated measures are modeled using wide data format within a structural equation framework [59, 61]. Constraints are imposed on the loadings of latent factors to model the observed variables (in this case the biennially measured cognitive outcomes) as a parametric or non-parametric function of time. In our case two, latent variables were estimated: an intercept and a slope; also, an estimate of the covariance between these two variables. Each latent variable is summarized by two parameters: a mean (the fixed effect) and a variance (the random effect). The estimated mean of the intercept is interpreted as the average cognitive score for the target population at the baseline. The estimated slope mean quantifies the average positive or negative change in cognitive scores over time. The estimated variance of the intercept is a measure of deviations of participant baseline scores from the estimated population mean. The variance of the slope measures average deviations in growth trajectories about the slope mean. A statistically significant variance estimate suggests that individual trajectories are not equivalent across participants. Since the intercept and slope are estimated as latent variables, rather than parameters, LGCM affords the flexibility of modeling these factors as a function of covariates. The interpretation of the associations between the covariates and the factors follow standard regression models [60, 61]. A linear function provided good fit to the data, and all models were estimated as such. In sensitivity analyses, we re-estimated the LCGM without imposing a parametric functional form on time (e.g., linear). Results were qualitatively unchanged. We used years since baseline wave as the time metric and estimate the models using full information maximum likelihood (FIML) models to account for all available outcome data. All LGCM models were fit using Mplus8.3 [64]. Supplementary Figure 3 provides a conceptual representation of the estimated model, and Table 3A includes the estimated fixed and random effects (variance estimates of growth parameters) and inferential statistics (SEs, and p-values). Evaluation of changes in the estimated variances/residual variances of the intercept and slope allows us to assess the contribution of the covariates to explaining individual differences in growth trajectories.

Fifth, we fit survey multinomial logit regression models to examine the association between a combined measure of cognitive classification and death status in 2014, using the normal cognitive group as the reference and the exposure and sequentially adjusting for covariates of interest as described in step 2 above. The estimated relative risk ratios derived from these models and their 95% confidence intervals are presented in Table 3B. As with step four above, we estimate and plot post-hoc average marginal probabilities of classification in each of the cognitive status groups and death derived from the sequentially adjusted models and corresponding 95% confidence intervals in Fig. 2.

OPEN IN VIEWER

Fig. 2.

Estimated average marginal probabilities of longitudinal (2014) Langa-Weir cognitive status classification (Normal, CIND, and Dementia) and death by race/ethnicity, nativity, and age of migration. Note: NLW refers to US-Born (USB) non-Latino Whites; USBL refers to USB Latinos; <18L refers to Foreign Born (FB) Latinos who migrated at age less than prior to age 18; 18–34L refers to FB Latinos who migrated between the ages of 18–34 years; 35+L refers to FB Latinos who migrated at age 35 and older. Full Adjustment includes age, sex, education, marital status, insurance, census region, alcohol consumption, smoking, BMI, count of number of chronic conditions, and number of participated waves.

Finally, we examined sex modification in baseline global cognitive performance, baseline cognitive impairment and longitudinal cognitive status, and death. To do so, we refit the sequentially adjusted models described in steps 2, 3, and 5 above and included an interaction term between sex and the race/ethnicity and age of migration indicator. Sex specific marginal mean estimates for global cognition, and marginal probabilities of baseline and longitudinal dementia classifications are presented in Figs. 3 and 4. Detailed estimates of sex specific post-hoc ANOVA contrasts (relative to NLWs) for the marginal means and probabilities and their corresponding p-values and 95% confidence intervals are included in Supplementary Tables 3 and 4.

OPEN IN VIEWER

Fig. 3

Sex specific average marginal estimates of mean baseline global cognitive function by race/ethnicity, nativity, and age of migration. Note: NLW refers to US-Born (USB) non-Latino Whites; USBL refers to USB Latinos; <18L refers to Foreign Born (FB) Latinos who migrated prior to age 18; 18–34L refers to FB Latinos who migrated between the ages of 18–34 years; 35+L refers to FB Latinos who migrated at age 35 and older. Full Adjustment includes age, education, marital status, insurance, census region, alcohol consumption, smoking, BMI, count of number of chronic conditions, and number of participated waves.

OPEN IN VIEWER

Fig. 4

Sex specific average marginal estimates of (a) prevalence of baseline dementia, and (b) prevalence of longitudinal (2014) dementia classification by race/ethnicity, nativity, and age of migration. Note: NLW refers to US-Born (USB) non-Latino Whites; USBL refers to USB Latinos; <18L refers to Foreign Born (FB) Latinos who migrated prior to age 18; 18–34L refers to FB Latinos who migrated between the ages of 18–34 years; 35+L refers to FB Latinos who migrated at age 35 and older. Full Adjustment includes age, education, marital status, insurance, census region, alcohol consumption, smoking, BMI, count of number of chronic conditions, and number of participated waves.

All analyses used the data in wide form. The linear models focusing on baseline outcomes used the 2006 cross-section as the baseline. The multinomial logit models focusing on cognitive impairment outcomes at the end year (2014) and the longitudinal growth models used wide form data to link baseline measures to end wave outcomes; only individuals who are in the universe of the baseline target population (given the use of baseline weights) are considered in these analyses.

Sensitivity analyses

We conducted several sets of analyses to ensure that our results are robust. First, we split the sample into two age groups (<70 and 70+) and examined group invariance in effects of race/ethnicity and age of migration groupings on the intercept and change parameters in the longitudinal growth models. Invariance testing (model with freely estimated effects of race/ethnicity and migration groups on growth parameters versus model with effects constrained to equality across the two age stratas) showed no decrement in fit by restricting the effects to be equal within the two groups. Second, we re-estimated all models using generalized estimating equations models (population average models that focus on the fixed effects of groups) using the overall study population (see description above) as well as the subsample of participants 70 years and older using time since baseline as the time metric. Our results indicate that the group differences reported are consistent for both target populations. Importantly, as with the results from the latent growth models the group differences were restricted to baseline differences, and the slopes of change (decline in cognition) over time were statistically equivalent. Third, we re-estimated all models using generalized estimating equations and a group by age interaction (to specify the latter as the time metric). We did so for both the overall and 70+ subpopulation. As with the second set of sensitivity analyses detailed above and the primary findings reported in the manuscript, we found largely equivalent group differences and that the main effects were specific to intercept differences between groups rather than differences in change over age (i.e., no significant differences in estimated age slopes).

Descriptive statistics (Table 1)

Participants reported a mean age of 66 years±9.8 with an average education of 13 years±2.8. Slightly more than half were female, two-thirds were married, and approximately 94% had health insurance. We found no group differences in sex or marital status. We found significant group variations in age, education, number of chronic conditions, and health behavior profiles among NLWs and Latino subgroups. Notably, FBLs35+ were older, had lower educational attainment, reported a higher number of chronic conditions, and exhibited lower levels of alcohol consumption and smoking behavior than NLWs and Latino subgroups. In addition, USBLs reported higher levels of educational attainment, alcohol consumption and smoking than their foreign-born counterparts. Detailed tests of differences (and their 95% confidence intervals) in characteristics among Latino groups relative to NLWs are presented in Supplementary Table 2.

Baseline cognitive function difference among NLWs and Latinos (Table 2A)

In age and sex adjusted models USBLs (β= –2.71; SE(0.22); p < 0.001), FBLs < 18 (β= –2.81; SE(0.28); p < 0.001), FBLs18–34 (β= –2.88; SE(0.46); p < 0.001), and FBLs35+ (β= –3.05; SE(0.59); p < 0.001) had lower cognitive scores relative to NLWs. Except for USBLs, (β= –1.23; SE (0.22); p < 0.001) group differences between all FBLs, irrespective of age of migration, were explained by adjusting for years of education. Additional adjustment for sociodemographic, health, and health behavior factors did not substantively attenuate the difference between USBLs and the other two groups (NLWs and FBLs). The estimated parameters for all foreign-born Latino groups remained statistically indistinguishable from NLWs. Similar results emerged when we modeled the memory and processing speed/attention domains separately. The effect sizes from these models were consistent with those derived for global cognition. Figure 1 provides a visualization of the estimated marginal means for each group and attenuation in the estimated means through adjustment for covariates.

Baseline cognitive impairment among NLWs and Latinos (Table 2B)

Similar to cognitive function, after adjusting for age and sex, USBLs (RRR = 3.57; 95% CI = [2.84;4.47]; p < 0.001), FBLs < 18 (RRR = 4.45; 95% CI = [3.50;5.66]; p < 0.001), FBLs18–34 (RRR = 3.51; 95% CI = [2.44; 5.03]; p < 0.001), and FBLs35+ (RRR = 3.52; 95% CI = [1.84;6.72]; p < 0.001) exhibited higher relative risk ratios for CIND in comparison to NLWs (see Fig. 1). Adjusting for education did not explain the difference between USBLs (RRR = 1.87; 95% CI = [1.41; 2.48]; p < 0.001) and NLWs. In contrast, years of education explained the difference between FBLs and NLWs, regardless of age of immigration. Additional adjustment for covariates did not reduce the estimated relative risk for USBLs (RRR = 1.78; 95% CI = [1.34;2.36]; p < 0.001). In age and sex adjusted models, all Latino groups were also at higher risk for dementia classification relative to NLWs; The estimated relative risk ratios were RRR = 4.23; 95% CI = [3.15;5.69]; p < 0.001, RRR = 4.52; 95% CI = [2.70; 7.57]; p < 0.001, RRR = 4.89; 95% CI = [2.69; 8.89]; p < 0.001, and RRR = 4.97; 95% CI = [2.73; 9.03]; p < 0.001 for USBLs, FBLs < 18, FBLs18–34, and FBLs35+, respectively. Adjusting for education completely attenuated the higher risk for dementia across all Latino groups relative to NLWs. Figure 1 provides a visualization of the estimated marginal probabilities for classification into (1) normal, (2) CIND, and (3) dementia for each group and attenuation in the estimated probabilities through adjustment for covariates.

Cognitive change (Table 3A)

The fixed effects estimates derived from the growth models were consistent with the results from the baseline regressions. The mean intercept was 16.0 (SE = 0.08; p < 0.001), and the average biennial decline (slope) estimate was –0.34 (SE = 0.01; p < 0.001). The random effects indicated significant and sizable variations around the mean estimates for the intercept and slope. The variances for the intercept and slope were 12.29 (SE = 0.33; p < 0.001) and 0.13 (SE = 0.02; p < 0.001), respectively. In age and sex adjusted models USBLs, FBLs < 18, FBLs18–34, and FBLs35+ had lower intercepts relative to NLWs. Apart from USBLs, group differences between all FBLs, irrespective of age of migration, were explained by adjusting for number of years of education. No additional attenuations in the estimates for USBLs were attained through additional covariates adjustment. The estimated slopes for all Latino groups were not significantly different from NLWs. This was independent of covariates adjustment. Latino ethnicity and age of migration explained 4% of the variance in the intercept. Adjusting for age and sex explained an additional 23.2% and education explained 16.1%. All other covariates explained an additional 5.5%. Age and sex adjustment explained 38.5% of the slope variance, but education did not contribute to slope variance explanation.

Longitudinal cognitive impairment (Table 3B)

In age and sex adjusted models USBLs (RRR = 3.30; 95% CI = [2.39; 4.56]; p < 0.001), FBLs < 18 (RRR = 3.80; 95% CI = [2.97; 4.87]; p < 0.001), FBLs18–34 (RRR = 4.32; 95% CI = [2.78; 6.72]; p < 0.001), but not FBLs35+, had higher relative risk ratios for CIND (versus normal cognition) relative to NLWs. After adjusting for education, there were no differences between NLWs and FBLs regardless of age of immigration. With the exception of USBLs (RRR = 1.85; 95% CI = [1.40; 2.44]; p < 0.001) adjusting for education explained the differences relative to NLWs. Additional adjustment for covariates slightly reduced the estimated relative risk for USBLs (RRR = 1.60; 95% CI = [1.17; 2.20]; p < 0.001). In age and sex adjusted models, all Latino groups were also at higher risk for dementia classification relative to NLWs; The estimated relative risk ratios were RRR = 5.31; 95% CI = [3.55; 7.94]; p < 0.001, RRR = 6.22; 95% CI = [3.82; 10.13]; p < 0.001, RRR = 5.23; 95% CI = [2.93; 9.34]; p < 0.001, and RRR = 4.70; 95% CI = [1.38; 16.05]; p < 0.05 for USBLs, FBLs < 18, FBLs18–34, and FBLs35+, respectively. Except for USBLs (RRR = 2.37; 95% CI = [1.62; 3.46]; p < 0.001), adjusting for education completely attenuated the higher risk for dementia across all Latino groups, relative to NLWs. Additional adjustment for covariates had minimal effects on the relative risk ratio estimates. Finally, in age and sex adjusted models USBLs and FBLs < 18 had higher relative risk of death (relative to survival with normal cognition). Again, these group differences were explained by adjusting for education. Figure 2 provides a visualization of the estimated marginal probabilities for classification into (a) normal cognition, (b) CIND, (b) dementia, and (d) death for each group and attenuation in the estimated probabilities through adjustment for covariates.

Sex modifications

We found consistent evidence supporting sex modifications in the associations between our primary exposure and baseline cognitive performance and impairment classification (Supplementary Tables 3A and 3B). Adjusting for covariates, Latinas, particularly USBLs, FBLs < 18, and FBLs35+ had more disadvantageous cognitive scores profiles relative to NLW females; whereas only USBL males exhibited poorer baseline global functioning compared to NLW males (Supplementary Table 3A, and Fig. 3). The higher rates of baseline dementia but not CIND among USBLs, and FBLs < 18 Latinas were explained by covariates adjustment (Supplementary Table 3B, also see Fig. 4A for prevalence estimates of baseline dementia classification). Sex did not significantly modify (insignificant test of interaction) the association between the primary exposure and longitudinal CIND and dementia classifications. Group contrasts tests for longitudinal classifications are presented in Supplementary Table 4, and dementia classification estimates in Fig. 4b.

Limitations

Our findings should be considered in the context of several limitations. First, the cognitive instruments used in HRS are brief and limited with regard to the level of detail and precision in measurement of cognitive function. These instruments may be impacted by bias due to demographic and sociocultural factors [82, 83], and dementia classification with these data have been shown to be less accurate for Latinos relative to non-Latino Whites [84]. Importantly, we were unable to account for direct measures of literacy and numeracy in this study. Recent findings [85] indicate that illiteracy is associated with an approximately three times higher risk of dementia. These limitations make it necessary for our findings to be replicated using detailed data that more precisely assess both fluid and crystallized knowledge, account for culturally appropriate cognitive measures, and include a broader assessment of cognitive domains. Despite this, the TICS-M and associated cognitive impairment classifications are used extensively in epidemiological studies to ascertain impairment and dementia prevalence and risks in the US [9, 84]. Second, the HRS is a biennial survey. More frequent interviews at smaller time intervals may capture more specific variations and nuances in cognitive change and possibly enhance precision of estimations. Third, the variables used to assess risk in our models are self-reported and in some instances crude and as such are subject to errors known to exist with self-reported data (e.g., recall bias, social desirability among others) [83, 86]. For example, evidence indicates that alcohol dosage may have differential effect on the risk of dementia, wherein low to moderate alcohol consumption may be protective but excessive alcohol consumption may elevate dementia risk. We used a crude measure for alcohol consumption which could have biased the estimates [87]. Future work, with more precise measures of behavioral risks, including alcohol, is needed to better clarify potential attenuations in effects. Fourth, we only focused on baseline risk factors in our longitudinal models. More complex models that account for time varying covariates can provide higher levels of precision in assessing the influence of some of the risk factors (e.g., comorbid conditions). Fifth, we did not include a measure capturing language of interview or respondent bilingualism because of confounding with the non-Latino White referent. Sixth, we were unable to consider participants’ country of origin (e.g., Mexicans and Puerto Ricans) and as such assume a homogenous risk profile across Latino subgroups. Existing studies suggest important variations in cognitive risks, comorbidities (e.g., higher incidence of diabetes among Mexican-origin Latinos) [14] and health behaviors profiles (e.g., higher rates of smoking among Puerto Rican-origin Latinos). Disaggregating Latino subgroups offer an important future line of research on differential risks within Latinos relative to non-Latino Whites. Finally, given the scope of the work we did not model discrete switching in cognitive status over time and focused only on baseline and endpoint estimates of CIND and dementia classification. Examining the correlates of maintenance of pre-clinical cognitive status (e.g., CIND), reversion to normal function, and mixed switching (e.g., CIND to normal and back) is critical to clarify the pathways from pre-clinical to clinical status. Future work should consider using longitudinal Latent Class Analyses or Latent Transition Models to better characterize these processes and their correlates.

Conclusion

US-born Latinos 50 years and older, regardless of sex exhibit poorer cognitive function than older non-Latino Whites and their foreign-born Latino counterparts. Social determinants (e.g., education) that affect cognitive reserve, particularly those that compensate for education as well as sex specific stressors, warrant further investigation. These factors may help to decrease the burden of cognitive dysfunction and ADRD among Latinos in the US and yield important population health benefits.