Cumulative impact of health deficits, social vulnerabilities, and protective factors on cognitive dynamics in late life: a multistate modeling approach

We used data from the Honolulu-Asia Aging Study (HAAS). The HAAS is a prospective cohort study of 3,845 Japanese-American men, free of dementia at baseline. It originated in the Honolulu-Heart Program, the participants of which were born between 1900 and 1919 and lived in Oahu, Hawaii, at baseline, as detailed elsewhere [17]. In 1991, the HAAS was established to focus on cognition and brain diseases. Follow-up examinations took place approximately every 3 years. Clinical assessments were conducted, as were structured interviews on participant characteristics. All dates of death for HAAS participants within the study period (1991 to 1997) were recorded by the study’s surveillance system.

The Cognitive Abilities Screening Instrument (CASI) [18], a 100-point scale designed for cross-cultural use, was administered at each study wave. It is a measure of global cognitive function that assesses attention, concentration, orientation, short-term memory, long-term memory, language abilities, visual construction, list-generating fluency, abstraction, and judgment. The CASI has been validated for use among Japanese and Western populations [19] and within the items subsets can be used to form the Hasegawa Dementia Screening Scale, the Folstein Mini-Mental State Examination, and the Modified Mini-Mental State Examination. The standardized protocol used for cognitive assessments in the HAAS has been described previously [17,20]. For this study, two manipulations were performed to prepare the CASI data for multistate transition modeling, as has been done previously [14,15]. Individual CASI scores were reverse scored so that lower scores represent better cognition (for example, a score of 100 on the CASI = 0 CASI “errors”). Secondly, CASI error scores were grouped by intervals of three (for example, 0–2 errors, 3–5 errors, and so forth) forming a series of cognitive states ranging from high cognition/low errors to impaired cognition/high errors. Each grouping is considered as a global cognitive state where transition models can be used to estimate the probabilities of transitioning, including improvement, stability, and decline, from baseline states to consequent cognitive states at follow-up assessments.

A frailty index (FI), based on the accumulation of deficits, was created using a standard procedure [21]. Variables were selected as health deficits if they were associated with health status, accumulated with age, and had a prevalence in the sample greater than 1% but less than 80%. Each variable was dichotomized (0 for absent; 1 for present). These deficits were then summed and here divided by 48, that being the total number of deficits considered (see Appendix A in Additional file 1). Thus, a participant with no missing measurements and 24 deficits would be given a FI score of 24/48 = 0.50. We treated the FI as missing for participants in whom data were missing on more than 20% of the items.

Using the procedure outlined to create a FI [21], 18 social variables were selected from the baseline wave to construct a measure of overall social vulnerability (Social Vulnerability Index; SVI). This approach has been demonstrated to predict a range of health outcomes, including cognition and survival [22,23]. All HAAS variables that could be considered as social deficits were included in a dichotomized format (0 for absent; 1 for deficit) including living alone, social networks, and marital status (see Appendix B in Additional file 1). As with the FI, the social deficits were summed and divided by the total number of health deficits considered, producing a score between 0 and 1.

To evaluate the cumulative impact of a number of protective factors, a protection index (PI) was generated using baseline variables that were considered to be related to positive health outcomes. A previous study using this approach demonstrated that the aggregate effect of protective factors can influence age-related health outcomes [24]. Within the HAAS, 20 baseline variables that might offer protection to age-related cognitive decline were selected including physical exercise [25,26], the use of antihypertensive medications [27], availability of health services, not smoking cigarettes [28], moderate alcohol consumption [29], self-rated good health [30], healthy body weight [31], and others (see Appendix C in Additional file 1). As with the FI and SVI, an accumulation approach was used [21], in which selected items were dichotomized, summed, and divided by the total number of items considered. Note that in this index, higher scores are positive, reflecting that more protective factors are present (whereas in the other two indices higher scores are negative).

Data collection was approved by the Kuakini Medical Center Institutional Review Board, with informed consent provided by all participants. Approval for the secondary analyses came from the Research Ethics Committee of the Capital District Health Authority, Halifax, Nova Scotia, Canada.

The primary outcomes of the analyses were transitions in cognitive status and mortality at the first and second follow-up waves. Analyses take into consideration the baseline state, and then estimate the probability of transitions (staying at the same score, as well as degrees of worsening and improvement) from each of the states at baseline to the follow-up cognitive states using a Poisson approximation [14,15]. The probability of death was simultaneously modeled using a logistic function [14,15]. The incorporation of death into the transition models is essential, as longitudinal studies with older subjects are inherently predisposed to attrition via death [32], and failing to incorporate death in analyses can lead to biased estimates [33].

For this study, the effects of the baseline level of frailty, social vulnerability, and protective factors were evaluated for two time intervals (3 and 6 years), controlling for age (in years, measured as a continuous variable, centered on the grand mean of 78 years) and education (measured in years, continuous, centered on the grand mean of 10.5 years). To aid in interpretation of betas (β) and odds ratios, each of the index variables were transformed by multiplying by 10 so that results relate to 10% increases in the index variables. Figure 1 contains the participant flow chart for the two series of models including number of decedents within each interval. For each interval, index variables were modeled separately to analyze the effects of each of the indices on the outcomes, independent of the other indices. Models included all covariates (age, education, frailty, social vulnerability, protective factors). The outputs of the multivariable Poisson regression models are parameters (β) that can be used to analyze the effects that covariates have on the transition from any baseline state to any follow-up cognitive state or death. A positive coefficient indicates that increases in the predictor variable are on average related to increases in CASI errors. When the β coefficients are negative, the variable can be considered to have, on average, a protective effect of changes in cognition. As with standard regression analyses, the β values represent the expected impact on the outcome variable based upon a one-step increase in the predictor variable. To ease interpretation, all coefficients from the logistic regression models were transformed to odds ratios; Akaike information criterion and the Bayesian information criteria were calculated for each Poisson model as indicators for relative quality of the model fit (lower scores indicate better model fit). Similarly, −2_log likelihood was used for logistic regression models. All analyses were performed in SPSS 18.0 (SPSS Inc., Chicago).