Participants

The Canadian Longitudinal Study on Aging (CLSA) is a study of approximately 50,000 individuals who were between the age of 45 and 85 years at entry with plans to follow them for over 20 years. ^{Reference Raina, Wolfson and Kirkland18} The CLSA consists of a Tracking Cohort (N = 21,241), in which questions are administered over telephone, and a Comprehensive cohort (N = 30,097), in which participants are assessed in person at one of eleven data collection sites across Canada. We used data from the Comprehensive cohort due to the more detailed cognitive assessments. The CLSA collects a wide range of data including detailed information on sociodemographics, health and functional status, and social supports, as well as physical measurements in the Comprehensive cohort such as blood pressure, body mass index, and motor function (see https://www.clsa-elcv.ca/researchers/data-collection for the description of the data collection overview and tools). Data are available from the CLSA (www.clsa-elcv.ca) for researchers who meet the criteria for access to de-identified CLSA data. For the current study, participants were included if they had baseline (2011–2015) and follow-up (2015–2018) assessments complete including neuropsychological testing at both time points (Figure 1).

Figure 1: Participant flowchart.

Neuropsychological Testing and Normative Scores

Cognitive tests include a modified Rey Auditory Verbal Learning Test immediate recall after one learning trial (REY I) and 5-minute delayed recall (REY II), the Mental Alternation Test (MAT), Animal Fluency (AF), the Controlled Oral Word Association Test (COWAT; total score of the letters F, A, and S), and the Victoria Stroop Test with the Interference ratio (defined as the performance on the “color” task divided by performance on the “dot” task). Scores for the Stroop test were reversed to be compatible with the other tests. The neuropsychological tests are described in more detail elsewhere. ^{Reference Tuokko, Griffith, Simard and Taler19,Reference Tuokko, Griffith and Simard20}

We generated normed neuropsychological scores in multiple steps. First, for each neuropsychological test, each participant’s raw score was transformed into standardized scores (mean 0 and SD 1) based on comparisons with the healthy normative subsample. Normative comparisons require correction because measurement of cognitive function can be biased; for example, few years of education can make someone who is cognitively intact appear impaired because the cognitive tests capture educational status in addition to cognitive status. Full regression models did not reduce measurement bias (e.g., due to education), instead stratification and regression models for demographic corrections were the only method that reduced measurement bias. ^{Reference O’Connell, Kadlec and Griffith21} In other words, any analysis that includes demographic variables as covariates in regression models will not remove measurement bias in the cognitive tests. The standardized normed scores in CLSA were adjusted for language of administration (French/English), age, sex, and years of formal education using stratification and regression models (hybrid regression models). ^{Reference O’Connell, Kadlec and Griffith21} Second, scores were normed for baseline and follow-up separately, and then, third, reliable change scores were calculated. Reliable change scores account for error in measurement, regression to the mean, aging, and expected practice effects based on the performance of the subsample of participants who were healthy at baseline and first follow-up (O’Connell et al., in preparation). For this paper, reliable change scores were calculated for the composite memory (REYI and REYII), composite executive function (AF, MAT, COWAT, Stroop), and overall cognitive composite for participants with complete data for tests included in the respective composite scores.

Functional Status

The Older Americans Resources and Services (OARS) Multidimensional Assessment scale ^{Reference Fillenbaum and Smyer22,Reference George and Fillenbaum23} was used to assess functional decline between baseline and follow-up. The OARS questionnaire obtains information about seven activities of daily living (dressing, feeding, taking care of appearance, walking, getting out of bed, bathing, trouble getting to the bathroom in time) and seven instrumental activities of daily living (using the telephone, ability to travel, shopping/groceries, housework, taking medicine, handling money, and meal preparation). ^{Reference Fillenbaum24} We calculated the total number of times the respondent indicated that they need help with an activity or are completely unable to do an activity. Our primary outcome was the numerical change in the number of dependent activities from baseline to follow-up. We also assessed the outcome as any increase in dependence versus no increase in dependence.

Stroke and Other Covariates

Prior history of stroke was the primary exposure. This variable was obtained from self-report. For stroke and other co-morbidities, participants were asked to respond yes/no to the question: “Has a doctor ever told you that you have (the chronic condition)?”

Covariates included age, sex, education, self-reported co-morbidities of heart disease, peripheral vascular disease, high blood pressure, diabetes, smoking, rural residence, ethnicity, income, marital status, physical activity, fruit and vegetable consumption, and calculated body mass index. Categorization of variables is shown in Table 1. Physical activity was measured with the Physical Activity Scale for the Elderly (PASE). ^{Reference Washburn, Smith, Jette and Janney25} PASE evaluates physical activity over the last 1 week and includes leisure, occupational, and household activities, ranging from 0 to 793 where a higher score indicates more activity. The PASE score was the only variable not collected at baseline, as it was collected as part of a maintaining contact questionnaire 18 months after the baseline assessment.

Table 1: Baseline characteristics of those with memory scores comparing people with and without stroke

Statistical Analysis

Presence of stroke was the independent variable, standardized cognitive decline was the mediator, and functional decline was the dependent variable. Figure 2 shows the hypothetical model where stroke is related to functional decline directly, as well as indirectly through cognitive decline. We first tested the three components of the pathway independently using linear regression – the association between stroke and cognitive decline (path a; first component of the indirect effect), the association between cognitive decline and functional decline (path b; second component of the indirect effect), and the association between stroke and functional decline (path c’; the direct effect). The simple association between stroke and functional decline is measured by path c, which is contrasted with path c’ where cognitive decline is modeled as a mediator removing some of the variance in the association between stroke and functional decline. Mediation was completed using the “medsem” package in Stata, which provides two approaches to the statistical analysis of mediation. The first approach to mediation was first developed by Baron & Kenny. ^{Reference Baron and Kenny26} Structural equation modeling is used to simultaneously estimate regression coefficients for the three pathways. If all three pathways are statistically significant (steps 1–3), and c’ is reduced to zero after controlling for the mediation (step 4), then there is “full mediation.” If steps 1–3 are met but step 4 is not, there is “partial mediation.” The statistical significance of the reduction in path c’ after controlling for the mediator is determined using the Sobel test. ^{Reference Sobel27} Medsem employs the modified approach of the above steps for use with structural equation modeling. ^{Reference Iacobucci, Saldanha and Deng28}

Figure 2: Theoretical causal model for mediation of the relationship between stroke and function decline by cognitive decline. a: association between stroke and cognitive decline; b: association between cognitive decline and functional decline; c’: association between stroke and functional decline in presence of mediator; c: association between stroke and functional decline without mediator present; a * b = indirect effect; a * b + c’ = total effect.

The second approach is based on Zhao et al., ^{Reference Zhao, Lynch and Chen29} in which mediation can be determined solely with a bootstrap test of the indirect effect (pathways a and b), and is thought to be a better option than the Sobel test. If the confidence intervals do not include the value of zero, the indirect effect is significant. If the bootstrap test of the indirect effect is significant but path c’ remains significant after controlling for the mediator, there is partial mediation. If the bootstrap test of the indirect effect is significant and path c’ no longer significant after controlling for the mediator, there is full mediation. Medsem uses the Monte Carlo approach, an alternative to bootstrapping which is less computationally intensive. ^{Reference Jose30–Reference Kenny33}

We used both the Baron & Kenny and the Zhao approaches and compared for compatibility. We computed the presence and degree of mediation for each cognitive domain and for both functional outcomes separately. We estimated the proportion mediated by cognitive decline by calculating the proportion of the bootstrapped indirect effect (a * b) divided by the total effect (a * b + c’).

Lastly, to verify the unique role of cognition in mediation versus other control variables we replaced cognitive decline with the following potential mediators in different domains: change in depression status using the Center for Epidemiologic Studies Depression scale (CES-D-10), change in systolic blood pressure, change in Timed Get Up and Go Test, and change in Social Support Scale from baseline to follow-up. The CES-D-10 is a 10-item Likert scale questionnaire assessing depressive symptoms in the past week. ^{Reference Andresen, Malmgren, Carter and Patrick34} It includes three items on depressed affect, five items on somatic symptoms, and two on positive affect. Options for each item range from “rarely or none of the time” (score of 0) to “all of the time” (score of 3). Total scores can range from 0 to 30. Higher scores suggest greater severity of symptoms. The Social Support Scale was adapted from the Medical Outcomes Study Social Support Survey ^{Reference Sherbourne and Stewart35} and has 19 items which address tangible social support, affection, positive social interaction, and emotional or informational support. An overall social support score is obtained by averaging the responses over all 19 items and transforming the score into a 0–100 scale, where higher score indicates better social support. The Timed Get Up and Go Test measures the time taken to stand up and walk a distance of 3 m, turn around, walk back to the chair, and sit down again.

In a separate sensitivity analysis, the associations with functional decline were assessed with logistic regression using a dichotomous outcome of any increase in dependence. In the main models, we adjusted for age, sex, education, heart disease, peripheral vascular disease, high blood pressure, diabetes, smoking, body mass index, rural residence, and ethnicity. Although cognitive scores were normed by age, sex, and education, these variables were included to ensure that the direct causal relationship between stroke and functional decline was appropriately adjusted for confounders. In a sensitivity analysis, we excluded age, sex, and education. In the full model, additional adjustment was made for income, marital status, physical activity, and fruit and vegetable consumption. In additional sensitivity analyses we (1) reversed the order of cognitive and functional impairment, to test whether functional change is a mediator of cognitive change, (2) assessed whether baseline cognition mediates functional change, and (3) changed the independent variable to Parkinson’s disease or myocardial infarction.

Analyses were conducted in Stata 17.0 (College Station, TX).