Harm avoidance scores ranged from 0 to 34 (mean=10.5, SD=6.6, skewness=0.7), with higher values indicating more of the trait. Harm avoidance was unrelated to age (r=.06, p=.07) or possession of an apolipoprotein E ε4 allele (χ2 =0.8, p=.36), inversely related to education (r=−.20, p<.001), and higher in women than men (χ2=9.3, p=.002). Higher trait score was associated with higher level of depressive symptoms (r=.42, p<.001), loneliness (r=.34, p<.001), and neuroticism (r=.66, p<.001) and lower level of global cognitive functioning (r=−.17, p<.001), cognitive activity (r=−.17, p<.001), and physical activity (r=−.11, p=.002).
Harm Avoidance and Incidence of Alzheimer’s disease
During a mean of 3.5 years of observation, 98 people developed AD. Those who became affected were older, more cognitively impaired, and more likely to have an ε4 allele than those who remained unaffected, and they differed in harm avoidance, loneliness, and cognitive activity (). We assessed the relation of harm avoidance score to risk of AD in a proportional hazards model that controlled for age, sex, and education. Higher level of the trait was associated with increased risk of AD (hazard ratio=1.045, 95% confidence interval: 1.013, 1.077). To visually examine this effect, we plotted the model-based estimates of risk of developing AD at different levels of harm avoidance (upper panel of ). A person with a high level of the trait (score=20, 90th percentile, red line) was more than twice as likely to develop AD as a person with a low level of the trait (score=3, 10th percentile, black line).
Baseline characteristics of participants who developed Alzheimer’s disease and those who did not*
Figure 1 Cumulative risk of developing Alzheimer’s disease (upper panel) and mild cognitive impairment (lower panel) in persons with different levels of harm avoidance (90th percentile, red; 75th percentile, green; 50th percentile, blue; 25th percentile, (more ...)
The scaled Schoenfeld residuals for harm avoidance, age, and education did not vary over time, consistent with the proportional hazard assumption of the Cox regression model. The residuals for sex were related to time, but in a subsequent Cox model stratified for sex, the association of harm avoidance with risk of AD was virtually identical to the original analysis (hazard ratio = 1.043, 95% confidence interval: 1,013, 1.073).
Because women tend to have higher levels of the trait than men (12
), we examined the possibility that gender might modify the association of harm avoidance with risk of AD. There was no interaction, however (estimate = −0.038, SE = 0.038, p = .31). In subsequent analyses, there was no evidence that the effect of harm avoidance varied by age (estimate = −0.002, SE = 0.002, p =.31) or education (estimate = 0.000, SE = 0.005, p = .94).
Depressive symptoms (26
), loneliness (27
), and the neuroticism trait (35
) have been linked to late life dementia and each was associated with harm avoidance in this cohort. To determine whether these associations affected results, we adjusted for each covariate in separate analyses. Harm avoidance continued to be associated with increased risk of AD after controlling for depressive symptoms (hazard ratio=1.038; 95% confidence interval: 1.005, 1.073), loneliness (hazard ratio=1.034; 95% confidence interval: 1.001, 1.068), and neuroticism (hazard ratio=1.046; 95% CI: 1.003, 1.091). With all 3 of these correlated covariates in the same model, the point estimate of the association was similar to previous analyses (hazard ratio = 1.037) but the standard error was increased (95% confidence interval: 0.996, 1.081), likely due to collinearity.
We considered additional factors that might have affected results. First, because physical activity is related to harm avoidance (10
) and dementia (37
), we repeated the analysis with a term added for self reported level of physical activity at baseline. Second, we repeated the analysis with a term for frequency of participation in cognitively stimulating activities and again with a term for possession of an apolipoprotein E ε4 allele because they are established risk factors for AD. The association of harm avoidance with risk of AD persisted in each analysis (hazard ratios for harm avoidance ranged from 1.028 to 1.045, each p<.05).
Substantial variation was evident in harm avoidance subscores: anticipatory worry (range: 0–11), fear of uncertainty (range: 0–7), shyness (range: 0–8), fatigability (range: 0–9). To determine whether the subscores were differentially associated with AD risk, we analyzed each in a separate model. As shown in , higher levels of anticipatory worry, fear of uncertainty, and fatigability were each related to increased risk of AD with a nearly significant effect for shyness.
Relation of harm avoidance subscales to risk of developing Alzheimer’s disease*
Harm Avoidance and Incidence of Mild Cognitive Impairment
We conducted additional analyses to assess whether harm avoidance is associated with incidence of MCI, one of the earliest clinical manifestations of AD (38
). Of 588 people without evidence of any cognitive impairment at baseline, 208 (35.4%) developed MCI during the study period. Higher level of harm avoidance was associated with increased incidence of MCI (hazard ratio =1.038; 95% confidence interval: 1.015, 1.061). The lower panel of shows the model-based estimates of risk of developing MCI at different levels of harm avoidance. Risk of MCI was increased by more than 80% with a high level of the trait (score =20, 90th
percentile, red line) compared to a low level (score=3, 10th
percentile, black line). In line with the proportional hazards assumption, there was no evidence that model coefficients varied over time.
Similar associations were observed for all four trait subscores (hazard ratio for anticipatory worry =1.109; 95% confidence interval: 1.036, 1.188; hazard ratio for fear of uncertainty =1.100; 95% confidence interval: 1.007, 1.202; hazard ratio for shyness = 1.066; 95% confidence interval: 1.001, 1.134; hazard ratio for fatigability =1.086; 95% confidence ratio: 1.022, 1.055).
Harm Avoidance and Cognitive Decline
Because of the insidious onset and gradual progression of AD, separating MCI from normality and dementia from MCI can be difficult. To ensure that results obtained with MCI and AD were not the result of diagnostic bias or baseline differences in level of cognitive function, we examined the relation of harm avoidance to cognitive decline, the primary manifestation of the disease. In a mixed-effects model adjusted for age, sex, and education, harm avoidance was associated with lower baseline score on a composite measure of global cognition (beta estimate = −0.007, SE=0.003, p=.007) and, with this baseline effect accounted for, more rapid global cognitive decline (beta estimate =−0.002, SE=0.001, p=.003). Rate of global cognitive decline was approximately 50% faster in a person with a high level of the trait (90th percentile, dotted line) compared to a low level (10th percentile, solid line), as shown in the upper left portion of . Plots of the conditional studentized residuals suggested adequate model fit though there were some outliers. The association of harm avoidance with MCI risk was unchanged, however, when the analysis was repeated with the outliers removed (beta estimate = −0.002, SE = 0.001, p < .001).
Rate of decline in different cognitive domains in persons with different levels of harm avoidance (90th percentile, dotted line; 50th percentile, dashed line; 10th percentile, solid line), adjusted for age, sex, and education.
In subsequent models, we controlled for other affective states and traits. The association of harm avoidance with global cognitive decline persisted in separate analyses that adjusted for depressive symptoms (beta estimate = −0.003, SE = 0.001, p = .004), loneliness (beta estimate = −0.002, SE = 0.001, p =.03), and neuroticism (beta estimate = −0.002, SE = 0.001, p = .049) and in an analysis that simultaneously adjusted for all 3 of these covariates (beta estimate=−0.002, SE=0.001, p=0.047).
To determine whether harm avoidance was related to decline in some cognitive systems but not others, we repeated the analysis with composite measures of specific cognitive functions in place of the global measure. Harm avoidance was associated with decline in episodic memory, working memory and perceptual speed () but not semantic memory or visuospatial ability (). Results for working memory were especially notable. As shown in (lower left), trait score was unrelated to baseline level of working memory, but those with a high level of the trait (90th percentile, dotted line) experienced twice the rate of working memory decline as persons low in the trait (10th percentile, solid line).
Relation of harm avoidance to change in different domains of cognitive function*
To see if the association of harm avoidance with cognitive decline was due to an association with MCI, we repeated analyses excluding those with MCI at baseline. Higher harm avoidance was associated with more rapid decline in global cognition (beta estimate = −0.003, SE = 0.001, p = .001), working memory (beta estimate = −0.004, SE = 0.001, p<.001) and perceptual speed (beta estimate = −0.003, SE = 0.001, p = .001) but not with decline in other cognitive measures.
Harm Avoidance and AD Pathology
We conducted a final series of analyses to determine whether harm avoidance was a subtle manifestation of the neuropathologic changes underlying AD. At the time of these analyses, 220 study participants had died; 182 (82.7%) underwent brain autopsy, the results of which were available in 116 of whom 74.1% were women. They had a mean age at death of 88.2 (SD = 5.9), a mean postmortem interval of 7.9 hours (SD = 7.6), and mean of 7.5 months (SD = 4.8) from last clinical evaluation to death. In separate linear regression models adjusted for age at death, sex, and education, a composite measure of plaques and tangles (mean = 0.52, SD = 0.49) was not related to harm avoidance (beta estimate = 1.37, SE = 1.37, p= .32) or its component scores (beta estimate for anticipatory worry = 0.12, SE = 0.47, p = .81; beta estimate for fear of uncertainty = 0.51, SE = 0.31, p = .10; beta estimate for shyness = 0.68, SE = 0.51, p = .19; beta estimate for fatigability = 0.08, SE = 0.50, p= .87).