Is cognitive aging predicted by one's own or one's parents' educational level? results from the three-city study

Abstract

The authors examined the associations of participants' and their parents' educational levels with cognitive decline while addressing methodological limitations that might explain inconsistent results in prior work. Residents of Dijon, France (n = 4,480) 65 years of age or older who were enrolled between 1999 and 2001 were assessed using the Isaacs' verbal fluency test, Benton Visual Retention Test, Trail Making Test B, and Mini-Mental State Examination up to 5 times over 9 years. The authors used random-intercepts mixed models with inverse probability weighting to account for differential survival (conditional on past performance) and quantile regressions to assess bias from measurement floors or ceilings. Higher parental educational levels predicted better average baseline performances for all tests but a faster average decline in score on the Isaacs' test. Higher participant educational attainment predicted better baseline performances on all tests and slower average declines in Benton Visual Retention Test, Trail Making Test B, and Mini-Mental State Examination scores. Slope differences were generally small, and most were not robust to alternative model specifications. Quantile regressions suggested that ceiling effects might have modestly biased effect estimates, although the direction of this bias might depend on the test instrument. These findings suggest that the possible impacts of educational experiences on cognitive change are small, domain-specific, and potentially incorrectly estimated in conventional analyses because of measurement ceilings.

Figure 1.

Trajectory of change in summary cognitive score over years of follow-up based on linear mixed (mean) regression models and quantile regression models at the 20th, 50th, and 80th percentiles, Dijon, France, Three-City Study, 1999–2010. Models included sex, baseline age, baseline age squared, first assessment indicator, years of follow up, and years of follow up squared. Predictions were for a female individual 65 years of age at baseline (enrollment) and were inverse probability weighted for survival and drop-out.

Figure 2.

Parental educational level quantile regression coefficients from the 20th to 80th percentiles for A) the Isaacs’ test, B) Benton Visual Retention Test, C) Trail Making Test B reversed, and D) Mini-Mental State Examination, Dijon, France, Three-City Study, 1999–2010. The dotted line connecting triangles shows differences by parental educational level in the 20th–80th percentiles of cognitive level at baseline. The solid line with solid circles shows predicted differences by parental educational level in the predicted rate of change per decade based on changes over the follow-up period at the 20th–80th percentiles of cognitive score. For all outcomes, lower quantiles corresponded to worse test performance. Models were also adjusted for sex, baseline age, first assessment indicator, interaction of parental educational level with baseline age, and interaction of parental educational level with first assessment. Predictions are for an individual who was 65 years of age at baseline (enrollment). For comparison, coefficients from random-intercept mixed models (contrasting mean values instead of quantiles) showing difference in predicted baseline score (large circle) and rate of change per decade (large triangle) by parental educational level for the same reference group are also shown on the plots. Models were inverse probability weighted for survival and drop-out.

Figure 3.

Quantile regression coefficients for participants’ educational levels from the 20th to 80th percentiles for A) the Isaacs’ test, B) Benton Visual Retention Test, C) Trail Making Test B reversed, and D) Mini-Mental State Examination, Dijon, France, Three-City Study, 1999–2010. The dotted line connecting triangles shows predicted differences by participants’ own educational levels in the 20th–80th percentiles of cognitive level at baseline. The solid line with solid circles shows differences by participants’ own educational levels in predicted rate of change per decade based on changes over the follow-up period at the 20th–80th percentiles of cognitive score. For all outcomes, lower quantiles corresponded to worse test performance. Models were also adjusted for sex, baseline age, first assessment indicator, interaction of parental educational level with baseline age, interaction of parental educational level with first assessment, interaction of parental educational level with years of follow up, interaction of participants’ own educational level with baseline age, and interaction of participants’ own educational level with first assessment indicator. Predictions are for an individual who was 65 years of age at baseline (enrollment). For comparison, coefficients from random-intercept mixed models (contrasting mean values instead of quantiles) showing difference in predicted baseline score (large circle) and rate of change per decade (large triangle) by the participants’ own educational levels for the same reference group are also shown on the plots. Models were inverse probability weighted for survival and drop-out.