Introduction

Age is a key risk factor for cognitive performance. Cognitive decline is common in older ages but recently there has been interest in understanding the age at which significant decline in cognitive abilities begins. Such knowledge has implications for the design of behavioral or pharmacological interventions since they are more likely to work if they are applied when, or even years before, individuals first begin to experience decline [1]. Efforts to date are often based on cross-sectional studies which may be confounded by ‘cohort effects’ [2, 3]. Longitudinal studies suggest evidence of cognitive decline in middle age but that age trajectories differ by sex and cognition domain or task [2, 4–7]. The age at peak performance also varies by task [8] and might be relevant to the age range examined for cognitive decline. Longitudinal data that span many decades generally report minimal cognitive decline before the age of 65 (reviewed in [2])[9], but such studies are rare and also subject to limitations including small sample size, selection attrition and retest or practice effects [3]. Hughes et al [7] examined cognitive decline among ~2,500 participants aged 25 to 95 years at recruitment in the Midlife in the Unites States (MIDUS) study, and all cognitive domains measured showed significant but small declines over 9 years, with differences in the timing and extent of change. The largest analysis to-date included a 10 year follow-up of ~7,400 participants aged 45–70 at recruitment of the Whitehall Study [4]. The design of this study allowed for cross-sectional and longitudinal analysis. For the former analysis, performance on several tests were progressively lower with older age categories. In longitudinal analyses, there was some evidence of greater decline at older ages and of a linear trend in decline with increasing age for some of the tests, particularly in men. However, the cross sectional data considerably overestimated cognitive decline among women but not in men.

UK Biobank is a large population cohort of adults aged 37 to 73 who underwent medical, sociodemographic, mental health and cognitive assessment in 2006–2010 and are being followed up at intervals. The large age-distribution and follow-up enables cross-sectional as well as longitudinal analysis of age as done by Singh-Manoux et al [4] but for a much larger sample. In a subsample of the UK Biobank, Hagenaars et al [10] presented simple age trends in standardized mean cognition scores, but the focus of that report was not age-related cognition, and it included only a subset of tests and participants at baseline. At the time the current study was under review, Kievit et al [11] used structural equation modeling to examine age and declines in fluid intelligence in the UK Biobank and reported a weak but significant association of cross-sectional age on the mean fluid intelligence score, such that older individuals scored slightly lower. The current study also leverages this valuable resource to investigate the impact of age on cognitive decline, as determined by several tests, in the largest population sample to date.

Materials and methods

Supporting Information (S1 File) provide an expanded description of the UK Biobank which is summarized below.

Results

Cross-sectional analysis

Descriptive statistics across age categories for participants completing at least one of the cognitive function tests at baseline are presented in Table 1. Older individuals were more likely to be current smokers and retired and to have lower income and educational achievement. In multivariable analysis (Table 2, Fig 1), FI scores were higher between ages 55 and 64 and lower at 65+ compared to age <45 (P<0.0001). The number of errors made on the Pairs and SDS tests were increasingly higher with older age groups (P<0.0001). RT and time taken to complete either Trail test were also increasingly higher with older age (P<0.0001). The odds of completing the PM test correctly on the first attempt was lower with older age (P<0.0001).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Baseline characteristics of study population^a.

https://doi.org/10.1371/journal.pone.0213948.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Cross-sectional associations between age and cognitive function tests measured at baseline (2006–10).

https://doi.org/10.1371/journal.pone.0213948.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Percentage difference/change in cognition scores with age based on cross-sectional (difference) and longitudinal (change) analysis.

To allow comparison across cognitive tests, cognitive difference (cross-sectional) or change (longitudinal) are expressed as a percentage difference/change as a function of the range of each test (See Analyses). Negative % difference/change in FI and SDS represents a lower/decline in performance. Positive % difference/change in Pairs, RT, Trail A and Trail B represents a lower/decline in performance. With increasing age category, Pairs, RT, Trail A, Trail B and SDS significantly changed by 5.9, 2.6, 6.0, 7.8 and 3.4% according to cross-sectional analysis. With increasing age category, Pairs, and RT significantly declined by 1.2 and 3.0% according to longitudinal analysis.

https://doi.org/10.1371/journal.pone.0213948.g001

Overall, men performed better than women on the FI, RT, Pairs, and Trail tests after multivariable adjustment (P<0.0001). Women performed better than men on the SDS test (P<0.0001). Sex×age interactions were observed for FI, RT, Pairs, and SDS tests (P<0.0001, S5 Table). Most notable is the lower FI scores for the 65+ age cohort compared to the <45 age cohort among men (β = -0.32, 2.4% difference) than among women (β = -0.08, 0.64% difference) according to multivariable models only.

Participants with higher educational attainment performed better than those with lower attainment for all examined tests (P<0.0001). APOE ε4 non-carriers performed significantly better than APOE ε4 carriers on the Pairs (P<0.0001), SDS (P<0.0001), Trail A (P<0.0001) and Trail B (P<0.0001) tests. Education×age interactions were observed for Trail B (P<0.0001) and PM (P<0.0001) tests (S6 Table), while significant APOE ×age interactions were observed for RT (P = 0.001) and SDS (P<0.0001) (S7 Table). Interpretations of these interactions were nevertheless unclear when reviewing stratified results and so they may be chance findings, have limited clinical significance or warrant further investigation.

Longitudinal analysis

Up to 26,005 participants returned for at least one follow-up assessment of FI, Pairs, RT and PM tests. These, but Overall, these participants were more likely white, male, nonsmokers, employed with lower Townsend deprivation indices and reporting higher income, higher educational attainment and lower alcohol consumption compared to those with no follow-up assessments (P<0.0001). However, these participants presented with similar cross-age characteristics (S8 Table) as those of the full sample (Table 1) and associations between age and cognitive function tests at baseline (S9 Table) were also similar to those of the full sample (Table 2). Mean performance on the FI, Pairs, and PM tests generally improved at the first follow-up (~4 years) across age groups, suggesting ‘practice effects’ for these tests. Some UK participants did not complete the first follow-up (~4 years) but returned for the second follow-up (~6 years). Because the mean scores were lower than baseline for these participants, practice effects were likely considerably weaker than that observed for the first follow-up. Table 3 and Fig 1 present results of the longitudinal analysis over an average follow-up time of 5.1 years. In multivariable analysis, FI decreased over follow-up among participants 65+ at baseline compared to those <45 years at baseline (Model 1, P = 0.002), but was no longer statistically significant in multivariable analysis once accounting for multiple testing (Model 2, P = 0.02). No significant declines were observed for participants aged 45–64 at baseline. RT and Pairs test scores worsened over the follow-up across the older age cohorts compared to the <45 age cohort (P<0.0001 for trends). For these two tests, trend tests for age-related declines in performance remained nominally significant when restricting analysis to participants <50 years at baseline and treating age as a continuous variable (P<0.05). There was no evidence that age impacted declines in PM test performance (Table 3). When comparing results from longitudinal and cross-sectional analysis (Fig 1), the latter yielded 10 to 12-fold higher changes in RT with age and 8 to 10-fold higher changes in Pairs with age.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Longitudinal analysis of age and cognitive change.

https://doi.org/10.1371/journal.pone.0213948.t003

The impact of age on FI, RT and Pairs tests were similar when analyses were performed separately for men and women (S10 Table), for lower and higher educational attainment (S11 Table), and for APOE ε4 carriers and non-carriers (S12 Table). Significant interactions with baseline score performance were observed for age-specific declines in Pairs (P<0.0001) and RT (P = 0.002) test performance (S13 Table). Although significant declines were observed in both strata for Pairs and RT, greater declines with age were observed among those with an initially low baseline test score.

We examined attrition and practice effects on FI and PM by comparing the characteristics and mean-level performance in the group of returnees with that of a previously untested group of participants (S1 File). We estimated that practice effects inflated FI and PM scores by up to ~10% and 18%, respectively, and that inflations were greater for older age cohorts.

Discussion

In the current study of a large cohort of individuals aged 38 to 73 at baseline, we observed significantly lower performance on memory, attention and processing tasks across successive age groups. Reasoning scores, based on the fluid intelligence test, were higher with successive age group until 60, then dropped to less than that of under 45 year olds. Longitudinal analysis of a subset of individuals with repeated measures of four tests showed linear declines in visual memory and processing speed tasks with age but of a much lesser degree than those observed in cross-sectional analyses. Decline rates in reasoning and prospective memory did not significantly differ with age. Taken together, our findings suggest that decline in cognitive abilities before age 65 is evident but small, and that observed cross-sectional differences in cognition from middle to older adult years may be due largely to age cohort effects.

The age at which cognitive decline occurs at the population level has been subject to debate. Estimates from cross-sectional studies reflect the effects of biological ageing but also cohort effects [2]. In the UK Biobank we observed marked differences in educational attainment between the youngest and oldest age groups. Income and employment status were also different but anticipated given life-stage of data collection. Accounting for education and measures of socioeconomic status influenced cross-age cohort differences, particularly those for FI, but not enough to narrow the differences in results from longitudinal analysis. Other factors or more detailed data on education and socioeconomic status likely contribute to the large age-cohort differences observed in the UK Biobank.

While longitudinal data are often considered superior to cross-sectional for characterizing decline in cognitive abilities, they may underestimate the effect of age because of selective sample retention (or attrition) and practice effects [2, 22–24]. Indeed, we observed evidence of both factors in the UK Biobank. The subset of participants undergoing repeat assessments presented with a different and potentially favorable risk factor profile to those providing only baseline measures. While some argue that samples in longitudinal studies are likely to be positively biased (i.e. presenting with less decline) in comparison with the samples in cross-sectional studies, it does not necessarily follow that this would lead to different patterns of age relations unless there is a relationship between the initial level of functioning and the rate of change [3]. In the current study, stratifying on educational attainment did not impact age-differences in baseline cognitive scores or in cognitive decline rates. Age-specific differences in declines for RT and Pairs test performance were observed when considering initial baseline scores; greater declines with age were observed among those with an initially low baseline test score. These findings contrast with those of other, but much smaller, studies which suggest that the age trends in cognitive functioning are parallel across different initial levels [3, 25–29].

Practice effects are defined as improvement on cognitive test performance resulting from learning during repeated testing. The magnitude of these practice effects may be modulated by the length of the test-retest interval; with shorter intervals yielding greater practice effects than longer intervals [3, 30–33]. In the UK Biobank, mean performance on tests generally improved at each follow-up, particularly the first, across age groups. The older age groups, nevertheless, appeared to benefit more from repeated testing (S1 File). This may be due to the low initial performance enabling potential for greater improvement compared to the higher initial scores of younger participants. This pattern might also reflect test score limitations at the top of the scale (i.e. ceiling effects).

In the current study we observed a non-linear relationship between age and FI in cross-sectional analysis and a similar but non-significant pattern in longitudinal analysis. Similar observations were reported by Kievet et al [11] who examined only FI in the UK Biobank and using structural equation modeling; results made available while the current study was under-review. Fluid ability is involved in processing novel information efficiently and with increased flexibility [5, 34, 35]. Smaller cross-sectional [8, 36] and longitudinal studies [5, 37, 38] have shown that advancing age is associated with a marked decrease in FI performance. Crystallized abilities rely on over-learned cognitive skills or accumulated knowledge and remain preserved or even improve with age in the cognitively intact [39]. Scores for FI also peak early in adulthood whereas scores for crystallized intelligence peak latter; middle age [8, 40–43]. The pattern of results we observed with FI would better align with those reported for crystallized intelligence. Previous UK Biobank investigators have also preferred the test label ‘verbal-numerical reasoning’ as opposed to the UK Biobank’s label ‘fluid intelligence” since performance on some items (i.e. “which number is the largest”, “age is to years as height is to?”, “relaxed means the opposite of?”) more likely rely on crystallized knowledge” [10, 18, 44–50].

Human and animal studies demonstrate sex differences in cognitive function in adulthood and ageing [6, 51, 52]. For example, on average, men perform better on spatial tasks, and women on verbal tasks [51]. Much of this literature is based on interviewer-based testing whereas tests administered in the UK Biobank were self-administered computer-based and, with the exception of a few fluid intelligence items, captured largely non-verbal tasks. The generally better performance of men over women in the UK Biobank is nevertheless consistent with previous research. However, for fluid intelligence we observed a greater difference in scores with increasing age among men compared to women and only when accounting for confounders. This sex×age interaction was not apparent in longitudinal analysis and further emphasizes that differences between sexes in cognitive decline are likely due to cohort differences in social factors rather than biological differences.

Most longitudinal studies covering a wide age range (~20 to 90 years) have had sample sizes less than 2000 but also generally report minimal cognitive decline before the age of 65 (reviewed in [2])[9]. Lundervold et al [9] recently reported a linear age-related decline via cross-sectional and longitudinal analysis with a lower change in the latter but included only 163 participants. Hughes et al [7] examined cognitive decline among ~2,500 participants aged 25 to 95 years at recruitment in the MIDUS study. Multiple cognitive domains were measured by a brief telephone battery and all showed significant but small declines over 9 years, with differences in the timing and extent of change. Processing speed showed the earliest and steepest declines [7]. Our findings generally align with those of a 10 year follow-up of ~7,400 participants aged 45–70 at recruitment of the Whitehall Study [4]. Performance on tests of memory, reasoning and phonemic and semantic fluency were progressively lower with older age categories. In longitudinal analyses, there was some evidence of greater decline at older ages and of a linear trend in decline with increasing age for some of the tests, particularly in men. However, the cross sectional data considerably overestimated cognitive decline among women but not in men, which the authors attributed to cohort differences in education.

The large sample size, age distribution, recency and availability of information on potential confounders are key strengths of the current study. However, a number of limitations should be considered when interpreting the results. The approach to data capture in the UK Biobank aimed to optimize the accuracy and completeness of the data collected, while also maximizing the efficiency of the process. To this end, brief tests of cognition that can be self-administered, are easily repeatable, and have associations with future cognitive decline were selected [53, 54] and further developed in the UK Biobank. Few tests were administered for different cognitive function domains and the tests employed are thus brief, non-standardized and lack external validation. Similarly, whether tests present with floor or ceiling effects has also not been vetted, and should they exist, may vary with age. The testing environment could not be fully standardized due to the self-administrative nature of these tests, particularly those completed at home. Baseline covariates for the web-administered cognitive tests were not collected at the same time as the tests. Taken together, measurement error contributing to type 2 error may be of concern as well as an incomplete representation of cognition. FI and PM tests were administered part way through recruitment and were only used in ten of the recruiting sites. Although power-loss attributable to imprecise measures of cognition may have been compensated for by the large sample size, it is unknown whether measurement error was random or, for example, varied by age or time of recruitment. The UK Biobank is also not representative of the sampling population, with evidence of a ‘healthy volunteer’ selection bias [55] and thus extrapolation of our findings to a more general population is limited. Indeed, this bias was especially evident in the current longitudinal study as demonstrated by characteristic differences between the larger baseline sample and follow-up sample as well as our attrition analysis.

In summary, our findings suggest that declines in cognitive abilities between the end of the fourth decade and age 65 are small. Because the pathophysiological process of dementia may begin years before clinical symptoms [1], the effectiveness of early intervention efforts aiming to treat or delay cognitive decline might only be clinically realized upon older age. Evidence of improving social factors for earlier generations may protect younger individuals from premature cognitive decline.

Supporting information

Acknowledgments

This research has been conducted using the UK Biobank Resource (Application #21394). Computations in this paper were run on the Quest cluster supported in part through the computational resources and staff contributions provided for the Quest high performance computing facility at Northwestern University, which is jointly supported by the Office of the Provost, the Office for Research, and Northwestern University Information Technology. The authors report no disclosures relevant to the manuscript.