Trajectories of self-reported hearing and their associations with cognition: evidence from the United Kingdom and United States of America

Abstract

Objective

This study aimed to investigate the relationships between trajectories of change in self-reported hearing over eight years with subsequent effects on cognition, measured using episodic memory.

Methods

Data were drawn from 5 waves (2008–2016) of the English Longitudinal Study of England (ELSA) and the Health and Retirement Study (HRS), involving 4,875 individuals aged 50+ at the baseline in ELSA and 6,365 in HRS. The latent growth curve modelling was used to identify trajectories of hearing over eight years, and linear regression models were performed to investigate the relationship between hearing trajectory memberships and episodic memory scores, controlling for confounding factors.

Results

Five trajectories of hearing (stable very good, stable fair, poor to fair/good, good to fair, and very good to good) were retained in each study. Individuals whose hearing remains suboptimal and those whose hearing deteriorates within suboptimal levels throughout eight years have significantly poorer episodic memory scores at follow-up than those with stable very good hearing. Conversely, individuals whose hearing declines but is within an optimal category at baseline do not see significantly poorer episodic memory scores than those with consistently optimal hearing. There was no significant relationship between individuals whose hearing improved from suboptimal baseline levels to optimal by follow-up and memory in ELSA. However, analysis using HRS data shows a significant improvement for this trajectory group (−1.260, P < 0.001).

Conclusions

Either stable fair or deterioration in hearing is associated with worse cognitive function, both stable good or improving hearing is associated with better cognitive function specifically episodic memory.

Key Points

• Few studies have examined longitudinal associations between hearing impairment and cognitive decline across countries.

• Five trajectories are retained for self-reported hearing over eight years.

• A continuously poorer hearing was associated with poorer memory than those whose hearing was consistently good.

Introduction

Prevention of cognitive decline in later life is a public health priority due to the growing burden placed on the economy and society by increasing diagnoses of dementia among an ageing population [1]. As cognitive decline is a strong yet potentially modifiable risk factor for dementia [2, 3], there is great importance in investigating interventions focused on cognitive decline to reduce diagnosed cases of dementia.

Hearing impairment in later life is also an increasingly important area for public policy consideration. Not only does hearing impairment itself place a burden on the individual in terms of quality of life [4] and society through loss of engagement in work and healthcare costs [5], evidence has independently linked hearing impairment to various poor health outcomes, including cognitive decline and the incident dementia [6, 7]. A study in the Lancet Commission has developed a life course model of dementia risk factors and found that midlife hearing impairment is a significant risk factor for dementia [8]. It found that 9% of dementia cases in high-income countries are attributable to hearing loss [6, 8], and evidence suggests this relationship may be modifiable [6, 9]. Thus, we must consider protecting cognitive function in older adults by taking a collective risk factor management approach similar to cardiovascular prevention [10], which could include a hearing assessment.

Many studies have examined the cross-sectional associations between hearing impairment and cognitive decline [11–14]. However, longitudinal studies have provided a mixed set of results. Some provide evidence of significant associations, and others suggest these associations may disappear once relevant confounding factors have been accounted for [15–17]. Moreover, few studies have examined longitudinal associations between hearing impairment and cognitive decline across countries [11]. Understanding how a given trajectory of hearing might be associated with a cognitive measure may offer opportunities for intervention to maximise cognitive function in older age.

This study used data from the English Longitudinal Study of Ageing (ELSA) and the Health and Retirement Study (HRS) based in the United States of America to investigate whether trajectories of change in self-reported hearing over eight years were associated with subsequent effects on cognition, measured using episodic memory. We include the two datasets as ELSA was designed to be a sister study to the HRS from the outset and the core design features of those two studies are thus comparable [18]. For instance, both studies are: (i) nationally representative of community-dwelling and older adults, (ii) longitudinal panel surveys and (iii) conducting the core interview every two years.

Data and methods

Data

This study was an observational study using two nationally representative datasets: the English Longitudinal Study of England (ELSA) and the US Health and Retirement Study (HRS), both of which provide data on the socio-demographic, economic and health circumstances of community-dwelling individuals aged ≥50 [18, 19]. The initial ELSA sample was taken from people aged ≥50 who had participated in the Health Survey for England in either 1998, 1999 or 2001. The initial HRS sample was drawn from community-dwelling adults aged ≥50 in the United States of America in 1992. In each survey, data are collected every two years. This study used data from waves 4 to 8 of ELSA and waves 9 to 13 of HRS (Supplementary Figure 1). We included those waves as they have the same time span (2008–2016) and have comparable variables, including episodic memory, self-reported hearing and covariates. The present analysis includes core respondents from ELSA Wave 4 and HRS Wave 9 aged 50 years and older who participated in all waves from baseline onwards and underwent the cognitive test in Wave 8 for ELSA and Wave 13 for HRS. The final sample consisted of 4,875 individuals in ELSA and 6,365 in HRS. We have included the comparison in descriptive statistics at baseline for the sample without and with missing data in Supplementary Table 2. As the descriptive statistics show similar characteristics, we consider that the missing data are completely at random.

Cognitive function

Cognitive function in ELSA Wave 8 and HRS Wave 13 was captured using episodic memory, measured identically in both ELSA and HRS. Respondents were read a list of 10 common nouns and then asked to recall as many of these words as possible: in the first instance immediately (immediate recall) and in the second at the end of the cognitive function module (delayed recall). Immediate recall demonstrates the ability to retain new information, and delayed recall demonstrates the ability to remember information after a period of distraction. Episodic memory has been shown to represent general cognition [20] and is more age-sensitive than other episodic measures of cognition [21]. Furthermore, episodic memory has been linked to increased risks of dementia [22]. Raw scores were calculated as the total words recalled immediately and delayed, with a minimum score of 0 and a maximum score of 20.

Self-reported hearing

Respondents in ELSA and HRS were asked to rate their hearing within the categories ‘excellent’, ‘very good’, ‘good’, ‘fair’ and ‘poor’. In addition, individuals with hearing aids were asked to rate their hearing based on when they used their hearing aid. The original categories were used in the analysis to capture distinct patterns of change in this variable over time.

Covariates

Fully-adjusted models control for various confounding variables, including age and sex. As socio-economic factors influence cognitive outcomes [23], models control for marital status, education and wealth. Marital status is included as a binary variable to show the effect of being married or in a couple relative to being single, divorced, separated or widowed. Wealth is measured as net total non-pension wealth at the household level and is split into quintiles to demonstrate the hierarchical effects of wealth, with the highest wealth quintile set as reference. In order to make an educational attainment variable comparable across both datasets, a binary variable was created to show those with no formal qualifications compared to those with any formal qualification. As cognition is also affected by lifestyle behaviours [24], models were adjusted by smoking, drinking status and physical activity. Smoking status was included as a categorical variable, comparing current and former smokers with those who reported never smoking. Alcohol use was again included categorically, comparing those who consumed alcohol more than once a week and less often than once a week with those who never consumed alcohol. Variables asking respondents how often they participate in vigorous, moderate or mild physical activity in their daily lives were combined and recoded to create a binary variable showing the effects of a sedentary lifestyle (no level of physical activity reported) compared to any other level of activity. Finally, as cognition is associated with general health [25], self-reported health is included as a categorical variable, with those reporting very good, good, fair or poor health compared to the excellent reference group. All the covariates were taken at baseline (ELSA Wave 8 and HRS Wave 13).

Statistical analysis

To compare results between ELSA and HRS, each stage of the analysis process was carried out separately for each dataset. The first stage of this analysis was to identify trajectories of hearing over the eight-year period by means of latent growth curve modelling. Latent trajectory models identified trajectories of self-reported hearing over the data period. Through these models, respondents could be classified into distinct ‘classes’ based on their patterns of response so that individuals within classes are similar and can be compared to those belonging to different classes. The intercept demonstrates baseline self-reported hearing, and the growth curves represent the change in hearing over time. Trajectories were created while accounting for individual baseline age and sex. Individuals were included if they had no missing data on self-reported hearing at any wave, and relevant longitudinal weights were used on both datasets.

The number of distinct trajectories identified and chosen for final analyses was dependent on various criteria: a lower Akaike’s Information Criteria (AIC) and Bayesian Information Criteria (BIC) with each class added to the model, an entropy value of at least 0.8, and at least 5% of the respondents contained within each class (class N is reported in Table 2). The goodness of fit criteria for both ELSA and HRS data are shown in Table 1. The number of trajectories was also based on the assumption that they demonstrated distinct patterns of change over time. Respondents were assigned to a trajectory based on their highest probability of trajectory membership.

Table 2.

Sample descriptive statistics by hearing trajectory (estimates are means or percentages)

	ELSA					HRS
	Stable very good (n = 2,893)	Optimal—suboptimal (decline) (n = 423)	Good to fair (n = 557)	Suboptimal—optimal (improve) (n = 304)	Stable fair (n = 698)	Stable very good (n = 3,821)	Optimal—suboptimal (decline) (n = 623)	Good to fair (n = 498)	Suboptimal—optimal (improve) (n = 387)	Stable fair (n = 1,034)
Female	60.39	58.04	47.23	43.07	39.63	68.70	65.06	63.48	46.07	48.46
Age	62.85	64.72	67.00	65.99	66.57	65.12	65.66	69.23	67.53	69.10
Married/in a couple	74.05	69.40	69.77	70.80	72.59	69.53	68.85	52.17	68.54	67.64
Wealth quintile
Poorest	13.02	14.84	15.93	18.52	18.95	13.57	15.06	27.83	25.84	19.76
Second quintile	16.63	21.61	21.06	12.59	20.20	17.73	18.61	26.09	24.72	21.93
Third quintile	20.32	18.39	19.05	28.15	20.58	19.11	20.11	19.13	14.61	19.89
Fourth quintile	22.55	18.06	21.79	16.30	20.20	26.04	22.02	8.70	14.61	18.32
Wealthiest	27.49	27.10	22.16	24.44	20.08	23.55	24.20	18.26	20.22	20.10
No qualifications	20.66	20.63	24.91	29.20	33.17	14.34	13.77	25.00	28.33	29.01
Smoker status
Current	12.21	13.38	10.89	13.33	13.04	8.98	10.65	16.95	11.86	12.49
Former	45.71	43.63	46.28	48.89	48.20	49.39	47.55	37.29	45.76	41.92
Never	42.08	42.99	42.83	37.78	38.76	41.63	41.81	45.76	42.37	45.59
Drinks alcohol
Frequently	64.76	63.07	67.57	61.98	62.46	45.56	41.00	30.00	37.29	33.64
Less frequently	26.95	26.29	22.18	23.97	25.63	18.95	17.12	18.33	16.95	19.35
Never	8.29	9.17	10.25	14.05	11.90	35.48	41.87	51.67	45.76	47.01
Sedentary lifestyle	5.09	4.73	8.77	8.76	8.77	3.65	5.69	7.83	9.64	9.78
Self-reported health
Excellent	17.00	16.24	8.36	10.37	7.50	19.39	12.50	6.96	6.74	5.60
Very good	35.09	38.85	27.82	17.78	22.38	41.55	36.52	28.70	30.34	21.98
Good	31.39	28.34	37.82	42.96	35.63	23.55	33.19	25.22	26.97	35.32
Fair	13.04	12.42	18.36	23.70	25.12	11.91	13.97	26.09	17.98	27.63
Poor	3.48	4.14	7.64	5.19	9.38	3.60	3.82	13.04	17.98	9.47
Episodic memory	10.96	10.31	9.27	9.74	9.03	9.60	9.55	7.04	7.81	7.92

Notes: Statistical group differences are estimated using Kruskal–Wallis one-way analysis of variance for numerical variables and ordinal chi-square tests for categorical variables. All group differences are statistically significant (P < 0.001).

Table 1.

Statistical criteria for choice of eight trajectory model

Trajectories identified	ELSA			HRS
	BIC value	AIC value	Entropy	BIC value	AIC value	Entropy
3	69446.654	69373.989	0.898	109608.877	109423.391	0.844
4	68572.847	68479.982	0.887	109514.534	109344.022	0.835
5	68044.610	67931.846	0.872	109383.138	109261.105	0.831
6	67963.279	67830.615	0.843	109257.091	109113.522	0.824
7	67870.069	67717.505	0.812	109105.169	108940.065	0.819
8	67843.028	67670.565	0.812	108976.754	108790.115	0.815

The final phase of the analysis used linear regression models to investigate the relationship between hearing trajectory membership and episodic memory score, controlling for confounding factors, including age, sex, marital status, education, wealth, smoking and drinking status, physical activity and self-reported general health. Supplementary Figure 2 shows that the episodic memory scores are normally distributed in both ELSA and HRS. Analyses were conducted using Mplus 8 and STATA 14.

Results

Identifying trajectories of self-reported hearing

Using AIC, BIC and entropy statistics, as well as ensuring at least 5% of the sample were included in each identified class, and distinct trajectories were identified, 8-class models proved to be the best fit of data from both ELSA and HRS (model fit statistics provided in Appendix 1). Within each dataset, five ‘stable’ trajectories of hearing were identified, each corresponding to the five categories of the original self-reported hearing variable (excellent, very good, good, fair and poor). Additionally, two trajectories of worsening hearing were identified (one from excellent or very good to fair, one from fair to poor) alongside one trajectory of poor to fair/good (from fair or poor to good). In order to simplify analyses, the eight trajectories were reduced to five using the individual cumulative probability of each group membership: stable very good hearing (excellent, very good or good), stable fair hearing (fair or poor) and the original three trajectories showing the change. The five trajectories identified are shown in Figure 1. The number of samples in each trajectory is: (i) stable very good (ELSA = 2,893; HRS = 3,821), (ii) optimal —suboptimal (decline) (ELSA = 423; HRS = 623), (iii) good to fair (ELSA = 557; HRS = 498), (iv) suboptimal—optimal (improve) (ELSA = 304; HRS = 387) and (v) stable fair (ELSA = 698; HRS = 1,034).

Figure 1.

Identified trajectories of self-reported hearing using ELSA and HRS (merged from the original eight).

Main analysis

Table 2 shows descriptive statistics of the ELSA and HRS samples by trajectory membership. Characteristics are compared between groups using Kruskal-Wallis one-way analysis of variance for numerical variables and ordinal chi-square tests for categorical variables. In both samples, the baseline characteristics of those with stable very good hearing and those whose hearing declines from optimal levels are reasonably similar: they are more likely to be female, younger, less likely than any other group to belong to the poorest wealth quintile and more likely to belong in the wealthiest, the least likely to have no qualifications, much more likely to report excellent or very good health and less likely to report fair or poor health. The highest episodic memory scores are observed among these two trajectory types in ELSA and HRS. Those with stable fair hearing, hearing which worsens from already suboptimal levels, and hearing which improves but from suboptimal levels are older, male, have lesser wealth, are more likely to have no qualifications, and are more likely to report suboptimal health and have lower episodic memory scores.

Table 3 shows the results of regression analyses of self-reported hearing trajectory on episodic memory. Unadjusted results are reported as model 1 and fully adjusted for model 2. Models were run using a stepwise approach, and the initial significant effects of hearing trajectories became non-significant after controlling for age and sex. Estimates of trajectory effects are relative to the reference group of individuals with stable very good hearing throughout the data period.

Table 3.

Regression of self-reported hearing trajectory (from waves 4 to 8 ELSA; waves 9 to 13 HRS) on episodic memory at wave 8 (ELSA) and wave 13 (HRS)

	ELSA		HRS
	Model 1	Model 2	Model 1	Model 2
Trajectorya (ref. stable very good hearing)
Stable fair	−1.928***	−0.531***	−1.035***	−0.492***
Poor to fair/good	−1.219***	−0.144	−1.417***	−1.260**
Good to fair	−1.689***	−0.553***	−2.099***	−0.680***
Very good to good	−0.646	−0.190	−0.088***	−0.190
Age	−0.178***	−0.176***	−0.153***	−0.153***
Female	0.766***	0.946***	0.672***	0.878***
Married		−0.025		0.608
Wealth (ref. poorest)
Second quintile		0.211		0.595***
Middle quintile		0.492**		0.977***
Third quintile		0.816***		1.248***
Wealthiest		1.374***		1.531***
No qualifications		−1.389***		−0.778***
Smoker status (ref. never)
Former smoker		0.001		0.097
Current smoker		−0.276*		−0.246*
Alcohol (ref. never)
Less than once a week		−0.071		0.204
More than once a week		0.298		0.380***
Self-reported health (ref. excellent)
Very good		−0.176		−0.338*
Good		−0.692***		−0.735***
Fair		−1.090***		−1.373***
Poor		−1.539***		−1.754***
Sedentary lifestyle		−0.117		−0.022

Notes: a = the change in hearing over time per trajectory.

After controlling for all covariates, including age, sex, marital status, education, wealth, smoking and drinking status, physical activity, and self-reported general health, individuals whose hearing remains suboptimal throughout the study period, as well as those whose hearing deteriorates within suboptimal levels, have significantly poorer episodic memory scores at follow-up than those with stable very good hearing throughout the study (ELSA stable fair −0.531, good to fair −0.553, P < 0.001; HRS stable fair −0.492, good to fair −0.680, P < 0.001). The magnitude of these effects is similar for both hearing trajectories and data from ELSA and HRS. Individuals whose hearing declines but is within an optimal category at baseline do not see significantly poorer episodic memory scores than those with consistently optimal hearing. Models using ELSA data show no significant relationship between individuals whose hearing improves from suboptimal baseline levels to optimal by follow-up and memory. However, models using HRS data show a significant improvement in memory scores for this trajectory group (−1.260, P < 0.001).

Discussion

Although independent effects of hearing ability on cognitive function have been established [6], little is known about how trajectories of self-reported hearing ability over time might impact cognition. The study shows that consistently reporting good hearing is associated with better episodic memory than consistently suboptimal hearing or declines from fair to poor over eight years. Results remained significant after controlling for several potential confounding variables, including age, sex, marital status, education, wealth, smoking and drinking status, physical activity, and self-reported general health. This research ties in with previous longitudinal work linking hearing impairment with cognitive decline [14, 26].

There are four key mechanisms through which hearing impairment might affect cognition. Firstly, the ‘cognitive load on perception’ hypothesis assumes cognitive decline precedes deterioration in hearing [27]. Secondly, the ‘sensory deprivation hypothesis’ assumes hearing impairment leads to cognitive decline, possibly through poorer levels and quality of social engagement. Thirdly, the common-cause hypothesis assumes another factor simultaneously influences sensory and cognitive ability. This study does not provide evidence for a conclusive understanding of the mechanisms between poorer trajectories of hearing and cognitive decline, although the social and health circumstances of those with the poorest hearing are significantly worse than those of individuals whose hearing remains at higher levels, possibly suggesting a link between hearing ability and societal inclusion—a lack of which might, in turn, lead to poorer cognitive outcomes as measured by episodic memory. However, the common cause hypothesis could equally explain the pattern of findings that we observed. The fourth mechanism is the diminished availability of cognitive resources that are occupied in support of listening in difficult conditions [28, 29]. It is likely that all mechanisms work in combination with some level [27, 30]. However, our results do not allow for a conclusive distinction between hypotheses. Further work is needed to establish the specific effects of hearing trajectory membership on social engagement prior to the examination of outcomes.

The finding that improvement in hearing was associated with significantly poorer memory outcomes in HRS was unexpected (n = 6,365). The plausible explanation is that as the number of respondents within this trajectory (n = 387) is considerably smaller than in the others, this finding may be down to measurement error, which may be further biased by the self-report nature of the hearing variable. One of the major sources of measurement error in data is self-reported bias. Those small proportion of respondents (6%) may be biased in reporting their hearing function over time and feel that their hearing is better over time as they get used to that condition. Descriptive statistics show the baseline socio-economic characteristics of those whose hearing improves to be poorer than those whose hearing is better at baseline, even those whose hearing declines over the data period. This is in line with other research that has shown that hearing aid use does not improve cognitive ability among individuals with hearing impairments [31, 32]. It should also be noted that the nature of the trajectory modelling does not show the length of time for which respondents within this group reported poor hearing before reporting an improvement. Therefore, a large number may have reported poor hearing across the majority of the eight-year study period, and the measured time after improvement has occurred may be too little to show any true effect.

The potential clinical impact of this study is that we need to ensure that we enquire about hearing problems when older adults connect with healthcare professionals to ensure they have necessary interventions. Strengths of the study include the large sample sizes used in both ELSA and HRS and the fact that both datasets are nationally representative. ELSA and HRS include very similar variables enabling a direct comparison between the two. There are some limitations to the study. Firstly, the hearing variable was self-reported rather than objectively measured, which may provide some level of bias in the results. Although some studies have shown a good correlation between self-reported and objectively measured hearing [33, 34], others have shown this correlation to be strongly dependent on socioeconomic factors, such as age, sex, income and educational attainment [35, 36] and that self-reported measures of hearing are influenced by factors such as mental wellbeing, personality and general health [37, 38]. Furthermore, it has been postulated that self-reported measures of health should be treated with caution as they measure ‘disability’ rather than ‘impairment’ [39], and impairments, which may bear a stronger impact on cognitive function, may be under-identified [40]. An impairment is an actual condition, while a disability is the restriction of ability caused by the condition. However, self-reported measures of sensory ability are commonly used in observational epidemiological studies, and so the results of this work are comparable with similar research [41]. Another limitation is that episodic memory does not capture all aspects of cognition, and it may be the case that other cognitive abilities are affected to a greater extent by patterns of hearing ability in older age. A study by Deal and colleagues has found that hearing loss is associated with a faster decline of three domains of cognitive function: memory, perceptual speed and processing speed [42]. However, episodic memory has been shown to have good validity [43] and correlate well with other measures of cognition as well as dementia [44, 45] and relate especially well to the everyday activities and decisions of older individuals [46]. Third, as with any longitudinal survey, both ELSA and HRS are subject to attrition, with those most likely to drop out of the studies those who have poorer health, including cognitive function and hearing ability. As the episodic memory scores were only measured at the final wave, the respondents who attended all waves and were included in this study may have better health and higher memory scores than those who dropped out. This may lead to a smaller magnitude of effects being observed than might be observed in a population unaffected by drop-out. Finally, the observational design of this study means the relationship between hearing trajectory and cognition may be affected by confounding factors uncaptured by the data. The directionality of the relationships between sensory impairment and cognitive decline thus could not be inferred from this study.

In this study, continuously poorer hearing was associated with a poorer cognitive outcome as measured by episodic memory than those whose hearing was consistently good. Poor hearing may have some detrimental effects even years before cognitive functioning is measured. From an interventional perspective, preventing hearing loss in the first instance, perhaps through early uptake of hearing aid use, might prevent the onset of poorer episodic memory, which, in turn, is associated with a high risk of subsequent dementia [47]. Further studies are needed to properly assess that point in the trajectory of hearing loss such intervention might prevent negative outcomes.

Declaration of Conflict of Interest

None.

Declaration of Sources of Funding

The English Longitudinal Study of Ageing was developed by a team of researchers based at University College London, NatCen Social Research, the Institute for Fiscal Studies, the University of Manchester and the University of East Anglia. The data were collected by NatCen Social Research. The funding is currently provided by the National Institute on Ageing in the USA (grant numbers: 2RO1AG7644 and 2RO1AG017644-01A1), and a consortium of UK government departments coordinated by the National Institute for Health Research. Funding has also been received by the Economic and Social Research Council. The current analysis was supported by an International Project Grant from the Royal National Institute for Deaf People (RNID) and Alzheimer’s Research UK. Piers Dawes was supported by the NIHR Manchester Biomedical Research Centre.

Data Availability Statement

Data are available in a public, open-access repository. The English Longitudinal Study of Ageing dataset is available in a public, open-access repository and can be accessed through the UK Data Service at: https://ukdataservice.ac.uk/. The data can be used after registration and acceptance of end user licence.