Is there a dose–response relationship between musical instrument playing and later-life cognition? A cohort study using EPIC-Norfolk data

Abstract

Introduction

Musical instrument playing provides intellectual stimulation, which is hypothesised to generate cognitive reserve that protects against cognitive impairment. Studies to date have classified musicianship as a binary entity. This investigation draws on the dataset of the European Prospective Investigation of Cancer Norfolk study to examine the effect of frequency of playing on later-life cognition.

Methods

We compared three categorisations of self-reported musical playing frequency in late mid-life (12-month period) against cognitive performance measured after a 4–11 year delay, adjusted for relevant health and social confounders. Logistic regression models estimated the adjusted association between frequency of musical playing and the likelihood of being in the top and bottom cognitive deciles.

Results

A total of 5,693 participants (745 musicians) provided data on music playing, cognition and all co-variables. Classification of musicianship by frequency of playing demonstrated key differences in socio-demographic factors. Musicians outperformed non-musicians in cognition generally. Compared with non-musicians, frequent musicians had 80% higher odds of being in the top cognitive decile (OR 1.80 [95% CI 1.19–2.73]), whereas musicians playing at any frequency had 29% higher odds (95% CI 1.03–1.62). There was evidence of a threshold effect, rather than a linear dose–response relationship.

Discussion

This study supports a positive association between late mid-life musical instrument playing and later-life cognition, although causation cannot be assumed. Musicians playing frequently demonstrated the best cognition. ‘Musicians’ are a heterogeneous group and frequency of music playing seems a more informative measure than binary classification. Ideally, this more nuanced measure would be collected for different life course phases.

Key points

• Dose–response relationship between late mid-life music playing and later-life cognition.

• Cognitive reserve literature should include nuanced measures of exposure.

• Musical instrument playing provides intellectual stimulation, which is hypothesised to generate cognitive reserve.

Introduction

It is hypothesised that high levels of intellectually stimulating leisure activities across the life course produces a cognitive reserve that mitigates against overt cognitive impairment in the face of neuropathology [1–3]. Playing a musical instrument or singing involves several high-level cognitive processes, as well as sensory and motor components, and is therefore considered a viable source of that intellectual stimulation [4–7].

To date, intellectually stimulating leisure activities have been grouped in published analyses in exploration of their relationship with cognition. Many have demonstrated promising results, but a limitation remains that the included leisure activities are heterogeneous in both quantity and quality of potential cognitive stimulation. A smaller set of publications has focussed specifically on musical instrument playing or singing, also suggesting an association with better later-life cognition [8–11]. However, the evidence to date has tended to consider musicianship as a binary entity, for example, including only ‘advanced’ players as musicians [12], considering people musicians if they report currently playing ‘frequently and/or occasionally’ [13] or considering people musicians only if they have played for ≥10 years [14]. The implication of these ascertainment measures is that all musicians are the same, whereas the reserve hypothesis states that it is high levels of intellectual stimulation across the life course that conveys cognitive protection [15]. A gap remains therefore about whether ‘dose’ of musical instrument playing might differentially be associated with later-life cognition.

Methods

This study investigated, in a longstanding mid- to later-life cohort in the United Kingdom, the association between time spent playing a musical instrument or singing, measured in minutes per week (ascertained 1998–2000) and cognition (ascertained 2004–11), adjusted for age, sex, education, social class, alcohol, smoking, general health, depression and physical activity.

The data came from the European Prospective Investigation of Cancer Norfolk (EPIC-Norfolk) study, described elsewhere [16]. Briefly, EPIC-Norfolk recruited 30,445 participants, commencing in 1993, aged 40 and 79 from Norfolk, United Kingdom. The response rate at baseline was 39.2%. The cohort was comparable with those recruited at the time for the Health Survey for England, except for a lower smoking profile [16].

Data on musical instrument playing or singing were collected by questionnaire during the 2nd postal follow-up. Of the 30,445 participants from baseline, 8,046 were not approached for the 2nd follow-up because they had died, requested no further correspondence or their address could not be found. Therefore, 22,399 participants were approached, of whom 15,672 (70.0%) completed the questionnaire. Participants were asked how often in the last 12 months they had engaged in playing a musical instrument or singing, with an eight-point answer scale ranging from ‘none’ to ‘6 times a week or more’. Participants were also asked to provide the average time per episode. Musical playing was converted to average minutes per week by combining data on frequency and duration of episodes. To aid interpretation, these data were categorised into ‘occasional musicians’ (<1 h/week on average), ‘moderate musicians’ (≥1 h/week on average, but not most days) and ‘frequent musicians’ (≥1 h on ≥4 days/week on average). Detailed information on this process is in Supplementary Material A1.

Cognition data were collected in person using the Short Form Extended Mental State Exam (SF-EMSE) [17], during the 3rd health examination, which was attended by 8,623 (46.9%) of the 18,380 invited. The full version of the EMSE [18] is an extension of the Mini-Mental State Examination (MMSE). The MMSE covers several cognitive domains [19], is ubiquitous in clinical practice [20] and is validated as a screening test for cognitive decline [21]. However, the MMSE has a ceiling effect [18]. A previous study comparing cognitive measures in EPIC-Norfolk reported that 27.0% of participants scored full marks on a shortened version of the MMSE, compared with 2.4% on the SF-EMSE [17]. The SF-EMSE is marked out of 37 and drops some items, which only test the lower end of the cognitive range. The nature and intensity of EPIC-Norfolk meant that very few participants in the early phases were likely to require proxy consent with a relatively healthy cohort at outset, and this was not therefore undertaken. The SF-EMSE was an appropriate adaptation to take this into account with an ability to differentiate across the cognitive spectrum and with wide coverage across cognitive domains.

Cognition is highly related to many factors, which could be considered confounders or highly correlated co-variables. These were collected by questionnaire during either the baseline stage (age, sex, education and social class) or during the 2nd postal follow-up (smoking, alcohol, general health, depression and physical activity). A previously published analysis of attrition between baseline and the 3rd health examination reported that retained participants were more likely to be younger, of higher socioeconomic status and generally healthier. However, the cohort retained in the 3rd health examination remained diverse with regards to socioeconomic and lifestyle factors [22].

Education level was characterised by self-reporting of whether the participant had achieved no qualifications, O levels, A levels or a degree. Social class was characterised by occupation, using the Registrar General’s Classification [23]. Smoking (pack years) and alcohol (units/week) were derived from simple questions. Participants were asked to rate their general health as either poor, moderate, good or excellent. Participants were also asked whether their doctor had ever told them that they had depression, which required treatment, with a yes/no response. Physical activity was scored on a four-point scale from inactive to active, using a set of questions validated against objective measures of energy expenditure and cardiorespiratory fitness [24].

Governance and statistical analysis

EPIC-Norfolk was approved by the Norfolk Local Research Ethics Committee (05/Q0101/191) and East Norfolk and Waveney National Health Service Research Governance Committee (2005EC07L). Participants gave signed informed consent. This study has been reported using the Strengthening The Reporting of OBservational studies in Epidemiology (STROBE) guidance, and the completed checklist is attached as Supplementary Material A1. The statistical analysis plan was determined a priori. All statistical analysis was performed using Stata SE 16.1.

Any nonsense or clearly outlying values for any variables were recoded as missing. Missing data for all co-variables were reported. Attrition was analysed using appropriate statistical testing. In keeping with previous cognitive assessments of this cohort [17], logistic regression was used to compare likelihood of being in the top and bottom deciles of cognition scores. Due to skewness of the cognition data, multiple linear regression analysis was performed after a Box-Cox transformation, and the results of this are included as an appendix in Supplementary Material A1.

Results

Participants

Of the 15,672 who took part in the 2nd postal questionnaire, 15,293 (97.6%) completed the question on musical playing. A total of 1,758 (11.5%) were classified as musicians, of whom 527 were classified as ‘occasional musicians’, 831 as ‘moderate musicians’ and 400 as ‘frequent musicians’. Of the 1,758 musicians, only 68 (3.9%) had missing music playing time data requiring imputation. The 68 participants with missing data were relatively even distributed across music playing frequencies. Of the 15,293 participants with exposure data, 6,735 (44.0%) provided outcome data. Of the 1,758 musicians, 871 (49.5%) provided cognition data (occasional musicians n = 285, moderate musicians n = 408 and frequent musicians n = 178).Those lost to follow-up were younger, less well educated and of lower social class, heavier smokers, in poorer general health and less physically active, on average compared with those retained (see Supplementary Material A1 for more details). There was no evidence of differential attrition between musicians and non-musicians (Supplementary Material A1). Missing data were not a large problem (generally <5% for each variable), and there were minimal differences in availability between musicians and non-musicians (Supplementary Material A1). The final analysis cohort included 5,693 participants.

Socio-demographic characteristics

Table 1 shows the demographic characteristics of the analysis cohort, (sex-stratified data are shown in Supplementary Material A1, and show some subtle differences in the trends for depression and physical activity by gender and musician group). Overall, musicians were generally healthier and more affluent than their non-musician counterparts. However, musicians were also more likely to report a history of depression. When compared with other musicians, frequent musicians were more likely to be older, male, better educated, to smoke more and drink more alcohol.

Table 1.

Demographics of the analysis cohort, stratified by musician group, with sensitivity testing

		Non-musicians	Occasional musicians	Moderate musicians	Frequent musicians
Number of participants		4,948	249	338	158
Age (years)	Mean (SD)	59.1 (7.9)	58.8 (8.3)	59.5 (8.2)	60.8 (8.1)
	P value		0.630	0.412	0.009
Sex (%)	Male	44.5	36.6	37.3	43.9
	P value		0.013	0.009	0.376
Education level (%)	Low	26.8	19.3	14.8	12.0
	O level	12.5	9.6	11.5	7.0
	A level	45.3	37.4	42.9	47.5
	Degree	15.5	33.7	30.8	33.5
	P value		<0.001	<0.001	<0.001
Social class (%)	1	3.4	0.4	0.9	2.5
	2	13.6	11.2	9.2	5.1
	3a	13.4	9.6	5.3	8.2
	3b	28.3	20.9	28.7	23.4
	4	36.3	48.2	48.5	53.8
	5	5.2	9.6	7.4	7.0
	P value		<0.001	<0.001	<0.001
Smoking (pack years)	Median (IQR)	0 (0–10)	0 (0–2.5)	0 (0–5)	0 (0–7.5)
	P value		<0.001	0.003	0.159
Alcohol (units/week)	Median (IQR)	4 (1.5–10.5)	3 (1–8.5)	3 (1.5–8.5)	4.5 (1.5–10)
	P value		0.003	0.026	0.904
General health (%)	Poor	0.6	0.0	0.3	1.3
	Moderate	12.0	8.8	10.4	11.4
	Good	67.5	70.3	63.3	60.8
	Excellent	19.9	20.9	26.0	26.6
	P value		0.284	0.048	0.122
Depression (%)	Yes	15.4	18.5	17.8	19.6
	P value		0.188	0.244	0.147
Physical activity (%)	Inactive	27.6	25.3	29.0	23.4
	M inactive	28.5	36.6	32.0	37.3
	M active	22.9	22.9	23.4	24.7
	Active	21.1	15.3	15.7	14.6
	P value		0.023	0.113	0.03

Social class: 1, unskilled; 2, partially skilled; 3a, skilled manual; 3b, skilled non-manual; 4, technical; 5, professional. Physical activity: M, moderately.

P values calculated using non-musicians as the reference category; using t-test for age; chi-squared test for sex, education, social class, general health, depression and physical activity; and Wilcoxon rank sum test for pack years and units of alcohol.

Cognition data

Scores ranged from 0–37, with a median score of 33 (interquartile range 31–35, skew −2.61). Only 121 (2.1%) of participants scored full marks and 12 participants (0.2%) scored <10. The EPIC-Norfolk team clarified that scores below 10 were feasible results, rather than data entry errors, with the caveat that they likely represent participants who started but declined to finish the test, in which case participants were given a score of how many questions they had answered correctly to that point. In light of this, these cognition scores were considered valid results and included in the analysis. As the SF-EMSE data were discrete, it was not possible to create clean deciles of performance. The best approximations of top and bottom deciles were that 12.0% of participants scored ≥36 and 8.2% of participants scored ≤28.

Being a woman, having better education and higher social class, reporting higher alcohol intake, reporting depression and increased physical activity were all associated with statistically significantly better cognition in univariate analysis, whereas increasing age and smoking pack years were both associated with poorer cognition. None of the interaction terms were statistically significant.

Regression analysis

Tables 2 and 3 show the results of fully adjusted logistic regression models for each music playing subgroup being in the top and bottom cognitive deciles, respectively. Musicians were statistically significantly more likely to be in the top decile for cognition scores, compared with non-musicians. This was particularly pronounced for frequent musicians, who had 80% greater odds of being in the top cognition decile than non-musicians. Similar trends were seen for the bottom cognitive decile, though these did not reach statistical significance.

Table 2.

Multiple logistic regression models for likelihood of being in the top cognitive decile for musicians (combined, and at each frequency subcategory) compared with non-musicians

		Unadjusted model [OR (95% CI) P value]	Fully adjusteda model [OR (95% CI) P value]
	n	5,693	5,693
Musicians (versus non-musicians)	All	1.64 (1.33, 2.03) P < 0.001	1.29 (1.03, 1.62) P = 0.025
	Occasional	1.83 (1.31, 2.54) P < 0.001	1.40 (0.98, 1.98) P = 0.061
	Moderate	1.33 (0.97, 1.82) P = 0.075	1.02 (0.73, 1.42) P = 0.922
	Frequent	2.07 (1.40, 3.07) P < 0.001	1.80 (1.19, 2.73) P = 0.005

^aAdjusted for age, sex, education, social class, smoking, alcohol, general health, depression and physical activity.

Table 3.

Multiple logistic regression models for likelihood of being in the bottom cognitive decile for musicians (combined, and at each frequency subcategory) compared with non-musicians

		Unadjusted model [OR (95% CI) P value]	Fully adjusteda model [OR (95% CI) P value]
	n	5,693	5,693
Musicians (versus non-musicians)	All	0.69 (0.50, 0.94) P = 0.021	0.84 (0.60, 1.17) P = 0.301
	Occasional	0.73 (0.44, 1.23) P = 0.242	0.89 (0.52, 1.52) P = 0.660
	Moderate	0.78 (0.51, 1.21) P = 0.266	0.97 (0.62, 1.53) P = 0.896
	Frequent	0.42 (0.19, 0.96) P = 0.040	0.51 (0.22, 1.17) P = 0.113

^aAdjusted for age, sex, education, social class, smoking, alcohol, general health, depression and physical activity.

Supplementary Material A1 reports the results of the multivariable linear regression model, after using a Box-Cox transformation to normalise the cognition data. As for the logistic regression models, musicians as a whole group had statistically significantly better cognition than non-musicians (P = 0.035), and this was most pronounced in frequent musicians (P = 0.030).

Sensitivity analysis

Occasional musicians surprisingly outperformed moderate musicians. It was noted that a small number of musicians who reported moderate (n = 4) and frequent (n = 1) music playing had outlying low values on the SF-EMSE. To investigate how much these had affected the results of the analyses, a post hoc sensitivity analyses was performed in which the adjusted regression models were re-run, with the five outliers excluded. The results can be seen in Supplementary Material A1 and show that, after exclusion of the outlying values, moderate musicians performed relatively similarly to occasional musicians, although occasional musicians remained a lot more likely to be in the top cognitive decile.

Discussion

Main findings

In this cohort study of 15,293 participants from EPIC-Norfolk, musical instrument playing and singing were found to be associated with better cognition after 4–13 years (mean 9.3 years) of follow-up. There was evidence of possible threshold effect, with frequent musicians (playing for an average of at least an hour most days) having the best cognitive performance across all analyses. Frequent musicians had 80% higher odds of being in the top decile of cognitive function, whereas musicians playing at any frequency had 29% higher odds, than non-musicians.

At lower frequencies of playing, occasional musicians (playing for an average of less than an hour per week) surprisingly outperformed moderate musicians (playing for an average of at least an hour per week, but not most days). A sensitivity analysis, which removed four outlying moderate musicians who scored much lower than the rest of the subgroup, found that the remaining moderate musicians were, on average, similar to occasional musicians. This suggests that, rather than demonstrating a linear dose–response relationship, there may be a baseline benefit for those playing at any frequency, with a threshold effect of much greater cognitive benefits for those playing most regularly. Alternatively, this could represent complex relationships between co-variables, such as age and socioeconomic status, and the cognitive benefits of musical playing.

Strengths and limitations

The strengths include that the study is large and was recruited from the population. The granularity of the exposure measure was high, with eight categories of frequency, and continuous data on time per episode. The adjustment for potential confounders was rigorous, and there were very small amounts of missing data. Finally, SF-EMSE differentiates between those across the cognitive spectrum and tests general cognitive function across multiple domains, meaning that there is greater applicability of findings compared with several previous studies, which measured very specific cognitive functions. Although the delay in measuring cognition of 4–13 years is strength of this study, cognitive decline towards dementia is mostly gradual, with longitudinal studies of cognition revealing decline for several years before diagnosis of dementia [25,26]; therefore, reverse causality is minimised rather than eliminated.

There are three main limitations of this study. First, we could not combine the exposure data with a measure of the number of musical playing years, or the age of onset, prohibiting assessment of total life course musical exposure. Second, all musical instrument playing and singing were grouped together for the purposes of the questionnaire. Playing different types of musical instrument require different types of cognitive processing, and singing may not stimulate visuospatial elements of cognition in the same way that playing an instrument does. This has led to calls for more research into the effects of different types of instrument playing [10]. Lastly, EPIC-Norfolk contains only one measurement of cognition, precluding any analysis of change in cognition over time, and data on dementia incidence were unavailable. Cognition is not a direct measure of cognitive reserve or of likelihood of cognitive decline.

Findings in the context of existing knowledge

Previous studies have tended to dichotomise music playing as ‘musician’ and ‘non-musician’, often based on years of playing, with a range of cut-offs applied. The subcategorisation of music playing by current frequency of playing produced meaningful results, both in terms of their characteristics and their cognition scores. Frequent musicians were older than other musicians, which may demonstrate that people are more likely to engage frequently in leisure activities when they are retired and have more free time. Frequent musicians were slightly more likely to have a degree, but were much less likely to be in a professional occupation than occasional musicians.

Implications for policy, practice and research

This study supports the growing evidence that musical instrument playing across the life course is associated with better cognition and protection against dementia. Clear benefits were observed for those playing frequently, over other musicians. Arts participation is heavily influenced by childhood experiences, and whether engagement is perceived as a social norm [27]. Therefore, policies should encourage activities like musical instrument playing across the life course and across society.

‘Musicians’ are a heterogeneous group, with regards to socio-demographic and cognitive factors. Therefore, research into the effects of musical instrument playing (or any other cognitively stimulating activity) on cognitive reserve, later-life cognition, or incident dementia risk, are likely to benefit from categorisation of the frequency and duration of musical instrument playing, in addition to the number of years played for, rather than just a binary classification of ‘musicians’ versus ‘non-musicians’. This is also likely to hold true for research on other sources of life course cognitive stimulation.

Declaration of Sources of Funding

S.W. is an National Institute for Health Research (NIHR) funded Academic Clinical Fellow, award number: ACF-2019-14-008 NIHR Academic Clinical Fellowship. This work was supported by the Medical Research Council, UK, http://www.mrc.ac.uk/ (MRC, Ref: G0401527) and Cancer Research UK, http://www.cancerresearchuk.org/ (CRUK, Ref: C864/A8257). The clinic for EPIC-Norfolk 3 was funded by Research into Ageing, now known as Age UK, http://www.ageuk.org.uk/ (Grant Ref: 262). The pilot phase was supported by MRC (Ref: G9502233) and CRUK (Ref: C864/A2883). The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.

Declaration of Conflict of Interest

None.