Results of the search

The total number of included studies for this update was 22. For the first version of this review (Vink 2003), we identified 354 references related to music‐based interventions and dementia (Figure 1). Of those, on the basis of the abstracts, 254 were discarded as they did not refer to a research study or were identified as anecdotal or reports of case studies. Hard copies were obtained for the initially remaining 100 studies in 2003. We then discarded a further 74 studies as they involved participant series or case studies. As a results, 26 studies remained in 2003, of which five met the criteria for inclusion at that time (Brotons 2000; Clark 1998; Gerdner 2000; Groene 1993; Lord 1993). In 2008, an additional 18 studies were reviewed, of which three studies met the criteria (Svansdottir 2006; Raglio 2008; Sung 2006). For the update of 2010, we retrieved 188 references of possible relevance. After a first assessment, 16 references remained which were further assessed, of which two studies met the criteria of this review (Guétin 2009; Raglio 2010a). In total, 10 studies were included in the previous update. In 2015, due to clarified criteria for eligibility of interventions, randomisation and more stringent application of criteria for analyses of outcomes after a minimum number of sessions, we excluded five of the 10 previously included studies (Brotons 2000; Gerdner 2000; Groene 1993; Raglio 2008; Sung 2006; see Characteristics of excluded studies table). However, we included 12 new studies after evaluating 121 references including 25 full‐text evaluations, which resulted in 17 included studies. A new search on 12 April 2016 identified eight potentially eligible additional studies which warranted review against inclusion criteria (Curto Prieto 2015; Hsiung 2015; Hsu 2015; Raglio 2015; Rouch 2017; Thornley 2016; 신보영, 황은영 2015; 채경숙 2015), in addition to one study for which we were waiting for clarification from the authors about the results (Hong 2011). The latest search was performed 19 June 2017. We identified a new eligible study (Cho 2016), and we included four studies that had been awaiting classification (Hsu 2015; Lyu 2014; Raglio 2015; Thornley 2016; from which we could extract data with the help of collaborators). We excluded 채경숙 2015 (see Characteristics of excluded studies table) and remaining potentially eligible studies are listed in the Characteristics of studies awaiting classification and Characteristics of ongoing studies tables.

Figure 1

---

---

Study flow diagram.

Included studies

Details of the included studies are presented in the Characteristics of included studies table. One article (Narme and colleagues 2012: Narme 2012‐study 1 and Narme 2012‐study 1a) reported on two studies with rather similar designs indicated with study 1 and study 2 in the article (note that study 2 is indicated with 1a in our analyses). More articles with additional results or background of the study were available for five studies (Cooke 2010; Lin 2011; Narme 2014; Raglio 2010a; Vink 2013).

Nineteen studies had a parallel‐group designs (Ceccato 2012; Cho 2016; Guétin 2009; Hsu 2015; Liesk 2015; Lin 2011; Lord 1993; Lyu 2014; Narme 2012‐study 1; Narme 2012‐study 1a (also referred to as study 2); Narme 2014; Raglio 2010a; Raglio 2010b; Raglio 2015; Sakamoto 2013; Sung 2012; Svansdottir 2006; Thornley 2016; Vink 2013); and three used a cross‐over design with first‐period data available for all (Clark 1998; Cooke 2010; Ridder 2013).

The 22 studies were performed in 14 countries. Whereas the two oldest studies and one recent study were from the USA (Cho 2016; Clark 1998; Lord 1993), the studies published after 1998 were from a variety of other regions and countries: 13 studies conducted in eight countries in Europe (Italy, France, Germany, the Netherlands, the UK and Iceland, including also one study performed in two countries, Denmark and Norway; Ridder 2013), four studies from three countries in Asia (Taiwan, Japan and China), one study from Australia and one from Canada. The studies were all performed in institutional settings of nursing homes, residential homes and hospital wards for older adults. Dementia severity varied. The total number of randomised participants varied between 14 (Narme 2012‐study 1a) and 120 (Raglio 2015), with a median number of 47 participants across the studies. Nine out of 22 randomised fewer than 40 participants, and only two had more than 100 participants. The total number of participants randomised over all studies was 1097.

The interventions were active (Cho 2016; Cooke 2010; Hsu 2015; Liesk 2015; Lyu 2014; Raglio 2010a; Raglio 2010b; Raglio 2015; Sung 2012; Thornley 2016); receptive (listening interventions while there was communication with the therapist, Clark 1998; Guétin 2009); or a mixture of the two forms (Ceccato 2012; Lin 2011; Lord 1993; Narme 2012‐study 1; Narme 2012‐study 1a; Narme 2014; Ridder 2013; Sakamoto 2013; Svansdottir 2006; Vink 2013). Appendix 2 describes the music‐based therapeutic intervention and other activities of all studies. Music included live or recorded music that met preferences of the group or individual. The active forms often combined playing of instruments and singing activities, and some also combined with movement such as clapping hands and dance. In seven studies, the intervention concerned an individual intervention. Sessions varied in duration between half an hour and two hours. The total number of sessions ranged from six (Narme 2012‐study 1) to 156 (Lord 1993), with a median total number of 14 sessions until the end of treatment assessment. The frequency ranged between one session per week (Guétin 2009; Hsu 2015; Sakamoto 2013) and seven sessions per week (daily, Lyu 2014) with a median and more typical number (mode) of two sessions per week (13 studies employed two per week). These figures probably reflected number of sessions offered, as the number of attended session may be lower. There were few reports about implementation fidelity including adherence and dose received. However, Ridder 2013 reported that a minimum of 12 sessions were offered, but the participants received a mean of 10 sessions, and Thornley 2016, in their study on an acute inpatient psychiatric unit within an academic hospital, mentioned that the participants enrolled in the study were generally hospitalised for two to three weeks, which limited the number of sessions attended.

In 12 of the studies, we could be sure from the report that the interventions had been delivered by an accredited music therapist (Ceccato 2012; Cho 2016; Hsu 2015; Lin 2011; Lyu 2014; Raglio 2010a; Raglio 2010b; Raglio 2015; Ridder 2013; Svansdottir 2006; Thornley 2016; Vink 2013). In four studies, it was unclear whether a music therapist was involved (no profession reported in the older studies, Lord 1993 and Clark 1998; probably delivered by trained music therapists but it was not stated explicitly in Guétin 2009; and delivered by musicians trained in the delivery of sessions and in working with older people with dementia but unclear if these were formally trained music therapists in Cooke 2010). In the other six studies, the intervention was not delivered by a music therapist (psychologist and other supervisor(s) with no training in music therapy: Narme 2012‐study 1; Narme 2012‐study 1a; Narme 2014; trained research assistants: Sung 2012; music facilitator: Sakamoto 2013; music teacher specialised in teaching older people: Liesk 2015). Nine studies selectively included people with agitation, mood or behavioural problems (Clark 1998; Cooke 2010; Guétin 2009; Hsu 2015; Raglio 2010a; Raglio 2015; Ridder 2013; Sung 2012; Vink 2013), while some studies (also) excluded people with major psychiatric conditions such as psychosis or major depression (Ceccato 2012; Cho 2016; Guétin 2009; Raglio 2015), or people with other medical conditions such as hearing impairment or acute illness.

Most studies compared the music intervention with an active control intervention with the same number of sessions and frequency as the music group. Two‐armed studies compared with the following interventions: reading (Cooke 2010; Guétin 2009), a cognitive stimulation intervention (Liesk 2015), painting (Narme 2012‐study 1), cooking (Narme 2012‐study 1a – also referred to as study 2; Narme 2014), or individual active engagement activities (Thornley 2016) or variable recreational activities which included handwork, playing shuffleboard, and cooking and puzzle games (Vink 2013). Five studies had three arms with the active control groups working on jigsaw puzzles (Lord 1993), reading familiar lyrics (Lyu 2014), television watching (Cho 2016), or receiving a passive group music intervention which did not meet our inclusion criteria for a therapeutic music‐based intervention (Cho 2016; Raglio 2015; Sakamoto 2013).

Outcomes that were assessed often were 'emotional well‐being' including quality of life, mood disturbance or negative affect (also as part of behavioural scales), and 'behavioural problems' (agitation or aggression, and behaviour overall) and 'cognition.' Social behaviour was less commonly assessed (Lord 1993; Narme 2012‐study 1; Narme 2012‐study 1a; Narme 2014); and the meta‐analyses of end‐of‐treatment scores included only the three studies from Narme and colleagues. The Cohen‐Mansfield Agitation Inventory (CMAI, for agitation; Cohen‐Mansfield 1986), Mini‐Mental State Examination (MMSE, for cognition; Folstein 1975), and the Neuropsychiatric Inventory (NPI, for behaviour; Cummings 1994) in particular were frequently used. Item‐level NPI outcome data were reported in the article or the author additionally provided data about depression, anxiety and agitation outcomes.

Excluded studies

We screened 769 records and we excluded 678 (Figure 1). Of the remaining 91 records examined in full text, we excluded 70 records (see Characteristics of excluded studies table for a selection of excluded studies which were close but did not qualify upon careful consideration). They were often excluded because the participants did not have dementia, or because of a trial design (i.e. not an RCT). Further, and often less obvious, we critically reviewed whether the intervention met the inclusion criteria for a music‐based therapeutic intervention, and whether the reported outcomes included any assessments after fewer than five sessions. There are a number of studies on group music interventions such as group music in addition to movement interventions (e.g. Sung 2006): these were excluded because music was not the main or only therapeutic element, or was not provided with individual therapeutic intent. Further, some studies assessed outcomes during the treatment sessions only, combining immediate effects, for example, on behaviour during the first session, with effects after multiple sessions (e.g. Gerdner 2000). Studies awaiting classification included conference abstracts and articles about studies in Asia which we could not retrieve or evaluate in time (see Characteristics of studies awaiting classification table).

Allocation

All included studies were RCTs. However, the randomisation procedure was not always described in detail (Figure 2). Moreover, allocation concealment was described and adequate in detail in six studies, all of which were published in 2010 or later (Cho 2016; Cooke 2010; Hsu 2015; Lin 2011; Raglio 2015; Ridder 2013). One older study stated that participants were "non‐systematically separated" into groups without further detail, which we considered posed a high risk of selection bias (Lord 1993). One study used cluster randomisation, but this study contributed only a maximum of 13 participants to the meta‐analyses (Hsu 2015).

Blinding

Blinding of therapists and participants to the intervention is not possible. Therefore, the studies were at high risk of performance bias even though therapists do not generally assess outcomes and participants may not be aware, have no specific expectations or were unable to self‐report. The outcomes were assessed unblinded, by the research team or unblinded nurses, in at least six studies (Figure 2). For example, Narme and colleagues described two studies differing in detection bias (Narme 2012‐study 1; Narme 2012‐study 1a). The first study involved a high risk of detection bias because the outcomes 'anxiety' (measured with the State‐Trait Anxiety Inventory for adults, STAI‐A) and, as assessed from the first two minutes of filmed interviews, 'emotions' (from facial expressions) and 'social behaviour' (discourse content), were assessed by nurses who were not blinded to the interventions (music intervention or painting) (Narme 2012‐study 1). By contrast, in the second study, risk of detection bias was low because five independent observers who were blinded for the type of intervention (music intervention or cooking) assessed the outcomes (Narme 2012‐study 1a). For all outcomes except for cognition, less than half of the number of patients participated in a study that was at high or unclear risk of detection bias (emotional well‐being including quality of life: 134/348 participants; depression: 140/503; anxiety: 117/478; agitation or aggression: 254/626; behavioural problems overall: 147/442; social behaviour: 22/70). For cognition, for 237/350 cases, risk of detection bias was unclear. Risk of performance bias, and for some outcomes also risk of detection bias, in several studies resulted in downgrading of the quality of the evidence for all end‐of‐treatment outcomes (summary of findings Table for the main comparison); and for all long‐term outcomes (summary of findings Table 2).

Incomplete outcome data

Self‐reported outcomes were rarely employed. Occasionally death, hospitalisation, acute illness or no interest in the therapy occurred across the different study arms; and cases with no outcome data were not included in the analyses. Incomplete outcome data were problematic in a few studies (Cho 2016; Hsu 2015; Thornley 2016). In Hsu 2015, three of nine participants in the intervention group died (and one of eight in the control group). In contrast, Cho 2016 lost nine of 17 participants in the television watching control group (and only a few in the other groups) and suggested this was because individual preferences for television programmes were not taken into account. Thornley 2016 did not perform their study in a long‐term care setting but in an inpatient psychiatric unit of a hospital and some participants were discharged after having attended a few sessions. The studies at high risk of attrition bias were three of the five studies added in this update. Newer studies often visualised cases lost to follow‐up and missing outcome assessment in detail using flow diagrams. The two oldest studies, and some newer studies, only reported the number of cases randomised (and analysed) and did not explicitly report reasons for missing outcome data by study arm, or how these were handled. Therefore, it was possible that attrition bias was problematic in more studies, but that the reporting of missing outcome data was better in newer studies.

Selective reporting

Most studies, including the newer studies, did not refer to initial plans, a study protocol or trial registration. Therefore, it was unclear to what extent bias due to selective outcome reporting was pertinent. We found some indication of inconsistent reporting of primary and secondary outcomes (Cooke 2010; Hsu 2015). Without these two studies, the pooled estimate for emotional well‐being and quality of life decreased from 0.32 to 0.23; other SMDs were similar. Only one study clearly referred to a change in initial plans (Ceccato 2012); and two studies referred to a trial registration, and outcome reporting was consistent with the registration for Sakamoto 2013 but not for Hsu 2015. We did not downgrade the quality of the evidence because of unclear risk of selective reporting.

Regarding publication bias, funnel plots for outcomes with sufficient studies (anxiety, 13 studies of which two with both a 'usual care' and 'other activity' control group, Figure 4; and agitation or aggression, 14 studies, also two with two types of control groups, Figure 5) did not clearly suggest possible publication bias.

Other potential sources of bias

We found some other potential sources of bias. Outcome assessment may be either imprecise or biased by the use of non‐validated outcome measures with suboptimal distributions (such as skewed distributions, e.g. number of times yelling was observed; Clark 1998) and different procedures for the baseline and outcome assessment (Sakamoto 2013). Further, we found problems with the reporting of outcomes or we suspected errors (Lord 1993; and for this reason, Hong 2011 is under Studies awaiting classification). Implementation fidelity, including non‐adherence, was infrequently described, but Liesk 2015, one of the few studies with null findings, reported on this in detail. Finally, there may be bias due to a financial or intellectual conflict of interest when funding was provided by a source with a potential interest in the effectiveness of music therapy. This may apply to two studies (Ceccato 2012; Ridder 2013), but it should be noted that no source of funding was reported for more studies (Clark 1998; Liesk 2015; Lin 2011; Lord 1993; Lyu 2014; Raglio 2010a; Raglio 2010b). Only six studies were both definitely delivered by a music therapist and funded by a source unrelated to music or music therapy (no potential financial conflict of interest, but at least for some, the music therapists (co)authored the article; Cho 2016; Hsu 2015; Raglio 2015; Svansdottir 2006; Thornley 2016; Vink 2013).

Emotional well‐being including quality of life

We included nine studies with 348 participants in the analysis of end‐of‐treatment scores for the critically important outcome of emotional well‐being and quality of life. Most studies used a validated quality‐of‐life or well‐being measure for more direct observation; the Dementia Quality of Life (DQOL) (Cooke 2010); a German translation of the Dementia Quality of Life Instrument (DEMQOL) (Liesk 2015); a Danish translation of the Alzheimer's Disease‐Related Quality of Life (ADRQL) (Ridder 2013); the Cornell‐Brown Scale for Quality of Life in Dementia (CBS‐QoL) – although it was unclear if this was a validated translated version (Raglio 2015); a Dementia Care Mapping Wellbeing score (Hsu 2015); and the Quality of Life‐Alzheimer's Disease (QOL‐AD) (Cho 2016). In the three studies conducted by Narme and colleagues, emotional well‐being referred to counts of positive and negative facial expressions as assessed from the first two minutes of filmed interviews (Narme 2012‐study 1; Narme 2012‐study 1a; Narme 2014). We found evidence of an effect at the end of treatment (SMD 0.32, 95% CI 0.02 to 0.62; Analysis 1.1; Figure 6; summary of findings Table for the main comparison). Heterogeneity was low to moderate (I² = 40%; Chi² P = 0.09). There was no blinding of outcome assessment in four of the nine studies. The overall quality for effects of music‐based interventions on emotional well‐being and quality of life at end of treatment was low, downgraded for serious risk of bias and imprecision (wide CI). The quality was also low for long‐term outcomes for which there were only four studies (180 participants; Hsu 2015; Narme 2012‐study 1a; Narme 2014; Raglio 2015). The SMD was similar to the SMD at the end of treatment but the imprecision was greater so we were less certain of the direction of the effect (SMD 0.34, 95% CI –0.12 to 0.80; I² = 46% Chi² P = 0.12; Analysis 2.1; summary of findings Table 2).

Figure 6

---

---

Forest plot of comparison: 1 Music‐based therapeutic interventions versus usual care or versus other activities: end of treatment, outcome: 1.1 Emotional well‐being and quality of life. CI: confidence interval; SD: standard deviation.

Mood disturbance or negative affect: depression

Eleven studies contributed 503 participants to the analysis on end‐of‐treatment effect (Figure 7), and six studies contributed 354 participants to the analysis on long‐term effects. Depression or depressive symptoms were measured with (translated versions of) the Geriatric Depression Scale (GDS), the Cornell Scale for Depression in Dementia, or with a subscale of the Behavioural Pathology in Alzheimer's Disease (BEHAVE‐AD) or the NPI. Heterogeneity was not important (I² = 0%) for either end‐of‐treatment or long‐term outcomes. We downgraded both outcomes for risk of bias, due to lack of blinding in many studies. Imprecision was more of an issue for long‐term outcomes. The overall quality of the evidence was moderate for end‐of‐treatment effects and low for long‐term outcomes. We found that music‐based therapeutic interventions probably reduced depressive symptoms at the end of treatment (SMD –0.27, 95% CI –0.45 to –0.09; Analysis 1.2; Figure 7; summary of findings Table for the main comparison). There was no evidence of a reduction in the longer term, with a smaller estimate and a CI including no effect (SMD –0.03, 95% CI –0.24 to 0.19; Analysis 2.2; summary of findings Table 2).

Figure 7

---

---

Forest plot of comparison: 1 Music‐based therapeutic interventions versus usual care or versus other activities: end of treatment, outcome: 1.2 Negative affect or mood disturbances: depression. BEHAVE‐AD: Behavioural Pathology in Alzheimer's Disease; NPI: Neuropsychiatric Inventory; SD: standard deviation.

Mood disturbance or negative affect: anxiety

The other mood item we considered was anxiety. For this outcome, at the end of treatment, we included 13 studies with 478 participants. A variety of (translated) outcome measures were used; Rating Anxiety in Dementia Scale (RAID), STAI‐A, Hamilton Anxiety Scale, and subscale scores of the BEHAVE‐AD and NPI. Heterogeneity was substantial for end‐of‐treatment effects (I² = 53%; Chi² P = 0.008) and longer‐term effects (I² = 63%; Chi² P = 0.01). In addition to serious inconsistency, we downgraded the quality for lack of blinding. We did not find clear evidence of publication bias (Figure 4). We judged the quality of the evidence as low at the end of treatment and, for the longer‐term outcome, very low because there was also imprecision. Therefore, we can have little or very little confidence in the results. Anxiety was lower in the music intervention group at the end of treatment (SMD –0.43, 95% CI –0.72 to –0.14; 13 studies, 478 participants; Analysis 1.3; Figure 8; summary of findings Table for the main comparison). In the longer term, we could not be certain of either the size or the direction of effect (SMD –0.28, 95% CI –0.71 to 0.15; 6 studies, 265 participants; Analysis 2.3; summary of findings Table 2).

Figure 8

---

---

Forest plot of comparison: 1 Music‐based therapeutic interventions versus usual care or versus other activities: end of treatment, outcome: 1.3 Negative affect or mood disturbances: anxiety. BEHAVE‐AD: Behavioural Pathology in Alzheimer's Disease; NPI: Neuropsychiatric Inventory; SD: standard deviation; STAI‐A: State‐Trait Anxiety Inventory for Adults.

Behavioural problems: agitation or aggression

Fourteen studies with 626 participants contributed to the end‐of‐treatment effect analysis, and five studies with 330 participants contributed to the long‐term effect analysis. Outcome measures used for agitation were (translated versions of) the CMAI and the agitation subscale of the NPI; and for aggression, the aggressiveness subscale of the BEHAVE‐AD and counts of observed aggressive behaviour. Heterogeneity was not important at the end of treatment (I² = 9%, Chi² P = 0.35) and longer term (I² = 6%, Chi² P = 0.38). Inconsistency and imprecision were not serious for effects on agitation or aggression at the end of treatment, but imprecision was serious for effects on the long‐term outcome. There was no evidence of publication bias (regarding end‐of‐treatment effect; Figure 5). We rated the quality of the evidence as moderate for the end‐of‐treatment outcome but low for the long‐term outcome. We found no evidence of an effect on agitation or aggression at the end of treatment (SMD –0.07, 95% CI –0.24 to 0.10; Analysis 1.4; Figure 9; summary of findings Table for the main comparison) or in the long term (SMD –0.10, 95% CI –0.33 to 0.13; Analysis 2.4; summary of findings Table 2).

Figure 9

---

---

Forest plot of comparison: 1 Music‐based therapeutic interventions versus usual care or versus other activities: end of treatment, outcome: 1.4 Problematic behaviour: agitation or aggression. BEHAVE‐AD: Behavioural Pathology in Alzheimer's Disease; CI: confidence interval; CMAI: Cohen‐Mansfield Agitation Inventory; NPI: Neuropsychiatric Inventory; SD: standard deviation.

Behavioural problems: overall

Ten studies with 442 participants contributed to the end‐of‐treatment effect analysis, and six studies with 351 participants contributed to the analysis of longer‐term effects. Outcome measures were (translated versions of) the BEHAVE‐AD and NPI. Heterogeneity was low for the end of treatment effect (I² = 19%, Chi² P = 0.25). The quality of the evidence was moderate due to lack of blinding. We found evidence of an effect of music‐based therapeutic interventions on problematic behaviour overall at the end of treatment (SMD –0.23, 95% CI –0.46 to –0.01; Analysis 1.5; Figure 10; summary of findings Table for the main comparison). There was no convincing evidence of a long‐term effect because of imprecision (SMD –0.19, 95% CI –0.51 to 0.14; I² = 51%, Chi² P = 0.05; Analysis 2.5; summary of findings Table 2). Therefore, heterogeneity was moderate, and the quality of the evidence was low due to imprecision in addition to lack of blinding.

Figure 10

---

---

Forest plot of comparison: 1 Music‐based therapeutic interventions versus usual care or versus other activities: end of treatment, outcome: 1.5 Problematic behaviour overall. NPI: Neuropsychiatric Inventory; SD: standard deviation.

Social behaviour: music versus other activities

The three studies of Narme and colleagues) contributed 70 participants to the end‐of‐treatment effect analysis (Narme 2012‐study 1; Narme 2012‐study 1a; Narme 2014), and two of them contributed 48 participants to the analyses of longer‐term effects (Narme 2012‐study 1a; Narme 2014). For all, the outcome was the contents of conversation (positive versus negative expressions when interviewed about current feelings and personal history). Lord 1993 reported effects on their self‐made questionnaire on social interaction, mood and recall (combined outcome), but there were no separate figures for social interaction and therefore we could not use the data for the meta‐analysis. We downgraded the evidence at both time points due to serious or very serious risk of bias and very serious imprecision. There was also moderate to substantial heterogeneity in the long‐term analysis (I² = 54%, Chi² P = 0.14). We considered the quality of the evidence to be very low for both outcomes and were therefore very uncertain about the result of more positive expressions in the music‐based interventions group at the end of treatment (SMD 0.54, 95% CI 0.06 to 1.02; 3 studies; I² = 0%, Chi² P = 0.70; Analysis 1.6; Figure 11; summary of findings Table for the main comparison). There was a similar SMD but an even wider CI in the analysis of long‐term effects (SMD 0.53, 95% CI –0.53 to 1.60; Analysis 2.6; summary of findings Table 2).

Figure 11

---

---

Forest plot of comparison: 1 Music‐based therapeutic interventions versus usual care or versus other activities: end of treatment, outcome: 1.6 Social behaviour: music vs other activities. SD: standard deviation.

Cognition

Seven studies contributed 350 participants to the end‐of‐treatment effect analysis and two studies with 193 participants assessed long‐term effects. Outcome measures used in the analyses were (translated versions of) the MMSE and the Severe Impairment Battery (SIB). We used the MMSE data if these were available in addition to other cognition measures such as Prose Memory tests, the FAS‐Test (Controlled‐Oral‐Word‐Association Test) or the Alzheimer's Disease Assessment Scale Cognitive subscale (ADAS‐cog). The end‐of‐treatment results were imprecise but not inconsistent. There was no important heterogeneity (I² = 0%; Chi² P = 0.89). There was serious risk of bias. The overall quality of the evidence was low for both time points and suggested that music‐based interventions may have had little or no effect on cognition at the end of treatment (SMD 0.15, 95% CI –0.06 to 0.36; Analysis 1.7; Figure 12; summary of findings Table for the main comparison) or at the long term (SMD 0.07, 95% CI –0.21 to 0.36; I² = 0%; Chi² P = 0.90; Analysis 2.7; summary of findings Table 2).

Figure 12

---

---

Forest plot of comparison: 1 Music‐based therapeutic interventions versus usual care or versus other activities: end of treatment, outcome: 1.7 Cognition. MMSE: Mini‐Mental State Examination; SD: standard deviation; SIB: Severe Impairment Battery.

Adverse effects

None of the trials reported adverse effects.

Effects of interventions delivered by a music therapist and sensitivity analyses

The sensitivity analyses with analyses restricted to studies where the intervention was definitely or possibly delivered by a qualified music therapist resulted in similar end‐of‐treatment effect estimates (there was no sensitivity analysis for the social behaviour outcome because no study remained). When restricting to studies that were definitely delivered by a music therapist, most effects were similar, but there was a smaller effect on anxiety. In the six of 13 studies in which the intervention was definitely delivered by a music therapist, the estimate for anxiety was –0.19 (SMD –0.19, 95% CI –0.52 to 0.13; with less heterogeneity; I² = 29%, Chi² P = 0.21; 242 participants).

When we restricted analyses further to studies definitely delivered by a music therapist, and having no potential financial conflict of interest or no funding source reported, we removed no studies from the anxiety analysis, and removed one or two studies for the remaining five outcomes. We found somewhat larger SMDs for the end of treatment outcomes. However, when we restricted analyses to studies at low risk of detection bias, the SMDs of six of the seven outcomes were smaller; all except for the SMD of behavioural problems overall, which was slightly larger. SMDs for individual therapy were similar to those for the main analyses (combined individual and group therapy) except for behavioural problems (both agitation or aggression and overall), for which SMDs for individual therapy were clearly larger.