Attention and vigilance advantages related to formal musical training across childhood, adolescence and young adulthood

Abstract

Attention and vigilance are fundamental cognitive abilities that develop throughout childhood and adolescence and have been associated with cognitively demanding activities such as formal musical training. This cross‐sectional study examined whether individuals engaged in long‐term musical training show superior attention and vigilance compared with matched controls without such training. Participants (N = 268; ages 8–34 years) were drawn from two samples of children, adolescents and adults, with musicians and nonmusicians matched on a wide array of demographic and lifestyle variables using multivariate propensity‐score matching. Attentional performance was assessed using the ANTI‐Vea, a validated computerized task that measures two components of vigilance – executive and arousal vigilance. Moreover, we applied advanced behavioural modelling given their sensitivity to the complex developmental trajectories in vigilance. For many indices, musically trained participants outperformed nonmusicians at all ages, with overall faster responses, fewer attentional lapses, detecting more vigilance targets and lower variability in arousal vigilance trials. Other group differences became more pronounced with age, indicating a possible dosage effect. Crucially, these findings remained after extensive control for confounders. Although the advantages associated with musical training were modest and the correlational study design, they are consistent with the hypothesis that music practice may foster domain‐general cognitive skills.

BACKGROUND

The capacity to maintain attentional focus over prolonged periods while engaging in a task (i.e., sustained attention) and the ability to detect and react quickly to infrequent yet crucial events during that time (i.e., vigilance) are both fundamental to everyday functioning. Based on the characteristics of the tasks used to assess vigilance, a recent framework proposed by Luna et al. (2018) distinguishes between two components of vigilance: arousal vigilance (AV) and executive vigilance (EV). AV is typically measured using tasks like the Psychomotor Vigilance Test (PVT; Dinges & Powell, 1985), which emphasize speeded automatic responses to infrequent, salient stimuli (e.g., pressing any button when a salient target appears). In contrast, EV involves more cognitively demanding tasks that require identifying rare targets among distractors and/or selecting between different responses. These tasks engage higher‐order cognitive processes, including working memory, response inhibition and decision‐making. Examples include the Sustained Attention to Response Task (SART; Robertson et al., 1997) and the Continuous Performance Test (CPT; Conners & MHS Staff, 2000). Sustained attention and vigilance (Hemmerich et al., 2025) have been shown to develop throughout childhood and adolescence (Fortenbaugh et al., 2015; Román‐Caballero, Trujillo, Cásedas, et al., 2025). Interestingly, with the use of advanced behavioural modelling of the decrement in executive vigilance (i.e., nonlinear logistic growth of hits), a two‐phase developmental pattern has been identified (Román‐Caballero, Trujillo, Cásedas, et al., 2025). While vigilance decrement in childhood and early adolescence seems to result from limitations in cognitive capacity (i.e., decline in target discrimination over time‐on‐task, TOT), during later developmental stages, the decline may increasingly reflect a reallocation of resources towards mind wandering, resulting in more frequent attentional lapses (i.e., failures to respond to the target when it is perceptually distinctive) over TOT.

While age plays a key role in the development of attention and vigilance, it is not the only factor that may influence these abilities. A central and complex question in cognitive science concerns whether cognitive performance can be enhanced through training and engagement in intellectually demanding activities. For instance, there is extensive evidence that computerized working memory and brain‐training programs improve performance on the specific task they target but offer limited transfer to distantly related tasks or everyday cognitive functioning (Melby‐Lervåg et al., 2016; Simons et al., 2016). Therefore, transfer is more likely when trained and untrained activities share contextual and content features (Barnett & Ceci, 2002) or when the cognitive routines acquired through the training apply to novel tasks (Gathercole et al., 2019).

In this regard, musical training, and in particular, learning to play a musical instrument, has been linked to brain plasticity (de Manzano & Ullén, 2018; Jünemann et al., 2022) and cognitive benefits in trained skills, such as auditory perception and fine motor dexterity (Herholz & Zatorre, 2012; Neves et al., 2022), as well as domain‐general abilities, including memory, attention and intelligence (Gustavson et al., 2021; Román‐Caballero et al., 2022). A priori, the singular characteristics of playing a musical instrument make it a promising candidate for transfer (Román‐Caballero et al., 2022). Unlike other cognitive activities that are confined to highly specific contexts and tasks, musical training engages multiple cognitive functions and sensory modalities. Moreover, it unfolds in environments that are rich and dynamic, involving a constantly renewed set of stimuli, shifting performance demands and a wide range of social contexts, from solitary practice to ensemble collaboration. Musical practice often coexists with school, work and home obligations that require organization and time management. This combination of cognitive breadth, variability and sustained challenge and planning is thought to increase the likelihood of far transfer.

However, it is difficult to prove the causality of the differences observed between long‐term musicians and nonmusicians (Schellenberg, 2020). The first challenge is the well‐established observation that individuals that choose to study music formally often differ from nonmusicians in their background and experiences (Schellenberg & Lima, 2024). For instance, factors such as parental education, openness to experience and intelligence determine the duration of musical training (Corrigall et al., 2013). Even overlooked factors, such as the month of birth and its consequences on the perception of parents over their child's cognitive abilities, may influence the decision to engage in this activity (i.e., relative age effect; Román‐Caballero, Trujillo, Martín‐Sánchez, et al., 2025). Over the past decades, a growing body of randomized controlled trials with children and older adults has demonstrated positive effects of learning to play a musical instrument on domain‐general cognitive abilities (Bugos & Wang, 2022; Frischen et al., 2021; Hallberg et al., 2017; James et al., 2020; Schellenberg, 2004; but see D'Souza & Wiseheart, 2018), as well as on brain structure (Jünemann et al., 2022; Marie et al., 2023; Worschech et al., 2022) and activity (Jünemann et al., 2023). Although the type of training in these studies resembles the early stages of musical instruction typically undertaken by musicians (including systematized learning of the technique of an instrument, and musical theory and language.), the relatively short duration of practice in these programs (from 5 weeks to 4 years) limits the extent to which their findings can be generalized to the long‐term differences observed between musicians with decades of experience and nonmusicians. Moreover, although randomization helps control for many potential confounds, it is difficult to achieve good levels of adherence over long periods, particularly when participants are randomly assigned to activities they may not find intrinsically motivating (Tervaniemi et al., 2022). Even if such participants remain in the program, their lack of engagement may compromise its effectiveness and possibly raise ethical concerns (Habibi et al., 2018).

Consequently, alternative designs are needed to investigate more effectively the causality of the effect of a long‐term musical practice such as that typically accumulated by musicians over several decades. When key confounding variables contributing to selection bias (e.g., socioeconomic status or personality) are carefully controlled, longitudinal designs with twins (Gustavson et al., 2021) and population‐based studies (Guhn et al., 2020), and well‐designed cross‐sectional approaches (Nichols et al., 2020; Román‐Caballero et al., 2021) can provide a closer estimate of the long‐term impact of music practice. While the effect sizes in uncontrolled cross‐sectional studies have typically reported moderate‐to‐large effects (Schellenberg, 2020; Talamini et al., 2017), well‐controlled cross‐sectional investigations have revealed much smaller differences, closer in magnitude to those observed in randomized controlled trials (d ≈ 0.2; Román‐Caballero et al., 2022). Thus, although experimental designs are feasible for examining the long‐term effects of musical training, the extensive cross‐sectional literature seems to have largely failed to disentangle genuine musicianship‐related differences from potential confounding factors. The present cross‐sectional study addressed previous limitations with extensive control of confounds, considering a large number of inclusion and matching variables. This study aimed to compare attentional and vigilance performance between musically trained and untrained participants (8–34 years old), after equating the groups on background variables through multivariate propensity‐score matching.

This study focused on differences in attention and vigilance as the core cognitive functions of interest. Playing an instrument implies multiple attentional demands, requiring individuals to process multiple stimuli simultaneously (e.g., reading a score, monitoring the movements of other performers and integrating overlapping melodic lines) while also sustaining focus and responding appropriately over extended periods (Rodrigues et al., 2013). In addition, improvements in core cognitive functions that are involved in many daily activities and tasks, such attention and executive functions, which are engaged across a wide range of everyday activities, might indirectly enhance performance in nonmusical domains (acting as a mediator factor between musical training and nonmusical skills; Hannon & Trainor, 2007; Moreno & Farzan, 2015).

Previous randomized controlled trials (Hallberg et al., 2017) and well‐controlled cross‐sectional studies (Román‐Caballero et al., 2021) have observed advantages in vigilance in adults with musical training. The present cross‐sectional study examines those differences with a developmental approach, investigating their trajectory over the years. Moreover, we applied advanced behavioural modelling (i.e., nonlinear logistic growth and ex‐Gaussian distribution with generalized additive models), which has shown greater sensitivity to the complex developmental trajectories of these skills than other traditional analytical approaches (such as the signal detection theory; Román‐Caballero, Trujillo, Cásedas, et al., 2025). A priori, one possibility is that differences emerge early, with a threshold effect, or differences may become more pronounced over the years, with a dosage effect following the accumulation of years of musical practice. Characterizing how these cognitive differences evolve is not only descriptively valuable but also theoretically relevant, as any potential cognitive benefits are likely to depend on the level of engagement (Talamini et al., 2025). Yet, the nature of the relationship between musical training and attentional differences (threshold vs. dosage effect) remains unclear. Previous research has addressed this question indirectly, by correlating cognitive measures with variables of engagement in adult participants (e.g., training duration or intensity). In the present study, we directly examine that relationship by using a developmental design.

METHOD

Participants

The study included two independent samples. Sample 1 comprised 301 Spanish children and adolescents (8–18 years old) with and without formal musical training. This sample comes from a larger dataset that also included 148 additional children and adolescents engaged in structured sport training (collected for other research purposes), totalling 449 participants (for details, see Román‐Caballero, Trujillo, Cásedas, et al., 2025; Román‐Caballero, Trujillo, Martín‐Sánchez, et al., 2025). To extend the age range into early adulthood, we included a second sample of 119 adult musicians and nonmusicians (Sample 2; 18–34 years old) previously reported in Román‐Caballero et al. (2021).

Participants were approached through the distribution lists of Spanish conservatories, high schools, sports centres (Sample 1) and universities (Sample 2). Data collection for Sample 1 took place between March 2022 and July 2023, whereas for Sample 2, it took place between December 2018 and March 2019. The study adhered to ethical standards set by the University of Granada (protocol numbers: 536/CEIH/2018 and 2614/CEIH/2022), in line with the Declaration of Helsinki (2013 revision, Brazil). The families of the sample of Sample 1 received €10 for their participation, while participants in Sample 2 were included in a raffle for a monetary reward.

Inclusion criteria

The participants of both samples met the general inclusion criteria, related to health status and behavioural habits that may have an impact on cognition (e.g., drug abuse), and the specific inclusion criteria of this study to be categorized as musicians and nonmusicians. To determine eligibility, participants completed a questionnaire prior to the attention task, completed by caregivers or parents in the case of Sample 1. All participants reported normal or corrected‐to‐normal vision and hearing, and none had a history of neurological, psychiatric, or chronic health conditions (including attention‐deficit/hyperactivity disorder, ADHD). In addition, children and adolescents did not have an extremely or very preterm birth (< 32 weeks), severe problems during pregnancy or childbirth, and chronic drug use or abuse by the mother during pregnancy or by the child/adolescent. Adults in Sample 2 were excluded in cases of chronic use of psychoactive medication (longer than 6 months), drug abuse, or alcoholism. Participants that did not complete the task were also excluded. This included children and adolescents with excessive non‐responses (> 20% of trials; 19 participants) and less than four full blocks completed for adults (2 participants; Figure 1a).

FIGURE 1.

(a) Flow diagram of the selection of participants and (b) the propensity‐score matching results.

For Sample 1, the participants in the musical training group participated in extracurricular musical training programs at the conservatory, while control children and adolescents did not have any extracurricular musical training. For Sample 2, participants in the musical training group were musicians that played an instrument and/or sang for 10 years or longer, had at least 5 years of formal musical training and had an age of musical training onset prior to 14 years old. Nonmusicians were adults that could not play any instrument or read scores (and therefore, without any formal musical training) and had no experience as singers.

Finally, participants with extremely low performance were excluded from the sample. Specifically, 23 participants from Sample 1 were excluded for having overall accuracies more than 2.5 standard deviations (SDs) below their corresponding cohort mean.

Propensity‐score matching

In addition to the inclusion criteria, we also controlled for confounds in both samples by matching groups on multiple variables with a potential influence on cognition (Reuter‐Lorenz & Park, 2014). Following the procedure in Román‐Caballero et al. (2021), we used propensity‐score matching to form pairs of participants (Adelson, 2013). Propensity‐score matching uses a logistic regression to predict group membership probability (i.e., propensity score; with a value ranging from 0, without formal musical training, to 1, with formal musical training) based on a set of background characteristics – potential confounds – and allows matching participants from one group to the other in pairs. Unlike the traditional univariate matching procedure, this approach also accounts for the multivariate interactions among confounding variables.

We used propensity‐score matching in Sample 1 to generate similar groups of children and adolescents considering multiple relevant background characteristics such as age (in years), sex (categorical), nationality (Spanish/other), bilingualism (yes/no), socioeconomic status (sum of ordinal scores with the highest level of education attained for each parent/caregiver, the salary of each parent/caregiver and the average income of the area of residence within the city according to the 2023 report of the Spanish National Institute of Statistics), physical exercise (Likert scale, from ‘never’ to ‘more than 9 hours weekly’ in relation to several activities: swimming, running, soccer, tennis, gym, aerobic exercise, dancing, cycling, walking, gymnastics and others), involvement in cognitively stimulating activities (Likert scale, from ‘never’ to ‘more than 9 h weekly’ in relation to several activities: chess, puzzles, reading, plastic arts and drawing, other nonmusical arts and others), video game playing (Likert scale, from ‘never’ to ‘more than 9 h weekly’) and the big five personality traits (neuroticism, extraversion, openness, agreeableness and conscientiousness; McCrae & John, 1992). In addition, we included the task version (in‐person vs. online via a supervised virtual session) in the model to balance the proportion of children/adolescents conducting the task online in each group.

For Sample 2, we used the same set of confounds as in Román‐Caballero et al. (2021): age, sex, level of education (ordinal scale with the educational degrees of the Spanish system, from primary education to PhD), lifelong tobacco consumption (Likert scale, from ‘none’ to ‘more than 20 cigarettes per day, in general’), physical exercise, bilingualism, second language use (Likert scale, from ‘none’ to ‘more than 9 hours per day’), involvement in cognitively stimulating activities and video game playing. For more details about how these variables were measured, see the Supplementary Material S1. We used the MatchIt R package (Ho et al., 2011) to conduct the analysis. Matching was performed with a 1:1 matching ratio and using the optimal matching method (maximally minimizing the sum of absolute distances between pairs). Cases with non‐overlapping, extreme propensity scores were excluded from the matched sample, and propensity scores were re‐estimated after discarding them. This approach allowed us to retain the best possible matched pairs while improving balance on the covariates used in the propensity‐score model. After matching, 85 pairs of children and adolescents, and 49 pairs of adults were formed (Figure 1b), leading to a total sample of 268 participants. In our previous work, we reported that both groups of adults were successfully matched in the set of potential confounding variables (Román‐Caballero et al., 2021). A comparable pattern was observed in the new sample of children and adolescents (Table 1). In the adult cohort (68.37% of women, 88.78% Spanish), the mean age was 22.9 years (SD = 3.9), with musically trained participants reporting an average of 14.8 years (SD = 3.8) of musical practice, having typically begun training at 7.4 years of age (SD = 1.7). In contrast, the final sample of children and adolescents had a mean age of 12.6 years (SD = 2.8), an average of 6.1 years (SD = 3.2) of musical training, and begun at 6.5 years (SD = 1.9) on average.

TABLE 1.

Potential confounding variables in both groups of Sample 1 after propensity‐score matching.

Variable	Musical training	Control	Test
N	85	85
Age (years)	12.59 (2.83)	12.53 (2.73)	t(168) = 0.14, p = .891
Sex (% woman)	47.06	48.24	χ2(1) ≈ 0, p ≈ 1
Nationality (% Spanish)	96.47	96.47	χ2(1) ≈ 0, p ≈ 1
Bilingualism (%)	12.94	11.76	χ2(1) ≈ 0, p ≈ 1
SES score	21.88 (3.85)	21.78 (3.86)	t(168) = 0.18, p = .858
Physical activity score	9.19 (4.77)	9.41 (5.16)	t(168) = −0.29, p = .770
Video game	1.96 (1.58)	2.15 (1.71)	t(168) = −0.74, p = .457
Cognitive activity score	5.66 (2.91)	5.40 (3.26)	t(168) = 0.55, p = .586
Conscientiousness	46.84 (8.77)	45.61 (8.40)	t(168) = 0.93, p = .354
Neuroticism	33.32 (8.09)	34.71 (8.35)	t(168) = −1.10, p = .273
Extraversion	49.53 (7.44)	48.93 (7.13)	t(168) = 0.54, p = .592
Openness	51.09 (6.37)	49.47 (6.80)	t(168) = 1.61, p = .110
Agreeableness	53.19 (6.03)	53.32 (6.12)	t(168) = −0.14, p = .890
Task version (% in‐person)	15.29	17.65	χ2(1) = 0.04, p = .836

Materials

Information regarding inclusion criteria, sociodemographic characteristics and relevant background variables was collected through an in‐house questionnaire, containing Likert‐type, open‐ended and binary (yes/no) questions. For Sample 1, the questionnaire was completed by a caregiver or parent. For Sample 2, adult participants provided the information themselves. In addition to this questionnaire, caregivers/parents in Sample 1 also completed two standardized instruments: the ADHD Rating Scale–IV (DuPaul et al., 1998; Servera & Cardo, 2007) to assess symptoms of ADHD, and the Big Five Questionnaire for Children (Barbaranelli et al., 2003; Carrasco Ortiz et al., 2005) to evaluate personality traits. The analysis of these self‐reported measures is beyond the scope of the present study, except for the variables used in participants' inclusion and matching.

Attentional performance was assessed using the online version of the ANTI‐Vea (https://anti‐vea.ugr.es; Luna et al., 2018), implemented with Javascript ES5, HTML5, CSS3 and Angular JS. The ANTI‐Vea is an open computerized task designed to measure several components of attention—phasic alertness, spatial orienting, executive control and vigilance (Posner & Petersen, 1990)—based on reaction time (RT) and accuracy. A detailed description of its characteristics and implementation is provided in Coll‐Martín et al. (2023).

Procedure

After giving consent to participate in the study, the parents/caregivers of the child/adolescent in Sample 1 or the participants themselves in Sample 2 completed an online survey via LimeSurvey. After the survey, participants proceeded to the online version of the ANTI‐Vea task. For Sample 1, the procedure was adapted to enhance feasibility with younger participants. Depending on the location of the families, the session with children and adolescents was conducted either in person at their school, conservatory, sports centre (28%), or online via a supervised virtual session (72%). In both formats, data collection took place in small groups under the supervision of one or more researchers, either physically present or connected via video call. Children aged 12 years and under completed the tasks with four blocks, while participants over 12 years completed the standard six‐block version. For online sessions, a video call was initiated on the caregiver's smartphone to provide technical support and supervision, while the task itself was run on the laptop or desk computer of the family. Each session lasted approximately 45–60 min, depending on the number of blocks completed.

As illustrated in Figure 2, the ANTI‐Vea task presents three intermixed trial types: ANTI, EV and AV trials. ANTI and EV trials share the same structure: a fixation cross, a black asterisk, a 2000‐Hz warning tone and an array of five arrows (one central target arrow flanked by four distractors). In EV trials (20% of all trials), the central arrow is displaced vertically by 8 pixels (up or down). In addition to this displacement, a random variability of ±2 px is applied to the horizontal and vertical positions of each flanker arrow to add noise, thus parametrically varying the signal–noise distinction (see Figure 2) and increasing task difficulty. In ANTI trials, a random variation of ±2 px is also added to the central arrow. In AV trials (also 20% of total trials), a red countdown timer appears at the fixation point, replacing the standard stimulus display.

FIGURE 2.

Schematic representation of the ANTI‐Vea task. (a) Sequence of stimuli in ANTI and EV trials, where the target display was a horizontal array of five arrows. In these trials, participants were instructed either (b) to indicate the direction of the central arrow (ANTI trials) or (c) to detect when it was largely vertically displaced (EV trials). (d) In ANTI trials, all arrows are randomly jittered by ±2 px to introduce noise, while in EV trials, the central arrow is displaced vertically by 8 pixels, producing a graded signal–noise similarity continuum based on the total pixel separation between the target and surrounding arrows (ranging from 0 to 4 px – noise– and from 5 to 10 – signal). (e) Sequence of AV trials, in which participants (f) were required to stop a red countdown timer as quickly as possible by pressing any key.

The ANTI‐Vea task started with a practice phase, followed by the experimental blocks. The majority of trials were either ANTI (60%) or executive vigilance (EV; 20%) types. In these trials, half included an auditory warning tone prior to the target display. Regardless of tone presence, the auditory event was followed by a visual spatial cue, an asterisk, appearing either at the upcoming target location (valid cue), at the opposite location (invalid cue), or not appearing at all (no cue condition), each comprising one‐third of trials. For both ANTI and EV trials, the visual array consisted of a central target arrow flanked by four distractors. In ANTI trials, participants were asked to indicate the direction of the central arrow by pressing ‘c’ (left) or ‘m’ (right), while ignoring the flanking arrows. In EV trials, the task instead involved detecting a large vertical displacement (upward or downward) of the central arrow by pressing the space bar, regardless of its direction. AV trials, which made up the remaining 20% of trials, did not include a warning signal or spatial cue. Instead, after a variable timed interval (900–2100 ms), a red millisecond countdown timer appeared at fixation, and participants were instructed to press any key as quickly as possible to stop it. Participants were instructed to keep their eyes on the fixation point and to respond as quickly and accurately as possible during the task.

During the practice phase, task instructions were introduced progressively to facilitate understanding of each type of trial. Participants completed four practice blocks that increased in both length and complexity. The sequence began with 16 ANTI trials, followed by a mixed block containing 16 ANTI and 16 EV trials. This was followed by a 48‐trial block combining 24 ANTI, 8 EV and 8 AV trials. Visual feedback was displayed in these first three practice blocks. The final practice block consisted of 40 trials (24 ANTI, 8 EV and 8 AV) and was presented without feedback to simulate the conditions of the main task. Prior to the experimental trials, participants could repeat the last practice block if needed. The experimental phase comprised four or six, depending on the age of the participant, blocks of 80 trials each: 48 ANTI, 16 EV and 16 AV. No feedback or rest breaks were provided during this phase, and the order of trials within each block was fully randomized.

Previous comparisons between the in‐person and online versions of the task (Luna et al., 2021) have shown that responses tend to be slower in online settings, and that the alerting and orienting effects may be slightly influenced by the version (presumably due to variability in headphone and screen displays). However, the version of the task did not affect the observed vigilance decrement, and both formats demonstrated comparable split‐half reliability. To account for minor overall differences, we balanced the proportion of children and adolescents completing the online version through our propensity‐score matching procedure.

Preprocessing

Due to the importance of TOT decrement in our analyses, trials from incomplete blocks were excluded from the dataset, representing 0.87% of all trials. Due to the parameters of the initial programming of the online task, unintentional right‐clicks allowed the right‐click menu to open and prevented the collection of responses to the task until it was closed. Generally, participants removed the menu by left‐clicking, but the response during the corresponding trials was lost. Thus, sequences of four or more consecutive trials with missing data were identified as corresponding to this failure to collect responses. In total, 0.25% of the trials were removed for this reason.

In line with standard analyses of RT, we excluded trials with incorrect responses as well as those with RT below or above 2 SDs of the mean of the participant in ANTI trials. Because we anticipated a nonlinear decrease in RTs with age, absolute differences between two conditions, such as in the indices of the three attentional networks (e.g., RT in incongruent minus congruent trials), might not be comparable across ages. Thus, we computed the proportional RT over the mean RT of the age cohort (i.e., RT in each trial divided by the grand mean of participants within the same age × 100) for the indices of the three attentional networks. This transformation is common in studies addressing age‐related differences (Bialystok et al., 2008; Colcombe et al., 2005; De Bruin & Sala, 2018), where groups differ in their overall mean RT.

Statistical analysis

Analyses were conducted separately for each type of trial. For ANTI trials, we analysed: (1) the overall RT, (2) the overall errors; and the indices for (3) phasic alertness (no tone minus tone, in no cue trials), (4) orienting (invalid minus valid trials) and (5) executive control (incongruent minus congruent trials) using RTs. In all these measures (except errors), we fitted linear mixed‐effects models with trial‐to‐trial RT using the lme4 R package (Bates, Mächler, et al., 2015). We included a logarithmic component of age, group and its interaction with age as fixed effects, and a random intercept of participant: RT ~ log(age) + group + group × log(age) + 1|participant. Note that, in some of the models, other more complex random‐effects structures (i.e., including random slopes) led to convergence issues. However, the results of the models that successfully converged were similar. For the sake of consistency across analyses, we report the results from the models with random intercepts only. For the three attentional network indices, tone/validity/congruency was added as a main factor (overall effect of alertness, orienting and executive control) and interacting with the components of age and group, as well as the three‐way interactions: prop. RT ~ manipulation + log(age) + group + manipulation × log(age) + manipulation × group + log(age) × group + manipulation × log(age) × group + 1|participant, where manipulation represents the attentional manipulation of the index. In the case of overall errors, we conducted a binomial generalized linear mixed‐effects model with trial‐to‐trial accuracy (0 = error, 1 = correct). To facilitate convergence, we used a bound optimization by quadratic approximation optimizer with all generalized mixed‐effects models in the study. In addition, we performed model selection using the buildmer R package (Voeten, 2025), which automates stepwise selection for mixed‐effects models. For each model, we used backward selection based on the Akaike Information Criterion.

For EV trials, we followed a double‐analytic strategy. First, for comparative purposes with previous studies in the literature, we conducted separate binomial generalized linear mixed‐effects models with trial‐to‐trial hits (0 = omission error, 1 = correct detection of the vertical displacement of the target) and false alarms (0 = true rejection, 1 = commission error or non‐displaced target identified as being vertically displaced). Then, we modelled participants' ‘signal’ responses (i.e., space bar presses) as a function of the target‐to‐flanker distance using a nonlinear mixed‐effects model with a logistic growth curve. The model was fitted using the nlme package in R (Pinheiro et al., 2023) and followed the formula: $a+\frac{b-a}{1+{\mathrm{e}}^{\frac{c-\text{distance}}{d}}}$, where a represents the lower asymptote, b represents the upper asymptote, c represents the inflection point of the curve and d represents the scale of the curve (with smaller values indicating higher sensitivity; 1/d corresponds to the slope of the curve). Fixed effects for age (logarithmic component), across‐block decrement, group and their interactions were included for each component, except a, for which it was assumed no decrement and only changes with age (McCarley & Yamani, 2021). To avoid convergence issues, a random intercept by participant was included for the inflection point component (Bates, Kliegl, et al., 2015). In the same vein, block was treated as a binary factor, with the first and the second half of the task (McCarley & Yamani, 2021). For participants who completed the 6‐block version (older than 12 years), blocks 1, 2 and 3 were coded as the first half. For those in the 4‐block version, blocks 1 and 2 were assigned to the first half.

Finally, we modelled an ex‐Gaussian distribution with the RTs in AV trials using generalized additive models for location, scale and shape with the gamlss R package (Rigby & Stasinopoulos, 2005). The ex‐Gaussian model decomposed RT into two components: a normal distribution characterized by mean (μ) and standard deviation (σ), and an exponential component (τ), which captures the positive skew commonly observed in RT data. Consistent with the EV trial analyses, we included age (logarithmic component), block, group and their interactions as fixed effects for each of the three distributional parameters (μ, σ and τ). In addition, we modelled random intercepts by participant for all three parameters. As in the logistic curve analysis, block was coded as a binary factor (first vs. second half of the task).

Model selection for both the logistic and ex‐Gaussian analyses followed a backward stepwise approach, using an exclusion criterion of α = .05. Each analysis began with a comprehensive model with all relevant main effects and interactions.

RESULTS

Attentional networks

Along with the anticipated nonlinear decrease in RT of ANTI trials with development, β_log(age) = −120.95, t(265.38) = −13.43, p < .001, participants with formal musical training were faster across ages: β_group = −35.54, t(264.62) = −2.50, p = .013 (~36 ms faster; Figure 3a; Table 2). Moreover, accuracy in ANTI trials nonlinearly increased with a similar trend in both groups, log β_log(age) = 0.58, OR _log(age) = 1.78, z = 10.09, p < .001 (Figure 3b). On the other hand, musical training did not significantly modulate the effects of spatial orienting (Figure 3c,d), but it was associated with superior phasic alertness, β_tone×group = −1.10, t(19650) = −2.07, p = .039 (i.e., smaller tone vs. no‐tone difference; Figure 3c), and a progressive enhancement in executive control across years, β_{log(age)×congruency×group} = −1.11, t(59210.30) = −2.52, p = .012 (i.e., smaller interference; Figure 3e).

FIGURE 3.

Development and between‐groups differences of the main measures of ANTI trials, including (a) overall RT, (b) overall accuracy and the effects of (c) phasic alertness, (d) orienting and (e) executive control. For visualization, means and standard errors of the mean were displayed by age cohort in the panels. Panels c–e show the attentional effects in proportional RTs.

TABLE 2.

Results of the best fit models with measures of ANTI trials.

Overall RT
RT ~ log(age) + group +1|participant	βlog(age) = −120.95, t(265.38) = −13.43, p < .001 βgroup = −35.54, t(264.62) = −2.50, p = .013
Overall accuracy
Accuracy ~ log(age) + 1|participant	log βlog(age) = 0.58, OR log(age) = 1.78, z = 10.09, p < .001
Effect of phasic alertness
prop. RT ~ tone + group + tone × group +1|participant	βtone = 5.55, t(19650) = 20.96, p < .001 βgroup = −0.09, t(265.30) = −0.05, p = .960 βtone×group = −1.10, t(19650) = −2.07, p = .039
Effect of spatial orienting
prop. RT ~ validity +1|participant	βvalidity = 5.85, t(39280) = 30.29, p < .001
Effect of executive control
prop. RT ~ log(age) + congruency + group + log(age) × congruency + log(age) × group + congruency × group + log(age) × congruency × group + 1|participant	βlog(age) = −0.92, t(264.74) = −0.75, p = .454 βcongruency = 10.34, t(59212.38) = 21.02, p < .001 βgroup = 0.99, t(265.16) = 0.19, p = .850 βlog(age)×congruency = −1.33, t(59210.30) = −6.02, p < .001 βlog(age)×group = −0.53, t(264.74) = −0.22, p = .830 βcongruency×group = 1.97, t(59212.39) = 2.00, p = .046 βlog(age)×congruency×group = −1.11, t(59210.30) = −2.52, p = .012

The effects corresponding to the three attentional networks remained relatively stable except for the congruency effect (i.e., executive control), which decreased with age, β_{log(age)×congruency} = −1.33, t(59210.30) = −6.02, p < .001, and indicated a development of executive control from the age of 8 years up to adulthood. Furthermore, although we observed a larger congruency effect (i.e., more interference) in early childhood for the musical training group, β_{congruency×group} = 1.97, t(59212.39) = 2.00, p = .046, the developmental decrease in the congruency effect (i.e., enhancement in executive control across years) was more pronounced in the musical training group, as indicated by the three‐way interaction. Thus, this developmental course led to the reversion of group difference: smaller congruency effect/less interference for musically trained participants in late adolescence and early adulthood (Figure 3e).

Executive vigilance

As expected, the percentage of correct detections of the ‘signal’ event (i.e., central arrow vertically displaced) increased with development, log β_log(age) = 0.25, OR _log(age) = 1.29, z = 3.58, p < .001 and musical training did not significantly predict overall larger proportion of hits, log β_group = 0.18, OR _group = 1.20, z = 1.65, p = .100 (Figure 4a; Table 3). Moreover, a vigilance decrement in hits was observed across TOT, log β_block = −0.13, OR _block = 0.88, z = −12.38, p < .001 (Figure 4b), without a modulation by group.

FIGURE 4.

Development and between‐groups differences of (a) hits and (c) false alarms. In addition, the effect of TOT in these measures (panels b and d, respectively). For visualization, means and standard errors of the mean were displayed by age cohort in the panels.

TABLE 3.

Results of the best fit models with hits and false alarms.

Hits
hits ~ log(age) + block + group +1|participant	log βlog(age) = 0.25, OR log(age) = 1.29, z = 3.58, p < .001 log βblock = −0.13, OR block = 0.88, z = −12.38, p < .001 log βgroup = 0.18, OR group = 1.20, z = 1.65, p = .100
False alarms
FAs ~ log(age) + block + group + block × log(age) + block × group + 1|participant	log βlog(age) = −0.33, OR log(age) = 0.72, z = −4.81, p < .001 log βblock = 0.04, OR block = 1.04, z = 0.96, p = .335 log βgroup = 0.07, OR group = 1.07, z = 0.64, p = .520 log βlog(age)×block = −0.04, OR log(age)×block = 0.96, z = −2.06, p = .040 log βblock×group = −0.05, OR block×group = 0.95, z = −1.92, p = .055

Results with false alarms were similar. The percentage of false alarms decreased across years, log β_log(age) = −0.33, OR _log(age) = 0.72, z = −4.81, p < .001 (Figure 4c) and a vigilance decrement, with a decrease over TOT, arose with development, log β_{log(age)×block} = −0.04, OR _{log(age)×block} = 0.96, z = −2.06, p = .040 (Figure 4d; Table 3). The overall effect of musical training, log β_group = 0.07, OR _group = 1.07, z = 0.64, p = .520, and its interactions with TOT, log β_block×group = −0.05, OR _block×group = 0.95, z = −1.92, p = .055, did not reach the significance level.

When EV was modelled with a nonlinear logistic model, the usual reduction in guesses and lapses was observed with development, indicating that participants committed fewer errors across years in rejecting the target event when the absence of signal is evident (i.e., guesses; β_log(age) = −0.02, t(67097) = −9.74, p < .001) and detecting the target event when the presence of signal is evident (i.e., lapses; β_log(age) = −0.02, t(67097) = −3.46, p < .001; Table 4). In the present sample, vigilance decrements were observed with the scale (i.e., sensitivity), β_block = 0.23, t(67097) = 3.51, p < .001 and the position of the curve (i.e., criterion), β_block = 0.46, t(67097) = 12.63, p < .001 (Figure 5a) but not for the parameter corresponding to the upper asymptote (i.e., lapses), which is a common finding in studies with adults and late adolescence (McCarley & Yamani, 2021; Román‐Caballero et al., 2023; Román‐Caballero, Trujillo, Cásedas, et al., 2025). In line with previous developmental studies of vigilance (Román‐Caballero, Trujillo, Cásedas, et al., 2025), the decrement with sensitivity was attenuated across childhood and adolescence, β_{log(age)×block} = −0.07, t(67097) = −2.18, p = .029.

TABLE 4.

Results of the best logistic curve model for EV trials.

Scale (sensitivity)
scale ~ block + group + log(age) × block + log(age) × group + log(age) × block × group	βblock = 0.23, t(67097) = 3.51, p < .001 βgroup = 0.16, t(67097) = 2.23, p = .026 βlog(age)×block = −0.07, t(67097) = −2.18, p = .029 βlog(age)×group = −0.09, t(67097) = −2.64, p = .008 βlog(age)×block×group = −0.09, t(67097) = −3.69, p < .001
Position (criterion)
position ~ block +1|participant	βblock = 0.46, t(67097) = 12.63, p < .001
Lower asymptote (guesses)
lower asymptote ~ log(age)	βlog(age) = −0.02, t(67097) = −9.74, p < .001
Upper asymptote (lapses)
upper asymptote ~ log(age) + group	βlog(age) = −0.02, t(67097) = −3.46, p < .001 βgroup = −0.03, t(67097) = −4.97, p < .001

FIGURE 5.

(a) Change in the estimated logistic curves between the two halves of the task. In addition, the effect of musical training on (b) lapses overall and (c) the improvement of the TOT decrement in slope (sensitivity) across years.

Importantly, the model predicted overall fewer lapses in participants with musical training than without, β_group = −0.03, t(67097) = −4.97, p < .001 (Figure 5b). Finally, mirroring the pattern observed for the congruency effect in ANTI trials, the model predicted lower sensitivity for the musically trained group in early childhood, β_group = 0.16, t(67097) = 2.23, p = .026, followed by a progressive improvement in this group's overall sensitivity with age, β_{log(age)×group} = −0.09, t(67097) = −2.64, p = .008. In addition, musical training was associated with more pronounced attenuation of the age‐related decrement in sensitivity, β_{log(age)×block×group} = −0.09, t(67097) = −3.69, p < .001. Together, these effects resulted in the disappearance of the decrement in sensitivity among musically trained participants by late adolescence and early adulthood (Figure 5c).

Arousal vigilance

The ex‐Gaussian modelling of RT in AV trials showed that the three parameters were modulated by age in an exponential decay fashion (Table 5): mean (μ), β_log(age) = −67.58, t = −88.31, p < .001; standard deviation (σ), log β_log(age) = −0.29, t = −44.56, p < .001; and exponential component (τ), log β_log(age) = −0.22, t = −21.96, p < .001. Moreover, we observed a decrement in variability (σ; more variability in the 2nd than the 1st half) at all ages, log β_block = 0.03, t = 3.40, p < .001. Similarly, a decrement in lapses (τ; more lapses in the 2nd than the 1st half) that progressively emerged across years, log β_{log(age)×block} = 0.13, t = 21.62, p < .001, increasing the exponential component over TOT throughout childhood and adolescence (Figure 6a). The reverse pattern arose for response speed (μ), with responses becoming faster over TOT with age, β_{log(age)×block} = −3.35, t = −8.84, p < .001, suggesting a learning effect (i.e., improvement throughout the task).

TABLE 5.

Results of the best ex‐Gaussian model for AV trials.

μ (mean of the normal distribution of RT)
μ ~ log(age) + group + log(age) × block + log(age) × group +1|participant	βlog(age) = −67.58, t = −88.31, p < .001 βgroup = 42.80, t = 11.82, p < .001 βlog(age)×block = −3.35, t = −8.84, p < .001 βlog(age)×group = −22.77, t = −14.90, p < .001
σ (variability)
σ ~ log(age) + block + group + block × group +1|participant	log βlog(age) = −0.29, t = −44.56, p < .001 log βblock = 0.03, t = 3.40, p < .001 log βgroup = −0.23, t = −8.01, p < .001 log βlog(age)×group = 0.07, t = 5.47, p < .001
τ (lapses)
τ ~ log(age) + group + log(age) × block +1|participant	log βlog(age) = −0.22, t = −21.96, p < .001 log βgroup = −0.15, t = −10.75, p < .001 log βlog(age)×block = 0.13, t = 21.62, p < .001

FIGURE 6.

(a) Development of the estimated ex‐Gaussian distribution of RT between the two halves of the task. In addition, the effect of musical training on the development of (b) overall response speed, (c) variability and (d) lapses across years.

Of relevance to this study, musical training predicted overall lower variability (σ), log β_group = −0.23, t = −8.01, p < .001 and fewer lapses (τ), log β_group = −0.15, t = −10.75, p < .001 (Figure 6c,d). For mean response speed (μ), the model predicted a more pronounced improvement throughout childhood and adolescence for participants with musical training, β_{log(age)×group} = −22.77, t = −14.90, p < .001, at the same time as a small initial disadvantage for the musically trained participants in our sample (with slower responses in the first years of the investigated age span), β_group = 42.80, t = 11.82, p < .001 (Figure 6b). Due to the age‐by‐group interaction, this initial difference was attenuated over the years and even reversed (adult musicians with a faster response mean than nonmusicians) at the older end of the age span. The reverse pattern arose with variability (σ), with less pronounced differences between musically trained and untrained participants across years, β_{log(age)×group} = 0.07, t = 5.47, p < .001 (Figure 6c).

Finally, the overall effect of task version (online vs. in person) was negligible for most measures, only with a smaller congruency effect and slower mean RT in AV trials for the online version of the task (see Supplementary Material S2, Tables S1–S4).

DISCUSSION

Previous randomized controlled trials (Frischen et al., 2021; Hallberg et al., 2017; James et al., 2020) and well‐controlled cross‐sectional studies (Nichols et al., 2020; Román‐Caballero et al., 2021) have reported benefits of formal musical training and music‐making on attention and vigilance. The present study adds to this growing body of evidence by showing that individuals enrolled in formal musical training at conservatories and private institutions showed advantages in attention and vigilance compared with individuals without such training. Although causality cannot be inferred from the present study, a key contribution of this work is its developmental approach, which allowed for the characterization of the trajectories of these differences across years. Notably, some of the differences between groups were consistent across all ages for many measures. Regardless of the age cohort, musically trained individuals were overall faster on the attention task, benefited less from warning signals, presumably because of a higher sustained state of vigilance throughout the task (Román‐Caballero et al., 2021), and showed fewer attentional lapses in both vigilance trials (EV and AV). This suggests a greater capacity to sustain endogenously a state of high preparation for long periods that would make individuals with formal musical training less prone to disengagement from the activity and attentional lapses (Román‐Caballero et al., 2021). Interestingly, the lack of an EV decrement in lapses, which is common in studies with adults and late adolescence (McCarley & Yamani, 2021; Román‐Caballero et al., 2023; Román‐Caballero, Trujillo, Cásedas, et al., 2025), indicates the prominence of this type of decrement in later stages of development and the lower representation of adults and older adolescents in our sample. Previous developmental evidence (Román‐Caballero, Trujillo, Cásedas, et al., 2025) suggests that, while vigilance decrement in childhood and early adolescence seems to result from limitations in cognitive capacity, during later developmental stages, the decline may increasingly reflect a reallocation of resources towards mind wandering.

The finding that the group differences are already present in the youngest cohort of our sample (8–9 years old) might align with a threshold effect, in which cognitive benefits emerge once a certain level of training is reached. An alternative explanation is the presence of a confounding variable that both influenced the choice to pursue formal musical training and contributed to the observed differences. In fact, critical approaches with a far transfer account (Talamini et al., 2025) propose that the lack of a dosage‐effect pattern (i.e., no relation between the observed cognitive differences and the duration and intensity of musical training) is a signature of selection bias. Although the official starting age of conservatory programs in the Spanish system is at 8 years, not all children usually begin instruction at the same time (often because families engaged their children in pre‐conservatory training). This produced variability in musical experience within the same age group. To address this variability, we repeated the analyses including only participants who began formal training no earlier than age 6. The results remained largely unchanged (Supplementary Material S3, Table S5), further supporting the idea of a threshold effect (Walsh et al., 2021). Given the cross‐sectional design of our study, we cannot fully determine whether the differences between the two groups were caused by or preexisting to conservatory training. In addition, as our assessment took place after a period of training in all cases, our design cannot fully address this point. Longitudinal studies would be better suited to investigate this question.

On the other hand, some differences became more pronounced throughout development, consistent with a dosage effect. This was the case for executive control, for which the reduction in interference from incongruent ANTI trials across years was more pronounced among participants with formal musical training. A similar pattern was observed in AV trials, where the overall mean response decreased more sharply with age among musically trained participants. For EV sensitivity, we observed a performance decrement from the beginning to the end of the task, which indicates an increase in errors in detecting the signal under perceptual uncertainty. Notably, this decrement attenuated progressively across childhood and adolescence, leading to the absence of any TOT effect on sensitivity in the adulthood (Gyles et al., 2023; Román‐Caballero et al., 2023). Crucially, a more pronounced attenuation was observed in individuals with musical training. Although these differences increased across age, which has been suggested in the literature as stronger evidence of a causal direction in cross‐sectional studies (Talamini et al., 2025), the influence of variables of selection bias such as socioeconomic status over neurocognitive development also seem to accumulate over development (Dennis et al., 2022; Noble et al., 2015). Therefore, these results cannot discard an explanation based on selection bias.

Importantly, the magnitude of the observed effects was modest and significantly smaller than in other cross‐sectional studies (Medina & Barraza, 2019), as expected given the high level of control in our design. Prior studies that aimed to isolate the specific contribution of musical training and separate it from confounding variables have similarly reported modest transfer effects (Guhn et al., 2020; Gustavson et al., 2021; Nichols et al., 2020). In the present study, a large number of inclusion variables and potential confounders were controlled for with propensity‐score matching. However, matching procedures have their own limitations and do not ensure a full isolation of the target effect. In our case, the control relies on the type of measurement of the confounding variables, in which issues such as measurement error, social desirability, or the person answering (caregivers for children/adolescents vs. the person themselves for adults) in self‐report might impact the matching outcome. Similarly, the method cannot account for unmeasured factors, such as genetic predispositions, which play a role in associations between musical engagement and cognitive abilities (Gustavson et al., 2021; Wesseldijk et al., 2023). Finally, given that musicians show specific characteristics (higher openness to experience, from families with higher socioeconomic status, etc.), the resulting reference group after matching will be less representative of the general population. It is important to note that statistical control does not necessarily protect against false causal conclusions. When measurement error in the dependent variable or the confounder is substantial, Type I error rates may exceed the nominal level (Westfall & Yarkoni, 2016), leading sometimes to erroneous inferences about an effect. Moreover, when the two variables in a presumed causal relationship are not directly associated but both relate to a third variable (i.e., a collider), statistically controlling for that variable can induce a spurious association, a phenomenon known as collider bias (Rohrer, 2018). Although prior research consistently identifies all third variables included in our propensity‐score matching procedure as confounders (e.g., Corrigall et al., 2013), measurement error in our observed variables warrants caution when generalizing these conclusions to the level of latent constructs (Westfall & Yarkoni, 2016).

Although our findings are compatible with the possibility that formal musical training, which involves mastering complex skills such as reading sheet music and playing an instrument, might confer benefits to domain‐general cognitive abilities like attention and vigilance, an important question remains regarding the practical significance of such effects. The standard Spanish conservatory program spans approximately a decade (longer still if pursued professionally). The small benefit of that training would equate to an improvement of 1–3 IQ points (Schellenberg, 2004). Such an effect is unlikely to impact meaningfully everyday functioning, and from this perspective, formal musical training may not be one of the first‐choice interventions if the sole aim is cognitive enhancement (Román‐Caballero et al., 2022). At the same time, evidence of neurocognitive changes in nonmusical domains caused by musical training brings back the notion that music can serve as a valuable framework for studying neuroplasticity, extending beyond the auditory, somatosensory and motor systems (Herholz & Zatorre, 2012). Also, the main reason music has been practiced across societies for centuries is not its role as a potential cognitive enhancer, but rather its ability to provide pleasure and foster social connection (Mas‐Herrero et al., 2013; Savage et al., 2021).

Although we have observed differences in many attentional and vigilance outcomes, it is important to note that musicians did not outperform nonmusicians in all the measured domains. Thus, the results for orienting and EV criterion were similar across groups, which is relevant to understanding that some cognitive domains may be less malleable with musical training or less likely to be part of the selection background for this activity, depending on the direction in which the correlation is interpreted. In the same vein, despite the large number of comparisons in the present study, only one of the eight attentional measures showing group differences (phasic alertness) did not remain significant after applying a Benjamini and Hochberg (1995) correction (from p = .039 uncorrected to p = .065 corrected), underscoring the robustness of these findings.

The ability to remain focused over extended periods might lead individuals with formal musical training to perform better on tasks of domains other than attention, such as episodic memory (Bugos & Wang, 2022) and academic achievement (Guhn et al., 2020). Therefore, the wide‐ranging improvements reported in prior research may be mediated by enhancements in a narrower set of core cognitive functions, whose benefits would generalize across domains (Hannon & Trainor, 2007; Moreno & Farzan, 2015). Future studies could explore this possibility by examining attention not only as an outcome but also as a potential mediator of group differences observed in other cognitive domains.

A plausible reason why musical training might enhance attentional abilities lies in the nature of music performance itself. Playing an instrument – particularly in structured, semi‐professional, or professional contexts such as concerts – requires sustained engagement and continuous monitoring of external cues, from other performers' signals to subtle deviations or mistakes. These high demands on vigilance could, in principle, serve as implicit practice for maintaining alertness in other, nonmusical tasks (Román‐Caballero et al., 2021). Still, this kind of ‘use‐it‐to‐improve‐it’ explanation has been criticized for oversimplifying how training effects emerge (Gathercole et al., 2019; Simons et al., 2016; Taatgen, 2013). An alternative perspective suggests that learning and performing music fosters the acquisition of complex cognitive routines and strategies that can later be applied in nonmusical domains (Gathercole et al., 2019), therefore increasing the efficiency in the use of cognitive resources rather than cognitive capacity itself (von Bastian et al., 2022). Nevertheless, as we have mentioned, the observed differences might come from preexisting differences and selection bias, so that musical training would be more a consequence than a cause of high cognitive functioning.

Furthermore, cognitive advantages in musicians might not only come from improvements in cognitive capacity, efficiency and skills. In the same vein, musical training, in contrast to other cognitive activities, constitutes a continuously rewarding activity with multiple feedback sources (teachers, audience, colleagues, the perceived accuracy of the produced sounds, etc.). We speculate that musical training may promote effort itself, by becoming a secondary reinforcer and being rewarding in its own right (i.e., learned industriousness; Inzlicht et al., 2018). Learning to play an instrument constitutes a prolonged successful experience that is worth investing effort in, as it leads to successful mastery of the instrument. As an alternative to accounts based on cognitive enhancements, people with musical training might become more willing to exert effort and approach any demanding tasks, such as cognitive tests, with a different motivational state. In this sense, musical training would improve performance by altering the effort willingness of the musicians, not their cognitive capacities themselves.

CONCLUSIONS

The present cross‐sectional study provides new evidence that formal musical training is associated with superior attention and vigilance across development. The thorough control of confounding variables in the design was intended to provide a closer estimate of the differences between musicians and nonmusicians in isolation from other factors of selection bias. While the observed benefits were modest, they are consistent with the idea that musical training can foster domain‐general cognitive skills, although the alternative explanation of preexisting differences cannot be discarded. Moreover, some group differences were already present in the youngest participants, while others became more pronounced with age, depending on the cognitive process, which suggests the possibility of both a threshold and dosage effects, and perhaps both directions of the observed association between musical training and attention: better attention might make individuals more likely to engage in musical training, which in turn could further enhance their attentional skills. This pattern would be consistent with the conclusions of recent meta‐analyses (Román‐Caballero et al., 2022) and longitudinal twin studies (Gustavson et al., 2021).

AUTHOR CONTRIBUTIONS

Rafael Román‐Caballero: Conceptualization; investigation; funding acquisition; writing – original draft; methodology; validation; visualization; writing – review and editing; software; formal analysis; project administration; data curation; supervision; resources. Laura Trujillo: Data curation; formal analysis; investigation; methodology; project administration; validation; writing – review and editing. Paulina del Carmen Martín‐Sánchez: Project administration; data curation; investigation. Elisa Martín‐Arévalo: Conceptualization; methodology; resources; supervision; writing – review and editing. Juan Lupiáñez: Writing – review and editing; conceptualization; formal analysis; funding acquisition; methodology; project administration; resources; supervision.

Supporting information

Data S1. Supporting Information.

DATA AVAILABILITY STATEMENT

The task and further documentation are available at https://anti‐vea.ugr.es. Data and analysis scripts are publicly available at https://osf.io/8vqyu/.