Identification and Selection of the Study Sample

A meta-analysis was conducted following PRISMA statement guidelines to ensure a relevant and accurate search of the study topic. The process consisted of three phases: identification, screening, and inclusion (^{Page et al., 2021}). In addition, recommendations for the appropriate reporting of meta-analysis results were followed (^{Rubio-Aparicio et al., 2018}).

Phase 3: Inclusion

Two investigators independently assessed the previous phases without discrepancies. The bias assessment also involved a third investigator, who used the PRISMA digital checklist (^{Page et al., 2021}) to assess the information incorporated in the manuscript sections. The management of bibliographic references and the elimination of duplicates were carried out using EndNote 21 software. Figure 1 shows the flow diagram with the phases followed.

Figure 1: Flow diagram of the article selection process 

Extraction of Information from the Studies

To extract information from the 18 articles, a protocol was established that included the following elements: ID, citation, study objective, country, age, educational level, sample size, educational content, duration of exposure to technology, learning model, instrument to measure learning, and descriptive statistics. The principal investigator performed the initial extraction of information, collecting the relevant data from each article according to the established protocol. The rest of the team then reviewed and verified the extraction to ensure the consistency and accuracy of the data. Table 1 ², ⁴, ⁵, ⁸, ¹³, ¹⁵, ²⁴, ³⁰, ³⁸, ⁴⁰, ⁴⁷, ⁴⁹, ⁵¹, ⁵⁴, ⁵⁵, ⁵⁶, presents the extraction matrix with the information associated with each study. The results from this table will be presented to answer RQ1.

Table 1: Extraction matrix: general characteristics 

Table 2 presents the descriptive statistics (mean, standard deviation, and number of cases) associated with each study. The results from this table will be presented to answer RQ2.

Table 2: Extraction matrix: descriptive statistics 

Data Analysis

The analysis was performed using the standardized mean difference as the outcome measure, which allows for the comparison of effects across studies using different outcome scales. A random-effects model was fitted to the data, assuming that the true effects vary between studies due to differences in context, intervention design, or population characteristics (^{Rubio-Aparicio et al., 2018}).

Heterogeneity (i.e., tau²) was obtained using the restricted maximum likelihood estimator (^{Viechtbauer, 2005}). In addition to the estimation of tau², the Q-test for heterogeneity and the I² statistic were calculated. If some degree of heterogeneity is detected (i.e., tau² > 0, regardless of the Q-test results), a prediction interval for the true effect is also provided. Studentized residuals and Cook's distances are used to examine whether the studies may be outliers and/or influential in the context of the model.

Publication bias was assessed using the Fail-Safe N, which estimates the number of unpublished or missing studies with null or non-significant results were included; the overall results of the meta-analysis would still be statistically significant or consistent.

In addition, bias was assessed with Egger's Regression, which examines the association between effect sizes and their precision. A non-significant p-value (e.g., p > 0.05) indicates no evidence of publication bias. However, a significant p value (e.g., p < 0.05) suggests the presence of publication bias but may also mean that the sample size is too small, or that there is substantial heterogeneity among the included studies (Egger et al., 1997).

Data analysis was performed using the MAJOR module of JAMOVI software version 2.3.13.0.

General Characteristics of the Studies (RQ1)

The studies come from Turkey (39%; n = 7), Taiwan (28%; n = 5), Saudi Arabia (17%; n = 3), Malaysia (6%; n = 1), China (6%; n = 1), and Denmark (6%; n = 1). The educational levels with the highest prevalence are 6th grade (50%; n = 9) and 7th grade (28%; n = 5). Sample sizes range from 22 to 102 participants, distributed across control (absence of technologies) and experimental (presence of immersive technologies) groups.

The interventions address educational content in science (67%; n = 12), English (28%; n = 5), and programming (6%; n = 1). The reported learning models include mobile learning (17%; n = 3), experiential learning (11%; n = 2), inquiry-based learning (6%; n = 1), the SMAR model (6%; n = 1), creative situated learning (6%; n = 1), and collaborative learning (6%; n = 1). Fifty percent of the studies do not report a specific learning theory guiding the integration of immersive technologies.

Exposure time is reported in two main formats: by number of weeks (61%; n = 11), ranging from 3 to 7 weeks, or by number of sessions (39%; n = 7), ranging from 1 to 6 sessions.

Regarding the instruments, the studies generally employ ad hoc content tests to measure academic achievement, mostly based on multiple-choice items (94%; n = 17). Only one study reports using a test previously developed by other authors.

Effectiveness of the Integration of Immersive Technologies (RQ2)

Table 3presents the results of the heterogeneity test, showing significant heterogeneity among the 18 studies (Q(17)=128.745, p<0.0001), considerable variability among the effects of the individual studies (tau²=0.7406) and observed variability (I²=89%), suggesting that the studies originate from different populations.

Table 3: Heterogeneity statistics 

Table 4 ¹⁷, presents the results of the publication bias; the Fail-Safe N indicates that at least 1,043 unfound or unpublished studies with null results would be necessary for the results of the current meta-analysis to be insignificant. For its part, the Egger's Regression test indicates that there is significant statistical evidence of asymmetry in the data. This suggests the possible presence of a publication bias, where studies with significant results are more likely to be published than those with non-significant results.

Table 4: Publication bias assessment 

Note. Fail-safe N calculation using the Rosenthal approach

An examination of the studentized residuals revealed that none of the studies had a value greater than ±2.9913, indicating no outliers in the context of this model. According to Cook’s distances, none of the studies could be considered overly influential. Figure 2 presents the funnel plot.

Figure 2: Funnel Plot 

Both the rank correlation test and the regression test indicate possible skewness (p = .0022 and p = .0005, respectively), suggesting the presence of publication bias. This asymmetry may reflect a tendency to publish studies with significant or positive results, while non-significant or negative findings remain unpublished or are less accessible.

Table 5presents the results of the random-effects model for the 18 studies. The analysis estimates a large effect size of 1.02 (^{Sawilowsky, 2009}). However, due to variability across studies, the true population effect may range from 0.591 to 1.443, with a 95% confidence interval.

Table 5: Random-Effects Model (k = 18) 

Note. Tau² Estimator: Restricted Maximum-Likelihood.

Figure 3 displays the forest plot, which includes 18 effect sizes, each represented by an asterisk in the central column, with horizontal lines indicating their corresponding 95% confidence intervals.

Figure 3: Forest Plot 

The dashed vertical line denotes the line of no effect, while the diamond at the bottom represents the overall pooled effect size. The position of the diamond to the right of the vertical line indicates that the experimental group achieved significantly better learning outcomes compared to the control group.

General Characteristics of the Studies (RQ1)

According to previous findings and systematic reviews (^{Altinpulluk, 2019}; ^{Garzón et al., 2019}; ^{Qiu et al., 2023}), there is greater scientific production on the use of immersive technologies in Asia. This corroborates that certain countries in that region are showing increasing interest and investment in integrating technologies to enhance the learning experience in primary education.

Regarding the age of the participants, the studies were conducted with students aged 11 and 12, corresponding to sixth and seventh grades. At this stage of cognitive and psychosocial development, students are undergoing changes in how they think, process information, and understand the world, as well as in how they interact socially (^{Erikson, 1985}; ^{Piaget, 1974}). The visualization features of AR and VR can have a significant impact on students' learning and development at this age, due to several factors identified in the literature: the exploration of concepts and scenarios in an interactive way; increased engagement in learning, as they are at an age when interest may fluctuate due to personal and contextual factors; development of cognitive and social skills; and sensory stimulation that supports different learning styles (^{Akçayır & Akçayır, 2017}; ^{Hamilton et al., 2021}; ^{Villena-Taranilla et al., 2022}).

Concerning the educational interventions, the studies mainly address science-related content. This finding is consistent with other reviews that highlight the widespread use of immersive technologies for learning about cells, human body systems, the planetary system, flora, and fauna (^{Garzón et al., 2019}; ^{Pellas et al., 2021}; ^{Mystakidis et al., 2022}; ^{Oyelere et al., 2020}). As for the time of exposure to the technology, it varies in how it is reported (e.g., weeks, sessions, hours), which constitutes a limitation. The lack of consistency in reporting exposure time makes it difficult to replicate interventions or draw generalizable conclusions. For example, the study by ^{Hashim et al. (2022}), which reports improvements in learning after several weeks of exposure, may not be comparable to the study by Tsai and Yu-^{Cheng (2022}), which only reports minutes of use in a single session.

The studies apply different learning theories to the design of technology-based interventions, with mobile learning and experiential learning being the most common. However, the lack of an explicit theoretical framework in 50% of the studies analyzed is a recurring limitation in educational research (^{Buchner & Kerres, 2023}; ^{Mystakidis et al., 2022}; ^{Qiu et al., 2023}). Learning theories provide the conceptual basis for selecting and designing pedagogical strategies aligned with educational goals and student needs. When theories are not referenced, it becomes difficult to understand the rationale behind the integration of AR and VR in the classroom (^{Sandoval-Henríquez & Badilla-Quintana, 2021}).

Regarding assessment, the studies employed content tests to measure academic achievement. None of the reviewed studies used indirect measures such as perceived learning. This aligns with the findings of previous reviews (^{Hamilton et al., 2021}). The inclusion of indirect measures alongside traditional assessments could offer a broader understanding of the motivational, emotional, and engagement-related dimensions associated with immersive technologies and their effect on learning. Pedagogical models such as co-association also promote the use of alternative forms of assessment, such as self-assessment and peer assessment, which help students become aware of their own progress and develop self-regulation skills (^{Prensky, 2013}).

Effectiveness of the Integration of Immersive Technologies (RQ2)

Individual studies report that the integration of immersive technologies has a positive impact on academic achievement compared to control conditions. In this regard, learning gains are a commonly reported educational benefit in previous systematic reviews (^{Buchner & Kerres, 2023}; ^{Garzón et al., 2019}; ^{Mystakidis et al., 2022}). The results of the meta-analysis indicate an effect size of 1.02, which is considered large (^{Sawilowsky, 2009}). However, due to variability across the 18 studies, the true population effect may range from 0.59 to 1.44 with a 95% confidence interval-i.e., a moderate to large effect.

These findings are consistent with those of other meta-analyses. ^{Villena-Taranilla et al. (2022}) reported a moderate effect size (g = 0.64), noting that shorter interventions (less than two hours) were associated with greater learning effects. Similarly, ^{Chang et al. (2022}) found a moderate effect size (g = 0.65) for the impact of immersive technologies on academic achievement. ^{Garzón et al. (2019}) also reported a moderate effect size (g = 0.65), with higher effectiveness observed in science, arts, and humanities. In contrast, ^{Cao and Yu (2023}) reported a significant difference between experimental and control groups, with the former achieving a large effect size (g = 0.85).

However, some meta-analyses have found more modest or inconsistent results. For example, ^{Coban et al. (2022}) reported a small effect size (g = 0.38) for the impact of VR on learning outcomes. Interestingly, their study also found that immersive technologies had a significantly larger effect in elementary education compared to higher education. These contrasting results underscore the need for further research that considers factors such as educational level, intervention duration, subject area, and the degree of immersion (fully immersive, semi-immersive, or non-immersive) to better understand the impact of these technologies on learning.