Augmented Reality in Higher Education: A Systematic Review and Meta-Analysis of the Literature from 2000 to 2023

Abstract

Augmented reality (AR) has been widely used in higher education because of its unique characteristics that combine virtuality and reality. This study systematically reviewed the literature on the application of AR in higher education between 2000 and 2023. A total of 237 articles were selected, and the key research findings were analyzed and synthesized based on the coding manual from five aspects: basic information, disciplines, technology features, instructional design, and research results. The results revealed interesting findings regarding AR literature in terms of publication trends and applied disciplines, technical features and affordances, instructional design, and learning outcomes and measurement methods. Furthermore, a meta-analysis was conducted on 60 experimental studies selected from the literature to examine the overall effectiveness of AR-based instructions. The results indicate that AR applications in higher education tend to have a large overall effect size (g = 0.896, 95% confidence interval = [0.685–1.107], p = 0.000), with two significant moderators: instructional function and learning outcomes. Additional sensitivity and publication bias analyses confirmed the robustness of the meta-analytic findings. Based on these results, four implications for educational practice and research investigating AR-supported higher education were proposed and discussed.

1. Introduction

The EDUCAUSE Horizon Report released in 2020 identified augmented reality (AR) as one of the six emerging technologies that will have the greatest impact on teaching and learning in the future. AR refers to a mixed environment that integrates digital information within a real-world setting (Wu et al., 2013), which emphasizes precise matching between virtual objects and 3D models of real-world objects to achieve a combination of reality and virtuality, that is, real-time interaction (Akçayır & Akçayır, 2017). Moreover, AR does not seek to replace physical reality with virtual elements; rather, it enhances physical reality by synchronously overlaying virtual objects, which take the form of text, still images, audio, video, or 2D or 3D models.

Based on these characteristics, the positive effect of AR-supported instruction was demonstrated in systematic reviews, manifested as learner outcomes, pedagogical contributions, and interactions (Akçayır & Akçayır, 2017; Radu, 2014). Previous meta-analyses showed a medium effect of AR implementation in educational settings (Garzón et al., 2020, 2019). AR technology has the potential to be an effective means of teaching innovation in higher education for two reasons. First, learners in higher education are generally proficient in using mobile devices (e.g., mobile phones or tablets) that can support AR-based education and teaching. Second, the varied self-directed learning paradigms inherent in higher education disciplines are highly amenable to the exploration and implementation of diverse AR-based instructional models. Consequently, AR applications have received increasing attention and have been widely applied in various fields including medicine (Mahrous et al., 2021), basic science (Wong et al., 2021), and engineering (Álvarez-Marín et al., 2021). The results of these studies indicate that AR-supported instruction has the potential to enhance learners’ learning experience by fostering a more personalized approach to learning, promoting collaboration, and improving problem-solving abilities (Ke & Hsu, 2015).

However, several issues exist regarding the effectiveness of AR-supported instruction in higher education. First, the researchers claimed that the empirical research studies were media comparison studies, there was a lack of consideration of explaining when and how AR is beneficial (Garzón et al., 2019; Buchner & Kerres, 2023; Sümer & Vaněček, 2024). Second, previous reviews have only focused on specific fields such as STEM (Mystakidis et al., 2021), medicine (Moro et al., 2020), or second language acquisition, but there has been a lack of discussion on the overall application and distribution across different disciplines, which leads to conclusions with limited representativeness. Additionally, no meta-analysis has examined the role of discipline as a moderating variable. Previous reviews have not adequately addressed the distinctive attributes of AR and associated pedagogical approaches, and have also lacked exploration of the long-term trajectory and evolution of emerging trends.

To address this research gap, this article provides a systematic review of the applications of AR in higher education from 2000 to 2023, focusing on four aspects: publication trends and applied disciplines, technical features and affordances, instructional design, learning outcomes and measurement methods. Experimental studies were then selected to examine the overall effectiveness of AR-based instruction and to explore potential moderating factors. We expect the systematic review can provide a comprehensive summary of the application status of AR in higher education based on the key aspects of a technology-enhanced learning environment, including context, pedagogy, technology, and effectiveness (Wang & Hannafin, 2005). These aspects were potential moderating factors in this meta-analysis. Additionally, the results of the meta-analysis should provide evidence (as in a primary research study) that supports the claims derived from the systematic review (Pigott & Polanin, 2020). Specifically, the research questions addressed in this study were as follows:

(1)

What are the overall publication trends and characteristics of the distribution of AR applications across different disciplines in higher education?

(2)

What are the essential technical features and affordances of AR in higher education applications, and how are they evolving over time?

(3)

What instructional design approaches are used to facilitate teaching and learning in AR-supported higher education?

(4)

What are the common types of learning outcomes supported by AR in higher education, and how are they measured?

(5)

What is the overall effectiveness of the application of AR in higher education, and what are the moderating variables?

2. Method

This study conducted a systematic review to gain insights into the application of AR in higher education. The review process consisted of three main stages: the initial literature search, manual screening, and analytical coding. A meta-analysis was also conducted to examine the overall effectiveness of AR applications in higher education and its moderating factors. This study followed the Preferred Reporting Items for Systematics and Meta-analysis statement for the selection and use of research methods. The protocol was registered and published on INPLASY (202430062), which can be accessed at http://doi.org/10.37766/inplasy2024.3.0062 (accessed on 17 March 2024).

2.1. Identification

The literature search was conducted using two databases: SCOPUS and Web of Science. SCOPUS was employed as the primary database to retrieve studies owing to its comprehensiveness and multidisciplinary coverage. It indexes over 15,000 high-quality journals across various fields, such as medicine, sociology, psychology, humanities, and the arts. The Web of Science covers high-quality peer-reviewed journals and thus serves as a supplementary database. We conducted an initial literature search by using random combinations of two clusters of strings. The first cluster included “augmented reality” and its abbreviation “AR”. The second cluster of search keywords consisted of phrases such as “higher education”, teaching*, “instruction*”, and “classroom”. The final date on which the literature search was conducted and sources consulted was 10 February 2024. The initial search yielded 3453 articles.

2.2. Manual Screening and Eligibility

Based on the initially retrieved literature, we manually screened the titles and abstracts of the articles using the decision diagram shown in Figure 1. We screened the articles for eligibility. Two researchers were involved in the literature screening process. Each researcher assessed the titles and abstracts of the articles independently to minimize bias. Subsequently, any articles with uncertainties were discussed weekly by the researchers. All included studies were recorded in an Excel spreadsheet. The specific inclusion criteria were as follows: (1) all articles must focus on AR, excluding articles on MR or VR; (2) the educational context must target higher education, excluding articles applied to other educational stages (e.g., K-12 education, special education, vocational education); and (3) peer-reviewed journal articles, excluding reports and dissertations.

2.3. Analytical Coding

After finalizing the main library, we analyzed and coded the selected articles based on the coding scheme developed by our research team. As shown in

Table 1. Coding scheme for the selected articles.

Category	Code	Description
Basic information	Title Authors Year
	Research type	Empirical/design case/theoretical/synthesis
	Empirical type	Experimental/quantitative/qualitative/mixed/survey/design-based research
Disciplines	/	Basic science/social science/engineering/health & medicine/language/mathematics/other
Technology features	Input	Natural: voice command/motion/haptic/location-based Artificial: mouse & keyboard/scan/stylus/controller/other
	Output	Monitor-based/video see-through/optical see-through
	Media representation	Symbol or indicator/text/data/2D image/3D object/animation/video/audio/fantasy
Instructional design	Instructional function	Content delivery/attention grabber/engagement/practice/assessment/other
	Pedagogy	Inquiry-based/game-based/collaborative/direct instruction/experiential learning/trial-and err
Research results	Learning outcomes	Knowledge/affective/behavior or skills/capacity
	Measurement format	Knowledge test/psychometric/conversation/performance/product/other
	Statistical results	Difference (T-test/analysis of variance/multi-variant analysis of variance/analysis of covariance/non-parametric) Association: structural equation modeling/regression/factor analysis

, the coding scheme comprises five parts: basic information, disciplines, technological features, instructional design, and research results. Basic information encompasses the main metadata for the selected articles, including the title, authors, publication year, research type, and empirical type. The other four codes corresponded to our research questions. Disciplines reveal the use of AR to facilitate teaching and learning in higher education. Technology codes describe the technical features of AR, including the input, output, and media representation. Instructional design focuses on the instructional function and pedagogy of AR-based instruction in higher education. Research codes concern empirical findings such as learning outcomes, measurement formats, and statistical results. Five researchers participated in the coding process to collect data from the selected articles, with each researcher working independently. Controversies were resolved by weekly discussions to ensure coding consistency according to the formula for coding reliability (Xu & Zhang, 2005): $R=(N\times K)/\left\{1+\left(N-1\right)\times K\right\}$ where N represents the number of coders, and K represents the average inter-rater agreement: $K=5\times S/\left(N1+N2+N3+N4+N5\right)$. S indicates the number of articles coded identically by the five coders and N1–N5 refer to the number of articles coded by each coder. The final average inter-rater agreement was K = 0.72, and the reliability coefficient was R = 0.93, indicating good coding reliability.

2.4. Methods for Syntheses and Results Presentation

Figure 2 described the detailed process to decide which studies were eligible for synthesis. A total of 237 articles were included in the main library for systematic review, including empirical studies (n = 198), design cases (n = 20), systematic reviews (n = 18), and theoretical article (n = 1). Reviews and theoretical article were included in this study to provide a theoretical perspective for analytical coding. Empirical studies included 6 common types of research design. Based on the coding manual, we conducted a descriptive statistical analysis of the articles to visually display the results of the syntheses. The descriptive analysis was performed in SPSS software (version 21). Specific coding of synthesis can be accessed at Mendeley Data at https://doi.org/10.17632/fgpjs74df9.3 (accessed on 19 December 2024).

2.5. Methods for Meta-Analysis

We conducted a meta-analysis of experimental studies from 2000 to 2023 to assess the overall effect size of AR applications in higher education and its moderating variables. Five researchers worked independently to contribute of the data coding for the meta-analysis. We calculated the effect size based on the means and standard deviations of the measurements from the experimental and control groups. Articles with missing statistics were deleted, while articles with multiple dependent variables as learning outcomes were treated as separate experimental studies. Finally, 43 articles were included in the meta-analysis, representing 60 experimental studies for the analysis of effect size.

There are three major approaches for meta-analysis: publication bias test, heterogeneity test, and moderating effect test. First, a publication bias test using the trim-and-fill method based on a random-effects model with a linear estimator. This non-parametric procedure estimates and adjusts for potentially missing studies on the asymmetric side of the funnel plot, helping to ensure that the pooled effect size is not skewed by selective publication. Second, a heterogeneity test evaluated variability across studies to ensure consistency and reliability. This helped determine whether the studies were measuring the same underlying effect and whether the pooled estimate was reliable. Third, a moderating effect test was conducted to reveal the potential moderator variables, thereby revealing how the effect varies under different conditions. Additionally, a leave-one-out sensitivity analysis was conducted to test the robustness of the pooled effect size. The calculation of the meta-analysis was performed in the CMA (version 2) software. Specific coding of meta-analysis was described in the Appendix A.

2.6. Quality Appraisal

To evaluate the methodological quality of the 237 studies included in this review, we adopted the Mixed Methods Appraisal Tool (MMAT, Version 2018), which is suitable for appraising empirical, design-based, and mixed-methods studies (Hong et al., 2018). The five core MMAT criteria includes: participant representativeness, measurement appropriateness, completeness of outcome data, control for confounders, and implementation integrity. Each criterion was rated using a three-point scale: “yes”, “no”, or “can’t tell”. Based on the clarity and sufficiency of information provided. A study was assigned an overall quality level as follows: high quality (5 “yes”), moderate (3–4 yes), low/unclear (2 or fewer yes), and not applicable (synthesis or theoretical articles where MMAT is not suitable). Among the included studies, 198 (83.5%) were rated as high quality, 20 (8.4%) as moderate quality, and 19 (8.0%) as not applicable. A complete list of 237 articles can be found in Supplementary File S2. A complete matrix of study-level quality ratings is provided in Supplementary File S3.

3. Systematic Review Results

3.1. Publication Trends and Applied Disciplines

3.1.1. Publication Trends

Figure 3 depicts the number of publications on AR-based instruction in higher education, including overall publication trends and specific research types, which has followed an overall upward trend in the past 24 years. The total number of publications provides a direct visualization of the overall growth of AR technology in the field of higher education, whereas the trends in publications by research type provide a deeper understanding of the application of various research methodologies.

The results indicated that the application of AR in higher education did not receive much attention until 2003. Two significant peaks were observed in 2015 and 2020. The first peak could be attributed to the introduction of AR games and the release of Google Glass in 2014. These developments have boosted the application and research on AR in higher education. The emergence of the second peak may be related to the commercialization of the Oculus head-mounted device (HMD). In 2020, with a decrease in prices, AR technology became widely used in various fields, especially in education. However, from 2021 to 2022, the literature on AR-based instruction in higher education has shown a downward trend. This may be due to the absence of technical advancements in AR and the lack of innovative products to support teaching methods, such as collaborative or inquiry-based learning.

3.1.2. Disciplines

As shown in Figure 4, AR-based instruction in higher education has primarily been utilized in the fields of health/medicine and engineering, which aligns with the results of previous reviews (Alvarez-Marin & Velazquez-Iturbide, 2021; Han et al., 2022). Healthcare and engineering education rely on experiential learning to meet professional standards, but in-person training can be limited by safety, cost, etc. AR provides a viable alternative with its interactive and visual features. In addition, the AR-based instruction is also widely used in basic disciplines and other informal disciplines, such as STEM and maker education. Some researchers have emphasized that AR can effectively support learning in these fields by improving learners’ spatial ability, practical skills, conceptual understanding and comprehension, and scientific inquiry abilities (Ibáñez & Delgado-Kloos, 2018).

In contrast, the results show that AR is not commonly used in the fields of language and mathematics in higher education, perhaps because of the characteristics of these disciplines. Mathematics, for example, emphasizes a high degree of abstraction and complex concepts, while second language acquisition requires students to engage in extensive oral practice with a focus on reading and writing. Currently, AR primarily overlays virtual objects on physical environments to visualize simple concepts or events. This may not effectively support learning in mathematics and language education, indicating a mismatch between specific learning tasks in these fields and AR functionalities. Additionally, higher teaching costs and cognitive overload for students could contribute to the infrequent use of AR in language and mathematics education.

3.2. Technical Features and Affordance

3.2.1. Input

User input is an important technical feature of AR in higher education. Based on user experience, technological dependency, and adaptability to educational scenarios, we classified input devices into two categories: natural and artificial. Natural input refers to the interaction between users and AR objects without the need for external equipment. Common types of natural inputs include voice commands, magnetic sensors, motion trackers, and haptic sensors. In contrast, artificial interaction relies on equipment to achieve interaction with the interface; this includes the use of a mouse and keyboard, a scan, a stylus, or a controller. A total of 175 research studies reported AR input, with 49 studies using natural interactions, while the remaining 126 reported artificial interactions.

As shown in Figure 5a, scans (n = 111) accounts for the largest proportion of the input from 2000 to 2023. Scans from 2000 to 2011 were mainly conducted in laboratory environments, for example, using computers to recognize and scan QR codes to obtain 3D graphics for instruction (Martín-Gutiérrez et al., 2010). In contrast, from 2012 to 2023, scans were mainly performed using the cameras of portable devices (e.g., mobile phones and tablets) to capture and scan QR codes. For example, some researchers designed and developed a collaborative and interactive AR learning environment for civil engineering students, enabling them to use handheld devices to scan and gain knowledge (Shirazi & Behzadan, 2015).

Additionally, inputs from haptic devices and motion sensors gradually declined from 2012 to 2017 and 2018 to 2023. Traditional input devices such as mouse and keyboard, stylus decreased, indicating a move away from desktop-based setups. Compared with the first 12 years of AR research, there has been an increase in diverse input methods, such as voice command and location-based features, from 2012 to 2023. Location-based features rely on geographic positioning. For example, researchers have used a geographic positioning system to establish a visual 3D model in architecture education, which is connected to a virtual information channel through a database and allows location-based geofencing (Sánchez Riera et al., 2015).

3.2.2. Output

AR commonly features three types of outputs: monitor-based, video see-through, and optical see-through. Monitor-based output includes the projection of images onto a screen, typically a computer monitor. Video see-through involves the presentation of outputs through HMDs, mobile/handheld devices, or glasses. Optical see-through refers to real-world images directly refracted into the user’s eyes after light modulation processing, while virtual channel information is refracted into the user’s eyes through projection and reflection without the need for other output devices. As shown in Figure 5b, monitor-based output was more commonly used in the first 12 years, whereas video see-through has been incorporated in a large proportion of studies in the last 12 years. From 2012 to 2017, video see-through technologies emerged as the dominant output method, increasing to 64%. This trend continued into the 2018–2023 period, where video see-through maintained its dominant position. As technology evolves, numerous HMDs and handheld devices have emerged in the market that can serve as better support output devices. The number of monitor-based and video see-through output devices has remained stable, whereas the application of optical see-through has been limited, which indicates a lack of breakthroughs in the development of 3D imaging.

3.2.3. Media Representation

Figure 5c identifies the nine common media representations in the AR environment. The most common media representation is 3D objects, particularly in the fields of architecture and medicine. Such images provide learners with a more intuitive experience and opportunities for manipulation. Another important information carrier is a 2D image, which is typically based on virtual simulation to present virtual characters or teaching content. However, this type of presentation offers learners limited affordances and interactions. It is worth noting that few studies have created virtual educational scenarios to provide learners with fantasy experience. For example, researcher designed an AR game called “environmental detective”, which enabled students to choose different roles and collaboratively explore sewage treatment using GPS positioning (Klopfer & Squire, 2008). This also reflects the characteristics of AR, which combines virtual and real worlds to promote collaboration and completion of exploratory tasks.

3.3. Instructional Design

To ensure the effective implementation of AR in educational settings, it is necessary to focus on the integration of technological applications and teaching methods as well as to understand the instructional functions and pedagogy supported by AR (see Figure 6).

3.3.1. Instructional Functions

We identified six instructional functions of AR in higher education, as shown in Figure 6a, among which content delivery accounted for the highest percentage (51%). This is related to the characteristics of AR, which emphasize the overlay of virtual objects onto the real world to provide supplementary material for learners to explore the real world. AR can also support learners’ interactions with learning materials from multiple perspectives in the real world (e.g., zooming, rotating, and editing). Compared with VR, AR has lower costs and simpler platform operation. However, it should be noted that some studies only used AR to present non-interactive teaching materials in a simple way, without fully utilizing its potential advantages (Mahmoudi et al., 2017).

AR can also be used as a practice tool (20%) for courses with strong operational requirements, such as medicine and computer programming, to provide operational guidance or real-time feedback. For example, researchers have used AR to teach medical students surgical location points (Logishetty et al., 2018). Chung et al. (2021) embedded visual programming problems in AR to display real-time operational results and cultivate learners’ ability to collaborate and solve problems through programming. The other two common instructional functions of AR are as an attention grabber (10%) and to increase learning engagement (14%). These two functions often co-occur with other instructional functions; however, they are rarely implemented independently (Chang et al., 2014).

3.3.2. Pedagogy

Figure 6b illustrates the six pedagogies that are supported by AR in higher education. Direct instruction is the most widely used (44%), which involves intuitively presenting 3D, microscopic, and procedural teaching content in a digital format for purposes such as explaining code, displaying the 3D model of the matrix, and demonstrating dynamic changes after running the code while learning matrix algorithms (Paredes-Velasco et al., 2023). In addition, AR can provide more natural and diverse ways of interacting with inquiry-based learning. For instance, researchers designed an AR-enhanced mobile robot course that provides guidance and feedback for the entire inquiry activity. This course guides learners to gradually accomplish machine assembly through arrow symbols, text, and colors; showcases the model and function of components; and displays virtual values and running status for different functions.

Other pedagogies have been applied to a certain extent but with minimal quantitative differences. It is worth mentioning that game-based learning has been applied to AR-based instructions. For example, Lee (2020) designed a role-playing educational game for second language learning, where learners solved a virtual campus murder case through investigation, communication, and collaboration.

3.4. Learning Outcomes and Measurement Methods

Figure 7 illustrates the learning outcomes supported by AR from the perspectives of knowledge, behavior, and affective outcomes. Knowledge (n = 106) and affective (n = 116) outcomes received the most attention. Researchers frequently employed knowledge tests (n = 61) to assess participants’ knowledge acquisition, whereas 34 studies used psychological scales. Only a few knowledge evaluation methods have adopted qualitative evaluation methods such as conversation, performance, and production. In terms of affective outcomes, most studies used psychological scales (n = 90). Researchers have used AR to explore its effects in assembling and exploring educational robots (AlNajdi et al., 2018), using the Intrinsic Motivation Inventory developed by Ryan and Deci (2000) to assess learners’ perception of usefulness, enjoyment, and ability improvement in AR experiments.

Additionally, 50 studies measured learners’ behaviors or skills; performance was the most frequently used measurement method, followed by psychometric measurement. For example, Wolf et al. (2021) used AR to educate medical students on the insertion of extracorporeal membrane oxygenation and then conducted an analysis of AR-captured error logs to assess their performance during AR learning.

Furthermore, 34 studies assessed learners’ capacity, with measurement methods primarily focused on knowledge tests and psychological scales. Only a few ability assessments used alternative methods, such as interviews (n = 4), behavioral observation (n = 3), or homework and product evaluation (n = 2). Researchers frequently rely on subjective self-reporting methods to assess AR learning outcomes in higher education, which may influence the evaluation of AR teaching effectiveness.

4. Meta-Analysis Results

4.1. Publication Bias

Publication bias was tested using funnel plots and Rosenthal’s Fail-safe N. Before conducting the overall effect test, we performed a publication bias test on the 60 included studies (shown in Figure 8).

The funnel plot indicated that the effect sizes were evenly distributed around the average effect size and concentrated at the top, which suggests a low level of publication bias. Additionally, Rosenthal’s Fail-safe N (N = 8550) was much larger than the tolerance level of 5 K + 10, where K refers to the total number of effect sizes included in the meta-analysis (K = 60). Thus, publication bias of the included studies was low and did not affect the overall results of the meta-analysis.

4.2. Heterogeneity Test

To further clarify the selection of a fixed-effects model or a random-effects model (REM) to calculate the effect size, a heterogeneity test was conducted on the selected experimental studies. The results showed significant heterogeneity among the selected sample studies (Q_B = 699.472, I² = 91.6%, p = 0.000). Higgins (2003) suggested that an I² value greater than 75% indicates high heterogeneity among research studies. Therefore, we chose REM for the overall calculation of effect size. The REM results showed that AR-based instruction had a significant effect on learning outcomes (g = 0.896, SE = 0.108, CI = [0.685–1.107], p = 0.000). The range of effect sizes fluctuated between −0.871 and 7.675. Only two study had a negative effect size, while the other 58 studies had positive effect sizes.

There were 23 studies (38.3%) with a large effect size (greater than 0.8), 12 studies (20%) with a moderate effect size (between 0.5 and 0.8), and 25 studies (41.7%) with a small effect size (less than 0.5).

4.3. Moderating Effect

This study employed a mixed effects analysis to further investigate the potential moderating factors that could influence the effectiveness of AR-based instruction. The results are presented in

Table 2. Moderation analysis of selected experimental studies.

Moderator	K	g	95% CI	QB	p-Value
Discipline				7.036	0.218
Basic science	21	0.740 b	[0.404–1.076]
Social science	11	1.062 a	[0.588–1.536]
Engineering	9	0.840 a	[0.298–1.381]
Health & medicine	14	0.920 a	[0.493–1.348]
Language	2	0.373 b	[−0.706–1.452]
Other	3	2.081 a	[1.019–3.143]
Input				2.071	0.150
Natural	25	0.713 b	[0.390–1.037]
Artificial	35	1.024 a	[0.751–1.298]
Output				0.373	0.830
Monitor-based	12	0.997 a	[0.515–1.480]
Video see through	46	0.883 a	[0.640–1.127]
Optical see through	2	0.637 b	[−0.520–1.793]
Pedagogy				6.254	0.282
Inquiry-based	7	0.755 b	[0.061–1.449]
Game-based	2	0.761 b	[−0.392–1.915]
Experiential learning	11	0.693 b	[0.192–1.194]
Direct instruction	26	1.213 a	[0.881–1.546]
Collaborative	10	0.569 b	[0.052–1.085]
Trial-and err	4	0.622 b	[−0.256–1.500]
Instructional function				11.939 **	0.008
Attention grabber	5	0.265 c	[−0.426–0.956]
Content delivery	31	1.224 a	[0.939–1.509]
Practice	9	0.746 b	[0.166–1.327]
Engagement	15	0.519 b	[0.118–0.920]
Learning outcomes				11.069 *	0.026
Knowledge	28	0.914 a	[0.062–1.225]
Behavior/skills	8	1.846 a	[1.199–2.494]
Affective	16	0.665 b	[0.248–1.803]
Capacity	6	0.476 b	[−0.186–1.139]
Other	2	0.731 b	[−0.468–1.930]

Note: K = number of independent studies; g = mean effect size; Q_B = between-group homogeneity. Effect sizes of 0.2, 0.5, and 0.8 were treated as small, medium, and large, respectively. ** p < 0.01, * p < 0.05. ^a Large effect; ^b Medium effect; ^c Small effect.

, which reveal a significant difference in the effectiveness of AR-based instruction in terms of instructional function (Q_B = 11.939, p = 0.008) and learning outcomes (Q_B = 11.069, p = 0.026). Regarding instructional function, content delivery (g = 1.224) showed the largest effect size, while attention grabber (g = 0.265) showed the smallest effect size. Regarding learning outcomes, behavioral/skills (g = 1.846) and knowledge tests (g = 0.914) showed the largest effect sizes, while capacity (g = 0.476) showed the smallest effect size.

However, no significant differences in the effectiveness of AR-supported instruction were observed across disciplines, inputs, outputs, and pedagogy. In terms of disciplines, AR applied to other fields (g = 2.081) and social science (g = 1.062) exhibited the largest effect size, while it exhibited the smallest effect sizes in language learning (g = 0.373). Regarding technical features, differences in input and output did not yield significant variations in effect sizes. In terms of the pedagogy used in AR-based instruction, direct instruction (g = 1.213) and game-based learning (g = 0.761) had a larger effect size, whereas collaborative learning showed the smallest effect size (g = 0.569).

4.4. Publication Bias and Sensitivity Analysis

Publication bias and sensitivity analysis were conducted to enhance the robustness of the meta-analysis results. First, the results of leave-one-out sensitivity analysis indicate that the overall effect size remained stable across all iterations, ranging from 0.812 to 0.925 (M = 0.896, SD = 0.021), suggesting that no single study disproportionately influenced the overall outcome. Second, the Trim-and-Fill analysis did not impute any missing studies, and the pooled effect size remained unchanged before and after adjustment (observed = 0.933; observed + imputed = 0.933; 95% CI: 0.623 to 1.244). These results suggest that the distribution of effect sizes is symmetric and that publication bias is unlikely to have distorted the overall conclusions.

5. Discussion

5.1. Discussion of the Systematic Review Results

This study conducted a systematic review of AR in the field of higher education that included four aspects: the trend of publications on AR and its applied disciplines, AR’s technological features, pedagogical approaches, AR-related learning outcomes, and measurement methods. Our findings revealed the following.

The trend in publications on AR has shown an overall upward trend over the past 24 years, with two periods of peak applications in 2015 and 2020. This growth aligns with the global rise in mobile AR adoption and increasing institutional and governmental support for immersive technologies. In higher education, AR is more commonly applied in the fields of health and medicine, whereas its use in the fields of language and mathematics remains comparatively scarce. This results consistent with previous systematic review results, which also reported the predominant of AR use in STEM-related disciplines and a lack of integration in humanities-oriented fields (Bacca et al., 2014; Garzón et al., 2019).

With updates and enhancements on AR technology, the portability of devices has improved, as reflected in the increase in the use of scanning as the input method and video see-through as the output method. These developments have improved the portability and usability of AR tools in instructional settings. Similar findings were found by Akçayır and Akçayır (2017), who emphasized the significance of device ergonomics and real-time interactivity in AR usability.

Instructional use of AR remains largely teacher centered, with content delivery as the most common instructional function and direct instruction as the dominant pedagogy. This highlights a persistent gap between AR’s interactive affordances and its actual integration in learner-centered or constructivist pedagogies. Few studies incorporated collaborative, inquiry-based, or problem-based learning approaches, which limits the exploration of AR’s full educational potential. This trend aligns with the conclusions drawn by Radu (2014), who found that while AR supports visualization and engagement, it is rarely implemented through active learning strategies.

The most common types of learning outcomes assessed in studies of AR in higher education are affective outcomes, followed by knowledge. The most commonly used measurement methods are the tests and psychometric scales. These findings reflect the concern raised by Cheng and Tsai’s (2013) regarding the insufficient development of rigorous assessment frameworks for evaluating higher-order competencies in AR-based learning contexts.

5.2. Discussion of the Meta-Analysis Results

The results of the meta-analysis indicate that AR-based instruction in higher education is highly effective. AR appears to have a positive impact on learners’ learning outcomes (g = 0.896). This finding is consistent with previous studies suggesting that AR, as a novel teaching method, can effectively enhance learners’ learning effects (Kalemkuş & Kalemkuş, 2022; Ni & Hu, 2019). There are two possible explanations. First, AR enhances motivation by visualizing abstract concepts and offering diverse learning scenarios (Wu et al., 2013). Second, AR integrates digital and physical elements to create immersive environments which can foster critical thinking, problem-solving, and collaboration skills (Akçayır et al., 2016).

In terms of instructional function, AR is more commonly used to convey teaching content, with the largest effect size, whereas it is used to attract learner attention with the smallest effect size. A possible reason for this result is that the number of research studies using AR as a tool to deliver teaching content is large (N = 31), whereas there are fewer studies on using AR to attract students’ attention (N = 5). This highlights a limitation in AR’s educational integration: overemphasis on content delivery rather than leveraging the technology to design engaging learning scenarios.

In terms of pedagogy used in AR, the effect size of direct instruction and game-based learning is relatively large, whereas the effect size of collaborative learning is the smallest. This result aligns with our systematic review results, which suggest that AR-based instruction in higher education remains predominantly focused on content delivery. The reduced effectiveness of collaborative AR teaching may stem from cognitive overload caused by managing complex tasks and technological challenges simultaneously (Wu et al., 2013). Another possible reason is that collaborative learning in AR places great demand on learners’ abilities. Compared with learners with comprehensive abilities, collaborative learning has a less significant effect on students with restrictive abilities.

The results indicate that AR technology can be effectively integrated into the instruction of most disciplines with a positive and significant effect size. Among the various disciplines studied, the largest effect size was observed in fields such as fashion design (Elfeky & Elbyaly, 2018), special craftsmanship (Yip et al., 2019), and mobile robotics (Lee, 2020). These disciplines have high requirements for learners’ hands-on engagement, and AR can offer a virtual learning environment that can foster learners’ creative and aesthetic skills. However, the application of AR in the field of language learning has been found to have a poor effect, which may be attributed to two factors. First, there are few relevant studies on the application of AR in the field of language learning, which results in a potential bias in the results. Second, there may also be a mismatch between the technical characteristics of AR and teaching tasks for language instruction. Language learning typically emphasizes auditory information and conversational practice, whereas AR primarily provides visual stimuli. Consequently, in language learning settings, teachers and students often prefer traditional methods.

Regarding technical features, there is little difference in the inter-group effect size of AR for the input and output methods, which indicates the universality and supportability of AR characteristics for higher education. AR’s support for learning appears, first, in its ability to integrate with the real-world environment, and second, in its capability to provide an immersive experience in dangerous, visually challenging, or abstract scenarios. During instruction, high-performance and portable devices can be chosen based on specific needs, thus enabling the standardization and sustainable development of AR teaching applications.

In terms of learning outcomes, AR has the potential to enhance learners’ knowledge, emotions, and skills; however, the degree of improvement in capacity remains limited. Here, “capacity” refers to learners’ spatial thinking ability, computational thinking ability, and other macro-literacies. There are two possible reasons for the small effect size of learning outcomes on capacity. First, there is a scarcity of research on the assessment of learners’ capacity, which makes it challenging to conduct measurements and evaluations, leading to potential biases in the results. Second, the enhancement of capacity is a long-term process, and it is difficult to achieve significant improvement in learners’ abilities through short-term AR-based instruction.

5.3. Implication

Based on the results of this systematic review and meta-analysis, this article proposes the following suggestions for the application of AR in higher education. First, an appropriate pedagogy should be selected to maximize the effectiveness of AR-based instruction. Instructional designers should not be limited to a single pedagogy, but should flexibly adopt a variety of pedagogies, such as collaborative and game-based learning.

Second, AR usage should fully consider the characteristics of the different disciplines. Practical disciplines such as medicine have better prospects for AR applications, while it is less common in fields such as mathematics and language. Thus, teachers should explore effective teaching models that integrate AR with disciplines based on discipline characteristics. Researchers and educators should address questions regarding technical compatibility, appropriate educational situations, and curriculum implementation paths.

Third, the evaluation of AR teaching effectiveness should focus on diversification, normalization, and sustainable tracking of the learning process. Researchers should incorporate the evaluation of phased learning tasks, move beyond quantitative data analysis, and utilize qualitative data as indicators for evaluation. Educators should also consider how to promote the deep integration of AR and teaching content, thereby enhancing learners’ comprehensive abilities including information literacy, spatial thinking, and computational thinking.

Finally, researchers should standardize the data-reporting format for AR education experimental research and improve data usability. The complete experimental data provides readers and future researchers with a significant reference material. This article identified the need for improved completeness in the reporting of experimental research data, particularly regarding the omission of fundamental details, such as effect and sample sizes. In future studies, researchers should appropriately report the complete data required for meta-analysis, thereby improving the academic standardization of articles while facilitating the calculation of comprehensive effects.

6. Limitations

There are several limitations in the included literature. First, the included studies predominantly focus on undergraduate students, with limited coverage of other populations such as graduate students. Second, the included articles lack sufficient exploration of certain instructional approaches, such as collaborative learning. Lastly, the review primarily incorporates studies reporting positive outcomes of AR, with relatively few studies addressing negative or null results.

Also, there are limitations in the review process. First, the literature search was conducted in only two academic databases, the possible omission of relevant studies can undermine the comprehensiveness of the review results. Second, the subjective coding process could introduce errors caused by confirmation bias and coding drift, leading to an incomplete or less accurate interpretation and synthesis of the results.