This systematic review and meta-analysis was conducted in line with the PRISMA (Preferred Reporting Items for Systematic Review and Meta-Analyses) guidelines [40].

We searched 4 electronic databases (PubMed, Scopus, PsycINFO, and Web of Science) from database inception until March 26, 2022, to identify relevant studies. A manual search was conducted of the retrieved publications to identify potentially missing studies. The search strategy consisted of 2 strings of keywords, including PA- and MPA-related terms. These included the following: (“cell phone” OR “cell phones” OR “cellular phone” OR “cellular phones” OR “cellular telephone” OR “cellular telephones” OR “mobile devices” OR “mobile phone” OR “smart phone” OR “smartphone”) AND (“addiction” OR “dependence” OR “dependency” OR “abuse” OR “addicted to” OR “overuse” OR “problem use” OR “compensatory use”) OR (“problematic smartphone use” OR “problematic smart phone use” OR “problematic mobile phone use” OR “problematic cell phone use” OR “problematic cellular phone use” OR “Nomophobia” OR “Phubbing” OR “fear of missing out” OR “FoMO” OR “smartphone separation anxiety” OR “smartphone use disorder” OR “compulsive mobile phone use”) AND (“physical activity” OR “walk*” OR “exercise*” OR “physical activity*” OR “strength training” OR “resistance training” OR “resistance exercise*” OR “conditioning muscle” OR “training” OR “leisure training” OR “leisure activities” OR “physical fitness” OR “motor activity”). The detailed search strategy is presented in Multimedia Appendix 1. We manually performed a complementary Google search using the abovementioned keyword combinations to broaden the results on September 20, 2022. Secondary searches were performed by manually screening reference lists of included studies and tracking cited articles to ensure no relevant study was omitted.

The identified and retrieved studies were imported into EndNote X7 software (Thompson Reuters). Duplicates were excluded using the deduplication function in Endnote. This screening and processing was conducted by 2 reviewers, who independently read the titles and abstracts and assessed the studies against predetermined inclusion criteria. The full text of the included studies was also independently examined by the 2 reviewers. Inclusion checklists were completed for each study, along with details on the decision to exclude. The reference list of each included study and the articles cited were thoroughly reviewed to ensure that no relevant studies were missed. At all stages, any discrepancies in the results obtained were resolved through consensus or by involving a third reviewer.

Case-control studies were excluded because they examined specific groups that were beyond the scope of this review. Furthermore, reviews, meta-analyses, commentaries, replies, clinical guidelines, conference abstracts, theses, and book chapters were also excluded.

A total of 892 studies were identified by reading the titles and abstracts. Among these, 46 candidate studies were identified after reading their full text. At this stage, 24 studies were excluded based on the above criteria. The remaining 22 studies were deemed eligible and included in the systematic review. The final meta-analysis included 17 of the 22 studies (Figure 1 shows the details of the article screening process).

Two reviewers independently extracted data from the included articles and entered the data into a form tailored to the requirements of this review. The extracted data included (1) publication details (author, year, and country); (2) sample characteristics (sample size, sex of participants, and type of participant); (3) time of data collection; (4) measurements of PA and MPA; and (5) the main study outcome (ie, correlation coefficient).

The Joanna Briggs Institute (JBI) appraisal checklist, which has 10 items, was used to examine the methodological quality of the included cross-sectional studies [42]. The studies were given a score of 0 to 2 for each item. Studies with an overall score higher than 70% were considered high quality with a low risk of bias. Details of the scoring criteria applied in the JBI appraisal checklist are presented in Multimedia Appendix 2.

All statistical analyses were conducted with Comprehensive Meta-Analysis software (version 3; Biostat Inc).

All data were extracted from the included studies. The pooled Pearson correlation coefficients (with the corresponding 95% CIs) between PA and MPA were calculated with the inverse variance method. Subsequently, the Pearson correlation coefficients were transformed to Fisher z scores before the pooled estimate was obtained to calculate variance-stabilized correlation coefficients, as described previously [43] The effect sizes were interpreted in line with recent suggestions concerning correlations for psychometrics with r: small (r=0.10-0.20), medium (r=0.21-0.35), and large (r>0.35) [44].

The Cochran Q test and the I² statistic were employed to measure heterogeneity across studies. The Cochran Q determines the conformity to the normal distribution of effect sizes. A significant value (P<.10) indicates heterogeneity. I² is an estimate of the ratio of true heterogeneity in the observed variation. I²>50% reflects statistically significant between-study heterogeneity [45]. For such studies, the random effects model was used to calculate the summary of the Pearson correlation coefficients with a P value <.05 or I²>50% [44]. Otherwise, the fixed effects model was used [44].

To determine potential moderators of heterogeneity, subgroup analyses were carried out for country or region, population (college students and adolescents), and time of data collection (before or during COVID-19). All subgroup analyses were conducted with a mixed effects analysis. The random effects model was used to summarize the studies within the respective subgroups, and the fixed effects model was used to test for significant differences between the subgroups [46]. Full details of coding forms for the subgroups are provided in Multimedia Appendix 3.

To determine the influence of individual studies on the summary correlation coefficients and test the robustness of the correlations between PA and MPA, sensitivity analyses were conducted by sequentially omitting one study at time [11].

Funnel plots were established to determine the existence of potential publication bias. Additionally, the Begg rank correlation test and Egger linear regression test were performed to determine publication bias, with P<.05 indicating significant publication bias [47,48]. In the case of publication bias, the trim-and-fill method was used to adjust for funnel plot asymmetry [11].

Other statistical analyses performed included valid measures of the association between PA and MPA, measured with the correlation coefficient (r), standardized regression coefficient (b), unstandardized regression coefficient (β), odds ratio (OR), mean, and SD. To include as many eligible studies as possible, several data transformation steps were used. For studies that reported the mean and SD, the Cohen d effect size was calculated and converted to a correlation [49]. Studies that reported relevant ORs with 95% CIs were converted to Cohen d effect estimates and then to correlations [50]. The authors of the eligible studies were contacted if potentially relevant data were missing.

Table 1 presents a summary of the characteristics of the included studies. Overall, 23,365 participants aged between 15 and 26 years were included. Eighteen studies (numbers 1 to 22) were included in the systematic review, and 17 studies (numbers 1 to 17) were included in the meta-analysis. Moreover, 17 studies reported a correlation between PA and MPA. Considering the high heterogeneity among studies (Q=468.050, P<.001; I²=96.582), the random effects model was used to estimate the effect size of summary r (r=–0.243; 95% CI –0.309 to –0.175; P<.001; Table 2 and Figure 2). This result showed that PA was moderately negatively correlated with MPA.

As shown in Table 3, the summary correlation coefficient between PA and MPA did not change when stratified by time of data collection, country or region, or type of population (all P^b>.05). However, to allow comparison with other studies, we present model-implied effect sizes for each level of the moderator.

The time of data collection did not significantly moderate the effect sizes (between-subgroup P^b=.14). Notably, the summary correlation coefficient of the studies reporting on data collected before COVID-19 was slightly higher compared with that for data collected during COVID-19. Specifically, the effect sizes for data collection before COVID-19 were moderate, with a 95% CI that did not overlap with 0 (r=–0.333, 95% CI –0.466 to –0.187; k=4; P^a<.001), whereas the effect sizes for data collected during COVID-19 were also moderate, with 95% CIs that overlapped with 0 (r=–0.207, 95% CI –0.285 to –0.126, k=13; P^a<.001).

Similarly, we did not find significant moderator effect sizes for country or region (between-subgroup P^b=.71). The summary correlation coefficient for both China and other developing regions showed a similarly moderate effect size (China: r=–0.201, 95% CI –0.311 to –0.127; k=7; P^a<.001; other developing regions: r= –0.217, 95% CI –0.326 to –0.103; k=8; P^a<.001). However, the effect sizes for developed regions with a 95% CI that overlapped with 0 (r=–0.446, 95% CI –0.616 to 236, k=2; P^a=.24) were not significant.

In addition, there were no significant moderator effect sizes for the type of population (between-subgroup: P^b=.26). Specifically, we found that the effect sizes for young adults were moderate, with a 95% CI that did not overlap with 0 (r=0.250, 95% CI –0.325 to –0.173, k=15; P^a<.001). However, the effect sizes for adolescents were not significant, with a 95% CI that overlapped with 0 (r=–0.129, 95% CI –0.333 to 0.086, k=2, P^a=.24).

In the analysis that removed studies one at a time, no evident outliers were identified. Thus, the correlation coefficient for removing each study was in the range of r=–0.195 to –0.248. This shows that no one study significantly skewed or changed the correlation coefficient or influenced the overall results of the meta-analysis. Therefore, the results were reliable.

Subjectively speaking, we could not determine the existence of publication bias from the funnel plots for the summary correlation coefficients, as shown in Figure 3A. Studies with a small sample size are unlikely to result in symmetrical distributions of scattered points. The Begg rank correlation tests and Egger linear regression tests showed no significant publication bias (P=.62 and P=.14, respectively). After the trim-and-fill analysis, the correlation between PA and MPA remained statistically significant (number to trim and fill 6, summary r=–0.319, 95% CI –0.40552 to –0.22771; Figure 3B). Therefore, the overall modeling results after correction remained unchanged. Thus, we concluded that there was no publication bias.

The methodological quality assessment results are shown in Table 4. Notably, the mean scores of the included studies and the other studies were 15.29 (SD 1.53) and 15.40 (SD 0.89), respectively. All included studies were of high quality with a low risk of bias.

To the best of our knowledge, this is the first meta-analysis to explore pooled correlation coefficients of PA and MPA. Our analysis of 17 studies found a moderately negative correlation between PA and MPA, with a summary Pearson correlation coefficient of r=–0.243. This is in line with a previous review [40]. Sensitivity analyses did not find significant publication bias, indicating that the pooled analyses of the correlation coefficients provided reliable and convincing results. In addition, all the included studies were high quality with a low risk of bias. Subgroup analysis showed that none of the hypothesized moderators (data collection, country or region, and type of population) significantly influenced the relationship between PA and MPA, as confirmed with a mixed effects analysis.

The target subjects of research into MPA are adolescents and young adults, who are relatively less self-disciplined in controlling their frequency of mobile phone use and are more susceptible to smartphone use addiction compared to middle-aged or older adults [29]. Lack of self-control is an essential factor influencing MPA among adolescents and young adults. Previous studies found that self-control regulates the correlation between PA and MPA. Factors such as negative emotions (eg, anxiety [69,70] and loneliness [71,72]) and mental toughness [73,74] have been shown to affect the relationship between PA and MPA. We speculate that these factors may modulate the relationship between PA and MPA.

Results of magnetic resonance imaging studies suggest that MPA is associated with structural brain abnormalities, like other types of addiction dependence. For example, the insula cortex participates in the formation of addictive behaviors, because these behaviors may influence the decision-making process in terms of choosing immediate rewards that are always associated with physiological state while eliciting strong interoceptive signals [75]. Two recent studies reported changes in gray matter volume in this region (ie, the insula cortex) among people with MPA [66,76]. Exercise has been shown to improve brain health [77,78]. Therefore, we hypothesize that the relationship between PA and MPA might be influenced by the structure and function of the insula and even other brain regions.

Notably, the time of data collection did not significantly influence the relationship between PA and MPA. Moreover, a moderate negative relationship was found between PA and MPA among adolescents and young adults before and during the COVID-19 pandemic. According to the compensatory internet use theory, when people encounter psychosocial problems in the real world, they are likely to use the internet or smartphones as a coping mechanism to alleviate negative emotions [79]. The restrictions imposed on participation in social activities and gatherings during the COVID-19 pandemic increased anxiety, depression, and stress levels in people [80]. Therefore, they were more likely to overly rely on their mobile phones to cope with stress. Moreover, recent studies have shown that adolescents and young adults have had low PA levels during the COVID-19 pandemic [81,82]. The target subjects in our study were adolescents and young adults, who are more inclined to use social media for physical training. A previous study found that young-adult Spanish university students used social media apps to improve their high-intensity interval training [82], mind-body activities, and strength exercises. In other words, adolescents and young adults used social media to facilitate their participation in PA during the pandemic [82]. This reduced the time spent sitting for long periods of time and reduced leisure-time screen activities, subsequently reducing the risk of MPA in young people [30].

The present findings demonstrate that country or region do not have a significant moderating role on the relationship between PA and MPA. Notably, a medium-strength negative relationship between PA and MPA has been reported in China and other developing counties among adolescents and young adults. However, this correlation was not found in developed countries. This finding should be interpreted with caution, because it is based on 2 studies from developed countries. These 2 studies were carefully reviewed elsewhere [30,34]. In addition, we found that one of these studies reported a weak negative correlation between PA and MPA [34], whereas the other found a significant, strong negative correlation [30]. It should be noted that the 2 studies were published around the same time. The difference in the correlation results may be due to the type of PA measurement tools used. For instance, one of the studies used a pedometer sensor to measure the level of PA, which is more precise [30]. The influence of measurement errors associated with self-reported PA questionnaires also needs to be acknowledged. The majority of the reviewed studies used self-reported scales or questionnaires; thus, we suggest that accelerometry should be adopted in future studies to obtain more reliable data.

Further analysis revealed that population type did not significantly affect the relationship between PA and MPA. This may be explained by the widespread use of mobile phones. This is especially true for young people, as their ownership rate for smartphones is very high. Additionally, subgroup analysis revealed that there was no significant correlation between PA and MPA among adolescents (P=.26). We presume that this might be influenced by the degree of external restrictions on the use of mobile phones. Compared with adults (eg, college students), adolescents are subjected to more control and restriction measures on mobile phone use by their parents, schools, and even by the mobile phone apps themselves. Therefore, they are less likely to influence the correlation [83]. These findings, however, should be interpreted with caution, because only 2 studies on adolescents were analyzed.

In conclusion, our study indicates that a low PA level contributes to MPA behavior. This is because low PA encourages a sedentary lifestyle among young adults. The PA guidelines of the World Health Organization encourage individuals of different ages to participate in PA. Previous studies have shown that increasing the PA level of young adults can reduce MPA behavior. We recommend higher PA levels than those stipulated in the guidelines of the World Health Organization, because more PA could bring more mental health benefits. From a practical perspective, the findings of this study may help to inform countermeasures to prevent MPA behavior among adolescents and young adults amid the COVID-19 pandemic and future public health crises.

All previous findings were objectively stated, analyzed, and interpreted using an appropriate research design. All original data were retained to provide a reference for future research. The repeatability and reproducibility of our analyses have been ensured. However, there are some limitations and potential sources of bias that need to be noted. First, only studies published in English were included in our meta-analysis. Second, the studies mainly provided cross-sectional data, which do not allow determination of causality in the relationship between PA and MPA. Third, we only analyzed a young population. Fourth, no study reported moderating variables between PA and MPA. Finally, although a sensitivity analysis was conducted, sources of bias were identified, and our results should thus be interpreted with caution. Further case-control and cohort studies are needed to test the benefits of PA on MPA in young adults.

Our findings demonstrate a moderate negative relationship between PA and MPA among young adults. The strength of the relationship between PA and MPA did not differ by time of data collection, country or region, or type of population.