In this study, a 2-step procedure was conducted: (1) data collection and (2) data analysis. In order to collect metadata of smartphone usage automatically, we developed an Android app named MobileSens [50,51]. MobileSens consists of 2 modules, one is for collecting metadata of smartphone usage, and the other is for filling the questionnaire, which can get smartphone usage data and the corresponding questionnaire results at the same time.

Descriptive statistics were computed as the mean and SD for continuous variables and absolute frequencies and percentage for categorical variables. The chi-square test and independent t-test were used to compare demographics between 2 groups. The 2-way ANOVA of depressive symptom (nondepressed group vs depressed group) × gender (male vs female) was conducted to examine the differences in social behaviors. Multiple logistic regression analysis with dichotomous depressive symptoms (nondepressed vs depressed) as dependent variables was used to find predictors of depressive symptoms. In this analysis, variable selection was performed using stepwise forward selection, subsequently including one by one the variables that were not statistically significant (α=.05). To improve the chances of retaining meaningful predictor variables, demographics that differed between the groups at P<.1, and behavior features that differed in the main effect of depressive symptoms in 2-way ANOVA at P<.1 were analyzed as independent variables via multiple logistic regression analysis [61]. For further data exploration of gender differences, multiple logistic regression models adjusted for gender were conducted. All statistical tests were performed using SPSS version 23 for Windows (IBM). The level of significance was set at .05.

Of the 120 participants who formed the study sample, 71 participants were divided into the nondepressed group and 49 participants were divided into the depressed group. Demographics of participants are summarized in Table 2. The results of the chi-square test and independent t-test on demographics between the nondepressed and depressed groups showed no significance (gender: P=.76; education: P=.34; living place: P=.99; occupation: P=.75; marital status: P=.36; and age: P=.31).

To examine differences in social behavior features between the nondepressed and depressed groups, we conducted 2-way ANOVA of depressive symptom (nondepressed group vs depressed group) × gender (male vs female). All social behavior features with significant results are shown in Table 3.

For “usage of traditional communication functions,” 2 behavioral features had significant differences. Specifically, in terms of “receiving calls from contacts all day,” the main effect of depressive symptom was significant (F_1,116=3.995, P=.048,η²=0.033). The frequency of receiving calls from contacts all day was significantly higher in the nondepressed group than in the depressed group (P=.048). By contrast, the main effect of gender and the interaction effect of depressive symptom and gender were not significant (F_1,116=0.005, P=.94, η²=0.000; F_1,116=0.010, P=.92, η²=0.000). As for “receiving calls from contacts in the afternoon,” the main effect of depressive symptom was significant (F_1,116=5.278, P=.02, η²=0.044). The frequency of receiving calls from contacts in the afternoon was significantly higher in the nondepressed group than in the depressed group (P=.02). However, the main effect of gender and the interaction effect of depressive symptom and gender were not significant (F_1,116=0.004, P=.95, η²=0.000; F_1,116=0.006, P=.94, η²=0.000).

In “usage of social apps,” 6 behavioral features had significant results. Specifically, in terms of “usage of social apps all day,” the main effect of depressive symptom was significant (F_1,116=6.801, P=.01, η²=0.055). The frequency of usage of social apps was significantly higher in the depressed group than in the nondepressed group (P=.01). By contrast, the main effect of gender and the interaction effect of depressive symptom and gender were not significant (F_1,116=0.675, P=.41, η²=0.006; F_1,116=1.654, P=.20, η²=0.014). As for “usage of social apps in the evening,” the main effect of depressive symptom was significant (F_1,116=6.902, P=.01, η²=0.056). People in the depressed group used social apps more frequently than people in the nondepressed group in the evening. However, the main effect of gender and the interaction effect of depressive symptom and gender were not significant (F_1,116=0.095, P=.76, η²=0.001; F_1,116=1.964, P=.16, η²=0.017).

In addition, Weibo was the only app with significant results among the 2 popular social apps. Under “usage of Weibo all day,” the main effect of depressive symptom, the main effect of gender, and the interaction effect of depressive symptom and gender were significant (F_1,116=6.689, P=.01, η²=0.055; F_1,116=6.642, P=.01, η²=0.054; F_1,116=7.118, P=.009, η²=0.058, respectively). Simple effect analyses showed that in the depressed group, females used Weibo more frequently all day than males (F_1,116=11.744, P=.001, η²=0.092; Figure 1). However, in the nondepressed group, there were no significant differences (F_1,116=0.005, P=.94, η²=0.000). As for “usage of Weibo in the morning,” the main effect of depressive symptom, the main effect of gender, and the interaction effect of depressive symptom and gender were significant (F_1,116=8.838, P=.004, η²=0.071; F_1,116=4.765, P=.03, η²=0.039; F_1,116=5.932, P=.02, η²=0.049, respectively). Simple effect analyses showed that in the depressed group, the frequency of usage of Weibo among females in the morning was significantly higher than that of males (F_1,116=9.105, P=.003, η²=0.073; Figure 1). However, in the nondepressed group, no significant gender differences existed (F_1,116=0.039, P=.85, η²=0.000).

Concerning “usage of Weibo in the afternoon,” the main effect of depressive symptom, the main effect of gender, and the interaction effect of depressive symptom and gender were significant (F_1,116=6.370, P=.01, η²=0.052; F_1,116=8.918, P=.003, η²=0.071; F_1,116=7.764, P=.006, η²=0.063, respectively). Simple effect analyses showed that in the depressed group, females used Weibo more frequently in the afternoon than males (F_1,116=14.224, P<.001, η²=0.109; Figure 1). However, in the nondepressed group, no significant gender differences existed (F_1,116=0.024, P=.88, η²=0.000). In terms of “usage of Weibo in the evening,” the main effect of gender and the interaction effect of depressive symptom and gender were not significant (F_1,116=4.531, P=.04, η²=0.038; F_1,116=4.889, P=.03, η²=0.040; F_1,116=5.731, P=.02, η²=0.047, respectively). Simple effect analyses showed that in the depressed group, females used Weibo more frequently than males in the evening (F_1,116=9.052, P=.003, η²=0.072; Figure 1). However, in the nondepressed group, no significant gender differences existed (F_1,116=.020, P=.89, η²=0.000).

The following 11 variables were entered into the multiple logistic regression model: “receiving calls from contacts all day,” “receiving calls from contacts in the afternoon,” “receiving calls from contacts in the evening,” “usage of social apps all day,” “usage of social apps in the morning,” “usage of social apps in the evening,” “usage of Weibo all day,” “usage of Weibo in the morning,” “usage of Weibo in the afternoon,” “usage of Weibo in the evening,” and “usage of WeChat in the evening.” Of the 11 variables, “usage of social apps in the evening” remained a significant independent predictor (P=.02) of depressive symptoms in the model established on the whole sample (Table 4). In addition, “usage of social apps in the evening” emerged as a significant factor for male (P=.01), and “usage of Weibo in the morning” emerged as a significant factor for female (P=.03).

Our results supported the hypothesis that nondepressed and depressed users have differences in social behaviors on smartphones, and there were gender differences in social behavior among depressed users. Specifically, we used MobileSens to obtain metadata of smartphone usage on 120 participants, and extracted social behavior features to investigate the social behavior of nondepressed and depressed users. The results found depressed users received less calls from contacts (all day, in the afternoon) and used social apps more frequently (all day, in the evening) than nondepressed users. In the depressed group, females used Weibo more frequently than males (all day, in the morning, in the afternoon, in the evening).

The results show that depressed users receive fewer calls from contacts throughout the day than nondepressed users. Consistent with our results, a cross-sectional study involving 6105 adults aged 18-84 years found that low frequency of contacts with friends and parents was a significant factor associated with depression [62]. In particular, social interactions with strong ties (eg, family, friends) play an important role in satisfying people’s need for social connectedness [63-66]. Furthermore, meta-analyses have also shown that interventions for depression by addressing social relationships, including couples therapy and peer support, may be effective in reducing depressive symptoms [67,68]. Therefore, it is plausible that social interactions with strong ties via calls can satisfy psychological needs essential for reduction of depression and human flourishing. In addition, the result can be supplemented by one of the major assumptions of the behavioral theory of depression, namely, that a low rate of response-contingent positive reinforcement acts as an eliciting stimulus for depression [69-71]. Response-contingent positive reinforcement is defined as pleasurable or positive outcomes that follow an individual’s behavior within his or her environment and increase the likelihood of those behaviors. Under such an assumption, social interactions with friends through calls could be perceived as a positive response contingent, whereas lack of social interaction with friends or family will increase depression.

Notably, only in the afternoon did the 2 groups differ significantly in the frequency of calls received from contacts over all periods (P=.02). Previous studies have found a circadian rhythm of positive affect in healthy individuals, peaking in the afternoon [72,73]. Moreover, this positive affect was positively related to the number of interactions in which participants engaged, and the amount of time spent engaged in social contact [74]. According to the social enhancement model [75], those who already have good mental state enjoy more benefits from social interactions, while problems are only compounded for those who have psychological deficits. Thus, people who are already in good mental states have more willingness to engage in social interactions, which may encourage others to call them. However, unlike healthy individuals, the diurnal variation of positive affect in depressed patients showed an inverse-U shape with a steeper overall increase from morning lows to evening highs [76]. Thus, it is not surprising that difference exists in the number of receiving calls in the afternoon between the 2 groups.

We found that depressed users use social apps throughout the day more frequently than nondepressed ones. There might be several reasons. First, depressed people expect that interpersonal interaction via the internet can alleviate their psychological problems, because they believe that it is less risky and easier to get support and build relationships online than face to face [77,78]. Given that 99.1% of internet users are mobile internet users in China [41], the positive association between depression and internet overuse would be replicated for a smartphone. Second, deficiency in self-regulation is a common manifestation among those with depression [79], which makes it difficult for them to maintain a healthy level and amount of smartphone use. Therefore, depressed users use social apps more frequently than nondepressed ones.

As for “usage of social apps in the evening,” we found that it is a predictor for depression on both the overall data (all participants) and male data (male participants), and the usage frequency of depressed users is higher than nondepressed ones. Previous studies have found that depression symptoms tend to be relatively more active during the evening and night [59,80-82], which may exacerbate the usage of social apps. Additionally, people prefer to spend time on social media rather than face-to-face activities [83-85], but the quality of communication over social media has been highlighted as a potential limiting factor in building strong, emotionally intense relationships [86-88]. Furthermore, frequent smartphone usage can make people vulnerable to negative outcomes, such as interpersonal isolation [89]. Therefore, the usage of social apps not only fails to lessen their negative emotions, but also actually worsens these outcomes.

In the usage of top 2 social apps, at any time, the depressed ones used Weibo more frequently than the nondepressed ones. In the depressed group, females used Weibo more frequently than males. As we discussed above, the depressed ones use social apps more than the nondepressed ones, so this conclusion applies equally to Weibo. What deserves our attention is the gender differences in Weibo usage in the depressed group. Similar to Twitter, but unlike WeChat, Weibo focuses on sharing of opinions and information exchange rather than on social interaction [90], and offers some anonymity in online communication [91]. A previous study found that females are more likely to report Weibo usage for communicating with peers, passing time, and entertainment, whereas male users are more likely than females to report Weibo use for social compensation and social identity gratifications [92]. So depressed females are more likely to overuse Weibo than depressed males. Given a worsening of symptoms in the morning hours and early morning waking are characteristics of depression [58], depressed females may try to eliminate psychological problems by using Weibo. This could explain the usage of Weibo in the morning, which could serve as a predictor of depression in females.

One limitation of this study is that we only use questionnaire-based scales to measure depression. Although the validity of the questionnaires as a screening tool in accessing depression severity has been well proven in the literature, the questionnaire score itself cannot be used as diagnostic. Second, the sample in this study was composed of healthy individuals rather than clinical patients, which means that there were few “real” patients with depression disorders in our sample. So our results are applicable to individuals with minor depression, and their applicability to patients with confirmed moderate to severe depression has not been verified. Moreover, given the convenience sampling method of this study and the majority of participants were students, the extent to which sampling bias affected the results cannot be determined. Therefore, replication and extension of this research using larger, more representative samples are desirable in future work.

Although the results found that depressed users used social apps more frequently than nondepressed ones, the psychological mechanism behind it is still unclear. For instance, because deficiency in self-regulation is a common manifestation among those with depression [79], future work could explore the mediating effect of self-regulation in the relationship between depression and social behavior on smartphones. Additionally, our results showed that compared with nondepressed users, depressed users received fewer calls from acquaintances. Future studies may need to pay closer attention to whether it is the reason that depressed users use social apps frequently. Furthermore, our result found gender differences in some social behaviors among depressed users (eg, the usage of Weibo), which suggests that social-based features should be treated as a personalized feature that should be assessed in a within-subject analysis in future studies.

In summary, by analyzing metadata of smartphone usage acquired by MobileSens, we found that relationships exist between depression and social behavior on smartphones, which means we may be able to implement an early diagnosis system of depression through smartphone usage. Besides, the results of this study provide useful suggestions to those who are depressive in daily lives.