The valence framework

Originally developed from the economics and psychology literature, the valence framework explained individuals’ behaviors that incorporated both the positive and negative utilities of their decision [10]. According to the valence framework, people act to minimize expected negative utility and maximize expected positive utility associated with a particular decision-making behavior [10, 11].

The valence framework has been applied in various research environments including online shopping [13], mobile payment [11], and sharing economy [14]. For instance, Kim, Ferrin [13] applied the valence framework in the context of online shopping, and found that perceived benefits as the positive utility positively affects willingness to purchase, while perceived risk as the negative utility negatively influences willingness to purchase. Yang, Lu [11] extended the valence framework to the mobile Internet environment to explain mobile payment continuance. They found that compatibility and relative advantage affect continuance intention, while perceived risk negatively influences continuance intention. More recently, Lee, Chan [14] found that perceived benefits, perceived risks, trust in the platform significantly affect Uber users participate intention in the sharing economy context. Based on the extant studies, the present research applied the valence framework in the mobile social network context to explore the impacts of both positive and negative utilities on mobile GSM continuance.

The positive utilities

According to the original valence framework, perceived benefits reflects the positive utility of users’ decision to use a particular technology. In the context of the present study, we should capture the positive unities based on the mobile GSM features, such as social value and hedonic value. Social value has long been regarded as the important factor that affects social media usage [9]. In the present study, when citizens can easily interact with government agencies or other citizens via the mobile GSM platform, they will be more likely to develop positive utility evaluations on the mobile GSM. Previous studies also found that social value for maintaining social relationships positively affect users’ continuance intention [9]. Therefore, we can propose the following hypothesis:

H1a: Social value will positively affect citizens’ mobile GSM continuance intention.

The impacts of hedonic value on citizens’ continuance intention have also validated by the extant studies [5, 15]. For instance, Ding, shuiqing [5] examined the impacts of citizens’ gratifications on mobile government social media participation behaviors. They found that hedonic value has positive impact on participation behaviors. In the present study, it is expected that citizens will tend to continue using mobile GSM if they have experienced a higher hedonic value. Therefore, following on the extant studies [5, 15], we can propose the following hypothesis:

H1b: Hedonic value will positively affect citizens’ mobile GSM continuance intention.

The negative utility

When using mobile GSM, citizens have to bear the privacy concerns related to their usage decision [16]. In the original valence framework, the negative utility is mainly reflected by risk associated with the usage decision. In the context of mobile government social media, privacy concerns occurred when citizens using mobile GSM services, which was not considered in the original valence framework. The significant negative impacts of privacy concerns on usage behaviors have been validated in the various environment including virtual health community’s usage [17], smart services resistance [18], and retail technologies usage [19]. Therefore, based on the extant studies [17, 19], we can propose the following hypothesis:

H2a: Privacy concerns will negatively affect citizens’ mobile GSM continuance intention.

In the present study, self-censorship will also have negative impact on mobile GSM continuance intention. Self-censorship refers to an individual suppresses her or his own attitudes where a pubic censor is either irrelevant or absent [20]. In the social media environment, self-censorship reflected the fact that users filter their own thoughts or expression before shared them to the social media platform [21]. In the context of the present study, when citizens experience a high level of self-censorship, they will be more likely to discontinue use of mobile GSM. Therefore, based on the extant studies [20, 21], we can propose that:

H2b: Self-censorship will negatively affect citizens’ mobile GSM continuance intention.

Instrument

All constructs in the present study were borrowed from the existing well-established studies to ensure content validity. The items of privacy concern were adapted from James, Wallace [16]. Three items of self-censorship were adapted from Ranzini and Hoek [22]. The items of social value were borrowed from Zhang, Li [23] and Guo, Liu [24]. Three items of hedonic value were adapted from Zhang, Li [23]. The items of continuance intention were borrowed from Bhattacherjee [25]. A seven-point Likert scale was used to measure the instrument, with response choices ranging from (1 = strongly disagree) to (7 = strongly agree). A back-translation procedure was performed to ensure the translation validity from the original English items to the Chinese scales. A pilot test was also conducted on 36 mobile GSM users to further examine the readability and clarity of the instrument. The final questionnaire is displayed in the S1 Table.

Data collection

The present study conducted an online survey on a well-known web-based survey platform (WJX.CN) to collect empirical data. The sample pool of WJX.CN has 2.6 million registered members from different cities. The sample services offered by this online survey site were adopted for our sample collection. Due to the objective of our study is to investigate citizens’ mobile GSM usage behaviors, the participants should have certain usage experience of mobile GSM. Therefore, the filter conditions in our survey project are limited to those who had mobile microblogging services usage experience in Sina microblog. WJX.CN then randomly invited eligible respondents to participate in the web-based survey. Participants in this study have been properly instructed and are voluntarily participating in our survey. Therefore, the participant’s consent has been obtained. In addition, the data were analyzed anonymously. The data collection process lasted for three weeks (from September 20, 2019 to October 12, 2019). All responses were scrutinized and discarded whose who given the same answer for all questions, and whose who did not have usage experiences of mobile GSM. Finally, 509 valid response were obtained. The demographic statistics of the sample are listed in Table 1.

Table 1

Sample demographics.

Measure	Item	Number(N = 509)	percentage (%)
Gender	Male	253	49.7%
	Female	256	50.3%
Age	<18 years	3	0.6%
	18~25 years	78	15.9%
	26~35 years	320	62.9%
	>35 years	108	21.2%
Education level	High school or below	12	2.4%
	Two-year college	47	9.2%
	Four-year college	407	80%
	Graduate school or above	43	8.4%
Occupation	Student	22	4.3%
	General Employee	308	60.5%
	Manager	154	30.3%
	Self-employed	13	2.6%
	Others	12	2.4%
Mobile GSM experience	<1 year	45	8.9%
	1~3years	246	48.3%
	3~5 years	162	31.8%
	>5 years	56	11%
Main followed GSM	Hangzhou Releases	97	19.1%
	Shanghai Releases	117	23%
	Beijing Releases	177	34.8%
	Guangzhou Releases	65	12.8%
	Others	53	10.3%

Reliability and validity

The reliability and validity of the scales were tested by conducted a confirmatory factor analysis. As shown in Table 2, for each construct, the coefficient of Cronbach’s alpha and composite reliability (CR) were all higher than the recommended value of 0.7, showing good reliability of the constructs. The loadings of the items are all above the recommended level of 0.7, and the constructs’ average variance extracted (AVE) values were all over the recommended benchmark of 0.5, indicating good convergent validity of the constructs.

Table 2

Scale properties.

Variable	Item	Standard Loading	Cronbach’s Alpha	CR	AVE
Social value (SOV)	SOV1	0.805	0.810	0.887	0.725
	SOV2	0.883
	SOV3	0.863
Hedonic value (HEV)	HEV1	0.866	0.827	0.896	0.742
	HEV2	0.852
	HEV3	0.865
Privacy concerns (PCO)	PCO1	0.850	0.883	0.928	0.811
	PCO2	0.939
	PCO3	0.909
Self-censorship (SCE)	SCE1	0.861	0.794	0.877	0.705
	SCE2	0.767
	SCE3	0.885
Continuance Intention (CON)	CON1	0.842	0.824	0.895	0.740
	CON2	0.844
	CON3	0.852

Note: PCO = Privacy concerns; SCE = self-censorship; SOV = social value; HEV = hedonic value; CON = Continuance intention.

The discriminant validity was then tested by compared each construct’s inter-construct correlation and its square root of the AVEs. Following the procedure conducted by [26], we adopted both Fornell-larcker Criterion and Heterotrait-Monotrait Ratio (HTMT) for testing discriminant validity. As displayed in Table 3, for all constructs, the square root of the AVEs were significantly higher than their corresponding inter-construct correlation coefficients, indicting good discriminant validity [27]. To further assess the validity of construct, the cross-loading matrix was listed in S2 Table. The internal-construct loading of each extracted factor was higher than its cross-loading on other factors, displaying a clear loading matrix.

Table 3

Fornell-larcker criterion and Heterotrait-Monotrait Ratio (HTMT).

		SOV	SCE	HEV	PCO	CON
Fornell-larcker Criterion	SOV	0.851
	SCE	-0.091	0.839
	HEV	0.035	-0.128	0.861
	PCO	-0.077	0.288	-0.110	0.900
	CON	0.381	-0.232	0.647	-0.171	0.860
Heterotrait-Monotrait Ratio (HTMT)	SOV
	SCE	0.093
	HEV	0.686	0.099
	PCO	0.160	0.536	0.261
	CON	0.651	0.178	0.767	0.253

*Diagonal elements are the square root of AVE. These values should exceed the Inter-Construct Correlations for adequate discriminant validity.

Note: PCO = Privacy concerns; SCE = self-censorship; SOV = social value; HEV = hedonic value; CON = Continuance intention.

The present study further tested the potential common method bias in our research. First, a Harman’s one-factor test was conducted. The results shown that the five factors were exerted and the largest covariance explained by the single factor was not more than 50%, suggesting that common method bias was not a serious concern in the present study. Second, the present study built a new measurement model that included an additional common method factor and compared it with the original model [24]. The results shown that the variable loading in the original model were all significant at the P<0.001 level, whereas the loadings of the common method factor in the new developed model were not significant. This again indicated that common method bias was unlikely a concern in our study.

Hypothesis testing

The proposed research model and its corresponding hypotheses were tested by using structure equation modeling (SEM) software LISREL 8.72. Table 4 listed the recommended and actual values of the model fit indices. As shown in the Table, the actual values of the model fit indices were all better than the recommended values, showing a good fitness between our research model and the data [27].

Table 4

Fit induces and recommended values (N = 509).

Fit index	x2/df	RMSEA	GFI	AGFI	CFI	NFI
Recommended value	<5	<0.08	>0.90	>0.80	>0.90	>0.90
Model value	1.98	0.044	0.961	0.940	0.988	0.976

Notes: RMSEA, root mean square error of approximation; GFI, goodness of fit index; AGFI, adjusted goodness of fit index; CFI, comparative fit index; NFI, normed fit index.

The results of model testing are displayed in the Fig 2, which shown that most proposed hypotheses are validated by the empirical data. The hypothesized paths from positive utilities factors including social value and hedonic value to mobile GSM continuance were significant at P<0.001, thus validating hypothesis H1 and hypothesis H2. The impact of self-censorship on mobile GSM continuance was negative significant, thus supporting hypothesis H4. However, the influence of privacy concerns on mobile GSM continuance were not significant, thus hypothesis H3 were not supported. The R2 for mobile GSM continuance intention was 0.641, which indicated a reasonable explanation of the variance for the dependent variable.

Fig 2

Test results of the research model.

* p<0.05; **p<0.01; ***p<0.001.

Neural network analysis

Following the procedure conducted by extant studies [26, 28, 29], the present study adopted a SEM-NN approach to test our proposed model. Specifically, the multilayer perceptron training algorithm was employed to train the neural network, which was calculated using SPSS 24.0 [30]. The significant predictors calculated by SEM were used as the input layer to build the neural network model. And the output of the neural network is mobile GSM continuance intention.

Following the procedure performed by Zhang, Chen, et al. [30], our study calculates the neural network using one to ten hidden nodes. To predict the accuracy of the NN model, the present study conducted a ten-fold cross-validation analysis [5, 30]. The Root Mean Square Error (RMSE) were calculated to test the prediction accuracy of the NN model. The two models’ average cross-validated RMSE were very small (training data: from 0.170 to 0.176; testing data: from 0.172 to 0.179). This indicated that the neural network model of our study is reliable and accurate prediction [5, 30].

The present study then conducted the sensitivity analyses [28, 29]. The importance of each predictor is listed in Table 5. hedonic value is the strongest factors that affect mobile GSM continuance, following by social value, and self-censorship.

Table 5

Sensitivity analysis.

	Output: CON
	SCE	SOV	HEV
1	0.245	0.344	0.411
2	0.251	0.314	0.435
3	0.259	0.368	0.374
4	0.224	0.400	0.375
5	0.202	0.349	0.415
6	0.245	0.340	0.571
7	0.230	0.291	0.479
8	0.234	0.433	0.333
9	0.203	0.398	0.400
10	0.190	0.374	0.436
Average	0.228	0.361	0.423
Importance	52.777	85.342	100

Note: SCE = self-censorship; SOV = social value; HEV = hedonic value; CON = Continuance intention.

Theoretical implications

The present study has several theoretical implications. First of all, many previous studies tend to focus on the technology performance-based enablers to explain mobile social media usage behaviors by extending the leading TAM. In the context of the present study, users of mobile GSM have dual roles included technology user and service consumer. The leading TAM and its extending models unable to explain the underlying mechanism of why and how citizens use mobile GSM to satisfy their various needs. Therefore, the present study applied the valence framework in the mobile GSM context to explain why and how citizens decide to continue using of mobile GSM.

Secondly, unlike the existing studies mainly focused on the benefits aspect to explain mobile social media usage, the present study explored the impact of both positive and negative utilities on citizens’ mobile GSM continuance. Specifically, the positive utilities included hedonic value and social value positively affect citizens’ mobile GSM continuance, while the negative utilities such as self-censorship negatively influence citizens’ mobile GSM continuance. The present study thus contributes to the existing literature by exploring how citizens’ positive and negative utilities might influence their continuance using of mobile GSM.

Finally, unlike the extant studies tend to examine mobile social media usage by adopting SEM approach, our study utilized a SEM-neural network method to explore the factors that affect mobile GSM continuance. The SEM-neural network method can take the advantages of both the theoretical validating-based SEM and non-linear relationships-centered neural network approaches. The present study thus can validate the results by employing multiple methods and find out the most significant factors that determine mobile GSM continuance.

Practical implications

The present study also has several practical contributions. First, mobile GSM managers should pay close attention to citizens’ positive utilities because our study found that citizens’ social value and hedonic value positively affect their mobile GSM. In terms of hedonic value, mobile GSM managers should provide citizens with pleasure and sense of fun during the process of their mobile GSM usage. For instance, mobile GSM managers can provide personalized responses to citizens based on their preferences and specific contexts by employed big data and artificial intelligence technologies. In terms of citizens’ social value, mobile GSM managers should maintain good relationships with citizens by promptly responding to the comments or suggestions provided by their citizens in mobile GSM platform.

Second, mobile GSM managers should also pay close attention to citizens’ negative utilities [31]. For instance, our study found that self-censorship negative influence their mobile GSM continuance. Mobile GSM managers can reduce citizens’ self-censorship by clearly stating which content or information should not be published online. Such clear explanation will help citizens to reduce their uncertainty perceptions and thus decreasing their self-censorship. For instance, government agencies and mobile GSM managers can classify citizens’ comments and discussions into different topics on the mobile GSM platform to manage them separately and identically. On the other hand, even though, the impact of privacy concerns on mobile GSM continuance is not significant. Mobile GSM managers should also take measures to reduce citizens’ privacy concerns during their usage process.

Finally, government agencies and mobile GSM managers should pay close attention to both positive and negative utilities citizen perceived in the process of their mobile GSM usage. Nowadays, social media regulation has been adopted by many countries and have receiving worldwide attention [2]. Government agencies should take measures to increase citizens’ positive utilities, and decrease their negative utilities. To minimize the negative impact, government agencies and GSM system designers should develop new algorithms based on the big data technology to protect citizens’ privacy. More specifically, the keyword filtering algorithms should be improved to ensure the normal posts and discussion. For instance, the keywork filtering algorithms should include the tendency evaluation based on a particular user’s historical content or logs to minimize the negative influence of self-censorship and perceived privacy.