The Present Study

The available empirical evidence, which is still somewhat fragmented, suggests that smartphone addiction may be a potential risk for cyberfraud victimization. This potential risk derives from the double vulnerability associated with smartphone-addicted users (^{Herrero, Torres, Vivas, Hidalgo, et al., 2021}). On the one hand, the addictive and deregulated use of smartphones can expose users to potential cybercriminals and to situations where there is no capable guardian. On the other hand, smartphone addiction also negatively affects the presence of protective factors in the user with a potential effect on cyberfraud victimization, such as the lack of social support.

Most of the evidence available to date on the relationship between smartphone addiction, social support, and cybercrime victimization is correlational in nature, which certainly limits the generalizability of the results. In the present study, we incorporate the temporal dimension in the examination of these relationships. To verify the empirical plausibility of this set of theoretical relationships, we followed up on smartphone addiction, social support, and cybercrime victimization (specifically, cyberfraud) over time in a representative sample of Spanish smartphone users. We tested the extent to which variations in smartphone addiction and social support over time were associated with the risk of being victimized by cyberfraud.

The evolution of smartphone addiction and social support were analyzed using latent growth models (LGMs). LGM is a longitudinal analysis technique to estimate "growth" over time. The evolution of cyberfraud victimization over time was modeled through discrete-time survival curves, which allowed us to investigate the time it takes for an event of interest to occur (i.e., cyberfraud). Finally, the latent growth mixture modeling (LGMM) technique was used to evaluate the association of different trajectories of smartphone addiction and social support and different survival curves of cyberfraud victimization (see Figure 1). LGMM allows the researcher to explain between-subject heterogeneity in growth on an outcome by identifying latent classes with different growth trajectories and survival rates. With these three combined statistical techniques, we aimed to empirically identify population subgroups with different trajectories in smartphone addiction and social support as well as different survival rates of cyberfraud victimization. A representation of the analysis strategy is presented in Figure 1 and discussed in more detail in the method section.

Note. SS1-SS6 = social support from T1 to T6; SA4-SA6 = smartphone addiction from T4 to T6; CF1-CF6 = cyber-fraud from T1 to T6; Iss = intercept for social support; Sss = slope for social support; Isa = intercept for smartphone addiction; Ssa = slope for smartphone addiction; SR = survival rates for cyber fraud over time; C = number of distinct classes.

Figure 1.  Discrete-time Survival Mixture Growth Model: Social Support and Smartphone Addiction Trajectories and Different Survival Rates over Time. 

We expected that the levels of smartphone addiction and social support and their evolution over time would allow the identification of participants with different vulnerabilities to cyberfraud. Our starting hypothesis was that a downward trajectory of smartphone addiction and increasing levels of social support would be associated with a lower risk of cyberfraud victimization.

Participants

Data from the Cybersecurity and Confidence in Spanish Households National Survey (CCSHNS) conducted by the National Observatory of Telecommunications and Information Society were used for this study. The CCSHNS is a nationally representative survey of Spanish internet users on cybersecurity conducted every six months (see ^{Herrero, Torres, Vivas, Arenas, et al., 2021)}. The survey has a fixed set of questions that are evaluated every semester and other modules that may undergo modifications in the different waves. Data from six time panels obtained between the first half of 2017 and the second half of 2019 were used for this study. Each panel of the study was conducted on a representative sample of the population of internet users aged 18 to 75. For this study, 716 participants had complete data on the study variables during the 36-month follow-up.

Social support and cyberfraud victimization were measured in all six panels of the study. In panels T1 and T3, smartphone addiction was assessed in a small number of participants (277 and 288, respectively), and from panel T4 onward, this variable was regularly assessed in all study participants. The reason for this strategy is that at the beginning of the study, only a subset of participants was assessed for smartphone addiction to pilot the potential assessment of addiction in all participants in the following waves. Once the pilot study was completed, starting from the T4 panel, smartphone addiction was assessed in all users. As a result of this strategy, data were available from six panels on social support and cyberfraud victimization (T1-T6) and from three panels for smartphone addiction (T4-T6).

Variables and Scales

Smartphone addiction. The survey used eight items from the Smartphone Addition Symptoms Scale (SAPS; ^{Bian & Leung, 2015}). The items originally have 5 response categories (from 1 = not true to 5 = extremely true), but for the calculation of smartphone addiction, the authors of the scale recommend dichotomizing the response categories in the following way: 4 (true) or 5 (extremely true) = 1; all remaining response categories were set to zero. Items were summed. The descriptive and internal consistency statistics are presented in Table 1.

Table 1.  Descriptive Statistics and Internal Consistency of Variables of the Study in the Six Waves (N = 716)¹ 

Note. ¹For cyberfraud, percentages of participants having experienced cyberfraud at each time point are shown. For smartphone addiction and social support, means and standard deviation (in parenthesis) are displayed.

Social support. The strong-tie support scale (^{Lin, et al., 1981}) was used to measure social support from intimate and confidant relationships with three items on a five-point scale from 1 = never to 5 = most of the time. The scale measures to what extent respondents felt their support needs were fulfilled by close companions. The three items of the strong-tie support scale are a recommended measure of social support for large-scale surveys (^{Herrero, et al., 2011}). Social support was assessed in all panels from T1 to T6. The descriptive and internal consistency statistics are presented in Table 1.

Cyberfraud. The survey evaluated cyberfraud with a dichotomous response question (0 = no, 1 = yes) addressed to participants: "Have you suffered any financial loss in the last 6 months due to possible cyberfraud?". Cyberfraud was assessed in all panels from T1 to T6. The descriptive statistics are presented in Table 1. In each wave, between 1.5% and 3.5% of the participants claimed to have suffered cyberfraud.

Sociodemographic variables: sex (male 56.4%, female 43.6%); age in five age groups: 15 to 24 years (3.2%), 25 to 34 years (16.1%), 35 to 44 years (37.7%), 45 to 54 years (27.5%), and more than 55 years (15.5%) (M = 3.28, SD = 1.22); educational background: highest educational level attainment, 1 = elementary (0.7%), 2 = secondary (46.9%), and 3 = university studies (52.4%) (M = 2.48, SD = 0.53); and size of locality: from 1-less than 10,000 to 6-more than 500,000 inhabitants (M = 3.94, SD = 1.77).

Analytical Strategy

For the study of the trajectories of smartphone addiction, social support, and cyberfraud over time, we used a combined strategy of latent growth mixture modeling and survival curve analyses. We studied the evolution of cyberfraud over the 36 months of follow-up (data from six panels) using discrete-time survival analysis. This technique is specially designed to analyze survival rates in a sample of individuals. Survival here refers to the occurrence or nonoccurrence of an event (having experienced cyberfraud) in each panel. Participants who did not experience cyberfraud throughout the study were considered survivors.

Discrete-time survival analysis allows modeling the survival rates of a set of participants and relates that survival rate to other processes of interest. In our case, we modeled the relationship of the survival rate with the evolution of smartphone addiction and social support. We first modeled trajectories in social support and smartphone addiction using latent growth curves (LGMs). With this strategy, we tested whether the average evolution of users' smartphone addiction and social support were upward, downward, or zero during the three years of follow-up.

LGMM was then applied to the study of trajectories of smartphone addiction and social support and cyberfraud survival data to characterize the heterogeneity in these processes (i.e., growth and survival factors) in terms of latent classes (see Figure 1). The aim was to identify different subgroups among the participants with a differential profile in terms of their evolution of smartphone addiction, social support, and cyberfraud survival rates. According to Figure 1, the trajectories described by the latent growth factors of smartphone addiction (Ssa) and social support (Sss) may not be the same for all participants. This heterogeneity is modeled by the Factor C corresponding to the number of latent classes (or groups) that best describes the heterogeneity of the trajectories of smartphone addiction and social support. This part of the model is estimated using the latent growth mixture model technique. These heterogeneous trajectories, moreover, are associated with cyberfraud survival rates, represented by the survival factor SR. Models were estimated using MPLUS (version 8.2) software (^{Muthén, & Muthén, 1998-2017}).

Discrete-time survival analysis predicts time-to-first cyberfraud events and considers cyberfraud to be a nonrepeatable event. The data did not fit exactly with these requirements, so a brief discussion is necessary to justify the analytical strategy. First, the study design did not allow us to completely rule out that some participants might have experienced cyberfraud prior to the beginning of the study. This is a circumstance that potentially limits the scope of the results. Second, it does not seem that cyberfraud is a nonrepeatable event, as might be death, the onset of drug use, or first school expulsion, to give just a few examples. Rather, it is a recurring event, in that a person could be a victim of cyberfraud multiple times. Prior to the implementation of the technique, we checked some issues related to the reoccurrence of cyberfraud. Out of 716 study participants, 118 suffered cyberfraud during the 36 months of follow-up. Of those 118 participants, 64% (n = 77) experienced cyberfraud in a single period, and 21% (n = 25) experienced it in two periods. It is interesting to note that all participants with a record of cyberfraud in two periods experienced it in contiguous periods. Given that approximately 85% of participants experienced cyberfraud in a very localized time span, the strategy of analyzing cyberfraud for the first time seemed plausible. With this strategy, the estimation of model parameters is less computationally difficult than the estimation of survival curves of repeatable events (^{Lougheed, et al., 2019}).

Attrition Analyses

Participants who remained in the study throughout the follow-up phase showed some significant differences with dropouts. There was a tendency for male (χ² = 14.08, p < .001, Cramer's V = .68) and older [F(1, 3575) = 29.13, p < .001; dropouts = 3.11, follow-ups = 3.36], participants to remain in the study. There was also a trend that participants who remained in the study throughout follow-up showed higher levels of social support at baseline, [F(1, 575) = 11.33, p < .001; dropouts = 3.41, follow-up = 3.54], lower initial levels of smartphone addiction [dropouts = 1.02, follow-ups = 0.81, F(1, 1766) = 5.96, p < .01], and lower cyberfraud victimization rates (χ² = 15.01, p < .001, Cramer's V = .65) than those who dropped out. A total of 7.5% of the participants who dropped out of the study experienced cyberfraud at T1. This percentage was reduced to 3.5% among those who completed the follow-up. It is interesting to note at this point that there appeared to be self-selection in users to participate in the study guided by low smartphone addiction, high social support, and lower rates of cyberfraud victimization at T1. Table 1 shows the descriptive statistics of the study variables. The percentage of participants victimized by cybercrime in each period remained between 1.5% and 3.5%. This suggested that the rate of cyberfraud did not fluctuate significantly over time. On average, smartphone addiction showed a decrease (from 0.81 to 0.70), while social support showed an increase (from 3.54 to 3.75) throughout the study.

Growth Trajectories

The trajectories of smartphone addiction and social support were estimated using LGM. The model with joint estimation of the trajectories of smartphone addiction and social support showed a good fit to the data (χ² = 47.08, df = 31, p = .003; CFI = .99; RMSEA = .04, 95% CI [0.03, 0.04]). For the overall sample, at baseline, smartphone addiction was 0.81 (SE = 0.13, p < .001), and social support was 3.54 (SE = 0.03, p < .001). Smartphone addiction showed a slight decrease (slope = -0.06, SE = 0.03, p = .03), and social support showed an average increase (slope = 0.05, SE = 0.01, p < .001) over the six waves. Both the slope variance for smartphone addiction (var. slope = 0.19, SE = 0.09, p =.03) and social support (var. slope = 0.01, SE = 0.00, p = .03) were statistically significant. Likewise, variances of the intercepts (smartphone addiction and social support at baseline) were statistically significant (smartphone addiction, 4.18, SE = 1.53, p = .04; social support, 0.37, SE = 0.04, p < .001). Therefore, not all participants started the study with the same levels of smartphone addiction and social support. The variability of their slopes suggested heterogeneity in the trajectories (not all participants decreased their smartphone addiction or increased their social support scores in the same way or at the same rate). Baseline levels of smartphone addiction and social support were significantly and negatively related (-0.24, SE = 0.11, p = .03): at baseline, the lower the score for smartphone addiction, the higher the score for social support.

Survival and Latent Growth Mixture Analyses

In the next step, we sought to examine this heterogeneity in trajectories of smartphone addiction and social support in relation to survival rates in cyberfraud over time. The aim was to find different groups or classes of participants with significant differences in their trajectories of smartphone addiction and social support and verify whether they also showed different cyberfraud survival rates (see Figure 1). Discrete-time survival mixture analysis with survival predicted by growth trajectory classes was estimated. LGMM proceeds with the estimation of different models with an increasing number of classes, and then the model with the optimal number of classes is selected. Although there is no single method for the selection of these models, the literature in this field has pointed out that the selection should be conceptually and empirically guided. On the one hand, the resulting models must be consistent with available theory. On the other hand, empirical criteria such as entropy, number of participants per class, and likelihood ratio test (LRT) between models with a class difference of k =1 can help the researcher select the most suitable models. Entropy can range from zero to one, with values closer to one indicating higher classification accuracy and the degree to which classes are distinct from each other. Models with higher values of entropy (and closer to 1) are preferable to models with values farther from 1. LRT allows to statistically compare the chi-square difference of a model with k classes versus a model with one fewer class (k-1). A significant p value rejects the k-1 class model in favor of the k class model. Both entropy and LRT serve to guide the decision on the selection of the optimal number of classes that best define the sample. However, models that either identify classes with too few members or whose classes lack a theoretical basis are candidates for rejection, even with adequate levels of entropy or supported by LRT (see ^{Bauer & Curran, 2003}; ^{Celeux, et al., 2018}; ^{Muthén & Muthén, 1998-2017}, for an analysis). Maximum likelihood estimation with robust standard errors was used for model estimation.

Models with 1, 2, 3, 4 and 5 classes were estimated. First, we performed a likelihood ratio test using bootstrap LRT χ² (BLRT). Instead of assuming that the difference distribution follows a known distribution (e.g., the chi-square distribution), the BLRT empirically estimates the difference distribution and outperforms other LRTs to help determine whether the inclusion of an additional latent class while maintaining the same model specifications significantly improves the model fit (^{Nylund, et al., 2007}). The 2-class model was a statistical improvement over the 1-class model, according to BLRT results (value = 541.81, p < .001). The 3-class model did not provide a statistically significant improvement over the 2-class model (value = 326.32, ns). Subsequent models with a number of classes from 4 to 5 also showed no statistical improvement. The 2-3- and 5-class models showed a similar level of entropy (entropy = 0.98), while the 4-class model showed a considerable reduction in entropy (0.77). In terms of the number of participants per class, the 5-class model had two very small classes of 12 and 13 participants, while the 3- and 4-class models had two classes with approximately 40 participants each. The 2-class model grouped the majority of participants (n = 653) in one class and the remaining (n = 63) in another.

In the 2-class model, each class had different growth rates in smartphone addiction and social support and different survival curves. This reinforced the idea that classes 1 and 2 allowed the identification of statistically distinct subgroups and was also in accordance with substantive theory on trajectories of smartphone addiction and social support among the general population (^{Herrero, Torres, Vivas, & Urueña, 2019a}). Overall, the model comparison suggested that it would be appropriate to analyze latent classes in which these trajectories and the survival curve were different across classes to explain the heterogeneity observed in the growth trajectories for smartphone addiction and social support. The 2-class model met these requirements and was retained for further analysis.

Table 2.  Unstandardized Parameter Estimates for the 2-latent Class Model¹ (N = 716) 

Note. ¹Standard errors in parenthesis.

Smartphone addiction. Class 1 participants showed higher levels of smartphone addiction at the beginning of the study than class 2 participants (class 1, smartphone addiction = 2.26, SE = 1.05, p < .01; class 2, smartphone addiction = 0.78, SE = 0.14, p < .001). The evolution of smartphone addiction was also different in the two classes: class 1, smartphone addiction showed a marginal increase (slope = 0.40, SE = 0.23, p = .09), while in class 2, it significantly decreased (slope = -0.11, SE = 0.03, p = .001).

Social support. Class 1 participants showed lower levels of social support at the beginning of the study (class 1, social support = 3.16, SE = 0.09, p < .001; class 2, social support = 3.58, SE = 0.03, p < .001). The evolution of social support was also different in the two classes. In class 1, social support did not change (slope = 0.04, SE = 0.03, ns), while in class 2, it increased (slope = 0.05, SE = 0.01, p < .001) during the study. In summary, class 1 encompassed participants with high levels of addiction and low social support, while class 2 encompassed participants with lower levels of addiction and higher levels of social support.

Survival rates. The statistical significance of survival rates across classes was estimated from the mean of the latent factor (SR) in Figure 1. The study of the survival rates between classes was performed by taking class 2 as the comparison group (with a mean of zero). Class 1 participants showed a significant drop in survival rates throughout the study (factor SR mean = 1.54, SE = 0.25, p < .001) compared to class 2 participants. In other words, class 1 participants experienced higher rates of cyberfraud throughout the study (see Figure 2). Figure 2 plots the Kaplan-Meier estimator of survival curves for class 1 and class 2. The Kaplan-Meier estimator is a nonparametric statistic used to estimate the survival function. The decreasing horizontal steps shown in Figure 2, with a sufficiently large sample size, approximate the true survival function of both classes.

Note. All survival estimates to a given point represent the cumulative probability of surviving to that time. The probability for each group at time 7 represents the cumulative probability of survival (not having suffered cyber fraud) over the three-year follow-up period. The curve begins at T1 (which would be T0 at follow-up) with a 100% survival rate prior to the start of the study. Thus, T2 in the figure would be equivalent to T1 in our study and so on.

Figure 2.  Survival Kaplan-Meier Estimates for Class 1 and Class 2. 

For class 1, the probability of surviving (to not experience cyberfraud) at the end of the six periods of follow-up (36 months) was 0.59; conversely, for class 2, the probability of surviving after six months was slightly more than 0.89. In terms of percentages, 41% of class 1 participants (n = 26) experienced cyberfraud throughout the study, while only 11% of class 2 participants (n = 66) did. Class 1 participants showed a higher risk of cyberfraud during the study.

Class 1 participants, with a higher risk for cyberfraud, showed high levels of smartphone addiction and low levels of social support. Class 2 participants, with a lower risk of cyberfraud, started with higher levels of social support, which increased during follow-up, and low levels of smartphone addiction, which decreased over time. These results suggested that the levels of smartphone addiction and social support and their evolution over time allow the identification of participants with a higher risk of being victims of cyberfraud.

Strengths and Limitations

A strength of the study is its longitudinal nature. Compared to other research in this area that uses cross-sectional data to tentatively explore explanatory hypotheses, the availability of panel data allowed for the combined study of trajectories in smartphone addiction, social support, and cyberfraud over time. This circumstance adds robustness to the results of the study while allowing for a dynamic interpretation of the processes under scrutiny. From a theoretical point of view, the study combines different explanations of cyberfraud victimization that have traditionally followed parallel paths, something that can be considered a strength of the research. Thus, situational explanations of cybercrime victimization (L-RAT) indirectly suggest that smartphone addiction could be related to cyberfraud. These theoretical contributions, however, do not anticipate a particular incidence of social support on cyberfraud victimization, as our results suggest.

Beyond the strengths of the study, we also detected some potential limitations. The first of these potential limitations has to do with the estimation of survival curves for cyberfraud. As we have already briefly discussed in the rationale for choosing the discrete-time survival analysis technique to study the evolution of cyberfraud, it is likely that this variable does not fully fit the type of nonrepeatable events recommended for this statistical technique. Although most of the study participants experienced cyberfraud in a very temporally localized period, a small percentage reported cyberfraud in noncontiguous periods. In these participants, only the first victimization was predicted, so the models tested did not take into account subsequent victimizations. This may have affected the results of the study. When events are repeatable, a survival curve should be calculated for each event, where the time interval is counted from the last episode of victimization suffered (^{Masyn, 2009}). Another potential limitation that may have affected the results of the study lies in the fact that levels of smartphone addiction, social support, and cyberfraud were significantly related to permanence in the study. Thus, at higher levels of smartphone addiction, lower levels of social support, and higher rates of cyberfraud, participation in the study was significantly lower. This may have affected the results of the study, mainly as a consequence of reduced variability in the study variables. Further research should elucidate whether this circumstance could have had a downward effect on the magnitude of the coefficients found.

To summarize, our study seems to confirm that the addictive use of smartphones not only negatively affects the psychosocial well-being of users, as has been recognized in the scientific literature. It may also lead to the vulnerability of these addicted users to cybercrime. This circumstance places this type of user in a double vulnerability (in their well-being and in their cybersecurity) that should be the subject of further scientific scrutiny in future research.