An observational study of Internet behaviours for adolescent females following sexual abuse

Jennie G Noll; Ann-Christin Haag; Chad E Shenk; Michelle F Wright; Jaclyn E Barnes; Mojtaba Kohram; Matteo Malgaroli; David J Foley; Michal Kouril; George A Bonanno

doi:10.1038/s41562-021-01187-5

. Author manuscript; available in PMC: 2022 Jul 6.

Published in final edited form as: Nat Hum Behav. 2021 Sep 27;6(1):74–87. doi: 10.1038/s41562-021-01187-5

An observational study of Internet behaviours for adolescent females following sexual abuse

Jennie G Noll ^1,^2,^8,^✉, Ann-Christin Haag ^3,^8,^✉, Chad E Shenk ^1,⁴, Michelle F Wright ⁵, Jaclyn E Barnes ², Mojtaba Kohram ², Matteo Malgaroli ⁶, David J Foley ⁷, Michal Kouril ^2,⁹, George A Bonanno ^3,⁹

¹College of Health and Human Development, The Pennsylvania State University, University Park, PA, USA.

²College of Medicine, Cincinnati Children’s Hospital Medical Center, University of Cincinnati, Cincinnati, OH, USA.

³Department of Counseling and Clinical Psychology, Columbia University Teachers College, New York, NY, USA.

⁴College of Medicine, The Pennsylvania State University, Hershey, PA, USA.

⁵College of Liberal Arts, The Pennsylvania State University, University Park, PA, USA.

⁶Department of Psychiatry, New York University, Grossman School of Medicine, New York, NY, USA.

⁷College of Information Systems & Technology, The Pennsylvania State University, University Park, PA, USA.

⁸These authors contributed equally: Jennie G. Noll, Ann-Christin Haag.

⁹These authors jointly supervised this work: Michal Kouril, George A. Bonanno.

Author contributions

J.G.N. was the principal investigator of the study, was involved in conceiving and designing the study, participated in data acquisition, directed the analysis and interpretation of the findings, produced the manuscript drafts and revisions and provided final approval of the manuscript. A.C.H. was involved in conceiving the manuscript, directed and performed the analysis and interpretation of the findings, produced the manuscript drafts and revisions and provided final approval of the manuscript. C.E.S. was a coinvestigator, was involved in conceiving and designing the study, participated in data acquisition, aided in the interpretation of the findings and contributed to the manuscript drafts and revisions. M.F.W. aided in the analysis and interpretation of the findings and contributed to the manuscript revisions. J.E.B. was involved in conceiving and designing the study, participated in data acquisition and aided in the interpretation of the findings. M.Kohram was involved in conceiving and designing the study, participated in data acquisition, aided in the interpretation of the findings and contributed to the manuscript drafts. M.M. aided in the analysis and interpretation of the findings and contributed to the manuscript revisions. D.J.F. aided in the interpretation of the findings and contributed to the manuscript revisions. M.Kouril was a coinvestigator, was involved in conceiving and designing the study, participated in data acquisition, aided in the interpretation of the findings and contributed to the manuscript drafts and revisions. G.A.B. was involved in conceiving the study, aided in the interpretation of the findings and contributed to the manuscript drafts and revisions. All authors approved the submitted version.

^✉

Correspondence and requests for materials should be addressed to Jennie G. Noll or Ann-Christin Haag. jgn3@psu.edu; ah3784@tc.columbia.edu

PMCID: PMC9258728 NIHMSID: NIHMS1817043 PMID: 34580439

Abstract

Child sexual abuse (CSA) is associated with revictimization and sexual risk-taking behaviours. The Internet has increased the opportunities for teens to access sexually explicit imagery and has provided new avenues for victimization and exploitation. Online URL activity and offline psychosocial factors were assessed for 460 females aged 12–16 (CSA = 156; comparisons = 304) with sexual behaviours and Internet-initiated victimization assessed 2 years later. Females who experienced CSA did not use more pornography than comparisons but were at increased odds of being cyberbullied (odds ratio = 2.84, 95% confidence interval = 1.67–4.81). These females were also more likely to be represented in a high-risk latent profile characterized by heightened URL activity coupled with problematic psychosocial factors, which showed increased odds of being cyberbullied, receiving online sexual solicitations and heightened sexual activity. While Internet activity alone may not confer risk, results indicate a subset of teens who have experienced CSA for whom both online and offline factors contribute to problematic outcomes.

Today’s youth represent the first generation to grow up in a world where high-speed Internet is the norm, with up to 93% reporting daily usage¹ and wide availability across all socio-economic strata². There is spirited debate^3–5 regarding the impact of Internet use on adolescent development, with some studies showing negative associations^6–10, while others show negligible associations and even benefits^11–16. Recent reviews^4,17, including one reanalysis of large-scale international social datasets¹⁸, concluded that associations between the amount of digital technology use and adolescent well-being are small relative to other adolescent problems and thus do not warrant policy reform nor prevention efforts. Hence, the Internet, in and of itself, may not be a deleterious force in the lives of adolescents⁴. On the other hand, the Internet may be impacting some teens more than others depending on the extent and types of their online behaviours as well as their individual, offline characteristics and the contexts in which they are developing.

Child maltreatment is a powerful context that is associated with problematic Internet and social media behaviours^19–22, with child sexual abuse (CSA) having notable potency^20,23,24. There are several reasons why CSA should warrant particular concern. CSA has been shown to significantly disrupt the normal course of development—for example, early pubertal timing²⁵, deficits in cognitive development²⁶, dropping out of high school²⁷ and teenage childbearing²⁸—thereby introducing unique challenges that can accentuate the already difficult and often volatile developmental struggles of adolescence. Given that the Internet has presented new challenges for today’s youth in terms of complexities in social discourse and the potential for exploitation and victimization, learning how to engage the Internet safety and appropriately has become a key task of adolescent development that is likely to be especially challenging for survivors of CSA.

The traumatic sexualization brought on by the betrayal, powerlessness, sexual boundary violations and stigma associated with CSA has long been upheld as one compelling reason why sexual abuse can be uniquely noxious and distinct from other forms of maltreatment and adversity²⁹. For instance, the sexually explicit nature of CSA can disrupt sexual schemas²⁸ thus making survivors particularly vulnerable to disruptions in sexual development that can manifest in elevations in risky sexual behaviours^28,30–33, including higher rates of pornography consumption, compared to their non-abused peers^34,35. The Internet has created the possibility for unfettered access to pornography, whether through intentional seeking or accidentally through spam and pop-up menus. Exposure to sexual content and images can have a detrimental impact on adolescent sexual development in terms of normalizing risky sexual behaviours and endorsing ‘sexual scripts’ that can influence sexual health, risk-taking, objectification and dysfunction³⁶. Although much of what we know about the impact of pornography comes from studies of traditional sexual media (television, music, movies, magazines), associations have been largely upheld in studies of online pornography^37–39, especially for females^38,40. Even as 20% of adolescents report unwanted online exposures to sexually explicit content⁴¹ and 59% report intentional consumption⁴², few studies have tracked the impact of pornography use over time^43–45. Moreover, extant research has relied exclusively on adolescent self-reports. Thus, it is unclear whether and how Internet pornography and other adult-content material might be influencing the sexual development of today’s teens over time or whether CSA confers added risk.

Navigating adolescence can be difficult for CSA survivors due to their high susceptibility to peer victimization and bullying^46–48. Given the substantial co-occurrence between offline and online bullying among teens in general^49,50, it is likely that the high rates of peer victimization observed in CSA samples translates to increases in cyberbullying for this population, as reflected in a recent Canadian national survey showing that teens who have experienced CSA are twice as likely to be cyberbullied than non-abused teens⁵¹. Although, a large population study reported that cyberbullying mediated the relationship between social media use and subsequent mental health for adolescent females⁵², effects sizes for associations between cyberbullying and mental health have been small¹⁷, definitions of cyberbullying victimization are not always comparable across studies and the assessment of social media use has been limited to self-reports⁵². It is also unknown whether teens who have experienced CSA are more likely to use social media compared to other teens or whether their use of social media plays a role in their propensity to experience online bullying over time.

Finally, females who experience CSA report relatively high rates of sexual revictimization, exploitation and sex trafficking^48,53,54. The many avenues by which strangers might gain access to teens online has generated increasing concerns about sexual victimization^55–58. Indeed, social media and social networking websites are places where most Internet sex crimes originate^41,59–61 since sexual predators seek out vulnerable teens who are willing to engage in online sexual solicitations including engaging in titillating chat, expressing sexual interest and sharing sexual images^61,62. Perpetrators first become acquainted through social media⁶³ and then groom potential victims by building trust⁶⁴, using flattery and deception⁶⁵ and then luring victims into offline encounters. While offline meetings with strangers can be benign⁶⁶, agreeing to meet someone in person whose identity has not been fully confirmed is arguably dangerous for adolescents. For example, in a Danish national survey²³, 43% of teens reported meeting face to face (offline) with a person that they had first met online. While most of these meetings were with peers, roughly 6% resulted in some type of offence, including sexual solicitations and sexual victimization. To date, few longitudinal studies have connected specific Internet behaviours to subsequent Internet-initiated victimization and directionality is unclear⁴. Some evidence exists that abused and neglected teens receive aggressive sexual solicitations²⁴ and meet strangers offline²² more frequently than non-maltreated teens; however, despite their propensity for sexual exploitation and revictimization⁶³, there is not yet clear evidence that CSA survivors are at particular risk.

While CSA indeed confers unique vulnerabilities for the development of various forms of psychopathology and other deleterious sequelae⁶⁷, large proportions of adolescent survivors do not develop trauma symptoms^68,69, with research showing personality styles and family support variables accounting for persisting resilience⁷⁰. Internet safety research has identified similar factors that account for variation in Internet-initiated victimization. Sociodemographic disadvantage^24,71–73, delinquency⁷⁴, emotional and conduct problems^23,73, substance use^74,75, low cognitive ability²², depression⁷⁴, poor emotional control⁷⁶, low self-esteem⁷⁷ and being female^71,78 have been associated with elevations in online victimization, whereas positive parent–child relationships, school engagement, academic achievement and involvement in prosocial opportunities such as sports and clubs appear to be protective⁷⁷. Taken together, these findings suggest that understanding problematic online behaviours and outcomes will be enhanced by models that consider how such behaviours operate within a larger offline context^79,80.

While the current study sought to understand if CSA survivors exhibit a greater propensity to use the Internet in ways that are consistent with traumatic sexualization, we also sought to discern whether there might be some survivors who require larger doses of oversight and intervention relative to others who may be less affected. Most current theories of child development agree that risk factors occur in combination, as opposed to isolation, and that the intersection or accumulation of risk, rather than any single risk factor, accounts for the preponderance of negative developmental sequelae^81–83. Given the interplay between online victimization and offline psychosocial factors^{50,73,84–86}, a recent analysis of prevention programming across several countries called for the integration of online safety education into offline violence prevention as a means to enhance the efficacy and sustainability of both⁸⁷. However, this review recognized the dearth of well-designed studies that can inform such an integration. To begin to address these gaps, we employed a relatively holistic assessment strategy that would allow the modelling of a constellation of online behaviours along with various offline individual-, family- and community-level characteristics, which might explain variations in the serious consequences of Internet use including Internet-initiated victimization and aberrant sexual development.

Recent reviews of the impact of technology on adolescent development^4,17,18 along with several recent commentaries^4,88 agree that there are major shortcomings in the extant Internet research. These include reliance on adolescent self-reports to quantify Internet activity, which show only moderate correlations with actual use^89–92, an over-reliance on cross-sectional research and failing to control for potential confounding factors. Reviews have also highlighted the need to move beyond a singular focus on the amount of time kids engage with technology towards a more nuanced assessment of the types and characteristics of various Internet activities. The present study sought to go beyond adolescent self-reporting by tracking actual Internet use patterns through direct observation of actual online activity. Participating teens were provided with laptop computers capable of recording all URL activity, which was then objectively quantified using open-source ‘blacklist’ databases into commonly understood classification categories (that is, entertainment, social media, gaming, pornography and education). Keywords associated with videos on the YouTube and Netflix platforms were likewise quantified for containing adult content such as illicit drug use, sexual material and violence. A sophisticated machine-learning authentication algorithm was also developed to ensure that only behaviours specific to those enrolled in the study were quantified. In-person interviews were also conducted with adolescents and caregivers assessing psychosocial risk and protective factors along with victimization histories both before the laptop assignments and then again annually for two follow-up assessments (Fig. 1).

Fig. 1 | — Females who experienced CSA were recruited from child welfare agencies for confirmed CSA within the previous 12 months. Comparison females were recruited from flyers posted in regional outpatient teen health centres, community centres and local middle schools and screened before enrolment for no history of CSA in statewide child welfare records. DMC females were enrolled to match the CSA subgroup on age, family income and minority ethnicity status. CMC females were enrolled to mirror the income and ethnicity composition of the catchment region based on 2010 census track information.

The sample consisted of 460 females aged 12–16 years. Females were the focus because, relative to males, they are more likely to be cyberbullied⁵⁸, experience CSA at higher rates⁹³, field the bulk of online sexual solicitations²⁰, are disproportionately impacted by social media⁹⁴ and are more likely to be targeted by online sexual predators^71,95 and lured into sex trafficking⁹⁶. The sample was constructed to more fully understand CSA as a risk context and therefore included females recruited through child welfare agencies for having experienced substantiated CSA confirmed within the previous year (n = 156). A sample of comparison females (n = 304) was also enrolled after being screened through child welfare records to ensure there was no substantiated CSA in this condition. A common design strategy in observational research where randomization is not possible, this comparison group was recruited to ensure sociodemographic similarity to the CSA group in terms of age, ethnicity and family income to provide naturalistic control over confounding, extraneous variability⁹⁷. Although an effective control over overt bias, the sociodemographic characteristics most often used to create balance across study groups in quasi-experimental research are often themselves well-established risk factors that have deleterious effects on child development akin to those attributable to child maltreatment⁹⁸. The intentional inclusion of a subsample with such sociodemographic characteristics can thus introduce alternative sources of hidden bias that, depending on the research question, could attenuate or inflate results⁹⁹. Therefore, the comparison condition for this study included (1) a subgroup of demographically matched comparisons (DMCs; n = 156), who were matched to the CSA sample on age, family income and minority ethnicity, and (2) a subgroup of census-matched comparisons (CMCs; n = 148), who were similar in terms of average income and ethnicity distributions to the region of the US where the study took place. The inclusion of multiple comparisons enhances the ability to parse out effects that are attributable to the sexual abuse per se versus other sources of overt or hidden biases introduced by sociodemographic matching¹⁰⁰.

Capitalizing on these methodological advances, the overall aims of this study were to objectively assess and describe Internet behaviours for CSA and comparison females and specifically test whether females who have experienced CSA would be more likely to engage in Internet behaviours consistent with traumatic sexualization, that is, more likely to consume pornography and experience Internet-initiated victimization than their non-abused peers. Furthermore, this study adopted a person-centred approach aimed at empirically deriving distinct subgroups (or profiles) of adolescents—based on constellations of online Internet use patterns as well as offline individual-, family- and community-level risk and protective factors—in an attempt to understand if certain subgroups would exhibit varying degrees of risk for subsequent Internet-initiated victimization and aberrant sexual development in longitudinal models. Based on the literature demonstrating how the traumatic sexualization of CSA can lead to disruptions in sexual schemas²⁸, risk profiles that include significantly high rates of pornography use were of particular interest. Similarly, because revictimization is commonly cited as a sequela of CSA⁶³ and because social media platforms are where cyberbullying is most likely to occur⁵², profiles that include significantly high rates of social media activity were also of particular interest. As such, the following hypotheses were tested: (1) independently of sociodemographic disadvantage, females who have experienced CSA will be more likely to consume pornography and experience Internet-initiated victimization including being cyberbullied, receiving online sexual solicitations and having offline meetings with strangers; (2) Internet use profiles characterized by pornography consumption, coupled with psychosocial risk factors and few protective factors will be predictive of being cyberbullied, receiving online sexual solicitations, having offline meetings with strangers first met online and heightened offline sexual behaviours; (3) Internet use profiles characterized by high levels of social media use will be predictive of subsequently being cyberbullied.

Results

Descriptive statistics.

The Venn diagram in Fig. 2 depicts all URL activity among the 23,839 total authenticated, 2-h Internet surfing sessions analysed. In general, most sessions included social media and entertainment; however, in virtually every single session, other URL activities were also engaged. Additional details regarding the number of authenticated sessions for this sample are included in Extended Data Fig. 1.

Table 1 includes the descriptive statistics for all variables used in the analyses. Overall, the sample had a mean age of 14.26 years, had 56.3% minority ethnicity and reported a mean annual family income level of approximately US$35,000. Consistent with the sampling strategy, there were no statistically significant differences across CSA and DMC subgroups with respect to minority ethnicity, income or age at study entry (time 1). However, CMC females were of higher income and lower percentage minority ethnicity than both CSA (t = 13.95, P < 0.001; χ²(1) = 50.48, P < 0.001) and DMC (t = 12.81, P < 0.001; χ²(1) = 38.57, P < 0.001) females. CMC females were also 4.5 months younger at time 1 than females who experienced CSA (t = 2.70, P = 0.01). As a result, income, minority ethnicity and age at time 1 were used as demographic control variables in subsequent analyses.

Table 1 |.

Descriptive statistics for the variables used in the analyses across CSA and comparison subgroups

	Total (n = 460)		CSA versus comparison contrasts								Post-hoc three subgroup comparisons^b
	Total (n = 460)		CSA (n = 156)		Comparison (n = 304)		Omnibus MANCOVA F_{(15, 424)} = 14.79, P < 0.001				DMC (n = 156) versus CSA		CMC (n = 148) versus CSA
	Mean	s.d.	Mean	s.d.	Mean	s.d.	Test statistic	s.e.^a	95% CI^a	ηp ²	Mean	s.d.	Mean	s.d.
Time 1 age	14.26	1.26	14.44	1.41	14.16	1.16	t₍₄₅₈₎ = −2.26, P = 0.03				14.27 P = 0.22	1.16	14.05 P = 0.01	1.17
Income	4.63	3.38	3.09	2.65	5.42	3.45	t₍₄₅₇₎ = 7.38, P < 0.001				3.46 P = 0.21	2.40	7.49 P < 0.001	3.16
Percentage minority	56.3		71.1		48.7		χ²(1) = 21.16, P < 0.001				66.0 P = 0.23		30.4 P < 0.001
Overall activity (AMD)	176.67	143.36	220.50	159.90	154.11	128.61	F_{(1, 446)} = 4.63, P = 0.33	4.79	0.11–18.54	0.06	187.51	143.45	118.66	99.56
Social media (AMD)	91.50	116.83	118.90	127.60	77.40	108.42	F_{(1, 446)} = 2.68, P = 0.48	3.81	0.01–13.76	0.01	107.00	127.48	45.99	71.65
Entertainment (AMD)	87.91	87.31	91.32	78.39	86.15	91.63	F_{(1, 446)} = 0.36, P = 0.83	1.65	0–5.513	0.001	100.21	101.37	71.22	77.62
Pornography (AMD)	7.99	21.31	11.62	29.08	6.08	15.63	F_{(1, 446)} = 4.46, P = 0.33	4.62	0.04–17.04	0.02	7.61	18.76	4.47	11.25
Gaming (AMD)	16.97	50.96	18.55	32.33	16.16	58.31	F_{(1, 446)} = 0.03, P = 0.99	1.74	0–5.52	0	22.58	72.81	9.34	36.26
Education (AMD)	3.01	6.15	1.82	2.85	3.63	7.21	F_{(1, 446)} = 8.04, P = 0.07	4.35	2.51–19.62	0.02	3.36	6.09	3.92	8.25
Adult-content videos (AND)	4.92	7.37	6.43	8.75	4.15	6.46	F_{(1, 446)} = 0.34, P = 0.86	2.00	0–7.21	0.001	5.35	7.15	2.87	5.37
Substance use	1.69	3.90	3.47	5.72	0.80	2.03	F_{(1, 435)} = 39.68, P = 0.001	12.10	17.64–65.77	0.12	1.04 P < 0.001	2.32	0.55 P < 0.001	1.64
Depression	15.35	10.69	19.46	11.89	13.28	9.33	F_{(1, 446)} = 16.55, P = 0.08	9.54	3.46–39.44	0.04	13.67	9.36	12.87	9.32
Low impulse control	34.26	9.80	36.38	9.83	33.20	9.63	F_{(1, 447)} = 6.30, P = 0.22	5.17	0.30–19.48	0.03	33.56	9.88	32.83	9.38
Poor executive functioning	14.84	10.29	19.20	11.27	12.60	8.98	F_{(1, 437)} = 28.30, P = 0.02	12.01	10.62–57.75	0.09	13.19 P < 0.001	9.09	11.99 P < 0.001	8.85
Self-esteem	13.84	3.58	12.82	3.53	14.36	3.50	F_{(1, 446)} = 8.30, P = 0.17	6.02	0.83–23.90	0.05	14.09	3.79	14.64	3.15
Grades	5.97	1.63	5.48	1.70	6.22	1.53	F_{(1, 446)} = 2.93, P = 0.48	4.11	0.01–15.11	0.01	5.87	1.62	6.59	1.34
Prosocial activities	7.53	6.76	4.18	4.81	9.24	6.97	F_{(1, 447)} = 30.65, P = 0.003	10.21	13.86–53.01	0.14	7.39 P < 0.001	6.51	11.20 P < 0.001	6.93
Parental quality	68.42	11.36	66.94	12.24	69.17	10.83	F_{(1, 447)} = 0.12, P = 0.94	1.59	0–5.81	0	68.69	11.34	69.69	10.28

Open in a new tab

AMD/AND, average minutes/number per day. All F tests (univariate bootstrap ANCOVAs) were controlled for time 1 age, income, ethnic minority status, authentication mode, contamination and other types of maltreatment. ηp, partial η².

Bootstrap s.e. and 95% CIs (1,000 iterations).

Post-hoc comparisons based on Sidak-adjusted bootstrap pairwise comparisons.

Supplementary Table 1 includes zero-order associations between URL activity and psychosocial variables as well as partial correlations with age at time 1, income level and minority ethnicity partialled. There were several non-trivial correlations between CSA (versus comparisons) and several URL activities at the zero-order and after partialling for demographics, that is, CSA was associated with higher overall activity (r = 0.124) and pornography consumption (r = 0.110) and lower educational activity (r = −0.128). However, there were no statistically significant differences across the CSA versus DMC or CMC subgroups for any of the URL activity variables when controlling for additional covariates identified in the Methods that could introduce bias including authentication mode, contamination and other types of maltreatment (see the results presented in Table 1 and Methods for a full description of these covariates). Females who experienced CSA showed higher rates of substance use (F_{(1, 435)} = 39.68, P = 0.001), lower executive functioning (F_{(1, 437)} = 28.30, P = 0.02) and lower involvement in prosocial activities (F_{(1, 447)} = 30.65, P = 0.003) relative to both DMC and CMC females (Table 1).

Latent profile analysis.

Latent profile analysis (LPA) was used as a person-centred approach to empirically derive distinct profiles of adolescents based on the constellation of their URL activities coupled with coexisting individual-level and family/community-level risk and protective factors that have been shown in previous research to be associated with Internet behaviours. LPA yielded a best-fitting stable model indicative of 3 profiles (Bayesian information criterion (BIC) = −13876.54 and integrated complete likelihood (ICL) = −13902.20) using Model VVE from R mclust¹⁰¹, indicating variable volume, variable shape, equal orientation and ellipsoidal distribution. Significant bootstrap likelihood ratio test (BLRT) results (P = 0.01) were generated for all possible alternative model comparisons suggesting the best fit for the 3-profile model. The average probability of latent profile membership was high, which implies good discrimination between the three profiles: 0.981 for profile 1; 0.968 for profile 2; and 0.981 for profile 3. An excellent entropy value probability of 0.978 indicated high accuracy in profile separation, supporting the resultant latent profiles. Detailed fit indices are presented in Supplementary Table 2.

As depicted in Fig. 3 and detailed in Table 2, the three resultant profiles consisted of females at various levels of risk: relatively low risk (profile 1); moderate risk (profile 2); and relatively high risk (profile 3). Profile 1 consisted of 139 females at relatively low risk based on statistically significant differences from the other profiles on all variables used in the LPA, that is, lower rates of all URL activities and high individual- and family/community-level psychosocial functioning. In contrast, profile 3 consisted of 93 females at relatively high risk based on significantly high URL activities and poor psychosocial functioning relative to profile 1. Profile 2 consisted of 200 females at moderate risk in terms of moderate URL activity but exhibiting several psychosocial risks. Like females in profile 3, profile 2 females showed statistically significant elevated risk on all LPA variables relative to profile 1 but were similar to profile 3 (that is, no statistically significant differences) in terms of social media use, entertainment, educational use, adult-content videos, impulse control, executive functioning, self-esteem, grades, prosocial activities and parental quality.

Table 2 |.

Profile differences for variables included in the LPA

Omnibus MANCOVA	Profile 1 (n = 139)			Profile 2 (n = 200)			Profile 3 (n = 93)			Post-hoc three subgroup comparisons^a
F_(30,424) = 24.05, P < 0.001	EMM	Bootstrap s.e.	Bootstrap 95% CI	EMM	Bootstrap s.e.	Bootstrap 95% CI	EMM	Bootstrap s.e.	Bootstrap 95% CI	Profile 1 versus 2 P	Profile 1 versus 3 P	Profile 2 versus 3 P
Overall activity	91.85	5.50	82.11–103.46	203.16	8.98	185.09–220.42	244.28	17.49	210.24–279.35	<0.001	<0.001	0.04
Social media	41.98	3.84	35.28–50.26	109.99	9.17	92.47–130.43	127.08	12.65	100.71–150.26	<0.001	<0.001	0.35
Entertainment	37.77	3.47	30.81–44.90	116.01	6.39	103.34–128.42	103.73	11.11	81.30–126.36	<0.001	<0.001	0.32
Pornography	1.27	0.53	0.29–2.31	3.14	0.39	2.37–3.95	29.32	4.05	22.12–37.63	0.02	<0.001	<0.001
Gaming	3.84	1.63	0.81–7.20	9.30	1.26	6.69–11.60	53.56	10.87	34.04–76.60	0.04	0.003	0.01
Education	0.78	0.32	0.07–1.36	3.34	0.34	2.71–4.03	5.32	1.20	3.08–7.92	<0.001	0.01	0.13
Adult-content videos	2.19	0.24	1.71–2.69	6.71	0.56	5.60–7.82	5.11	0.66	3.87–6.40	<0.001	<0.001	0.08
Substance use	0.21	0.13	−0.34–0.47	1.24	0.15	0.95–1.55	4.87	0.69	3.65–6.33	<0.001	<0.001	<0.001
Depression	11.28	0.67	9.95–12.62	15.67	0.79	14.20–17.29	19.90	1.39	17.58–23.01	<0.001	<0.001	0.005
Low impulse control	30.30	0.80	28.74–31.94	36.29	0.68	34.94–37.56	37.00	1.01	35.04–38.97	<0.001	<0.001	0.52
Poor executive functioning	12.37	0.81	10.89–13.94	15.44	0.76	13.95–17.00	17.11	1.11	14.84–19.27	0.01	0.003	0.23
Self-esteem	14.82	0.29	14.25–15.38	13.73	0.25	13.26–14.20	13.05	0.39	12.17–13.70	0.007	<0.001	0.10
Grades	6.32	0.11	6.10–6.53	5.76	0.012	5.53–5.98	5.94	0.17	5.61–6.26	<0.001	0.07	0.38
Prosocial activities	9.50	0.57	8.29–10.50	6.76	0.47	5.87–7.80	6.43	0.66	5.24–7.83	<0.001	0.003	0.98
Parental quality	72.29	0.72	70.89–73.76	67.71	0.78	66.21–69.18	65.91	1.38	62.94–68.36	<0.001	<0.001	0.22

Open in a new tab

EMM, estimated marginal mean, controlled for time 1 age, income, ethnic minority status and authentication mode.

Post-hoc comparisons based on Sidak-adjusted bootstrap pairwise comparisons (1,000 iterations).

As shown in Fig. 3, CMC females had the highest percentage representation in profile 1 (58.3%), DMC females in profile 2 (42.5%) and females who experienced CSA in profile 3 (54.8%). As depicted in the forest plots of Fig. 4, when compared to profile 1, the odds for females who experienced CSA to be represented in profile 3 were 4.58 times greater than for comparison females (model 1a), 5.72 times greater than for CMC females and 3.64 times greater than for DMC females (model 1b). Compared to profile 1, the odds for females who experienced CSA to be represented in profile 2 were 2.21 time greater than for comparison females (model 1a), 3.79 times greater than for CMC females but there were no statistical differences than the odds for DMC females (model 1b). Further details of the multinomial regression analyses provided in Supplementary Table 3 indicated significantly increased odds for CMC females to be represented in profile 1.

Fig. 4 | — ORs (Exp(B)) were adjusted for covariates of time 1 age, income, minority ethnicity, contamination, other forms of child maltreatment and authentication mode. Covariates are fully described in the Methods. ORs were inversed for ease of interpretation. Profiles 1, 2 and 3 refer to the profiles derived in the LPA of Fig. 3. 95% confidence intervals (CIs) are presented for the binomial logistic regression models (models 2a through 7). Wald CIs are presented for the multinomial regression models (models 1a and 1b). CSA serves as the reference class in models 1a, 1b, 2a, 2b, 4a, 4b, 6a and 6b. Profile 3 serves as the reference class in models 3a, 5 and 7. Profile 2 serves as an additional reference class in model 3b to show the contrast between profiles 2 and 1, which is essential for testing hypothesis 3. Exact P values are presented when CIs do not include 1.0. Full models that include the parameter estimates for the covariates are included in Supplementary Tables 3–5.

Main effects for CSA controlled for by sociodemographics.

Pornography: The results reported in Table 1 indicated that there were no statistically significant differences in the levels of pornography use for CSA versus comparison females (F_{(1, 446)} = 4.46, P = 0.33). As such, this aspect of hypothesis 1 was not confirmed. Results for the remaining aspects of hypothesis 1 involving dichotomous outcomes are in depicted Fig. 4 with full models reported in Supplementary Table 4. Cyberbullied: Overall, 24.2% of the sample (CSA = 35.8%, comparisons = 18.4%) reported being cyberbullied at follow-up. As seen in Fig. 4, model 2a, the odds for females who experienced CSA to be cyberbullied were significantly greater than for comparison females (odds ratio (OR) = 2.84). Model 2b further shows these odds to be significantly greater than for both CMC (OR = 2.97) and DMC (OR = 2.76) females, confirming this aspect of hypothesis 1; sexual abuse confers risk for being cyberbullied independent of sociodemographic disadvantage. Online sexual solicitations: Overall, 12.3% of the sample (CSA = 17.1%, comparisons = 9.8%) reported receiving upsetting online sexual solicitations at follow-up. As seen in Fig. 4, models 4a and 4b, there were no statistically significant differences in the odds for females who experienced CSA receiving online sexual solicitations compared to either DMC or CMC females, rendering this aspect of hypothesis 1 unconfirmed. Meeting strangers offline: At follow-up, 10.0% of the sample (CSA = 17.1%, comparisons = 6.3%) reported meeting a stranger offline whose identity had not been fully confirmed. Figure 4, model 6a, shows that the odds for females who experienced CSA to meet strangers offline were 2.05 times greater than for comparison females. However, as seen in model 6b, these odds were significantly greater for females who experienced CSA versus CMC females (OR = 4.97, P = 0.02) but the difference in odds for females who experienced CSA versus DMC females were not statistically significant (Supplementary Table 4). This aspect of hypothesis 2 was therefore not confirmed.

Profile including heavy pornography use predicting outcomes.

As seen in Table 2, profile 3 exhibited significantly higher rates of pornography use relative to both profiles 1 and 2. Therefore, profile 3 was used as the reference class to test hypothesis 2. The results of the logistic regression analyses are depicted in the forest plots of Fig. 4 and detailed in Supplementary Table 5. Cyberbullied: The results indicated that the odds for profile 3 females to be cyberbullied at follow-up were 3.14 times greater than for profile 1; however, there were no statistically significant differences in the odds for being cyberbullied across profiles 2 and 3 (model 3a). Online sexual solicitations: For females in profile 3, the odds of receiving online sexual solicitations at follow-up were 2.63 times greater than for profile 1 and 2.64 times greater than for profile 2 (model 5). Meeting strangers offline: There were no statistically significant differences in the odds for females in profile 3 to meet strangers offline at follow-up relative to females in profiles 1 or 2 (model 7). Offline sexual activity: Table 3 includes the longitudinal bootstrap ANCOVA model predicting the level of offline sexual activity at follow-up. The results indicated that females in profile 3 reported significantly higher levels of offline sexual activity compared to females in profile 1 (t = 3.57, P < 0.001) and profile 2 (t = 4.75, P < 0.002). There were no statistically significant differences observed between profiles 1 and 2 (t = 1.93, P = 0.05).

Table 3 |.

Bootstrap ANCOVA model of latent profiles predicting sexual activity at follow-up

Bootstrap ANCOVA
	d.f.	F	s.e.^a	95% CI^a	P	ηp ²
Profile	2	11.70	5.48	3.80–25.54	0.03	0.13
Time 1 age	1	22.23	9.39	8.91–46.11	0.02	0.18
Income	1	2.31	3.48	0.01–12.50	0.51	0
Minority ethnicity	1	7.49	5.66	0.69–22.30	0.19	0.01
Time 1 sexual activity	1	112.35	26.36	67.59–167.96	<0.001	0.22
Authentication mode	1	0.08	1.50	0–4.86	0.96	0
Residuals	408
Bootstrap post-hoc comparisons ^b
	EMM	s.e.^a	95% CI^a	Profile 1 versus 2	Profile 1 versus 3	Profile 2 versus 3
Profile 1	6.02	0.51	5.04–7.06
Profile 2	7.38	0.46	6.51–8.30	P = 0.05	P < 0.001	P = 0.002
Profile 3	9.99	0.74	8.54–11.49

Open in a new tab

Bootstrap s.e. and 95% CIs (1,000 iterations).

ηp², partial η².

Sidak-adjusted post-hoc bootstrap pairwise comparisons.

Adjusted R² = 0.39. n = 416.

Profile including heavy social media use and cyberbullying.

As seen in Table 2, profiles 2 and 3 exhibited significantly higher levels of social media use relative to profile 1. As such, rates of being cyberbullied for both profiles 2 and 3 were contrasted with rates for profile 1 to test hypothesis 3. As seen in Fig. 4 and Supplementary Table 5, the odds for females in profiles 3 to be cyberbullied were 3.14 greater than for profile 1 (model 3a). The odds for females in profile 2 were 2.64 greater than for profile 1 (model 3b). There were no statistically significant differences in the odds of being cyberbullied for profile 2 versus profile 3 (model 3a).

Discussion

Results from this study indicate that, on average, females who experienced CSA did not use the Internet in ways that were dissimilar to their non-abused peers. Females who experienced CSA were, however, at significantly increased odds for being cyberbullied at follow-up controlling for potential confounding factors. Results further indicated additional aspects of the larger social ecology that work in conjunction with Internet use to compound vulnerability for a subset of CSA survivors. For example, relative to comparison females, females who experienced CSA were at increased odds for being represented in the highest-risk LPA profile (profile 3). Females in this profile were at increased odds for Internet-initiated victimization, in terms of being cyberbullied and receiving upsetting online sexual solicitations and reported significantly heightened offline sexual activity at follow-up. Although pornography has particular relevance to the study of the traumatic sexualization accompanying CSA, and was elevated in profile 3, this finding does not indicate that pornography, in and of itself, is the sole driving force behind the risk for subsequent Internet-related victimization and exploitation. Instead, the person-centred approach used in this study identified a subgroup of teens who are homogeneous with respect to the constellation of multiple, interacting risk factors that, taken together, are associated with significant elevations in risk for deleterious outcomes.

Females who experienced CSA were also at significantly increased odds for being represented in the moderate-risk profile (profile 2), which was characterized by moderate Internet usage (including elevations in social media use relative to profile 1) as well as moderate to high rates of individual- and family/community-level psychosocial risks. Females in profile 2 were at increased odds for being cyberbullied relative to the lowest risk profile (profile 1) but showed similar rates of cyberbullying compared to profile 3. These results suggest that there may be multiple typologies of females who experienced CSA and who are vulnerable to being cyberbullied, that is, those clustering in profile 3 and those clustering in profile 2.

As opposed to variable-centred approaches, which are designed to maximize the overall percentage of variance explained by a competing set of independent variables, person-centred approaches such as LPA maximize ecological validity through an intuitive summary of how a relatively large set of risk factors is organized within individuals, families and contexts to confer risk. Variable-centred approaches assume that, for the average member of a population, certain behaviours should be modified to reduce the average risk for developing a certain outcome, whereas person-centred approaches identify constellations of characteristics that differentiate persons from one another, thereby identifying subgroups of individuals who can be targeted to maximize the potency and cost-effectiveness of preventive interventions¹⁰². Thus, results suggest that there is a subset of females who have experienced CSA who exhibit relative elevations in URL activity coupled with problematic psychosocial factors, such as substance use and depression, where targeted intervention is warranted to stave off being cyberbullied, receiving online sexual solicitations and subsequent risky sexual behaviours.

The goal of person-centred approaches is not to discern which input variables contribute most to the prediction of an outcome. However, given the current debate regarding whether or not the Internet has had an appreciable impact on the development of today’s youth and whether psychosocial risk factors might uniformly outperform Internet-use variables in models of risk prediction, we performed a post-hoc exploratory analysis to determine which variables used in the LPA would be retained as important features in the optimal prediction of outcomes. Least absolute shrinkage and selection operator (LASSO)¹⁰³ regression is a machine learning approach whereby redundant or non-relevant (in terms of a lack of unique or interactive contribution to the combined prediction) variables are shrunk to zero to optimize the selection of retained variables (or features) that lead to a model where prediction error is minimized. As reported in Supplementary Table 10, the LASSO regression results indicate that when considered in the same model with all Internet and psychosocial variables, social media use was a non-zero feature in the prediction of being cyberbullied and receiving online sexual solicitations and pornography was a non-zero feature in the prediction of heightened offline sexual activities at follow-up. Because the focus of LASSO is on the best combined prediction, parameter estimates are not reliably interpretable in terms of their independent contributions. However, results show that parameter estimates for psychosocial variables, especially substance use, are consistently larger than those for Internet-use variables, even though several aspects of both sets are retained as important features in optimal models.

This study employed methodological advances entreated by recent reviews^4,17,18,88 to advance digital technology research including: recording actual Web behaviour; authentication of Web activity; objective classification of URLs into common-use categories; quantification of video keywords; using validated instruments and structured interviews to assess psychosocial factors and victimization histories; and a longitudinal design. However, several caveats should be considered when interpreting the results. First, our sample excluded teen males. While we intentionally focused on teenage females for several reasons^20,58,94, males reliably report qualitatively different Internet-use patterns compared to females (for example, social media use is higher among females but gaming is higher among males¹⁰⁴) and the risk pathways to online victimization for males are likely quite distinct. Second, results are not generalizable to all teens given the focus on the impact of the Internet for females who have experienced CSA and the self-selected nature of the sample. Third, this study took place between 2012 and 2015 and the landscape of Internet access has changed dramatically in recent years. For example, teens increasingly use their smartphones to interface with the Internet and the websites frequented by teens transition in and out of popularity¹. However, our reliance on laptop computers to capture Internet behaviours was consistent with the larger population of teens in the US according to a 2015 technology overview¹⁰⁵ showing that adolescents aged 13–14 most often used laptop and desktop computers to access the Internet. By quantifying URLs into broad categories of behaviours that continue to be relevant (social media, entertainment, pornography, gaming), results are arguably applicable to today’s teens. Moreover, the extent to which participants reported using alternative platforms, such as mobile phones, to access the Internet during the study period was assessed (36.7%) and a sensitivity analysis was performed excluding these teens; all results were replicated (Supplementary Tables 6–8). This study was also initiated before the establishment of the Open Science Forum and was therefore not preregistered. Fourth, this study did not include a reliable measurement of how teens interacted with their peers in ways that could influence Internet use or representation in LPA profiles, which is a glaring omission of an important feature of adolescent development. Finally, this study does not provide causal evidence that the Internet impacts the development of females who have experienced CSA because the reverse could also be true, that is, females who have been sexually abused might actually seek out the Internet as a means to cope with or make sense of sexual trauma.

The results presented further indicate that there is wide variation in how teens who have experienced CSA might be approaching the Internet and that problematic Internet-related outcomes for survivors are by no means inevitable. Indeed, over 13% of females who experienced CSA were represented in the lowest risk profile group (profile 1) characterized by a lack of both online and offline risk factors. Although this percentage is small, these findings are nonetheless consistent with abundant evidence documenting resilience and flexible adaptation following even the direst circumstances¹⁰⁶. Still, females who experienced CSA were shown to be at increased odds for being victims of cyberbullying and were overly represented in profile 3, where subsequent online exploitation was elevated. This suggests that a large portion of CSA survivors could benefit from enhanced intervention approaches to ward off and cope with Internet-initiated victimization in terms of cyberbullying and sexual advances. For example, such information and training could be seamlessly integrated into the most widely disseminated evidence-based treatment for CSA, trauma-focused cognitive behavioural therapy¹⁰⁷, where the final phase is devoted to enhancing safety through teaching appropriate boundaries, identifying potential triggers and warning signs of danger and devising a safety plan to reduce revictimization. Likewise, universal Internet safety prevention programmes designed to raise awareness (akin to those delivered in school settings) could be enhanced by equipping providers with the knowledge to recognize detectable trauma symptoms and refer such teens to more intensive interventions¹⁰². Such approaches will likely have a substantial benefit for a large proportion of the population given that CSA is a persistent, international public health issue¹⁰⁸, with pooled national prevalence estimates of 15% for females and 8% for males¹⁰⁹ across 24 nations and recent incidence rates up by 6% in the US for the first time in over 15 years^110,111.

Methods

All methods and procedures were approved by the institutional review board (IRB) at the regional hospital where the study took place (IRB no. 2012-0613; Federalwide Assurance no. 00002988). A federal certificate of confidentiality was also secured (CC-HD-12-83). Caregivers provided written informed consent and adolescents provided written assent. Monetary compensation was provided to both caregivers and adolescents according to IRB approval. Adolescents received US$120 for lab visits and 4 weeks of laptop participation. Caregivers received US$20 for lab visits. Caregivers were reimbursed for transportation expenses. Additional details regarding participant safety are included in the Supplementary Methods.

Sample.

Participants were aged 12–16 (mean = 14.25, s.d. = 1.25) at the initial assessment (time 1) which took place in December 2012 through to September 2015. As depicted in Fig. 1, two additional follow-up assessments commenced annually at time 2 (n = 434; 94.3%) and time 3 (n = 421; 91.5%) when the sample was mean-aged 15.32 (s.d. = 1.25) and 16.42 (s.d. = 1.26). The sample was drawn from the catchment area of a large urban children’s hospital located in the Midwest region of the US. Inclusion criteria were: (1) female adolescents aged 12–16; (2) the ability to read and understand English; and (3) a legal guardian or caregiver who could provide written informed consent and participate as an additional informant. The study enrolled a total of 460 adolescent females, 156 with substantiated CSA and 304 non-abused comparisons. Females who experienced CSA were recruited from local county child protective service agencies for having experienced substantiated sexual abuse within the previous 12 months. This 12-month period was deemed necessary to reduce heterogeneity with respect to the timing of CSA victimization. Comparisons were enrolled into two subgroups: DMCs (n = 156) and CMCs (n = 148). DMC females were demographically matched to females who experienced CSA on ethnicity, family income (within US$10,000) and age (within 1 year). According to the 2010 census, CMC females were enrolled to mirror the sociodemographic make-up of the children’s hospital catchment region in terms of household income and ethnicity. DMC and CMC females were recruited from flyers posted in regional outpatient teen health centres, community centres and local middle schools located within the hospital catchment area. At the time of recruitment, verbal consent was obtained from the caregivers of eligible DMC and CMC females to search the statewide child welfare information administrative records system for child protective service involvement of eligible females. Screened through statewide child welfare records, DMC and CMC females were excluded from enrolment if they showed any previous history of sexual abuse (Fig. 1). Collectively, females who experienced CSA and DMC females were mean-aged 14.35 (s.d. = 1.28), reported a median household income of US$20,000–29,000 and had an ethnicity breakdown of 31.4% White, 54.2% African-American, 1.6% Native American, 9.9% Asian, 2.9% multiracial and 4.0% Hispanic. CMC females were of similar age (mean = 14.05; s.d. = 1.17) but were systematically enrolled to mirror the census make-up of the catchment area and thus reported a median household income of US$60,000–69,000 and an ethnicity breakdown of 69.6% White, 21.6% African-American, 1.4% Native American, 6.1% Asian, 1.4% multiracial and 4.0% Hispanic.

Procedures.

Overview.

Figure 1 outlines the sampling and procedural flow for ease of referral. Adolescents and caregivers travelled to dedicated lab spaces within the children’s hospital facilities to complete in-person lab sessions conducted by clinically trained interviewers who were blind to study group designation. The initial lab session lasted approximately 2 h and included structured interviews and questionnaires administered to adolescents and caregivers. At the end of this lab session, participants were provided with a password-protected 15-inch Apple MacBook Pro laptop. The CD-ROM drives on these computers were replaced with a 4G LTE Verizon modem and an accompanying antenna that allowed continuous connectivity and unlimited data usage to ensure that every participant would have unlimited high-speed Internet access anywhere within their catchment area.

Laptops were equipped with software especially designed by the research team and partially based on an existing open-source toolset Squid version 3.1.20 (http://www.squid-cache.org/) modified to include firewall updates and additional complex scripts. Coined ‘infoSquid,’ this software was transparent and non-disruptive to the user while locally recording all URL activity plus YouTube and Netflix keywords. These data were continuously uploaded to a secure server via an encrypted channel. In the event of connectivity issues, data were cached on the laptop and uploaded at the next opportunity. At the time the laptops were assigned, adolescents were asked to use the laptop as their primary Internet surfing device for a four-week period. On returning the laptops, adolescents were asked whether they mainly used the study laptop to access the Internet. Most indicated ‘yes’ while 36.7% indicated ‘no’, most of whom described using either smartphones or tablets in addition to the study laptop (see Supplementary Tables 6–8 for a sensitivity analysis of findings excluding the 36.7% of participants who indicated the use of other devices to access the Internet during the study period).

After approximately four weeks, adolescents and caregivers returned to the lab for a second in-person session to complete questionnaires and return the laptops. The psychosocial interviews were repeated annually for two additional time points at one-year intervals (times 2 and 3). Retention rates were 94.3% and 91.5% for times 2 and 3, respectively. There were no significant differences across the three study groups with regard to retention rates (time 2: χ² = 675.72, P = 0.08; time 3: χ² = 683.61, P = 0.21).

Objective quantification of URL activity in ‘sessions’.

Once data were uploaded, postprocessing automatically commenced where infoSquid aggregated all collected data to determine and record the length of activity in all user-initiated ‘sessions’. A session began each time the user logged on to the computer and lasted until the computer was shut down or if there was consistent non-use for a period of 60 min. When laptops were in continuous use, a new session was delineated every 2 h. A total of 32,323 sessions were available for analysis.

Within each session, the ‘active’ interaction time on each individual URL was recorded in seconds. Active usage was defined as consistent or intermittent interaction with a Web page in terms of page transitions, clicks, refreshes or data inputs. If there was no active interaction with a Web page for longer than 60 min, the time remaining on that Web page, even if it remained open, was considered ‘inactive’. Only active URL activity was quantified for content. URLs were then matched to a commercial blacklist database (Squidblacklist.org) to objectively classify each based on broad category types including ‘social media’, ‘entertainment’, ‘pornography’, ‘education’ and ‘gaming’. Just over 93% of all URLs were classified in at least one of these categories (Fig. 2). This classification approach is commonly used as continuous support for various parental control filtering software such as Cisco Umbrella. Resultant data were quantified as the number of minutes participants spent actively surfing each category type within a given session.

Quantification of YouTube and Netflix videos.

To increase precision in quantifying Web activity on URLs classified as ‘entertainment’, infoSquid ‘drilled down’ to extract keywords associated with videos viewed on the two most common video platform websites: YouTube and Netflix. During the 32,323 sessions, a total of 137,025 YouTube or Netflix videos were viewed. A subset of these videos (22,456) did not have associated keywords. From the remaining 114,569 videos, 102,665 keywords were extracted. Two coders were trained by two senior project coordinators to rate these keywords in terms of their relative appropriateness for teen girls. Keywords were deemed ‘inappropriate’ if they were associated with sex, violence or illicit drug use. Discrepancies were reconciled by group consensus with input from trained coders and project coordinators. A total of 7,544 keywords were judged to be inappropriate, 947 (12.6%) of which required reconciliation. Additional details are provided in the Supplementary Methods.

Session authentication.

Participants were provided with a unique password when laptops were assigned, which was required at power-up to bring the laptop back from sleep mode and/or every 2 h during active surfing. During the first six months of the study, interviewers casually asked whether the password had been shared. Anecdotally, interviewers reported that about half of the adolescents admitted to sharing the passwords. At this point, the research team used machine learning to employ a sophisticated keystroke dynamics authentication algorithm to better ensure that the observed URL activity could indeed be attributable to the adolescent with a greater degree of certainty.

Machine-learning authentication.

At the time of laptop assignment, an authentication protocol was also administered to gather ‘training’ data when adolescents typed a sentence (‘a rat in the house may eat the ice cream’) into the computer 20 times in front of an interviewer during the lab visit. This allowed the capturing of idiosyncratic keystroke data unique to the adolescent on which to authenticate the subsequent typing of the same sentence. During laptop use, adolescents were prompted to type the authentication sentence again at each point when the password was required (that is, at login, at the beginning of each surfing session and every 2 h) to obtain ‘field’ data. When the laptop was returned, the adolescent again typed the same authentication sentence another ten times during the lab visit in front of the interviewer to obtain another set of ‘training’ data. By comparing training data with field data, Gaussian mixture modelling of keystroke dynamics¹¹² was used via machine learning to evaluate sessions that were deemed consistent with the participant versus from other sources with 97% specificity. Additional details regarding the machine-learning authentication procedures are included in the Supplementary Methods and Supplementary Table 9.

Human authentication.

As described above, the study was initially designed with password protection as a means of ensuring that laptops were used exclusively by the participating adolescent. However, several of the first enrollees shared their passwords with others. For sessions where machine-learning authentication was not yet implemented or when there was inadequate reliability data, a data visualization interface was developed to facilitate manual, human evaluation of authentication. This interface was designed to aid human coders in judging whether patterns of Internet activity were consistent with that of the participant and included: (1) time and type of day. Activity occurring during sessions inconsistent with typical use by the adolescent was judged to be unlikely that of the participant, for example, Web activity during school hours on a day when it was confirmed that school was in session; (2) activity during confirmed Facebook logins. As a means to keep in touch with participants over the course of the longitudinal study, adolescents confirmed their most consistent Facebook login name(s) and ‘friended’ the study Facebook page so that the study team could send notifications and private messages about study updates and visit reminders. Activity occurring during sessions when Facebook logins were confirmed was judged to be consistent with that of the participant; (3) Web activity. Word clouds were produced for each session as visual tools to characterize time spent on three distinct Web activities: URLs visited (by type); search engine search terms; and video keywords. Trained coders judged each session according to these visualization tools (see Supplementary Methods for the details regarding coder training and reliability).

Authentication results.

Results from both machine-learning and human authentication indicated that 26.3% of the 32,323 sessions included activity deemed inconsistent with the participating adolescent and were excluded, resulting in a total of 23,839 authenticated sessions that were used in subsequent analyses. The rates of authentication were similar between machine-learning and human authentication (71.8 and 78.3%, respectively; Phi coefficient = 0.06). However, of the participants receiving human authentication, significantly more (56.0%) were in the DMC group compared to 29.0 and 15.0% in the CSA and CMC groups, respectively (F_(2,257) = 16.51; P < 0.01). Due to this difference, a binary variable defining ‘authentication mode’ (1 = human authentication; 0 = machine-learning authentication) was added to each set of analyses as a covariate.

Variables and variable definitions.

URL activity.

For each of the URL activity variables defined above, activity in all authenticated sessions was summed and then averaged across all days when there was laptop activity; days when there was zero URL activity were not included in this denominator. The average number of active days was 21.21 (s.d. = 7.98) with no statistical difference across the CSA, DMC and CMC subgroups. One participant did not have any authenticated sessions and was dropped from final models (n for analyses = 459).

Psychosocial risk and protective factors assessed at time 1.

Psychosocial factors were assessed via widely used and validated measures as detailed in the Supplementary Methods. Psychosocial measures included substance use, depression, impulse control, executive functioning, self-esteem, school engagement (grades), prosocial activities and parental quality.

Outcomes for longitudinal follow-up analyses.

When time 3 was unavailable due to attrition, the same variables assessed at time 2 (1-year after time 1) were considered the outcome. This was the case for 23 participants, 10 from CSA, 6 from DMC and 7 from CMC with no significant differences across these groups. As detailed in the Supplementary Methods, incidences of Internet-initiated victimization (that is, being cyberbullied, online sexual solicitations and meeting strangers offline) were assessed via a validated, semistructured interview conducted by trained clinical interviewers. Offline sexual activities were assessed via a valid, standardized questionnaire delivered via multi-media computers with headphones to maximize anonymity (Supplementary Methods).

Covariates.

Family income level, minority ethnicity status, URL authentication mode, contamination (that is, sexual abuse occurring in the comparison groups) and other types of child maltreatment experiences were used as covariates in the analyses (see Supplementary Methods for full details).

Analytical plan.

All analyses were performed in R¹¹³ v.3.5.1 and v.4.0.4 or SPSS v.26 for Mac. All tests were two-sided. This study was powered to detect small-to-moderate effect sizes (0.10–0.20). A power analysis using G*Power¹¹⁴ with α = .05 and power = 0.85 indicated a required total sample size of 460 to detect small-to-moderate group difference effects in the three subgroups with up to 6 covariates.

A Shapiro–Wilk test for normality showed that, with the exception of the variable measuring impulse control, all other continuous variables (including Internet-use variables and psychosocial risk/protective factors) were sufficiently non-normally distributed at P < 0.05. Levene’s test for heterogeneity of variance showed that, with the exception of entertainment, games, impulse control, parental quality and self-esteem, homogeneity of variances was violated across the three subgroups (CSA, DMC and CMC). With the exception of impulse control, prosocial activities and self-esteem, homogeneity of variances was violated across the LPA profiles (profiles 1, 2 and 3) at P < 0.05. With tests of normality and homogeneity of variance violated, robust tests were conducted using bootstrapping (1,000 iterations) of means, F tests and post-hoc pairwise comparisons (Sidak-adjusted for multiple comparisons). Due to sufficiently small proportions of missingness for all study variables (missingness = 1.1%; Supplementary Table 1), data were assumed to be missing at random and the bias introduced by missingness was assumed to be minimal¹¹⁵.

Descriptive statistics.

A Venn diagram was produced as a visual representation of how various categories of URL activity were represented in the 23,839 authenticated sessions. Demographic differences across subgroups were evaluated via t and χ² tests. With covariates (time 1 age, income, minority ethnicity, authentication mode, contamination and other types of maltreatment) controlled, an omnibus multivariate analysis of covariance (MANCOVA) model was used to describe the sample in terms of CSA versus comparison differences for the variables used in the analyses. In general, overall MANCOVAs were followed by univariate bootstrapped analysis of covariance (ANCOVA) models to contrast the three study subgroups (CSA, DMC and CMC). These three subgroup contrasts evaluated whether observed CSA versus comparison differences could be attributed to the experience of sexual abuse per se or merely to sociodemographic influences. For example, if the three subgroup contrasts showed that CSA was significantly different from both DMC and CMC, then the CSA versus comparison difference observed in the omnibus tests could be attributed to the experience of CSA independently of any hidden sociodemographic bias that might be introduced by demographic matching procedures.

LPA.

LPA is a widely used tool in the developmental sciences for characterizing large multivariate systems spanning multiple levels of functioning^116,117. Unlike more traditional, variable-centred approaches (for example, general linear models) that estimate the main effects of independent variables and hold each constant in the prediction of a dependent variable, LPA is a person-centred approach that characterizes individuals into empirically derived like-groupings, or profiles, based on both mean levels and interactions among large sets of independent variables. Because the resultant small subsets of people then supplant individual variables for further analysis, LPA is highly useful for summarizing the cumulative and interactive effects operating within a large multivariate system^116,117. While general linear model approaches are valuable when parsing out unique variation, when large sets of independent variables are considered, these models can be overly conservative in terms of illuminating important features of main effects and can fail to capture the configurations of independent variables that might jointly explain behavioural processes. Relative to bivariate, multiple regression and cumulative risk index approaches, personalized approaches have been shown to provide added value to the quantification of risk in several domains of child development⁸³. LPA also makes no assumptions about linearity among independent variables or interactions.

Using the R mclust package v.5.4.5 (ref. ¹⁰¹), LPA was used to identify groups of adolescents who showed similar patterns of Internet-use variables coupled with co-occurring individual-level psychosocial risk variables (substance use, depression, low impulse control, poor executive functioning and self-esteem) and family/community-level protective factors (grades, offline prosocial activities and parental relationship quality). Based on finite Gaussian mixture modelling, LPA models are fitted in a series of steps starting with a one-profile model and subsequently increasing the number of profiles until there is no further improvement in the model. The mclust package allows one to fit profiles with covariances differing in their distribution structure type, volume, shape and orientation¹⁰¹. Based on simulations¹¹⁸, BIC and BLRT are the best indicators to determine the optimal number of profiles in the studied sample. In mclust, higher BIC values represent better-fitting models. The BLRT compares model fit between k−1 and k profile models. A P < 0.05 indicates that the specified model k provides better fit to the data than the model with one fewer profile. Values of the ICL¹¹⁹ were compared to discern the number of profiles to extract. Finally, entropy, or the average accuracy in assigning individuals correctly to the appropriate profiles, was assessed and shown to range from 0 to1 with values from 0.80 being considered high¹²⁰.

To aid in the interpretation of the LPA results, all input variables used in the analyses were standardized (mean = 0; s.d. = 1) for better comparability within and across the resultant profiles. Difference in these standardized estimated marginal means (EMMs) were evaluated across the profiles via an omnibus MANCOVA model controlling for covariates (time 1 age, income, minority ethnicity and authentication mode), followed by univariate bootstrapped ANCOVAs and post-hoc comparisons corrected for multiple testing using Sidak adjustment to control for type 1 errors.

To empirically describe the distribution of females who experienced CSA across resultant profiles, multinomial logistic regression analyses evaluated the differential odds for being represented in profiles for CSA versus comparisons followed by a multinomial regression model evaluating the differential odds for the three subgroups (CSA, DMC or CMC) controlling for covariates (time 1 age, income, minority ethnicity, authentication mode, contamination and other types of maltreatment).

Analyses for hypothesis 1.

As part of the omnibus MANCOVA tests used in the descriptive analyses described above, CSA versus comparison group differences in pornography use were tested via a univariate bootstrapped ANCOVA controlling for covariates and followed by a post-hoc ANCOVA with the three subgroup comparisons (CSA, DMC and CMC). With CSA as the reference class, binary logistic regression analyses, controlling for covariates, were performed to test the differences in the odds of CSA versus comparisons for experiencing each of the three Internet-initiated victimization dichotomous outcomes: being cyberbullied; receiving online sexual solicitations; and meeting strangers offline.

Analyses for hypothesis 2.

For any LPA profile group with significantly elevated levels of pornography use (relative to other profiles), three binary logistic regression models were performed to investigate whether females in such a profile had increased odds for experiencing the three dichotomous Internet-initiated victimization outcomes (being cyberbullied, receiving online sexual solicitations or meeting strangers offline), controlling for covariates. A similar approach was used to evaluate the continuous outcome (offline sexual activities) with a bootstrapped ANCOVA model followed by post-hoc comparisons to discern mean differences across profiles with covariates controlled and Sidak-adjusted for multiple testing.

As post-hoc tests to LPA, LASSO-regularized regressions¹⁰³ using the R glmnet package version 4.1–1¹²¹ were conducted as exploratory analyses to determine important non-zero features that should be retained to maximally predict outcomes. LASSO regression is a machine-learning approach whereby redundant or non-relevant (in terms of their unique or interactive contribution to the combined prediction) variables are shrunk to zero to optimize the selection of retained variables (or features) that lead to a model where prediction error is minimized. LASSO models included all 15 LPA input variables and were ‘trained’ on a random subset of the data using tenfold cross-validation to select the optimal lambda parameter estimates. To determine model fit and accuracy, a standard accuracy estimate (that is, percentage of overlap between predicted and observed outcomes) was employed as well as estimating the area under the receiving operating characteristics curve (AUC) for binary outcomes (logistic regressions) and the root mean square error for continuous outcomes (standard regressions). These post-hoc tests were performed on models where LPA profiles were shown to be significant predictors in logistic regression or ANCOVA analyses.

Analyses for hypothesis 3.

For any LPA profile group with significantly elevated levels of social media (relative to other profiles), binary logistic regression models were performed to investigate the odds for being cyberbullied controlling for covariates and followed by post-hoc contrasts across profiles. Post-hoc LASSO logistic regression was conducted as an exploratory analysis to determine important non-zero features that should be retained in the models to maximally predict cyberbullying.

Reporting Summary.

Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Extended Data

Supplementary Material

NIHMS1817043-supplement-1.pdf^{(1.3MB, pdf)}

Acknowledgements

J.G.N., M.Kouril and C.E.S. acknowledge support from a grant from the National Institutes of Health (NIH) (grant no. R01HD052533). J.G.N. and C.E.S. acknowledge support from a grant from the NIH (grant no. P50HD089922). The research was also supported by the National Center for Advancing Translational Sciences (grant no. UL1TR001425). M.M. acknowledges support from the National Center for Advancing Translational Sciences, NIH (grant no. 2KL2TR001446-06A1) and the American Foundation for Suicide Prevention (grant no. PRG-0-104-19). We thank J. D. Long, S. Lanza and J. Buchheim for their statistical advice.

Footnotes

Competing interests

The authors declare no competing interests.

Code availability

The data analysis script is available from A.C.H. upon request.

Additional information

Extended data is available for this paper at https://doi.org/10.1038/s41562-021-01187-5.

Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41562-021-01187-5.

Peer review information Nature Human Behaviour thanks Amy Orben and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Reprints and permissions information is available at www.nature.com/reprints.

Data availability

The data reported in the current article are not publicly available because they contain extremely sensitive information that could compromise research participant privacy and confidentiality. We cannot provide individual-level data from this project due to limits to our confidentiality agreement with participants. Data are available upon request from J.G.N. by qualified scientists. Requests require a concept paper describing the purpose of data access, ethical approval at the applicant’s university in writing and provision for secure data access.

References

1.Anderson M & Jiang J Teens, social media and technology 2018. Pew Research Center https://www.pewinternet.org/2018/05/31/teens-social-media-technology-2018/ (2018).
2.Robison KK & Crenshaw EM Reevaluating the global digital divide: socio-demographic and conflict barriers to the Internet revolution. Sociol. Inq 80, 34–62 (2010). [Google Scholar]
3.Orben A & Przybylski AK Reply to: Underestimating digital media harm. Nat. Hum. Behav 4, 349–351 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Orben A Teenagers, screens and social media: a narrative review of reviews and key studies. Soc. Psychiatry Psychiatr. Epidemiol 55, 407–414 (2020). [DOI] [PubMed] [Google Scholar]
5.Twenge JM, Haidt J, Joiner TE & Campbell WK Underestimating digital media harm. Nat. Hum. Behav 4, 346–348 (2020). [DOI] [PubMed] [Google Scholar]
6.van den Eijnden RJJM, Meerkerk G-J, Vermulst AA, Spijkerman R & Engels RCME Online communication, compulsive Internet use, and psychosocial well-being among adolescents: a longitudinal study. Dev. Psychol 44, 655–665 (2008). [DOI] [PubMed] [Google Scholar]
7.Hunt MG, Marx R, Lipson C & Young J No more FOMO: limiting social media decreases loneliness and depression. J. Soc. Clin. Psychol 37, 751–768 (2018). [Google Scholar]
8.Stavropoulos V, Burleigh TL, Beard CL, Gomez R & Griffiths MD Being there: a preliminary study examining the role of presence in Internet gaming disorder. Int. J. Ment. Health Addict 17, 880–890 (2019). [Google Scholar]
9.Salmela-Aro K, Upadyaya K, Hakkarainen K, Lonka K & Alho K The dark side of Internet use: two longitudinal studies of excessive Internet use, depressive symptoms, school burnout and engagement among Finnish early and late adolescents. J. Youth Adolesc 46, 343–357 (2017). [DOI] [PubMed] [Google Scholar]
10.Mazzer K, Bauducco S, Linton SJ & Boersma K Longitudinal associations between time spent using technology and sleep duration among adolescents. J. Adolesc 66, 112–119 (2018). [DOI] [PubMed] [Google Scholar]
11.Borca G, Bina M, Keller P, Gilbert LR & Begotti T Internet use and developmental tasks: adolescents’ point of view. Comput. Hum. Behav 52, 49–58 (2015). [Google Scholar]
12.Best P, Manktelow R & Taylor B Online communication, social media and adolescent wellbeing: a systematic narrative review. Child. Youth Serv. Rev 41, 27–36 (2014). [Google Scholar]
13.Subrahmanyam K, Smahel D & Greenfield P Connecting developmental constructions to the Internet: identity presentation and sexual exploration in online teen chat rooms. Dev. Psychol 42, 395–406 (2006). [DOI] [PubMed] [Google Scholar]
14.Ellison NB, Steinfield C & Lampe C Connection strategies: social capital implications of Facebook-enabled communication practices. New Media Soc. 13, 873–892 (2011). [Google Scholar]
15.Thom RP, Bickham DS & Rich M Internet use, depression, and anxiety in a healthy adolescent population: prospective cohort study. JMIR Mhealth Uhealth 5, e44 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Dolev-Cohen M & Barak A Adolescents’ use of instant messaging as a means of emotional relief. Comput. Hum. Behav 29, 58–63 (2013). [Google Scholar]
17.Odgers CL & Jensen MR Annual research review: adolescent mental health in the digital age: facts, fears, and future directions. J. Child Psychol. Psychiatry 61, 336–348 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Orben A & Przybylski AK The association between adolescent well-being and digital technology use. Nat. Hum. Behav 3, 173–182 (2019). [DOI] [PubMed] [Google Scholar]
19.Maas MK, Bray BC & Noll JG Online sexual experiences predict subsequent sexual health and victimization outcomes among female adolescents: a latent class analysis. J. Youth Adolesc 48, 837–849 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
20.Mitchell KJ, Finkelhor D & Wolak J Online requests for sexual pictures from youth: risk factors and incident characteristics. J. Adolesc. Health 41, 196–203 (2007). [DOI] [PubMed] [Google Scholar]
21.Negriff S & Valente TW Structural characteristics of the online social networks of maltreated youth and offline sexual risk behavior. Child Abuse Negl. 85, 209–219 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Noll JG, Shenk CE, Barnes JE & Haralson KJ Association of maltreatment with high-risk Internet behaviors and offline encounters. Pediatrics 131, e510–e517 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Helweg-Larsen K, Schütt N & Larsen HB Predictors and protective factors for adolescent Internet victimization: results from a 2008 nationwide Danish youth survey. Acta Paediatr. 101, 533–539 (2012). [DOI] [PubMed] [Google Scholar]
24.Mitchell KJ, Finkelhor D & Wolak J Youth Internet users at risk for the most serious online sexual solicitations. Am. J. Prev. Med 32, 532–537 (2007). [DOI] [PubMed] [Google Scholar]
25.Noll JG et al. Childhood sexual abuse and early timing of puberty. J. Adolesc. Health 60, 65–71 (2017). [DOI] [PubMed] [Google Scholar]
26.Noll JG et al. Receptive language and educational attainment for sexually abused females. Pediatrics 126, e615–e622 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Trickett PK, Noll JG & Putnam FW The impact of sexual abuse on female development: lessons from a multigenerational, longitudinal research study. Dev. Psychopathol 23, 453–476 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Noll JG et al. Is sexual abuse a unique predictor of sexual risk behaviors, pregnancy, and motherhood in adolescence? J. Res. Adolesc 29, 967–983 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Browne A & Finkelhor D Impact of child sexual abuse: a review of the research. Psychol. Bull 99, 66–77 (1986). [PubMed] [Google Scholar]
30.Negriff S, Schneiderman JU & Trickett PK Child maltreatment and sexual risk behavior: maltreatment types and gender differences. J. Dev. Behav. Pediatr 36, 708–716 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Noll JG, Shenk CE & Putnam KT Childhood sexual abuse and adolescent pregnancy: a meta-analytic update. J. Pediatr. Psychol 34, 366–378 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Widom CS & Kuhns JB Childhood victimization and subsequent risk for promiscuity, prostitution, and teenage pregnancy: a prospective study. Am. J. Public Health 86, 1607–1612 (1996). [DOI] [PMC free article] [PubMed] [Google Scholar]
33.Wilson HW & Widom CS Sexually transmitted diseases among adults who had been abused and neglected as children: a 30-year prospective study. Am. J. Public Health 99, S197–S203 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
34.Noll JG, Trickett PK & Putnam FW A prospective investigation of the impact of childhood sexual abuse on the development of sexuality. J. Consult. Clin. Psychol 71, 575–586 (2003). [DOI] [PMC free article] [PubMed] [Google Scholar]
35.Burton DL, Leibowitz GS & Howard A Comparison by crime type of juvenile delinquents on pornography exposure: the absence of relationships between exposure to pornography and sexual offense characteristics. J. Forensic Nurs 6, 121–129 (2010). [DOI] [PubMed] [Google Scholar]
36.Collins RL et al. Sexual media and childhood well-being and health. Pediatrics 140, S162–S166 (2017). [DOI] [PubMed] [Google Scholar]
37.Doornwaard SM et al. Sex-related online behaviors and adolescents’ body and sexual self-perceptions. Pediatrics 134, 1103–1110 (2014). [DOI] [PubMed] [Google Scholar]
38.Kohut T & Štulhofer A Is pornography use a risk for adolescent well-being? An examination of temporal relationships in two independent panel samples. PLoS ONE 13, e0202048 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Owens EW, Behun RJ, Manning JC & Reid RC The impact of Internet pornography on adolescents: a review of the research. Sex. Addict. Compulsivity 19, 99–122 (2012). [Google Scholar]
40.Cheng S, Ma J & Missari S The effects of Internet use on adolescents’ first romantic and sexual relationships in Taiwan. Int. Sociol 29, 324–347 (2014). [Google Scholar]
41.Madigan S, Ly A, Rash CL, Ouytsel JV & Temple JR Prevalence of multiple forms of sexting behavior among youth: a systematic review and meta-analysis. JAMA Pediatr. 172, 327–335 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
42.Peter J & Valkenburg PM Adolescents and pornography: a review of 20 years of research. J. Sex Res 53, 509–531 (2016). [DOI] [PubMed] [Google Scholar]
43.Doornwaard SM, van den Eijnden RJJM, Baams L, Vanwesenbeeck I & ter Bogt TFM Lower psychological well-being and excessive sexual interest predict symptoms of compulsive use of sexually explicit Internet material among adolescent boys. J. Youth Adolesc 45, 73–84 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
44.van Oosten JMF Sexually explicit Internet material and adolescents’ sexual uncertainty: the role of disposition-content congruency. Arch. Sex. Behav 45, 1011–1022 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
45.Brown JD & L’Engle KL X-rated: sexual attitudes and behaviors associated with U.S. early adolescents’ exposure to sexually explicit media. Commun. Res 36, 129–151 (2009). [Google Scholar]
46.Messman-Moore TL & Long PJ The role of childhood sexual abuse sequelae in the sexual revictimization of women: an empirical review and theoretical reformulation. Clin. Psychol. Rev 23, 537–571 (2003). [DOI] [PubMed] [Google Scholar]
47.Shields A & Cicchetti D Parental maltreatment and emotion dysregulation as risk factors for bullying and victimization in middle childhood. J. Clin. Child Psychol 30, 349–363 (2001). [DOI] [PubMed] [Google Scholar]
48.Barnes JE, Noll JG, Putnam FW & Trickett PK Sexual and physical revictimization among victims of severe childhood sexual abuse. Child Abuse Negl. 33, 412–420 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]
49.Modecki KL, Minchin J, Harbaugh AG, Guerra NG & Runions KC Bullying prevalence across contexts: a meta-analysis measuring cyber and traditional bullying. J. Adolesc. Health 55, 602–611 (2014). [DOI] [PubMed] [Google Scholar]
50.Cosma A et al. Bullying victimization: time trends and the overlap between traditional and cyberbullying across countries in Europe and North America. Int. J. Public Health 65, 75–85 (2020). [DOI] [PubMed] [Google Scholar]
51.Hébert M, Cénat JM, Blais M, Lavoie F & Guerrier M Child sexual abuse, bullying, cyberbullying, and mental health problems among high school students: a moderated mediated model. Depress Anxiety 33, 623–629 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
52.Viner RM et al. Roles of cyberbullying, sleep, and physical activity in mediating the effects of social media use on mental health and wellbeing among young people in England: a secondary analysis of longitudinal data. Lancet Child Adolesc. Health 3, 685–696 (2019). [DOI] [PubMed] [Google Scholar]
53.Roodman AA & Clum GA Revictimization rates and method variance: a meta-analysis. Clin. Psychol. Rev 21, 183–204 (2001). [DOI] [PubMed] [Google Scholar]
54.Lederer LJ & Wetzel CA The health consequences of sex trafficking and their implications for identifying victims in healthcare facilities. Ann. Health Law 23, 61–91 (2014). [Google Scholar]
55.boyd D & Hargittai E Connected and concerned: variation in parents’ online safety concerns. Policy Internet 5, 245–269 (2013). [Google Scholar]
56.Cole DA et al. Longitudinal and incremental relation of cybervictimization to negative self-cognitions and depressive symptoms in young adolescents. J. Abnorm. Child Psychol 44, 1321–1332 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
57.Garett R, Lord LR & Young SD Associations between social media and cyberbullying: a review of the literature. Mhealth 2, 46 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
58.Hamm MP et al. Prevalence and effect of cyberbullying on children and young people: a scoping review of social media studies. JAMA Pediatr. 169, 770–777 (2015). [DOI] [PubMed] [Google Scholar]
59.Dowdell EB, Burgess AW & Flores JR Online social networking patterns among adolescents, young adults, and sexual offenders. Am. J. Nurs 111, 28–36 (2011). [DOI] [PubMed] [Google Scholar]
60.Malesky LA Jr. Predatory online behavior: modus operandi of convicted sex offenders in identifying potential victims and contacting minors over the Internet. J. Child Sex. Abus 16, 23–32 (2007). [DOI] [PubMed] [Google Scholar]
61.Black P, Wollis M, Woodworth M & Hancock JT A linguistic analysis of grooming strategies of online child sex offenders: implications for our understanding of predatory sexual behavior in an increasingly computer-mediated world. Child Abuse Negl 44, 140–149 (2015). [DOI] [PubMed] [Google Scholar]
62.Whittle H, Hamilton-Giachritsis C, Beech A & Collings G A review of online grooming: characteristics and concerns. Aggress. Violent Behav 18, 62–70 (2013). [Google Scholar]
63.Wolak J, Finkelhor D, Mitchell KJ & Ybarra ML Online ‘predators’ and their victims: myths, realities, and implications for prevention and treatment. Am. Psychol 63, 111–128 (2008). [DOI] [PubMed] [Google Scholar]
64.Lorenzo-Dus N, Izura C & Pérez-Tattam R Understanding grooming discourse in computer-mediated environments. Discourse Context Media 12, 40–50 (2016). [Google Scholar]
65.Marcum CD Interpreting the intentions of Internet predators: an examination of online predatory behavior. J. Child Sex. Abus 16, 99–114 (2007). [DOI] [PubMed] [Google Scholar]
66.Livingstone S & Helsper EJ Children, internet and risk in comparative perspective. J. Child. Media 7, 1–8 (2013). [Google Scholar]
67.Noll JG Child sexual abuse as a unique risk factor for the development of psychopathology: the compounded convergence of mechanisms. Annu. Rev. Clin. Psychol 17, 439–464 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
68.Nooner KB et al. Factors related to posttraumatic stress disorder in adolescence. Trauma Violence Abuse 13, 153–166 (2012). [DOI] [PubMed] [Google Scholar]
69.Lauterbach D & Armour C Symptom trajectories among child survivors of maltreatment: findings from the Longitudinal Studies of Child Abuse and Neglect (LONGSCAN). J. Abnorm. Child Psychol 44, 369–379 (2016). [DOI] [PubMed] [Google Scholar]
70.Collishaw S et al. Resilience to adult psychopathology following childhood maltreatment: evidence from a community sample. Child Abuse Negl. 31, 211–229 (2007). [DOI] [PubMed] [Google Scholar]
71.Mitchell KJ, Wolak J & Finkelhor D Trends in youth reports of sexual solicitations, harassment and unwanted exposure to pornography on the Internet. J. Adolesc. Health 40, 116–126 (2007). [DOI] [PubMed] [Google Scholar]
72.Livingstone S & Helsper E Balancing opportunities and risks in teenagers’ use of the internet: the role of online skills and internet self-efficacy. New Media Soc. 12, 309–329 (2009). [Google Scholar]
73.George MJ et al. Young adolescents’ digital technology use, perceived impairments, and well-being in a representative sample. J. Pediatr 219, 180–187 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
74.Mitchell KJ, Ybarra M & Finkelhor D The relative importance of online victimization in understanding depression, delinquency, and substance use. Child Maltreat. 12, 314–324 (2007). [DOI] [PubMed] [Google Scholar]
75.Díaz KI & Fite PJ Cyber victimization and its association with substance use, anxiety, and depression symptoms among middle school youth. Child Youth Care Forum 48, 529–544 (2019). [Google Scholar]
76.Hemphill SA, Tollit M, Kotevski A & Heerde JA Predictors of traditional and cyber-bullying victimization: a longitudinal study of Australian secondary school students. J. Interpers. Violence 30, 2567–2590 (2015). [DOI] [PubMed] [Google Scholar]
77.Zych I, Farrington DP & Ttofi MM Protective factors against bullying and cyberbullying: a systematic review of meta-analyses. Aggress. Violent Behav 45, 4–19 (2019). [Google Scholar]
78.de Santisteban P & Gámez-Guadix M Prevalence and risk factors among minors for online sexual solicitations and interactions with adults. J. Sex Res 55, 939–950 (2018). [DOI] [PubMed] [Google Scholar]
79.Ferrari M & Schick A Teenagers, screens and social media: a commentary on Orben’s narrative review. Soc. Psychiatry Psychiatr. Epidemiol 55, 973–975 (2020). [DOI] [PubMed] [Google Scholar]
80.Hillis S, Mercy J, Amobi A & Kress H Global prevalence of past-year violence against children: a systematic review and minimum estimates. Pediatrics 137, e20154079 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
81.Appleyard K, Egeland B, van Dulmen MH & Sroufe LA When more is not better: the role of cumulative risk in child behavior outcomes. J. Child Psychol. Psychiatry 46, 235–245 (2005). [DOI] [PubMed] [Google Scholar]
82.Deater-Deckard K, Dodge KA, Bates JE & Pettit GS Multiple risk factors in the development of externalizing behavior problems: group and individual differences. Dev. Psychopathol 10, 469–493 (1998). [DOI] [PMC free article] [PubMed] [Google Scholar]
83.Lanza ST, Rhoades BL, Greenberg MT & Cox M Modeling multiple risks during infancy to predict quality of the caregiving environment: contributions of a person-centered approach. Infant Behav. Dev 34, 390–406 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
84.Görzig A Adolescents’ viewing of suicide-related Web content and psychological problems: differentiating the roles of cyberbullying involvement. Cyberpsychol. Behav. Soc. Netw 19, 502–509 (2016). [DOI] [PubMed] [Google Scholar]
85.Williams KR & Guerra NG Prevalence and predictors of Internet bullying. J. Adolesc. Health 41, S14–S21 (2007). [DOI] [PubMed] [Google Scholar]
86.Kardefelt-Winther D & Maternowska C Addressing violence against children online and offline. Nat. Hum. Behav 4, 227–230 (2020). [DOI] [PubMed] [Google Scholar]
87.Finkelhor D, Walsh K, Jones L, Mitchell K & Collier A Youth Internet safety education: aligning programs with the evidence base. Trauma Violence Abuse, 10.1177/1524838020916257 (2020). [DOI] [PubMed] [Google Scholar]
88.Prinstein MJ, Nesi J & Telzer EH Commentary: an updated agenda for the study of digital media use and adolescent development—future directions following Odgers & Jensen (2020). J. Child Psychol. Psychiatry 61, 349–352 (2020). [DOI] [PubMed] [Google Scholar]
89.Boase J & Ling R Measuring mobile phone use: self-report versus log data. J. Comput. Mediat. Commun 18, 508–519 (2013). [Google Scholar]
90.Gold JE, Rauscher KJ & Zhu M A validity study of self-reported daily texting frequency, cell phone characteristics, and texting styles among young adults. BMC Res. Notes 8, 120 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
91.Mireku M et al. Total recall in the SCAMP cohort: validation of self-reported mobile phone use in the smartphone era. Environ. Res 161, 1–8 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
92.Scharkow M The accuracy of self-reported Internet use—a validation study using client log data. Commun. Methods Meas 10, 13–27 (2016). [Google Scholar]
93.Finkelhor D, Shattuck A, Turner HA & Hamby SL The lifetime prevalence of child sexual abuse and sexual assault assessed in late adolescence. J. Adolesc. Health 55, 329–333 (2014). [DOI] [PubMed] [Google Scholar]
94.Twenge JM More time on technology, less happiness? Associations between digital-media use and psychological well-being. Curr. Dir. Psychol. Sci 28, 372–379 (2019). [Google Scholar]
95.Tener D, Wolak J & Finkelhor D A typology of offenders who use online communications to commit sex crimes against minors. J. Aggress. Maltreat. Trauma 24, 319–337 (2015). [Google Scholar]
96.Deshpande NA & Nour NM Sex trafficking of women and girls. Rev. Obstet. Gynecol 6, e22–e27 (2013). [PMC free article] [PubMed] [Google Scholar]
97.Rubin DB Matching to remove bias in observational studies. Biometrics 29, 159–183 (1973). [Google Scholar]
98.Petersen AC, Feit MN & Joseph J New Directions in Child Abuse and Neglect Research (The National Academies Press, 2014). [PubMed] [Google Scholar]
99.Rosenbaum PR Discussing hidden bias in observational studies. Ann. Intern. Med 115, 901–905 (1991). [DOI] [PubMed] [Google Scholar]
100.Rosenbaum PR Impact of multiple matched controls on design sensitivity in observational studies. Biometrics 69, 118–127 (2013). [DOI] [PubMed] [Google Scholar]
101.Scrucca L, Fop M, Murphy TB & Raftery AE mclust 5: clustering, classification and density estimation using Gaussian finite mixture models. R. J 8, 289–317 (2016). [PMC free article] [PubMed] [Google Scholar]
102.Dodge KA Annual research review: universal and targeted strategies for assigning interventions to achieve population impact. J. Child Psychol. Psychiatry 61, 255–267 (2020). [DOI] [PubMed] [Google Scholar]
103.Tibshirani R Regression shrinkage and selection via the lasso. J. R. Stat. Soc. B 58, 267–288 (1996). [Google Scholar]
104.Twenge JM & Farley E Not all screen time is created equal: associations with mental health vary by activity and gender. Soc. Psychiatry Psychiatr. Epidemiol 56, 207–217 (2021). [DOI] [PubMed] [Google Scholar]
105.Lenhart A Teens, social media & technology overview 2015. Pew Research Center https://www.pewresearch.org/internet/2015/04/09/teens-social-media-technology-2015/ (2015).
106.Galatzer-Levy IR, Huang SH & Bonanno GA Trajectories of resilience and dysfunction following potential trauma: a review and statistical evaluation. Clin. Psychol. Rev 63, 41–55 (2018). [DOI] [PubMed] [Google Scholar]
107.Cohen JA & Mannarino AP Trauma-focused cognitive behavior therapy for traumatized children and families. Child Adolesc. Psychiatr. Clin. N. Am 24, 557–570 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
108.Mathews B New International Frontiers in Child Sexual Abuse: Theory, Problems and Progress (Springer International Publishing, 2019). [Google Scholar]
109.Barth J, Bermetz L, Heim E, Trelle S & Tonia T The current prevalence of child sexual abuse worldwide: a systematic review and meta-analysis. Int. J. Public Health 58, 469–483 (2013). [DOI] [PubMed] [Google Scholar]
110.U.S. Department of Health and Human Services, Children’s Bureau. Child Maltreatment 2018 (U.S. Government Printing Office, 2020). [Google Scholar]
111.Finkelhor D, Saito K & Jones L Updated Trends in Child Maltreatment, 2018. (Crimes Against Children Research Center, 2020); http://unh.edu/ccrc/pdf/CV203%20-%20Updated%20trends%202018_ks_df.pdf [Google Scholar]
112.Hosseinzadeh D, Krishnan S & Khademi A Keystroke identification based on Gaussian mixture models. In Proc. 2006 IEEE International Conference on Acoustics, Speech, and Signal Processing III.1144–III.1147 (2006). [Google Scholar]
113.R: a language and environment for statistical computing (R Foundation for Statistical Computing, 2021). [Google Scholar]
114.Faul F, Erdfelder E, Lang A-G & Buchner A G*Power 3: a flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behav. Res. Methods 39, 175–191 (2007). [DOI] [PubMed] [Google Scholar]
115.Dong Y & Peng CJ Principled missing data methods for researchers. Springerplus 2, 222 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
116.Gibson WA Three multivariate models: factor analysis, latent structure analysis and latent profile analysis. Psychometrika 24, 229–252 (1959). [Google Scholar]
117.Lanza ST Latent class analysis for developmental research. Child Dev. Perspect 10, 59–64 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
118.Nylund KL, Asparouhov T & Muthén BO Deciding on the number of classes in latent class analysis and growth mixture modeling: a Monte Carlo simulation study. Struct. Equ. Modeling 14, 535–569 (2007). [Google Scholar]
119.Bertoletti M, Friel N & Rastelli R Choosing the number of clusters in a finite mixture model using an exact integrated completed likelihood criterion. Metron 73, 177–199 (2015). [Google Scholar]
120.Clark S & Muthén B Relating Latent Class Analysis Results to Variables Not Included in the Analysis (2009); http://www.statmodel.com/download/relatinglca.pdf
121.Friedman J, Hastie T & Tibshirani R Regularizing paths for generalized linear models via coordinate descent. J. Stat. Softw 33, 1–22 (2010). [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

NIHMS1817043-supplement-1.pdf^{(1.3MB, pdf)}

Data Availability Statement

[R1] 1.Anderson M & Jiang J Teens, social media and technology 2018. Pew Research Center https://www.pewinternet.org/2018/05/31/teens-social-media-technology-2018/ (2018).

[R2] 2.Robison KK & Crenshaw EM Reevaluating the global digital divide: socio-demographic and conflict barriers to the Internet revolution. Sociol. Inq 80, 34–62 (2010). [Google Scholar]

[R3] 3.Orben A & Przybylski AK Reply to: Underestimating digital media harm. Nat. Hum. Behav 4, 349–351 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R4] 4.Orben A Teenagers, screens and social media: a narrative review of reviews and key studies. Soc. Psychiatry Psychiatr. Epidemiol 55, 407–414 (2020). [DOI] [PubMed] [Google Scholar]

[R5] 5.Twenge JM, Haidt J, Joiner TE & Campbell WK Underestimating digital media harm. Nat. Hum. Behav 4, 346–348 (2020). [DOI] [PubMed] [Google Scholar]

[R6] 6.van den Eijnden RJJM, Meerkerk G-J, Vermulst AA, Spijkerman R & Engels RCME Online communication, compulsive Internet use, and psychosocial well-being among adolescents: a longitudinal study. Dev. Psychol 44, 655–665 (2008). [DOI] [PubMed] [Google Scholar]

[R7] 7.Hunt MG, Marx R, Lipson C & Young J No more FOMO: limiting social media decreases loneliness and depression. J. Soc. Clin. Psychol 37, 751–768 (2018). [Google Scholar]

[R8] 8.Stavropoulos V, Burleigh TL, Beard CL, Gomez R & Griffiths MD Being there: a preliminary study examining the role of presence in Internet gaming disorder. Int. J. Ment. Health Addict 17, 880–890 (2019). [Google Scholar]

[R9] 9.Salmela-Aro K, Upadyaya K, Hakkarainen K, Lonka K & Alho K The dark side of Internet use: two longitudinal studies of excessive Internet use, depressive symptoms, school burnout and engagement among Finnish early and late adolescents. J. Youth Adolesc 46, 343–357 (2017). [DOI] [PubMed] [Google Scholar]

[R10] 10.Mazzer K, Bauducco S, Linton SJ & Boersma K Longitudinal associations between time spent using technology and sleep duration among adolescents. J. Adolesc 66, 112–119 (2018). [DOI] [PubMed] [Google Scholar]

[R11] 11.Borca G, Bina M, Keller P, Gilbert LR & Begotti T Internet use and developmental tasks: adolescents’ point of view. Comput. Hum. Behav 52, 49–58 (2015). [Google Scholar]

[R12] 12.Best P, Manktelow R & Taylor B Online communication, social media and adolescent wellbeing: a systematic narrative review. Child. Youth Serv. Rev 41, 27–36 (2014). [Google Scholar]

[R13] 13.Subrahmanyam K, Smahel D & Greenfield P Connecting developmental constructions to the Internet: identity presentation and sexual exploration in online teen chat rooms. Dev. Psychol 42, 395–406 (2006). [DOI] [PubMed] [Google Scholar]

[R14] 14.Ellison NB, Steinfield C & Lampe C Connection strategies: social capital implications of Facebook-enabled communication practices. New Media Soc. 13, 873–892 (2011). [Google Scholar]

[R15] 15.Thom RP, Bickham DS & Rich M Internet use, depression, and anxiety in a healthy adolescent population: prospective cohort study. JMIR Mhealth Uhealth 5, e44 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R16] 16.Dolev-Cohen M & Barak A Adolescents’ use of instant messaging as a means of emotional relief. Comput. Hum. Behav 29, 58–63 (2013). [Google Scholar]

[R17] 17.Odgers CL & Jensen MR Annual research review: adolescent mental health in the digital age: facts, fears, and future directions. J. Child Psychol. Psychiatry 61, 336–348 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R18] 18.Orben A & Przybylski AK The association between adolescent well-being and digital technology use. Nat. Hum. Behav 3, 173–182 (2019). [DOI] [PubMed] [Google Scholar]

[R19] 19.Maas MK, Bray BC & Noll JG Online sexual experiences predict subsequent sexual health and victimization outcomes among female adolescents: a latent class analysis. J. Youth Adolesc 48, 837–849 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R20] 20.Mitchell KJ, Finkelhor D & Wolak J Online requests for sexual pictures from youth: risk factors and incident characteristics. J. Adolesc. Health 41, 196–203 (2007). [DOI] [PubMed] [Google Scholar]

[R21] 21.Negriff S & Valente TW Structural characteristics of the online social networks of maltreated youth and offline sexual risk behavior. Child Abuse Negl. 85, 209–219 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R22] 22.Noll JG, Shenk CE, Barnes JE & Haralson KJ Association of maltreatment with high-risk Internet behaviors and offline encounters. Pediatrics 131, e510–e517 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R23] 23.Helweg-Larsen K, Schütt N & Larsen HB Predictors and protective factors for adolescent Internet victimization: results from a 2008 nationwide Danish youth survey. Acta Paediatr. 101, 533–539 (2012). [DOI] [PubMed] [Google Scholar]

[R24] 24.Mitchell KJ, Finkelhor D & Wolak J Youth Internet users at risk for the most serious online sexual solicitations. Am. J. Prev. Med 32, 532–537 (2007). [DOI] [PubMed] [Google Scholar]

[R25] 25.Noll JG et al. Childhood sexual abuse and early timing of puberty. J. Adolesc. Health 60, 65–71 (2017). [DOI] [PubMed] [Google Scholar]

[R26] 26.Noll JG et al. Receptive language and educational attainment for sexually abused females. Pediatrics 126, e615–e622 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R27] 27.Trickett PK, Noll JG & Putnam FW The impact of sexual abuse on female development: lessons from a multigenerational, longitudinal research study. Dev. Psychopathol 23, 453–476 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R28] 28.Noll JG et al. Is sexual abuse a unique predictor of sexual risk behaviors, pregnancy, and motherhood in adolescence? J. Res. Adolesc 29, 967–983 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R29] 29.Browne A & Finkelhor D Impact of child sexual abuse: a review of the research. Psychol. Bull 99, 66–77 (1986). [PubMed] [Google Scholar]

[R30] 30.Negriff S, Schneiderman JU & Trickett PK Child maltreatment and sexual risk behavior: maltreatment types and gender differences. J. Dev. Behav. Pediatr 36, 708–716 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R31] 31.Noll JG, Shenk CE & Putnam KT Childhood sexual abuse and adolescent pregnancy: a meta-analytic update. J. Pediatr. Psychol 34, 366–378 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R32] 32.Widom CS & Kuhns JB Childhood victimization and subsequent risk for promiscuity, prostitution, and teenage pregnancy: a prospective study. Am. J. Public Health 86, 1607–1612 (1996). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R33] 33.Wilson HW & Widom CS Sexually transmitted diseases among adults who had been abused and neglected as children: a 30-year prospective study. Am. J. Public Health 99, S197–S203 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R34] 34.Noll JG, Trickett PK & Putnam FW A prospective investigation of the impact of childhood sexual abuse on the development of sexuality. J. Consult. Clin. Psychol 71, 575–586 (2003). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R35] 35.Burton DL, Leibowitz GS & Howard A Comparison by crime type of juvenile delinquents on pornography exposure: the absence of relationships between exposure to pornography and sexual offense characteristics. J. Forensic Nurs 6, 121–129 (2010). [DOI] [PubMed] [Google Scholar]

[R36] 36.Collins RL et al. Sexual media and childhood well-being and health. Pediatrics 140, S162–S166 (2017). [DOI] [PubMed] [Google Scholar]

[R37] 37.Doornwaard SM et al. Sex-related online behaviors and adolescents’ body and sexual self-perceptions. Pediatrics 134, 1103–1110 (2014). [DOI] [PubMed] [Google Scholar]

[R38] 38.Kohut T & Štulhofer A Is pornography use a risk for adolescent well-being? An examination of temporal relationships in two independent panel samples. PLoS ONE 13, e0202048 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R39] 39.Owens EW, Behun RJ, Manning JC & Reid RC The impact of Internet pornography on adolescents: a review of the research. Sex. Addict. Compulsivity 19, 99–122 (2012). [Google Scholar]

[R40] 40.Cheng S, Ma J & Missari S The effects of Internet use on adolescents’ first romantic and sexual relationships in Taiwan. Int. Sociol 29, 324–347 (2014). [Google Scholar]

[R41] 41.Madigan S, Ly A, Rash CL, Ouytsel JV & Temple JR Prevalence of multiple forms of sexting behavior among youth: a systematic review and meta-analysis. JAMA Pediatr. 172, 327–335 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R42] 42.Peter J & Valkenburg PM Adolescents and pornography: a review of 20 years of research. J. Sex Res 53, 509–531 (2016). [DOI] [PubMed] [Google Scholar]

[R43] 43.Doornwaard SM, van den Eijnden RJJM, Baams L, Vanwesenbeeck I & ter Bogt TFM Lower psychological well-being and excessive sexual interest predict symptoms of compulsive use of sexually explicit Internet material among adolescent boys. J. Youth Adolesc 45, 73–84 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R44] 44.van Oosten JMF Sexually explicit Internet material and adolescents’ sexual uncertainty: the role of disposition-content congruency. Arch. Sex. Behav 45, 1011–1022 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R45] 45.Brown JD & L’Engle KL X-rated: sexual attitudes and behaviors associated with U.S. early adolescents’ exposure to sexually explicit media. Commun. Res 36, 129–151 (2009). [Google Scholar]

[R46] 46.Messman-Moore TL & Long PJ The role of childhood sexual abuse sequelae in the sexual revictimization of women: an empirical review and theoretical reformulation. Clin. Psychol. Rev 23, 537–571 (2003). [DOI] [PubMed] [Google Scholar]

[R47] 47.Shields A & Cicchetti D Parental maltreatment and emotion dysregulation as risk factors for bullying and victimization in middle childhood. J. Clin. Child Psychol 30, 349–363 (2001). [DOI] [PubMed] [Google Scholar]

[R48] 48.Barnes JE, Noll JG, Putnam FW & Trickett PK Sexual and physical revictimization among victims of severe childhood sexual abuse. Child Abuse Negl. 33, 412–420 (2009). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R49] 49.Modecki KL, Minchin J, Harbaugh AG, Guerra NG & Runions KC Bullying prevalence across contexts: a meta-analysis measuring cyber and traditional bullying. J. Adolesc. Health 55, 602–611 (2014). [DOI] [PubMed] [Google Scholar]

[R50] 50.Cosma A et al. Bullying victimization: time trends and the overlap between traditional and cyberbullying across countries in Europe and North America. Int. J. Public Health 65, 75–85 (2020). [DOI] [PubMed] [Google Scholar]

[R51] 51.Hébert M, Cénat JM, Blais M, Lavoie F & Guerrier M Child sexual abuse, bullying, cyberbullying, and mental health problems among high school students: a moderated mediated model. Depress Anxiety 33, 623–629 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R52] 52.Viner RM et al. Roles of cyberbullying, sleep, and physical activity in mediating the effects of social media use on mental health and wellbeing among young people in England: a secondary analysis of longitudinal data. Lancet Child Adolesc. Health 3, 685–696 (2019). [DOI] [PubMed] [Google Scholar]

[R53] 53.Roodman AA & Clum GA Revictimization rates and method variance: a meta-analysis. Clin. Psychol. Rev 21, 183–204 (2001). [DOI] [PubMed] [Google Scholar]

[R54] 54.Lederer LJ & Wetzel CA The health consequences of sex trafficking and their implications for identifying victims in healthcare facilities. Ann. Health Law 23, 61–91 (2014). [Google Scholar]

[R55] 55.boyd D & Hargittai E Connected and concerned: variation in parents’ online safety concerns. Policy Internet 5, 245–269 (2013). [Google Scholar]

[R56] 56.Cole DA et al. Longitudinal and incremental relation of cybervictimization to negative self-cognitions and depressive symptoms in young adolescents. J. Abnorm. Child Psychol 44, 1321–1332 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R57] 57.Garett R, Lord LR & Young SD Associations between social media and cyberbullying: a review of the literature. Mhealth 2, 46 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R58] 58.Hamm MP et al. Prevalence and effect of cyberbullying on children and young people: a scoping review of social media studies. JAMA Pediatr. 169, 770–777 (2015). [DOI] [PubMed] [Google Scholar]

[R59] 59.Dowdell EB, Burgess AW & Flores JR Online social networking patterns among adolescents, young adults, and sexual offenders. Am. J. Nurs 111, 28–36 (2011). [DOI] [PubMed] [Google Scholar]

[R60] 60.Malesky LA Jr. Predatory online behavior: modus operandi of convicted sex offenders in identifying potential victims and contacting minors over the Internet. J. Child Sex. Abus 16, 23–32 (2007). [DOI] [PubMed] [Google Scholar]

[R61] 61.Black P, Wollis M, Woodworth M & Hancock JT A linguistic analysis of grooming strategies of online child sex offenders: implications for our understanding of predatory sexual behavior in an increasingly computer-mediated world. Child Abuse Negl 44, 140–149 (2015). [DOI] [PubMed] [Google Scholar]

[R62] 62.Whittle H, Hamilton-Giachritsis C, Beech A & Collings G A review of online grooming: characteristics and concerns. Aggress. Violent Behav 18, 62–70 (2013). [Google Scholar]

[R63] 63.Wolak J, Finkelhor D, Mitchell KJ & Ybarra ML Online ‘predators’ and their victims: myths, realities, and implications for prevention and treatment. Am. Psychol 63, 111–128 (2008). [DOI] [PubMed] [Google Scholar]

[R64] 64.Lorenzo-Dus N, Izura C & Pérez-Tattam R Understanding grooming discourse in computer-mediated environments. Discourse Context Media 12, 40–50 (2016). [Google Scholar]

[R65] 65.Marcum CD Interpreting the intentions of Internet predators: an examination of online predatory behavior. J. Child Sex. Abus 16, 99–114 (2007). [DOI] [PubMed] [Google Scholar]

[R66] 66.Livingstone S & Helsper EJ Children, internet and risk in comparative perspective. J. Child. Media 7, 1–8 (2013). [Google Scholar]

[R67] 67.Noll JG Child sexual abuse as a unique risk factor for the development of psychopathology: the compounded convergence of mechanisms. Annu. Rev. Clin. Psychol 17, 439–464 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R68] 68.Nooner KB et al. Factors related to posttraumatic stress disorder in adolescence. Trauma Violence Abuse 13, 153–166 (2012). [DOI] [PubMed] [Google Scholar]

[R69] 69.Lauterbach D & Armour C Symptom trajectories among child survivors of maltreatment: findings from the Longitudinal Studies of Child Abuse and Neglect (LONGSCAN). J. Abnorm. Child Psychol 44, 369–379 (2016). [DOI] [PubMed] [Google Scholar]

[R70] 70.Collishaw S et al. Resilience to adult psychopathology following childhood maltreatment: evidence from a community sample. Child Abuse Negl. 31, 211–229 (2007). [DOI] [PubMed] [Google Scholar]

[R71] 71.Mitchell KJ, Wolak J & Finkelhor D Trends in youth reports of sexual solicitations, harassment and unwanted exposure to pornography on the Internet. J. Adolesc. Health 40, 116–126 (2007). [DOI] [PubMed] [Google Scholar]

[R72] 72.Livingstone S & Helsper E Balancing opportunities and risks in teenagers’ use of the internet: the role of online skills and internet self-efficacy. New Media Soc. 12, 309–329 (2009). [Google Scholar]

[R73] 73.George MJ et al. Young adolescents’ digital technology use, perceived impairments, and well-being in a representative sample. J. Pediatr 219, 180–187 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R74] 74.Mitchell KJ, Ybarra M & Finkelhor D The relative importance of online victimization in understanding depression, delinquency, and substance use. Child Maltreat. 12, 314–324 (2007). [DOI] [PubMed] [Google Scholar]

[R75] 75.Díaz KI & Fite PJ Cyber victimization and its association with substance use, anxiety, and depression symptoms among middle school youth. Child Youth Care Forum 48, 529–544 (2019). [Google Scholar]

[R76] 76.Hemphill SA, Tollit M, Kotevski A & Heerde JA Predictors of traditional and cyber-bullying victimization: a longitudinal study of Australian secondary school students. J. Interpers. Violence 30, 2567–2590 (2015). [DOI] [PubMed] [Google Scholar]

[R77] 77.Zych I, Farrington DP & Ttofi MM Protective factors against bullying and cyberbullying: a systematic review of meta-analyses. Aggress. Violent Behav 45, 4–19 (2019). [Google Scholar]

[R78] 78.de Santisteban P & Gámez-Guadix M Prevalence and risk factors among minors for online sexual solicitations and interactions with adults. J. Sex Res 55, 939–950 (2018). [DOI] [PubMed] [Google Scholar]

[R79] 79.Ferrari M & Schick A Teenagers, screens and social media: a commentary on Orben’s narrative review. Soc. Psychiatry Psychiatr. Epidemiol 55, 973–975 (2020). [DOI] [PubMed] [Google Scholar]

[R80] 80.Hillis S, Mercy J, Amobi A & Kress H Global prevalence of past-year violence against children: a systematic review and minimum estimates. Pediatrics 137, e20154079 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R81] 81.Appleyard K, Egeland B, van Dulmen MH & Sroufe LA When more is not better: the role of cumulative risk in child behavior outcomes. J. Child Psychol. Psychiatry 46, 235–245 (2005). [DOI] [PubMed] [Google Scholar]

[R82] 82.Deater-Deckard K, Dodge KA, Bates JE & Pettit GS Multiple risk factors in the development of externalizing behavior problems: group and individual differences. Dev. Psychopathol 10, 469–493 (1998). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R83] 83.Lanza ST, Rhoades BL, Greenberg MT & Cox M Modeling multiple risks during infancy to predict quality of the caregiving environment: contributions of a person-centered approach. Infant Behav. Dev 34, 390–406 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R84] 84.Görzig A Adolescents’ viewing of suicide-related Web content and psychological problems: differentiating the roles of cyberbullying involvement. Cyberpsychol. Behav. Soc. Netw 19, 502–509 (2016). [DOI] [PubMed] [Google Scholar]

[R85] 85.Williams KR & Guerra NG Prevalence and predictors of Internet bullying. J. Adolesc. Health 41, S14–S21 (2007). [DOI] [PubMed] [Google Scholar]

[R86] 86.Kardefelt-Winther D & Maternowska C Addressing violence against children online and offline. Nat. Hum. Behav 4, 227–230 (2020). [DOI] [PubMed] [Google Scholar]

[R87] 87.Finkelhor D, Walsh K, Jones L, Mitchell K & Collier A Youth Internet safety education: aligning programs with the evidence base. Trauma Violence Abuse, 10.1177/1524838020916257 (2020). [DOI] [PubMed] [Google Scholar]

[R88] 88.Prinstein MJ, Nesi J & Telzer EH Commentary: an updated agenda for the study of digital media use and adolescent development—future directions following Odgers & Jensen (2020). J. Child Psychol. Psychiatry 61, 349–352 (2020). [DOI] [PubMed] [Google Scholar]

[R89] 89.Boase J & Ling R Measuring mobile phone use: self-report versus log data. J. Comput. Mediat. Commun 18, 508–519 (2013). [Google Scholar]

[R90] 90.Gold JE, Rauscher KJ & Zhu M A validity study of self-reported daily texting frequency, cell phone characteristics, and texting styles among young adults. BMC Res. Notes 8, 120 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R91] 91.Mireku M et al. Total recall in the SCAMP cohort: validation of self-reported mobile phone use in the smartphone era. Environ. Res 161, 1–8 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R92] 92.Scharkow M The accuracy of self-reported Internet use—a validation study using client log data. Commun. Methods Meas 10, 13–27 (2016). [Google Scholar]

[R93] 93.Finkelhor D, Shattuck A, Turner HA & Hamby SL The lifetime prevalence of child sexual abuse and sexual assault assessed in late adolescence. J. Adolesc. Health 55, 329–333 (2014). [DOI] [PubMed] [Google Scholar]

[R94] 94.Twenge JM More time on technology, less happiness? Associations between digital-media use and psychological well-being. Curr. Dir. Psychol. Sci 28, 372–379 (2019). [Google Scholar]

[R95] 95.Tener D, Wolak J & Finkelhor D A typology of offenders who use online communications to commit sex crimes against minors. J. Aggress. Maltreat. Trauma 24, 319–337 (2015). [Google Scholar]

[R96] 96.Deshpande NA & Nour NM Sex trafficking of women and girls. Rev. Obstet. Gynecol 6, e22–e27 (2013). [PMC free article] [PubMed] [Google Scholar]

[R97] 97.Rubin DB Matching to remove bias in observational studies. Biometrics 29, 159–183 (1973). [Google Scholar]

[R98] 98.Petersen AC, Feit MN & Joseph J New Directions in Child Abuse and Neglect Research (The National Academies Press, 2014). [PubMed] [Google Scholar]

[R99] 99.Rosenbaum PR Discussing hidden bias in observational studies. Ann. Intern. Med 115, 901–905 (1991). [DOI] [PubMed] [Google Scholar]

[R100] 100.Rosenbaum PR Impact of multiple matched controls on design sensitivity in observational studies. Biometrics 69, 118–127 (2013). [DOI] [PubMed] [Google Scholar]

[R101] 101.Scrucca L, Fop M, Murphy TB & Raftery AE mclust 5: clustering, classification and density estimation using Gaussian finite mixture models. R. J 8, 289–317 (2016). [PMC free article] [PubMed] [Google Scholar]

[R102] 102.Dodge KA Annual research review: universal and targeted strategies for assigning interventions to achieve population impact. J. Child Psychol. Psychiatry 61, 255–267 (2020). [DOI] [PubMed] [Google Scholar]

[R103] 103.Tibshirani R Regression shrinkage and selection via the lasso. J. R. Stat. Soc. B 58, 267–288 (1996). [Google Scholar]

[R104] 104.Twenge JM & Farley E Not all screen time is created equal: associations with mental health vary by activity and gender. Soc. Psychiatry Psychiatr. Epidemiol 56, 207–217 (2021). [DOI] [PubMed] [Google Scholar]

[R105] 105.Lenhart A Teens, social media & technology overview 2015. Pew Research Center https://www.pewresearch.org/internet/2015/04/09/teens-social-media-technology-2015/ (2015).

[R106] 106.Galatzer-Levy IR, Huang SH & Bonanno GA Trajectories of resilience and dysfunction following potential trauma: a review and statistical evaluation. Clin. Psychol. Rev 63, 41–55 (2018). [DOI] [PubMed] [Google Scholar]

[R107] 107.Cohen JA & Mannarino AP Trauma-focused cognitive behavior therapy for traumatized children and families. Child Adolesc. Psychiatr. Clin. N. Am 24, 557–570 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R108] 108.Mathews B New International Frontiers in Child Sexual Abuse: Theory, Problems and Progress (Springer International Publishing, 2019). [Google Scholar]

[R109] 109.Barth J, Bermetz L, Heim E, Trelle S & Tonia T The current prevalence of child sexual abuse worldwide: a systematic review and meta-analysis. Int. J. Public Health 58, 469–483 (2013). [DOI] [PubMed] [Google Scholar]

[R110] 110.U.S. Department of Health and Human Services, Children’s Bureau. Child Maltreatment 2018 (U.S. Government Printing Office, 2020). [Google Scholar]

[R111] 111.Finkelhor D, Saito K & Jones L Updated Trends in Child Maltreatment, 2018. (Crimes Against Children Research Center, 2020); http://unh.edu/ccrc/pdf/CV203%20-%20Updated%20trends%202018_ks_df.pdf [Google Scholar]

[R112] 112.Hosseinzadeh D, Krishnan S & Khademi A Keystroke identification based on Gaussian mixture models. In Proc. 2006 IEEE International Conference on Acoustics, Speech, and Signal Processing III.1144–III.1147 (2006). [Google Scholar]

[R113] 113.R: a language and environment for statistical computing (R Foundation for Statistical Computing, 2021). [Google Scholar]

[R114] 114.Faul F, Erdfelder E, Lang A-G & Buchner A G*Power 3: a flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behav. Res. Methods 39, 175–191 (2007). [DOI] [PubMed] [Google Scholar]

[R115] 115.Dong Y & Peng CJ Principled missing data methods for researchers. Springerplus 2, 222 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R116] 116.Gibson WA Three multivariate models: factor analysis, latent structure analysis and latent profile analysis. Psychometrika 24, 229–252 (1959). [Google Scholar]

[R117] 117.Lanza ST Latent class analysis for developmental research. Child Dev. Perspect 10, 59–64 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R118] 118.Nylund KL, Asparouhov T & Muthén BO Deciding on the number of classes in latent class analysis and growth mixture modeling: a Monte Carlo simulation study. Struct. Equ. Modeling 14, 535–569 (2007). [Google Scholar]

[R119] 119.Bertoletti M, Friel N & Rastelli R Choosing the number of clusters in a finite mixture model using an exact integrated completed likelihood criterion. Metron 73, 177–199 (2015). [Google Scholar]

[R120] 120.Clark S & Muthén B Relating Latent Class Analysis Results to Variables Not Included in the Analysis (2009); http://www.statmodel.com/download/relatinglca.pdf

[R121] 121.Friedman J, Hastie T & Tibshirani R Regularizing paths for generalized linear models via coordinate descent. J. Stat. Softw 33, 1–22 (2010). [PMC free article] [PubMed] [Google Scholar]

PERMALINK

An observational study of Internet behaviours for adolescent females following sexual abuse

Jennie G Noll

Ann-Christin Haag

Chad E Shenk

Michelle F Wright

Jaclyn E Barnes

Mojtaba Kohram

Matteo Malgaroli

David J Foley

Michal Kouril

George A Bonanno

Abstract

Fig. 1 |. Sample and procedures flow.

Results

Descriptive statistics.

Fig. 2 |. Venn diagram of the various online URL activities occurring within authenticated sessions.

Table 1 |.

Latent profile analysis.

Fig. 3 |. Unadjusted standardized means and bootstrapped s.e. for all variables used in the LPA for the three resultant profiles.

Table 2 |.

Fig. 4 |. Forest plots of the results from the logistic and multinomial logistic regression models.

Main effects for CSA controlled for by sociodemographics.

Profile including heavy pornography use predicting outcomes.

Table 3 |.

Profile including heavy social media use and cyberbullying.

Discussion

Methods

Sample.

Procedures.

Overview.

Objective quantification of URL activity in ‘sessions’.

Quantification of YouTube and Netflix videos.

Session authentication.

Machine-learning authentication.

Human authentication.

Authentication results.

Variables and variable definitions.

URL activity.

Psychosocial risk and protective factors assessed at time 1.

Outcomes for longitudinal follow-up analyses.

Covariates.

Analytical plan.

Descriptive statistics.

LPA.

Analyses for hypothesis 1.

Analyses for hypothesis 2.

Analyses for hypothesis 3.

Reporting Summary.

Extended Data

Extended Data Fig. 1 |. Histogram of the number of “authenticated” sessions for all participants by study subgroup.

Supplementary Material

Acknowledgements

Footnotes

Data availability

References

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases