Corresponding Author:

Mohammad Belal, MCA

Abstract

Background: In today’s digital era, the internet plays a pervasive role in daily life, influencing everyday activities such as communication, work, and leisure. This online engagement intertwines with offline experiences, shaping individuals’ overall well-being. Despite its significance, existing research often falls short in capturing the relationship between internet use and well-being, relying primarily on isolated studies and self-reported data. One major contributor to deteriorated well-being is stress. While some research has examined the relationship between internet use and stress, both positive and negative associations have been reported.

Objective: This study aimed to identify the associations between an individual’s internet use and their stress.

Methods: We conducted a 7-month longitudinal study. We combined fine-grained URL-level web browsing traces of 1490 German internet users with their sociodemographics and monthly measures of stress. Further, we developed a conceptual framework that allows us to simultaneously explore different contextual dimensions, including how, where, when, and by whom the internet is used. We applied linear mixed-effects models to examine these associations.

Results: Our analysis revealed several associations between internet use and stress, varying by context. Increased time spent on social media, online shopping, and gaming platforms was associated with higher stress. For example, the time spent by individuals on shopping-related internet use (aggregated over the 30 days before their stress was measured via questionnaires) was positively associated with stress on both mobile (β=.04, 95% CI 0.00‐0.08; P=.04) and desktop devices (β=.03, 95% CI −0.00 to 0.06; P=.09). In contrast, time spent on productivity or news websites was associated with lower stress. Specifically, in the last 30 days of mobile usage, productivity-related use showed a negative association with stress (β=−.03, 95% CI −0.06 to −0.00; P=.04). In addition, in the last 2 days of data, news usage was negatively associated with stress on both mobile (β=−.54, 95% CI −1.08 to 0.00; P=.048) and desktop devices (β=−.50, 95% CI−0.90 to −0.11; P=.01). Further analysis showed that total time spent online (β=.01, 95% CI 0.00‐0.02; P<.001), social-media usage (β=.02, 95% CI 0.00‐0.03; P=.02), and gaming usage (β=.01, 95% CI 0.00‐0.02; P=.02) were all positively associated with stress in high-stress Perceived Stress Scale (PSS>26) individuals on mobile devices.

Conclusions: The findings indicate that internet use is associated with stress, and these associations differ across various usage contexts. In the future, the behavioral markers we identified can pave the way for designing individualized tools for people to self-monitor and self-moderate their online behaviors to enhance their well-being, reducing the burden on already overburdened mental health services.

J Med Internet Res 2026;28:e78775

doi:10.2196/78775

Keywords

Introduction

Stress is an unavoidable part of human life, arising from the demands and challenges we face daily. It is a significant factor in health issues such as cardiovascular disease, weakened immune function, and mental health challenges [,]. The internet, now an integral part of modern life, has sparked debates about its impact on stress levels and psychological well-being [,], as well as whether this influence is predominantly positive or negative. As our online and offline lives become increasingly interconnected, understanding the relationship between internet use and stress has gained considerable attention.

While the internet offers numerous advantages, such as enhanced connectivity and easy access to information, excessive or problematic use has been linked to various stress-related factors [,,]. For example, heavy internet use has been associated with higher levels of anxiety [], while the amount of time spent online has been linked to sleep loss and withdrawal []. On the other hand, past research suggests that not all forms of internet use are detrimental; certain online activities have been associated with reduced stress and improved psychological well-being [-].

Despite the internet’s widespread influence, research on psychological well-being, including stress, has primarily focused on offline activities, leaving a critical gap in understanding how online behaviors impact stress and well-being []. For instance, a comprehensive review of 99 commonly used psychological well-being scales identified 196 dimensions, yet none explicitly addressed online activities or behaviors []. Moreover, studies on online engagement have often faced limitations, including short study durations, small sample sizes, and an over-reliance on questionnaires to capture internet use patterns. These approaches can introduce biases and fail to provide a complete picture of the connection between online and offline experiences [,].

In this paper, we first review previous research on the association between internet use and stress and examine the methodologies used in these studies. We then outline our longitudinal multimodal study design, which integrates actual internet usage data with monthly questionnaires to measure stress, and discuss our study’s potential impact.

Conflicting Findings on Associations Between Stress and Internet Use

The relationship between internet use and stress is complex, with previous research showing contrasting associations depending on the type and context of digital engagement, as well as individual characteristics. High levels of internet and smartphone use have been linked to increased stress [,], often due to digital overload (ie, the cognitive strain caused by constant notifications and an endless stream of information) [,]. In contrast, internet use through computers has been associated with less burnout compared to smartphone use []. However, these associations are not consistent. For instance, age influences the impact of digital multitasking: younger users report higher stress than older adults when handling multiple digital tasks, yet they appear less affected by communication overload []. Experimental evidence shows that multitasking increases perceived stress levels in both younger and older adults, with no significant differences between the age groups []. Other studies have shown no association [] or even a negative association between time spent online and stress, particularly in young adults [].

Moreover, the type of online activity plays a crucial role in stress outcomes. Social networking and entertainment-related use have been associated with higher stress levels, while internet use for work-related tasks has been linked to greater life satisfaction in the middle-aged population []. Research also indicates that communication overload from emails and messages is positively associated with perceived stress in the age group of 50‐85 years []. Studies on social media show similarly nuanced findings. While Pew Research found no association between social media use and stress in men, a negative association was observed in women []. A large-scale study also showed slightly higher perceived stress among high social media users than nonusers []. Other digital behaviors, such as problematic news consumption [] and adult content addiction [], have also been linked to heightened stress and emotional distress. Similarly, concerns such as cyberbullying, online harassment [], work-life boundary erosion [], and data privacy issues [] have been widely documented as stress-inducing. Further studies have found positive associations between stress and various digital behaviors, such as online shopping addiction in young people [], negative information seeking [], interpersonal communication in older adults [], misinformation sharing [], and excessive gaming in adolescents [,].

Conversely, the internet can also act as a buffer against stress [], offering access to supportive communities [], relaxation tools, and leisure activities []. Online entertainment and social interaction, in particular, have been shown to reduce stress and enhance well-being among older adults [,,]. In addition, internet use has been recognized as a coping mechanism. Several online activities, including social media [], entertainment [,], shopping [], and gaming [,], have been identified in previous studies as strategies for managing stress.

User Characteristics Shape the Relationship Between Online Activity and Stress

Studies show that age, gender, income, and baseline stress levels influence how online activities relate to stress [,,]. Social media use has been negatively associated with stress among females [], though females overall report higher stress than males [,]. Older adults tend to report lower stress than younger groups but experience stronger associations between communication load, frequent messaging, and stress [,]. Higher income is generally linked to lower stress levels [], though findings differ by context—for instance, higher income was associated with fewer mental health issues in Germany but with more issues in China []. Social media use has also been linked to slower recovery from real-world stressors, suggesting possible stress maintenance in already stressed individuals [].

Methods for Identifying Connections Between Internet Use and Stress

Past research on internet use and stress has used various methodologies. Many studies rely on cross-sectional designs and self-reported surveys [,,,,,]. These studies often focus on specific populations, such as university or medical students [,]. Some have used larger samples [,,] but still rely on questionnaires to capture internet use. A smaller number of experimental studies are available [,], and some have adopted alternative approaches, such as analyzing social media data to infer psychological states []. Studies using actual web browsing data [,] are limited and tend to capture general metrics, such as total time spent online [], and are based on relatively small samples (92 and 107 participants), indicating how difficult it is to conduct such studies.

Contributions and Impact of Our Study

Previous research on the relationship between internet use and stress has shown mixed findings, revealing both negative and positive associations depending on the type of internet activity. However, much of this evidence remains fragmented, as previous studies have largely relied on self-reported internet use data (which often lacks granularity) and have focused on limited aspects of internet use.

To address these limitations and to provide a more comprehensive understanding of how internet use relates to stress, we conducted a longitudinal multimodal study involving 1490 internet users in Germany over 7 months. Our study integrates fine-grained, passively collected web trace data from both desktop and mobile devices with participants’ monthly responses to a validated stress scale (refer to ). Using objective behavioral data, we move beyond subjective self-reports and introduce a data-driven framework for revealing long-term usage patterns and identifying digital markers of stress.

‎**Figure 1. ** Overview of the study design and contextual dimensions. The top panel shows our longitudinal study design combining desktop and mobile web‐trace data with monthly stress questionnaires. The middle panel depicts the internet use and well-being features extracted from web browsing traces and monthly questionnaires. The bottom panel shows the contextual dimensions we consider for examining associations between internet use and stress.

Building on existing research, we further identify 4 key dimensions that shape the relationship between digital behavior and stress:

1. How: the type and pattern of internet behaviors—such as usage volume [,], temporal rhythms [], and content categories (eg, social media, productivity, entertainment, or shopping)—may influence stress in different ways [,-].
2. Where: the device context plays a significant role. Previous research suggests that desktop use is often more goal-oriented and structured, whereas mobile use tends to be more fragmented and reactive [].
3. When: the timing of online behaviors in relation to stress assessments is important, as short-term engagement with digital content may have immediate effects on stress responses [,].
4. By whom: individual differences—such as age, gender, income, and baseline stress levels—can moderate the impact of digital engagement on stress, with some groups being more susceptible than others [-].

By adopting this multidimensional perspective, our study seeks to bring coherence to the scattered evidence in existing research and provide a more thorough understanding of the link between digital behaviors and stress. In addition, identifying behavioral markers of stress in internet use may inform the design of future tools for real-time stress monitoring, complementing traditional self-reported measures. This approach contributes to a deeper understanding of digital well-being and supports the development of targeted interventions for healthier online behaviors.

Methods

In this section, we describe our study design, participants, collected data, extracted features, and analysis models that allowed us to overcome the challenges of previous work described in the “Contributions and Impact of Our Study” section.

Study Design

We conducted a longitudinal multimodal study over 7 months that combined passively collected fine-grained web browsing traces with repeated monthly online questionnaires (as provided in ). The web browsing traces for desktop users included URL-level traces, while for mobile users, both URL-level and application-level traces were included (throughout the paper, we will use “app” to mean mobile app). We measured the perceived stress of our panelists using the validated Perceived Stress Scale (PSS)-10 in 6 monthly waves. In the first wave, we also collected the sociodemographic characteristics of the panelists, including age, gender, and income.

Participants

The study was conducted on a sample of German internet users, recruited through a General Data Protection Regulation–compliant panel company (Bilendi GmbH), which provided access to participants who had already installed tracking software on their devices to capture their internet use. The company also managed survey coding and distribution, sending email notifications to relevant participants each time a survey was launched. Participation was voluntary; panelists were informed about the study and chose to take part. Compensation ranged from €1‐€3 (US $1.10-US $3.30) US $3.30) per month, depending on the number of devices tracked, plus an additional amount based on the company’s standard rate (€6/h or US $6.60/h) for each survey completion. All currency values in this study are reported in euros. The exchange rate at the time of the study was €1=US $1.10. All panelists from the company were invited via email for the first wave of questionnaires, yielding 1490 completed responses. In the subsequent 5 waves, which were approximately 1 month apart, all these 1490 respondents were invited via email to participate in each wave. Across all waves, the average range between the earliest and latest survey completion dates was 15 days, and the average completion time for the baseline survey was about 37 minutes (90% CI 32.7-42.3), while for the remaining 5 waves, it was approximately 19.6 minutes (90% CI 17.8-21.5). reports the number of participants with completed responses for each wave that we retained for further analysis. We excluded 1 panelist who reported their gender as nonbinary and 52 panelists who reported their income as “other,” because these categories had too few respondents. The sample for Wave 1 is therefore 1437. We observe that about 23% of the panelists dropped out by the sixth wave. In , we also depict the distribution of the panelists across age, gender, and income for the 6 waves of questionnaires.

Next, we examined how closely our sample’s sociodemographic distributions match the German population margins for gender, age, and income (provided in ). We observe that our sample’s gender distribution matches closely with that of the German population (Destatis []). However, for both age and income, the middle ranges are overrepresented in our sample, while the extremes are underrepresented.

**Table 1. **Descriptive characteristics of participants across 6 survey waves. The table presents the number of participants, gender distribution, age groups, income categories, and mean perceived stress scores (with SDs) for each wave. Percentages are shown in parentheses for categorical variables.

Wave123456 Participants (n)143713141212119812051107 Gender, n (%) Male738 (51.35)688 (52.36)639 (52.72)635 (53.01)628 (52.12)593 (53.57) Female699 (48.65)626 (47.64)573 (47.28)563 (46.99)577 (47.88)514 (46.43) Age group (years), n (%) 18‐30119 (8.28)101 (7.67)91 (7.51)84 (7.01)92 (7.63)76 (6.67) 31‐45462 (32.15)414 (31.51)385 (31.77)382 (31.89)379 (31.45)330 (29.81) 46‐60569 (39.60)528 (40.18)483 (39.85)483 (40.32)483 (40.08)460 (41.55) >60287 (19.97)271 (20.62)253 (20.87)249 (20.78)251 (20.83)241 (21.77) Income (euros/month), n (%) <1000 (Tier I)126 (8.77)119 (9.06)116 (9.57)110 (9.18)103 (8.55)94 (8.49) 1000‐2000 (Tier II)300 (20.88)271 (20.62)244 (20.13)244 (20.37)249 (20.66)240 (21.68) 2001‐3000 (Tier III)364 (25.33)339 (25.80)319 (26.32)310 (25.88)309 (25.64)281 (25.38) 3001‐4000 (Tier IV)294 (20.46)271 (20.62)243 (20.05)245 (20.45)251 (20.83)229 (20.69) >4000 (Tier V)353 (24.57)314 (23.90)290 (23.93)289 (24.12)293 (24.32)263 (23.76) Perceived Stress Score, mean (SD)16.19 (7.19)15.83 (7.43)15.76 (7.56)15.65 (7.41)15.54 (7.50)14.89 (7.58)

**Table 2. **Distribution of the adult population in Germany (2023) by gender, age group, and monthly income level.

CategoryAdult population, n (%) Sex Male41.2 (48.8) Female42.3 (51.2) Age group (years) 18‐3014.2 (17) 31‐4519.2 (23) 46‐6021.7 (26) >6028.4 (34) Monthly income level (euros) <€1250 (<US $1375)21.1 (25.3) €1250-€2080 (US $1375-US $2288)13 .7 (16.4) €2080-€2920 (US $2288-US $3212)12.4 (14.8) €2920-€4170 (US $3212-US $4587)

13.6 (16.3) >€4170 (>US $4587)22.6 (27.1)

Data Collection

The panelists of the panel company had already consented to install tracking software on their desktops or mobile devices. Some participants consented to install it on both devices. Through this tracking software, the company provided fine-grained traces of visited URLs and mobile apps, including the time of visit and duration of each visit. During the 7 months, we recorded 47,100,701 URL visits from both desktop and mobile users, covering 236,955 unique web domains. For mobile apps, we captured 13,553,645 app visits across 13,476 unique apps.

Data Cleaning and Preprocessing

First, we removed the bottom 20% of panelists in each wave, ranked by total time spent browsing, since they did not have sufficient data to extract meaningful internet use patterns. Second, we identified a group of panelists who appeared to be “professional survey takers,” spending more than 25% of their online time on survey domains. To focus on users with more typical internet use, we excluded these individuals from the sample. Notably, applying the above time threshold to their nonsurvey activities would have led to the removal of more than 29% of these panelists. Finally, to ensure that our internet use measures accurately capture user behavior, we only included panelists for whom we could categorize at least 80% of their web visits (refer to the “Data Enrichment” section for details). summarizes the number of participants excluded at each step, resulting in a set of distinct panelists across waves comprising 656 mobile users and 526 desktop users. presents the sociodemographic characteristics of the remaining users included in the analysis. In the mobile data, the proportion of users aged 31‐45 years increased compared to the baseline questionnaire, as provided in . In the desktop data, the proportion of males and users aged 46‐60 years increased, while the proportion of users aged 31‐45 years decreased relative to the baseline questionnaire.

**Table 3. **Overview of panelists with matched passive web data from desktop and mobile devices across 6 survey waves. The table shows the number of users before and after data cleaning for both device types. The final row indicates the total number of distinct users retained in the cleaned dataset.

Survey wave	Number of panelists	Desktop users	Desktop users (cleaned)	Mobile users	Mobile users (cleaned)
1	1437	981	359	907	519
2	1314	848	321	806	470
3	1212	762	284	728	426
4	1198	717	257	717	418
5	1205	714	265	697	399
6	1107	656	227	649	368
Total distinct users	—	—	526	—	656

aNot applicable.

**Table 4. **Sociodemographic characteristics of users included in the analysis after data cleaning for both mobile and desktop datasets.

CharacteristicMobile (n=656)Desktop (n=526) Gender, n (%) Male334 (50.91)289 (54.94) Female322 (49.09)237 (45.06) Age group (years), n (%) 18‐3053 (8.08)38 (7.22) 31‐45246 (37.5)137 (26.05) 46‐60247 (37.65)232 (44.11) >60110 (16.77)119 (22.62) Income (euros/month), n (%) <1000 (Tier I)60 (9.15)59 (11.22) 1000‐2000 (Tier II)129 (19.66)121 (23) 2001‐3000 (Tier III)166 (25.30)128 (24.33) 3001‐4000 (Tier IV)137 (20.88)87 (16.54) >4000 (Tier V)164 (25)131 (24.9)

Data Enrichment

To understand “how” the panelists are using the internet, we categorized their online visits into semantic categories. The goal was to group domains, subdomains, and apps based on their primary function into categories such as “social media” (eg, facebook.com [Meta Platforms, Inc] and TikTok app [ByteDance]) and “productivity” (eg, Gmail [Google LLC] and calendar.google.com [Google LLC]). We derived the set of categories (provided in ) by combining categories used by app stores and web domain classification services such as Webshrinker.com. For platform domains such as google.com, we also categorized their subdomains. For instance, google.com was categorized as “search,” while mail.google.com was classified as “productivity.” provides the complete list of semantic categories we considered, along with some examples. Two researchers from our team first independently annotated the categories for all domains and apps that constituted around 85% of web visits made by our panelists. Later, disagreements were resolved collaboratively. We observed a substantial interannotator agreement with a Cohen ĸ agreement [] score of 0.7, based on annotations for a random subset of 200 domains. Following this process, we classified 3777 domains and 989 apps into semantic categories, capturing 85% of visits from mobile devices and 84% from desktops.

**Table 5. **Categorization of web domains and mobile apps based on semantic usage type. The table lists representative examples of domains and apps across various categories, grouped separately for desktop web domains and mobile apps.

CategoryExample of domains, subdomains, and apps in the category Domains Entertainmentyoutube.com, twitch.tv, disneyplus.com, and netflix.com Shoppingamazon.de, ebay.de, kleinanzeigen.de, and temu.com Social mediafacebook.com, twitter.com, and instagram.com Messagingwhatsapp.com, knuddels.de, and fdating.com Productivitymail.google.com, outlook.live.com, navigator.web.de, and docs.google.com Gamesgameduell.de, anocris.com, forgeofempires.com, and spielaffe.de Adultpornhub.com, xvideos.com, xnxx.com, and romeo.com Newsbild.de, focus.de, welt.de, and wunderweib.de Apps EntertainmentYouTube (Google LLC), Netflix (Netflix, Inc), and Spotify Music (Spotify Technology) ShoppingAmazon Shopping (Amazon.com, Inc), eBay (eBay, Inc), Vinted.fr (Vinted Group), and Lidl Plus (Schwarz Group) Social mediaFacebook (Meta Platforms, Inc), Instagram (Meta Platforms, Inc), and Twitter (X Corp), TikTok – Make Your Day (ByteDance) MessagingWhatsApp (Meta Platforms, Inc), Facebook Messenger (Meta Platforms, Inc), and Telegram (Telegram FZ-LLC) ProductivityGmail (Google LLC), GMX Mail (Global Mail eXchange), WEB.DE Mail (United Internet Group), and Google Calendar (Google LLC) GamesCandy Crush Saga (King), Coin Master (Moon Active), Royal Match (Dream Games), and Pokémon GO (Niantic) Newsn-tv Nachrichten (RTL Group), kicker online (Olympia -Verlag GmbH), AOL – News (AOL Media LLC), and BILD: Immer aktuell informiert (Axel Springer SE)

Measures

As described in the “Study Design” section, we combined repeated monthly stress questionnaires with web browsing data. Perceived stress was measured through PSS-10 questionnaire responses, and internet use features were derived from passively collected web traces. For each panelist in each survey wave, we calculated internet usage features based on their activity during the period “T” preceding the stress measurement (ie, the time of questionnaire response). To address the question of “when” the internet is used, we extracted features for either 30 or 2 days to examine both long-term and short-term effects. The resulting measures were then used to examine the associations between internet use and stress.

Measures From Web Traces

To measure “how” individuals use the internet, we created features that span from coarse- to fine-grained measures, as provided in . We captured overall web activity at the coarse-grained level, such as total time spent online. We also accounted for the time of the day when panelists were browsing the web by including the difference between the time spent online during daytime (6 AM-6 PM) and nighttime (6 PM-6 AM) hours. At a finer granularity, we analyzed how panelists distributed their time across online activities such as social media, entertainment, and news. For each survey wave, if a participant completed the survey on a given date (eg, July 31), we summed their time spent on each activity over the period (T=30 or 2 days; eg, July 1‐30 or July 29‐30) preceding the day of completing the survey. For instance, time spent on news represents the aggregated time on news domains (desktop) and news domains and apps (mobile) during that period. Each participant had up to 6 time points, 1 per wave.

**Table 6. **Features and their descriptions. Time spent online is measured in hours in the period T (30 days or 2 days) before the measurement of stress. These features are computed for online activity on each device (desktop or mobile) separately.

FeaturesDescription Coarse-grained

• Total time spent online Total time spent online in period T

• Daytime nighttime difference Difference of time spent online during daytime (6 AM-6 PM) and nighttime hours (6 PM-6 AM) Fine-grained

• Time spent on entertainment

• Time spent on social media

• Time spent on messaging

• Time spent on news

• Time spent on adult content

• Time spent on games

• Time spent on shopping

• Time spent on productivity Time spent on different semantic classes of online activities. For instance, time spent on entertainment domains or apps such as YouTube.com or Amazon Prime is classified as entertainment use. Control variables

• Gender

• Age

• Income

• Survey wave Sociodemographic characteristics of individuals and seasonality

Measures From Questionnaires

We used the PSS-10 [] in our monthly questionnaires to measure the stress levels of our panelists. The PSS-10 is a widely used, validated scale designed to assess how stressed individuals feel. It captures aspects such as the unpredictability of life, perceived control over situations, and general stress levels over the past month. Participants rate their responses on a scale from 0 (never) to 4 (very often), producing a total score between 0 and 40 across the 10 items. Higher scores indicate greater perceived stress, with scores typically grouped into 3 levels: 0‐13 (low stress), 14‐26 (moderate stress), and 27‐40 (high stress) [,]. In addition, we collected each participant’s self-reported sociodemographics, including age, gender, and income, in the first wave of the questionnaires.

Statistical Analysis

We used linear mixed-effects models (LMMs) [] to examine the relationship between internet use and stress. We chose LMMs for analyzing data from our longitudinal study since they account for repeated measurements of individuals and incorporate fixed and random effects. Fixed effects included internet use features provided in . Random intercepts were added to account for individual-specific differences in baseline stress levels across participants.

We formally describe the models as follows. For an individual i at questionnaire wave j∈ {1,2...6}, we denote Yij as the variable of interest, xij as the covariate, and the intercept for the random effect as uj. Therefore, we consider the following LMM:

Yij=β0+β1xij1+β2xij2+…+βnxijn+uj+ϵij

where:

• Yij is the perceived stress level of the individual i measured at the questionnaire wave j.
• β0 is the fixed intercept.
• β1,…,βn are the fixed effect coefficients for each covariate xijn.
• xij1=(total time spent online, xij2=daytime-nighttime difference, xij3=time spent on entertainment ... xijn=survey wave), where xijn corresponds to each feature provided in .
• uj is the random effect for individual i, capturing individual-level variability.
• ϵij is the residual error term for the individual i at wave j.

We conducted model diagnostics to validate the assumptions of LMMs, including checks for multicollinearity (Variance Inflation Factor (VIF) <2.0). All statistical analyses were implemented using Python’s statsmodels package (version 0.14.1; Python Software Foundation*)* [].

To understand whether the granularity of the extracted features affects model performance and the associations identified, we developed 2 models. The first model (Model 1) focused on coarse-grained measures of internet use such as total time spent online and daytime-nighttime difference. The second model (Model 2) extended the first model and also incorporated finer-grained measures of internet use across semantic classes. For Model 2, we dropped the total time spent online feature to avoid multicollinearity.

Previous work has shown that both sociodemographics [-] and seasonal variations [,] can significantly influence individuals’ stress levels. Accordingly, we included the sociodemographics and seasonality measures as control variables for both models, as provided in .

Ethical Considerations

Our study was approved by Aalto University’s Research Ethics Committee (approval ID D/894/03.04/2023). Data collection was conducted via a General Data Protection Regulation–compliant European company, and informed consent was obtained from participants for both the surveys and web-trace datasets, with the option to withdraw consent at any time during or after the study. To protect participants’ privacy, we implemented strict data privacy measures. The web dataset was anonymized by the panel company by removing personal information such as email addresses and usernames to prevent participant identification. In addition, the dataset was stored and analyzed solely on the university’s secure server, with access restricted to the research team. We will make the anonymized data and code available to support the open-source community and to spur further research at the intersection of internet use and well-being.

Results

Overview

Our study examined various internet use behaviors associated with stress, and in this section, we present our results across 4 key contextual dimensions (as outlined in the “Contributions and Impact of Our Study” section). We first analyzed “how*”* internet-based features relate to stress. We then explored the remaining dimensions: device-based differences (where) by comparing desktop and mobile usage, temporal patterns (when) using internet activity from the 2 days before the survey, and individual differences (by whom) through subgroup analyses based on age, gender, income, and baseline stress levels.

Behavioral Patterns (How)

To understand how internet usage is associated with stress, we ran LMMs on 2 sets of features, progressing from coarse-grained (amount and timing of usage) to fine-grained (also including semantic category usage) measures. An ANOVA test was conducted to determine whether the more complex model explained significantly more variance than the simpler model. The results showed no statistically significant improvement when using the more complex model, although the more complex model provided important information on the nuanced relationship between internet use and stress.

Analysis of 30-day mobile data (number of panelists, N=656) and observations, n=2600), as provided in , revealed that Model 2—which includes both timing and semantic category usage—identified significant associations with stress. Specifically, shopping-related usage was positively associated with stress (β=.04,, 95% CI 0.00‐0.08; P=.04), while productivity usage showed a negative association (β=–.03, 95% CI −0.06 to −0.00; P=.04). In contrast, Model 1, which included only total usage and timing, did not show any significant associations.

**Table 7. **Results from linear mixed-effects models for all participants, based on 30-day mobile data. Model 1 includes coarse-grained features, while Model 2 incorporates fine-grained usage categories (described in the “Statistical Analysis” section). Estimates, CIs, and P values are reported for each predictor. Statistically significant P values are in bold.

Predictors	Model 1, estimate (95% CI)	P value	Model 2, estimate (95% CI)	P value
Intercept	20.77 (19.09‐22.44)	<.001	20.69 (19.01‐22.37)	<.001
Survey wave	−0.10 (−0.19 to −0.02)	.01	−0.11 (−0.19 to −0.02)	.01
Gender	1.79 (0.80‐2.77)	<.001	1.68 (0.68‐2.68)	<.001
Age	−1.62 (−2.20 to −1.04)	<.001	−1.57 (−2.16 to −0.99)	<.001
Income	−1.11 (−1.50 to −0.73)	<.001	−1.08 (−1.47 to −0.69)	<.001
Total time spent online	0.00 (−0.00 to 0.01)	.39	—	—
Daytime nighttime difference	−0.01 (−0.02 to −0.00)	.12	−0.01 (−0.02 to 0.00)	.18
Time spent on entertainment	—	—	0.01 (−0.01 to 0.02)	.37
Time spent on social media	—	—	0.00 (−0.01 to 0.02)	.71
Time spent on messaging	—	—	0.00 (−0.02 to 0.02)	.70
Time spent on games	—	—	0.00 (−0.00 to 0.01)	.30
Time spent on shopping	—	—	0.04 (0.00‐0.08)	.04
Time spent on productivity	—	—	−0.03 (−0.06 to −0.00)	.04
Time spent on news	—	—	−0.03 (−0.09 to 0.03)	.33

aσ²=11.67; τ₀₀=36.54pid; ICC=0.76; N=656pid; Observations=2600.

bσ²=11.65; τ₀₀=36.48pid; ICC=0.76; N=656pid; Observations=2600.

cP<.001.

dP<.05.

eNot applicable.

In addition, sociodemographic factors such as age, gender, and income consistently predicted stress across both models. Age (β=–1.57, 95% CI −2.16 to −0.99; P<.001) and income (β=–1.08, 95% CI −1.47 to −0.69; P<.001) were negatively associated with stress, while women reported higher stress levels (β=1.68, 95% CI 0.68‐2.68; P=.001).

Device Matters (Where)

To observe device differences, we analyzed 30-day desktop data. For desktop data (N=526 and n=1713), the results are shown in . Model 2, which incorporates semantic and temporal features, showed a weaker positive association between shopping usage and stress (β=.03, 95% CI −0.0 to 0.06; P=.09). As observed with mobile data, the simpler Model 1 did not reveal any significant associations with internet usage features. Similarly, the sociodemographic results were consistent with those observed in the mobile data.

**Table 8. **Results from linear mixed-effects models for all participants, based on 30-day desktop data. Model 1 includes coarse-grained features, while Model 2 incorporates fine-grained usage categories (described in the “Statistical Analysis” section). Estimates, CIs, and P values are reported for each predictor. Statistically significant P values are in bold. Random effects, intraclass correlation coefficient (ICC), number of participants (N), and total observations are also provided.

Predictors	Model 1, estimate (95% CI)	P value	Model 2, estimate (95% CI)	P value
Intercept	20.62 (18.70‐22.55)	<.001	20.49 (18.58‐22.41)	<.001
Survey wave	−0.14 (−0.24 to −0.03)	.009	−0.15 (−0.26 to −0.04)	.006
Gender	1.83 (0.66‐3.00)	.002	1.73 (0.55‐2.90)	.004
Age	−1.72 (−2.41 to −1.04)	<.001	−1.74 (−2.43 to −1.05)	<.001
Income	−0.94 (−1.38 to −0.51)	<.001	−0.92 (−1.36 to −0.48)	<.001
Total time spent online	−0.00 (−0.01 to 0.00)	.40	—	—
Daytime nighttime difference	0.00 (−0.01 to 0.01)	.46	0.00 (−0.01 to 0.01)	.46
Time spent on entertainment	—	—	−0.00 (−0.01 to 0.01)	.64
Time spent on adult content	—	—	−0.01 (−0.03 to 0.00)	.11
Time spent on social media	—	—	0.00 (−0.02 to 0.02)	.94
Time spent on messaging	—	—	0.01 (−0.04 to 0.06)	.61
Time spent on games	—	—	0.00 (−0.03 to 0.03)	.84
Time spent on shopping	—	—	0.03 (−0.00 to 0.06)	.09
Time spent on productivity	—	—	−0.00 (−0.03 to 0.02)	.86
Time spent on news	—	—	−0.02 (−0.06 to 0.02)	.28

aσ²=11.10; τ₀₀=40.62pid; ICC=0.79; N=526pid; Observations=1713.

bσ²=11.09; τ₀₀=40.77pid; ICC=0.79; N=526pid; Observations=1713.

cP<.001

dP<.01.

eNot available.

Time Period of Data (When)

To investigate the relationship between short-term versus long-term internet usage patterns and stress, we analyzed associations between various online activities performed on mobile and desktop devices in 2 time periods—30 days and 2 days—and individual stress levels. We used the same features and models as in our previous 30-day data analyses. Here, we specifically focused on web activity recorded during the 2 days immediately preceding the PSS-10 survey.

For both mobile and desktop data (as provided in Tables S1 and S2 in ), news usage showed a negative association with stress (β=–.54, 95% CI −1.08 to 0.00; P=.048) and (β=–.50, 95% CI −0.90 to −0.11; P=.01), respectively, in Model 2. In addition, in desktop data, messaging usage demonstrated a weak negative association with stress (β=–.59, 95% CI −1.24 to 0.06; P=.07) in model 2.

Individual Differences (By Whom)

To explore how internet usage varies by individual characteristics, we conducted subgroup analyses, running models separately for categories such as gender (male and female) to understand how associations differ based on these characteristics. In the following subsections, we first examined the relationship between internet use and stress based on baseline stress levels by analyzing high-stress and low-stress groups. We then explored differences by gender, age, and income categories.

Stress Levels

We identified 2 groups of panelists from our data—high-stress and low-stress—based on their reported PSS-10 scores in the online questionnaires. Panelists who scored more than 26 in any wave they participated in were included in the high-stress group, and panelists who scored below 14 in any wave were included in the low-stress group.

For the high-stress population (PSS-10 score > 26), several notable results were observed (as provided in Tables S3-S6 in ). In the 30-day mobile data, time spent (β=.01, 95% CI 0.0‐0.02; P<.001) in Model 1, and social media usage (β=.02, 95% CI 0.0‐0.03; P=.02), and gaming usage (β=.01, 95% CI 0.0‐0.02; P=.02) in Model 2 were positively associated with stress. In the 2 days of data, the daytime-nighttime difference showed a weak positive association (β=.11, 95% CI −0.01 to 0.23, P=.08) in Model 1. For desktop data, no significant variables, including sociodemographics, were found to be associated with stress in the high-stress subgroup.

In the low-stress population (PSS-10 score <14 in the baseline survey), as provided in Tables S7-S10 in , adult-content usage was negatively associated with stress in 30-day desktop data (β=–.02, 95% CI −0.04 to 0.0; P=.07). Similarly, for the low-stress group, in the 2 days data, time spent (β=–.07, 95% CI −0.14 to 0.0; P=.04) was significant for desktop, while gaming usage was weakly significant for mobile data (β=.08, 95% CI −0.01 to 0.17; P=.1).

When analyzing the 30-day data for all participants, sociodemographic factors were strongly associated with stress in both mobile and desktop settings. However, within the high-stress group, income was the only sociodemographic variable that remained significant in mobile data, showing a negative association with stress in both the general population (β=–1.08, 95% CI −1.47 to −0.69; P<.001) and the high-stress subgroup (β=–.52, 95% CI −0.91 to −0.12; P=.01). In contrast, gender and age—which were significant predictors in the overall population—did not show statistical significance in the highly stressed subgroup for either mobile or desktop data.

Gender Differences

Subgroup analysis by gender revealed distinct patterns in feature significance for both desktop and mobile data. In the 30-day mobile data (as provided in Table S11 in ), shopping usage (β=.07, 95% CI 0.01‐0.14; P=.02) and productivity features (β=–.05, 95% CI −0.09 to −0.01; P=.02) were significant only for male users (N=334 panelists and n=1334 observations) in Model 2. In contrast, these features were not significant for female users (N=322 panelists and n=1266 observations) as provided in Table S12 in . No features were significant for males and females in the 30-day desktop data.

In the 2-day data (as provided in Tables S15-S16 in ), news consumption was negatively associated with stress for males in both desktop (β=–.52, 95% CI −1.02 to −0.01; P=.04) and mobile (β=–.58, 95% CI −1.23 to 0.07; P=.08) data. For females, in mobile data, daytime-nighttime difference (β=.10, 95% CI −0.0 to 0.21; P=.06) and messaging (β=.23, CI −0.02 to 0.48; P=.08) were weakly positively associated.

Age Differences

Subgroup analysis by age revealed distinct patterns in web-based associations. For the 30-day mobile data (as provided in Tables S19-S22 in ), shopping was positively associated with stress in age groups of 18‐30 years (β=.12, 95% CI −0.02 to 0.25; P=.09) and older than 60 years (β=.01, 95% CI 0.02‐0.19; P=.02). In the age group of 30‐45 years, weak positive associations were found for entertainment (β=.02, 95% CI −0.00 to 0.04; P=.08) and messaging (β=.03, 95% CI −0.00 to 0.06; P=.07), whereas productivity was negatively associated (β=–.08, 95% CI −0.14 to −0.02; P=.007). In the older than 60 years group, time spent (β=.01, 95% CI 0.00‐0.03; P=.03) and messaging (β=.05, CI −0.00 to 0.09; P=.06) were positively associated, while the daytime-nighttime difference (β=–.03, 95% CI −0.05 to −0.01; P=.01) and shopping (β=.10, 95% CI 0.02‐0.19; P=.02) were negatively associated.

In the 30-day desktop data (as provided in Tables S23-S26 in ), adult content usage was negatively associated in 18‐30 years (β=–.87, 95% CI −1.77 to 0.03; P=.06) and older than 60 years (β=–.19, 95% CI −0.38 to 0.00; P=.06) age groups. In addition, in the 18‐30 age group, the daytime-nighttime difference (β=.07, CI −0.00 to 0.14; P=.06) was positively associated, while news usage (β=.026, CI −0.42 to −0.10; P=.002) was negatively associated. Shopping was positively associated in the age group of 45‐60 years (β=.04, 95% CI −0.00 to 0.09; P=.06).

For the 2-day mobile data (as provided in Tables S27-S30 in ), news was negatively associated in both the 30‐45 years (β=–1.01, 95% CI −2.22 to 0.19; P=.10) and 45‐60 years (β=–.85, 95% CI −1.76 to 0.06; P=.07) age groups. In addition, in the age group of 30‐45 years, the daytime-nighttime difference (β=.14, 95% CI 0.02‐0.25; P=.02) and entertainment usage (β=.22, 95% CI 0.01‐0.43; P=.04) were positively associated, whereas shopping was negatively associated (β=–.54, 95% CI −1.08 to −0.0; P=.05). In the older than 60 years of age group, time spent on gaming (β=.26, 95% CI 0.01‐0.50; P=.04) was positively associated.

Similarly, for the 2-day desktop data (as provided in Tables S31-S34 in ), messaging was negatively associated in the 45‐60 years age group (β=–.78, 95% CI −1.48 to −0.07; P=.03), but positively associated for the older than 60 years age group (β=7.63, 95% CI 0.39‐14.87; P=.04). In the age group of 18‐30 years, entertainment (β=1.14, 95% CI −0.18 to 2.46; P=.09) was positively associated, and social media (β=1.22, 95% CI 0.15‐2.29; P=.03) was positive for the 30‐45 age group. In addition, time spent (β=–.14, 95% CI 0.28‐0.0; P=.06) and news usage (β=.46, 95% CI −0.96 to 0.04; P=.07) were negatively associated in the older than 60 years age group.

Income Differences

For the 30-day mobile data (as provided in Tables S35-S39 in ), messaging (β=–.06, 95% CI −0.12 to −0.01; P=.03) was negatively associated with stress in participants earning less than €1000 (US $ 1100) per month (Tier I). In the €2001-€3000 (US $2201.10-US $3300) income group (Tier II), productivity (β=–.05, 95% CI −0.10 to 0.0; P=.06) and news usage (β=–.10, 95% CI −0.20 to 0.01; P=.08) were both negatively associated. Time spent (β=.01, CI 0.0‐0.02; P=.02) and shopping (β=.08, 95% CI 0.01‐0.16; P=.02) were positively associated with stress for participants earning €3001-€4000 (US $3301.10-US $4400; Tier IV). No other significant internet-based features were observed for other income categories.

For the 30-day desktop data (as provided in Tables S40-S44 in ), news usage was positively associated with stress in Tier I income group (β=.28, 95 CI 0.05‐0.50; P=.02), while productivity was negatively associated (β=–.07, 95% CI −0.13 to −0.01; P=.03). For participants in Tier III income, news usage (β=–.13, 95% CI −0.25 to 0.0; P=.06) was negatively associated. For participants in Tier IV income, shopping was positively associated with stress (β=.08, 95% CI 0.02‐0.15; P=.02). For Tier V participants, social media use was positively associated with stress (β=.06, 95% CI 0.01‐0.12; P=.03), while news use (β=–.16, 95% CI −0.25 to −0.07; P<.001) and time spent were negatively associated (β=–.01, 95% CI −0.03 to 0.0; P=.03). No significant associations were identified for other income categories.

For the 2-day mobile data (as provided in Tables S45-S49 in ), gaming (β=.21, 95% CI −0.02 to 0.43; P=.07) was positively associated in the Tier II income group and negatively associated (β=–.19, 95% CI −0.37 to 0.0; P=.047) in the Tier V income group. News use was negatively associated in the Tier III (β=–.75, 95% CI −1.58 to 0.09; P=.