1 General Introduction
There is ample debate in society over the impact of (social media) screen time on well-being (Ferguson et al., 2022; Twenge & Farley, 2021; Valkenburg, 2022). What fuels this debate is that studies on the topic provide conflicting results. For instance, Twenge and colleagues identified screen time as a direct predictor of mental health problems (e.g., Twenge et al., 2018), while others find positive dynamics between smartphone use and well-being (Marciano et al., 2022; Marciano & Viswanath, 2023). In a recent umbrella review synthesizing systematic reviews and meta-analyses on this association, Valkenburg and colleagues (2022) concluded that the extant evidence is indeed mixed, with methodological shortcomings stated as a key reason.
In particular, the field has long relied on self-reported measures of screen time, whose validity has been repeatedly questioned: such reports are inaccurate (Hodes & Thomas, 2021; Molaib et al., 2025; Ohme et al., 2021; Parry et al., 2021; Tkaczyk et al., 2024; Verbeij et al., 2021) and systematically biased (Sewall et al., 2020). For instance, Sewall and colleagues (2020) found that the accuracy of self-reports is associated with well-being and usage amount.
The concerns over the validity of self-reported screen time have motivated a noticeable paradigm shift over the past decade (Ellis et al., 2019; Parry et al., 2021), in which researchers are increasingly moving away from retrospective self-reports and towards the use of behavioral data (e.g., Buda et al., 2023; de Segovia Vicente et al., 2024; Elmer et al., 2025; große Deters & Schoedel, 2024; Rodman et al., 2024; Schenkel et al., 2024).
The collected behavioral data, also commonly referred to as digital trace data (Aalbers, 2023), typically fall within the realm of ‘big data’, leveraging thousands of data points per participant. Although these data often require extensive cleaning and pre-processing, their use enables more sophisticated methods for feature extraction, analysis and modeling (Bjerre-Nielsen & Glavind, 2022; Boeschoten et al., 2022; Parry & Toth, 2025; Peters et al., 2024; Saito et al., 2015). As such, the paradigm shift towards the use of passively sensed device data can be argued to also be part of the ‘computational turn’ in the broader field of communication, and even the social sciences (Berry, 2011; Es et al., 2021; Hepp et al., 2021).
Within this computational turn, the smartphone has become a main device of interest when it comes to the study of screen time and well-being. Smartphone trace data offer higher temporal resolution and reduce reliance on participants’ retrospective recall, enabling data collection within the context of daily life (Harari et al., 2017; Harari & Gosling, 2023). When combined with (mobile) Experience Sampling Methods (ESM), this provides the opportunity to carry out more precise analyses, taking within-person differences, temporal dimensions, and even momentary associations into account (Elmer et al., 2025; Harari & Gosling, 2023; Keusch & Conrad, 2022; Klingelhoefer et al., 2025).
The methods used to collect behavioral data are varied. For instance, researchers have used platform-specific (e.g., Facebook, Twitter/X) API (Application programming interface) approaches, Data Download Packages (DDPs) from those same platforms, screenshots from smartphone battery usage or from Screen Time (iPhone) or Digital Wellbeing (Android) tools, network traffic monitoring, and device-level logging. While these methods all provide objective measures, they differ in granularity and the type of research question they can help answer best.
To address the research questions posed in this dissertation, I specifically focus on utilizing event logs collected from both smartphones (Android) and personal computers (Windows/Mac). These event logs include timestamped data on app launches, screen activation, and incoming notifications. The decision to use event logs stems from the context of the broader project in which the empirical chapters of this dissertation are situated. This project, called project DISCONNECT, aimed to ‘unravel the science behind digital well-being’. The project, which is still ongoing, places the “always-on” nature of our digital technologies at the center. The fine-grained timing of these event logs was expected to support the modelling of short-term and within-person processes. Yet, the project was also conceived as an opportunity to critically examine whether such data truly deliver on their promise, or whether the shift from self-report to behavioral measurement introduces new challenges of its own.
This critical examination is warranted, as despite their advantages, behavioral data come with their own limitations, or, more precisely, with a set of validity considerations that differ from, but do not necessarily improve upon, those of self-reports. This dissertation focuses on three such challenges. The first concerns that of operationalization, or – in reverse – the theoretical interpretation of digital trace data: although digital trace data provide precise records of digital behavior, they lack information on aspects such as who performed the behavior, in which state or context, or what content was consumed. This means they do not speak for themselves, but sometimes require theoretical interpretation to become meaningful. For instance, Birk and Samuel (2020) argue in their sociological analysis of the use of trace data for digital phenotyping, that the lack of information on these aspects may lead to misinterpretations: An individual who spends hours voice calling, for example, may be interpreted as ‘being highly social’ while in reality the participant may have been on the line with a helpdesk. In short, without insights into how individuals experience or interpret digital behaviors, digital traces risk being used as proxies of cognitive-behavioral states that they fail to reliably represent. This raises the question of how such indicators can be meaningfully linked to subjective experiences of what they are supposed to represent.
The second challenge concerns granularity: At which level of detail should behavioral indicators be operationalized. Many studies have relied on daily or weekly aggregated metrics at the device level (e.g., total daily smartphone duration), which risks obscuring short, repeated or fragmented interactions which may increasingly characterize everyday digital behavior. Although emerging work suggests that more granular and dynamic indicators may offer greater inferential value (Rodman et al., 2024; Siebers et al., 2024), it remains unclear whether this increased complexity translates into meaningfully stronger or more theoretically coherent associations with well-being outcomes. This dissertation therefore treats granularity as an empirical question to be tested.
The third challenge concerns measurement scope: current research often employs a narrow measurement scope by focusing almost exclusively on smartphones as proxies for digital media use more broadly, rarely considering trace data from alternative digital devices such as laptops, tablets or televisions – leaving the question about ‘total screen time’ underspecified. Although this issue is often mentioned in discussions of study limitations (e.g., Johannes et al., 2021; Mahalingham et al., 2023; Verbeij et al., 2021), its implications for the validity of empirical claims remain underexplored. Excluding other digital devices should therefore not be seen as a minor gap, but rather as a potential source of systematic bias in estimating both the extent of digital engagement and its association with well-being.
These three challenges show that advances in behavioral measurement may not automatically translate into new empirical or theoretical insight. Instead, they show the need to align theoretical interpretation, granularity and data scope with the research questions put forward. The present dissertation addresses these challenges by examining how different measurement choices shape what can be validly inferred about digital media use and well-being in everyday life. To this end, I draw from a dataset collected within the ERC Starting Grant project DISCONNECT. Before presenting the empirical studies, I first present a literature review in which I elaborate on the paradigm shift towards behavioral data, to then introduce the central research questions guiding this dissertation.
1.1 Literature Review
In this brief literature review, I reflect on the paradigm shift from self-reported screen time behavior to behavioral data, to then discuss the challenges of theoretical interpretation, granularity and data scope.
1.1.1 Paradigm shift: From self-reported screen time to behavioral data
Accurately recalling the time spent on digital devices, apps or platforms is a cognitively demanding task: Digital interactions are frequent, distributed across multiple devices, and often fragmented throughout the day, making retrospective estimation inherently unreliable (Andrews et al., 2015; Deng et al., 2019; Tkaczyk et al., 2024). Digital devices are used for many different purposes, to engage with many different platforms, and often in short bursts, possibly even entailing many task switches (Deng et al., 2019; Jeong et al., 2020; Toth et al., 2025). Apart from this complexity, a normative1 dimension to self-reporting digital behavior has been shown to be present as well. This can lead participants to both over- and underestimate digital behavior, based on what a “normal” amount of time should look like (Boyle et al., 2022; Johannes et al., 2021).
Validation research confirms the discrepancies between self-reports and behavioral data. In a recent meta-analysis, Parry and colleagues (2021) found only moderate correlations between the two, concluding that self-reports often are unable to accurately represent actual behavior. These discrepancies are present in both directions, with studies such as Ohme and colleagues (2021) finding signs of clear underestimations, while studies such as Verbeij et al. (2021) showing overestimation of time spent on social media. The occurrence of these inaccuracies will hence impact the validity of any associations found between self-reported digital device use and well-being.
As a response, scholars have increasingly turned towards the collection of passively gathered behavioral data as a more objective alternative. These data are often considered the “gold(en) standard” (e.g., Goedhart et al., 2015; Leitgöb & Keusch, 2026). The use of these data, however, is not without problems or challenges. I will focus on three of these in the current literature review.
1.1.2 The first challenge: From behavioral objectivity to subjective experience
Digital trace data record when, what, and how long someone interacted with their device, but they do not reveal how these interactions were experienced. In research on well-being this distinction can be critical, as these outcomes-of-interest are often inherently subjective. Consider, for instance, that thirty minutes of social media use may be experienced by one person as a relaxing activity, and by another as procrastination that interferes with more important goals. In both cases, the behavioral record is identical, but the psychological implications may differ in ways that trace data alone cannot capture. Consequently, the theoretical interpretation of objective behavior as a representation of a particular psychological state is never straightforward, even when the underlying behavioral data are collected at a high temporal resolution.
The concept of digital wellbeing has recently been proposed to represent a dynamic system, where subjective experiences in relation to digital media should be seen as a function of person-, context-, and device-specific factors that dynamically interact over time (Vanden Abeele, 2021). Many of these factors, such as affective and cognitive appraisals of digital connectivity, or momentary motivational states, are inaccessible through behavioral observation alone, yet come with an underlying assumption that digital traces may give expression to them, as a ‘digital phenotypic’ manifestation of them. For instance, studies imply that ‘availability pressure’, defined as “a perceived obligation to respond quickly to incoming calls or messages and to regularly check the smartphone for potentially incoming notifications” might be represented by the behavior of frequently checking the phone for incoming notifications (e.g., 2026, p. 1). This underlying assumption suggests that the more fundamental question for digital well-being research is perhaps not whether behavior can be accurately measured, but whether these measures can be meaningfully translated into the subjective psychological states that they are meant to represent.
Recent theoretical work has begun to formalize this distinction. Wolfers (2024), for example, proposes a conceptual separation between objective media use and media use perceptions. Objective media use refers to externally observable behavior, e.g., duration and frequency, which can be captured through external agents such as logging apps. Media use perceptions, by contrast, refer to the mental model an individual has of their own media use, which Wolfers further differentiates into a knowledge and an evaluative component. This knowledge component encompasses descriptive perceptions and categorizations, such as one’s own estimations of media use duration or classifying that media use as work or non-work, for example. The evaluative component consists of moral judgements and self-conscious emotions surrounding one’s own media use: for example, evaluating it as excessive, labeling it as procrastination or recovery, and experiencing guilt in relation to this digital behavior. This distinction is consequential because, as Wolfers argues, evaluating media use negatively may be more likely to affect mental health than simply holding a descriptive perception of usage time. Earlier empirical work supports this claim, for example by showing that perceiving media use as procrastination versus recovery has a different association with well-being (Reinecke & Hofmann, 2016). These perceptions are not merely biased proxies for objective behavior; they represent distinct constructs with independent explanatory relevance (Sewall et al., 2020; Wolfers, 2024).
A growing body of empirical work provides support for the notion that subjective perceptions shape the relationship between digital behavior and well-being. Ernala et al. (2022) demonstrated that the association between time spent on the platform and subjective well-being was significantly moderated by users’ beliefs about whether Facebook use was good or bad for themselves and for society: When participants believed that Facebook was good, time spent on Facebook was not significantly associated with well-being; when they believed it was bad, increased time spent was associated with lower well-being. This moderating effect was stronger when the time-spent measure was self-reported than when it was derived from server logs, suggesting that the subjective framing of one’s behavior matters more for well-being than the behavior itself.
Work by Lee and Hancock (2024) further developed this line of work, coining the term ‘social media mindsets’, defined as “the core beliefs held by individuals that orient their expectations, behaviors, attributions, and goals regarding the role of social media in their lives” (p.2). They identified two dimensions: agency (perceived control over one’s social media use) and valence (perceived effects as enhancing or harmful). These mindsets explained more variance in psychological well-being than conventional measures of social media use such as self-reported time spent or intensity. Lee and Hancock proposed two routes through which mindsets relate to well-being: an appraisal route, where mindsets moderate how individuals interpret their use (e.g., social media use predicted lower well-being for those with low-agency or negative-valence mindsets), and a behavioral route, where mindsets orient people toward different patterns of engagement (e.g., more agentic mindsets were associated with less passive use).
More recently, Parry and Coetzee (2025) incorporated both self-reported and logged social media use alongside social media mindsets. Their findings partially corroborated and partially challenged the earlier work. Agency mindsets were robustly associated with well-being across all four indicators, replicating Lee and Hancock’s findings. However, valence mindsets were not consistently linked to well-being once logged behavior was controlled for. Importantly, when social media use was assessed through log data rather than self-reports, the associations between mindsets and well-being weakened substantially, and the evidence for the appraisal route became inconsistent. The authors concluded that while subjective experiences may be a better predictor of well-being than the logged amount of use, it may be perceived control over one’s social media use, rather than general evaluative attitudes, that is most consequential (Parry & Coetzee, 2025).
This body of work yields several insights directly relevant for the use of digital trace data in well-being research. First, subjective perceptions of media use are not merely noisy reflections of objective behavior; they are likely separate constructs with their own explanatory power (Wolfers, 2024). Second, how users interpret and evaluate their digital behavior can moderate, or sometimes even overpower, the association between objective use and well-being (Ernala et al., 2022; Lee & Hancock, 2024). Third, not all subjective perceptions may be equally consequential. Perceptions of control appear to be more robustly linked to well-being than general evaluative judgements, and the strength of these associations depends on whether behavior is measured through self-report or through log data (Parry & Coetzee, 2025). This pattern aligns with Vanden Abeele’s (2021) conceptualization of digital well-being as a dynamic system in which person-specific appraisals, rather than behavioral exposure per se, shape how media use becomes consequential for well-being. Both behavioral traces and subjective perceptions can hence be seen as complementary sources of evidence, each capturing a different facet of the media use experience (Wolfers, 2024). The question for digital well-being research is therefore not whether to use behavioral or subjective data, but how to meaningfully integrate both in order to better understand the mechanisms through which digital behavior relates to well-being.
If subjective experiences complement behavioral data, the methodological implication is that research designs should aim to capture both data streams in tandem, ideally at compatible temporal resolutions. This has motivated researchers to combine passively collected trace data with experience sampling methods, producing “linkage” designs that temporally align behavioral logs with momentary self-reports of subjective states and experiences (cf., Harari & Gosling, 2023; Klingelhoefer et al., 2025). Such designs allow researchers to examine not only whether behavior is associated with well-being, but through which subjective mechanisms associations emerge. In doing so, they operationalize complementarity between behavioral and subjective indicators that the theoretical and empirical work reviewed above calls for.
Emerging empirical work using such combined designs illustrates both the potential and current limits of this approach. The study by De Segovia et al. (de Segovia Vicente et al., 2024), to which I contributed, combined passively logged smartphone use with mobile momentary assessments to examine the downstream consequences of mindless scrolling, a specific form of social media consumption characterized by a reduced awareness and the absence of intentional goals. The findings revealed that longer periods of mindless scrolling were associated with increased guilt over smartphone use, a relationship that was partially mediated by the experience of goal conflict. Guilt accumulated over the course of a day and predicted lower end-of-day well-being. This psychological pathway – from behavior to goal conflict, through guilt, to diminished well-being – would be entirely invisible in behavioral data alone. The log data capture when and how long a person scrolled, but understanding whether and why that scrolling became consequential for well-being required access to the subjective mediators that connected the behavior to its outcomes.
Similarly, Elmer et al. (2025) combined smartphone log data with ESM to examine bidirectional within-person associations between smartphone usage and momentary well-being. Their results indicated that smartphone use in the hour before an ESM assessment, and particularly social media use, was associated with reduced affect valence and increased loneliness on the within-person level. Loneliness, in turn, predicted increased subsequent smartphone use, suggesting a bidirectional dynamic. However, the effects were consistently small, and the authors observed that individuals who generally scored higher on loneliness were more sensitive to the momentary effect of social media use. This between-person moderation, which aligns with susceptibility-based accounts of media effects (cf. Valkenburg et al., 2021), again shows the importance of capturing subjective states alongside behavioral indicators: the same amount of logged social media use related differently to well-being depending on the person’s broader psychological profile.
The first aim of this dissertation is to contribute to the growing body of work focusing on the subjective mechanisms through which behavioral patterns become consequential for well-being (or perhaps rather fail to do so?). The theoretical and empirical work reviewed above establishes that digital trace data, while providing precise records of device interaction, lack the capacity to capture how users interpret and evaluate their own digital behavior – processes that appear to play a meaningful role in shaping well-being. Research designs that combine logged behavior with momentary subjective assessments can help address this gap, yet the field is still in the early stages of understanding which subjective mechanisms matter most, how they interact with specific behavioral patterns, and what this implies for the status of trace data as indicators of well-being. The first sub-research question of this dissertation therefore asks:
S-RQ1: How do passively logged behavior indicators correspond with the cognitive-behavioral states they supposedly represent, and what does this imply for the status of log data as theoretically interpretable indicators of well-being?
1.1.3 The second challenge: Granularity and the operationalization of digital behavior
While including subjective experiences is essential for interpreting the psychosocial meaning of digital behavior, it does not by itself resolve the decision over which digital traces should be represented, and how they should be calculated. This shifts attention to an issue that perhaps precedes subjective appraisal: the level of granularity at which digital behavior is operationalized into ‘features’. The behavioral features selected for analysis serve as more than predictors; they define the space for possible associations with subjective states. Research that only operationalizes total duration, for instance, is unable to detect whether fragmented versus sustained use patterns of specific app categories relate differently to momentary experiences of stress or distraction.
Much of the existing research, whether using self-report or trace data, has relied on highly aggregated measures, most prominently total duration or frequency over a given time window (Parry et al., 2021). While these types of measures are perhaps more analytically convenient or intuitive, they treat digital media use as a homogenous concept (Kaye et al., 2020). This homogeneity assumption clashes with theoretical models that conceptualize digital well-being as a dynamic system in which person-, context-, and device-specific factors interact over time (Vanden Abeele, 2021). If well-being is shaped by the temporal patterns and contextual embedding of digital engagement, beyond its total volume, then aggregate indicators may be unable to capture the processes they are meant to represent. Recent empirical and theoretical work has increasingly questioned this assumption, and has indicated that, rather than the amount of time spent using digital technologies, how we spend or distribute that time is most relevant (Kaye et al., 2020; Meier, 2021).
Indeed, studies combining trace data with experience sampling show that momentary associations between well-being and digital media are often weak or inconsistent when operationalized in terms of total duration, yet become more meaningful when examined through the lens of more specific usage patterns. For instance, Rodman et al. (2024) provide empirical support for this claim, demonstrating that while overall smartphone use showed very weak or no associations with well-being, more granular features—such as engagement with distinct categories or platforms—were more predictive. Similarly, Elmer et al. (2025) found that total smartphone duration showed limited associations with well-being, while, when looking at specific app category analysis, social media appeared to be the main driver for these associations, with other categories largely showing no association. These findings lead to the conclusion that undifferentiated measures of screen time may systematically diminish observable associations with well-being outcomes.
Moreover, using digital traces allows researchers to go beyond coarse measures of “screen time”, enabling data collection at a level which can capture sequences or fluctuations of use (and non-use). This opens new possibilities to operationalize dimensions of digital behavior which would be very difficult to assess through the use of self-reports. Examples of such dimensions are fragmentation (Siebers et al., 2024), burstiness (Jo et al., 2012), or checking behavior or “glances” (Parry & Toth, 2025). These dimensions might be mapped more cleanly onto theoretically relevant behavior or psychological states, such as procrastination or distraction, than total duration or frequency alone (e.g., (Aalbers et al., 2021; Siebers et al., 2024). Many of these features are inherently dynamic: they capture not just the volume of behavior within a given period, but its temporal structure, i.e., how use is distributed over time. This temporal dimension aligns with theoretical calls for measurement approaches that match the timescale at which digital well-being processes unfold (Klingelhoefer et al., 2025; Vandenbosch et al., 2025), and its importance can be recognized in intensive longitudinal designs increasingly being adopted (e.g., Elmer et al., 2025; große Deters & Schoedel, 2024; Rodman et al., 2024).
These findings from extant research suggest that granularity should not just be reflected in data collection and afterwards flattened in highly aggregated measures, but rather be a core part of the operationalization of digital behavior. Aggregate indicators may obscure short-term processes, within-person variability and the broader context wherein this behavior takes place. Given that research embracing more granular operationalizations of screen time is still growing, whether this increased granularity translates into empirically stronger or more theoretically coherent associations with well-being, however, remains somewhat of an open question. The empirical chapters in this dissertation address this question by comparing how aggregate versus granular behavioral features – including category-specific duration, but also fragmentation and notification-based indicators – relate to momentary well-being states such as time pressure (Chapter 4) and mood (Chapter 5). Consequently, my second sub-research question asks:
S-RQ2: How do granular and dynamic features of digitally logged behavior relate to momentary well-being outcomes in daily life?
1.1.4 The third challenge: Measurement scope and device coverage
Finally, a third limitation of much log-based research on digital media use and well-being concerns the limited scope of behavioral data collection. Despite methodological advances in passive sensing, digital trace data are most often collected from a single device, most commonly the smartphone, thereby implicitly treating this device as a proxy for someone’s broader digital engagement (Parry et al., 2021). In measurement terms, this amounts to a form of construct underrepresentation: the operationalization of “digital media use” systematically excludes portions of the behavioral domain it aims to capture. This risks providing only a partial view of someone’s everyday media behavior – a concern that extends beyond smartphone logging studies to other digital trace traditions, such as web tracking research, where the failure to capture data from all devices participants use to go online has been termed tracking undercoverage (Bosch et al., 2025).
A growing body of work has begun to acknowledge this limitation. For instance, a recent meta-analysis by Parry et al. (2021) shows that nearly all studies investigating log-based digital media use relied exclusively on smartphone data, with only a handful of studies incorporating passive sensing data on the use of computers, gaming devices or general internet use. Similarly, while studies have increasingly emphasized the importance of differentiating between different types of media use, differences in device use have often not been accounted for when examining well-being outcomes (e.g., Liu et al., 2024). Even work comparing the accuracy of self-reported social media use to log-based counterparts has explicitly acknowledged that their conclusions were limited to smartphone data alone, thereby excluding any usage on tablets, laptops or web-based versions of social media platforms (Verbeij et al., 2021).
Emerging evidence suggests that these are not merely theoretical concerns. In their web tracking study, Bosch et al. (2025) demonstrated that 74% of participants in an online panel had at least one device untracked, and that this undercoverage produced substantial biases in estimates of online media exposure, with the magnitude and even direction of bias varying depending on the statistic of interest. Similarly, Pankowska et al. (2025) showed, using hidden Markov models applied to Facebook tracking data and self-reports, that digital trace measures severely underestimated the frequency of Facebook use for approximately one-third of their sample, driven specifically by incomplete device coverage. While these studies focus on web tracking rather than smartphone logging, they provide a direct empirical warning: when the measurement instrument fails to capture behavior across all relevant devices, the resulting data can be substantially biased.
Consequences of this narrow measurement scope extend beyond underestimating overall levels of digital engagement. Research that only focuses on smartphone-based activity is unable to capture how cross-device patterns might shape associations with well-being. For example, receiving and reading a short email message on the smartphone can initiate a much longer working session on the computer. As we will demonstrate in Chapter 5, more than half of email screen time in our sample occurred on computers rather than smartphones, meaning that smartphone-only data would miss the majority of this particular category of use. Prior research already shows that being interrupted by, and needing to manage, overlapping roles and tasks can lead to negative well-being outcomes, such as exhaustion (Derks et al., 2021) and stress (Cornwell, 2013). In the above example, for instance, a smartphone-only study would therefore capture the trigger of such a sequence while remaining blind to the bulk of the resulting behavior and its potential well-being consequences.
When activities performed on certain devices are omitted, any observed association would therefore reflect only a subset of relevant behavior, leading to imprecise estimates or biased inferences of these effects. In such cases, these estimates could misrepresent the magnitude or even the direction of associations between digital media use and well-being. Critically, such bias need not be uniform or predictable. Bosch et al. (2025) demonstrated through simulations of device undercoverage that the direction and magnitude of bias depend on the statistic of interest: while simple count measures (e.g., total time online) are consistently underestimated when devices are missed, proportional or relational measures can be biased in either direction. In their data, undercoverage weakened certain correlations while inflating others. In the context of digital media use and well-being research, this means that single-device estimates could misrepresent not only the size but also the direction of associations.
More broadly, focusing on a single device constrains the kinds of theoretical questions that can be addressed. People’s daily media repertoires are increasingly characterized by media multiplexity (Chan, 2015; Haythornthwaite, 2005), where communication takes place across multiple channels, platforms, and devices. Individuals do not interact with devices in isolation, but navigate a device-ecology, a system of interconnected screens across which attention, tasks and media engagement are distributed throughout the day. Capturing digital traces from computers alongside smartphones opens up the possibility of modelling cross-device dynamics, such as device switching, parallel use and multi-device workflows. As we will show in Chapter 5, such dynamics are far from rare: in our data participants switched between their smartphone and computer approximately 66 times per day, with nearly 90% of these switches also involving a change in the category of media activity. These patterns are invisible in single-device datasets, but could be central in emerging theories of media multitasking and attentional fragmentation (Abrahamse et al., 2025).
Ignoring computer-based log data should therefore not be seen as a minor omission, but rather a substantial limitation on what can validly be inferred about digital media use and its relationship with well-being. The evidence reviewed above, ranging from near-universal reliance on smartphone-only data in the literature (Parry et al., 2021), to the demonstrated biases that device-incomplete tracking introduces in media exposure estimates Pankowska et al. (2025), suggests that single-device measurement threatens both the content validity of digital media use constructs and the accuracy of their estimated associations with well-being outcomes. Hence, as a final sub-research question, this dissertation asks:
S-RQ3: To what extent does reliance on single-device (smartphone) log data create systematic validity gaps in estimating digital media use and its association with well-being?
By situating this question at the intersection of measurement scope and theoretical inference, this dissertation aims to contribute to the methodological debate, but also to clarify the boundaries of what digital trace data can – and cannot – tell us about digital media use and well-being. Across the three sub-research questions developed above, a common thread emerges. Whether the challenge concerns the interpretative gap between behavioral indicators and cognitive-behavioral states (S-RQ1), the level of granularity at which digital behavior is operationalized (S-RQ2), or the scope of devices from which behavioral data are collected (S-RQ3), each reflects a different facet of the same underlying problem: the choices researchers make in turning raw digital traces into analyzable constructs shape, and possibly constrain, the validity of the conclusions that follow. Taken together, the main research question of this dissertation therefore asks:
RQ: To what extent do the theoretical interpretation, granularity, and scope of digital trace data shape what can be validly inferred about digital media use and well-being?
1.2 Chapter Overview
This dissertation presents an inquiry into the above research questions, in the form of several empirical investigations. It includes three empirical chapters (Chapters 3-5) in which variations of digital trace measures are combined with self-reported well-being outcomes, contextual information, and subjective mechanisms linking digital behavior to experience. Each chapter addresses one or more sub-research questions, focusing respectively on psychological interpretation (Chapter 3), behavioral granularity (Chapter 4), and measurement scope (Chapter 5). Table 1.1 provides an overview of how each chapter contributes to each sub-research question.
In addition to the empirical chapters, Chapter 2 outlines the overarching methodological framework – including experience sampling design, digital trace data collection procedures, and analytic strategies – and situates the studies within the broader DISCONNECT project. Together, these chapters examine how digital trace data can be used, methodologically and theoretically, to study the relationship between digital media use and well-being.
Chapter 3 (“Connected Yet Cognitively Drained”) investigates whether online vigilance promotes mental fatigue, especially under perceived availability pressure, and whether passively sensed smartphone behavior can serve as a digital proxy for this subjective state. By combining passively sensed behavioral features with self-reported measures of online vigilance and mental fatigue, this chapter tests whether behavioral indicators capture the same psychological phenomenon that self-reports do, and whether both pathways lead to comparable well-being outcomes. The findings reveal that behavioral smartphone features are only weakly associated with self-reported online vigilance, suggesting that trace data alone may not reliably represent the subjective experience thereof. This chapter thus primarily relates to S-RQ1, questioning the conditions under which passively logged measures can validly stand in for psychologically meaningful states. Like Chapter 4, this chapter computes its behavioral indicators exclusively from smartphone data, and the absence of computer-based traces is considered a scope limitation that Chapter 5 subsequently addresses.
Chapter 4 (“Always On, Always Rushed for Time”) explores how momentary patterns of mobile communication, operationalized through four behavioral features (duration, frequency, fragmentation and notifications received) across four app categories (email, social media, chat, and work communication), relate to feeling rushed for time, both directly and indirectly through perceived task juggling load. As such, it shifts attention to the level of granularity at which digital behavior is operationalized. By demonstrating that granular, category-specific features reveal associations that aggregate measures obscure, this chapter primarily contributes to S-RQ2. Its finding that the link between behavioral patterns and time pressure is mediated by a subjective mechanism (perceived juggling load) also connects to S-RQ1, reinforcing that trace data require interpretative framing to become psychologically meaningful. As with Chapter 3, this study operates within a single-device (smartphone) measurement scope.
Chapter 5 (“Computers Matter”) combines experience sampling with both passively logged smartphone and computer trace data, and investigates how the inclusion of computer-based and cross-device behavioral features – such as device switching, context switching, and overlapping use – changes both the descriptive estimates of screen time and the observed associations between digital media use and momentary and daily mood. This chapter directly addresses the measurement scope limitation that Chapters 3 and 4 acknowledged but could not resolve. It further demonstrates that smartphone-only measures substantially underestimate total digital engagement, that cross-device dynamics constitute a common yet previously underexamined dimension of everyday digital behavior, and that including computer-based and cross-device features meaningfully alters the observed associations between media use and affect. This chapter primarily contributes to S-RQ3, while its introduction of novel cross-device behavioral features also advances S-RQ2 by expanding the repertoire of granular indicators available for modelling digital media use.
Chapter 6 (“Making the Impossible Possible”) complements the empirical work by addressing a practical barrier to the multi-device measurement strategies that our other chapters call for. In response to Apple’s restrictive data access policies, which have historically prevented researchers from obtaining detailed behavioral data from iOS devices, this chapter presents a novel data donation procedure that leverages the built-in Screen Time synchronization feature across Apple devices to extract app-level usage data from iPhones, iPads, Apple Watches, and Macs linked to the same Apple ID. In addition, we provide an open-source parsing tool that converts the resulting system-level files into researcher-ready datasets. By offering a non-intrusive, privacy-sensitive, and technically accessible method for capturing iOS screen time data, this chapter directly contributes to S-RQ3, extending measurement scope beyond the Android-only focus of Chapters 3 and 4 (and much of the digital trace literature), and supporting multi-device data collection approaches that Chapter 5’s findings call for.
Chapter 7 (Discussion) synthesizes the findings of the preceding chapters and discusses their implications for the use of trace data in digital-wellbeing research, for broader debates within the field, and for the societal context in which screen time research is situated, before addressing the limitations of this dissertation.
Finally, beyond the chapters included in this dissertation, the collaborative nature of the DISCONNECT project has also allowed me to contribute to several related papers by colleagues within the project team.
For instance, Johannes et al. (2021, p. 2) state how “Broadly speaking, inaccuracy in the form of measurement error can be exclusively random or a mix of random and systematic error. For example, people systematically overestimate their physical activity (Klesges et al., 1990) and underestimate their smoking (Connor Gorber et al., 2009)—both are examples of social desirability, one important factor in shaping inaccurate self-reports of behavior (for an overview of other factors, see Schwarz & Oyserman, 2001).”↩︎