Data extraction

In total, active data collection methods were used in more than half of the included studies (61%, see online Supplementary Table S2). Twenty-nine publications (17%) reported findings from dedicated sensors (see online Supplementary Table S3), and 8 studies (5%) from mobile sensing (see online Supplementary Table S4). In 30 studies (17%), a combination of active and passive methods for collecting intensive time series data was used (see online Supplementary Table S5).

Active data collection methods

The most commonly used active data collection method was the ESM (in 96 of 108 studies). Various sampling techniques were used in these studies, including event-contingent designs (e.g. Tasca et al., Reference Tasca, Illing, Balfour, Krysanski, Demidenko, Nowakowski and Bissada2009), time-contingent designs (e.g. Collip et al., Reference Collip, Wigman, van Os, Oorschot, Jacobs, Derom and Myin-Germeys2014) with (semi-) random or fixed sampling schedules, or hybrid designs (i.e. combining event- and time-contingent designs) (e.g. Smyth et al., Reference Smyth, Wonderlich, Sliwinski, Crosby, Engel, Mitchell and Calogero2009). In ESM studies, the sampling frequency ranged from three to ten assessments per day, whereas in the twelve telephone/ email studies included, the sampling frequency was between four times per day to once per week. The assessment period ranged between two days to two years (see online Supplementary Table S2). Notably, there was considerable heterogeneity in sampling designs and ESM measures.

Applications and target constructs

Next, we extracted the most common target constructs in the identified studies.

Most studies (i.e. 75 studies), used the ESM to capture self-reported positive and negative affect (e.g. Collip et al. Reference Collip, Nicolson, Lardinois, Lataster, van Os and Myin-Germeys2011a; Fitzsimmons-Craft et al. Reference Fitzsimmons-Craft, Accurso, Ciao, Crosby, Cao, Pisetsky and Wonderlich2015; Hartmann et al. Reference Hartmann, Wichers, Menne-Lothmann, Kramer, Viechtbauer, Peeters and Simons2015; Haynos et al. Reference Haynos, Crosby, Engel, Lavender, Wonderlich, Mitchell and Le Grange2015; Lavender et al. Reference Lavender, Utzinger, Crosby, Goldschmidt, Ellison, Wonderlich and Le Grange2016). To this end, e.g. items from the Positive and Negative Affect Schedule (PANAS; Watson, Clark, and Tellegen, Reference Watson, Clark and Tellegen1988) have been used. In addition, systematic variation in affective states over time (sometimes referred to as emotional instability) was frequently investigated (e.g. Johns et al. Reference Johns, Di, Merikangas, Cui, Swendsen and Zipunnikov2019, Solhan, Trull, Jahng, & Wood, Reference Solhan, Trull, Jahng and Wood2009; Wonderlich et al. Reference Wonderlich, Lavender, Wonderlich, Peterson, Crow, Engel and Crosby2015). As an alternative to assessing discrete emotions, ratings of valence and arousal have been used to capture affective states (Becker, Fischer, Crosby, Engel, & Wonderlich, Reference Becker, Fischer, Crosby, Engel and Wonderlich2018). Affective processes have been examined in at-risk samples, or samples of service users diagnosed with various mental disorders, such as depressive (Hartmann et al., Reference Hartmann, Wichers, Menne-Lothmann, Kramer, Viechtbauer, Peeters and Simons2015; Kordy et al., Reference Kordy, Wolf, Aulich, Burgy, Hegerl, Husing and Backenstrass2016; Simons et al., Reference Simons, Hartmann, Kramer, Menne-Lothmann, Hohn, van Bemmel and Wichers2015), bipolar (Tsanas et al., Reference Tsanas, Saunders, Bilderbeck, Palmius, Osipov, Clifford and De Vos2016), psychotic (Collip et al., Reference Collip, van Winkel, Peerbooms, Lataster, Thewissen, Lardinois and Myin-Germeys2011c; Lataster et al., Reference Lataster, van Os, de Haan, Thewissen, Bak, Lataster and Myin-Germeys2011; Oorschot et al., Reference Oorschot, Lataster, Thewissen, Lardinois, van Os, Delespaul and Myin-Germeys2012), eating (Berner et al., Reference Berner, Crosby, Cao, Engel, Lavender, Mitchell and Wonderlich2017; Fitzsimmons-Craft et al., Reference Fitzsimmons-Craft, Accurso, Ciao, Crosby, Cao, Pisetsky and Wonderlich2015; Haynos et al., Reference Haynos, Crosby, Engel, Lavender, Wonderlich, Mitchell and Le Grange2015), anxiety (Silk et al., Reference Silk, Tan, Ladouceur, Meller, Siegle, McMakin and Ryan2018), and personality (Andrewes, Hulbert, Cotton, Betts, & Chanen, Reference Andrewes, Hulbert, Cotton, Betts and Chanen2017; Chapman, Rosenthal, Dixon-Gordon, Turner, & Kuppens, Reference Chapman, Rosenthal, Dixon-Gordon, Turner and Kuppens2017) disorders (see supplementary tables).

ESM has also been applied to assess cognitive processes in daily life. In psychosis research, psychotic experiences [e.g. subclinical expressions of disordered thinking, paranoia, delusions (Collip et al., Reference Collip, Nicolson, Lardinois, Lataster, van Os and Myin-Germeys2011a; Collip et al., Reference Collip, van Winkel, Peerbooms, Lataster, Thewissen, Lardinois and Myin-Germeys2011c; Klippel et al., Reference Klippel, Myin-Germeys, Chavez-Baldini, Preacher, Kempton, Valmaggia and Reininghaus2017; Reininghaus et al., Reference Reininghaus, Gayer-Anderson, Valmaggia, Kempton, Calem, Onyejiaka and Morgan2016b; Reininghaus et al., Reference Reininghaus, Kempton, Valmaggia, Craig, Garety, Onyejiaka and Morgan2016c)], threat anticipation (Klippel et al., Reference Klippel, Myin-Germeys, Chavez-Baldini, Preacher, Kempton, Valmaggia and Reininghaus2017; Perez Arribas, Goodwin, Geddes, Lyons, & Saunders, Reference Perez Arribas, Goodwin, Geddes, Lyons and Saunders2018), and aberrant salience (Klippel et al., Reference Klippel, Myin-Germeys, Chavez-Baldini, Preacher, Kempton, Valmaggia and Reininghaus2017; Reininghaus et al., Reference Reininghaus, Kempton, Valmaggia, Craig, Garety, Onyejiaka and Morgan2016c) are important target constructs that have been captured using ESM. In eating disorder research, momentary assessments of social comparison processes and thoughts of compensatory behaviour added evidence to existing theories with high ecological validity (Leahey, Crowther, & Ciesla, Reference Leahey, Crowther and Ciesla2011). In addition, ESM has been used to capture worrying or rumination (Khazanov, Ruscio, & Swendsen, Reference Khazanov, Ruscio and Swendsen2019; Ruscio et al., Reference Ruscio, Gentes, Jones, Hallion, Coleman and Swendsen2015), and dissociation (Mason et al., Reference Mason, Lavender, Wonderlich, Steiger, Cao, Engel and Crosby2017). More recently, experimental experience sampling tasks have been developed to measure momentary cognition (Reininghaus et al., Reference Reininghaus, Oorschot, Moritz, Gayer-Anderson, Kempton, Valmaggia and Myin-Germeys2019). However, compared to affective processes, overall, cognitive processes have been less frequently studied using ESM in mental health research. This might be due to more frequent fluctuations in cognitive constructs and to difficulty accessing these processes using traditional ESM measures (Daniels et al., Reference Daniels, Bartels, Verhagen, Van Knippenberg, De Vugt and Delespaul2020).

In 31 studies, ESM has also been used to assess the occurrence of specific – often dysfunctional – behavioural patterns in daily life [e.g. self-harm (Muehlenkamp et al., Reference Muehlenkamp, Engel, Wadeson, Crosby, Wonderlich, Simonich and Mitchell2009)]. Momentary behaviour has either been assessed by event-contingent or time-contingent sampling schemes. Offering an appealing alternative to retrospective self-report measures, some studies assessed momentary behaviour, such as substance use (Fatseas, Serre, Swendsen, & Auriacombe, Reference Fatseas, Serre, Swendsen and Auriacombe2018; Serre, Fatseas, Denis, Swendsen, & Auriacombe, Reference Serre, Fatseas, Denis, Swendsen and Auriacombe2018), intoxication (Mackesy-Amiti & Donenberg, Reference Mackesy-Amiti and Donenberg2020; Pisetsky et al., Reference Pisetsky, Crosby, Cao, Fitzsimmons-Craft, Mitchell, Engel and Peterson2016), and mode of consumption (Mackesy-Amiti & Donenberg, Reference Mackesy-Amiti and Donenberg2020) in the daily lives of individuals with substance use disorders. Moreover, ESM has gained increasing attention in research focusing on dysfunctional behaviour in the spectrum of eating disorders to assess, for example, restrictive eating, binge eating, and purging (Fitzsimmons-Craft et al., Reference Fitzsimmons-Craft, Accurso, Ciao, Crosby, Cao, Pisetsky and Wonderlich2015; Lavender et al., Reference Lavender, Utzinger, Crosby, Goldschmidt, Ellison, Wonderlich and Le Grange2016; Schaefer et al., Reference Schaefer, Smith, Anderson, Cao, Crosby, Engel and Wonderlich2020; Smyth et al., Reference Smyth, Wonderlich, Sliwinski, Crosby, Engel, Mitchell and Calogero2009; Zunker et al., Reference Zunker, Peterson, Crosby, Cao, Engel, Mitchell and Wonderlich2011).

In the included studies, a strong emphasis was placed on contextual factors such as participants' current location (Mackesy-Amiti & Donenberg, Reference Mackesy-Amiti and Donenberg2020; Rintala, Wampers, Myin-Germeys, & Viechtbauer, Reference Rintala, Wampers, Myin-Germeys and Viechtbauer2019) and activities (Leendertse et al., Reference Leendertse, Myin-Germeys, Lataster, Simons, Oorschot and Lardinois2018; Oorschot et al., Reference Oorschot, Lataster, Thewissen, Lardinois, van Os, Delespaul and Myin-Germeys2012), but also social context [e.g. being alone or in company, interpersonal stressful events (Collip et al., Reference Collip, Oorschot, Thewissen, Van Os, Bentall and Myin-Germeys2011b; Morgan et al., Reference Morgan, Lee, Wright, Gilchrist, Forbes, McMakin and Silk2017; Tasca et al., Reference Tasca, Illing, Balfour, Krysanski, Demidenko, Nowakowski and Bissada2008)]. Moving beyond the assessment of context, appraisals of the context have gained increasing attention. For example, appraisals of the unpleasantness of events, activities, and social situations have been used to operationalise different types of stress (Collip et al., Reference Collip, Nicolson, Lardinois, Lataster, van Os and Myin-Germeys2011a; Klippel et al., Reference Klippel, Viechtbauer, Reininghaus, Wigman, van Borkulo and Wichers2018; Peerbooms et al., Reference Peerbooms, Rutten, Collip, Lardinois, Lataster, Thewissen and van Winkel2012; Reininghaus et al., Reference Reininghaus, Gayer-Anderson, Valmaggia, Kempton, Calem, Onyejiaka and Morgan2016b). In addition, specific processes such as social satisfaction (Collip et al., Reference Collip, Wigman, van Os, Oorschot, Jacobs, Derom and Myin-Germeys2014) or perceived rejection (Scott et al., Reference Scott, Wright, Beeney, Lazarus, Pilkonis and Stepp2017) have been examined using ESM.

Thus, overall, ESM research to date has commonly examined a combination of affective, cognitive, and behavioural processes taking into account contextual factors to further elucidate the role of candidate momentary mechanisms (as situated in context) in the development and maintenance of mental health problems and their subjective experience (Erwin, Dennis, Coughlin, Calhoun, & Beckham, Reference Erwin, Dennis, Coughlin, Calhoun and Beckham2019; Khazanov et al., Reference Khazanov, Ruscio and Swendsen2019; Oorschot et al., Reference Oorschot, Lataster, Thewissen, Lardinois, van Os, Delespaul and Myin-Germeys2012). For example, the construct of affective and psychotic stress reactivity, defined as an increased intensity of negative affect and psychotic experiences in response to minor daily stressors, has been widely studied as an important putative momentary mechanism in psychosis research (Collip et al., Reference Collip, van Winkel, Peerbooms, Lataster, Thewissen, Lardinois and Myin-Germeys2011c; Frissen et al., Reference Frissen, Lieverse, Drukker, Delespaul, Lataster, Myin-Germeys and van Os2014; Reininghaus et al., Reference Reininghaus, Kempton, Valmaggia, Craig, Garety, Onyejiaka and Morgan2016c) and in other fields [e.g. eating disorders (Pearson et al., Reference Pearson, Lavender, Cao, Wonderlich, Crosby, Engel and Crow2017), personality disorders (Glaser, Van Os, Mengelers, & Myin-Germeys, Reference Glaser, Van Os, Mengelers and Myin-Germeys2008)], and with respect to transdiagnostic phenotypes (Rauschenberg et al., Reference Rauschenberg, van Os, Cremers, Goedhart, Schieveld and Reininghaus2017)). Investigating the interplay of affective processes and contexts in daily life offers the advantage of operationalising important symptom domains and their subjective experience, as has been done, for example, for negative symptoms such as anhedonia (Oorschot et al. (Reference Oorschot, Lataster, Thewissen, Lardinois, van Os, Delespaul and Myin-Germeys2012). Insights on the momentary dynamics of affective experiences and dysfunctional behaviour have advanced our understanding of the emergence of disordered eating behaviour such as restrictive or binge eating episodes and purging (Berg et al., Reference Berg, Cao, Crosby, Engel, Peterson, Crow and Wonderlich2017; Engel et al., Reference Engel, Wonderlich, Crosby, Mitchell, Crow, Peterson and Gordon2013; Fitzsimmons-Craft et al., Reference Fitzsimmons-Craft, Accurso, Ciao, Crosby, Cao, Pisetsky and Wonderlich2015; Goldschmidt et al., Reference Goldschmidt, Wonderlich, Crosby, Engel, Lavender, Peterson and Mitchell2014; Haynos et al., Reference Haynos, Crosby, Engel, Lavender, Wonderlich, Mitchell and Le Grange2015; Schaefer et al., Reference Schaefer, Smith, Anderson, Cao, Crosby, Engel and Wonderlich2020). ESM measures of temporal dynamics in affect, craving, and substance use behaviour in daily life can now be used to inform treatment and relapse prevention in substance use disorders (Fatseas et al., Reference Fatseas, Serre, Swendsen and Auriacombe2018; Serre et al., Reference Serre, Fatseas, Denis, Swendsen and Auriacombe2018). An overview on psychometric quality of ESM measures is provided in the supplementary material.

Passive, sensor-based data collection methods

Overall, accelerometers were the most frequently used dedicated sensors to capture time intensive data in the included studies (i.e. in 27 of 29 studies, see online Supplementary Table S3). Only two studies reported pedometer data, i.e., estimating individuals' physical activity based on step count. Other studies using sensors capable of detecting other psychophysiological parameters, such as heart rate or electrodermal activity were not included. The assessment period in the included studies utilising dedicated sensors ranged from two hours to 22 days, and the sampling frequency was between two to 60 s epochs (see online Supplementary Table S3).

Applications and target constructs

Most studies (i.e. 18 out of 29, see online Supplementary Table S3) used sensors to passively monitor physical activity. In the included studies, daytime physical activity was operationalised as a gradient of intensity that includes sleep, sedentary behaviour, light physical activity, moderate to vigorous physical activity, and high intensity physical activity. In addition, the included studies reported frequency and duration of specific types of physical activity, such as sedentary behaviour assessed by a pedometer (Piette et al., Reference Piette, Valenstein, Himle, Duffy, Torres, Vogel and Richardson2011) or accelerometer (Stubbs, Ku, Chung, & Chen, Reference Stubbs, Ku, Chung and Chen2017), as well as gross motor activity (Difrancesco et al., Reference Difrancesco, Lamers, Riese, Merikangas, Beekman, van Hemert and Penninx2019). Further, some of the included studies quantified physical activity on a daily basis using step count (Baerg et al., Reference Baerg, Cairney, Hay, Rempel, Mahlberg and Faught2011), or by calculating a mean activity score from accelerometer data (Benard et al., Reference Benard, Etain, Vaiva, Boudebesse, Yeim, Benizri and Geoffroy2019). Others (Geoffroy et al., Reference Geoffroy, Micoulaud Franchi, Maruani, Philip, Boudebesse, Benizri and Etain2019) reported the average activity during the most active 10-h period.

Another construct that has been extensively investigated using sensors is sleep (11 out of 29 studies, see online Supplementary Table S3). There are different operationalisations for sleep duration. As shown in online Supplementary Table S3, different parameters have been applied to measure sleep quality and duration including e.g., sleep efficiency and start and end time of the rest period (e.g. Blake et al. Reference Blake, Snoep, Raniti, Schwartz, Waloszek, Simmons and Allen2017, Reference Blake, Blake, Schwartz, Raniti, Waloszek, Murray and Allen2018; Fang et al. Reference Fang, Suzuki, Lim, Chung, Ku and Chen2016; Goodlin-Jones, Waters, & Anders, Reference Goodlin-Jones, Waters and Anders2009; McCrae et al. Reference McCrae, Williams, Roditi, Anderson, Mundt, Miller and Robinson2019; Owens et al. Reference Owens, Sangal, Sutton, Bakken, Allen and Kelsey2009; Robillard et al. Reference Robillard, Naismith, Smith, Rogers, White, Terpening and Hickie2014; Verkooijen et al. Reference Verkooijen, Stevelink, Abramovic, Vinkers, Ophoff, Kahn and van Haren2017; Wallace et al. Reference Wallace, McMakin, Tan, Rosen, Forbes, Ladouceur and Silk2017; Wallen, Park, Krumlauf, & Brooks, Reference Wallen, Park, Krumlauf and Brooks2019; Wichniak et al. Reference Wichniak, Skowerska, Chojnacka-Wojtowicz, Taflinski, Wierzbicka, Jernajczyk and Jarema2011). Among the included studies, several measures intended to assess sleep disruption, including the fragmentation index, which quantifies sleep continuity (Benard et al., Reference Benard, Etain, Vaiva, Boudebesse, Yeim, Benizri and Geoffroy2019; Geoffroy et al., Reference Geoffroy, Micoulaud Franchi, Maruani, Philip, Boudebesse, Benizri and Etain2019). Further, two parameters were used to describe the transition period between sleep and wakefulness: sleep onset latency, which is the time required to fall asleep after going to bed (e.g. Bergwerff, Luman, & Oosterlaan, Reference Bergwerff, Luman and Oosterlaan2016; Blake et al. Reference Blake, Blake, Schwartz, Raniti, Waloszek, Murray and Allen2018), and sleep inertia, which is the time spent awake between sleep offset and getting out of bed [e.g. (Verkooijen et al., Reference Verkooijen, Stevelink, Abramovic, Vinkers, Ophoff, Kahn and van Haren2017)]. As demonstrated by the included studies, sensor data can also be informative to determine individuals' day-night rhythm or circadian patterns, when physical activity and sleep data are combined. The inter-daily variability (e.g. Benard et al. Reference Benard, Etain, Vaiva, Boudebesse, Yeim, Benizri and Geoffroy2019; Geoffroy et al. Reference Geoffroy, Micoulaud Franchi, Maruani, Philip, Boudebesse, Benizri and Etain2019; Shou et al. Reference Shou, Cui, Hickie, Lameira, Lamers, Zhang and Merikangas2017 is one measure to quantify consistency in sleep-wake pattern across days, whereas intra-daily variability (e.g. Geoffroy et al., Reference Geoffroy, Micoulaud Franchi, Maruani, Philip, Boudebesse, Benizri and Etain2019) represents an indicator for rhythm fragmentation, which relates to daytime napping or night time activity. Difrancesco et al. (Reference Difrancesco, Lamers, Riese, Merikangas, Beekman, van Hemert and Penninx2019) developed an index for circadian rhythm, also known as chronotype, or the proclivity to be asleep at a particular time of the day.

Active and passive data acquisition methods combined

The findings of our review further indicated that, to date, it is primarily sleep research that has pioneered the joint use of active and passive data acquisition methods in mental health research (i.e. 26 out of 30 included studies, see online Supplementary Table S5). The validation of measures can be accomplished by combining sensor data with self-report data e.g. on sleep (e.g. Lovato, Lack, Wright, & Kennaway, Reference Lovato, Lack, Wright and Kennaway2014; McMakin et al., Reference McMakin, Ricketts, Forbes, Silk, Ladouceur, Siegle and Dahl2019 or on other constructs such as pain (McCrae et al., Reference McCrae, Williams, Roditi, Anderson, Mundt, Miller and Robinson2019), affect (Merikangas et al., Reference Merikangas, Swendsen, Hickie, Cui, Shou, Merikangas and Zipunnikov2019; Wallace et al., Reference Wallace, McMakin, Tan, Rosen, Forbes, Ladouceur and Silk2017), or stress (Wallace et al., Reference Wallace, McMakin, Tan, Rosen, Forbes, Ladouceur and Silk2017).

Analysis

Intensive longitudinal data typically has a multilevel structure, with repeated measurements nested within individuals. Therefore, associations among the constructs of interest can be examined on at least two levels. Analyses at the cluster level (i.e. individuals or groups) reveal information on between-person differences in individuals' average responses (e.g. those who experience more stress in their daily life are, on average, more likely to report psychotic experiences (e.g. Glaser, Van Os, Thewissen, & Myin-Germeys, Reference Glaser, Van Os, Thewissen and Myin-Germeys2010; Reininghaus et al., Reference Reininghaus, Gayer-Anderson, Valmaggia, Kempton, Calem, Onyejiaka and Morgan2016b, Reference Reininghaus, Kempton, Valmaggia, Craig, Garety, Onyejiaka and Morgan2016c). Analyses at the within-person level account for potential variability in individuals' experience and behaviour over time (i.e. from one measurement occasion to another). These analyses therefore allow for investigating temporal trajectories and uncovering event-related or context-dependent relations among the constructs under scrutiny (e.g. whether an individual has a high risk for binge eating when experiencing high levels of negative affect (e.g. Berg et al., Reference Berg, Cao, Crosby, Engel, Peterson, Crow and Wonderlich2017; Crosby et al., Reference Crosby, Wonderlich, Engel, Simonich, Smyth and Mitchell2009; Selby et al., Reference Selby, Doyle, Crosby, Wonderlich, Engel, Mitchell and Le Grange2012).

To date, ESM in the field of mental health research has primarily reported findings based on the analyses of between-person differences i.e., aggregating ratings on target constructs across measurement occasions (e.g. Blum et al., Reference Blum, Vakhrusheva, Saperstein, Khan, Chang, Hansen and Kimhy2015; Engel et al., Reference Engel, Wonderlich, Crosby, Mitchell, Crow, Peterson and Gordon2013, Kimhy et al., Reference Kimhy, Vakhrusheva, Khan, Chang, Hansen, Ballon and Gross2014; Kuepper et al., Reference Kuepper, Oorschot, Myin-Germeys, Smits, van Os and Henquet2013; Muehlenkamp et al., Reference Muehlenkamp, Engel, Wadeson, Crosby, Wonderlich, Simonich and Mitchell2009; Pearson et al., Reference Pearson, Pisetsky, Goldschmidt, Lavender, Wonderlich, Crosby and Peterson2016; Pisetsky et al., Reference Pisetsky, Crosby, Cao, Fitzsimmons-Craft, Mitchell, Engel and Peterson2016). Most of the included studies conducted these types of analyses to examine the effectiveness of an intervention (e.g. comparing treatment v. control conditions (e.g. Chapman et al., Reference Chapman, Rosenthal, Dixon-Gordon, Turner and Kuppens2017; Kordy et al., Reference Kordy, Wolf, Aulich, Burgy, Hegerl, Husing and Backenstrass2016; Schlam, Baker, Smith, Cook, & Piper, Reference Schlam, Baker, Smith, Cook and Piper2020; Silk et al., Reference Silk, Tan, Ladouceur, Meller, Siegle, McMakin and Ryan2018; Simons et al., Reference Simons, Hartmann, Kramer, Menne-Lothmann, Hohn, van Bemmel and Wichers2015), or to examine differences in target constructs (e.g. the experience of stress, or negative affect) across service users and healthy controls (e.g. Blum et al., Reference Blum, Vakhrusheva, Saperstein, Khan, Chang, Hansen and Kimhy2015; Goldschmidt et al., Reference Goldschmidt, Peterson, Wonderlich, Crosby, Engel, Mitchell and Berg2013; Johns et al., Reference Johns, Di, Merikangas, Cui, Swendsen and Zipunnikov2019; Leraas et al., Reference Leraas, Smith, Utzinger, Cao, Engel, Crosby and Wonderlich2018; Morgan et al., Reference Morgan, Lee, Wright, Gilchrist, Forbes, McMakin and Silk2017; Oorschot et al., Reference Oorschot, Lataster, Thewissen, Lardinois, van Os, Delespaul and Myin-Germeys2012; Reininghaus et al., Reference Reininghaus, Gayer-Anderson, Valmaggia, Kempton, Calem, Onyejiaka and Morgan2016b; Tsanas et al., Reference Tsanas, Saunders, Bilderbeck, Palmius, Osipov, Clifford and De Vos2016). However, the full benefit of analysing intensive longitudinal data collected using ESM, arguably, comes into play when also considering temporal fluctuations in the relationship between an independent variable [e.g. affective experience (Anestis et al., Reference Anestis, Selby, Crosby, Wonderlich, Engel and Joiner2010; Berner et al., Reference Berner, Crosby, Cao, Engel, Lavender, Mitchell and Wonderlich2017; Karr et al., Reference Karr, Crosby, Cao, Engel, Mitchell, Simonich and Wonderlich2013)] and some outcome of interest [e.g. maladaptive behaviour (Anestis et al., Reference Anestis, Selby, Crosby, Wonderlich, Engel and Joiner2010; Berner et al., Reference Berner, Crosby, Cao, Engel, Lavender, Mitchell and Wonderlich2017; Karr et al., Reference Karr, Crosby, Cao, Engel, Mitchell, Simonich and Wonderlich2013; Ruscio et al., Reference Ruscio, Gentes, Jones, Hallion, Coleman and Swendsen2015)] that unfold at the within-person level. This approach also provides a means of identifying processes and situations that precede a critical event [e.g. incidents of self-injury (Muehlenkamp et al., Reference Muehlenkamp, Engel, Wadeson, Crosby, Wonderlich, Simonich and Mitchell2009), dietary restrictions (Engel et al., Reference Engel, Wonderlich, Crosby, Mitchell, Crow, Peterson and Gordon2013), aggressive urges or behaviour (Scott et al., Reference Scott, Wright, Beeney, Lazarus, Pilkonis and Stepp2017)]. Multi-level mixed-effect models further allow for the inclusion of random effects to account for person- and day-level differences, for example, in the association between negative affect and aggressive urges by modelling random intercepts and slopes. In this way, it can be shown that there are between-person differences in complex within-person associations. For example, it has been reported (Scott et al., Reference Scott, Wright, Beeney, Lazarus, Pilkonis and Stepp2017) that an increase in perceived rejection was associated with an increase in the experience of negative affect (i.e. within-person association). This association was stronger for individuals with more pronounced borderline personality symptoms (i.e. between-person difference). Finally, examining time-lagged associations between independent variables and outcomes provide insights into the development of these associations over time. Despite this advantage, only a minority of the included studies used time-lagged analyses (24 studies, e.g. Jahng et al., Reference Jahng, Solhan, Tomko, Wood, Piasecki and Trull2011; Klippel et al., Reference Klippel, Schick, Myin-Germeys, Rauschenberg, Vaessen and Reininghaus2021; Wigman et al., Reference Wigman, van Os, Borsboom, Wardenaar, Epskamp, Klippel and Wichers2015). Gerritsen et al. (Reference Gerritsen, Bagby, Sanches, Kiang, Maheandiran, Prce and Mizrahi2019), for instance, showed that high levels of activity-related stress experienced at time t _n−1 predicted increases in anhedonia at time t _n. Another study revealed that post-traumatic stress disorder symptom severity at time t _n was not predicted by the experience of negative affect at time t _n−1, but conversely, that symptom severity at time t _n−1 predicted the experience of negative affect at time t _n (Erwin et al., Reference Erwin, Dennis, Coughlin, Calhoun and Beckham2019). More recently, Klippel et al. (Reference Klippel, Schick, Myin-Germeys, Rauschenberg, Vaessen and Reininghaus2021) applied cross-lagged moderated multilevel mediation analyses in order to systematically test the temporal association between momentary stress, negative affect, and psychotic experiences.

Common approaches to analyse sensor data also include multi-level modelling. However, in most studies, parameters are aggregated prior to analysis by, for example, calculating the mean score for approximating individuals' physical activity from step counts collected on several consecutive days (Benard et al., Reference Benard, Etain, Vaiva, Boudebesse, Yeim, Benizri and Geoffroy2019). There is also a recent move towards utilising more complex methodological approaches, including supervised machine learning algorithms (e.g. Wallen et al., Reference Wallen, Park, Krumlauf and Brooks2019, Zebin, Peek, & Casson, Reference Zebin, Peek and Casson2019). In particular, in long time series derived from multiple sources (e.g. several sensors) machine learning approaches using prediction models including Bayesian networks and recurrent neural networks may be applied (Koppe, Guloksuz, Reininghaus, & Durstewitz, Reference Koppe, Guloksuz, Reininghaus and Durstewitz2019) These novel approaches are also increasingly being used for classifying individuals, e.g., into individuals with mental health problems and controls, based on mobile sensing data (Birnbaum et al., Reference Birnbaum, Ernala, Rizvi, De Choudhury and Kane2017). To this end, different algorithms have been applied including Support Vector Machines, Bayesian classifiers, random forest, and other decision trees.