The study took place online in November and December 2020. Participants were recruited via classes for physical education and psychology students who could collect experimental credits. Four groups were assigned in a randomised controlled way based on the e‑mail addresses that were provided in a random order separately from the survey data. Control group 1 (CG1) answered only the pre and post questionnaires 2 weeks apart, while control group 2 (CG2) and the intervention groups (IG) filled in online sleep diaries (OSD) every morning and evening for 2 weeks. CG2 did not receive further instructions. IG1 was invited to install the app Sleep Score (SleepScore Labs™, Carlsbad, CA, USA) and IG2 Sleep Cycle (Sleep Cycle AB, Gothenburg, Sweden) for free.

In week 1 (W1), intervention group participants used the sleep tracking feature and observed the apps’ reports, while in W2, participants additionally used the ‘smart alarm’. Sleep Score assesses body movements and respiration rates via the smartphone’s acceleration sensors and microphone. The sleep tracking feature calculates a score from 0 to 100, with 100 indicating optimal sleep quantity and quality. Continuity and duration of the sleep stages are also reported. With regular application, a graphical representation of the data can be obtained. The ‘smart alarm’ is activated by indicating the final wake time and a timeframe in which to be woken up. Using the sleep tracking feature, the app detects light sleep within this timespan and initiates the alarm. Sleep Cycle works in a similar way, yet without generating an overall sleep score. Instead, it provides more graphical analyses of sleep parameters. Sleep Cycle was shown to perform poorly in comparison to PSG; therefore, the app data were not used in the statistical analyses [7, 11].

Initially, 141 participants were recruited. Only those who answered the pre–post questionnaires and those in IG1 and IG2 who completed the OSD on at least 5 days during W2 were included in the analysis. Thus, the final sample consisted of 98 participants (n = 61 female; mean 21 ± 1.7 years, range 18–28 years). Regular physical activity was reported on 5.8 ± 3.6 d/week, with a mean duration of 6.9 ± 5.3 h/week. The majority participated in individual sports (n = 71) and 50% competed regularly in competitions (n = 49). The final groups consisted of n = 33 (CG1), n = 21 (CG2), n = 20 (IG1), and n = 24 (IG2).

Informed consent was obtained in the pre-study questionnaire and only those who consented to participate voluntarily could continue the survey. Individual feedback was provided to interested participants. Ethical clearance was obtained from the faculty’s local ethics committee prior to the start of the study.

IG1 and IG2 rated their expectations and experiences with the respective app. Specifically, they rated (a) whether the tracking feature resembled their own sleep perception, (b) whether they considered the app’s results as reliable, (c) whether the ‘smart alarm’ indeed chose the optimal timing, (d) whether they felt more refreshed upon waking compared to an ordinary alarm and whether they were willing to continue using (e) sleep tracking as well as (f) ‘smart alarm’. Furthermore, they judged (g) whether they consider sleep self-tracking via app as useful, and (h) whether they used the app’s data to optimise their sleep behaviour. These items were assessed on a scale ranging from 0 (not at all) to 4 (totally). Finally, participants gave (i) an overall evaluation of their app experience from 0 (very negative) to 4 (very positive).

The pre–post questionnaires contained the Pittsburgh Sleep Quality Index (PSQI) to assess sleep quality [6]. Hereby, 19 items are summarised to seven components with scores ranging from 0 to 3 (i.e. subjective sleep quality, sleep latency, sleep duration, sleep efficiency, sleep disorders, sleep medication and daytime sleepiness). The sum of these component scores generates the total score (0–21). Scores ≤ 5 indicate good sleep quality [6, 15]. While the original version covers the previous 4 weeks, the current study used a 2-week timeframe. The applicability of this modification has been shown previously [1]. Internal consistency in this sample was Cronbach’s α = 0.60 at pre. In addition, information on bedtime, get-up time and sleep onset latency was used to calculate the sleep parameters “total sleep time” (TST) and “sleep efficiency” (SE; relation of time spent in bed to TST), which were integrated in the pre–post analyses.

At the pre and post timepoints, the nine-item Bedtime Procrastination Scale (BPS) was used to assess bedtime procrastination [22]. The frequency of postponing bedtime is rated on a scale ranging from 1 ([almost] never) to 5 ([almost] always). The mean value represents the BPS score, with higher values indicating a higher prevalence of BP. Internal consistency was Cronbach’s α = 0.88 at pre.

The OSD consisted of an adapted evening and morning protocol according to Hoffmann et al. [18]. Before going to bed, daytime activities and mood states were documented, as was the planned bedtime. Within 30 min of getting up, participants rated restfulness and the events of last night’s sleep. Specifically, they documented the actual time of lights out, when they awoke, and when they finally got up in the morning. They estimated sleep onset latency and the frequency and duration of awakenings. Using these variables, the sleep parameters “time in bed” (TIB), TST and SE were calculated. The difference (in minutes) between planned bedtime and actual time of lights out was used as an indicator of daily bedtime procrastination (DBP).

Data preparation and descriptive and statistical analyses were performed with Microsoft Excel 2019 (Microsoft Corporation, Redmond, WA, USA), the statistical software SPSS (version 26, IBM Corp., Armonk, NY, USA) and RStudio (version 4.1.2, PBC, Boston, MA, USA).

It must be mentioned that different features and functions of Sleep Score (IG1) between Android and iOS smartphone systems became apparent only during the data collection. Android users received more limited feedback compared to iOS users. In IG2, n = 16 were iOS and n = 5 were Android users. Thus, in a preliminary step, the effect of smartphone system (SYS) on the subjective evaluation of the app was analysed. Within IG2 (Sleep Cycle), n = 11 were iOS and n = 12 were Android users. To examine the first study aim, moderated regression analyses were conducted with the items of the app evaluation as dependent variables. Dummy-coded variables app (0 = Sleep Score, 1 = Sleep Cycle) and SYS (0 = iOS, 1 = Android) served as predictors, and expectations at pre (item g and h) as covariates.

Since some effects of SYS were identified, this variable was included in subsequent analyses. The second study aim was investigated via linear mixed models with random intercepts using the maximum likelihood method [10]. Timepoints served as level 1 (pre = 0 vs. post = 1, W1 = 0 vs. W2 = 1) and individuals as level 2 predictors. Dependent variables were PSQI score, pre–post sleep parameters (TST, SE) and OSD parameters (TIB, TST, SE, DBP). Intercept-only models were calculated first (M0). Subsequent models included time (M1), group*time interaction with CG1 as the reference group (M2) and SYS (M3) as predictors. As the information about the system was obtained only for IG1 and IG2, this variable was coded as ‘unknown’ in CG1 and CG2, which also served as the reference group. M4 included the grand mean of the centred BPS pre score as a covariate. As M3 did not show significant improvements to the model, the more parsimonious M2 was chosen for building M4. For each model, intraclass correlation coefficients (ICC) were determined to show the relation between inter- and intraindividual variance. Descriptive model parameters, i.e. the Akaike information criterion (AIC) and Bayesian information criterion (BIC) as well as chi2 likelihood ratio tests, were used to compare nested models. However, in case of missing data for single variables, direct model comparisons were not possible, as those models were not considered as nested. The significance of single fixed effects coefficients was tested via Satterthwaite-approximated t-tests. The analyses were conducted with the lme4 package [3]. Level of significance was set to p < 0.05.