A sample of 34 children (17 ADHD + CD/ODD, 17 Controls) took part in the study. Healthy controls were recruited via advertisements in local journals. Patients were referred to our study from the Clinic of Child and Adolescent Psychiatry and Psychotherapy of the University of Kiel. All parents of the participants gave written informed consent. All children gave informed assent and they were reimbursed with a voucher for their participation. The study protocol was approved by the ethics committee of the medical faculty of the University of Kiel and followed the ethical standards of the Helsinki declaration.

According to DSM-IV-TR criteria (American Psychiatric Association [APA], 2013), all patients suffered from attention deficit hyperactivity disorder (15 patients with combined type, 314.01; two patients with predominantly inattentive type, 314.0). Furthermore, three patients suffered from comorbid CD (312.81) and 14 from oppositional defiant disorder (313.81). Two of the three patients suffering from CD also fulfilled the diagnostic criteria for an oppositional defiant disorder. One patient also suffered from enuresis (307.6). Six patients fulfilled the diagnostic criteria of a combined disorder of reading and written expression (315.0 and 315.2, ICD-10 F81.0) and five more patients had only subclinical symptoms. No further comorbidities were diagnosed. In all, 13 patients took MPH but discontinued medication 48 h (approximately 12 half-lives) prior to the experimental sessions. None of the control children suffered from any psychiatric disorder. To secure the diagnoses in the patients and to preclude any psychiatric disorders in the controls, all children and their parents were interviewed using a German translation of the Revised Schedule for Affective Disorders and Schizophrenia for School-Age Children: Present and Lifetime Version (K-SADS-PL;) (Kaufman et al., 1997; Delmo et al., 2000). Furthermore, the Child Behavior Checklist (CBCL) (Achenbach, 1991) was filled out by the parents to assess any psychiatric symptoms of their children. The patients received significantly higher ratings on the internalizing problems scale (t₃₂ = 3.05, p = 0.005, d = 1.05; descriptive statistics are reported in Table 1) and especially on the externalizing problems scale (t₃₂ = 9.02, p < 0.001, d = 3.09).

Since ADHD with comorbid disorders of social behavior is more often diagnosed in boys than in girls (Nussbaum, 2012), only boys were included in the sample. The age of patients and healthy controls ranged from 8 to 12 years and did not differ between the groups (t₃₂ = 0.49, p = 0.629, d = 0.17; for descriptive statistics see Table 1). As assessed using the Edinburgh Handedness Inventory (Oldfield, 1971), 15 of the patients and 16 of the healthy controls were right-handed. All participants had normal or corrected-to-normal vision. Candidates were excluded from the study if they had: (1) below average intelligence with an IQ < 85, as measured by the Culture Fair Intelligence Test Revised Version (CFT-20-R) (Weiß, 2006), (2) significant memory impairment as measured by the Diagnosticum für Cerebralschädigung (DCS) with scores below the 16th percentile (Lamberti and Weidlich, 1999), (3) advanced puberty as measured by Pubertal Development Scale (PDS; total score >7) (Watzlawik, 2009), (4) any medical condition or impairment that would interfere with the ability to participate in the study as assessed by interview, or (5) any sleep disorders. We screened for sleep disturbances using the parent-reported Children’s Sleep Habits Questionnaire (CSHQ) (Owens et al., 2000; Schlarb, 2011) and the children’s Sleep Self-Report questionnaire (SSR) (Schwerdtle et al., 2010). Moreover, the polysomnograms of the adaptation nights were examined for symptoms of sleep disorders by a trained sleep lab technician. No abnormal sleep patterns or sleep disorders were detected. The demographic and clinical characteristics of the remaining 34 participants are reported in Table 1. Patients and healthy controls were comparable with respect to IQ (t₃₂ = –1.54, p = 0.134, d = –0.53) and pubertal state (t₃₂ = 1.12, p = 0.270, d = 0.39), but the patients showed lower memory performance in the DCS (t₃₂ = –2.13, p = 0.041, d = –0.73) and worse sleep quality reported in the SSR (t₃₂ = 2.47, p = 0.019, d = 0.846) or the CSHQ (t₃₂ = 3.21, p = 0.003, d = 1.10).

In the so-called “pirate game” the children are asked to explore treasure islands (see Figure 1). Two equivalent versions of the pirate game with different stimulus sets were programmed in Presentation^® software (Version 14.9, Neurobehavioral Systems inc.) and the versions were approximately counterbalanced over conditions and order. In each trial, two pictures of islands are presented and the child has to decide which island to explore. Participants indicate their choice by pressing the corresponding mouse buttons. If the “correct” island is chosen, the picture of the island is replaced by a picture of a treasure, a sound of children cheering “yeah” is played, and the treasure counter is colored green and incremented by one (reward). If the “wrong” island is chosen, the island is replaced by a jolly roger, a disappointed voice uttering “ohhh” is played, and the treasure counter is colored red and decremented by one (punishment). During a block of 33 trials, the same islands are repeatedly shown on the left or right side of the monitor in pseudorandom, counterbalanced order. The participants were instructed to learn by trial and error to approach the island on which a treasure is hidden more often and to avoid the island which is inhabited by pirates more often. Probabilistic feedback was provided according to a reinforcement schedule with increasingly valid feedback: In the first third of the trials, the target island was correct with a frequency of 7/11 (≈63.6%) and wrong with a frequency of 4/11 (≈36.4%). In the second third the reward frequency was increased to 8/11 (≈72.7%) and the punishment frequency decreased to 3/11 (≈27.3%). Finally, in the last third, the reward frequency reached 9/11 (≈81.8%) and the punishment frequency 2/11 (≈18.2%). This schedule was chosen to allow the assessment of performance differences between the groups in the beginning of each block while ensuring that both groups encode the “correct” island equally well until the end of the block (compare results section).

During the encoding session before the retention interval, the children played three blocks of 33 trials (encoding). In each block, a unique set of island pictures was used. After the retention interval containing sleep or wake, the children played three blocks with entirely new picture sets (learning) to obtain a control measure for the influence of sleep vs. wake on learning performance. The crucial reversal learning block was identical to one of the encoding blocks using the same islands and the same reinforcement frequencies (reversal). However, the reinforcement schedule was reversed: now, formerly “correct” islands were “wrong” and vice versa. The rationale behind the reversal learning block was to test whether sleep helps to consolidate the stimulus-response mapping and therefore make it harder to reverse it during reversal learning. In other words, participants are expected to persist in preferring the formerly “correct” island as an indicator of the consolidation of rewarded behavior.

All participants spent an adaptation night and a test night in the sleep laboratory separated by at least one night for recovery from potential sleep loss during the first night. The adaptation night’s purpose was to exclude severe sleep disorders and to help participants adapt to the conditions in the sleep laboratory. During both nights sleep was recorded by standard procedures using a digital electroencephalogram (EEG), electromyogram (EMG) and electrooculogram (EOG). To amplify and record the data, a SOMNOscreen PSG plus (SOMNOmedics, Randersacker, Germany) was used. The EEG was recorded at a sampling rate of 128 Hz with a band-pass filter of 0.2–35 Hz using multi-use Ag/AgCl-electrodes attached to the positions C3 and C4 according to the 10–20 system referenced to an electrode on the bridge of the nose and with a ground placed at Fpz. A diagonal EOG was recorded at a sampling rate of 128 Hz with a band-pass filter of 0.2–75 Hz using single-use Ag/AgCl-electrodes attached to the lower right and upper left canthi. A bipolar EMG was recorded at a sampling rate of 256 Hz with a band-pass filter of 0.2–128 Hz using single-use Ag/AgCl-electrodes attached to the chin. Only during the adaptation night did we additionally record an EMG from the anterior tibial muscles, a bipolar electrocardiogram, nasal air flow using a thermistor, and respiratory thorax excursions using a belt sensor. All sleep data were visually scored according to the criteria by Rechtschaffen and Kales (1968) by a certified rater unaware of the hypotheses. The following macro-sleep parameters were obtained: sleep stages 1 to 4 and REM sleep (in min), time in bed (in min), lights off, lights on, sleep-onset latency (time in min from lights off to first epoch of sleep stage 2), total sleep time (in min), sleep efficiency (ratio of total sleep time to time in bed in percent), number of awakenings, and duration of wakefulness after sleep onset (in min).

To control for effects of sleep on wakefulness and mood, the participants kept sleep logs, rated their wakefulness on a visual analog scale (ranging from 0 to 100) and their mood on the valence and arousal scales of the self-assessment manikin (SAM) (Bradley and Lang, 1994) in the morning and in the evening in the sleep as well as the wake condition. Furthermore, before every encoding or reversal session, the alertness test from the test battery of attentional performance for children (KiTAP) (Zimmermann et al., 2002) was administered.

The participants took part in two diagnostic sessions and two experimental conditions, each comprised of two sessions with an interposed delay of 12 h. During the first diagnostic session, the children and their parents were interviewed independently by trained psychologists, and tests and questionnaires were administered. The second diagnostic session was the adaptation night and took place at least two days before the experimental sleep condition. In the experimental wake condition the children arrived at the laboratory at 8:00 a.m., filled out the sleep log and rating scales, completed the alertness test, and played three blocks (i.e., pairs of islands) of the pirate game (encoding). The participants were instructed not to sleep during the day and attend to their usual daily routines. Twelve hours later (8:00 p.m.) the children came back to the laboratory, filled out log and rating scales again, performed the alertness test, and played three new blocks of the pirate game (learning) as well as on one block with reversed contingencies (reversal).

In the sleep condition, the children arrived at the laboratory at 8 p.m. on the day of the experimental night and worked on the logs, scales, test and pirate game as described above. The electrodes for the PSG were attached around 9 p.m., lights off was at 9:30 p.m. and lights on at 7:00 a.m. Again, the second part of the testing took place 12 h after encoding. The sleep and the wake conditions were at least two weeks apart and the conditions and parallelized stimulus sets were approximately counterbalanced over groups.

First, we describe the preprocessing of the behavioral data from the probabilistic learning and reversal task. Second, we delineate the inference statistics and how we dealt with the control variables.

In the probabilistic learning paradigm, we counted the choices of the “correct” target stimulus which was followed by reward more often. In a few trials, some participants clicked the mouse multiple times so that responses were carried over to the next trial. These responses, indicated by reaction times shorter than 100 ms in the following trial, were deleted and interpolated with a random choice. The patients showed significantly more multiple button presses than healthy controls (t_16.8 = 2.19, p = 0.043, d = 0.75). However, multiple reactions were very scarce, amounting to only 0.65% ± 0.27% (MEAN ± SEM) in the patients and 0.06% ± 0.04% in the healthy controls. The resulting learning curves were then filtered with a two-way moving average with a window size of five to reduce random noise in the choice data (Smith, 1997). To measure encoding success, we calculated the relative frequency of correct choices during the last thirds of the three encoding blocks (see Figure 2). To obtain a measure for the consolidation of the stimulus-response mappings, we calculated the difference of the relative frequency of correct responses during the first five trials of the reversal learning block and the encoding success (reversal – encoding; see Figure 2). To control for influences of time of day on learning performance per se, we calculated the correct responses during the first five trials of the three new learning blocks after the retention interval.

To control whether the participants learned to prefer the island associated with reward more often during encoding, we computed t-tests comparing the performance during the last thirds of the three encoding blocks against guessing frequency (0.5). To evaluate whether sleep effected the consolidation of learned behavior differently in the ADHD + CD/ODD patients compared to healthy controls, we calculated an ANOVA using SLEEP (sleep vs. wake) and LEARNING (encoding vs. reversal) as within-subject factors and GROUP (ADHD vs. control) as a between-subject factor. Significant effects were resolved using Bonferroni-adjusted post hoc contrasts. Also, the difference between the performance during encoding and reversal in the sleep condition as a measure of sleep-dependent memory retention was correlated with the amounts of REM sleep and non-REM sleep. To control for differences in learning performance, we computed an ANOVA of the control measure learning using SLEEP as within and GROUP as between factors. To test whether the manipulation had any effect on subjective wakefulness, valence, arousal or objective alertness and learning performance, we computed ANOVAs with TIME (morning vs. evening) as a within-subject and GROUP (ADHD vs. control) as a between-subject factor with post hoc contrasts. In the case of a significant main effect of GROUP or a significant interaction of TIME and GROUP, the main analysis of the data from the probabilistic learning and reversal task was repeated using the respective control variables as covariates. To compare the groups regarding sleep parameters and questionnaire data, t-tests were used. Descriptives are reported as a mean ± standard error of the mean (SEM). For all t-tests, Cohen’s d was computed as a measure of effect size. Pearson’s correlation coefficients and partial η² from the analyses of variance were converted to d values according to Cohen (1988) and Rosenthal (1994).

Children in both groups slept roughly 8 h during the experimental night (ADHD: 475.6 ± 9.0, Control: 482.7 ± 8.9, t₃₂ = –0.56, p = 0.578, d = –0.19). The amount of sleep in the sleep stages 1 to 4 and REM sleep did not differ between groups (all p > 0.213, also see Table 2). By trend, the amount of slow-wave sleep (S3 + S4) was slightly higher in the control group (117.0 ± 5.5) than in the patients (105.3 ± 3.8, t₃₂ = –1.75, p = 0.091, d = –0.60). All sleep parameters were within normal range (Scholle et al., 2011).

All participants learned to prefer the island associated with reward more often during encoding. The relative frequencies of correct choices during the last thirds of the encoding blocks were significantly higher than the guessing frequency 0.5 in healthy controls (prior to wake: t₁₆ = 10.15, p < 0.001, d = 2.46 /prior to sleep: t₁₆ = 8.59, p < 0.001, d = 2.08) as well as in patients (prior to wake: t₁₆ = 10.54, p < 0.001, d = 2.56/prior to sleep: t₁₆ = 9.56, p < 0.001, d = 2.32, also see Table 3 and Figure 2). Please note that on a descriptive level all values of individual participants were greater than the guessing frequency 0.5. This confirms that all participants encoded the rewarded behavior, namely the preference for the correct island.

The performance was significantly higher at the end of the encoding blocks as compared to the beginning of the reversal block (main effect of LEARNING, F_1,32 = 76.90, p < 0.001, d = 3.10; also see Table 3 and Figure 2). This illustrates that the previously learned preferences were retained in memory and had to be relearned during reversal. There were no main effects of GROUP (F_1,32 = 0.68, p = 0.417, d = 0.29) or SLEEP (F_1,32 = 1.61, p = 0.214, d = 0.49) on the performance as well as no interaction effects of LEARNING and SLEEP (F_1,32 = 0.48, p = 0.495, d = 0.24) or LEARNING and GROUP (F_1,32 = 1.12, p = 0.298, d = 0.37). However, there was a significant interaction effect of SLEEP and GROUP (F_1,32 = 4.39, p = 0.044, d = 0.74) which was qualified by a significant three-way interaction of SLEEP, LEARNING and GROUP (F_1,32 = 5.28, p = 0.028, d = 0.81). Comparing the performance of encoding and reversal by post hoc contrasts, we found that performance of the control participants dropped under the wake (p = 0.027, d = 0.74) as well as under the sleep condition (p < 0.001, d = 1.42). In patients performance significantly dropped in the wake (p < 0.001, d = 1.12) and, by trend, in the sleep condition (p = 0.094, d = 0.62). Again, this confirms that in general contingencies were retained in memory in both groups under both conditions. However, the drop of performance did not differ between sleep and wake in the patients (p = 0.264, d = 0.28) but was stronger during sleep than during wake in the control participants (p = 0.042, d = –0.51), and this double difference was significant (p = 0.028, d = 0.79, see Figure 2). Hence, only the control participants showed sleep-dependent consolidation of rewarded behavior (hypothesis 1).

Note that learning performance per se did not differ between conditions or groups. The ANOVA of the additional control measure learning revealed no main effect of SLEEP (F_1;32 = 0.01, p = 0.930, d = 0.03), no main effect of GROUP (F_1,32 = 1.47, p = 0.234, d = 0.43), and no interaction of SLEEP and GROUP (F_1,32 = 1.41, p = 0.243, d = 0.42). This also illustrates that time of day did not influence learning performance as this would have produced a main effect of condition (SLEEP) in this control measure. Furthermore, our data preprocessing did not bias the results: The interaction effect reported above would remain significant (p = 0.026, d = 0.86) if the two outlier-patients with the highest rate of multiple reactions were excluded from the analysis. Filtering slightly improved the significance of the interaction effect (p = 0.044 to p = 0.028/d = 0.74 to d = 0.81) but did not change the results.

It was assumed that REM sleep fosters the consolidation of rewarded behavior in healthy controls (Perogamvros and Schwartz, 2012). Therefore, REM sleep was expected to correlate with the magnitude of the drop in performance during the night, whereas non-REM sleep should not (hypothesis 2). In ADHD there were only non-significant correlations of the performance drop with REM sleep (r = 0.322, n = 17, p = 0.207, d = 0.68) and non-REM sleep (r = 0.138, n = 17, p = 0.597, d = 0.28), and the difference between these correlations was not significant either (z = 0.640, p = 0.261). In contrast, the control participants showed a significantly higher (z = –1.819, p = 0.034) correlation of performance drop with non-REM sleep (r = 0.441, n = 17, p = 0.076, d = 0.98) than with REM sleep (r = –0.233, n = 17, p = 0.368, d = –0.48). Finally, we analyzed REM latency and REM density but did not find any significant correlations with the consolidation of rewarded behavior in healthy controls (all p > 0.276, all d < | 0.58|) or in patients (all p > 0.215, all d < | 0.61|).

To exclude the possible interpretation that the diminished sleep-dependent consolidation of rewarded behavior is caused by overall psychopathology we repeated the analysis reported above using LEARNING and SLEEP as within-subject factors, GROUP as a between-subject factor and the overall score of the CBCL a covariate. However, the three-way interaction reported remains significant (p = 0.015, d = 0.94). The same is true when we use the score of the nonverbal memory test DCS as a covariate (p = 0.038, d = 0.78).

Furthermore, we controlled for effects of TIME of day on alertness and subjective valence, arousal and wakefulness. The TIME of day (F_1,32 = 6.90, p = 0.013, d = 0.93) and GROUP (F_1,32 = 6.12, p = 0.019, d = 0.87) had an impact on the reaction times in the alertness test and there was a trend toward an interaction (F_1,32 = 3.39, p = 0.075, d = 0.65). Post hoc contrasts revealed that patients were slower than controls in the mornings (p = 0.046, d = –0.71) and in the evenings (p = 0.014, d = –0.90; descriptive data not shown). Furthermore, patients slowed down during the day (p = 0.003, d = –0.62) as opposed to the controls (p = 0.582, d = –0.20). Therefore, we used the performance in the alertness test in the morning and evening as covariates in an ANCOVA of the performance in the probabilistic learning and reversal task. However, the interaction effect reported above stays stable (F_1,30 = 5.87, p = 0.022, d = 0.89).

Neither TIME of day nor GROUP had any influence on the valence (all p > 0.159, all d < 0.51) or arousal ratings (all p > 0.139, all d < 0.54). However, the analysis of the subjective wakefulness revealed a significant interaction of TIME of day and GROUP (F_1,32 = 9.60, p = 0.004, d = 1.10) but no main effects (all p > 0.270, d < 0.40). Post hoc contrasts showed that the groups’ wakefulness did not differ in the mornings (p = 0.123, d = –0.54) but in the evenings (p = 0.047, d = 0.71) and that the patients’ wakefulness declined during the day (p = 0.005, d = –0.71) as opposed the controls’ wakefulness (p = 0.172, d = 0.35). Again, we used the subjective wakefulness in the mornings and evenings as covariates in an ANCOVA of the performance in the probabilistic learning and reversal task. The interaction effect reported above was not significant any longer (F_1,30 = 2.50, p = 0.125, d = 0.58) but on a descriptive level, the adjusted means pointed in the same direction as reported above.