Investigating the moderators and mediators of an effective sleep intervention in the Prevention of Overweight in Infancy (POI) randomized controlled trial: Exploratory analyses

Summary The Prevention of Overweight in Infancy (POI) sleep intervention halved obesity risk at 2 years of age. However, the intervention mechanisms are unclear. Consequently, the objective of the current work was to use exploratory analyses to investigate potential moderators and mediators of the sleep intervention on obesity outcomes at age 2 years. Data were collected between 2009 and 2012. The effect of demographic and study design variables on body mass index z‐score (BMI z‐score) and obesity was compared in moderator subgroups at 2 years of age (n = 683, 85%). Mediating effects of child and parent–household variables assessed whether the sleep intervention resulted in meaningful changes in the mediating variable (defined as changes which were statistically significant [p < .05] or where the effect size was ≥0.15 SD), followed by assessing relationships with obesity outcomes. The sleep intervention appeared most effective in children in higher deprivation areas (effect on BMI z‐score −0.25 [−0.53, 0.04], effect on obesity odds ratio [OR] 0.43 [0.16, 1.13]), and with mothers of non‐European, non‐Māori ethnicity (effect on BMI z‐score −0.27 [−0.73, 0.20], effect on obesity OR 0.13 [95% confidence interval 0.01, 1.11]). This suggested moderation by deprivation and ethnicity. Aspects of sleep improved meaningfully in children after intervention but did not significantly relate to obesity outcomes, and other outcomes were not meaningfully affected by the sleep intervention. Thus, mediation was not indicated. Overall, the POI sleep intervention improved obesity outcomes at 2 years, and the current work identified some potential moderators, but no mediators.

• The Prevention of Overweight in Infancy (POI) sleep intervention which was conducted antenatally and in early infancy, halved obesity risk at 2 years of age. However, the mechanisms behind this effect are unclear.
• Moderation and mediation analyses are useful for exploring potential intervention mechanisms.
What does this study adds?
• Exploratory analyses, investigating a range of potential moderators and mediators of the Prevention of Overweight in Infancy (POI) sleep intervention on body mass index z-score and obesity at 2 years of age, were undertaken.
• Some potential moderators were identified, including maternal ethnicity and area-level deprivation.
• Aspects of sleep improved in children following the POI sleep intervention. However, they did not significantly relate to obesity outcomes. Other outcomes were not meaningfully impacted by the sleep intervention. Thus, no mediators were identified.

| INTRODUCTION
Internationally, approximately 38 million children under 5 years of age are affected by overweight or obesity. 1 Excess weight in childhood adversely affects physical and psychological health, contributes to behavioural and emotional difficulties, reduces educational attainment, 2 and often persists into adulthood, 3 where it is associated with an increased risk of serious non-communicable diseases. 4 In recent years a number of interventions have targeted obesity prevention from the earliest days of life, 5 when metabolic and behavioural patterns are still developing. They have had only relatively modest effects, 5,6 which has led to calls for testing alternative approaches to the more commonly targeted lifestyle behaviours of diet, physical activity and media use. 6,7 Sleep is an alternative approach which is supported by consistent observational evidence indicating that short sleep duration 8 and poor sleep quality 9 are associated with an increased risk of childhood obesity, although interventions to date are limited. 10 Our Prevention of Overweight in Infancy (POI) trial 11 showed that a brief sleep intervention in infancy substantially reduced the risk of obesity at 2 12 years of age. However, earlier analyses did not specifically report differences in lifestyle behaviours of interest (such as diet, 13,14 sleep, 12,14,15 and activity 12,14,16 ) for the groups that received the sleep intervention compared to the groups that did not. Thus, there is a need for new analyses, which explore the potential mechanisms behind the observed benefits of the POI sleep intervention on obesity.
Moderation and mediation analyses (Figures S1 and S2) provide a method for identifying potential causal pathways with moderation describing when or for whom an independent variable most strongly causes a dependent variable, and mediation explaining how and why the independent variable causes the effect on the dependent variable. 17 In this context, there were several factors that we sought to explore through moderation and mediation.
With regard to moderation ( Figure S1), we were aware that mothers who participated in the POI study were predominantly welleducated and of New Zealand European ethnicity, and that fewer families lived in areas of high deprivation than is observed in the New Zealand population as a whole. 12,14 It is possible that the intervention was received differently by, or had different impacts on, participating families with other demographic characteristics. Furthermore, outcomes could be moderated by certain study design features. For example, different levels of support were available to families within the sleep intervention, and the POI study was designed as a 2 Â 2 factorial trial, meaning that some participants in the sleep group also received a Food, Activity and Breastfeeding (FAB) intervention.
With regard to mediation ( Figure S2), we hypothesized that there were a range of 'child' and 'parent-household' factors with potential to mediate the effects of the POI sleep intervention on obesity outcomes.
The sleep intervention educated parents about normal infant sleep development, and emphasized that parents should give infants opportunities to learn to settle to sleep unaided 11 in an effort to promote the ability for children to self-regulate their sleep behaviours. As such, we hypothesized that this focus may have influenced 'child factors' such as how they slept, and their general self-regulatory abilities, as well as a range of 'parent-household factors' potentially associated with childhood obesity, such as parenting style, 18 stress related to parenting, 19 parental feeding practices, 20 and parental mental health. 21 The objectives of the current work were to determine whether demographic and study design factors moderated, and child and parent-household factors mediated, the effect of the POI sleep intervention on obesity outcomes at 2 years of age. All analyses are of an exploratory nature, for the purpose of setting directions for future early childhood sleep interventions.

| POI trial
Information about the POI trial and the subsequent follow-up study are available elsewhere. 11 gestation, able to communicate in English or Te Reo M aori (indigenous language of New Zealand) and planned to live locally for 2 years. After assessment, 1458 women were eligible; of these, 611 (42%) declined to participate. Infants were excluded after birth if gestation was <36.5 weeks, or if they had a congenital abnormality or physicalintellectual disability likely to affect feeding, physical activity or growth.
The final sample size was 802. Retention was 86% at 2 years. Families were randomly allocated to 1 of the four study arms within 6 strata depending on area-level deprivation 23 (3 levels) and parity (2 levels) by using a block size of 12. Participants in all four arms received standard government-funded well child care. 24 Those in the Usual Care group did not receive any additional intervention. Families in the intervention groups received additional guidance and support related to Sleep, FAB or both. Those delivering and receiving interventions could not be blinded, but the main outcome measurements were performed by researchers blinded to group allocation. Ethical approval for the POI trial was obtained from the Lower South Regional Ethics Committee (LRS/08/12/063). Written informed consent was obtained from the parent-guardian of all child participants. The trial was registered at clinicaltrials.gov as NCT00892983. provided with a copy of the booklet and encouraged to refer back to it as often as necessary. At 6 months, and 1 and 1.5 years, parents who indicated that their child's sleep was problematic were offered additional assistance from a research nurse who worked alongside the family to implement one or more of four approaches, which differed in complexity. At the first level, families received simple advice (e.g., about changes to sleeping arrangements or feeding). The second level involved advice on settling techniques only. The third level involved a partial sleep intervention, and the fourth level a full sleep intervention. 11 Each family chose the level, or combination of levels, which they felt were most appropriate for their family, and used the relevant tools as often as they liked.

| Measurements
Comprehensive details of all measurement procedures are described in Table 1. To summarize briefly here, anthropometric measures were obtained by trained measurers blinded to intervention status at 2 years of age. Family demographic characteristics were collected by questionnaire at baseline (late pregnancy). Children's sleep and other behaviours were assessed by actigraphy, questionnaires or laboratory assessments at time points between the ages of 1 and 2 years. 11,22 Similarly, parent and household factors were assessed by questionnaire between the ages of 1 and 2 years, as described in Table 1. Moderation effects for demographic and study variables (sex, deprivation, parity, maternal education, maternal ethnicity, maternal obesity, FAB intervention and additional sleep support) were assessed by subgroup analyses. These were exploratory analyses (not pre-specified) and while tests of interaction were undertaken, the study is likely to be underpowered to detect statistically significant (p < .05) interactions. 25 Linear regression models were used to estimate the effect of the sleep intervention by subgroup, with body mass index (BMI) z-score as the outcome variable and intervention group as the predictor variable. An interaction term between intervention group and demographic subgroup was then included in the model to determine the p-value for interaction. For each subgroup, the mean difference in BMI z-score for those who received the sleep intervention compared to those who did not receive a sleep intervention was calculated along with the 95% confidence interval (CI). Logistic regression models were used in the same way for obesity outcomes with odds ratios (OR) reported.

| Statistical analysis
All potential mediating variables (except for sleep variables) were standardized to be in units of SD to allow identification of the variables that were most strongly affected by the sleep intervention.
Potential mediating variables were decided a priori to be: sleep variables (total sleep time, time awake after sleep onset, sleep efficiency, number of night wakings and sleep problem score), child temperament at 2 years of age, maternal depression, parental stress, parenting style, parental feeding practices, discipline strategies and family quality of life. Mediation was explored using a three-step process 26 : (1) First, we determined whether the potential mediating variable differed by sleep intervention group. Linear regression models were run with the T A B L E 1 Methods of measurement for potential moderating and mediating factors in the Prevention of Overweight in Infancy (POI) study  27 Linear regression models were used with BMI z-score as the outcome variable, the potential mediating variable as the predictor variable and adjustment for sleep intervention, FAB intervention, parity and deprivation; and if the sleep intervention had a meaningful effect on a potential mediating variable, and was related to BMI z-score in a meaningful way (decided a priori to be statistically significant or a mean difference of 0.05 or greater), then this variable was considered a mediator. The extent of mediation was assessed by the percent of the effect size that was reduced after adjustment for the mediator. The same method was used for obesity outcomes but using logistic regression and reporting odds ratios.

| RESULTS
Overall, the prevalence of overweight and obesity among children in the POI study at 2 years of age was 40.0% (273/683) while the prevalence of obesity alone was 16.5% (114/683). 12

| Moderation analyses
The results of the moderation analysis for BMI z-score and obesity are displayed in Table 2 Method of data collection and timing Measurement procedures and use of data encourage nutrient-dense foods, discourage energy-dense foods, mealtime behaviour). Each factor consisted of several items, and each item was scored using a five-point scale from 1 (never) to 5 (always). Higher scores indicated greater use of the relevant parental feeding practice.
Discipline practices Questionnaire, 1.5 years Positive discipline practices were assessed using an age-appropriate list of behaviours developed by the POI research team. 49 Parents were asked to indicate how often over the past 7 days they had employed each behaviour to get their child to do, or stop doing, something. Higher scores indicated greater use of positive discipline practices The baseline questionnaire was completed by the mother in late pregnancy.
T A B L E 2 Demographic and study design factors and assessment of effect moderation on obesity outcomes at 2 years of age   Adjusted for parity, deprivation and whether they received the FAB intervention. Standardized mean differences are presented for all questionnaire measures but not sleep outcomes. b p-values are not adjusted for multiple tests. We recommend using the standardized mean differences and precision of these estimates (95% CI) to judge whether an association is meaningful. c From actigraphy. Time spent awake after sleep onset presented as median (25th, 75th percentile), with differences in the median estimated using quantile regression. d From questionnaire (eight-point scale from 0 = no problem to 7 = large problem).

| Mediation analysis: Parent and household factors
compared to those who did not. The sleep intervention did not appear to have any impact on parental feeding practices (measured at 1.  Table 5 displays the relationships between those factors indicated as potentially important mediators in Tables 3 and 4 and the mean difference in BMI z-score and odds of obesity for every 1 SD increase in T A B L E 4 Differences in parent and household factors for those who received the sleep intervention compared to those who did not Mean (SD) a n p-values are not adjusted for multiple tests. We recommend using the standardized mean differences and precision of these estimates (95% CI) to judge whether an association is meaningful.

| Potential mediators and their relationships with obesity outcomes
the relevant factor. The effect sizes were uniformly small, ranging from À0.05 to 0.04 for BMI z-score and 0.77 to 1.12 for obesity.  Before considering the findings of the moderation and mediation analyses more closely, it is relevant to note that the overall outcomes reported in Table 2 (lower mean difference BMI z-score and reduced risk of obesity at age 2 years among those who received the sleep intervention compared to those who did not) are important for several reasons. First, the POI sleep intervention was brief, consisting in most cases of only two perinatal parent contacts (group education session in late pregnancy and home visit at 3 weeks post-partum).
While it has been suggested that interventions of long duration may be necessary to create sustained change, 28 these findings indicate that there is potential for sleep interventions to be brief, T A B L E 5 Relationships between potential mediators and body mass index (BMI) z-scores at 2 years of age Very little work appears to have investigated the moderators and mediators of obesity outcomes in early childhood obesity interventions.
One Australian exploration of potential moderators and mediators of an online healthy lifestyle programme (incorporating nutrition, physical activity, screen time and sleep for 2-5 year old children) on BMI change did not find any significant moderating or mediating effects. 32 However, in that study there were also no significant between-group differences in child BMI at trial completion. 33 The authors recommended that other childhood obesity interventions also investigate a wide range of factors to allow comparisons between studies and develop a better understanding of the factors contributing to successful interventions. 33 With regard to potential moderators, several of those identified in the current work, particularly deprivation level and maternal ethnicity, warrant further investigation. In New Zealand, children of M aori ethnicity, and children from more disadvantaged backgrounds, are over-represented in obesity statistics. 34 That the POI sleep intervention appeared to be more effective among those living in more deprived areas at 2 years, is potentially promising. However, conversely, while we are reluctant to place emphasis on the ethnicity findings due to the comparatively low numbers in the M aori and 'Other' groups, it is possible that the POI sleep intervention was less effective in children of M aori mothers, than in children of mothers with other ethnicities. Notably, the messages in the POI sleep intervention were consistent with existing sleep health recommendations for New Zealand children. In the time period since the POI sleep intervention was completed (8-9 years ago), recognition that some of these recommendations are inconsistent with the worldviews, sociocultural contexts and realities of some groups, including M aori wh anau (families), has received more attention. 35,36 We suggest that future sleep interventions aim to recruit a greater range of ethnically and socioeconomically diverse participants, and carefully consider how intervention messages are framed, to ensure that they do not increase inequities in vulnerable populations. 37 A key strength of this work is that it contributes to a very small, but important and growing area of research. Our findings are intriguing because while some differences in children's sleep behaviours were observed at 1 year of age between children who received the sleep intervention compared to children who did not, these differences were small (overnight sleep efficiency improved by 1.2%, time awake after sleep onset decreased by 8 min and number of night waking decreased by 0.14) and were not maintained at 2 years of age; yet, the sleep intervention clearly improved obesity outcomes at 2 years. Thus, there is a clear need to identify the mechanisms responsible for the effect of the sleep intervention on obesity outcomes to ensure that future research can be targeted to act on them. Sleep interventions also have potential as obesity prevention interventions because sleep may be potentially a less stigmatizing behaviour to focus on than diet and activity. 35 A key limitation of this work is that all presented analyses are of an exploratory nature, and thus the findings should be interpreted with caution. The lack of identified mediators indicates a need to think more broadly to identify other potential mediators. Weight-related behaviours such as dietary intake and physical activity did not differ following the intervention 13,16 . However, the POI study did not measure some parental feeding behaviours that may have contributed to the reductions in obesity observed. For example, the presence or absence of night-time milk feedings after 6 months of age was not measured. It is possible that the sleep intervention's focus on improving children's selfregulation of sleep behaviours resulted in less night-time milk feeding and lower overall energy intake among children who received the sleep intervention compared to children who did not. Thus, future work could focus more on specific parental feeding practices and eating behaviours, which might be influenced by a sleep intervention. Alternatively, it may be that the tools used to measure potential mediating effects in the POI study were not sensitive enough to detect the relatively small changes in energy-related behaviours that might be required to explain the differences in BMI z-score observed over the course of a 2 year intervention. 38 Other, more precise measurement techniques may be required. Furthermore, there was considerable homogeneity within the POI study population (mostly New Zealand European families living in areas of low or moderate deprivation), which was particularly evident in the moderation analyses, where numbers for some demographic characteristics were very low.
In conclusion, the exploratory analyses in this paper identified few potential moderators, and no mediators, of the effect of the POI sleep intervention on BMI z-score and obesity at 2 years of age.
Deprivation and maternal ethnicity were identified as potential moderators, and should be further considered alongside other potential moderators and mediators in future sleep intervention research in more ethnically and socioeconomically diverse populations.