This study is a secondary analysis of data from a pragmatic, 2-arm, single-blind cluster randomized controlled trial evaluating a universal digital-human parenting intervention delivered through the preschool system in China [28].

The trial was conducted in a large public preschool in a lower-middle-income city in Jiangxi Province, central China. All 21 classes within the preschool were included as clusters and randomly assigned in a 1:1 ratio to either the intervention group (10 classes) or a waitlist control group (11 classes) using a computer-generated sequence by an independent researcher.

Eligible participants were primary caregivers of children aged 3 to 6 years enrolled in the preschool. Inclusion criteria were minimal to reflect the universal nature of the intervention: caregivers had to be aged 18 years or older, have access to a smartphone, and provide informed consent. No additional exclusion criteria were applied. Caregivers were recruited through routine school communication channels, with one caregiver participating per family.

A total of 541 caregivers were enrolled, including 272 in the intervention group and 269 in the control group. Caregivers had a mean age of 36.6 (SD 5.4) years, and 74.8% (404/541) were female. Children had a mean age of 4.9 (SD 0.9) years, and households had an average of 1.9 (SD 0.6) children.

Data were collected at 4 time points: baseline (T0), immediate postintervention (T1), 6-month follow-up (T2), and 12-month follow-up (T3). Due to the waitlist design, the control group received the intervention after T1 and was not followed as a comparison group at later time points. Accordingly, between-group comparisons are limited to postintervention effects (T1), while longer-term changes (T2-T3) are examined within the intervention group. The participant flow is presented in Multimedia Appendix 1.

The intervention was a universal digital-human parenting program, named “Keyushiguang,” delivered through the preschool system. It was based on the Parenting for Lifelong Health (PLH) programs and adapted from the PLH ParentText intervention for delivery via WeChat (Tencent Holdings Limited).

The program comprised 37‐39 (depending on child age) brief interactive chatbot modules delivered over approximately 2.5 months, covering 8 core parenting topics (eg, parent-child relationships, child development, behavior management, and child safety). Content was delivered daily in a self-directed format and tailored to caregiver and child characteristics.

The digital content was complemented by human support, including weekly or twice-weekly message-based group discussions facilitated by preschool headteachers and social workers within class-based WeChat groups, to reinforce key messages, support reflection, and encourage peer interaction.

Additional strategies included facilitator reminders and a child-centered incentive mechanism linking caregiver participation to small rewards for the participating children. The incentive was implemented at the family or index-child level rather than separately for each child in the household. Together, these components were designed to promote sustained engagement and integration of parenting practices into daily routines. Further details about the intervention are provided in related publications [28,29].

The intervention group received the program following baseline assessment, while the waitlist control group received no formal support during this period and began the intervention after the postintervention assessment.

Baseline characteristics and outcomes were summarized by number of children using means (SD) for continuous variables and frequencies (%) for categorical variables. Group differences were assessed using 1-way ANOVA or chi-square tests. For the missing data, we assessed whether the missingness pattern was consistent with the assumption of missing completely at random (MCAR) using the Little MCAR test. We also summarized the extent and pattern of missing baseline data overall and by trial arm and number of children. Differences between families included in and excluded from complete-case baseline analyses were assessed using chi-square or Fisher exact tests for categorical variables and 2-tailed t tests or 1-way ANOVA for continuous variables, as appropriate. Multivariable regression models, adjusted for sociodemographic covariates, were used to examine baseline associations between the number of children and outcomes. To account for the potential influence of extended family support on parenting and resource allocation, we additionally adjusted for grandparent caregiving in baseline regression models.

Moderation of intervention effectiveness at postintervention was analyzed following the intention-to-treat principle. We used multivariable mixed-effects models (linear, Poisson, or negative binomial, depending on outcome distribution) to account for clustering at the class level and repeated assessments within families (random intercepts). Model selection for violence perpetration outcomes was informed by the distribution test (Table S1 in Multimedia Appendix 1). Moderation by family size was assessed using a group×time×family size interaction term. Subgroup analyses stratified by number of children were also conducted, and sensitivity analyses treated family size as a continuous variable.

The moderation analysis assessed whether family size modified the average short-term intervention effect, defined as the between-group difference in change from baseline to immediate postintervention. This differs from the trajectory analysis, which examined whether families in the intervention group followed different patterns of change over time after receiving the program. These analyses address related but distinct questions and may therefore yield different findings. Specifically, moderation analyses focus on average short-term treatment effects, whereas trajectory analyses examine patterns of change over time. Differences between the two may emerge when intervention responses are nonlinear, delayed, or differentially sustained.

For the intervention group, longitudinal trajectories from T0 to T3 were examined using mixed-effects models including time, family size, and their interaction, adjusting for baseline sociodemographic covariates. We additionally adjusted for grandparent caregiving in the longitudinal regression models to account for the potential influence of extended family support. Furthermore, 2 sensitivity analyses were conducted to assess the robustness of the findings: (1) estimating unadjusted longitudinal models including only time, number of children, and their interaction; and (2) applying multiple imputation by chained equations (MICE) with 30 imputed datasets to address missing data in the fully adjusted models. Estimated marginal means were used to visualize adjusted trajectories over time. These models above were interpreted as within-intervention-group postintervention trajectories.

To contextualize the limited precision of the moderation and trajectory analyses, particularly for families with 3 or more children, we calculated minimum detectable effect sizes (MDES) for key interaction terms. MDES were calculated from the model-based SEs of the interaction estimates, assuming a 2-sided α of .05 and 80% power. For continuous outcomes, MDES are presented as absolute differences in scale units. For count outcomes, MDES were reported as incidence rate ratios (IRRs) after backtransformation from the log scale. These estimates were used to quantify the smallest interaction effects the study was adequately powered to detect rather than to reinterpret statistical significance. All models used robust SEs. A 2-tailed α level of .05 was used to determine statistical significance. Analyses were conducted using Stata 17.0 (StataCorp, LLC) and R (R Foundation for Statistical Computing; refer to Multimedia Appendix 1 for detailed statistical methods).

Ethical approval for the original trial was obtained from Beijing Normal University (SSDPP-HSC-2024003) and the University of Oxford (SPI DREC 25 006). The trial was prospectively registered on the Chinese Clinical Trial Registry (ChiCTR2400081911). All participants provided informed consent prior to data collection. This study is a secondary analysis of deidentified trial data and was conducted in accordance with the original approvals; no additional ethical approval was required.

This study examined whether the number of children is associated with baseline differences and heterogeneous responses to a universal digital-human parenting intervention delivered in a preschool setting.

First, family size was associated with systematic baseline differences: families with more children reported lower levels of early learning and stimulation and positive parenting, alongside higher parenting stress and greater endorsement of corporal punishment, while child behavioral outcomes and caregiver-perpetrated violence were broadly comparable.

Second, family size did not significantly moderate immediate postintervention effects. The program showed similar short-term effectiveness across 1-child, 2-child, and 3-or-more-child families.

Third, differences emerged in postintervention trajectories. Although all groups improved, patterns of change varied over time: 2-child families showed the most consistent improvements, 3-or-more-child families showed larger but more variable gains in some domains, and 1-child families showed more limited changes.

The present findings can be interpreted by viewing the number of children as a multidimensional construct that simultaneously imposes constraints and creates opportunities for adaptation within family systems, rather than a purely demographic characteristic [30].

From a constraint perspective, the baseline differences observed are consistent with the resource dilution framework, which posits that parental resources are finite and must be distributed across children [15,16,18]. As family size increases, reduced per-child investment may contribute to lower levels of early learning and stimulation and positive parenting, as well as higher parenting stress. Early learning and stimulation may be particularly sensitive to family size because many of its core activities, including reading, playing, and guided interaction, rely on sustained one-to-one caregiver-child engagement. Such activities are likely to be more affected by competing demands on caregiver time and attention than broader parenting practices. This aligns with evidence that multichild families face greater role strain and competing demands [25,31] and that parenting stress may undermine parenting quality and parent-child interactions [21]. However, for caregiver-perpetrated violence, the interpretation may be more complex. Although caregiver age was adjusted for in the models, the older age profile of caregivers in larger families may be associated with generational differences in parenting norms, including attitudes toward physical discipline. Baseline differences in caregiver-perpetrated violence may therefore reflect not only resource-related pressures associated with raising multiple children, but also cohort-related variation in disciplinary beliefs and practices.

However, the trajectory findings suggest a more nuanced pattern than predicted by deficit-oriented accounts of family size. The stronger improvements observed in some outcomes among multichild families are consistent with an amplification pathway. Family systems theory emphasizes that changes in parenting behaviors may generalize across children within the household, with intervention-related skills extending beyond the initially targeted child [32].

Multichild families may provide more opportunities to apply and reinforce newly acquired skills across children and contexts. Sibling interactions may also create natural settings for social learning, potentially facilitating broader changes in family functioning [22]. These processes may help explain why families with more children, despite lower baseline levels, showed larger gains in some domains.

These dual pathways help reconcile the contrast between baseline disadvantage and postintervention gains. While greater caregiving demands may constrain initial conditions and implementation capacity, they may also support the diffusion of intervention effects within the family system. This is consistent with evidence suggesting that larger families may, under some conditions, function as “micro-systems” that facilitate relational adaptation and resilience [24].

Despite differences in baseline conditions and postintervention trajectories, this study did not detect statistically significant moderation by the number of children at the immediate postintervention. This pattern is consistent with prior research showing that well-designed parenting interventions can produce broadly similar short-term changes across diverse family contexts, including those facing structural disadvantages [4,6]. However, given the small number of families with 3 or more children and the wide CIs surrounding several interaction estimates, these findings should be interpreted as inconclusive rather than as evidence of equivalent short-term effects across groups defined by number of children.

One explanation is that differences related to caregiving demands may not manifest immediately but emerge over time. Parenting interventions often target proximal behaviors that can be modified relatively quickly, leading to similar short-term gains across families. However, differences in implementation burden and resource constraints may become more consequential for sustaining these gains, resulting in divergence in longer-term trajectories. Methodologically, these findings address a different question from the moderation analysis. Whereas the T0-T1 analyses estimated between-group differences in short-term intervention effects, the T2-T3 analyses examined patterns of change within the intervention group after the waitlist control group had crossed over. These trajectories may reflect delayed or nonlinear intervention-related change, but they may also be influenced by child maturation and broader environmental factors. This interpretation is consistent with evidence that intervention effects may follow nonlinear patterns over time, with initial improvements followed by divergence, and that trajectories may vary across family contexts [33], and it aligns with the heterogeneous trajectories observed at 6- and 12-month follow-ups.

A second explanation relates to statistical power. The relatively small number of families with 3 or more children may have limited the ability to detect moderation effects, particularly in interaction models. As a result, the absence of significant moderation should be interpreted with caution.

Finally, these findings highlight a distinction between average effects and patterns of change. While intervention effects may appear homogeneous at the immediate postintervention time point, heterogeneity may become evident when trajectories are examined over time, underscoring the importance of longitudinal and context-sensitive approaches to evaluation.

The role of family size observed in this study may be particularly salient in the context of digital-human parenting interventions. Compared to traditional face-to-face programs, these interventions are designed to be scalable and accessible, relying more heavily on self-directed engagement with structured content and less intensive human support [11].

This shift may increase the extent to which family context shapes intervention processes. Caregivers managing multiple children may face greater challenges in consistently engaging with materials and integrating recommended practices into daily routines, as competing demands and resource constraints influence the uptake and maintenance of new parenting strategies [18,21].

At the same time, the flexibility of digital-human interventions may enable adaptation to diverse family contexts. Self-paced delivery and repeated access to materials may facilitate practice across caregiving situations, while group-based components may support peer learning and engagement.

The findings should be interpreted within the broader demographic and sociocultural context of contemporary China, where rapid fertility policy changes have been accompanied by shifts in family structures and caregiving arrangements. Following decades of the 1-child policy, recent policy relaxations have led to the reemergence of multichild families within a relatively short period. However, this transition has occurred alongside a generational discontinuity in caregiving experience, as many current parents did not grow up in multichild households and therefore may have limited intergenerational guidance for managing sibling dynamics [25,26].

In this context, caregiving demands may be particularly salient. These demands are further shaped by broader social pressures, including high educational expectations and intensive parenting norms, which may increase the challenges of raising multiple children [34,35].

At the same time, the Chinese context helps illustrate how both constraints and adaptive processes may coexist. Limited experience with multichild parenting and uneven access to formal support may heighten resource constraints, particularly in the early stages of adjustment. Conversely, strong family interdependence and the involvement of extended kin, such as grandparents, may facilitate within-family learning and adaptation over time [36,37].

More broadly, these findings have implications beyond the Chinese context. As family structures evolve in many settings, including through declining fertility, delayed childbearing, and changing caregiving arrangements, variation in family size may interact with social and institutional contexts to shape how interventions are implemented and sustained. Understanding these contextual dynamics is important for designing parenting interventions that are both scalable and responsive to diverse family needs.

First, this study highlights the importance of moving beyond treating family structure as a control variable toward recognizing it as a meaningful dimension shaping the intervention process. The observed differences in baseline conditions and postintervention trajectories underscore the need for more context-sensitive research approaches that integrate heterogeneity with longitudinal dynamics.

Second, while parenting interventions may appear broadly effective in the short-term across family types, sustaining gains may require additional support, particularly for families facing higher caregiving demands. Incorporating strategies that reinforce long-term practice of skills—such as booster sessions, ongoing digital reminders, or structured follow-up support—may help maintain intervention effects over time.

Third, multichild families should not be viewed solely as resource-constrained. They may also provide opportunities for amplifying intervention impact, as parenting strategies can be applied across children. Designing interventions that leverage these within-family dynamics (such as promoting positive sibling interactions) may enhance overall impacts.

Fourth, the findings have particular implications for scalable digital-human interventions, where implementation relies heavily on caregivers. Differentiated pacing, targeted support for families with multiple children, or guidance on managing competing caregiving demands may help reduce implementation burden while preserving scalability.

Finally, in rapidly changing contexts such as China, policies and services need to better align with evolving family structures. Policies aimed at creating family-friendly environments may benefit from focusing not only on fertility levels but also on strengthening support systems that enable families to sustain positive parenting practices over time. Targeted support for multichild families—through parenting resources, community-based support, and integration with early childhood services—may help mitigate caregiving challenges and improve the sustainability of intervention benefits.

This study has several limitations. First, the relatively small number of families with 3 or more children, particularly within the intervention group, limited statistical power for the moderation and trajectory analyses. This increased the risk of type II error and reduced the precision of interaction estimates. Because only 10 families with 3 or more children contributed to the intervention-group trajectory analyses, the observed changes in this subgroup may be sensitive to individual cases and should not be interpreted as representative or confirmatory subgroup trends. Accordingly, these findings are best viewed as exploratory and hypothesis-generating. Second, all outcomes were based on caregiver self-report and may therefore be subject to reporting bias, particularly for sensitive behaviors such as violent parenting. For example, reductions in caregiver-perpetrated violence at T1 may partly reflect social desirability bias, as caregivers may have felt expected to report improvements in parenting behaviors after receiving the intervention. Third, due to the waitlist design, longer-term follow-up data were available only for the intervention group, limiting causal inference on sustained effects. Fourth, the number of children captures only one dimension of family size and caregiving complexity; other contextual factors, such as coparenting dynamics, may also shape intervention processes and should be examined in future research. This study also has several strengths. It integrates baseline differences, immediate postintervention effectiveness, and longitudinal trajectories within a single analytical framework, thereby providing a comprehensive assessment of intervention heterogeneity. The use of a cluster randomized controlled trial embedded in a real-world preschool system enhances both internal and external validity. In addition, the use of a standardized digital-human intervention supports consistency in delivery while maintaining contextual relevance.

This study provides evidence that the number of children is associated with variation in parenting intervention processes across baseline conditions, short-term effects, and longer-term trajectories. While intervention effectiveness was similar across family types in the short term, differences emerged in the sustainability of outcomes over time.

These findings suggest that caregiving demands may be less relevant for whether intervention effects occur than for how they are sustained and evolve over time. They highlight the importance of moving beyond average treatment effects to consider family context and longitudinal dynamics when understanding intervention responses.

By examining family size as a source of variation in caregiving contexts rather than solely as a control variable, this study contributes to a more nuanced understanding of heterogeneity in parenting intervention outcomes. It further integrates baseline differences, short-term effectiveness, and longer-term trajectories within a single analytical framework. Situated in contemporary China, where family structures and caregiving demands are rapidly changing, the findings also have practical implications for the design of parenting interventions, particularly the need for sustained and context-responsive support to promote long-term gains.