Associations between paternal engagement and child development outcomes in Lebanon: a multilevel cross-sectional analysis of national data

Abstract

Background

Fathers play meaningful and often under-recognised roles in supporting early child development (ECD) but evidence from crisis-affected settings is limited. In Lebanon, ongoing economic and social instability places pressure on the conditions that support children’s development, emphasising the value of examining caregiving practices that promote ECD. We examined whether paternal engagement (PE) in caregiving relates to children’s developmental outcomes and whether these relationships vary by socioeconomic context.

Methods

We used cross-sectional data from the 2023 Lebanon Micronutrient, Anthropometric and Child Development Survey (N=1181). Children, caregivers and households were nested within sampling units. The outcome was a pooled, age-standardised ECD z-score using the Global Scales of Early Development-Household Form (0–23 months) and the Early Childhood Development Index-2030 (24–59 months), constructed by standardising scores by age in months within each instrument and then pooling across instruments as a relative index of developmental status. PE was a composite of five stimulation and three caregiving activities (α=0.73). Hierarchical linear models estimated variance, individual associations, cross-level interactions with cluster-level household assets and PE decomposition within and between clusters. Analyses used weights and multiple imputation.

Results

PE was associated with higher ECD z-scores. Each one-unit increase in engagement was associated with a 0.22 SD rise in child ECD (SE 0.11; p<0.05), with sensitivity estimates from 0.13 to 0.32 SD. We found no evidence that the association differed by cluster-level household assets or by household nationality. Engagement components within and between clusters were positive but not significant.

Conclusions

In Lebanon, PE modestly correlates with higher ECD, with little variation across communities or levels of household assets. These findings emphasise fathers’ contributions to early development and suggest that supporting PE, alongside broader social protections, may help strengthen child development.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Paternal engagement in caregiving has been linked to small-to-moderate gains in early child development, yet most evidence comes from stable, higher-income settings rather than from countries facing concurrent economic and social crises.

WHAT THIS STUDY ADDS

• Using nationally representative data from Lebanon, a country in protracted crisis, we show that higher paternal engagement is associated with modestly higher developmental scores in children aged 0–59 months, independent of maternal caregiving and household assets, with little variation across communities or socioeconomic groups.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Programmes and policies seeking to strengthen early child development in crisis-affected settings could deliberately support fathers as everyday caregivers, alongside mothers and other kin and broader social protection.

Introduction

Early child development (ECD) is a foundational period for cognitive, emotional and social growth, and the conditions of early life have lasting consequences across the life course. In 2017, The Lancet estimated that approximately 250 million children in low- and middle-income countries (LMICs) were at risk of failing to reach their developmental potential by age 5.¹ Extensive evidence shows that optimal child development is shaped by nurturing care practices, such as responsive care, stimulation and early learning opportunities.² While much of this evidence has focused on maternal caregiving, there is increasing recognition of the importance of paternal caregiving practices in promoting child development outcomes.³

Despite this emerging focus, empirical evidence remains limited on whether and how these paternal engagement (PE) practices interact with socioeconomic circumstances to shape ECD outcomes.^{4 5} The socioeconomic context is known to influence parenting behaviours and child development. Low socioeconomic status (SES), a marker of inequitable access to resources and opportunities, is associated with reduced stimulation and responsive caregiving.^{6 7} At the same time, higher SES can buffer the influence of household stressors on caregiving.^{4 6 7} However, few studies have explored whether PE interacts with socioeconomic conditions to influence developmental outcomes.

This evidence gap is especially pressing in Lebanon, a country facing a polycrisis of economic collapse, political instability, conflict and a protracted refugee crisis, all of which have the potential to jeopardise young children’s developmental potential. Lebanon is a West Asian country on the eastern Mediterranean coast, home to a population of around 6 million people within an area of 10 452 km².^{8 9} Since 2019, Lebanon has experienced one of the most severe peacetime economic contractions recorded globally, with real gross domestic product falling by approximately one-third between 2019 and 2021 and the World Bank reclassifying the country from upper-middle-income to lower-middle-income status in 2022. Most of the population now lives below the national poverty line. This collapse has coincided with the 2020 Beirut Port explosion, the COVID-19 pandemic, recurrent political instability and successive periods of conflict and insecurity affecting multiple parts of the country.^{8 9} Although the intensity of hostilities has varied geographically, their consequences have been felt nationwide through population displacement, disruption of education and health services, lost livelihoods, economic uncertainty and widespread psychosocial stress. Lebanon also hosts one of the highest numbers of refugees per capita worldwide, including an estimated 1.5 million Syrian refugees alongside long-standing Palestinian communities.¹⁰

Together, these stressors have affected not only systems and services but also caregivers themselves, through increased stress, uncertainty and economic hardship, all of which erode the conditions that enable nurturing care (including good health, adequate nutrition, responsive caregiving, opportunities for early learning and security and safety), placing young children at heightened risk for not meeting their developmental potential.²

Lebanon is also highly heterogeneous, with its geographic regions differing markedly in socioeconomic composition, political governance, infrastructure, service access and cultural norms, including gender norms related to care across communities experiencing displacement.¹¹ Many children in this study have spent most, if not all, of their lives amid overlapping economic, social, political and security shocks. This makes the study particularly relevant, as it helps identify caregiving correlates that may support child development during protracted crises.

This study investigates these questions by analysing nationally representative data from Lebanon to assess the relationship between PE and ECD outcomes. We employ a multilevel modelling approach to account for the nested structure of the data, enabling us to separate influences at different levels of the social ecology. In a diverse country like Lebanon, ignoring this hierarchy could conflate individual-level associations with broader contextual influences, potentially obscuring important dynamics.

Our overarching research question is: How are PE practices associated with ECD outcomes in Lebanon, and does this association vary across different communities and socioeconomic contexts? To answer this, we pursue four specific aims: (1) assess the extent to which ECD outcomes vary across geographic clusters in Lebanon; (2) examine the individual-level association between PE practices and standardised ECD outcomes; (3) investigate whether cluster level household assets moderate this association and, as an exploratory analysis, we (4) distinguish between community-level and individual contributions of PE to ECD outcomes, including whether the strength of the individual-level association varies by cluster-level household assets.

Methods

Data

Data for this project come from the Lebanon Integrated Micronutrient, Anthropometric and Child Development Survey (LIMA 2023), a cross-sectional, nationally representative dataset. The study employed a two-stage stratified sampling design. In the first stage, primary sampling units (PSUs) were selected using probability proportional to size sampling, with hexagonal grids used to define the sampling area. In the second stage, households were randomly selected within each selected PSU. Children (aged 0–59 months) were recruited from the first eight households visited in the second stage. Further details on sampling are available in the report.¹² For this analysis, children, their caregivers and their households are considered level 1 units, while the PSUs (household clusters) are treated as level 2 units. Participants were not involved in the design, conduct, reporting or dissemination plans of this analysis. Data were collected between July and October 2023 by multidisciplinary field data-collection teams comprising interviewers, an anthropometrist, a phlebotomist and a team leader. Teams visited the selected households and administered the household, caregiver and child questionnaires in the home, interviewing the head of household or the child’s primary caregiver. The caregiver-reported measures analysed here were drawn from an hour-long nurturing-care questionnaire administered for one eligible child per selected household. Where more than one child aged 0 to 59 months was eligible, one was selected at random. For the module containing the PE items, the child’s mother or stepmother responded for 1127 children (95.43%), the father or stepfather for 27. Respondents were therefore overwhelmingly women, and we treat the exposure as predominantly maternal report of paternal behaviour. Interviewers initially requested that the nurturing-care questionnaire be administered in a private setting. When complete privacy could not be achieved, questions concerning attitudes towards domestic violence and the self-reporting Questionnaire (SRQ-20) were not read aloud. Instead, the interviewer turned the tablet towards the mother and asked her to enter her responses directly. No other member of the data-collection team was present during the administration of the nurturing-care questionnaire, although a field supervisor responsible for this component was occasionally present for supervisory observation. Other household members, including family members, were sometimes present. Their presence was occasionally documented in field notes, particularly when the interviewer considered the circumstances potentially relevant, but this information was not collected systematically for all interviews.

Key measures

The outcome (level 1) is pooled age-standardised ECD z-scores for children aged 0–59 months, calculated using the Global Scales of Early Development-Household Form (GSED-HF, 0–23 months) and the Early Childhood Development Index-2030 (ECDI2030, 24–59 months). Scores are standardised by age in months within each instrument. These instrument-specific z-scores were pooled to form a single continuous index of relative developmental status across the 0–59 month age range. The distributions were inspected for approximate normality and comparable variance. We excluded outlier observations with ECD z-scores >|3| SDs. The GSED Short, Long and Psychosocial Forms have been validated in seven countries, and the WHO officially launched the suite in February 2023.^{13 14} We used the HF (GSED-HF), whose items are drawn from the GSED Short Form and which is designed for caregiver administration in household surveys. The ECDI2030 has been field-tested in household surveys in two countries and will become the recommended tool for tracking ECD in the Multiple Indicator Cluster Surveys, Version 7 and future versions.¹⁵ The GSED-HF and ECDI2030 differ in content and developmental emphasis. Therefore, we interpret the pooled z-score as a standardised indicator of overall developmental standing rather than as evidence of strict metric equivalence between the instruments. PE (level 1), the main covariate, was calculated as an average composite score combining five stimulation activities (reading or looking at pictures in a book, storytelling, taking the child out, singing songs or lullabies and playing structured games) and three caregiving activities (daily care such as bathing, disciplining when the child misbehaves and healthcare visits). Stimulation items were drawn from the caregiver activities subscale of the Early Learning Indicators,¹⁶ while the LIMA 2023 research team developed caregiving items to capture culturally relevant paternal roles. Each stimulation activity was asked twice, once with a 3-day reference period and once with a 24-hour reference period, and the 2 responses were combined additively on a 0 to 2 scale: 2 if the father performed the activity in both reference periods, 1 if in 1, and 0 if in neither. Neither reference period was used alone or discarded. The five stimulation scores were then averaged to create a single stimulation score. Each of the three caregiving activities was rated on the same 0 to 2 scale (0=never/rarely, 1=1–2 times per week, 2=3–7 times per week). The stimulation composite score was combined with three individual caregiving activity scores, resulting in a total of four equally weighted components. The Cronbach’s alpha was 0.73. All item-test correlations exceeded 0.4 (range: 0.43–0.72).

Covariates included child-level (level 1) characteristics such as sex (male/female), age in months, a stunting indicator (more than 2.0 SDs below the median according to the WHO Growth Standards,¹⁷ low birth weight status (<2500 g at birth), and above average exposure to stress (above the stratum mean); caregiver-level (level 1) characteristics including maternal age (in years) and education (grades completed), maternal psychological distress (measured by the SRQ-20),¹⁸ maternal engagement in stimulation, paternal education (grades completed) and paternal age (in years); and household-level (level 1) characteristics such as the number of children under 5 years of age in the household, whether the household is food secure (using the Household Food Insecurity Access Scale),¹⁹ a standardised household asset index, access to stimulation and learning materials in the home (ranging from 0 to 7) and household size. Mean cluster-level household assets (level 2) is a continuous variable calculated by averaging household asset index scores within the PSU. A value of 0 can be interpreted as the national average. Nationality of the household head and co-residence of the child’s biological father are used in the sensitivity analyses described below, and the former should not be interpreted as measures of individual refugee or displacement status. We did not include a rural-urban classification because PSUs were defined using hexagonal grids rather than administrative locality type. Geographic context is instead captured by stratum fixed effects, which include dedicated strata for Syrian informal tented settlements and Palestinian camps and by mean cluster-level household assets. Exposure to stress was derived from an adapted 13-item Child Trauma Questionnaire capturing family traumatic events and stressful life events experienced by the child, and was dichotomised at above the stratum mean.²⁰ Maternal psychological distress was assessed with the WHO SRQ-20 and dichotomised at a cut-off of 8 or more out of 20 to indicate elevated depressive and anxiety symptoms.^{20 21} Full descriptions of all measures, including their items, scoring, translation and cultural adaptation for the Lebanese context, as well as available psychometric properties, are provided in online supplemental table S1.

Hierarchical modelling

We estimated hierarchical linear models to account for the nested structure of the data, where children (level 1) were clustered within PSUs (level 2). All models included a random intercept for PSU and fixed effects for survey strata to adjust for regional heterogeneity. We followed a sequential model-building approach aligned with our aims.

Model 0 corresponds to aim 1 and establishes a baseline by partitioning the variance in children’s ECD z-scores into between-cluster and within-cluster (PSU) components, while adjusting for fixed effects of survey strata. Model 1 corresponds to aim 2 and estimates the individual-level association between PE and child ECD outcomes, adjusting for child, caregiver and household (level 1) covariates, cluster-level household assets (level 2) and including random intercepts for PSUs. Model 2 addresses aim 3 by incorporating a cross-level interaction between PE and cluster-level household assets to test whether the association between PE and ECD outcomes differs across levels of local household assets. Model 3 addresses our exploratory aim 4 by decomposing PE into a between-cluster component (average level of PE in each PSU) and a within-cluster component (deviation of each child’s PE from their cluster mean) and includes a cross-level interaction between the within-cluster engagement score and PSU-level household assets.

Our primary models used the weighted, partially imputed (PMI-W) specifications, in which PE was observed (non-missing) and other incomplete covariates were imputed. We did not impute engagement because missingness in this variable is plausibly related to engagement itself (eg, non-co-residence), making missing-at-random assumptions less credible. To respect the survey’s two-stage design, we applied decomposed weights at their respective levels—a household design weight and a conditional child weight (child interview weight divided by household interview weight) using the svy2lme package in R.^{22 23}

We designated the PMI-W specification as our primary model because it (1) respects the complex two-stage sampling design of the LIMA survey and (2) avoids unverifiable assumptions about missingness in PE, which minimises bias from non-response. In contrast, complete case (CC) analyses restrict inference to a selective subset of children with no missing data. Given evidence that missingness in this context is plausibly related to socioeconomic and caregiver characteristics, CC models risk biased estimates and reduced generalisability. Nonetheless, each multilevel model was estimated across six specifications to assess robustness to analytic decisions: (1) CC unweighted, (2) CC weighted, (3) imputed unweighted and (4) imputed weighted. In addition, we ran two auxiliary versions of the imputed models, restricted to children with observed PE data (both unweighted and weighted), to verify that the exposure imputation did not drive the findings. We also examined variability in the random intercepts across strata and across clusters of different sizes to assess potential heterogeneity in baseline ECD outcomes that may reflect underlying geographic or demographic patterns, ensuring that observed variance patterns were not artefacts of model misfit.

Models

Model 0 establishes a baseline by estimating the proportion of total variance in children’s ECD scores attributable to between-cluster and within-cluster (PSU) components. Here, we are interested in the intraclass correlation coefficient (ICC), the percentage of variance in ECD z-scores explained by PSUs. For child/caregiver/household i in cluster (PSU) j and stratum k:

$${\displaystyle {\displaystyle {\displaystyle y_{ij}=\underset{\text{intercept}}{\underbrace{{\gamma }_{00}}}+\underset{\text{strata fixed effects}}{\underbrace{\sum _{k=1}^{K-1}{\gamma }_{0k}{s}_{kj}}}+\underset{\text{cluster random intercept}}{\underbrace{u_{0j}}}+\underset{\text{residual error}}{\underbrace{{\epsilon }_{ij}}}}}}$$

$${\displaystyle {\displaystyle u_{0j}\sim \mathcal{N}(0,{\tau }_{00}),{\epsilon }_{ij}\sim \mathcal{N}(0,{\sigma }^2)}}$$

Where:

$y_{ij}$: Age-standardised ECD z-score for child i in cluster j.

${\gamma }_{00}$: Intercept for the reference stratum (Akkar).

${\delta }_k$: Fixed effect of stratum k. Captures how much the mean ECD z-score in stratum k differs from the reference stratum (Akkar).

$s_{kj}$: Stratum-k dummy indicator for cluster j (K strata total, where K ranges from 0 to 9),

$u_{0j}$ ∼ ${\rm N}(0,{\tau }_{00})$: Random intercept for cluster j.

${\epsilon }_{ij}$ ∼ ${\rm N}(0,{{\sigma }_{\epsilon }}^2)$: Individual (within-cluster) error.

${\tau }_{00}$: Random intercept variance.

${{\sigma }_{\epsilon }}^2$: Residual variance.

Given the model, the ICC is then:

$${\displaystyle {\displaystyle \mathrm{I}\mathrm{C}\mathrm{C}=\frac{{\tau }_{00}}{{\tau }_{00}+{\sigma }^2}}}$$

The same error structure is assumed in models 1–3.

We next extend model 0 to model 1, which explores the individual-level association between PE practices and ECD outcomes, controlling for other level-1 covariates and including random intercepts at the cluster (PSU) level.

$${\displaystyle {\displaystyle {\displaystyle y_{ij}=\begin{array}{l}\underset{\text{intercept}}{\underbrace{{\gamma }_{00}}}+\underset{\text{paternal engagement}}{\underbrace{{\gamma }_{10}{\text{PE}}_{ij}}}+\underset{\text{other level-1 covariates}}{\underbrace{{\text{X}}_{ij}^{\mathrm{\prime }}{\gamma }_n}}+\underset{\text{cluster-level SES }}{\underbrace{{\gamma }_{01}{\text{SES}}_j}}\\ +\underset{\text{strata fixed effects}}{\underbrace{\sum _{k=1}^{K-1}{\gamma }_{0k}{s}_{kj}}}+\underset{\text{cluster random intercept}}{\underbrace{u_{0j}}}+\underset{\text{residual error}}{\underbrace{{\epsilon }_{ij}}}\end{array}}}}$$

Where our additional parameters are:

${PE}_{ij}$: Paternal engagement composite score for child i in cluster j.

${\gamma }_{10}$: Association between paternal engagement and ECD z-scores.

${X`}_{ij}$: Vector of other Level 1 covariates.

${\gamma }_n$: Coefficient vector for those covariates.

${\gamma }_{01}:\mathrm{}$Cluster-level household assets (continuous, level 2) and its respective coefficient.

Model 2 investigated whether cluster-level household assets moderate the association by incorporating a cross-level interaction.

$${\displaystyle {\displaystyle {\displaystyle y_{ij}=\begin{array}{l}\underset{\text{intercept}}{\underbrace{{\gamma }_{00}}}+\underset{\text{cluster-level SES }}{\underbrace{{\gamma }_{01}{\text{SES}}_j}}+\underset{\text{paternal engagement}}{\underbrace{{\gamma }_{10}{\text{PE}}_{ij}}}\\ +\underset{\text{cross-level interaction}}{\underbrace{{\gamma }_{11}\text{(}{\text{PE}}_{ij}\times {\text{SES}}_j)}}+\underset{\text{other level-1 covariates}}{\underbrace{{\text{X}}_{ij}^{\mathrm{\prime }}{\gamma }_n}}\\ +\underset{\text{strata fixed effects}}{\underbrace{\sum _{k=1}^{K-1}{\gamma }_{0k}{s}_{kj}}}+\underset{\text{cluster random intercept}}{\underbrace{u_{0j}}}+\underset{\text{residual error}}{\underbrace{{\epsilon }_{ij}}}\end{array}}}}$$

Where our additional parameters are:

$Y_{11}$: Cross-level interaction (how the PE association varies by cluster-level household assets).

Model 3 explored the contextual and cross-level effects of PE using group mean centring of paternal engagement, which explicitly decomposes paternal engagement into ${PE}_{ij}^W,$ its within-cluster (child-level paternal engagement relative to the PSU) and ${\stackrel{-}{PE}}_j$, the between-cluster (PSU average engagement) components.

$${\displaystyle {\displaystyle {\displaystyle {\overline{\mathrm{P}\mathrm{E}}}_j=\frac{1}{n_j}\sum _{i=1}^{n_j}{\mathrm{P}\mathrm{E}}_{ij},{\mathrm{P}\mathrm{E}}_{ij}^W={\mathrm{P}\mathrm{E}}_{ij}-{\overline{\mathrm{P}\mathrm{E}}}_j.}}}$$

Our Model 3 is then

$${\displaystyle {\displaystyle {\displaystyle y_{ij}=\begin{array}{l}\underset{\text{intercept}}{\underbrace{{\gamma }_{00}}}+\underset{\text{between-cluster (contextual)}}{\underbrace{{\gamma }_B\overline{{\text{ PE}}_j}}}+\underset{\text{within-cluster}}{\underbrace{{\gamma }_W{\text{PE}}_{ij}^W}}\\ +\underset{\text{cross-level interaction}}{\underbrace{{\gamma }_{11}\text{(}{\text{PE}}_{ij}^W\times {\text{SES}}_j)}}+\underset{\text{other level-1 covariates}}{\underbrace{{\text{X}}_{ij}^{\mathrm{\prime }}{\gamma }_n}}\\ +\underset{\text{strata fixed effects}}{\underbrace{\sum _{k=1}^{K-1}{\gamma }_{0k}{s}_{kj}}}+\underset{\text{cluster random intercept}}{\underbrace{u_{0j}}}+\underset{\text{residual error}}{\underbrace{{\epsilon }_{ij}}}\end{array}}}}$$

Where (additional):

${\gamma }_b$: Between-cluster association of average paternal engagement.

${\gamma }_w$: Within-cluster association of average paternal engagement.

$Y_{11}$: Cross-level interaction (how the within-cluster association varies by cluster-level household assets).

Across all models, fixed effects for strata are included, and child-level weights are scaled and applied appropriately. The coefficients of primary interest are the following: ${\gamma }_{10}$, representing the individual-level association between PE and child ECD outcomes; and ${\gamma }_{11},$ capturing the cross-level interaction between the strength of the child-level PE outcome and community-level household assets. Variance components of interest include ${\tau }_{00}$ (the variance of cluster-level random intercepts, representing variation in outcome across clusters beyond what is explained by our included covariates).

Sensitivity analyses

We conducted a series of sensitivity analyses to assess the robustness of our findings and evaluate the potential for selection bias due to missing data. First, we examined whether children with missing PE data systematically differed from those with complete data. We compared these groups across key child, caregiver and household characteristics using independent-samples t-tests for continuous variables and chi-squared tests for categorical variables. These comparisons revealed meaningful differences, suggesting that the potential for missing data to be missing completely at random may not hold, which motivated our use of multiple imputation. Missing covariate data (~0%–12% across covariates) were handled via multiple imputation using chained equations with the mice package,²⁴ incorporating all analysis variables (outcome, exposure, covariates and cluster identifiers) in the imputation model, together with the level-1 (child) weight. We generated eight imputed datasets with five iterations each and combined results using Rubin’s rules. One limitation of this package is that we cannot incorporate the two-stage design (ie, household weights) into the model’s imputation; therefore, coefficients should be interpreted conservatively. We ran four further sensitivity analyses on the primary specification. The first added nationality of the household head as a covariate, and the second added an interaction between PE and Syrian nationality. The third restricted the sample to the children whose module was answered by the mother, and the fourth to the children whose biological father lived in the household. Results are reported in online supplemental table S5.

Results

The final analytic sample included 1181 children (table 1) across 360 PSUs, with an average of 5 children per cluster (ranging from 1 to 12) and 10 regional strata. Approximately half were male (50.6%), with a mean age of 32.7 months (about 3 years old). Nearly one-quarter of children were born with low birth weight (24%), and 12.5% were stunted. Above-average stress exposure was reported among 28% of children. More than half of mothers experienced mental distress (60%). Parental education levels were generally low, with nearly half of mothers and over half of fathers having no formal schooling or having completed only primary school. Maternal stimulation scores averaged 2.86 (SD=1.58) out of a maximum possible score of 6, while PE was lower, averaging 0.85 out of a maximum possible score of 2 (SD=0.44). At the household level, families reported an average of 3.23 (SD=2.32) stimulation and learning materials, 20.73% of households were food secure and the mean number of children under five per household was 1.20 (SD=0.86). Just over half of households were headed by a Lebanese national and just under half by a Syrian national, with a small number of Palestinian and other households (50.1%, 46.8% and 3.1% weighted). The biological father lived with the child in 95.5% of cases. The child’s mother answered the PE module in 95.47% of cases and the father in 2%. CC data (no missing covariates) were available for 767 children, and non-missing PE data were available for 1058 children.

Table 1. Survey weighted descriptive statistics.

	Overall mean (SD); N (%)*	Missing N (%)
Child variables
Male	597 (50.55)	0 (0.00)
Age (in months)	32.68 (16.57)	0 (0.00)
Low birth weight	247 (23.91)	131 (11.09)
Stunted	137 (12.45)	78 (6.60)
Above average exposure to stress	471 (28.33)	13 (1.10)
Caregiver variables
Maternal age (in years)	31.3 (7.01)	7 (0.59)
Maternal psychological distress	522 (60.29)	137 (11.60)
Maternal education		2 (0.17)
No formal schooling	151 (12.79)	–
Primary school completed	406 (34.5)	–
Middle/high school completed	352 (29.89)	–
University or higher completed	269 (22.82)	–
Paternal education		57 (4.83)
No formal schooling	176 (15.66)	–
Primary school completed	430 (38.33)	–
Middle/high school completed	345 (30.78)	–
University or higher completed	171 (15.23)	–
Maternal stimulation score (range 0–6)	2.86 (1.58)	4 (0.34)
Paternal engagement score (range 0–2)	0.85 (0.44)	123 (10.4)
Paternal age (in years)	36.9 (7.96)	57 (4.8)
Paternal engagement module answered by the child’s mother	1127 (95.47)	10 (0.85)
Biological father resident in household (HH)	1136 (95.49)	1 (0.08)
Nationality of HH head		0 (0.00)
Lebanese	636 (50.13)	–
Syrian	499 (46.80)	–
Palestinian or other	46 (3.08)	–
HH variables
Access to stimulation and learning materials (mean number of item types, 0–7)	3.23 (2.32)	3 (0.25)
Food secure HHs	275 (20.73)	0 (0.00)
No. of children <5 in HH	1.20 (0.86)	6 (0.51)
HH size	5.36 (2.01)	0 (0.00)
HH asset index	0.30 (2.65)	0 (0.00)

Survey design weighted distribution of each covariate. N (%) for binary. N=1181.

^*Mean (SD) for continuous variables.

Figure 1 shows that children in clusters with above average household assets exhibited higher average developmental z-scores. Within each group, those with above-average PE had higher z-scores than those with below-average PE.

Figure 1. Mean early childhood development z-scores by paternal engagement and cluster-level household assets. SES, socioeconomic status.

Model results

Model 0 (no covariates) established the baseline partitioning of the variability in ECD z-scores across clusters. The results showed a residual SD ($\sigma$) of 0.88 and a cluster-level intercept SD ($\sqrt{{\tau }_{00}}$) of 0.13, with roughly 2.2% of the variability in ECD scores attributable to differences between clusters (ICC; table 2). Despite this low ICC, we retained the multilevel specification because our research questions explicitly concern level-2 covariates (cluster-level household assets and contextual PE), and failure to model clustering can bias level-2 effects and understate uncertainty. When including covariates (model 1), variance components decreased slightly ($\sigma$=0.79; $\sqrt{{\tau }_{00}}$=0.11; ICC=0.02). Model 1 shows evidence of a positive relationship between PE(${\gamma }_{10}$) and ECD outcomes (0.22 SDs, SE=0.11, p<0.05).

Table 2. Multilevel model estimates across models.

Parameter	Model 0 null model (baseline)			Model 1 PE (individual-level) association			Model 2 PE×cluster household assets interaction			Model 3 Decomposed PE+household assets SES moderation
	Est.	SE	P value	Est.	SE	P value	Est	SE	P value	Est.	SE	P value
${\gamma }_{10}$	–	–	–	0.22	0.11	<0.05	0.22	0.11	<0.05
${\gamma }_{11}$	–	–	–	–	–	–	−0.07	0.04	>0.05	0.00	0.04	>0.05
${\gamma }_W$										0.23	0.12	>0.05
${\gamma }_B$										0.22	0.24	>0.05
Random effects
σ	0.88			0.79			0.79			0.79
$\sqrt{{\tau }_{00}}$	0.13			0.11			0.07			0.11
ICC	0.02			0.02			0.01			0.02

These estimates are drawn from the primary weighted, partially imputed specification, restricted to children with paternal engagement data. γ10 = PE fixed effect; γ11 = cluster household assets*PE interaction coefficient; γW = PE GMC coefficient; γB = average cluster PE coefficient.

GMC, group mean center; ICC, intraclass correlation coefficient; PE, paternal engagement; SES, socioeconomic status.

Model 2 introduced a cross-level interaction between individual-level PE and cluster-level household assets. The main association of PE remained robust (${\gamma }_{10}$= 0.22 SDs, SE=0.11, p<0.05), indicating that, at the national average level of SES, higher PE was associated with higher ECD z-scores. The interaction term itself was statistically insignificant (${\gamma }_{11}$ =−0.072 SDs, SE=0.04, p>0.05), providing no evidence that the strength of the PE association differs between clusters with higher and lower household assets. Random effects remained stable.

For model 3, decomposition showed that the within-cluster association (${\gamma }_w$ = 0.23, SE=0.12, p=0.13) was not statistically significant, suggesting that children whose fathers were more engaged than the local average did not necessarily have higher ECD scores. The cross-level interaction between the individual deviation and cluster-level household assets was close to zero in magnitude and did not reach statistical significance: ${\gamma }_{11}$ = 0.003, SE=0.04, p>0.05. Taken together, these models indicate that PE is positively associated with development at the individual level, but we lack sufficient evidence to conclude that this association is meaningfully stronger or weaker in different community contexts or at different levels of household assets. Similarly, we did not find statistical evidence for a relationship between-cluster PE and child z-scores (${\gamma }_B$= 0.22 SDs, SE=0.24, p>0.05). Random-effects magnitudes remained essentially unchanged.

Sensitivity analyses

Results for all sensitivity specifications are presented in online supplemental tables S2–S4. Across complete-case, fully imputed and partially imputed models, with both unweighted and weighted estimation, the direction and magnitude of the PE coefficients are somewhat stable. Child development estimates range from 0.13 to 0.32 SD per one-unit increase in PE. In model 1, we observe no statistical evidence in the unweighted models, which may indicate that population reweighting is towards under-represented strata. In model 3, both within-cluster and between-cluster components remain directionally positive across specifications, reaching significance only in complete-case weighted analyses, but not in the others. Across all specifications, the variance in random intercepts is modest, with ICC values consistently ranging from 0.02 to 0.03, indicating that only a small proportion of variability in child development occurs between clusters. Four further analyses addressed reviewer questions about country of origin and sample composition (online supplemental table S5). Adding household-head nationality left the association essentially unchanged (0.22 SD, SE 0.11, p=0.052), and neither nationality contrast was distinguishable from zero. An interaction between engagement and Syrian nationality gave no evidence of moderation (0.10 SD, SE 0.22). Restricting the sample to the children whose module was answered by the mother gave a larger and more precisely estimated association (0.29 SD, SE 0.12, p<0.05), and restricting it to the children whose biological father lived in the household gave a similar magnitude with wider uncertainty (0.21 SD, SE 0.12).

Discussion

In a national sample of children aged 0–59 months in Lebanon, PE in caregiving and stimulation activities was consistently positively associated with ECD scores across most analytical approaches. Our primary weighted, partially imputed multilevel model (PMI-W) estimated a ~0.22 SD increase in the child development z-score per unit increase in the PE score. The consistency of evidence across models strengthens confidence that this relationship is not an artefact of any one analytic decision. In practical terms, the magnitude of the PE association was modest but meaningful and comparable to or slightly larger than the gains from other paternal inputs (eg, stimulation) in LMIC settings.³ Notably, the association persisted even after adjusting for mothers’ stimulation activities, supporting the view that PE provides unique, additive benefits beyond maternal engagement in caregiving.^{3 5} A multicountry analysis of 38 LMICs found that children with highly engaged fathers scored ~0.14 SD higher on early development indices than those with minimally engaged fathers.³ Systematic reviews and meta-analyses have similarly concluded that PE is positively associated with children’s cognitive and socioemotional outcomes.^{5 25} Together, these findings suggest that PE’s association with child outcomes in Lebanon is comparable to international estimates and that fathers’ contributions are complementary rather than redundant.

Contrary to our hypothesis, the strength of the association between PE and development did not differ systematically across levels of household assets. Cross-level interactions were small and statistically null, indicating that the developmental advantage of PE is not confined to clusters with higher household assets. This null finding contrasts with evidence from other regions. For example, Sun et al²⁶ reported that parental engagement buffered SES-related developmental gaps in East Asia. Several explanations are plausible. First, our level-2 measure of household assets is a proxy, capturing material assets but not safety, service quality, displacement or local norms—measurement error at level-2 tends to attenuate cross-level interactions.²⁷ The absence of a rural-urban classification compounds this, since settlement type shapes service access and childcare arrangements in Lebanon and could not be separated from the governorate-level variation absorbed by our stratum fixed effects. Second, contextual differences may be less readily detectable amid Lebanon’s ongoing polycrisis. In such contexts, widespread crises may compress differences between clusters, leaving little between-cluster signal and making traditional SES gradients less distinct than in more stable settings. Finally, both the modest between-cluster variance (ICC ≈0.02–0.03) and small average cluster size constrained statistical power to detect moderation, even if true moderation existed. However, the consistency suggests that PE was associated with higher child development in both disadvantaged and more affluent communities. From a substantive perspective, this lack of moderation by household assets may indicate that PE is a universally valued resource for children, not contingent on external community factors. Nonetheless, the absence of detectable moderation should be interpreted as limited evidence for heterogeneity by household assets under current data constraints rather than proof of equivalence across levels of household assets.

The association was also similar across nationality groups. Neither nationality contrast was distinguishable from zero, and an interaction with Syrian nationality was small and imprecisely estimated (online supplemental table S5). Syrian-headed households were 499 of the 1181 in the sample, so the association is not confined to Lebanese-headed households. This is not, however, evidence about displaced or refugee families. Displacement status was not recorded, and in Lebanon nationality is a poor substitute, because Syrian nationals have lived and worked in the country in large numbers since well before 2011. The Syrian category therefore mixes long-settled residents, labour migrants and families displaced by the war. Whether PE operates differently for displaced families remains an important question that these data cannot answer.

We observed minimal clustering of developmental outcomes by PSU, with ICCs around 0.02 across models, suggesting that most variation in early development is within-cluster rather than between-cluster. One interpretation is that outcomes are primarily driven by family-level caregiving rather than place-based context. Methodologically, the very low ICC highlights the difficulty of detecting cross-level interactions with modest sample sizes. Simulation studies show that slope-by-context interactions require large level-2 samples and appreciable between-cluster variance to achieve power.²⁷

We chose the PMI-W model as our primary analysis to strike a balance between methodological rigour and representativeness. This approach applied sample weights to account for the complex survey design and used multiple imputation to handle missing covariates, thereby maximising the use of available data. Notably, complete-case analyses yielded somewhat larger and statistically significant father-child development associations in our study. However, complete-case results should be interpreted cautiously, as children with complete data tend to come from slightly more advantaged households. Therefore, those models may overestimate associations by excluding vulnerable groups. In contrast, our PMI-W estimate, while more conservative, is likely less biased and more generalisable to the national population. We acknowledge that imputing within a multilevel, weighted framework carries complexities. Our imputation procedure was unable to account for the survey’s clustering and stratification fully. Despite these challenges, the robustness checks (including fully imputed and non-imputed models, with and without weights) all pointed in the same direction, suggesting that the positive association of PE is a stable finding.

It is important to note the limitations inherent in our cross-sectional observational design. We cannot establish causality. It is possible that fathers engage more with children who exhibit advanced development or easier temperaments, or that unmeasured confounders (eg, parental personality, overall quality of the home environment) contribute to both higher father involvement and better child outcomes. We attempted to control for many such factors (maternal stimulation, household characteristics, etc), but residual confounding may remain. In particular, our models did not capture support from grandparents, extended kin or paid childcare support, even though caregiving in Lebanon and other Arab contexts often occurs within multigenerational, collectivist family systems in which these actors play central roles.²⁸ Evidence from Palestinian refugees in Lebanon and Bahraini families suggests that live-in grandparents and especially maternal grandmothers play a significant role in children’s health, schooling and day-to-day care.^{29 30} Unmeasured variation in these broader caregiving networks may therefore partly confound the observed association between PE and ECD and should be explicitly examined in future work. Longitudinal or intervention studies would be needed to confirm a causal relationship. A further limitation is that both the exposure and the outcome were obtained by caregiver report, and correlated reporting error can inflate the observed association. Independent measurement of the exposure or the outcome would be needed to rule out shared-reporter bias. Because interviews took place in the home, other household members, including fathers, were sometimes present. Although the most sensitive items were self-entered when privacy could not be ensured, this was noted only occasionally in field notes rather than systematically and the presence of others may have influenced responses to sensitive items, including reports of PE and caregiver mental health. Our PE composite, while internally consistent, measured whether specific paternal activities occurred within a short reference period (eg, whether the father sang to the child in the past 24 hours) and not how often, how long or how sensitively those interactions occurred. This limitation may attenuate observed associations, as qualitative features of caregiving are often most consequential.³¹ We also acknowledge that our outcome was a standardised developmental score. While it captures broad developmental status, we could not examine whether PE relates differently to motor, language, cognitive or socioemotional development. Socioemotional outcomes in particular may be under-represented. Future research could explore specific developmental domains and longer-term outcomes (such as socioemotional competency) in relation to PE.

Implications for policy and early child programmes in Lebanon

Our findings carry several implications for child development initiatives in Lebanon and similar contexts. First, they emphasise that fathers may be an underused resource for ECD. In Lebanon, as in many countries in the West Asia region, normative gender roles have traditionally positioned fathers as economic providers and mothers as primary caregivers. At the same time, there are now explicit national efforts to involve fathers in early caregiving, including, for example, Program Abb (meaning ‘father’ in Arabic), which the Lebanese Ministry of Social Affairs recently adopted to encourage fathers’ active participation in caring for young children.³² Lebanon has also seen the introduction of several complementary initiatives designed to support fathers’ caregiving skills, including RISING Initiative (UNICEF Lebanon), which has a dedicated ‘fatherhood module’ delivered through frontline workers and Ahlan Simsim (Sesame Workshop & International Rescue Committee), one of the region’s growing parenting support programmes offering sessions that include fathers with a focus on responsive parenting, emotional well-being and learning through play.^{33 34} The Parenting Toolkit for Lebanon (Save the Children) similarly provides structured guidance to help parents (including fathers) support children’s safety, psychosocial well-being and early learning.³⁵

Our study provides empirical support for such efforts—even amid Lebanon’s ongoing ‘polycrisis’—we find that more involved fathering is associated with better developmental outcomes. Encouraging and enabling fathers to participate in daily childcare and early learning may therefore be a viable strategy to improve child development at the population level. Given the cross-sectional nature of the study, we cannot conclude that increasing fathers’ engagement would, by itself, cause improvements in development, nor can we identify which specific programme models are most effective. However, the consistency and magnitude of the association suggest that existing platforms that already reach families—such as routine child health and nutrition visits, social development centres and early childhood programmes—are logical entry points for systematically inviting fathers into caregiving and normalising shared parenting roles. Crucially, the association appears to hold across levels of household assets, suggesting that father-engagement programmes need not be targeted solely at the poorest communities, but instead can have broad relevance. Supporting fathers and other caregivers should complement, not replace, investments in health, nutrition, education and social protection systems that enable families to provide nurturing care during protracted crises. Ultimately, promoting engaged fatherhood should be seen as an investment in human capital and a relatively low-cost, culturally adaptable means of enriching children’s early experiences. With nearly 250 million children in LMICs at risk of not reaching their developmental potential, strengthening fathers’ participation in early caregiving in Lebanon and comparable settings is a timely and promising strategy to help more children not only survive, but also thrive.

The funder had no role in survey design, data collection, data interpretation, report writing, or the decision to submit for publication. The views expressed are those of the authors and not those of UNICEF Lebanon Office or the Lebanon Ministry of Public Health.

Data availability statement

Data are available on reasonable request.