Early-life maternal attachment and risky health behaviours in adolescence: findings from the United Kingdom Millennium Cohort Study

Using data from the MCS, we investigated associations between early-life MA at 9 months and adolescent MRBs at 17 years old in the UK. The MCS is a nationally representative prospective cohort study following the lives of 18,296 singleton children born from September 2000 to January 2002 [18]. A total of seven data collection sweeps were conducted at 9 months, 3, 5, 7, 11, 14 and 17 years old. Responses were received from cohort members and parents, where mothers were typically the main parent respondent. Oversampling participants from ethnic minorities and disadvantaged socioeconomic backgrounds was conducted to ensure sufficient statistical power in these smaller subgroups. Stratified clustered sampling was employed to broadly represent all four countries in the UK (England, Wales, Scotland and Northern Ireland) [19]. Study weightings were applied to account for the sampling design and adjust for non-response at each cross-sectional sweep. More information on the sampling methods and the cohort profile is detailed online (https://​cls.​ucl.​ac.​uk/​cls-studies/​millennium-cohort-study/​). MCS datasets were accessed and downloaded from the UK Data Service website, and all analyses were performed on Stata/IC 16.0 [20].

The MCS secured ethical approval from the National Health Service Multi-Centre Research Ethics Committee (NHS MREC) for each sweep. All methods comply with the relevant national and institutional committees on human experimentation, including the ethical standards outlined in the 1975 Helsinki Declaration, as revised in 2013 [21, 22]. With cohort members under 18, written informed consent was required and obtained from parents/legal guardians. Following interviews, ongoing support was also made available to cohort members and parents through information booklets, where links to professional support services and helplines were provided [23]. This study required no further ethical approval.

The first data collection sweep at age 9 months included 18,296 singleton-born children. However, by the seventh sweep, participating families had declined to 10,496 due to attrition commonly affecting cohort studies (Fig. 1). The MCS has weightings for attrition to account for follow-up loss from previous cross-sectional sweeps of surveys [24]. Nonetheless, the overall response rate was an acceptable 74% at 17 years, suggesting that missingness was largely due to follow-up loss rather than questionnaire response [23]. Additional comparison between the omitted and analytical complete case samples was conducted and presented in Additional file 1 – List of Appendices: Appendix Table 1. No significant differences were observed, so a Missing Completely At Random assumption was undertaken.

We undertook a complete case analysis following an acceptable response rate of 74% and a comparison between omitted and analytical samples (Appendix Table 1) [23]. Attrition weights were applied at each studied sweep to account for missing data due to follow-up loss up to the seventh sweep at age 17.

We first estimated the prevalence of risky behaviours among cohort members at 17 years by MA at 9 months. Bivariate associations between risky health behaviours, such as smoking and vaping, were examined. These logistic regressions in Appendix Table 5 assessed the co-occurrence of risky behaviours and that a sum score of MRBs was suitable for this study. MA was studied as a binary variable following methods by Wadman et al. (2020), yet associations as a continuous variable are also presented in Appendix Table 3 [29].

Guided by the logic model (Appendix Fig. 1), we used multivariate logistic regression analysis to compare MRB uptake among adolescents at 17 by higher or lower MA scores at 9 months. The baseline model first assessed univariate associations between MA and MRBs. Adjusting for blocks of potential explanatory factors in a stepwise manner, we then examined how the odds ratio (OR) changed on the inclusion of each block of potential explanatory factors in five models through mediation analysis.

Model 1 adjusted for low attachment mechanisms (multiple births, unplanned pregnancy, infant prematurity, breastfeeding and maternal age at birth), Model 2 adjusted for maternal depression, Model 3 adjusted for poor adolescent mental health, Model 4 adjusted for socioeconomic factors (single-parent status, SECs by maternal education and household income by UK quintiles) and Model 5 fully adjusted for all explanatory factors. All models except the baseline adjusted for infant sex. To estimate the change in OR of MRBs for those with lower compared to higher attachment, we calculated the difference as 100 x (baseline OR – adjusted OR) / (OR – 1), whereby adjusted ORs accounted for blocks of explanatory factors [38, 39].

Causal pathways are challenging to understand. Explanatory factors are expected to cluster and are difficult to interpret in isolation. We, therefore, took a sequential “en-block” approach through mediation analysis to examine the cumulative effect of blocks of potential mediating and confounding factors, considering low attachment mechanisms, maternal mental health, adolescent mental health, socioeconomic factors [40]. Pearson goodness-of-fit tests were applied after logistic regression models to assess whether there was no departure from logistic regression assumptions and no interactions were present, provided that the resulting p-value was non-significant (> 0.05). Multi-collinearity, variables explaining the same variation in a model, was also examined through mean Variance Inflation Factors. All six logistic regression models (including the baseline) in this study met their required assumptions. Analyses were conducted on Stata/IC 16.0.

As alcohol consumption tends to be patterned differently than other risky behaviours, with adolescents from high-income backgrounds more likely to consume alcohol than low-income peers, we conducted sensitivity analyses of the association between MA and MRBs without alcohol consumption [41].

A total of 7796 children had complete case observations on exposure, outcome and potential explanatory variables from the first (9 months), sixth (14 years) and seventh (17 years) sweeps (Fig. 1). No significant differences between omitted and complete case samples (see Appendix Table 1) were observed following an acceptable survey response rate of 74% before employing a Missing Completely at Random assumption [23].

At 9 months, 6907 mothers had higher MA (88.72, 95% Confidence Intervals (95% CI): 87.73–89.64%), while 889 had lower attachment (11.28, 95% CI: 10.36–12.27%). Table 1 shows that 2588 cohort members (31.43, 95% CI: 29.84–33.06%) reported no MRB uptake, while 5208 (68.57, 95% CI: 66.94–70.16%) had engaged in 2 or more risky activities at 17 years old. Children of mothers who reported lower attachment were more likely to engage in MRBs at 17 years old (72.36, 95% CI: 68.13–76.22%) than mothers reporting higher attachment (68.09, 95% CI: 66.43–69.70%).

With the exception of alcohol consumption, all individual risky behaviours had increased odds of partaking in other risky behaviours (Appendix Table 5). For example, adolescents who regularly smoked had 12.7 times the odds of regularly vaping (OR: 12.70, 95% CI: 10.55–15.48). The highest association between risky behaviours was between regular smoking and illegal drugs (OR: 13.22, 95% CI: 11.20–15.61).

Adolescent alcohol consumption was associated with decreased odds of regular vaping (OR: 0.87, 95% CI: 0.57–1.33) and gambling (OR: 0.86, 95% CI: 0.61–1.22).

Lower attachment at 9 months among mothers was associated with a higher OR of MRBs at 17 years in the unadjusted baseline model (OR: 1.23, 95% CI: 1.00–1.50). Table 2 shows the extent to which the elevated OR of MRBs by attachment during early life was attenuated when adjusting for blocks of potential explanatory factors in a stepwise manner.

Adjustment for low attachment mechanisms (Model 1: multiple siblings, prematurity, breastfeeding, maternal age at birth, unexpected pregnancy and infant sex) led to a 13% attenuation in the OR of MRBs in adolescents from mothers with lower attachment (OR: 1.20, 95% CI: 0.98–1.47). Adjustment for maternal depression (Model 2) led to a 26% attenuation in the OR of MRBs in adolescents from mothers with lower attachment (OR: 1.17, 95% CI: 0.95–1.44). Adjusting for adolescent mental health (Model 3) saw a 17% attenuation in the baseline OR (OR: 1.19, 95% CI: 0.97–1.45). Adjustment for social inequality factors (Model 4, including single-parent status, household income by UK quintiles, maternal education according to NVQ levels) led to a 17% attenuation in the odds ratio of MRBs (OR: 1.19, 95% CI: 0.97–1.45). In the final model, adjusting for all blocks of explanatory factors together, the overall reduction in OR was 30%.

Sensitivity analyses findings without alcohol consumption were consistent with the associations observed by the main models in Table 2 (Appendix Table 6).

Having lower MA at 9 months increased the likelihood of partaking in MRBs in adolescence by 23%. Almost a third (30%) of the excess risk was attributable to child, maternal and socioeconomic characteristics during the early years of life, with the leading explanatory factor, maternal depression, accounting for 26% of the association between MA and MRBs.

The majority of individual risky behaviours were significantly associated with each other, supporting recommendations from Hale and Viner (2016) and a systematic review by Meader et al. (2016) to target MRBs concurrently [6, 8]. Whitaker et al. (2021) found that substance use such as alcohol consumption, smoking and illegal drugs clustered in unhealthy adolescents, while risky sexual activity and gambling typically occurred in isolation [42]. By contrast, this study finds reduced odds in alcohol consumption with vaping and gambling, suggesting that adolescents who drink may not engage in other risky behaviours. A possible explanation from Hale and Viner (2016) cites the problem behaviour theory by Jessor et al. (1998), hypothesising that alcohol consumption is perceived as socially acceptable among adolescents and is adopted independently without other MRBs [6, 43].

Sensitivity analysis findings without alcohol consumption were consistent with previous research, suggesting that adolescents from wealthier backgrounds had a higher prevalence of drinking than poorer households [41]. An explanation from Collins (2016) proposes that higher disposable income and financial support from wealthier parents may enable adolescents to spend more money on alcohol than their socioeconomically disadvantaged peers [44]. As a result, taking perspectives from both Jessor et al. (1998) and Collins (2016), increased alcohol consumption among adolescents from affluent families may be due to a combination of higher financial opportunity and social acceptability [41, 43].

By contrast, other risky behaviours in this study were associated with low SECs, showing a reverse trend compared to alcohol consumption. Similar findings by Delker et al. (2018) from the United States National Institute of Child Health Development study show that insecure-disorganised children at 15 months with poor SECs were five times more likely to adopt MRBs at 15 years than securely attached children from affluent backgrounds (n = 1364) [45]. This study further reinforces social inequalities as a confounder between MA and MRBs and comparatively investigates a longer timeframe using a larger UK cohort. However, Delker et al. studied attachment from the child’s perspective, while this study focuses on the mothers’ feelings and thoughts. Uniting both perspectives could provide further insight into the influence of attachment or emotional ties between mothers and children on later health outcomes. Another UK study by Kipping et al. (2015) using the Avon Longitudinal Study of Parents and Children also found that 6406 adolescents aged 15–16 were more likely to engage in MRBs with each incremental decrease in SEC (OR: 1.22, 95% CI: 1.15–1.29), maternal education (OR: 1.15, 95% CI: 1.09–1.21) and income (OR: 1.12, 95% CI: 1.08–1.16) [46]. Such social patterning of risky behaviours associated with low SECs is also characteristic of a lack of health literacy and access to health facilities in deprived neighbourhoods, potentially contributing to increased MRB uptake among adolescents from lower-income families [46]. A systematic review by Reiss (2013) further emphasises the influence of socioeconomic inequalities on child mental health, where low SECs was strongly associated with higher mental health problems [47]. As a result, additional efforts to support mothers working longer hours due to socioeconomic disadvantage is essential to nurture secure infant attachment, as well as healthy mental and physical child development.

Attenuating the highest proportion of the association between MA and MRBs, our study further highlights the influence of maternal depression on child health. Previous studies suggest that maternal depression may affect how mothers behave and interact with infants, thereby influencing MA. A meta-analysis by Barnes and Theule (2019) using 42 eligible studies found a small but significant association between maternal depression and attachment insecurities overall [16]. Findings showed 20% more cases of lower attachment among infants with depressed mothers compared to mothers without depressive symptoms. Decreased affection and sensitivities demonstrated by depressed mothers is a potential explanation, as depression is characterised by desolate, irritable or vacant moods [48]. Flouri and Ioakeimidi (2018) using the MCS also found that children who experienced maternal depression at 3, 5, 7 and 11 years reported more risky behaviours at 11 years old than children with non-depressed mothers, which was consistent with previous studies [17, 49, 50]. This study consequently underscores the importance of maternal depression as an early-life risk factor mediating an increased risk of MRB uptake through lower MA.

To our knowledge, this is the first longitudinal study to test whether MA formed during infancy is associated with subsequent risky health behaviours during adolescence in the UK, whilst exploring a wide range of potential confounders and mediators. The significant strengths of our study include the use of data from a large, UK-representative cohort, with longitudinal follow-up from birth to adolescence.

However, as with all observational studies, attrition may be a potential source of bias. The MCS at baseline had a representative sample of 18,296 singleton children born in the UK between September 2000 and January 2002. Yet, by age 17 years, 10,496 children had participated in all subsequent waves. To minimise attrition bias, we used response weights to account for the loss of respondents up to age 17 years, yielding a 74% response rate [23, 51]. The attrition weights adjust the sample composition by taking into account the selective loss of respondents, yet some population groups may still be overlooked. For example, low-income families may be less likely to remain in the cohort or attrition due to MA may underrepresent mothers and infants with severe attachment complications. More information on the MCS attrition weights is available in Ketende and Jones’ (2011) “User Guide to Analysing MCS Data using Stata” [51]. Yet, comparisons between omitted and analytical sample characteristics were similar (Appendix Table 1), indicating no relationship between the missing and observed data.

Combining risky behaviours into a sum score increased statistical power to examine similar explanatory mechanisms together, but the inquiry into links specific to each risky behaviour is complicated. Bivariate regression results presented in Appendix Table 5 investigating associations between individual risky behaviours showed increased odds of most risky behaviours occurring together. However, following anomalous reduced odds in alcohol consumption with gambling and vaping, we also conducted sensitivity analyses to examine any differences without alcohol from the main results. These findings were consistent with the associations observed by the main models and found that maternal depression attenuated the highest proportion of baseline odds (25%) following full adjustment (29%) (Appendix Table 6).

To identify any reporting bias of substance use reported by adolescents, a question on a fictional drug, “Semeron”, was administered in the illegal drug questions [52]. Four cohort members responded “yes” to Semeron attempts alongside illegal drug use. These members were a small proportion (0.04%) of the total final sample (n = 7796 after exclusion) and were subsequently removed from the analyses to maintain response validity (Fig. 1).

There is an opportunity to integrate and deliver both child and maternal health needs under the Global Strategy for Women’s, Children’s and Adolescent’s Health currently implemented by the World Health Organisation and Every Woman Every Child movement [53]. Aligned with the Sustainable Development Goals, an ongoing target is to reduce premature deaths due to non-communicable diseases by a third and promote mental health by 2030. Present recommended evidence-based intervention postnatal packages include early breastfeeding initiation, responsive caregiving and post-partum maternal depression screening, while adolescent health focuses on psychosocial support and promoting healthy behaviour. Regularly monitoring and supporting these needs throughout the life course by recognising early-life MA as an overarching child development need could innovate health delivery through integrated MRB interventions during adolescence.

A systematic review by MacArthur et al. (2018) highlighted that current adolescent interventions targetting MRBs yield mixed success, as school-based programs were more effective than individual- and family-based agendas [54]. Universal school programmes targeting a larger population of adolescents effectively reduced smoking (OR: 0.77, 95% CI: 0.60–0.97) alcohol consumption (OR: 0.72, 95% CI: 0.56–0.92), illegal drug use (OR: 0.74, 95% C: 0.55–1.00) and antisocial behaviour (OR: 0.81, 95% CI: 0.66–0.98), but not other MRBs [54]. By contrast, despite most family-based interventions focusing on improving parent-child relationships to reinforce family attachment, studies were low to moderate quality, and findings yielded no or little effect on MRBs. These findings suggest a need for better quality family intervention studies to inform policies on early-life MA and adolescent MRBs [54]. In low- and middle-income countries, Gilmore and McAuliffe (2013) also note the implementation of community health workers as an essential cost-effective resource providing maternal and child support promoting secure MA and aiding the recovery of maternal depression in intervention groups [55]. Increasing awareness and guidance from healthcare workers on the shared influence of early-life factors, including socioeconomic, mental and psychosocial influences on mothers and children, could optimise current patient-centred strategies reducing MRBs.