Site

The Drakenstein Child Health Study (DCHS) is a multidisciplinary population-based birth cohort study investigating the early-life determinants of child health and development. The study is located in Paarl, a periurban area, 60 km outside of Cape Town in the Western Cape of South Africa [14]. It is a stable, low-socioeconomic community comprising approximately 200,000 people, characterised by a high prevalence of a range of health risk factors such as depression, childhood trauma, IPV, and poverty [15]. The DCHS is representative of many periurban regions in South Africa, as well as in other LMICs. The majority of the population accesses healthcare in the public sector. Pregnant women were recruited from two public sector primary healthcare clinics, one serving a predominantly mixed ancestry population (TC Newman) and the other serving a predominantly black African population (Mbekweni).

Population

Pregnant women were enrolled into the DCHS between 20 and 28 weeks’ gestation while attending routine antenatal care and are being prospectively followed through childbirth and early childhood until children are 10 years of age. Recruitment was unfiltered, and eligibility criteria included (1) attendance at one of the two study clinics, (2) being at least 18 years of age, and (3) intending to remain in the study area for at least 1 year. Between March 2012 and March 2015, 1,225 pregnant women were enrolled into the DCHS antenatally; 88 (7.2%) mothers were lost to follow-up antenatally, had a miscarriage, or had a stillbirth. In total, there were 1,143 live infants.

Measures

All measures were performed as part of the main study, and child development outcomes form a primary aim of the original DCHS [14,15]. Expanded descriptive detail and rationale may be found in the associated Methods paper [16].

Sociodemographic and environmental variables

Sociodemographic variables were measured using validated questionnaires administered by trained study staff at an antenatal visit at 28 to 32 weeks’ gestation. Sociodemographic variables including household factors (running water, flushing toilet, electricity in home, and household income) and maternal demographics (age at enrolment, any secondary versus only primary education, married or cohabiting with partner, employed, and whether this was the first pregnancy) were collected using an interviewer-administered questionnaire adapted from items used in the South African Stress and Health (SASH) Study [17,18].

Child and maternal physical health

Gestational age at delivery was calculated in weeks, based on ultrasound results when these were available and otherwise based on fundal height measurements and maternal report of last menstrual period. Prematurity was defined as <37-week gestational age. Birth weight was abstracted by trained study staff from hospital records at birth and was taken as a continuous measure. Duration of exclusive breastfeeding was derived based on maternal report of feeding practices at birth; 6, 10, and 14 weeks; and 6 and 9 months of child age. Exclusive breastfeeding was defined as occurring until the first maternal report of introduction of solid foods or formula. We included exclusive breastfeeding for 6 months, as per current recommendations by WHO [19].

Maternal haemoglobin was tested for in pregnancy. Haemoglobin levels < 10 g/dL in pregnancy were classified as moderate to severe iron deficiency anaemia as per WHO guidelines [20]. Maternal HIV status was established during routine HIV testing of women in pregnancy as per the Western Cape of South Africa guidelines for prevention of mother-to-child transmission of HIV. Stunting was investigated separately because of the known association with delayed development, but it was not included as a risk factor in the final model, because of the association with the other risk and protective factors [21]. Stunting was defined here as <−2 standard deviations below WHO z-score height for age at 2 years of age.

Substance use

Alcohol use during pregnancy was assessed using a composite, dichotomous measure (exposure versus no exposure) using the Alcohol, Smoking and Substance Involvement Screening Test (ASSIST) and retrospectively collected data on hazardous alcohol use during pregnancy [22,23]. The ASSIST has shown good reliability and validity in international, multisite studies. Total scores were obtained for alcohol by summing individual items related to maternal alcohol use during pregnancy; a score of >10 indicates moderate to high levels of risk for alcohol problems (reflecting weekly or daily/almost daily alcohol use and negative consequences related to the quantity of alcohol consumed). We categorised children as alcohol-exposed versus not alcohol-exposed. We used the following inclusion criteria to define alcohol exposure in order to optimise numbers: (1) scoring greater than moderate on the ASSIST performed antenatally as per reference Myers et al. [22] and Donald et al. [23] AND/OR (2) 2 or more drinks a week on the neonatal alcohol questionnaire (to further quantify alcohol use on top of the ASSIST) AND/OR (3) 2 or more drinks a week on a retrospective alcohol questionnaire regarding alcohol use during pregnancy (as mothers may be more likely to respond to this than during/immediately after pregnancy). Tobacco exposure during pregnancy was objectively measured by maternal urine cotinine collected antenatally or at birth. Cotinine, a metabolite of nicotine measurable in the urine, was measured using the IMMULITE 1000 Nicotine Metabolite Kit (Siemens Medical Solutions Diagnostics, Glyn Rhonwy, Llanberis, United Kingdom). Mothers were categorised as active smokers during pregnancy if either the antenatal or birth cotinine levels were ≥500 ng/ml [18].

Psychosocial measures

Mothers completed a battery of psychosocial measures, administered by trained study staff, at an antenatal visit between 28 and 32 weeks’ gestation. The IPV questionnaire used in this study was adapted from WHO’s multicountry study [24] and the Women’s Health Study in Zimbabwe [25]. Mothers were asked about exposure to partner behaviour and frequency of occurrence for emotional, physical, or sexual abuse behaviours; when behaviours were experienced ‘many times’, mothers were categorised as exposed to lifetime IPV. The Edinburgh Postnatal Depression Scale (EPDS) [26] was used to measure depression; this scale has been validated for use with pregnant women and in a South African population [27–29]. The EPDS consists of 10 items referring to the past 7 days; a total score was obtained by summing responses for all items, and a cutoff score of ≥13 was used to dichotomise participants into below and above threshold for depression. The Childhood Trauma Questionnaire [30] Short-Form was used to assess abuse and neglect experienced as a child. Each item was responded to on a 5-point scale; a cutoff score of >36 was used to dichotomise mothers into above or below threshold for childhood abuse. The Self Reporting Questionnaire-20 (SRQ-20) is a WHO-endorsed measure of psychological distress [31]. The SRQ-20 consists of 20 items, which assess nonpsychotic symptoms, including symptoms of depressive and anxiety disorders; items are summed, and a cutoff score of >8 was used to dichotomise mothers into below or above threshold for psychological distress [32].

Child development: Bayley Scales of Infant and Toddler Development

Child development was assessed on all available children across the full cohort at 24 months of age using the Bayley Scales of Infant and Toddler Development (Third Edition) (BSID-III) [33]. The BSID-III is a gold-standard observational measure of development for children from 0 to 42 months. It has been validated for a South African population [34,35] and found to be culturally appropriate without modifications. However, the tool may slightly underestimate delay in this population [35]. The tool measures development by direct observation across five subscales: cognition, receptive and expressive language, and fine and gross motor [36]. These scales were measured by direct observation by a trained physiotherapist and occupational therapist blinded to the child and family risk factors, overseen by a paediatric neurologist with specialist developmental expertise [16]. Quality control and monitoring processes were implemented to ensure accuracy. All data were entered into the BSID-III scoring programme, and the data were exported to Excel. In this analysis, both the raw scores and scaled scores were used. The scaled scores are calculated using a normal US population, scaled to a mean of 10 and standard deviation of 3. We assessed poor developmental outcomes [33] by categorising the scores into ‘delay’ or ‘no delay’, defined by scoring <−1 standard deviation below the mean scaled score (using a cutoff of 7).

Ethical considerations

Ethical approval was obtained from the Faculty of Health Sciences Research Ethics Committee, University of Cape Town (401/2009), and by the Western Cape Provincial Research committee (2011RP45). Mothers gave written informed consent at enrolment and were reconsented annually for study involvement. Consent was done in the mother’s preferred language: English, Afrikaans, or isiXhosa. When significant developmental delays were identified by study staff, mothers and children were referred to local healthcare services for further assessment and management.

Statistical strategy

The analyses followed a defined approach that was decided upon prior to running the models, though no prespecified analysis plan exists. Thus, data-driven approaches to analysis were not used at any stage. Early childhood development outcomes’ scores were measured by five developmental domains using the BSID-III: cognitive development, language (receptive and expressive) development, and motor (fine and gross) development. We used a cutoff scaled score of less than 7 (<−1 SD below the mean scaled score) to define developmental delay.

BSID-III raw scores were analysed using linear regression, and developmental-delay data were analysed using logistic regression separately for each domain. Odds ratios are presented for logistic regression models, and regression coefficients are presented for linear regression models with 95% confidence intervals (CIs) for both. We used a missing-at-random approach and generated a complete-case dataset, which omitted records that were missing data in any of the covariates that we considered plus any domain (apart from gross motor, which was not analysed). Multivariable regression models were built using the complete-case dataset in a hierarchical approach. Initial models adjusted for maternal education, child age, and sex forced into the model as known and well-proven factors that impact child developmental outcomes. Subsequently, other covariates were added in blocks following the hierarchical order from sociodemographic and environmental variables to maternal and child physical variables to maternal psychosocial variables based on prior literature [37] as illustrated in Fig 1. Covariates added from each block were selected using a best-subsets variable-selection approach that aimed to minimise Akaike information criterion (AIC; using the ‘gvselect’ command in Stata 14) [38,39]. That is, the model for one outcome considered forced variables first and compared the fit of the model to the fit of the models with all possible combinations of variables from the sociodemographic block. Only the subset of these variables that reduced the AIC were retained for consideration with the next block (this includes an empty set, which would arise if no variables reduced the AIC from the previous model). Collinearity was assessed using variance inflation factors after each block, and then these steps were repeated for the next two blocks of variables. After creating a final model assessing the associations between risk and protective factors and each developmental outcome, we assessed whether including interaction terms between child sex and any variable retained from the best-subsets regression reduced the AIC further, to explore whether child sex altered the association between any of these variables and developmental outcomes. We report both the final model and the model including interaction terms. Finally, we used the same process to assess the outcome of global developmental delay in all domains consisting of delay in cognition, language (combining delay in receptive or expressive language), fine motor, and gross motor.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Methodological conceptual framework: Hierarchical model.

This was adapted from the model by Chopra and colleagues [36].

https://doi.org/10.1371/journal.pmed.1002920.g001

All analyses were performed with Stata version 14 [40].

Child development

A total of 369/731 children (50.5%) were categorised as having cognitive delay, 402/723 (55.6%) with receptive language delay, 389/702 (55.4%) with expressive language delay, 169/730 (23.2%) with fine motor delay, and 267/696 (38.4%) with gross motor delay (Table 2). Overall, 405 children from the 734 (55.2%) were classified as having delay in two or more domains; the combination of cognitive and language delay was the most common, with 304 children (41.4%). There were 75 children (10.2%) who were delayed in all four domains.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Child developmental outcomes at 24 months of age for all children who completed at least one domain of the BSID-III (n = 734).

https://doi.org/10.1371/journal.pmed.1002920.t002

On raw scores, boys exhibited lower total scores than girls for all domains except gross motor on bivariate analyses (Table 2 and S2 Table), as well as in adjusted analyses in cognition (β = −0.74; 95% CI −1.46 to −0.03, p = 0.042), receptive language (β = −1.10; 95% CI −1.70 to −0.49, p < 0.001), expressive language (β = −1.65; 95% CI −2.46 to −0.84, p < 0.001), and fine motor (β = −0.70; 95% CI −1.20 to −0.20, p = 0.006) raw scores (Table 3). Likewise, boys were at increased risk of delay in all domains except gross motor (Table 2 and S3 Table) (p < 0.05) in bivariate analyses, and this association held when adjusting for confounders in the multivariable models (S4 Table): cognitive delay (adjusted odds ratio [aOR] 1.35; 95% CI 0.95–1.91, p = 0.098), receptive language delay (aOR 2.40; 95% CI 1.68–3.42, p < 0.001), expressive language delay (aOR 2.12; 95% CI 1.49–3.03, p < 0.001), and fine motor delay (aOR 1.92; 95% CI 1.25–2.95, p = 0.003).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Multivariable linear regression results demonstrating the association of risk and protective variables with developmental domain outcome for total raw scores and exploring interaction with child sex.

https://doi.org/10.1371/journal.pmed.1002920.t003

At 2 years of age, 19.3% of children were classified as stunted (22.4% boys and 15.8% girls), and stunting was found to be associated with poor cognitive development (β = −1.13; 95% CI −2.10 to −0.15) and with cognitive developmental delay (OR = 1.77, 95% CI 1.18–2.65). As expected, across the group, children who were older tended to perform better in the developmental assessment, particularly in the cognitive and language subscales. However, the age window was very small, so the variance may not be meaningful in this analysis. In the gross motor domain, we investigated sex, education, and child age and we did not see any of the expected bivariate associations and therefore we did not run multiple regression analyses on this outcome variable.

Risk and protective factors for child development

Bivariate analyses (S2 and S3 Tables) and the final model multiple regression results are shown in Table 3, S4 Table, Fig 2, and S1 Fig. We first describe the final model examining associations between risk and protective factors with child developmental outcomes and then explore the interactions with child sex.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Marginal effects of birth weight and sex on cognition.

https://doi.org/10.1371/journal.pmed.1002920.g002

Sociodemographic and environmental factors

On bivariate analysis, protective factors associated with better development and reduced risk of delay comprised having higher maternal education, older child age, a primigravid mother, and better-resourced households (higher household income, flushing toilet, running water). Higher maternal education (any secondary versus only primary) was associated with increased total raw scores for all domains on bivariate analyses and in the final model for cognition (β = 1.70; 95% CI 0.45–2.95, p = 0.008) as well as lower odds of cognitive delay (aOR 0.52; 95% CI 0.28–0.99, p = 0.045) and receptive language delay (aOR 0.51; 95% CI 0.26–0.99, p = 0.045). Children of primigravid mothers had higher expressive language scores (β = 0.93; 95% CI 0.06–1.80, p = 0.036) and lower odds of expressive language delay (aOR 0.57; 95% CI 0.36–0.90, p = 0.016) in multiple regression models. In the interaction models, sex impacted the association between primigravida and expressive language outcomes. We calculated estimates of the interaction effects of primigravida in both boys and girls separately. For girls, the beta coefficient was 2.07 (95% CI 0.82–3.31, p = −0.01), and for boys, it was −2.88 (95% CI −4.64 to −1.12, p = 0.002). This suggests that primigravida status had a positive effect on expressive language for girls but a negative effect for boys in this cohort.

Flush toilets in household were associated with increased total raw scores for expressive language (β = 1.62; 95% CI 0.39–2.86, p = 0.010) in the multivariable regression. Tap water showed an interaction with child sex, and we calculated estimates of the effects of tap water in both boys and girls separately in the interaction model. For girls, the beta coefficient was −3.10 (95% CI −4.70 to −1.49, p < 0.001), and for boys, it was −3.72 (95% CI −5.41 to −2.02, p < 0.001). The effect direction is unexpected and is perhaps caused by the additional inclusion of flushing toilet in the model, although these together reduced the AIC.

Mother and child physical factors

On bivariate analysis, key maternal and child physical health risk factors for lower developmental scores or higher odds of delay included maternal anaemia, prematurity, maternal HIV, alcohol or tobacco use during pregnancy, and maternal depression and lifetime IPV; protective factors included higher birth weight. Increasing birth weight was associated with significantly greater scores and lower odds for delay for all domains in the total cohort in bivariate and multivariable models (Table 3 and S4 Table). There was also evidence of an interaction with child sex in cognition and language outcomes. We calculated marginal effects for the interaction of birth weight and cognition, and at a birth weight of 3 kg (mean of cohort is 3.05 kg), the marginal effect was −0.83 (95% CI −1.55 to −0.12, p = 0.022). See Fig 2 for all marginal effects. At lower birth weights, boys did less well than girls (negative coefficients indicate a lower score); and at higher birth weights, boys did better. Likewise, we calculated marginal effects for the interaction of birth weight and receptive language, and at a birth weight of 3 kg, the marginal effect was −1.17 (95% CI −1.77 to −0.56, p < 0.001). See S1 Fig for all marginal effects. As with cognition, at lower birth weights boys did less well than girls.

In the multivariable final model, maternal anaemia in pregnancy was associated with lower cognitive scores (β = −1.38; 95% CI −2.33 to −0.44, p = 0.004), receptive language scores (β = −1.12; 95% CI −1.93 to −0.31, p = 0.007), and expressive language scores (β = −1.11; 95% CI −2.18 to −0.04, p = 0.043). There was some evidence for interaction of child sex in this association. We calculated estimates of the effects of anaemia in both boys and girls separately for receptive and expressive language. For girls (receptive language), the beta coefficient was −1.78 (95% CI −2.89 to −0.67, p = 0.002), and for boys, it was −5.68 (95% CI −9.25 to −2.11, p = 0.002). For girls (expressive language), the beta coefficient was −1.99 (95% CI −3.46 to −0.52, p = 0.008), and for boys, it was −2.79 (95% CI −4.79 to −0.78, p = 0.007). This suggests that anaemia in pregnancy had a greater negative effect on receptive and expressive language for boys than for girls, although it still had an effect in the girls. Maternal anaemia was also robustly associated with increased odds of cognitive delay (aOR1.66; 95% CI 1.04–2.66, p = 0.033), receptive language delay (aOR1.77; 95% CI 1.09–2.88, p = 0.020), and expressive language delay (aOR1.83; 95% CI 1.12–2.98, p = 0.015) overall. Maternal anaemia demonstrated a stronger interaction with these outcomes in boys as above.

Maternal psychosocial factors

On bivariate analysis, maternal psychosocial risk factors associated with poorer developmental outcomes comprised maternal antenatal depression or lifetime maternal exposure to IPV. In multivariable regression, antenatal depression was associated with poorer cognitive scores (β = −1.03; 95% CI −1.04 to −0.12, p = 0.027). For maternal IPV, though the model without interactions was not significant in the adjusted analysis, there was an association with higher odds of expressive language delay, which came out in the interaction model (main effect aOR 1.88; 95% CI 1.13–3.13, p = 0.015).

Developmental delay in all four domains

In bivariate analyses, maternal education and higher birth weight were protective against developmental delay in four domains, and preterm birth was associated with poorer outcomes.

On adjusted analyses, maternal education (aOR = 0.40; 95% CI 0.17–0.94, p = 0.035) was found to be protective against developmental delay in all four domains across the total cohort (S5 and S6 Tables)