Data
This study analysed data from the [masked] data set. The [masked] is a prospective birth-cohort study conducted in eight low and middle-income cities in diverse regions, including Kingston (Jamaica), Koforidua (Ghana), Worcester (South Africa), Cluj-Napoca (Romania), Tarlai Kalan (Pakistan), Ragama (Sri Lanka), Hue (Vietnam), and Valenzuela City (the Philippines). The [masked] was designed to provide high-quality longitudinal evidence to support effective interventions to tackle violence against women and children. The [masked] currently consists of three completed waves of data collection; the first wave – the focus of the current study – was conducted when participating women were in the third trimester of pregnancy.
The [masked] questionnaires were first developed in English and then translated into nine different languages (Urdu, Afrikaans, IsiXhosa, Romanian, Filipino (Tagalog), Sinhala, Tamil, Vietnamese, and Twi) guided by the WHO Guidelines on Translation. The protocol of this study provides further details of the data collection procedures.25
Ethics
Ethical approval was obtained following national specific procedures in each of the eight participating study sites and the coordinating site prior to the start of data collection.
Sample
A convenience sampling method was employed to recruit participants through direct contact by the fieldworkers in the clinic waiting room, or by a health worker who would then refer potential participants to the fieldworkers. Recruitment strategies were adapted in each site. Pregnant women were invited to participate in the study if they were: i) in their third trimester of pregnancy (i.e., weeks 29-40); ii) aged 18 and over; iii) residing within the study area and with no plans to migrate during the first three months post-birth; and iv) able to give informed consent. The total baseline sample consisted of 1,208 participants (150 approx. per site) that were on average 28.27 years old (SD = 5.81 years, range: 18–48 years). For further participant demographic characteristics see [masked].
Procedure
Eligible women were invited to provide written informed consent in their relevant language. When necessary, alternative means of providing consent were offered, following the WHO’s Research Ethics Committee.25
Female fieldworkers interviewed participants after receiving 40 hours of standardised in-person training.25 Baseline data collection started in February 2019 and ended in July 2019. Interviews combined Computer-Aided Personal Interviews (CAPI) and Computer-Assisted Self-Interviewing (CASI) for the more sensitive items. Interview settings varied by site and included primarily the project office or clinic designated space, and the participants’ houses.
Measures
Neighborhood characteristics scales.
To measure neighborhood characteristics, an instrument was specifically developed by the [masked] Consortium, with items adopted from existing measures as described below.
Neighborhood cohesion: Five items from Mujahid et al.’s10 scale ask about presence of support and help, positive relationships, trust, and shared values within the neighborhood. Responses are measured on a four-point Likert scale ranging from 1 (strongly agree) to 4 (strongly disagree) (Cronbach’s alpha = 0.83).
Intergenerational Closure: Four items from Sampson et al.18 assess intergenerational closure, measured on a four-point Likert scale ranging from 1 (strongly agree) to 4 (strongly disagree). Items ask about the presence of adults in the neighborhood that watch out children, if these adult figures can be looked up to by children, and relationship among parents (Cronbach’s alpha = 0.63).
Neighborhood disorder: Four items from the Neighborhood Disorder Observation Scale9 assess the dimensions of neighborhood disorder, with a four-point response scale ranging from 1 (not a problem) to 4 (large problem). Items to measure neighborhood disorder ask about litter in the streets, smells and fumes, noise from traffic or other homes, and traffic and road safety (Cronbach’s alpha = 0.80).
Social disorder: Five items from the Neighborhood Disorder Observation Scale9 assess the dimensions of social disorder, with a four-point response scale ranging from 1 (not a problem) to 4 (large problem). Items inquire about vandalism, people being drunk on the streets, gangs, fights and arguments on the streets, and whether people are afraid of going out at night (Cronbach’s alpha = 0.87).
Nomological net measures.
Perceived Stress Scale (PSS)26: This 10-item instrument measures how stressful certain life situations are rated by respondents during the last month. Responses were measured on a four-point Likert scale ranging from 1 (not at all) to 4 (nearly every day) (Cronbach’s alpha = 0.76).
WHO (Five) Well-Being index (1998 version) 27: The WHO Well-Being Index is a five-item screening questionnaire to determine subjective psychological well-being of respondents within the past two weeks. Response categories include a six-point Likert scale ranging from 1 (at no time) to 4 (all the time). (Cronbach’s alpha = 0.84).
Patient Health Questionnaire (PHQ-9)28: To measure the severity of depressive symptoms in the last two weeks, the PHQ-9 employs nine items covering anhedonia, dysphoria, sleep disturbances, fatigue, changes in eating, low self-esteem, concentration difficulties, hypo-or-hyper-active behaviours, and suicide ideation. Response categories include four points ranging from 1 (not at all) to 4 (nearly every day) (Cronbach’s alpha = 0.76).
Analytical strategy
Factorial validity and measurement invariance
To evaluate the factorial structure of the selected measures, we conducted a confirmatory factor analysis including four factors: neighborhood cohesion, intergenerational closure, neighborhood disorder, and social disorder. This factorial structure was fitted individually for each country. Correlations were permitted between the factors, with non-significant correlations removed from the final models. Model identification was specified using the ‘marker indicator’ method, whereby the loading of one item was fixed to 1 and the variance of the factor was freely estimated. To account for the categorical nature of the items, all models were estimated using the weighted least squares with robust means and variances (WLSMV) estimator. Model fit was evaluated using the comparative fit index (CFI), the Tucker-Lewis index (TLI), the root mean square error of approximation (RMSEA), and the root mean square residuals (SRMR). Good model fit was indicated by CFI ≥ .95, TLI ≥ .95, RMSEA ≤ .06 and SRMR ≤ .08; adequate model fit was indicated by CFI ≥ .90, TLI ≥ .90, RMSEA ≤ .08 and SRMR ≤ .08.29-31
To examine whether the measures operated equivalently across countries, we conducted measurement invariance tests. Importantly, on several of the items, category 4 of response was not endorsed by any participant or was endorsed by very few participants, in at least some of the countries. Therefore, to balance the aim of comparing responses across all eight countries with the aim to maximise scale integrity (i.e., retaining the entire spectrum of possible response choices), we employed the following rule: we collapsed categories 3 and 4 where not more than 10 people endorsed response category 4 in at least 4 of the countries (i.e., 50% of the sites) or where category 4 was not endorsed by any participant in at least 1 country (because such a case impedes invariance testing). Consequently, the social cohesion and intergenerational closure subscales contained 3 categories for the purpose of this analysis, whereas 4 categories were retained for neighborhood and social disorder.
Measurement invariance was tested at three levels: configural, metric, and scalar. First, the configural model – serving as the baseline - inspected the extent to which the same factorial structure (i.e., pattern of loading) was applicable across the eight sites, namely whether the item-to-factor relationships was observed across sites. This is the weakest form of invariance. Secondly, the metric model was specified as a nested model within the configural model, with loadings of corresponding items constrained to be equal across groups. Finally, the scalar model was specified as a nested model within the metric model, with each threshold of each item constrained to be equal across groups, allowing latent factor means to be compared. To ensure model identification, the configural and metric models contained several parameters that were fixed in all groups: (a) the means of all factors were fixed to zero and (b) the item scale factors were fixed to 1. The scalar model retained these fixed parameters only for the ‘reference’ group. As there was no specific rationale for choosing one country over another as the ‘reference’ group, this was given by the first country in an alphabetically ordered list of country names (i.e., Ghana).
Configural invariance was achieved if the configural model fitted the data well according to the CFI, TLI, RMSEA and SRMR. Metric invariance was achieved if the fit indicators did not deteriorate by values greater than: .10 for CFI and TLI, .015 for RMSEA and .03 for SRMR.32 Scalar invariance was achieved if the fit indictors did not deteriorate by values greater than: .10 for CFI and TLI, .015 for RMSEA and .01 for SRMR.32 In the case of non-invariance, partial invariance was sought by releasing constraints on the loadings (in the metric model) or thresholds (in the scalar model), under guidance from the model modification indices. To retain the maximum level of invariance possible, such modifications were only performed for the countries where non-invariance was observed (i.e., it was possible for an item’s thresholds to exhibit invariance across six out of eight countries).
All models were conducted using the software Mplus v.8.8.33 To aid the identification of the best possible solution, we employed 10 random starts and 10,000 iterations.
Nomological networks
We obtained Pearson’s correlations between all the variables for each of the study sites. Then, we used qgraph with R34 to create diagrams with the aim of illustrating the association between the four neighborhood characteristics that concern the current study and the maternal outcomes of interest and compare these associations across countries. In the figures, the edge thicknesses are proportional to the magnitude of the Pearson’s correlations between variable, which allows visual comparison across study sites.