We performed secondary data analyses on the Future Families and Child Wellbeing (FFCW) data, previously named Fragile Families and Child Wellbeing, a publicly accessible dataset (Reichman et al., 2001). Institutional Review Board approval was not needed for the present study. The original FFCW study sampled 4,898 parents who shared in the birth of one child during the years 1998-2000. Participants were sampled across several densely populated cities in the United States using stratified sampling methods to oversample unmarried parents. The original study aimed to understand the risks associated with family dissolution.
Participants
Of the 4,898 children (2,341 girls), most children (93%) spent all or most of their time living with the primary caregiver, typically the mother, at 9 years old. Some primary caregivers changed across waves. Of families with complete data on relationship to the child, mothers were listed as the primary caregiver who participated in the parenting observation 99% of the time when children were 3 years old, 98% of the time when children were 5 years old, and 92% of the time when children were 9 years old. Only twelve mothers were missing data on race and ethnicity, while 33% of children were missing data. Of children with complete data, 49% were Black and non-Hispanic, 25% were Hispanic, 18% were White and non-Hispanic, 5% were multi-racial, and 3% were of another race or ethnicity not listed here. Of mothers with complete data, 48% non-Hispanic and Black, 27% Hispanic, 21% non-Hispanic and White, and 4% of mothers were of another race or ethnicity not listed here. On average, mothers were about 25 years old when children were born. Of mothers with complete data, more than half had a high school degree or equivalent (65%), and 28% of children lived in homes at or below the poverty line when the child was 1 year old.
Measures
The Infant Toddler Home Observation of the Environment (IT-HOME) was used to measure responsive parenting when children were 3 years old, and the Early Childhood HOME was used to measure responsive parenting when children were 5 and 9 years old (EC-HOME) (Bradley et al., 1988). The IT-HOME and EC-HOME Parental Responsivity subscale indexes a number of parent’s positive and affectionate behaviors during observation believed to contribute positively to child development of children’s self-regulatory mechanisms; a sum score was created as recommended for each age (Bradley, 2015). Parental responsivity observations were completed by trained researchers of the original FFCW study, and ratings were scored on a scale of 0-1 to create a total score (e.g., Parent positively praises child at least twice = 1, Parent does not positively praise the child twice = 0). The participating caregiver was observed without prompting on directed activities and behavioral interactions were coded by the researcher during the visit. FFCW used nine of the eleven items on the Responsivity subscale at 3 years old and eight items at 5 and 9 years old. Two reviews of the psychometric properties of the HOME found that the average interrater reliability of observers across studies was 90% (Elardo & Bradley, 1981; Totsika & Sylva, 2004). Information on inter-rater reliability of observers within the original FFCW is not presently provided. Given that the HOME is an index of formative rather than reflective indicators with no underlying latent construct, responsivity was treated as an observed variable (Bradley, 2015).
Internalizing Problems were assessed using the Child Behavior Checklist (CBCL), which consists of eight syndromes or subscales (Achenbach, 1996). Items are rated on a 0 - 2 scale (0 = never true, 1 = sometimes true, 2 = very often true). A latent variable reflective of Anxious/Depressed and Withdrawn problems known to commonly co-occur in young children were included for each age (3, 5, and 9 years old) using the primary caregiver’s report on the Anxious/Depressed and Withdrawn subscales. For the current sample, the Anxious/Depressed problems subscale maintains reasonable reliability at 3 years old (α =.70), 5 years old (α = .62), and 9 years old (α = .78). For the current sample, the Withdrawn subscale maintains reasonable reliability at 3 years old (α = .74), 5 years old (α = .60), and 9 years old (α = .69). The overall Internalizing problems scale maintains reasonable reliability using 25 items at 3 years old (α = .82), 20 items at 5 years old (α = .72), and 21 items at 9 years old (α = .84).
Child negative emotionality was assessed using maternal reports when children were between 1-1.5 years old using the Emotionality and Shyness Temperament Survey (EAST) (Buss & Plomin, 1984). When maternal report was not available, the biological father’s report was used. We used the three items (questions b, d, and e) from the five-item Emotionality subscale that were used in the original FFCW study (α = .60). Mothers rated child emotionality using a five-point Likert scale from 1 (not at all like my child) to 5 (very much like my child). Mothers were asked questions like, “Child reacts intensely when upset,” “Child often fusses and cries,” “Child gets upset easily.” The mean ranged from 2.16 – 2.71 (SD =1.38 – 1.64) across the three items. A latent variable was created with these three items to reflect the underlying and unobserved trait of negative emotionality or phenotypic expression of the trait. Higher values indicate higher levels of negative emotionality.
Covariates for the present study included: maternal race, child’s gender, and the Material Hardship Scale (MHS). Maternal race was dummy coded into two groups: non-Hispanic and Black, and non-Hispanic and White, with Hispanic and other racial and ethnic groups treated as one reference group. Children's gender was included as a covariate reported by the mother shortly after birth (0 = boys, 1 = girl). To account for the effect of poverty on responsivity and children’s internalizing problems, we used the Material Hardship scale, which has been used previously in large, nationally representative studies (Survey on Income and Program Participation, 1999). The Material Hardship scale (α = .64) assesses a family’s ability to meet basic needs. The primary caregiver answered twelve questions as either yes or no (1 = no; 2 = yes). like, “In the past year, did you receive food or free meals because there was not enough money?” A mean score was calculated and included as a covariate (M = 1.83, SD = .17).
Data Analysis
Data preparation was done using SPSS V. 25, and All hypothesis testing was done using Mplus V8. Maximum likelihood estimation with robust standard errors (MLR) was used as it is robust to the multivariate non-normality inherent to latent variables. We used a measurement model to define latent variables, assess the psychometric properties of constructs, and assess correlations between constructs. We created four latent variables: negative emotionality at 1 year old and internalizing problems at ages 3, 5, and 9 years old. The residuals for each of the two indicators for internalizing problems, the Anxious/Depressed and Withdrawn subscales, were correlated within each age. Factor loadings for both indicators were set equal to one another within each time point; thus, the structural model was just identified. Concurrent residuals between constructs were correlated. Residual variances of the latent variables were scaled to one.
A path-analytic structural model with latent and observed variables tested the bidirectional associations specified in hypothesis 1 (see Figure 1). The structural model estimated autoregressive effects and cross-lagged effects while controlling for the other. In the structural model, we treated negative emotionality as a covariate. Overall model fit was assessed using comparative fit index (CFI) and the Root Mean Square Error of Approximation (RMSEA) with adequate fit for CFI considered greater than .95 and RMSEA below .06 considered adequate levels of misfit (West et al., 2012). A latent-interaction model tested hypothesis 2: moderation by negative emotionality for the hypothesized bidirectional associations. In the latent-interaction model, we defined the four previously described latent variables and four latent interaction terms, the interaction between 1.) negative emotionality and internalizing problems at 3 years old, 2.) negative emotionality and internalizing problems at 5 years old, 3.) negative emotionality and responsivity at 3 years old, 4.) negative emotionality and responsivity at 5 years old. The latent interaction terms were included to test for moderation by children’s negative emotionality for the cross-lagged paths. We followed Klein and Moosebrugger’s (2000) example of maximum likelihood estimation for latent-moderated structural modeling using a two-step approach that can be implemented in Mplus. We then compared the null model without the interactions (Model 1), and second model with the hypothesized interactions (Model 2), using a log likelihood test for significant loss in model fit (Maslowsky et al., 2016). The Mplus STANDARDIZE command was used to standardize beta coefficients which provides a standardized effect size for the interaction effects. Graphing of the standardized regression coefficients were used to interpret significant interactions (Dawson, 2014).