Association of Neonatal Pain-Related Stress and Parent Interaction With Internalizing Behaviors Across 1.5, 3.0, 4.5, and 8.0 Years in Children Born Very Preterm

Key Points Question Does supportive parenting ameliorate the association between neonatal pain-related stress and child internalizing behaviors in children born very preterm? Findings In this cohort study of 186 children born very preterm, internalizing behaviors increased across ages 1.5, 3.0, 4.5, and 8.0 years, and more neonatal pain-related stress was associated with greater internalizing behaviors across ages. At 1.5 years, parenting stress was associated with more internalizing behaviors, whereas at age 3.0 years, a more supportive parenting environment was associated with fewer internalizing behaviors across development. Meaning These findings suggest that supportive parenting is associated with reduced child anxiety and depressive behaviors from toddlerhood through school-age in children born very preterm.


eAppendix 1. Supplemental Methods
Study Population. Of the n = 227 survivors, 21 participants (6 withdrawn, 5 moved away, 7 lost to follow-up) did not have CBCL follow-up data at any age. We excluded 6 infants with severe brain injury on neonatal ultrasound (cystic periventricular leukomalacia [n = 2], intraventricular hemorrhage grade 3 or 4 [n = 4]); at age 4.5 and/or 8 years 1 sensorineural hearing impairment, 1 non-ambulatory cerebral palsy, 2 cognitive impairment with IQ <70, 1 severe autism, 2 genetic abnormality (chromosome 15q11.2 duplication). Exclusion criterion were consistent with our work in a prior, independent cohort. 1 As well, one infant had incomplete (>50%) missing neonatal clinical data (moved hospital during NICU stay) and was therefore excluded.
Parent and Child Interactive Behaviors. At 1.5 and 3-years, the primary caregiver and child engaged in a 5-minute semi-structured teaching session. 1 During the session, the caregiver helped guide the child as they performed two tasks of varying difficulty. At 1.5 years, the easier and more familiar task consisted of the child stacking or nesting cups of varying sizes. In the second task, which was considered novel and more difficult, participants were directed to sort plastic pig and cow figurines into two separate bowls. At 3 years, the first task consisted of arranging a wooden puzzle of a family, while the more challenging task involved sorting plastic shapes by both colour and size. Parents were instructed to interact with their child typically as they would at home.
One primary and two reliability coders rated parent and child EA, blinded to all participant information. Interrater reliability assessed via intraclass correlation (ICC) on 20% of the sample was high: Parent Sensitivity, Structuring, Non-intrusiveness, and Non-hostility, and Child Responsiveness and Involvement at 18 months CA ICC's were 0.87, 0.86, 0.93, 0.84, 0.86, and 0.87 respectively, and at 3 years ICC's were 0.87, 0.85, 0.90, 0.83, 0.83, and 0.87.

eAppendix 2. Statistical Analysis
Hypothesis 2: Latent profile analysis. Gaussian mixture modelling estimate model parameters via maximum likelihood estimation. We used the Bayes Information Criteria (BIC), Integrated Complete-data Likelihood (ICL), classification uncertainty, along with Bootstrapped Likelihood Ratio Test Statistic (LRTS) to determine the number of classes that best fit the data. Smaller values of BIC indicate better fit. LRT tests whether an increase in profiles increases fit. We examined plots for uncertainty, favouring models with fewer points of large (> 0.25) uncertainty. Conclusions as to the best fitting model were drawn from all three fit indices, 2 theoretical models of Emotional Availability, 3 and prior empirical work examining interrelationships between parent behavior, parenting stress and child behavior. 1,4 Hypotheses 1 and 3: Multilevel models. Multilevel models examining the role of parent environment at 1.5 years used all data across 1.5, 3, 4.5, and 8 years. Subsequent models were run using behaviors from 3, 4.5 to 8 years when considering parenting behaviors and stress at age 3-years.
Multilevel models are ideal for analysis of repeated measures within observational study designs where data is missing at random. 5 For analysis with complete data, we fitted models using restricted maximum likelihood estimation which estimates parameters of analysis model by maximizing the observed-data likelihood. Model significance of effects was tested using the Satterthwaite approximations for degrees of freedom. These methods have been found to produce acceptable bias parameter estimates and Type 1 error rates even for smaller samples. [6][7][8] However, in multilevel models, missing data on the dependent variable is dealt with, while missing data on explanatory variables are not accounted for. 9 Using a joint modelling imputation approach, imputations were simultaneously generated from specified models and the clustered structure of the data was accounted for. This approach has been shown to produce less biased parameter estimates than other multiple imputation techniques particularly in the case of categorical and non-normally distributed longitudinal data collected at unevenly spaced time-points. 10 Child age at visit (time-dependant measure) was specified as a first level variable and participant-specific measures (e.g. gestational age, neonatal pain-related stress, parent behavior, parenting stress) as second level variables. First an intercept-only model was established by allowing the intercept to enter as a random effect in order to examine the amount of variance in internalizing behaviors that is accounted for within and between individuals and to ensure the data required multi-level modelling analysis. Next, we tested whether internalizing behaviors were predicted by child age by adding this variable to the model (Model 1). We then examined whether the effect of age on internalizing behaviors differed between participants, entering age of the child as a random effect. If the random effect of age did not improve model fit, it was removed from the model prior to entering second-level variables. Next, second-level variables were entered into the model. First, we added gestational age at birth to confirm established reports of degree of prematurity in relation to internalizing behaviors (Model 2). Next, to test Hypothesis 1, we added clinical factors associated with prematurity, neonatal pain-related stress and child sex to the models (Model 3). We entered the interaction terms of neonatal pain-related stress and age into the model, and explored whether this interaction was subsequently dependant on child sex, or gestational age. To test Hypothesis 3, parent environment (parent behavior and parent stress) at each age were included as predictors in separate models, and subsequent models tested whether parenting factors interacted with neonatal pain-related stress in relation to internalizing behaviors across age (2-way interaction of Parenting X neonatal pain-related stress and 3-way interaction of Parenting X neonatal pain-related stress X Child Age). Finally, we tested whether effects were independent of child interactive behaviors and maternal education level (index of socioeconomic status) when examining relationships between parent interactive behaviors and child internalizing behaviors.
Reported models met all assumptions of multilevel models.
© 2022 McLean MA et al. JAMA Network Open.