Statistical Analyses

To prevent the loss of information, improve the statistical power to detect associations in the data, and reduce bias in our effect estimates, we performed multiple imputation using chained questions (MICE package in R, version 3.13.0) for missing data (40). Multiple imputation using chained questions is an iterative procedure for computing missing data. It operates by systematically calculating the missing values for each variable using a regression model containing the non-missing data from other specified variables available in the dataset. For our MICE procedure, we implemented this iterative process 25 times for each of the 20 imputed datasets we created. Each of the 20 datasets contained all ALSPAC participants with SMFQ scores available at age 18 (n = 3,263). We then performed our inference using the pooled estimates from those 20 imputed datasets (41).

All confounders and exposures were imputed for in the imputation model. Data on the child's depressive symptoms at age 18 as well as variables pertaining to the child's friendships throughout adolescence, a strong predictor of social support levels (42), were included in the imputation procedure to provide additional information in an effort to yield more accurate estimates for missing covariate and exposure data.

We ran three linear regression models. Model 1 tested the association between levels of childhood emotional neglect and depressive symptom scores at 18. Model 2 added to Model 1 by including peer social support, allowing us to assess the effect of both emotional neglect and levels of peer social support on depressive symptoms at age 18. Using an interaction term between childhood neglect and the levels of social support, Model 3 added to Model 2 by assessing whether social support was an effect modifier of the relationship between childhood emotional neglect and depressive symptoms. To increase interpretability of the results, we standardized the SMFQ depression outcome variable with a mean of 0 and SD of 1, such that beta values for each of the independent variables in the model reflected standard deviation changes in the depressive symptoms. All models were adjusted for the covariates as described in the Measures (e.g., home ownership status, child sex, race).