Data Analysis
Analyses were conducted in Mplus 8.8 (Muthén & Muthén, 2017) using Mplus Automation (Hallquist & Wiley, 2018). Mplus’ MLR estimator was used with full information maximum likelihood of handling missing data for unconditional models. The data analyses were conducted in four steps, following guidelines by Ryoo and colleagues (2018). First, Confirmatory Factor Analyses were conducted to examine the properties of maternal mental health assessments in one model (i.e., stress, anxiety, and depression; Appendix A). CFA model fit was assessed using absolute and relative indices. A Root Mean Square Error of Approximation (RMSEA; Steiger & Lind, 1980) of less than.05 to .08, a Tucker-Lewis Index (TLI; 1973) of above .95 and a Comparative Fit Index (CFI; Bender, 1990) above .95 was considered evidence of adequate model fit (see Appendix A for more information).
Then Latent Profile Analyses (LPA; 2-5 models) were estimated separately at each timepoint to investigate configural measurement invariance LPA enumeration was completed using statistical criteria such as Bayesian Information Criteria (BIC), sample-size adjusted BIC (aBIC), bootstrapped likelihood ratio test (BLRT), adjusted Lo-Mendell-Rubin likelihood ratio test (aLMR), standardized entropy, and successful model estimation; substantive considerations such as interpretability and number of students in each latent profile were also considered (Wang & Wang, 2020). Third, unconditional Latent Transition Analyses (LTA) were conducted, testing measurement invariance across time points. Then, conditional LTAs were conducted, incorporating the main effects. Further, longitudinal LPA with formal tests of measurement invariance using corrected likelihood ratio tests comparing the constrained and unconstrained models; BIC was also used to compare these two sets of models (Wang & Wang, 2020). Finally, to avoid local maximum, all LTA start values were increased to 10,000, 2000, and 400 for the longitudinal models (Gillet et al., 2017). Model specification of the LPA and LTA models followed traditional LPA modelling with the assumption of local independence (i.e., a diagonal residual covariance matrix) and item variances constrained equal across profiles.
Finally, we tested differences in family and child characteristics (e.g., children’s temperament and mental health measured) across our different profiles within the wave. Due to limited sample sizes within profiles, we did this using a series of ANOVAs. Post-hoc analyses were conducted using Tukey’s B or Dunnett’s C (based on homogeneity of variance) and chi-square tests. In cases where the homogeneity of variance was met, to account for the unequal sample sizes across the maternal mental health profiles, the harmonic mean was used to measure the central tendency for the ANOVAs (Bortz, 2005; Rankin, 1974). Parenting information was only collected prior to the pandemic and is therefore only included in the comparison across profiles in our earlier wave.