2.4. Data analysis

Statistical analysis was performed using R (R Core Team, 2022, version 4.2.1). Through visual exploration of boxplots, we manually checked for outliers in the dependent variables (PPVT and TROG-D raw scores). No outliers were identified and the different scores were approximately normally distributed. Descriptive statistics were generated for all variables for both measurement occasions (T₁ and T₂). Raw scores were used as dependent variables. For caregivers, five different SprachKoPF-scores were calculated (total-score, knowledge-score, linguistic-knowledge-score, practical-knowledge score, skills-score).

Hypotheses (a) and (b): To perform mean comparisons from T₁ to T₂, we first conducted paired t-tests for both language variables in the IG and for receptive vocabulary in the CG. For mean comparison of caregivers’ SprachKoPF-scores, we conducted the non-parametric Wilcoxon signed-rank test due to a small sample size (N = 15).

For main analysis, we estimated separate multilevel linear mixed-effects models predicting fixed and random effects on children’s language scores (T₁ and T₂) using the lmerTest package (Kuznetsova et al., 2020). Children with incomplete observations were excluded from the main analysis. Alpha-error probability was set to 5%, i.e., we considered significance at α < 0.05. All metric variables were standardized using their grand mean and standard deviation. Children’s characteristics (age, gender, and length of exposure at T₁) were used as covariates. For the SprachKoPF total-score, we calculated a mean score for each ECD group for T₁ and T₂ and assigned them to each participating child. For visualization of our results, and especially interpretation of cross-level interactions, we used estimated marginal means of fixed effects and created interaction plots using the emmeans-package (Lenth et al., 2022). To indicate the proportion of variance explained by random effects, intraclass correlation coefficients were calculated for all variables.

Hypotheses (c): Addressing our first and second research questions about children’s receptive vocabulary and grammar growth in relation to caregivers’ language support competencies in interaction with time, we created two models, i.e., regressing on PPVT- and TROG-D-scores (repeated measurement, level 1 within-child). In the two models, we considered the effect of measurement occasion (time, level 1), caregivers’ SpachKoPF total-score (level 2: between children) nested in participants and caregivers, gender (level 2), age (level 2) and length of exposure (level 2) and a cross-level interaction between time and SprachKoPF total-score (level 2).

Hypotheses (d): For our third research question about children’s receptive vocabulary growth compared to a control group, we regressed PPVT-scores on time, age, gender, length of exposure and group affiliation and the cross-level interaction with time (T₁ and T₂). For this model, we regressed children’s receptive vocabulary (repeated measurement, level 1) on measurement occasion (time, level 1), group affiliation (level 2), gender (level 2), age (level 2) and length of exposure (level 2) and a cross-level interaction between time and group affiliation (level 2).