We included clustering by school (random intercept) and stratifying by region, as suggested by Chen and Chantala (Chen and Chantala 2014), to adequately account for the sampling design.
In the initial 'null' three-level model of random intercepts, the variation in the mental health score was modelled by random intercept terms for wards and unitary authorities and an individual random error term.
Here the variation in the mental health score was modelled by random intercept terms for households, LSOAs and unitary authorities, and a random error term for individuals.
We began by assuming a random intercept and selecting the degrees of freedom for the fixed spline terms using the AIC and BIC.
In each multilevel model, we will allow the outcome to vary by geographic area (random intercept) and, in the all children model, the Aboriginal to non-Aboriginal ratio to vary (random slope), enabling us to identify areas that are performing better or worse in terms of early childhood developmental outcomes, and areas with greater or lesser inequality in these outcomes.
The contrast-only GLMM included contrast as the only fixed effect (in addition to the intercept), and by-item random intercepts and slopes along with by-subject intercepts.
In each model, we controlled for pseudoreplicaton by adding nest as a random intercept and day of incubation (quadratic) as a random slope.
To do so, we incorporated random effects by rat (random intercepts and slopes) into linear models using PROC MIXED.
Given the small sample size, we endeavoured to control for Type I error by fitting random intercepts and slopes for time.
Variability in GDM between hospitals within states (assessed by random intercepts for hospitals) contributed to 14.7%.
