In addition to calculating correlations among BPb variables to assess potential for collinearity, we also ran an artificial multiple regression with the full mixed-model variables to calculate the variance inflation factors (VIFs) (Hardin 1995) for the lead terms.
We fitted a multilevel generalised mixed model, adding the variables of hospital treatment quarters (hospital level variables) to the 17 case mix variables (patient level variables), to compare the case mix adjusted risk ratios between quarters of hospital treatment use and mortality risk in Sweden and the UK.
20 Therefore, we estimated linear time trends from 1998 to 2008 using a linear mixed model for continuous variables (SAS PROC MIXED) and a generalised linear mixed model for binary variables (PROC GLIMMIX, using the logit link function) in which we adjusted for age and gender.
To take into account the clustering of outcomes (twin pregnancies), we compared birth outcome measures using a linear mixed model for quantitative variables and a generalised linear model with a logit link function for qualitative variables.
The final generalised additive mixed model included 15 variables in addition to year and region indicators.
Bayesian mixed models and variable selection models were used and, based on parameter samples from the model posterior distributions, explained variances were further partitioned at the level of chromosomes and genome segments.
Patient-level characteristics were compared amongst the three models while adjusting for clustering within a practice using generalized estimating equations for dichotomous variables and mixed models for continuous variables.
Univariate comparisons between the two study groups with respect to outcome measures and covariates at baseline will be conducted using Chi-square tests for categorical variables, Kruskal-Wallis tests for ordinal variables, and linear mixed models for continuous variables.
