The model assumes that the observations are conditionally independent at the lowest level given the higher level GP and practice random effects and the predictor variables.
30 Random effects Poisson regression models were estimated to relate the number of psychiatric admissions per practice to its QOF achievement, conditioning on potential confounding factors and a normally distributed GP practice random effect with zero mean and constant variance.
The use of practice random effects ensures that adjustments are limited to differences observed within practices and are not biased by the confounding of overall practice performance with between-practice differences in patient characteristics.
In all multilevel analyses, we adjusted for age, gender and educational level of the patient (all fixed effects) and GP-practices (random effect) and the patient group without comorbidity was used as reference group.
If the fixed effect for women (vs men) is a and the random effect for women by practice random effects has a variance of b, approximately 95% of practices will have women (vs men) coefficients between a-1.96×sqrt(b) and a+1.96×sqrt(b).
Secondary outcomes will be analysed using a mixed modelling strategy, accounting for between practice variability by defining practices as random effects.
For the primary endpoint a generalized multi-level model, that takes the randomized clusters (practices) as random effect into account, with antibiotic prescription rate at T2 as dependent and random group as independent variable, will be fitted to the data.
Hierarchical linear regression models with random-effects levels for practice, individuals within practices and repeated measurements per individual (slope and intercept) will be used to estimate the effects of the intervention on SF-36 and MLHFQ scores for the post-baseline time points.
To account for clustering of individuals within practices, we included practice level random effects.
However, there was a significant interaction between gender and deprivation status (accounting for gender as spline, and practices as random effects, p<0.001); the gradient in consultation rates across deprivation quintiles was apparent in women, but not men.
