Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
From | Abdelouahid Tajar <a_tajar@hotmail.co.uk> |
To | statalist <statalist@hsphsun2.harvard.edu> |
Subject | RE: st: Mixed model for longitudinal data: Time discrete or continuous? |
Date | Mon, 4 Jun 2012 14:48:47 +0100 |
Thanks Nick, I have a pilot study of 39 patients, 18 in group A (treatment 1) and 21 in group B (treatment 2). I have 6 time points baseline then a response after each four weeks the time variable is coded as: 5 ,9,13,17,21 and 25. I have a binary outcome (yes=1 and No=0) I used gllamm in stata because it allows computing predicted probabilities taking into account the random effect using gllapred, I have considered 3 types of models: Model1 ( Time discrete) the covariates: are group+ 5 dummy variables for time (1 time value was included as reference) and 5 interactions of group and the 5 time-dummies, in total 11 parameters. Model2 ( time as continuous) 3 parameters, group time and interaction group*time (time centered at t=13) Model 3: adds to model 2, a square term for time, time^2 and an interaction between group*time^2 which gives 5 parameters in model 3. The models criteria are as follows: Model1: AIC=211.48 and condition number =7.49 Model 2:AIC=204.78 and condition number =79.85 Model3: AIC=206.91 and condition number =673.20 According to gllamm book, the condition number is the square of the largest to the smallest eigenvalues of the Hessian matrix. Lower condition number are preferred for model identification.