winbugs notes

1. in winbugs script, the file path cannot be longer than 108 characters; otherwise, there will be an incompatible copy error message
2. from R2winbugs, even seed is set to NULL (the default), chains from winbugs are identical given identical initial values and data sets. Guess winbugs figure out a fixed seed based upon these input.
3.  a good reference for specify covariance matrix prior for multivariate normal. However, there are comments that promote scaled inverse Wishart instead of inverse Wishart as the prior for precision. An illustration against inverse Wishart can be found here. 
4. scaled inversed wishart recommended by Gelman. Need set df = K+ 1 to make individual correlation coefficient uniform on [-1,1], where K is the dimension of the Cov matrix (although df=K is the minimum allowed df). A winbugs implementation can be found here: it actually scales the y variable, which indirectly scale the covariance matrix.
5. From winbugs manual: "Note that WinBUGS simulates each node in turn: this can make convergence very slow and the program very inefficient for models with strongly related parameters, such as hidden-Markov and other time series structures." Thus, strongly related parameters should be put into one multivariate node.
6. For parameters for repeated measures, reference coding of effect (treating a visit as a reference, and other visits as deviation from the reference) seems to perform worse than cell mean effect coding (just a separate parameter for each visit). This is probably because the later gives rise to a orthogonal design matrix. 
7. to convert a object (eg. wb) returned from function 'bugs' in R2WinBUGS (codaPkg=FALSE ), we can use coda1 <- as.mcmc.list(wb). See the documentation of 'codaPkg' for more details.
8. openbugs is not necessary better than winbugs in perfomance and scalibility. 
1. openbugs has no automatic graphic output, which is good when we are not worried about convergence and when we run it through many cycles (eg. in simulations)
2. openbugs runs in unix.
9. Missing values are treated as a stochastic node and effectively 'imputed' by winbugs. However, such imputed data will not inflate precision (since they are not treated as data, but parameters by winbugs). It is possible to take this imputation procedure out: run multiple imputation and feed full data into winbugs to mimic the internal imputation procedure implemented by winbugs.We will have additional data from imputation (hence increase precision), but the variation among different imputed datasets, which will be fed one by one and posterior distributions merged, may compensate for the increased precision.
10. A good resource for DIC

mmrm model example

proc mixed data=dsin;
class treat;
model change = baseline treat genotype treat*genotype;
estimate 'additive effect' genotype 1 /e;
run;

proc mixed data=dsin;

class visit;

model change = baseline visit genotype visit*genotype;

estimate 'additive effect at the last visit' genotype 1 genotype*visit 0 0 0 0 1/e;

run;