I'm looking at underlying risk as a predictor of treatment effect in a
meta-analysis, as described by Thompson, Smith and Sharp, Stat.Med.
1997;16:2741-2758. As modified (correctly, I hope) for WinBugs, the
code is:
model
{
for( i in 1 : N ) {
rc[i] ~ dbin(pc[i],nc[i])
rt[i] ~ dbin(pt[i],nt[i])
logit(pc[i]) <- mu[i]
logit(pt[i]) <- mu[i]+delta[i]
delta[i] <- deldash[i] + (betac*(mu[i]-mean(mu[])))
deldash[i] <- delt+(sigdelt*z[i])
z[i] ~ dnorm(0,1)
mu[i] ~ dnorm(0,.1)
}
delt ~ dnorm(0,0.1)
sigdelt ~ dnorm(0,0.1)I(0,)
betac ~ dnorm(0,.0001)
}
where n and r are the totals and events in control (c) and treatment
(t) groups.
(As an aside, when I specify inits as
list(delt = 1, sigdelt = 1, betac = 1)
I get a message that the model contains uninitialized node: for which
other nodes do I need to specify initial values?)
I would like to change this model to look at the same question for risk
ratio and risk difference measures of treatment effect. This can be
done, I think, by changing the lines
logit(pc[i]) <- mu[i]
logit(pt[i]) <- mu[i]+delta[i]
to
log(pc[i]) <- mu[i]
log(pt[i]) <- mu[i]+delta[i]
for risk ratio, and
pc[i] <- mu[i]
pt[i] <- mu[i]+delta[i]
for risk difference.
However, when I do this, I get estimation errors for the risk
difference, and the model won't even compile for risk difference.
I assume that this is because for treatment effect measures other than
the odds ratio we have to constrain the procedure to ensure that only
feasible values of delta[i] are estimated. However I'm not sure how to
do this.
I'd also welcome guidance on whether the priors in the program above
are appropriate (my personal priors are completely vague).
Can anyone help?
Thanks
Jonathan Sterne
----------------------
Jonathan Sterne
Department of Social Medicine
University of Bristol
Canynge Hall, Whiteladies Road
Bristol BS8 2PR
Phone 0117 928 7396
Fax 0117 928 7325
E-mail [log in to unmask]
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|