dmphillippo
commented
6 months ago Hi @bct0022. The model you have fitted is not recommended: this is not regressing the model parameter $\muj$ on the treatment effects, but is instead regressing the observed baseline risk $\hat{p}{j1}$. This is the approach warned against in TSD3 (p42), and discussed by Thompson et al. (1998).
To do this correctly requires a modification to the package. As it happens, this is in progress in PR #36, and hopefully will be available soon.
yipeng18
Hi Phillippo,
We wanted to use meta-regression with baseline risk adjustment (NICE TSD3, section 4.4, page 45) and replicate example 6 (NICE TSD3, page 74) with multinma package. To do this, we took the logit of placebo response rate for each trial (mu i) and centered by the logit of placebo response rate of all placebo arms across trials (mu bar). The centered value was plugged into the NMR as an effect modifier
ra_RE_fit <- nma(ra_net, trt_effects = "random", regression = ~.trt:centered_pbo_rate, likelihood = "binomial", class_interactions = "common", prior_intercept = normal(scale = 100), prior_trt = normal(scale = 100), prior_reg= normal(scale = 10), prior_het = half_normal(scale = 5), prior_het_type = "sd", warmup = 1000, iter = 3000, chains = 3, seed = 12345, adapt_delta = 0.99)
The outputs of RE model is a bit different from the results from TSD. I wonder if there is any way in multinma where we can specify baseline risk adjustment and whether our approach is appropriate. If not, do you have any plan to include it in the next package version?
Thanks,