Each study should have a single reference/baseline treatment,
against which relative effects in the other arm(s) are given. For the
reference arm, include a data row with continuous outcome y
equal to
NA
. If a study has three or more arms (so two or more relative effects),
set the standard error se
for the reference arm data row equal to the
standard error of the mean outcome on the reference arm (this determines
the covariance of the relative effects, when expressed as differences in
mean outcomes between arms).
By default, trt_ref = NULL
and a network reference treatment will be chosen
that attempts to maximise computational efficiency and stability. If an
alternative reference treatment is chosen and the model runs slowly or has
low effective sample size (ESS) this may be the cause - try letting the
default reference treatment be used instead. Regardless of which treatment is
used as the network reference at the model fitting stage, results can be
transformed afterwards: see the trt_ref
argument of
relative_effects()
and predict.stan_nma()
.
The sample_size
argument is optional, but when specified:
Enables automatic centering of predictors (center = TRUE
) in nma()
when a regression model is given for a network combining IPD and AgD
Enables production of study-specific relative effects, rank probabilities,
etc. for studies in the network when a regression model is given
Nodes in plot.nma_data()
may be weighted by sample size