library(multinma)
options(mc.cores = parallel::detectCores())
This vignette describes the analysis of 6 trials comparing transfusion of granulocytes (white blood cells) to control for preventing mortality in patients with neutropenia or neutrophil dysfunction (Stanworth et al. 2005; Turner et al. 2012). The data are available in this package as transfusion
:
head(transfusion)
#> studyc trtc r n
#> 1 Bow 1984 Transfusion 5 13
#> 2 Bow 1984 Control 4 11
#> 3 Herzig 1977 Transfusion 1 13
#> 4 Herzig 1977 Control 3 14
#> 5 Higby 1975 Transfusion 2 17
#> 6 Higby 1975 Control 14 19
Turner et al. (2012) previously used this dataset to demonstrate the application of informative priors for heterogeneity, an analysis which we recreate here.
We begin by setting up the network - here just a pairwise meta-analysis. We have arm-level count data giving the number of deaths (r
) out of the total (n
) in each arm, so we use the function set_agd_arm()
. We set “Control” as the reference treatment.
<- set_agd_arm(transfusion,
tr_net study = studyc,
trt = trtc,
r = r,
n = n,
trt_ref = "Control")
tr_net#> A network with 6 AgD studies (arm-based).
#>
#> ------------------------------------------------------- AgD studies (arm-based) ----
#> Study Treatment arms
#> Bow 1984 2: Control | Transfusion
#> Herzig 1977 2: Control | Transfusion
#> Higby 1975 2: Control | Transfusion
#> Scali 1978 2: Control | Transfusion
#> Vogler 1977 2: Control | Transfusion
#> Winston 1982a 2: Control | Transfusion
#>
#> Outcome type: count
#> ------------------------------------------------------------------------------------
#> Total number of treatments: 2
#> Total number of studies: 6
#> Reference treatment is: Control
#> Network is connected
We fit two random effects models, first with a non-informative prior for the heterogeneity, then using the informative prior described by Turner et al. (2012).
We fit a random effects model using the nma()
function with trt_effects = "random"
. We use \(\mathrm{N}(0, 100^2)\) prior distributions for the treatment effects \(d_k\) and study-specific intercepts \(\mu_j\), and a non-informative \(\textrm{half-N}(5^2)\) prior for the heterogeneity standard deviation \(\tau\). We can examine the range of parameter values implied by these prior distributions with the summary()
method:
summary(normal(scale = 100))
#> A Normal prior distribution: location = 0, scale = 100.
#> 50% of the prior density lies between -67.45 and 67.45.
#> 95% of the prior density lies between -196 and 196.
summary(half_normal(scale = 5))
#> A half-Normal prior distribution: location = 0, scale = 5.
#> 50% of the prior density lies between 0 and 3.37.
#> 95% of the prior density lies between 0 and 9.8.
Fitting the RE model
<- nma(tr_net,
tr_fit_RE_noninf trt_effects = "random",
prior_intercept = normal(scale = 100),
prior_trt = normal(scale = 100),
prior_het = half_normal(scale = 5))
Basic parameter summaries are given by the print()
method:
tr_fit_RE_noninf#> A random effects NMA with a binomial likelihood (logit link).
#> Inference for Stan model: binomial_1par.
#> 4 chains, each with iter=2000; warmup=1000; thin=1;
#> post-warmup draws per chain=1000, total post-warmup draws=4000.
#>
#> mean se_mean sd 2.5% 25% 50% 75% 97.5% n_eff Rhat
#> d[Transfusion] -1.13 0.03 0.96 -3.23 -1.64 -1.10 -0.55 0.73 1195 1
#> lp__ -134.26 0.09 2.96 -141.17 -136.06 -133.91 -132.11 -129.30 1103 1
#> tau 1.84 0.03 1.03 0.54 1.15 1.61 2.27 4.55 1348 1
#>
#> Samples were drawn using NUTS(diag_e) at Thu Feb 24 09:05:43 2022.
#> For each parameter, n_eff is a crude measure of effective sample size,
#> and Rhat is the potential scale reduction factor on split chains (at
#> convergence, Rhat=1).
By default, summaries of the study-specific intercepts \(\mu_j\) and study-specific relative effects \(\delta_{jk}\) are hidden, but could be examined by changing the pars
argument:
# Not run
print(tr_fit_RE_noninf, pars = c("d", "mu", "delta"))
The prior and posterior distributions can be compared visually using the plot_prior_posterior()
function:
plot_prior_posterior(tr_fit_RE_noninf, prior = "het")
The posterior distribution for the heterogeneity variance \(\tau^2\) is summarised by
<- as.array(tr_fit_RE_noninf, pars = "tau")
noninf_tau <- noninf_tau^2
noninf_tausq names(noninf_tausq) <- "tausq"
summary(noninf_tausq)
#> mean sd 2.5% 25% 50% 75% 97.5% Bulk_ESS Tail_ESS Rhat
#> tausq 4.47 6.2 0.29 1.33 2.59 5.16 20.66 1368 1765 1
Keeping the rest of the model setup the same, we now use an informative \(\textrm{log-N}(-3.93, 1.51^2)\) prior for the heterogeneity variance \(\tau^2\). We can examine the range of parameter values implied by this prior distribution with the summary()
method:
summary(log_normal(-3.93, 1.51))
#> A log-Normal prior distribution: location = -3.93, scale = 1.51.
#> 50% of the prior density lies between 0.01 and 0.05.
#> 95% of the prior density lies between 0 and 0.38.
Fitting the RE model, we specify the log_normal
prior distribution in the prior_het
argument, and set prior_het_type = "var"
to indicate that this prior distribution is on the variance scale (instead of the standard deviation, the default).
<- nma(tr_net,
tr_fit_RE_inf trt_effects = "random",
prior_intercept = normal(scale = 100),
prior_trt = normal(scale = 100),
prior_het = log_normal(-3.93, 1.51),
prior_het_type = "var")
Basic parameter summaries are given by the print()
method:
tr_fit_RE_inf#> A random effects NMA with a binomial likelihood (logit link).
#> Inference for Stan model: binomial_1par.
#> 4 chains, each with iter=2000; warmup=1000; thin=1;
#> post-warmup draws per chain=1000, total post-warmup draws=4000.
#>
#> mean se_mean sd 2.5% 25% 50% 75% 97.5% n_eff Rhat
#> d[Transfusion] -0.77 0.01 0.44 -1.72 -1.02 -0.73 -0.47 0.01 2405 1
#> lp__ -140.27 0.07 2.75 -146.58 -142.00 -139.95 -138.26 -135.85 1377 1
#> tau 0.47 0.01 0.35 0.05 0.20 0.40 0.67 1.32 1605 1
#>
#> Samples were drawn using NUTS(diag_e) at Thu Feb 24 09:05:55 2022.
#> For each parameter, n_eff is a crude measure of effective sample size,
#> and Rhat is the potential scale reduction factor on split chains (at
#> convergence, Rhat=1).
By default, summaries of the study-specific intercepts \(\mu_j\) and study-specific relative effects \(\delta_{jk}\) are hidden, but could be examined by changing the pars
argument:
# Not run
print(tr_fit_RE_inf, pars = c("d", "mu", "delta"))
The prior and posterior distributions can be compared visually using the plot_prior_posterior()
function:
plot_prior_posterior(tr_fit_RE_inf, prior = "het")
Note: The heterogeneity variance \(\tau^2\) is plotted here since the prior was specified on \(\tau^2\).
The posterior distribution for the heterogeneity variance \(\tau^2\) is summarised by
<- as.array(tr_fit_RE_inf, pars = "tau")
inf_tau <- inf_tau^2
inf_tausq names(inf_tausq) <- "tausq"
summary(inf_tausq)
#> mean sd 2.5% 25% 50% 75% 97.5% Bulk_ESS Tail_ESS Rhat
#> tausq 0.34 0.5 0 0.04 0.16 0.45 1.73 1490 3055 1