Skip to contents

Rankings use a multivariate normal approximation to the fitted coefficient distribution. They describe uncertainty in average effects, not new-study outcomes. Rank 1 is best under the specified direction. Interpret rankings alongside effect estimates and intervals, not probability of best alone.

Usage

network_rankings(
  object,
  benefit_direction,
  strata = NULL,
  nsim = 10000L,
  seed = 1L
)

# S3 method for class 'nma_ranking_table'
print(
  x,
  ...,
  digits = getOption("digits"),
  detailed = FALSE,
  row.names = FALSE
)

# S3 method for class 'nma_ranking_table'
summary(object, ..., digits = getOption("digits"), detailed = FALSE)

Arguments

object

A fitted network model; for the ranking-table summary method, the rankings data frame returned by network_rankings().

benefit_direction

Whether higher or lower effects are preferable.

strata

Named outcome/time selection; exactly one stratum is required.

nsim

Number of draws, at least 100.

seed

Optional local seed; the caller's random-number state is restored.

x

A ranking table from network_rankings().

...

Additional arguments passed to print.data.frame().

digits

Decimal places, from 1 to 22. P-values below 0.0001 print as <0.0001; other p-values use at least four decimal places. Stored values retain full precision.

detailed

Display the duplicate output name sucr_a as well as sucra.

row.names

Display row names (default FALSE).

Value

Ranking summaries, a rank-probability matrix, and direction metadata.

Details

Joint Gaussian draws. For the selected stratum, form basic effects relative to the reference at moderator zero. If their prediction design is \(X_b\), use \(\hat b=X_b\hat\beta\) and the full covariance \(V_b=X_b\operatorname{Var}(\hat\beta)X_b^T\). Draw \(b^{(s)}\sim N(\hat b,V_b)\) and set the reference-relative effect to zero. The zero reference is a coordinate choice, not an error-free absolute treatment effect. Draws use an eigen-based covariance square root; negative eigenvalues are truncated to zero. Treatment effects are not sampled independently.

Covariance and conditioning. Joint fits use object$fit$vcov, including CR1 covariance when selected. Rank sampling still uses a multivariate normal approximation for CR1 fits, not a multivariate t distribution. Heterogeneity remains at its fitted value; draws do not re-estimate variance components or propagate their estimation uncertainty. Multiplicity-adjusted treatment p-values do not change this ranking distribution.

Direction and summaries. Rank 1 is best: sort ascending for lower and descending for higher, using average ranks for ties. mean_rank and median_rank summarize draws. p_best and the rank_probability matrix are on the 0–1 scale. sucra (also sucr_a) is on the 0–100 percentage scale: \(100\{K-\overline R_i\}/(K-1)\), where \(K\) is the treatment count. sucr_a is an duplicate output name, hidden by default when printing the ranking table. Use detailed=TRUE to display it. Each row of rank_probability gives the probabilities of ranks 1 through K for one treatment; its first column equals that treatment's p_best. Rows sum to one for continuous draws without ties. These probabilities are comparable in meaning to the probabilities printed by a netmeta rankogram. sucra is simulated, not an analytic P-score; repeated seeds/draw counts can assess Monte Carlo stability. seed is local and restores the caller's RNG state; seed=NULL uses and advances that state.

Why preserve covariance? Contrast uncertainty is \(\operatorname{Var}(d_i-d_j)=V_{ii}+V_{jj}-2V_{ij}\). Independent treatment draws are equivalent only when they preserve the relevant contrast distribution. For a continuous Gaussian distribution and lower-is-better effects, fractional expected SUCRA is $$\frac{1}{K-1}\sum_{j\ne i} \Phi\!\left(\frac{\hat d_j-\hat d_i}{\sqrt{\operatorname{Var}(d_i-d_j)}}\right).$$ This equals the corresponding Gaussian P-score computed from the same contrast means and SEs; finite simulated SUCRA has Monte Carlo error. Probability of being best and full rank distributions require a compatible joint distribution. Interpret rankings together with effect sizes and intervals; high SUCRA does not demonstrate a clinically important advantage or establish transitivity.

References

Rücker, G., and Schwarzer, G. (2015). Ranking treatments in frequentist network meta-analysis works without resampling methods. BMC Medical Research Methodology, 15, 58. doi:10.1186/s12874-015-0060-8 .

Salanti, G., Ades, A. E., and Ioannidis, J. P. A. (2011). Graphical methods and numerical summaries for presenting results from multiple-treatment meta-analysis: an overview and tutorial. Journal of Clinical Epidemiology, 64, 163–171. doi:10.1016/j.jclinepi.2010.03.016 .

Examples

dat <- data.frame(study = 1:6, a = rep(c("A", "A", "B"), 2),
                  b = rep(c("B", "C", "C"), 2),
                  yi = c(.2, .5, .3, .3, .6, .4), vi = .04)
fit <- network_meta(dat, "study", "a", "b", "yi", "vi")
network_rankings(fit, benefit_direction = "higher", nsim = 500)$rankings
#>  treatment mean_effect std_error mean_rank median_rank      sucra    p_best
#>          C   0.5666667 0.1154701 1.0020000   1.0000000 99.9000000 0.9980000
#>          B   0.2333333 0.1154701 2.0320000   2.0000000 48.4000000 0.0020000
#>          A   0.0000000 0.0000000 2.9660000   3.0000000  1.7000000 0.0000000