Repeated-Effects Network Meta-analysis

network_meta_repeated() is an independent arm-level continuous-outcome NMA function. It leaves network_meta() unchanged. It internally derives treatment contrasts and their within-study covariance, including shared-arm covariance in multi-arm trials.

library(mars)
arm_data <- data.frame(
  study = rep(c("S1", "S2", "S3"), each = 3),
  treatment = rep(c("A", "B", "C"), 3),
  mean = c(10, 12, 13, 9, 11, 12, 11, 14, 15), sd = 2, n = 30
)
fit <- network_meta_repeated(arm_data, "study", "treatment", "mean", "sd", "n")
fit$coefficients
#>                term estimate std_error ci_lower ci_upper
#> B::common B::common 2.333333 0.2981424 1.748985 2.917682
#> C::common C::common 3.333333 0.2981424 2.748985 3.917682

Outcomes and times

When outcomes or times are repeated within treatment arms, supply their identifiers. The function constructs arm-mean covariance using the selected outcome and time correlation assumptions. These are assumptions, retained in fit$assumptions, not correlations inferred from marginal standard deviations.

repeated_data <- expand.grid(
  study = c("S1", "S2", "S3"), treatment = c("A", "B"),
  outcome = c("pain", "function"), time = c(1, 2), stringsAsFactors = FALSE
)
repeated_data$mean <- with(repeated_data, 10 + 2 * (treatment == "B") + time)
repeated_data$sd <- 2
repeated_data$n <- 25

repeated_fit <- network_meta_repeated(
  repeated_data, "study", "treatment", "mean", "sd", "n",
  outcome = "outcome", time = "time", treatment_effects = "by_outcome_time",
  outcome_correlation = "exchangeable", rho_outcome = .5,
  time_correlation = "ar1", rho_time = .7
)
repeated_fit$coefficients
#>                          term estimate std_error ci_lower ci_upper
#> B::function::1 B::function::1        2 0.3265986 1.359878 2.640122
#> B::function::2 B::function::2        2 0.3265986 1.359878 2.640122
#> B::pain::1         B::pain::1        2 0.3265986 1.359878 2.640122
#> B::pain::2         B::pain::2        2 0.3265986 1.359878 2.640122
repeated_fit$assumptions
#> $outcome_correlation
#> [1] "exchangeable"
#> 
#> $time_correlation
#> [1] "ar1"
#> 
#> $rho_outcome
#> [1] 0.5
#> 
#> $rho_time
#> [1] 0.7
#> 
#> $arms_independent
#> [1] TRUE

Benchmark

For a common-effect mean-difference model, the script inst/benchmarks/network-meta-repeated-benchmark.R compares the MARS coefficients with netmeta when that optional package is installed. The MARS function additionally retains the internally constructed covariance blocks, which are needed for repeated outcomes and times.

mars authors

2026-07-17

Outcomes and times

Benchmark