Skip to contents

Intended question: how much between-design disagreement remains when each treatment design, in turn, is allowed its own treatment effects?

Usage

network_design_detachment(object)

# S3 method for class 'nma_design_detachment'
print(
  x,
  ...,
  digits = getOption("digits"),
  estimable_only = TRUE,
  detailed = FALSE
)

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

Arguments

object

A fitted joint coherent network model; for summary(), a result returned by network_design_detachment().

x

A result from network_design_detachment().

...

Additional arguments forwarded to print.data.frame(); reserved additional arguments for summary.

digits

Decimal places, an integer from 1 to 22; defaults to getOption("digits"). Stored precision is unchanged. P-values below 0.0001 print as <0.0001; other p-values use at least four decimal places.

estimable_only

Logical; show only detachable designs with positive remaining df (default TRUE). FALSE retains all designs in the display.

detailed

Logical; show all diagnostic columns and assumptions (default FALSE). Combine with estimable_only=FALSE for every design.

Value

A list of class nma_design_detachment with global (baseline Q, df, p-value), detached (design, remaining Q, df, p-value, added_df, design_df, detachable, estimable, and reason), question, and assumption. All designs are retained in the returned data. Print/summary support decimal places, estimable-only selection and detailed diagnostic fields.

Details

A design is the set of treatments observed in a study. For each design, add interactions between its indicator and the basic treatment-effect columns, retaining only columns that increase the mean-model rank. Studies and sampling covariance are retained; detachment is not deletion or leave-one-design-out refitting. Existing moderator slopes remain shared and mean strata retain their fitted structure.

Let X be the coherent mean matrix, A the full design-by-treatment mean matrix, and B the coherent mean augmented for the selected design. Using sampling covariance S alone, define Q(M) as the GLS residual quadratic form for M. The baseline between-design statistic is Q(X)-Q(A). After detachment it is Q(B)-Q(A), with df equal to rank(A)-rank(B). The reference is chi-squared. No random-effects covariance is used or re-estimated, even for random-effects input. This corresponds to the sampling-only Q.inc.detach assessment in netmeta::decomp.design() for equivalent inputs and mean models, rather than its random-effects global Higgins test.

The printout follows the question with the null hypothesis that no between- design disagreement remains after the selected design is allowed separate effects. The baseline test separately states its null of no between-design disagreement in the original network.

Each p-value tests the remaining between-design disagreement, not whether the detached design is individually inconsistent. These overlapping exploratory results are unadjusted. A change in p-value also reflects changing df; it does not establish that a design is erroneous or justify excluding its studies. Designs adding no independent parameters, or whose full design effects cannot be separated from the remaining network, have detachable=FALSE. They are retained for inspection but have no detachment p-value. Zero remaining df has no p-value. Default printing shows only detachable designs with a testable remaining statistic, ordered by increasing Q.

References

Krahn, U., Binder, H. and Koenig, J. (2013). A graphical tool for locating inconsistency in network meta-analyses. BMC Medical Research Methodology, 13, 35. doi:10.1186/1471-2288-13-35 .

Examples

d <- data.frame(s=1:9,a=rep(c("A","A","B"),3),b=rep(c("B","C","C"),3),
  g=rep(c(.2,.6,.2),3),v=.04)
fit <- network_meta_coherent(d,"s","a","b","g","v",diagnostics=FALSE)
network_design_detachment(fit)
#> Design detachment: remaining between-design Q
#> Question: How much between-design disagreement remains when each design is allowed its own treatment effects?
#> H0 (each detachment): No between-design disagreement remains after the selected design has separate effects.
#> Sampling covariance only; studies retained; remaining-disagreement p-values are unadjusted.
#> 
#> Baseline between-design test
#> H0 (baseline): No between-design disagreement in the original network.
#>          Q df   p_value
#>  1.0000000  1 0.3173105
#> 
#> Between-design Q after detaching each design
#> No detachable designs with a testable remaining between-design statistic.
#> 
#> Each p-value tests remaining disagreement, not the detached design.