Skip to contents

Intended question: how much do study effects vary within the same treatment design, after allowing each design its own treatment means?

Usage

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

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

network_heterogeneity(object)

Arguments

x

A result from network_heterogeneity().

...

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 designs with positive independently testable residual df (default TRUE). FALSE displays every design.

detailed

Logical; show all diagnostic columns and additional returned components (default FALSE). Combine with estimable_only=FALSE for all rows.

object

A fitted coherent network model; for summary(), a result from network_heterogeneity().

Value

A list with overall (within-design Q, df, p-value and I2_percent), by_design (design, study/contrast counts, Q, residual df, p-value, I2_percent and reason), decomposition (total/within/between Q), fitted_heterogeneity (the input model's heterogeneity summary), question and assumption, with class nma_heterogeneity. Default printing shows overall Q and only independently testable designs, under the heading Within-design Q, using design, Q, df, p-value and I2 (%) columns. Print/summary accept digits, estimable_only, and detailed; filtering changes only the display. Shared moderator slopes can preclude independent design tests. Zero residual df has a missing p-value. Q contributions are additive; their p-values are not.

Details

The printout follows its research question with the null hypothesis of no residual within-design heterogeneity beyond sampling variation. The overall Q test applies across designs; each testable design row applies the same null within that design. I-squared is descriptive, with no separate test.

A design is the set of treatments observed in a study. Fit the augmented design-by-treatment mean by GLS using sampling covariance alone, preserving shared-arm dependence. The residual Q is within-design heterogeneity; between- design disagreement is excluded from this Q. This is the within-design part of network_q_decomposition(), not a sum of local incoherence tests.

Design-specific contributions use residual quadratic forms in each design. With independent design mean spaces their residual df and chi-squared p-values are reported. If mean parameters, such as moderator slopes, are shared across designs, the Q contributions remain additive but design-specific df and p-values are missing, with a reason. The overall residual test remains valid under its usual Gaussian known-sampling-covariance assumptions.

Overall and design-specific I2_percent use the Q-based descriptive measure \(100\max(0,1-df/Q)\). For positive independently testable df, Q less than or equal to df (including zero) gives zero. Zero or unavailable df gives NA, including design-specific df unavailable because moderator slopes are shared. I-squared describes excess residual variation relative to sampling variation after allowing design-specific means. It is not a fitted variance component, a direct/indirect incoherence measure, or an additive partition: design-specific I-squared values cannot be summed or averaged to obtain the overall within-design I-squared. Small numbers of studies limit its precision.

Q uses no fitted random-effect covariance, even for a random-effects input. fitted_heterogeneity separately reports the input model's variance summaries; it is not a new estimate under the design-specific mean model. For uncertainty in one common variance use network_heterogeneity_profile(). Heterogeneity and incoherence are distinct; small within-design Q p-values do not establish direct/indirect disagreement, and large values do not establish homogeneity.

References

Krahn, U., Binder, H., and König, 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)
f <- network_meta(d,"s","a","b","g","v")
network_heterogeneity(f)
#> Within-design heterogeneity
#> Question: How much do study effects vary within the same treatment design?
#> H0 (overall): No residual within-design heterogeneity beyond sampling variation.
#> Sampling covariance only.
#> 
#> Overall within-design Q
#>          Q df   p_value    I2 (%)
#>  0.0000000  6 1.0000000 0.0000000
#> 
#> Within-design Q
#> H0 (each design): No residual heterogeneity within the named design beyond sampling variation.
#>  design         Q df   p_value    I2 (%)
#>   A | B 0.0000000  2 1.0000000 0.0000000
#>   A | C 0.0000000  2 1.0000000 0.0000000
#>   B | C 0.0000000  2 1.0000000 0.0000000
#> 
#> I2 (%) is Q-based within-design residual heterogeneity; values are not additive.