
Independent direct and indirect evidence by node-splitting
network_nodesplit.RdEstimate a direct-minus-indirect treatment difference using separate study allocation, with explicit handling of multi-arm trials and nonestimable splits.
Arguments
- object
A fitted joint network model.
- method
conditionalholds fitted heterogeneity fixed;MLrefits heterogeneity jointly in each split model. Both use normal Wald inference.- comparisons
Optional data frame with
treatment_1andtreatment_2columns selecting unordered pairs. Selection includes all fitted mean strata; reversing a requested pair does not reverse the returned contrast direction.- level
Coverage of pointwise normal-approximation intervals for direct effects, indirect effects and their difference.
- adjust
none,holm, orbonferronifor the requested node-split pair/stratum family, retaining nonestimable hypotheses in its size. Output addsp_adjustedand anadjustmentattribute. Intervals are pointwise.
Value
One row per selected pair/mean stratum, including:
- network, network_se
Coherent-model estimate and model-based SE in the same treatment_2-minus-treatment_1 direction.
- prop
Direct evidence proportion: coherent-model variance divided by direct-only variance, both evaluated at the primary fitted covariance. It describes precision contribution, not the fraction of studies or patients. It remains on the primary covariance scale even for ML splits. Missing when direct precision is unavailable or no valid proportion can be formed.
- z
Signed normal Wald statistic: difference divided by its SE.
- direct, indirect, direct_se, indirect_se
Split mean effects and model-based standard errors.
- direct_lower, direct_upper, indirect_lower, indirect_upper
Pointwise normal confidence limits at
levelfor each evidence estimate. Direct limits remain available when the indirect route is not estimable.- direct_indirect_covariance
Covariance used in the difference variance.
- difference, se, lower, upper, p_value
Direct-minus-indirect estimate, its SE, normal interval and two-sided Wald p-value.
- direct_studies
Number of studies furnishing an estimable direct contrast.
- indirect_studies
Number of all retained indirect-side studies, including studies in other strata; not necessarily the number on a particular path.
- estimable, reason, method
Eligibility/status and calculation. A direct estimate may be available even when an independent split is not estimable.
Failure reasons include unavailable independent indirect evidence, aliased mean
parameters, nonidentifiable covariance or failed optimization. Missing split
results are not nonsignificant results. The returned allocation and inference
attributes describe study allocation and inference assumptions.
Details
For each pair and mean stratum, construct one direct contrast per eligible study from its observed contrast basis using sampling-covariance-aware GLS weights. The target contrast must be estimable within that study/stratum. Thus a three-arm study contributes direct B/C evidence even if its supplied rows are A/B and A/C. Moderators must be constant within each direct study/stratum; reported effects are evaluated at moderator zero, as in the original model.
Every study containing both target treatments in the target stratum is excluded from the indirect side, including all its other outcomes or times. Remaining studies must identify an indirect treatment contrast. This is conservative whole-study allocation: it can discard otherwise useful evidence in other strata, but prevents the same multi-arm study from providing both direct and indirect evidence. A network containing only A/B/C trials supplies no independent indirect route for those pairs after allocation.
A linear transformation T creates the direct study contrasts and retains the
indirect observations. Effects, sampling covariance, random-effect design and
covariance kernels transform as \(Ty\), \(TST^T\), and their corresponding
matrix products. A direct-evidence mean indicator is added to the transformed
network model. The conditional method uses transformed fitted total covariance;
ML estimates heterogeneity jointly in that split model unless originally fixed.
The latter uses a shared covariance specification, rather than fitting unrelated
heterogeneity models separately on the two sides.
For a returned pair, all effects are treatment_2 minus treatment_1. Compute
\(\Delta=\hat d_{direct}-\hat d_{indirect}\) and
\(SE(\Delta)^2=SE(\hat d_{direct})^2+SE(\hat d_{indirect})^2-
2\operatorname{Cov}(\hat d_{direct},\hat d_{indirect})\).
Direct and indirect intervals are each estimate plus or minus
\(z_{(1+level)/2}\) times its own SE. Difference intervals are
\(\Delta\mathbin{\pm}z_{(1+level)/2}SE(\Delta)\); the two-sided
p-value is the standard normal Wald test of \(\Delta=0\). Even with ML, this
p-value is not a likelihood-ratio test. Covariance estimates are treated as fixed
for interval calculation, and primary CR1 inference is not propagated.
A positive difference means the direct numerical effect exceeds the indirect
one. Whether that favors a treatment depends on outcome and benefit direction.
For logarithmic ratios, exponentiating the difference gives the ratio of the
direct and indirect ratio estimates, not a pooled treatment effect. Inspect the
magnitude and interval on the substantive effect scale, not only the p-value.
An interval containing zero does not establish coherence. Pair/stratum tests
are not multiplicity adjusted unless adjust is supplied.
References
Dias, S., Welton, N. J., Caldwell, D. M., and Ades, A. E. (2010). Checking consistency in mixed treatment comparison meta-analysis. Statistics in Medicine, 29(7–8), 932–944. doi:10.1002/sim.3767 .
van Valkenhoef, G., Dias, S., Ades, A. E., and Welton, N. J. (2016). Automated generation of node-splitting models for assessment of inconsistency in network meta-analysis. Research Synthesis Methods, 7(1), 80–93. doi:10.1002/jrsm.1167 .
Examples
d <- data.frame(study=1:9,a=rep(c("A","A","B"),3),
b=rep(c("B","C","C"),3),effect=rep(c(.2,.6,.2),3),variance=.04)
fit <- network_meta(d,"study","a","b","effect","variance")
network_nodesplit(fit,comparisons=data.frame(treatment_1="A",treatment_2="B"))
#> treatment_1 treatment_2 network network_se prop direct direct_se
#> A B 0.2666667 0.0942809 0.6666667 0.2000000 0.1154701
#> indirect indirect_se direct_indirect_covariance difference se
#> 0.4000000 0.1632993 0.0000000 -0.2000000 0.2000000
#> lower upper p_value z direct_studies indirect_studies
#> -0.5919928 0.1919928 0.3173105 -1.0000000 3 6
#> estimable reason method direct_lower direct_upper indirect_lower
#> TRUE conditional -0.0263171 0.4263171 0.0799392
#> indirect_upper p_adjusted
#> 0.7200608 0.3173105
# A/B: direct .2, indirect .6 - .2 = .4, difference -.2.