制定网状荟萃分析方案
根据研究问题规划治疗节点、适用性与统计模型,并安排一致性检验、证据分级和报告。
完整提示词
<context>
Network meta-analysis (NMA) combines direct and indirect evidence to compare several treatments at once, including pairs never compared head to head. Its validity rests on transitivity: the trials for different comparisons must be similar enough in effect modifiers (severity, dose, background therapy, era, outcome definition) that the treatments could in principle have been randomised together. Inconsistency is the statistical sign of a transitivity problem, tested locally (node-splitting, loop-specific) and globally (design-by-treatment interaction), but tests are underpowered, so judgement about effect modifiers comes first. Experienced analysts define nodes carefully (lumping or splitting doses and classes), inspect network geometry (sparse or star-shaped networks rely heavily on indirect evidence), choose frequentist or Bayesian models with a shared heterogeneity variance and sensible priors, report league tables with credible or confidence intervals, and treat rankings (SUCRA, P-scores, probability of being best) as secondary because they hide uncertainty and differences that are not important. Certainty is rated with CINeMA or the GRADE approach for NMA, and reporting follows PRISMA-NMA.
</context>
<task>
Plan a network meta-analysis for:
<question>
{{question}}
</question>
{{#evidence}}
<evidence>
{{evidence}}
</evidence>
{{/evidence}}
1. Decide whether NMA is appropriate: enough connected evidence, a decision that needs a comparison of several treatments, plausible transitivity. If not, recommend pairwise meta-analyses or a narrative comparison.
2. Define the nodes (treatments, doses, classes), with the lumping and splitting rules and reasons, and describe the expected network geometry; flag disconnected parts and comparisons informed only indirectly.
3. List the likely effect modifiers and how transitivity will be assessed: tabulate them by comparison, decide on exclusion or adjustment (network meta-regression) in advance.
4. Choose the model: effect measure, frequentist or Bayesian, random effects with a common heterogeneity variance, prior choices if Bayesian, handling of multi-arm trials, and software options to confirm.
5. Plan inconsistency checks (global and local), what you will do if inconsistency is found, and heterogeneity reporting with prediction intervals.
6. Plan outputs and certainty: network plot, league table, forest plot against a common comparator, rankings reported with their uncertainty and with a caution, and CINeMA domains (within-study bias, reporting bias, indirectness, imprecision, heterogeneity, incoherence).
7. Plan sensitivity analyses and reporting (PRISMA-NMA, protocol registration).
</task>
<constraints>
- Do not invent effect estimates or rankings; with no results supplied, describe the outputs only.
- Never present a ranking as the main result or say a treatment is best without the effect sizes and certainty behind it.
- Treat a non-significant inconsistency test as weak reassurance, not proof of consistency.
- Advise involving a statistician experienced in NMA when the user's team has none.
</constraints>
<output_format>
## Is NMA appropriate
Recommendation and reasons.
## Network and nodes
Node definitions and a text sketch of the network.
## Transitivity
A table: effect modifier | why it matters | how assessed | action if imbalanced.
## Model and estimation
Bullets.
## Consistency and heterogeneity
Checks and pre-specified responses.
## Ranking and certainty
Output list, ranking caveats and CINeMA plan.
## Reporting
Sensitivity analyses, PRISMA-NMA and registration.
</output_format>可替换内容
- {{question}}
- {{#evidence}}
- {{evidence}}
- {{/evidence}}
分类与标签
学习教育、循证研究、荟萃分析、研究方法