A prioritized agenda

What should change in practice now, what to build next, and what is worth the risk

The catalog lists problems. This page ranks them, in three tiers plus a list of work the field does not need. Tier 1 requires no new methodology: every item is something an analyst could do differently in the next analysis. Tiers 2 and 3 are research programs.

Priorities carried in the source material are weak evidence and are treated as such. Of the 46 rows in the source’s own master matrix, 30 are graded “Very high” and 15 “High”, which is under half a bit of information. Where a priority appears in this catalog it reflects three observable things instead: whether the problem blocks a decision-grade estimand, whether any method addresses it at all, and whether any software implements it.

Tier 1: change current practice now

These need no new theory. They are conventions the field could adopt in its next submission.

Declare one decision target population. Do not let data availability choose it implicitly. See estimands and target populations.
Separate conditional and marginal estimands. Do not combine a conditional regression coefficient with a published marginal effect without justifying the combination.
Prefer anchored analyses whenever a credible randomized path exists. Label unanchored methods as cross-study causal reconstructions, not routine adjustments. See the method landscape.
Require formal quantitative bias analysis for any decision-grade unanchored comparison. At minimum quantify omitted prognostic and modifier bias plus the most credible selection, measurement, or design threat. Otherwise label the result exploratory. See quantitative bias analysis.
Report support and influence, not effective sample size alone. An adjusted study that does not influence the decision contrast is not evidence for that contrast. See positivity, overlap and extrapolation.
Expose the bridge in a disconnected network. State exactly which parameter carries information across the gap. See disconnected networks.
Use target-standardized outcomes for non-collapsible and survival measures. Do not treat odds ratios and hazard ratios as population-invariant decision inputs.
Distinguish sampling, target-distribution, sensitivity, prior, and bridge uncertainty. An ordinary confidence or credible interval usually conditions on every major structural assumption being true.
Make sensitivity analysis decision-based, and joint where biases could interact. Report how much plausible bias changes the recommendation, and whether biases cancel or reinforce. One-at-a-time analyses can make a cancellation accident look like robustness.

The one-page test

A committee-facing summary should answer nine questions. If an analysis cannot, the problem is not the statistics.

What is being estimated? For which population? Which comparisons are supported by randomization? What connects any disconnected evidence? Which variables are assumed to remove population differences? Where is overlap weak? What is data-driven versus prior-driven or bridge-driven? Which residual biases were quantified, from what parameter sources? How much assumption violation changes the decision?

Tier 2: develop and validate next

Tractable programs where a first publishable contribution is within reach.

Estimand-specific deterministic and probabilistic bias analysis for matching, outcome regression, and unanchored meta-regression
Multiple-bias analysis combining confounding, selection, measurement, censoring, and design drift, with dependence preserved between bias parameters
Target-standardized multilevel network meta-regression, including the component case
Doubly robust network transport with study-level cross-fitting, nested inside explicit residual-bias analysis rather than presented as a substitute for it
Partial-identification and bias-function modules in mainstream software
Separation of within-study from between-study effect modification
Subnetwork-specific component drift with robust borrowing
Non-proportional-hazards, restricted-mean-survival, and multistate population adjustment, plus survival-specific bias analysis
Target-distribution uncertainty and distributionally robust standardization
Optimal individual-patient-data acquisition and bridging-trial design
Public validation datasets and cross-package benchmarks

Tier 3: high risk, high reward

Where the payoff is large and the probability of a clean result is not.

Federated cross-sponsor evidence synthesis
Mechanistic and statistical hybrid bridges
Invariant component effects across environments
Probabilistic component ontologies and treatment embeddings
Proximal methods using validated proxies for unmeasured modifiers
Graph-learning tools for design and diagnostics
Sequential and dynamic-regime indirect comparisons
Privacy-preserving release of target joint-distribution summaries

What the field does not need

Stated plainly, because an agenda that only adds is not an agenda.

Another weight optimizer with no estimand contribution
Another package reproducing conventional matching under different function names
Simulation studies restricted to linear models, good overlap, measured modifiers, and correctly specified outcomes, which is to say studies run where every method works
Disconnected-network analyses reporting one narrow interval and no bridge sensitivity
Black-box machine-learning outcome regression evaluated only by predictive accuracy
Component models that add every pairwise interaction without identification or hierarchy
Ranking metrics presented without uncertainty or decision relevance

The organizing discipline

No effect estimate without an estimand, no disconnected estimate without a stated bridge, no population-adjusted estimate without a named target, and no structural bridge without fragility or partial-identification analysis.