MetaFrontierLab

MetaFrontierLab is a prototype meta-analysis framework for rapid frontier-method development.

It does not pretend to be a validated statistical breakthrough. It is a research-grade implementation that takes several strong recent developments in meta-analysis and pushes them into one combined workflow:

• exact sparse-data likelihoods for binary outcomes
• model ensembling for publication-bias adjustment
• transport-to-target weighting for target-population inference
• power-style discounting of lower-trust evidence
• shrinkage meta-regression for many moderators

The main implemented method is called TBEMA:

Transport-Bias Exact Meta-Analysis

TBEMA combines four ideas in one estimator:

1. For sparse binary studies, it uses an exact conditional likelihood rather than relying only on asymptotic normal log-odds approximations.
2. It fits multiple bias-adjusted submodels instead of betting everything on one publication-bias correction.
3. It targets a user-defined population through relevance weights, so the estimand can move away from the average study population.
4. It discounts lower-credibility evidence through fractional likelihood weights, inspired by recent power-likelihood work.

Implemented Components

1. Exact sparse-data random-effects layer

For binary treatment-control studies, the code uses a conditional exact likelihood built from the 2x2 table margins and integrates over a random-effects distribution with Gauss-Hermite quadrature.

That choice is motivated by recent work showing that sparse meta-analysis can break standard normal-normal random-effects approximations.

2. Selection-tempered evidence weighting

Instead of a single publication-bias correction, TBEMA fits an ensemble:

• baseline
• small-study adjusted
• mildly selection-tempered
• strongly selection-tempered

The selection tempering is a new extension in this prototype: studies with highly significant results are smoothly downweighted according to a tunable significance-response curve, then the resulting submodels are stacked with information-criterion weights.

This is not identical to a classical Copas or full selection-model likelihood. It is a deliberately robust pseudo-likelihood extension intended as a frontier prototype.

3. Transport-to-target estimation

If study-level population descriptors and a target profile are supplied, TBEMA applies Gaussian-kernel relevance weights so the pooled effect is explicitly targeted to the chosen population.

This pushes meta-analysis away from “what is the average study effect?” toward “what is the best estimate for the population I actually care about?”

4. Power-style design discounting

Each study can be assigned a design_strength between 0 and 1.

• 1.0 means full trust
• values below 1.0 discount the study through a fractional-likelihood contribution

This is useful when combining randomized and observational evidence, or when some studies are much less credible than others.

5. Moderator shrinkage

Moderator effects are estimated with ridge regularization so the method remains usable when there are many candidate moderators and not many studies.

What Is New Here

The prototype goes beyond any one cited paper by combining:

• exact sparse-data meta-analysis
• publication-bias model ensembling
• target-population transportability
• power-likelihood discounting
• moderator shrinkage

in one objective function and one result object.

In practical terms, the project is trying to answer:

What if a meta-analysis had to be sparse-data aware, publication-bias pessimistic, target-population specific, and willing to partially borrow from lower-trust studies at the same time?

That joint problem is the frontier this prototype is aiming at.

Source Trail

The design draws directly from these recent primary sources:

• Bartoš et al. (2023), Research Synthesis Methods: RoBMA model-averages complementary publication-bias adjustments rather than choosing one method.
  https://pmc.ncbi.nlm.nih.gov/articles/PMC10087723/
• Hu et al. (2024), Biometrics: extends exact-likelihood sparse-data meta-analysis to publication-bias sensitivity analysis.
  https://academic.oup.com/biometrics/article/80/3/ujae092/7754376
• Lin, Tarp, and Evans (2025), Biometrics: uses a power likelihood to borrow observational information while controlling how much it contributes.
  https://academic.oup.com/biometrics/article/81/1/ujaf008/8016472
• Dahabreh et al. (2022), Clinical Trials: transport treatment-effect estimates from multiple trials to a defined target population.
  https://pmc.ncbi.nlm.nih.gov/articles/PMC9066547/
• Gronsbell et al. (2025), Stats: exact inference for random-effects meta-analysis with small, sparse data.
  https://www.mdpi.com/2571-905X/8/1/5
• Rose (2024), Stata Journal: sparse multivariate meta-analysis becomes tractable via penalized low-dimensional structure.
  https://ageconsearch.umn.edu/record/361294

Files

• metafrontier/core.py: main estimator
• metafrontier/simulation.py: data generator and naive comparator
• metafrontier/benchmark_methods.py: benchmark comparators and method adapters
• metafrontier/benchmarking.py: scenario engine and summary aggregation
• metafrontier/reporting.py: plot and report generation
• run_demo.py: end-to-end simulation and example run
• run_benchmarks.py: multi-scenario benchmark runner
• generate_benchmark_report.py: report generation from existing benchmark outputs
• generate_benchmark_pdf.py: polished PDF export from benchmark outputs
• results/: generated outputs

Run

python run_demo.py

To run the benchmark suite:

python run_benchmarks.py --replications 4

To generate the report from benchmark outputs:

python generate_benchmark_report.py --benchmark-dir results/benchmarks

Or in one step:

python run_benchmarks.py --replications 4 --report

To create a PDF report from an existing benchmark directory:

python generate_benchmark_pdf.py --benchmark-dir results/benchmarks_scaled_full

Output

The demo writes:

• results/demo_summary.json
• results/observed_studies.csv
• results/submodel_table.csv

The benchmark runner writes:

• results/benchmarks/benchmark_runs.csv
• results/benchmarks/benchmark_summary.csv
• results/benchmarks/benchmark_metadata.json
• results/benchmarks/benchmark_summary.json

The report generator writes:

• results/benchmarks/report/benchmark_report.md
• results/benchmarks/report/benchmark_report.html
• results/benchmarks/report/figures/*.png

Benchmark Methods

The suite currently benchmarks:

• tbema: the canonical TBEMA preset exposed by make_tbema_analyzer()
• exact_baseline: exact sparse-data baseline without transport or selection adjustments
• dersimonian_laird: conventional normal-normal random-effects pooling
• henmi_copas: a Python translation of the Henmi-Copas publication-bias-robust interval method from metafor
• metafor_trimfill_external: trim-and-fill via metafor
• metafor_selmodel_external: step-function selection model via metafor::selmodel()
• copas_selection_external: Copas selection model via metasens

There is also an optional external adapter:

• robma_bibma_external: calls RoBMA::BiBMA() through Rscript if a suitable R + RoBMA environment exists

The benchmark runner detects external methods per package, so metafor, metasens, and RoBMA adapters appear independently when their R dependencies are available.

Caveat

This is a serious prototype, not a claim of settled methodology. If you want publication or applied deployment, the next step is formal simulation benchmarking against established methods under many data-generating regimes.