Dataset Viewer
Auto-converted to Parquet Duplicate
table_id
string
preset
string
config_id
string
replicate
int64
n
int64
truth
float64
archive_path
string
train_0001
he_song_xu
train_c0001
1
500
0
train/tables/train_0001.csv
train_0002
he_song_xu
train_c0001
2
500
0
train/tables/train_0002.csv
train_0003
shan
train_c0002
1
500
-0.085001
train/tables/train_0003.csv
train_0004
shan
train_c0002
2
500
-0.085001
train/tables/train_0004.csv
train_0005
he_song_xu
train_c0003
1
500
0
train/tables/train_0005.csv
train_0006
he_song_xu
train_c0003
2
500
0
train/tables/train_0006.csv
train_0007
shan
train_c0004
1
500
0.569888
train/tables/train_0007.csv
train_0008
shan
train_c0004
2
500
0.569888
train/tables/train_0008.csv
train_0009
he_song_xu
train_c0005
1
500
-0
train/tables/train_0009.csv
train_0010
he_song_xu
train_c0005
2
500
-0
train/tables/train_0010.csv
train_0011
shan
train_c0006
1
500
0
train/tables/train_0011.csv
train_0012
shan
train_c0006
2
500
0
train/tables/train_0012.csv
train_0013
he_song_xu
train_c0007
1
500
0
train/tables/train_0013.csv
train_0014
he_song_xu
train_c0007
2
500
0
train/tables/train_0014.csv
train_0015
shan
train_c0008
1
500
2.320861
train/tables/train_0015.csv
train_0016
shan
train_c0008
2
500
2.320861
train/tables/train_0016.csv
train_0017
he_song_xu
train_c0009
1
500
-0
train/tables/train_0017.csv
train_0018
he_song_xu
train_c0009
2
500
-0
train/tables/train_0018.csv
train_0019
shan
train_c0010
1
500
3.196939
train/tables/train_0019.csv
train_0020
shan
train_c0010
2
500
3.196939
train/tables/train_0020.csv
train_0021
he_song_xu
train_c0011
1
500
-0
train/tables/train_0021.csv
train_0022
he_song_xu
train_c0011
2
500
-0
train/tables/train_0022.csv
train_0023
shan
train_c0012
1
500
1.915269
train/tables/train_0023.csv
train_0024
shan
train_c0012
2
500
1.915269
train/tables/train_0024.csv
train_0025
he_song_xu
train_c0013
1
500
0.026385
train/tables/train_0025.csv
train_0026
he_song_xu
train_c0013
2
500
0.026385
train/tables/train_0026.csv
train_0027
shan
train_c0014
1
500
0.833813
train/tables/train_0027.csv
train_0028
shan
train_c0014
2
500
0.833813
train/tables/train_0028.csv
train_0029
he_song_xu
train_c0015
1
500
-0
train/tables/train_0029.csv
train_0030
he_song_xu
train_c0015
2
500
-0
train/tables/train_0030.csv
train_0031
shan
train_c0016
1
500
0.413399
train/tables/train_0031.csv
train_0032
shan
train_c0016
2
500
0.413399
train/tables/train_0032.csv
train_0033
he_song_xu
train_c0017
1
500
-0
train/tables/train_0033.csv
train_0034
he_song_xu
train_c0017
2
500
-0
train/tables/train_0034.csv
train_0035
shan
train_c0018
1
500
0
train/tables/train_0035.csv
train_0036
shan
train_c0018
2
500
0
train/tables/train_0036.csv
train_0037
he_song_xu
train_c0019
1
500
0
train/tables/train_0037.csv
train_0038
he_song_xu
train_c0019
2
500
0
train/tables/train_0038.csv
train_0039
shan
train_c0020
1
500
3.257138
train/tables/train_0039.csv
train_0040
shan
train_c0020
2
500
3.257138
train/tables/train_0040.csv
train_0041
he_song_xu
train_c0021
1
500
-0.047494
train/tables/train_0041.csv
train_0042
he_song_xu
train_c0021
2
500
-0.047494
train/tables/train_0042.csv
train_0043
shan
train_c0022
1
500
2.537791
train/tables/train_0043.csv
train_0044
shan
train_c0022
2
500
2.537791
train/tables/train_0044.csv
train_0045
he_song_xu
train_c0023
1
500
0
train/tables/train_0045.csv
train_0046
he_song_xu
train_c0023
2
500
0
train/tables/train_0046.csv
train_0047
shan
train_c0024
1
500
0
train/tables/train_0047.csv
train_0048
shan
train_c0024
2
500
0
train/tables/train_0048.csv

Statistical Inference Arena · Causal Mediation Analysis

Given a simulated study with 500 observations, how much of a treatment's effect passes through its mediators? Estimate the population natural indirect effect (NIE) and return a two-sided 95% interval. Each study is an independent draw from a specified data-generating process and contributes one estimate. Both classical estimators and tabular foundation models can enter. Benchmark and submission · Leaderboard · Release v1.0

What do you want to do? Tables Ground truth Download
Fit on the historical MediPFN adapter studies 8,000 Public Adapter training
Train on the error-estimation head's fitting studies 10,000 Public Error-head training
Train on the fresh studies from fine-tuning 2,000 Public Fine-tuning
Try a complete workflow, from training to local scoring 48 train, 16 dev, 32 test Public Starter kit
Evaluate a fixed method for the leaderboard 2,000 Maintainers score submissions Official inputs and submission CSV

Target and observations

The outcome treatment stays at 1 while the mediators shift jointly from their values under control to their values under treatment.

NIE=E ⁣[Y(1,M(1))−Y(1,M(0))]. \mathrm{NIE}=E\!\left[Y^{(1,M^{(1)})}-Y^{(1,M^{(0)})}\right].

The expectation is over people in the population represented by a study's generating mechanism. Give the estimator the study's baseline covariates $X$, binary treatment $A$, post-treatment mediator vector $M$ and outcome $Y$. Report the effect in outcome units. For a binary outcome the effect is a difference in probabilities.

Run your first study comparison

Extract the starter kit and run these commands from its directory.

python -m pip install numpy
python baseline.py --tables dev/tables --output dev_predictions.csv
python score.py --truth dev/truth.csv --predictions dev_predictions.csv

The example estimates path coefficients with two additive linear regressions and forms a percentile interval from 200 row-bootstrap samples. On the 16 development tables it gives bias −0.3154, RMSE 0.6645, coverage 75% and mean interval length 0.1626. The kit includes the full example output, dev_predictions_example.csv. Replace the estimator, keep the four output columns and run the scorer again. Once you have chosen your settings, use the kit's independent public test/ split. The official leaderboard also includes a cross-fitted random-forest comparator.

Sampling and files

Historical fitting data. The 20,000 distinct tables were regenerated from the frozen MediPFN Arena fitting plan and samplers. They cover eight source families and ten presets. The adapter archive spans the full fitting run, including steps after the selected checkpoint. Its NIE labels are provided here for supervised use by other researchers. The original adapter learned from observed variables. The fine-tuning archive contains fresh tables from that stage. Replayed studies are available in their original training archive. Each ZIP includes a table manifest, truth.csv with population NIE labels, and a NumPy loading example. Each tables/<table_id>.npz stores four observed arrays (x, a, m, y) plus covariate and mediator types. Labels can be joined by table_id. Historical sampling and fitting stages

Public starter. The starter adapts two source designs. He–Song–Xu has linear mediator and outcome models and includes composite-null settings. Shan–Li–Ai has nonlinear covariate effects and a treatment–mediator interaction. Train, dev and public test use disjoint parameter configurations, and each configuration yields two independent studies. The kit's generator can draw further training studies. The starter_catalog viewer lists one row per study with its preset, configuration, replicate, row count, population NIE (truth) and CSV path inside the kit. The viewer calls the kit's dev/ split validation. Equations, priors and source papers

Official test. The 2,000 observed studies come from 1,000 configurations with two independent draws each. Each study has an opaque table ID. The release provides the observed CSV, variable types and five treatment-stratified outer folds for each study, plus a blank predictions template. Maintainers keep the table-to-generator assignments, parameter values, simulation seeds and true indirect effects private. Official studies have 0–4 baseline covariates and 1–100 mediators, with binary or continuous outcomes. The mediator columns are named M1 through Mq, including M1 for one mediator. The starter uses M for its single continuous mediator. The official_inputs viewer lists the observed dimensions and variable types.

Submit a method

Freeze the method and settings before running the official inputs. For cross-fitting, use the five folds in folds/<table_id>.csv in their supplied row order. Fill predictions_template.csv with one row for each of the 2,000 table IDs. The three numeric fields must be finite, with ci_lower at or below ci_upper.

table_id,estimate,ci_lower,ci_upper

Maintainers compare predictions to the held-out population effects. Bias, RMSE, 95% interval coverage and mean interval length weight each study equally. The official comparison also reports within-configuration variance using paired estimates, and runtime when available. Describe the estimator, interval method, training exposure and hardware so another group can reproduce the submission. Submit a method

References

  • Bo and Hu. MediPFN: Amortized In-Context Nuisance Learning and Inference for Causal Mediation Analysis. Paper · Model and worked examples
  • He, Song and Xu. Adaptive Bootstrap Tests for Composite Null Hypotheses in the Mediation Pathway Analysis. Paper
  • Shan, Li and Ai. Efficient Nonparametric Inference for Mediation Analysis with Nonignorable Missing Confounders. Paper
Downloads last month
-

Paper for shibo769/StatsInference-Arena-Mediation