File size: 8,020 Bytes
996dbb8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 | ---
pretty_name: Statistical Inference Arena — Instrumental Variables (LATE)
language:
- en
tags:
- tabular
- synthetic
- causal-inference
- instrumental-variables
- late
- benchmark
configs:
- config_name: starter_catalog
default: true
data_files:
- split: train
path: starter_train.csv
- split: validation
path: starter_dev.csv
- split: test
path: starter_test.csv
- config_name: evaluation_inputs
data_files:
- split: test
path: test_schema.csv
---
# Statistical Inference Arena · Instrumental Variables (LATE)
Estimate the local average treatment effect (LATE) from a study with a binary instrument and a binary treatment, and return a two-sided 95% confidence interval. Each study contains 500 observations and contributes one estimate. The benchmark compares a frozen TabPFN-IV adapter checkpoint and a cross-fitted random forest on the same 2,000 studies. [Model](https://huggingface.co/Brownight/tabpfn-iv-adapter-v0.3)
| Purpose | Studies | Download |
|---|---:|---|
| Historical adapter fitting | 20,000 | [Training archive](adapter-train.zip) |
| Historical checkpoint selection | 1,000 | [Validation archive](adapter-validation.zip) |
| Try estimation and scoring | 48 train, 16 dev, 32 test | [Starter kit](iv-starter-v1.zip) |
| Evaluate a fixed method | 2,000 | [Observed inputs, folds and prediction template](iv-test-inputs-v1.zip) |
| Reproduce the comparison | Same 2,000 | [Population LATE labels, scorer and reference predictions](iv-test-scoring-v1.zip) |
## Target and observations
For each study, hold its data-generating mechanism and parameters fixed. The target is the average effect of treatment among compliers, the units whose treatment follows the instrument.
$$
\mathrm{LATE} = \mathbb{E}\!\left[Y(1)-Y(0)\mid D(1) > D(0)\right].
$$
Here D(z) is a unit's treatment under instrument value z and Y(d) its outcome under treatment d. The expectation is over compliers in the population that generated the study. Every design satisfies exclusion, monotonicity (D(1) ≥ D(0)) and independence of the instrument given the covariates, so the LATE equals the ratio of the covariate-adjusted instrument effects on the outcome and on the treatment. Report one estimate in outcome units (a probability difference for binary outcomes) and one two-sided 95% confidence interval per study.
Each NPZ holds baseline covariates `X`, a binary instrument `Z`, a binary treatment `D` and the outcome `Y`, in matching row order. An array with shape (500, 0) denotes a study with no observed covariates. Supply only these four arrays to an estimator.
## Run a first comparison
Extract the starter kit, install NumPy and run the following from the extracted directory.
```sh
python baseline.py dev/tables > predictions.csv
python score.py dev/truth.csv predictions.csv
```
The example computes the unadjusted Wald ratio (difference in mean outcomes over difference in mean treatment between the two instrument arms) with a delta-method normal interval. It ignores the covariates, so designs in which the instrument depends on X bias this simple estimate. On the 16 development studies it gives bias 0.4994, RMSE 1.1440, coverage 68.75% and mean interval length 0.7004. The starter draws 96 separate studies from the three-family prior with seed 20261119 and includes the generator (`generate.py` with `iv_data.py` and `reference/`; needs numpy and scipy). Its three viewer splits describe the files inside the starter ZIP.
## Historical data and evaluation design
The historical files contain the original adapter training and validation studies. For each study the generator draws one of three families with equal probability (Ding 2017, Qiu 2021, Kennedy 2020), then the family's parameters from our prior, then 564 rows: 500 context rows followed by 64 query rows. Each NPZ also stores the analytic nuisance values at the query rows (`targets`: e(x), mD(x,0), mD(x,1), mY(x,0), mY(x,1)), the column permutation the adapter saw and metadata with the parameters, seeds and true LATE. Use only the observed arrays as model inputs. Root seed 20261116.
The evaluation uses one pooled split. For each study the generator draws one of the same three families uniformly, draws fresh parameters from the same prior, then samples 500 units. Seed 20261117 selects the exact tables used for the published comparison; the realized counts are Ding 665, Qiu 658 and Kennedy 677. There is no in-distribution / out-of-distribution split and one parameter draw per table, so the error variance is across studies. The input archive preserves the observed NPZ bytes and row order under opaque table IDs. The viewer's `evaluation_inputs` configuration lists one row per study with its size, number of covariates, outcome type and the SHA-256 of the input file. [DGP families, priors and realized counts](DGP.md)
## Evaluate a method
Freeze the estimator before opening the scoring archive. The evaluation labels are public for reproducibility. Keep the supplied five folds for cross-fitting: one fixed random 5-fold split of the 500 rows, identical for every study and used by both reference methods. Row indices start at 0 and fold labels run from 1 to 5. Write one row per table using the template columns below.
```csv
table_id,estimate,ci_lower,ci_upper
```
Run `python score.py truth.csv predictions.csv` after extracting the scoring archive. Each study has equal weight. Bias and RMSE use estimate minus population LATE. Error variance uses the sample variance of those errors (denominator 1,999). Coverage counts intervals containing the truth, including either endpoint. Mean interval length averages upper minus lower. Measure elapsed seconds per study during estimation (point estimate and interval, including any bootstrap; excluding model loading and warm-up) and record the hardware if reporting runtime.
TabPFN-IV and the random-forest comparator estimate the instrument propensity e(x), the treatment means mD(x,z) and the outcome means mY(x,z) with the same 5-fold cross-fitting, then apply the same augmented inverse-probability-weighted ratio for the LATE. TabPFN-IV (bf16, one process on one NVIDIA L4) uses an influence-function interval. The random forest (scikit-learn, 200 trees, minimum leaf size 5, one single-threaded process per core on an AMD EPYC 7B13) uses 50 whole-row bootstrap resamples with refitting and a percentile interval; the trimmed variant clips e(x) to [0.05, 0.95] before the ratio. Their predictions, timings and the resulting metrics are in the scoring archive (`predictions_pfn.csv`, `predictions_rf.csv`, `predictions_rf_trim.csv`, `reference_metrics.json`):
| Method | RMSE | Bias | Error var. | 95% coverage | CI length | Time (s/table) |
|---|---|---|---|---|---|---|
| TabPFN-IV | 0.268 | 0.012 | 0.072 | 95.65% | 0.875 | 0.366 |
| RF + bootstrap | 3.408 | 0.079 | 11.613 | 91.40% | 5.390 | 182.6 |
| RF (trimmed) + bootstrap | 0.316 | 0.022 | 0.099 | 90.25% | 1.111 | 179.9 |
The untrimmed forest fails on the Kennedy design, whose single covariate lets leaf frequencies reach a propensity of exactly 0 or 1; 260 of its 677 studies have an absolute error above 1.
## References
- Ding, VanderWeele and Robins (2017). *Instrumental variables as bias amplifiers with general outcome and confounding.* Biometrika 104(2), 291–302. [Paper](https://doi.org/10.1093/biomet/asx009)
- Qiu, Tao and Zhou (2021). *Inference of heterogeneous treatment effects using observational data with high-dimensional covariates.* JRSS-B 83(5), 1016–1043. [Paper](https://doi.org/10.1111/rssb.12469)
- Kennedy, Balakrishnan and G'Sell (2020). *Sharp instruments for classifying compliers and generalizing causal effects.* Annals of Statistics 48(4). [Paper](https://arxiv.org/abs/1801.03635)
- The Ding and Qiu generators are verbatim copies of the Arena registry files `iv-ding-2017-binary-bias-v1.py` and `iv-qiu-2021-principal-stratum-poisson-v1.py` in [stats-inference-arena](https://github.com/Statsinference/stats-inference-arena).
|