tf-inference

tf-inference computes tF critical values and confidence intervals for instrumental-variable (IV) models with one endogenous regressor and one excluded instrument. It provides generated tables for common confidence levels and numerical calibration for other levels.

The tF method and its theoretical results are due to David S. Lee, Justin McCrary, Marcelo J. Moreira and Jack Porter, authors of Valid t-Ratio Inference for IV (2022). This package provides an independent Python implementation of their Appendix B.3 construction and reusable numerical tables. Critical values depend on the first-stage F statistic and a two-sided significance level. The code constructs the curve from its boundary expansion and recurrence; the published rounded tables serve as test references.

This is an initial numerical implementation. Its calibration uses deterministic integration, a finite nuisance-parameter grid and local searches. These checks are documented and reproducible; they are not a proof of a uniform size bound. The inferential framework is asymptotic under the single-instrument IV model, not a finite-sample coverage guarantee or a remedy for invalid instruments.

Installation

Install with Python 3.12 or later:

python -m pip install tf-inference

NumPy, SciPy and PyArrow are required at runtime. No Stata, external data download, API key or network access is needed to evaluate or generate critical values.

Use the generated tables

from tf_inference import critical_value, confidence_interval

critical_value([3.8, 3.9, 10, 1000], alpha=0.05)
# Approximately [inf, 30.93, 3.43035, 1.95996]

confidence_interval(coefficient=1, se=0.4, f_stat=10, alpha=0.05)
# Approximately [-0.37214, 2.37214]

alpha is the significance level: 0.05 means 95% confidence. The return value of critical_value is the threshold for absolute t, not t squared or a standard-error multiplier. Divide it by the normal two-sided critical value to obtain the corresponding standard-error multiplier for that alpha.

Inputs broadcast as NumPy arrays. Confidence intervals put lower and upper bounds on the final axis. Complex inputs, nonfinite or negative F statistics, nonfinite coefficients, and nonpositive standard errors are rejected.

The 95% boundary is about 3.84145882. F at or below the boundary returns an infinite critical value; F strictly above it has a finite critical value. There is no artificial cutoff at F=4. The boundary is computed separately for each alpha as scipy.special.ndtri(alpha / 2)**2. Very close to the boundary, rounding in the input F itself can materially affect the result.

Bundled tables cover alpha values 0.0001, 0.0005, 0.001, 0.002, 0.005, 0.01, 0.02, 0.025, 0.03, 0.04, 0.05, 0.06, 0.075, 0.1, 0.125, 0.15, 0.2 and 0.25. Each contains at least 16,385 rows, concentrated near the boundary. List the installed levels with available_alphas(). Tables are never interpolated across alpha values.

The Python convenience functions and CLI default to method="table". A missing level raises an error; numerical calibration requires an explicit method="compute" or a direct call to calibrate. The auto method is not supported.

Calculate an unconventional level

from tf_inference import calibrate

calibration = calibrate(alpha=0.037)
calibration.critical_value([4.5, 10, 100])
calibration.confidence_interval(1, 0.4, 10)
print(calibration.diagnostics)

The bundled Parquet dataset is read and validated once; interpolators are cached per alpha and shared by load_table, critical_value and confidence_interval. Repeated lookup does not read files, verify checksums again or rebuild the interpolator. Explicit method="compute" calls use a separate cache of up to 32 calibrations. Direct calibrate calls always compute afresh. Tables loaded from a custom directory are read and checksum-verified on each call so changes are detected.

For large inputs, pass a NumPy array in one call. For repeated work at one level, retain load_table(alpha) or the result of calibrate(alpha) and call its methods directly.

The numerical alpha domain is [0.0001, 0.25]. It covers confidence levels from 75% to 99.99%. This is an explicit implementation range, not a theoretical limit of tF. Calculations outside it are rejected rather than extrapolated.

calibrate accepts epsilon, seeds, size_tolerance, strength_limit, strengths and correlations. Smaller initialization epsilon and more seeds allow convergence checks. These settings are not claims of certified error. Failed recurrence, integration or optimization raises ConvergenceError. Inspect the returned diagnostics when doing new calibrations, and validate beyond the search settings before relying on an unconventional level for research.

Generate constants for other software

The command-line interface exposes the same calculation:

tf-inference critical 3.9 10 100 --alpha 0.05
tf-inference critical 10 --alpha 0.037 --method compute
tf-inference generate --alpha 0.05 0.01 0.037 --output tables

The generator writes one calibrations.parquet dataset and a calibrations.json manifest. Adding or replacing one alpha preserves the other levels in that directory. The Parquet schema has three nonnullable float64 columns: alpha, f_stat and regularized_squared. Calibration settings, plateaus and diagnostics are embedded as JSON in the Arrow schema metadata. The schema-versioned manifest records the dataset's SHA-256 checksum. Loading validates the checksum, schema, finite values, ordered grids, boundary and plateau invariants, and agreement with the metadata.

export_table(calibration, directory, points=16385) and load_table(alpha, directory=directory) provide the same operations in Python. For a portable gzip CSV, add --csv to the generation command or pass include_csv=True to export_table. CSV files are optional exports, not packaged runtime data.

For accurate boundary evaluation, interpolate the regularized quantity Z(F) = (F - boundary) * critical(F)**2 using a monotone cubic interpolator. It is the Parquet regularized_squared column (and the optional CSV's branch_regularized_squared column). Then compute sqrt(Z(F) / (F - boundary)). Cap F at the metadata's plateau_f before evaluation and return infinity when F is at or below the boundary. The raw branch extends to the Property (C) endpoint; at large F use the selected plateau. The CSV's first critical value is infinite and must not be interpolated as an ordinary finite row.

The export measures relative interpolation differences at every cell midpoint against the computed curve, and rejects tables exceeding 1e-6 there. This is a measured check, not an interval-arithmetic error bound. The shipped validation also compares randomly chosen F values against a separately refined curve.

How the calculation works

1. Compute the boundary q and initialize the squared critical value near it using the expansion in Appendix B.3, step 1.
2. Apply the inverse map in step 2 to extend the decreasing curve. The code interpolates Z(F), whose finite boundary limit is q cubed, rather than interpolating a divergent critical value.
3. Locate the tangency that limits the decreasing segment under Property (C). Its admissible plateau is also constrained to be at least the normal two-sided critical value.
4. Integrate null rejection probabilities in the Gaussian weak-IV experiment. Search the nuisance parameters and move the plateau earlier if it over-rejects beyond the requested numerical tolerance.

The size search can change the plateau. For example, alpha=0.037 requires an adjustment to the first candidate. For alpha=0.01 the plateau is about 2.72549, above the normal 99% critical value of about 2.57583. Replacing the 95% normal quantile with another quantile does not implement general tF.

The default size search covers strength f0 from 0 to 80 and absolute correlation rho from 0 to 1. Here E[F] = 1 + f0 squared; f0 is not F. It includes singular rho=1 calculations, correlations approaching one, and local searches between grid points. The strong-instrument limit is checked separately by the normal critical-value constraint. The unbounded nuisance domain is not exhaustively certified. The recorded integration error does not include curve approximation or missed extrema in the parameter search.

Performance

Table lookup avoids numerical calibration. For large inputs, evaluate an array in one call. make benchmark records separate timings for warm scalar and array lookup, loading the complete Parquet dataset with empty package caches, and fresh calibration. Imports and operating-system file-cache effects are excluded. Results and environment details are in benchmarks/results.json; benchmarks/baseline.json preserves the earlier implementation's measurements. These are local observations, not CI timing assertions. Tests enforce that default calls never calibrate and warm calls avoid file access and interpolator construction.

Validation and development

Install development dependencies and run the checks:

uv sync --all-groups
make ci

Tests cover the published rounded-table values and plateau landmarks, initialization and grid refinement, the defining two-intersection equations, monotonicity and the exact boundary, near-singular numerical integration, independent Gaussian Monte Carlo calculations, unconventional-alpha plateau adjustment, interval broadcasting, table roundtrips and corrupted files, Parquet schema violations, complex inputs and large-strength numerical stability. make ci runs Ruff, Pyright, pydoclint, tests with a 90% coverage floor (including branch measurement), documentation and distribution checks. preen check --strict checks adoption of the shared py-canon standard.

validation/validate.py evaluates bundled tables against finer recurrence calculations and held-out nuisance-parameter grids extending to f0=160. validation/results.json records all cases and the environment. Reproduce it:

uv run python validation/validate.py --output validation/results.json

Rebuild the bundled constants from scratch:

make tables

make ci-docker uses the standard python:3.12 image. Documentation includes this README directly; it is not maintained as a second copy.

Scope and integration

This package takes summary quantities. It does not fit an IV model or verify that an instrument is valid. For use with an estimator, require one endogenous regressor and one excluded instrument, and use the appropriate first-stage squared t statistic and compatible covariance estimates. It supplies the numerical calibration; estimator-specific eligibility checks belong in the calling package.

This implements the 2022 tF construction. VtF is a different procedure using additional information and is outside the present implementation. No generic p-value API is provided: numerical nesting across arbitrary alpha levels would need separate validation before inverting the family of tests.

Citation and sources

Please cite the paper when using this package for research:

Lee, David S., Justin McCrary, Marcelo J. Moreira, and Jack Porter. 2022. Valid t-Ratio Inference for IV. American Economic Review 112 (10): 3260-3290.

@article{lee2022valid,
  author = {Lee, David S. and McCrary, Justin and Moreira, Marcelo J. and Porter, Jack},
  title = {Valid {t}-Ratio Inference for {IV}},
  journal = {American Economic Review},
  year = {2022},
  volume = {112},
  number = {10},
  pages = {3260--3290},
  doi = {10.1257/aer.20211063}
}

The authors provide the construction and supporting materials:

• Author manuscript with Appendix B.3, February 16, 2022.
• Published supplemental appendix.
• Data and code for the paper, the authors' replication package.
• Authors' supplementary material and Stata implementation.

AI assistance was used in this package's development and validation. The numerical limitations and executed checks are recorded above and in the repository's validation results.