confseq 0.0.1

Confidence sequences and uniform boundaries

Confidence sequences and uniform boundaries

This library supports calculation of uniform boundaries, confidence sequences, and always-valid p-values. These constructs are useful in sequential A/B testing, best-arm identification, and other sequential statistical procedures. The main reference is

S. R. Howard, A. Ramdas, J. McAuliffe, J. Sekhon. Uniform, nonparametric, non-asymptotic confidence sequences. Preprint, arXiv:1810.08240.

This library is in early-stage development and should not be considered stable. I have tested it only on Python 3.7.0 on macOS Mojave. The implementation is in C++ and a compiler with C++14 support is required to build the package.

Installing the python package

Run pip install confseq.

Demos

Estimating average treatment effect in a randomized trial

demo/ate_demo.py illustrates how to compute a confidence sequence for average treatment effect in a randomized trial with bounded potential outcomes, along with an always-valid p-value sequence. The method is based on Corollary 2 of the paper and uses the gamma-exponential mixture boundary. This demo requires numpy and pandas.

Uniform boundaries

The confseq.boundaries module implements several uniform boundaries from the paper.

• There are four mixture boundaries. These are implemented by the functions <TYPE>_log_mixture() and <TYPE>_mixture_bound(), where <TYPE> is one of normal (Propositions 4 and 5), gamma_exponential (Proposition 8), gamma_poisson (Proposition 9), or beta_binomial (Propositions 6 and 7).

  • <TYPE>_log_mixture(s, v, ...) returns the logarithm of the mixture supermartingale when called with S_t, the martingale, and V_t, the intrinsic time process. The reciprocal of the exponential of this value is an always-valid p-value. These functions are denoted log(m(s,v)) in the paper.
  • <TYPE>_mixture_bound(v, alpha, ...) returns the uniform boundary with crossing probability at most alpha, evaluated at intrinsic time v.

  Each function takes another required argument v_opt and an optional argument alpha_opt=0.05. These arguments are used to set the tuning parameter for each mixture, denoted by rho or r in the paper, optimizing the uniform boundary with crossing probability alpha_opt for intrinsic time v_opt. Such tuning is discussed in section 3.5 of the paper.

  The gamma-exponential and gamma-Poisson mixtures also require a scale parameter c. The beta-binomial mixture requires range parameters g and h. Finally, the normal_* and beta_binomial_* functions accept an optional boolean parameter is_one_sided which is True by default. If False, the two-sided variants of these mixtures are used (Propositions 4 and 6).

• The polynomial stitching boundary (see Theorem 1 and the subsequent example) is implemented by poly_stitching_bound. Besides v and alpha, this function requires the tuning parameter v_min as well as optional parameters c, s, and eta, all documented in the paper.

C++ library

The underlying implementation is in a single-file, header-only C++ library in src/confseq/uniform_boundaries.h. The top of the file defines a simplified interface mirroring the Python interface described above. Below that is an object-oriented interface useful for more involved work. The confseq.boundaries Python module is a wrapper generated by pybind11.

Unit tests

Run make -C /path/to/confseq/tests runtests to run the C++ unit tests.