earlysign 0.1.7

Early signs, faster decisions. A Python library for sequential/safe testing (alpha-spending, e-processes, etc.).

EarlySign

Early signs, faster decisions.

---

What is this?

EarlySign is a Python library for sequential/safe testing (alpha-spending, e-processes, etc.).

1. Group sequential tests for interim analysis
   • By using alpha-spending functions to control the overall Type I error rate, you can stop early for efficacy or futility, making your experiments more efficient without compromising statistical integrity. This approach allows for a pre-specified number of interim analyses during an experiment.
2. e-processes for anytime-valid inference
   • It allows you to continuously monitor your experiments and make decisions as soon as the evidence is strong enough, without waiting for a predetermined sample size. This can lead to faster conclusions, saving time and resources, while maintaining statistical rigor.

Install

pip install earlysign

Usage

This library supports the following steps in your experimentation.

1. Planning / Designing
2. Executing / Analyzing
3. Reporting / Visualizing
4. (optionally) Educating

Overview

from earlysign import designs

##################
# Planning phase #
##################
# A) assisted planning
design = designs.GroupSequential.from_power_analysis(
    mde=0.01,
    power=0.8,
    alpha=0.05,
    outcome_type="binary",
    num_interims=4 # analyze 4 times
)

# B) or... manually specify your experimental design
design = designs.GroupSequential(
    name="example-run", variants=["A", "B"],
    outcome_type="binary",
    alpha_spending_fn="obrien-fleming",
    max_batches=1000,
    alpha=0.05,
    efficacy=True, futility=False
)

###################
# Execution phase #
###################
# 1) execute the experiment in object-oriented style
from earlysign import Experiment
experiment = Experiment(design)  # stateful experiment object
experiment.initialize()          # init state
experiment.update(observation_1) # first batch of data
experiment.update(observation_2) # second batch of data
experiment.conclude()            # form the best advice with the current state
report = experiment.report()     # generate a report

# 2) or in functional style if you prefer
from earlysign import handlers
handler = handlers.get_handler(design)        # design-specific handler
state = handler.initial_state()              # {} -> State
state = handler.update(state, observation_1) # (Design, State, Obs) -> State
state = handler.update(state, observation_2) # (Design, State, Obs) -> State
conclusion = handler.conclude(state)         # (State) -> Conclusion
report = handlers.get_reporter(design).report(conclusion) # Conclusion -> Report

Internally, the object-oriented API is a wrapper around the functional-programming API.

Documentation

For detailed usage and available methods, see the docs.

You can also find useful tutorial here.

References

• R 'gsDesign' package
• R 'gsDesign' Technical Manual
• Stata 19 Adaptive Designs
• R 'safestats' package

Other References

• R 'rpact' package
• R 'ldbounds' package

Initial TODO

• どんな統計量を使いたいか（これによってどんなデータを受け取れるかが変わる）
  • Wald's Z
  • Bernoulli two proportions
• どんなシグナルを使いたいか（どんな決定領域定義を使うかも選択）
  • GroupSequentialFutilitySignal(criterion=AsymptoticNormalBoundary())
  • AnytimeValidEValueSignal(alpha)
• 実験プロセスを流す機能
• 実験の途中経過をレポートする機能
• 上記の設定を設定ファイルから行える
• 上記の途中経過を保存・再開できる
• 既存パッケージのAPIを解析して，パラメーター類を分配
• 既存パッケージのユースケースを調べて，内容とは別にまずAPIを定義付けていく
• ExperimentBuilderを実装して，実験設計のプロセスを支援する
• PyPI