rct 0.0.4

design robust balanced randomized experiments

rct

What this package does

This package provides tools to generate robust and balanced random assignments following Banerjee, Chassang, Montero, and Snowberg (2019).

The RCT, KRerandomizedRCT, and QuantileTargetingRCT classes of the rct.design module implement RCT, K-rerandomized RCT, and Quantile Targeting RCT designs described in Banerjee, Chassang, Montero, and Snowberg (2019).

For each design, assignment_from_iid draws designs selected from i.i.d. assignments; assignment_from_shuffled draws designs selected from exchangeable assignments guaranteed to exactly match desired sampling weights (up to integer issues).

The package allows for an arbitrary number of treatment arms, specified via the weights argument in each design.

rct implements various balance objectives, including:

• minimizing the Mahalanobis distance between the mean of selected covariates across treatment arms;
• maximizing the minimum p-value for the regression of covariates on treatment dummies;
• soft blocking on selected covariates;
• linear combinations of existing objectives.

Customizing balance objectives, besides linear combinations of existing balance functions, is straightforward. First, you can pass different aggregating functions to the BalanceObjective constructor. For instance, this would allow to maximize the mean p-value rather than the minimum p-value. Second, you can simply define a new class inheriting from BalanceObjective and implementing the abstract method _balance_func.

Citation

To cite rct in publications, use

Banerjee, Abhijit, Sylvain Chassang, Sergio Montero, and Erik Snowberg.   
A theory of experimenters. No. w23867. National Bureau of Economic Research, 2017.

The corresponding bibtex entry is:

@techreport{banerjee2017theory,   
  title={A theory of experimenters},   
  author={Banerjee, Abhijit and Chassang, Sylvain and Montero, Sergio and Snowberg, Erik},   
  year={2017},   
  institution={National Bureau of Economic Research}   
}

Installation

This package is tested for python 3.6 and python 3.7 under Ubuntu Linux 16.04.

You may download the package via pip:

$ pip install rct

this will install all required dependencies.

Alternatively, you can clone ([e-mail]:sylvaingchassang/rct.git) or download a .zip of the repo. If you do so you must install requirements for the package manually. With pip, run

./rct$ pip install -r requirements.txt

Running tests

Before using the package, you may want to check that unit and integration tests pass on your machine. To this end, run

./rct$ pytest --cov=. --cov-report=term-missing

Examples

Example notebooks illustrate the use of rct modules:

• rct/notebooks/examples_rct.ipynb shows how to generate traditional i.i.d. and shuffled RCT assignments for binary and ternary treatments. The pvalue_report function provides a useful summary of assignment balance by reporting the (#treatments -1, #covariates) matrix of p-values obtained from regressing covariates on different treatment dummies.

• rct/notebooks/examples_k_rerandomized_rct.ipynb shows how to obtain k-rerandomized i.i.d. and shuffled assignments under various balance objectives.

• rct/notebooks/examples_quantile_targeting_rct.ipynb performs a similar exercise for quantile targeting experiment designs.

Integration tests located at rct/tests/test_integration.py replicate the content of these notebooks.

Contribute

If you want to improve rct please reach out!

Whether you are a programmer who wants to improve our code, or an experiment designer with a practical comment, or a new design idea, we want to talk to you!

On our current todo list (2019/09/20):

• adding type hints to improve readability;
• profiling & speed improvement;
• implementing sequential designs.