pydtr 0.0.2

Python library of Dynamic Treatment Regimes

pydtr

Description

This is a python library to conduct a dynamic treatment regime (DTR), pydtr.

A DTR is a paradigm that attempts to select optimal treatments adaptively for individual patients.

Pydtr enables you to implement DTR methods easily by using sklearn-based interfaces.

Method	Single binary treatment	Multiple treatments	Multinomial treatment	Continuous treatment	Modeling flexibility	Interpretability
IqLearnReg (with sklearn)	:white_check_mark:	:white_check_mark:	:white_check_mark: (with pipeline)		:white_check_mark: (with arbitrary regression models)
IqLearnReg (with statsmodels)	:white_check_mark:	:white_check_mark:	:white_check_mark:		limited to OLS	:white_check_mark: (with confidence intervals)
GEstimation	WIP		WIP	WIP	WIP	WIP

IqLearnReg means a regression method of iterative q-learning.

When there are categorical independent variables and you use a sklearn model as a regression function, you need to encode the categorical variables before using the model.

We recommend to encode categorical variables by category_encoders and combine the encoders with the sklearn model by sklearn.pipeline.

G-estimation, a famous method of DTR, is now unavailable.

Requirements

• python>=3.6
• pandas>=1.1.2
• scikit-learn>=0.23.2
• numpy>=1.19.2
• statsmodels>=0.12.0

Installation

From pypi

pip install pydtr

From source

git clone https://github.com/fullflu/pydtr.git
cd pydtr
python setup.py install

Usage

Iterative Q Learning (IqLearnReg)

You need to import libraries and prepare data.

# import
import numpy as np
import pandas as pd
from sklearn.ensemble import RandomForestRegressor

from pydtr.iqlearn.regression import IqLearnReg

# create sample dataframe
n = 10
thres = int(n / 2)
df = pd.DataFrame()
df["L1"] = np.arange(n)
df["A1"] = [0, 1] * int(n / 2)
df["A2"] = [0] * int(n / 2) + [1] * int(n / 2)
df["Y1"] = np.zeros(n)
df["Y2"] = np.zeros(n)

You can use sklearn-based models.

# set model info
model_info = [
    {
        "model": RandomForestRegressor(),
        "action_dict": {"A1": [0, 1]},
        "feature": ["L1", "A1"],
        "outcome": "Y1"
    },
    {
        "model": RandomForestRegressor(),
        "action_dict": {"A2": [0, 1]},
        "feature": ["L1", "A1", "Y1", "A2"],
        "outcome": "Y2"
    }
]
# fit model
dtr_model = IqLearnReg(
    n_stages=2,
    model_info=model_info
)
dtr_model.fit(df)

# predict optimal atcions
opt_action_stage_1 = dtr_model.predict(df, 0)
opt_action_stage_2 = dtr_model.predict(df, 1)
opt_action_all_stages = dtr_model.predict_all_stages(df)

You can also use statsmodels-based models.

# set model info
model_info = [
    {
        "model": "p_outcome ~ L1 * A1",
        "action_dict": {"A1": [0, 1]},
        "feature": ["L1", "A1"],
        "outcome": "Y1"
    },
    {
        "model": "p_outcome ~ L1 + A1 + Y1 * A2",
        "action_dict": {"A2": [0, 1]},
        "feature": ["L1", "A1", "Y1", "A2"],
        "outcome": "Y2"
    }
]
# fit model
dtr_model = IqLearnReg(
    n_stages=2,
    model_info=model_info
)
dtr_model.fit(df)

# predict optimal atcions
opt_action_stage_1 = dtr_model.predict(df, 0)
opt_action_stage_2 = dtr_model.predict(df, 1)
opt_action_all_stages = dtr_model.predict_all_stages(df)

Please see examples to get more information.

Authors

• @fullflu

Contributors

Please feel free to create issues or to send pull-requests!

If all checkes have passed in pull-requests, I will merge and release them.

License

BSD

Structure

├── .circleci
│   ├── config.yml
├── .github
│   ├── CODEOWNERS
├── LICENSE
├── MANIFEST.IN
├── Makefile
├── README.md
├── examples
│   ├── ...several notebooks...
├── setup.cfg
├── setup.py
├── src
│   ├── pydtr
│   │   ├── __init__.py
│   │   └── iqlearn
│   │       ├── __init__.py
│   │       ├── base.py
│   │       └── regression.py
└── tests
    ├── test_iqlearn_sklearn_predict.py
    └── test_iqlearn_sm_predict.py

References

• Chakraborty, B, Moodie, EE. Statistical Methods for Dynamic Treatment Regimes. Springer, New York, 2013.