CausalNLP

CausalNLP is a practical toolkit for causal inference with text as treatment, outcome, or "controlled-for" variable.

Install

1. pip install -U pip
2. pip install causalnlp

Usage

Example: What is the causal impact of a positive review on a product click?

import pandas as pd
df = pd.read_csv('sample_data/music_seed50.tsv', sep='\t', error_bad_lines=False)

The file music_seed50.tsv is a semi-simulated dataset from here. Columns of relevance include:

• Y_sim: outcome, where 1 means product was clicked and 0 means not.
• text: raw text of review
• rating: rating associated with review (1 through 5)
• T_true: 1 means rating less than 3, 0 means rating of 5, where T_true affects the outcome Y_sim.
• T_ac: an approximation of true review sentiment (T_true) created with Autocoder from raw review text
• C_true:confounding categorical variable (1=audio CD, 0=other)

We'll pretend the true sentiment (i.e., review rating and T_true) is hidden and only use T_ac as the treatment variable.

Using the text_col parameter, we include the raw review text as another "controlled-for" variable.

from causalnlp.causalinference import CausalInferenceModel
from lightgbm import LGBMClassifier
cm = CausalInferenceModel(df, 
                         metalearner_type='t-learner', learner=LGBMClassifier(num_leaves=500),
                         treatment_col='T_ac', outcome_col='Y_sim', text_col='text',
                         include_cols=['C_true'])
cm.fit()

outcome column (categorical): Y_sim
treatment column: T_ac
numerical/categorical covariates: ['C_true']
text covariate: text
preprocess time:  1.1179866790771484  sec
start fitting causal inference model
time to fit causal inference model:  10.361494302749634  sec

Estimating Treatment Effects

CausalNLP supports estimation of heterogeneous treatment effects (i.e., how causal impacts vary across observations, which could be documents, emails, posts, individuals, or organizations).

We will first calculate the overall average treatment effect (or ATE), which shows that a positive review increases the probability of a click by 13 percentage points in this dataset.

Average Treatment Effect (or ATE):

print( cm.estimate_ate() )

{'ate': 0.1309311542209525}

Conditional Average Treatment Effect (or CATE): reviews that mention the word "toddler":

print( cm.estimate_ate(df['text'].str.contains('toddler')) )

{'ate': 0.15559234254638685}

Individualized Treatment Effects (or ITE):

test_df = pd.DataFrame({'T_ac' : [1], 'C_true' : [1], 
                        'text' : ['I never bought this album, but I love his music and will soon!']})
effect = cm.predict(test_df)
print(effect)

[[0.80538201]]

Model Interpretability:

print( cm.interpret(plot=False)[1][:10] )

v_music    0.079042
v_cd       0.066838
v_album    0.055168
v_like     0.040784
v_love     0.040635
C_true     0.039949
v_just     0.035671
v_song     0.035362
v_great    0.029918
v_heard    0.028373
dtype: float64

Features with the v_ prefix are word features. C_true is the categorical variable indicating whether or not the product is a CD.

Text is Optional in CausalNLP

Despite the "NLP" in CausalNLP, the library can be used for causal inference on data without text (e.g., only numerical and categorical variables). See the examples for more info.

Documentation

API documentation and additional usage examples are available at: https://amaiya.github.io/causalnlp/

How to Cite

Please cite the following paper when using CausalNLP in your work:

@article{maiya2021causalnlp,
    title={CausalNLP: A Practical Toolkit for Causal Inference with Text},
    author={Arun S. Maiya},
    year={2021},
    eprint={2106.08043},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
    journal={arXiv preprint arXiv:2106.08043},
}