This package contains tools for causal analysis using observational (rather than experimental) datasets.
Assuming you have pip installed, just run
pip install causality
The simplest interface to this package is probably through the CausalDataFrame object in causality.analysis.CausalDataFrame. This is just an extension of the pandas.DataFrame object, and so it inherits the same methods.
The CausalDataFrame current supports two kinds of causal analysis. First, it has a CausalDataFrame.zmean method. This method lets you control for a set of variables, z, when you're trying to estimate the effect of a discrete variable x on a continuous variable, y. It supports both returning the y estimates at each x value, as well as providing bootstrap error bars. For more details, check out the readme here.
The second kind of analysis supported is plotting to show the effect of discrete or continuous x on continous y while controlling for z. You can do this with the CausalDataFrame.zplot method. For details, check out the readme here.
the causality.estimation module contains tools for estimating causal effects from observational and experimental data. Most tools are parametric, like PropensityScoreMatching, and can be found in causality.estimation.parametric. Other models are non-parametric, and rely on directly estimating densities and using the g-estimation approach.
The causality.inference module will contain various algorithms for inferring causal DAGs. Currently (2016/01/23), the only algorithm implemented is the IC* algorithm from Pearl (2000). It has decent test coverage, but feel free to write some more! I've left some stubs in tests/unit/test\_IC.py.
To run a graph search on a dataset, you can use the algorithms like (using IC* as an example):
import numpy
import pandas as pd
from causality.inference.search import IC
from causality.inference.independence_tests import RobustRegressionTest
# generate some toy data:
SIZE = 2000
x1 = numpy.random.normal(size=SIZE)
x2 = x1 + numpy.random.normal(size=SIZE)
x3 = x1 + numpy.random.normal(size=SIZE)
x4 = x2 + x3 + numpy.random.normal(size=SIZE)
x5 = x4 + numpy.random.normal(size=SIZE)
# load the data into a dataframe:
X = pd.DataFrame({'x1' : x1, 'x2' : x2, 'x3' : x3, 'x4' : x4, 'x5' : x5})
# define the variable types: 'c' is 'continuous'. The variables defined here
# are the ones the search is performed over -- NOT all the variables defined
# in the data frame.
variable_types = {'x1' : 'c', 'x2' : 'c', 'x3' : 'c', 'x4' : 'c', 'x5' : 'c'}
# run the search
ic_algorithm = IC(RobustRegressionTest)
graph = ic_algorithm.search(X, variable_types)Now, we have the inferred graph stored in graph. In this graph, each variable is a node (named from the DataFrame columns), and each edge represents statistical dependence between the nodes that can't be eliminated by conditioning on the variables specified for the search. If an edge can be oriented with the data available, the arrowhead is indicated in 'arrows'. If the edge also satisfies the local criterion for genuine causation, then that directed edge will have marked=True. If we print the edges from the result of our search, we can see which edges are oriented, and which satisfy the local criterion for genuine causation:
>>> graph.edges(data=True)
[('x2', 'x1', {'arrows': [], 'marked': False}),
('x2', 'x4', {'arrows': ['x4'], 'marked': False}),
('x3', 'x1', {'arrows': [], 'marked': False}),
('x3', 'x4', {'arrows': ['x4'], 'marked': False}),
('x4', 'x5', {'arrows': ['x5'], 'marked': True})]We can see the edges from 'x2' to 'x4', 'x3' to 'x4', and 'x4' to 'x5' are all oriented toward the second of each pair. Additionally, we see that the edge from 'x4' to 'x5' satisfies the local criterion for genuine causation. This matches the structure given in figure 2.3(d) in Pearl (2000).
The causality.nonparametric module contains a tool for non-parametrically estimating a causal distribution from an observational data set. You can supply an "admissable set" of variables for controlling, and the measure either the causal effect distribution of an effect given the cause, or the expected value of the effect given the cause.
I've recently added adjustment for direct causes, where you can estimate the causal effect of fixing a set of X variables on a set of Y variables by adjusting for the parents of X in your graph. Using the dataset above, you can run this like
from causality.estimation.adjustments import AdjustForDirectCauses
from networkx import DiGraph
g = DiGraph()
g.add_nodes_from(['x1','x2','x3','x4', 'x5'])
g.add_edges_from([('x1','x2'),('x1','x3'),('x2','x4'),('x3','x4')])
adjustment = AdjustForDirectCauses()Then, you can see the set of variables being adjusted for by
>>> print adjustment.admissable_set(g, ['x2'], ['x3'])
set(['x1'])If we hadn't adjusted for 'x1' we would have incorrectly found that 'x2' had a causal effect on 'x3' due to the counfounding pathway x2, x1, x3. Adjustment for 'x1' removes this bias.
You can see the causal effect of intervention, P(x3|do(x2)) using the measured causal effect in adjustment,
>>>from causality.estimation.nonparametric import CausalEffect
>>>admissable_set = adjustment.admissable_set(g,['x2'], ['x3'])
>>>effect = CausalEffect(X, ['x2'], ['x3'], variable_types=variable_types, admissable_set=list(admissable_set))
>>>x = pd.DataFrame({'x2' : [0.], 'x3' : [0.]})
>>>effect.pdf(x)
0.268915603296Which is close to the correct value of 0.282 for a gaussian with mean 0. and variance 2. If you adjust the value of 'x2', you'll find that the probability of 'x3' doesn't change. This is untrue with just the conditional distribution, P(x3|x2), since in this case, observation and intervention are not equivalent.
This repository is in its early phases. The run-time for the tests is long. Many optimizations will be made in the near future, including
- Implement fast mutual information calculation, O( N log N )
- Speed up integrating out variables for controlling
- Take a user-supplied graph, and find the set of admissable sets
- Front-door criterion method for determining causal effects
Pearl, Judea. Causality. Cambridge University Press, (2000).