Get startedYour first causal effect

Your first causal effect

Learn a causal graph from data, check that it answers your question, and estimate the effect with an interval. The example runs in seconds on a bundled dataset.

Run the program#

The data comes from the textbook burglar-alarm network (Pearl, 1988). A burglary or an earthquake can set off an alarm, and the alarm makes John and Mary call. Because the network is known, you can check the answer.

from ergodic import discover, estimate_effect, identify_effect
from ergodic.datasets import load_earthquake

# 1. Load 10,000 rows sampled from the burglar-alarm network.
alarm = load_earthquake(n=10_000, seed=0)
rows = alarm.data.to_pandas()
print(rows.shape)
print((rows == "True").mean().round(3))

# 2. Learn the graph from the data alone.
graph = discover(alarm.data, method="pc").graph
for edge in sorted(edge.glyph() for edge in graph.edges):
    print(edge)

# 3. Ask whether the data can answer the question.
identification = identify_effect(graph, "Burglary", "JohnCalls")
print(identification)

# 4. Estimate the effect of a burglary on John calling.
estimate = estimate_effect(graph, alarm.data, "Burglary", "JohnCalls", treated="True")
print(estimate)

(10000, 5)
Burglary      0.011
Earthquake    0.022
Alarm         0.017
JohnCalls     0.067
MaryCalls     0.022
dtype: float64
Alarm --> JohnCalls
Alarm --> MaryCalls
Alarm <-- Burglary
Alarm <-- Earthquake
Identification(adjustment: P(JohnCalls | do(Burglary)) = sum_z P(JohnCalls | Burglary, z) P(z), with z = {Earthquake})
EffectEstimate(aipw, analytic: ate=0.7982, se=0.0333, 95% CI [0.733, 0.863], n=10000)

• Step 1. Five yes/no columns. Burglaries are rare: about 1 row in 100.
• Step 2. discover runs the PC algorithm. It finds all four true edges and their directions.
• Step 3. The effect is identifiable, by adjusting for Earthquake.
• Step 4. A burglary raises the chance that John calls by about 0.80, with a 95% interval of 0.73 to 0.86.

Check the answer#

The network's own tables give the true effect: John calls with probability 0.849 after a burglary and 0.056 without one. The difference, 0.793, is inside the interval.

Read the identification#

identify_effect answers a yes/no question before any estimation: can this graph and this data recover the effect?

identification = identify_effect(graph, "Burglary", "JohnCalls")
print(identification)

It returns the formula it will estimate. Here that is a sum over Earthquake: the canonical adjustment set, every ancestor of the treatment or the outcome that the treatment does not affect.

treated="True" names the treated level, because the columns hold the strings "True" and "False".

Next steps#