Complementary Events

Deep Breakdown

What it tells me (Situation Awareness): The complement of event $A$ is "everything except $A$." Since $A$ and $A^c$ together cover the entire sample space, their probabilities must sum to 1. This is one of the most-used probability identities in practice: it is often easier to compute $P(A^c)$ and subtract from 1 than to compute $P(A)$ directly.

Flowchart

Real-world use

Where it is used

"At least one" probability calculations: $P(\text{at least one success}) = 1 - P(\text{no successes})$. Also: $P(\text{churn}) = 1 - P(\text{retain})$, $P(\text{fraud detected}) = 1 - P(\text{missed})$.

Problem it solves

When computing $P(A)$ directly is complex (e.g., "at least one" requires summing many cases), computing $P(A^c)$ is often trivial.

How it solves the problem

Leverages $P(\Omega) = 1$ and the disjointness of $A$ and $A^c$: $P(A) + P(A^c) = 1$.

Example in a business

Probability that at least one of 10 email A/B tests detects a false positive (at $\alpha=0.05$ each) if all null hypotheses are true: $1 - (1-0.05)^{10} = 1 - 0.95^{10} \approx 0.401$. This motivates multiple testing corrections.

The formula, piece by piece

$P(A^c) = 1 - P(A)$

• $A^c$: complement of event $A$ (all outcomes NOT in $A$).
• $P(A)$: probability of event $A$.
• $1$: total probability from $P(\Omega) = 1$.

The formula in words

"The probability that $A$ does NOT occur equals 1 minus the probability that $A$ does occur." If $P(\text{churn}) = 0.15$, then $P(\text{retain}) = 1 - 0.15 = 0.85$.

Critical Concepts

• Requires $A$ and $A^c$ to be exhaustive (cover all outcomes) and mutually exclusive.
• "At least one" trick: $P(\geq 1 \text{ success in } n) = 1 - (1-p)^n$.
• False discovery rate calculations depend on this.
• In code: 1 - p — the complement is always this simple.

Step-by-step calculation (tiny example)

Step 1: Known probability $P(\text{click}) = 0.03$ (3% CTR).

Step 2: Apply complement $P(\text{no click}) = 1 - P(\text{click}) = 1 - 0.03 = 0.97$

Step 3: "At least one" extension $P(\text{at least one click in 5 impressions}) = 1 - 0.97^5 = 1 - 0.859 = 0.141$

Code examples

Beginner example (Python)

# Multiple testing: probability of at least one false positive
alpha = 0.05
n_tests = 10
p_at_least_one_fp = 1 - (1 - alpha) ** n_tests
print(f"P(>= 1 false positive in {n_tests} tests) = {p_at_least_one_fp:.3f}")
# 0.401 - motivates Bonferroni correction

Beginner example (R)

alpha <- 0.05
n_tests <- 10
p_at_least_one_fp <- 1 - (1 - alpha)^n_tests
cat(sprintf("P(>= 1 false positive in %d tests) = %.3f\n", n_tests, p_at_least_one_fp))

Professional example (Python)

import numpy as np

def family_wise_error_rate( per_test_alpha: float, n_tests: int ) -> float:
    """Compute FWER: P(at least one false positive) = 1 - (1 - alpha)^n. Uses complement rule: P(A^c) = 1 - P(A). Args: per_test_alpha: Type I error rate per individual test. n_tests: Number of independent tests performed. Returns: Family-wise error rate. """
    return 1.0 - (1.0 - per_test_alpha) ** n_tests

def bonferroni_correction(target_fwer: float, n_tests: int) -> float:
    """Return per-test alpha to achieve target FWER via Bonferroni."""
    return target_fwer / n_tests   # A/B testing multiple comparisons analysis
ALPHA = 0.05
test_counts = [1, 5, 10, 20, 50, 100]
print(f"Complement rule: FWER analysis at alpha={ALPHA}")
print(f"{'n_tests':>10} {'FWER':>12} {'Bonferroni alpha':>18}")
for n in test_counts:
    fwer = family_wise_error_rate(ALPHA, n)
bonf_alpha = bonferroni_correction(ALPHA, n)
print(f"{n:>10} {fwer:>12.4f} {bonf_alpha:>18.6f}")

Source: AIO p. 84

#probability-statistics#probability-theory#axioms-and-basic-probability