Conditional Independence and Conditioning Effects

Why Conditioning Can Change Independence

Two events can be independent in the full probability model but dependent after attention is restricted to a particular condition. The reverse is also possible: events that are dependent overall may be independent within a specified group. For actuarial work, this distinction matters whenever probabilities are calculated for a selected risk class, operating state, or other restricted population.

The key is to use probabilities from the same restricted sample space throughout the calculation. Conditional independence does not mean merely that two events are individually possible under a condition; it means their joint conditional probability factors into the product of their conditional probabilities.

Definition: Let $A$, $B$, and $C$ be events with $P(C)>0$. Events $A$ and $B$ are conditionally independent given $C$ if $$ P(A\cap B\mid C)=P(A\mid C)P(B\mid C). $$ Equivalently, under the conditional probability model restricted to $C$, the events $A$ and $B$ are independent.

Conditional Joint Probabilities and the Factorization Rule

The condition $P(C)>0$ is needed so that probabilities given $C$ are defined. From the definition of conditional probability, $$ P(A\cap B\mid C)=\frac{P(A\cap B\cap C)}{P(C)}, \qquad P(A\mid C)=\frac{P(A\cap C)}{P(C)}, \qquad P(B\mid C)=\frac{P(B\cap C)}{P(C)}. $$ A direct test of conditional independence compares the first quantity with the product of the other two. In particular, the unconditional probabilities $P(A)$, $P(B)$, and $P(A\cap B)$ are not the quantities to compare unless the condition is the whole sample space.

There is also a useful equivalent test when $P(B\cap C)>0$: conditional independence holds exactly when $P(A\mid B\cap C)=P(A\mid C)$. The condition $B\cap C$ gives the probability of $A$ after learning both that $B$ occurred and that $C$ holds. If that extra information about $B$ makes no difference, then $A$ and $B$ are conditionally independent within $C$.

Why the Conditional Product Rule Holds

The factorization criterion follows from conditional probability, not from an additional assumption about the events. If $P(C)>0$, then $$ P(A\cap B\mid C)=P(A\mid C)P(B\mid A\cap C) $$ whenever $P(A\cap C)>0$. Thus, if learning $A$ does not change the conditional probability of $B$ within $C$, so that $P(B\mid A\cap C)=P(B\mid C)$, then $$ P(A\cap B\mid C)=P(A\mid C)P(B\mid C). $$ Conversely, if the product rule holds and $P(A\mid C)>0$, dividing both sides by $P(A\mid C)$ gives $P(B\mid A\cap C)=P(B\mid C)$. When a conditional probability in this division has a zero denominator, use the factorization definition directly; it remains the appropriate test.

Worked Example: Finding a Conditional Joint Probability

In a hypothetical monitoring model, $C$ is the event that a device is operating in a high-demand state, with $P(C)=0.40$. Given that state, event $A$ is a warning from one sensor and event $B$ is a warning from a second sensor. The model specifies $P(A\mid C)=0.30$, $P(B\mid C)=0.20$, and that the warnings are conditionally independent given $C$. Find the probability of both warnings given $C$, the probability of all three events $A$, $B$, and $C$, and $P(A\mid B\cap C)$.

Method and assumption. Use the conditional independence factorization for the joint probability given $C$. The condition has positive probability, and the model explicitly supplies conditional independence; no assumption about independence outside $C$ is needed.

First calculate the conditional joint probability: $$ P(A\cap B\mid C)=P(A\mid C)P(B\mid C) =(0.30)(0.20)=0.06. $$

To obtain the probability of $A$, $B$, and $C$ together, multiply by $P(C)$: $$ P(A\cap B\cap C)=P(A\cap B\mid C)P(C) =(0.06)(0.40)=0.024. $$

Next, $P(B\cap C)=P(B\mid C)P(C)=(0.20)(0.40)=0.08$. Therefore, $$ P(A\mid B\cap C)=\frac{P(A\cap B\cap C)}{P(B\cap C)} =\frac{0.024}{0.08}=0.30. $$ This agrees with $P(A\mid C)=0.30$, as conditional independence requires. Common mistake avoided: multiplying the conditional joint probability by $P(C)$ is necessary to get the unconditional probability of all three events; $0.06$ alone is not $P(A\cap B\cap C)$.

Conditioning Can Create Dependence

Even if $A$ and $B$ are independent before conditioning, restricting attention to $C$ can make them dependent. A common mechanism is conditioning on an event whose occurrence depends on both $A$ and $B$. The restricted cases then have a different balance of outcomes from the full model.

Worked Example: Independent Alerts Become Dependent Given an Alert

In a hypothetical system, two separate alert mechanisms generate events $A$ and $B$ independently. Their probabilities are $P(A)=0.40$ and $P(B)=0.50$. Let $C=A\cup B$, the event that at least one alert occurs. Determine whether $A$ and $B$ are conditionally independent given $C$.

Method and assumption. The model states that $A$ and $B$ are independent in the full sample space. Calculate $P(C)$, the conditional marginals, and the conditional joint probability, then compare the joint probability with the product of the marginals.

By independence, $P(A\cap B)=(0.40)(0.50)=0.20$. The probability of at least one alert is $$ P(C)=P(A)+P(B)-P(A\cap B)=0.40+0.50-0.20=0.70. $$ Because $A\subseteq C$ and $B\subseteq C$, $$ P(A\mid C)=\frac{0.40}{0.70}=\frac{4}{7}, \qquad P(B\mid C)=\frac{0.50}{0.70}=\frac{5}{7}. $$ The conditional joint probability is $$ P(A\cap B\mid C)=\frac{P(A\cap B)}{P(C)} =\frac{0.20}{0.70}=\frac{2}{7}. $$ The product of the conditional marginals is $$ P(A\mid C)P(B\mid C)=\frac{4}{7}\frac{5}{7}=\frac{20}{49}. $$ Since $2/7=14/49$, the conditional joint probability is not the product of the conditional marginals. Thus $A$ and $B$ are not conditionally independent given $C$.

Common mistake and correction. It would be wrong to use $P(A)P(B)=0.20$ as the product for the conditional independence test. The test uses $P(A\mid C)P(B\mid C)=20/49$, because the question restricts attention to cases in $C$. The alerts are independent overall but dependent among cases with at least one alert.

Conditioning Can Reveal Independence

Conditioning can also remove an apparent overall dependence. In a mixture of different operating conditions, the combined probabilities may not factor even when they do factor within one condition. The next example uses a finite model so each probability can be checked by counting equally likely outcomes.

Worked Example: Dependence Overall but Independence Within a Condition

A hypothetical system has eight equally likely scenarios. In four scenarios, condition $C$ holds; those four have the four possible combinations of events $A$ and $B$, each occurring once. In the other four scenarios, $C$ does not hold: two have both $A$ and $B$, and two have neither. Determine whether $A$ and $B$ are independent given $C$, and whether they are independent overall.

Method and assumption. Count the scenarios in each group. Since all eight scenarios are equally likely, each has probability $1/8$, and $P(C)=4/8=1/2$.

Within $C$, one of the four scenarios has both events, two have $A$, and two have $B$. Therefore, $$ P(A\mid C)=\frac{2}{4}=\frac12,\qquad P(B\mid C)=\frac{2}{4}=\frac12,\qquad P(A\cap B\mid C)=\frac{1}{4}. $$ Since $(1/2)(1/2)=1/4$, the events are conditionally independent given $C$.

For the overall probabilities, $A$ occurs in two scenarios within $C$ and two outside $C$, so $P(A)=4/8=1/2$. The same count gives $P(B)=1/2$. Both events occur in one scenario within $C$ and two outside $C$, so $P(A\cap B)=3/8$. But $$ P(A)P(B)=\frac12\frac12=\frac14\neq\frac38=P(A\cap B). $$ Thus the events are dependent overall, despite being independent given $C$. Common mistake avoided: conditional independence in one group does not establish independence across all groups; the overall joint probability must be checked separately.

How to Choose and Apply the Test

When a problem names a condition, write it explicitly before multiplying probabilities. A reliable calculation follows these steps:

Identify the restricted event.
Write down $C$ and verify that $P(C)>0$, so conditional probabilities given $C$ are defined.

Find probabilities within that condition.
Use $P(A\mid C)$, $P(B\mid C)$, and $P(A\cap B\mid C)$. If starting from probabilities in the full model, divide the relevant intersections with $C$ by $P(C)$.

Compare the two sides.
Check whether $P(A\cap B\mid C)=P(A\mid C)P(B\mid C)$. Equality means conditional independence given $C$; inequality means conditional dependence.

This test concerns the named condition only. It does not imply that $A$ and $B$ are independent without the condition, nor does unconditional independence guarantee conditional independence for every possible $C$. A useful check, when $P(B\cap C)>0$, is whether $P(A\mid B\cap C)$ equals $P(A\mid C)$. A difference shows that learning $B$ changes the probability of $A$ within the restricted group.

Common Mistakes and Why the Distinction Matters

Mixing conditional and unconditional probabilities. For independence given $C$, compare probabilities given $C$; do not substitute $P(A)$ or $P(B)$.
Assuming independence is preserved by conditioning. The alert example shows that conditioning on a union can make previously independent events dependent.
Assuming dependence must remain. The finite-scenario example shows that events dependent overall can be independent within a particular condition.
Forgetting the condition's probability. When converting a conditional joint probability to an unconditional intersection, use $P(A\cap B\cap C)=P(A\cap B\mid C)P(C)$.
Declaring independence from one matching quantity. The required equality is between the conditional joint probability and the product of both conditional marginals.

In risk analysis, the condition may describe a class or state that changes the relevant probabilities. A calculation that combines probabilities from different conditions can misstate how events relate within the group being studied. The essential habit is to define the condition, stay within its restricted probability model, and test the factorization there. This topic is aligned with the SOA VEE Mathematical Statistics syllabus.

Check Your Understanding

For each problem, use probabilities from the stated condition when testing conditional independence.

In a hypothetical model, $P(C)=0.30$, $P(A\mid C)=0.40$, and $P(B\mid C)=0.25$, and $A$ and $B$ are conditionally independent given $C$. Find $P(A\cap B\mid C)$ and $P(A\cap B\cap C)$. (Answer: $0.1000$ and $0.0300$, respectively.)
Events $A$ and $B$ are independent with $P(A)=0.30$ and $P(B)=0.60$. Let $C=A\cup B$. Find $P(C)$, $P(A\cap B\mid C)$, and decide whether $A$ and $B$ are conditionally independent given $C$. (Answer: $P(C)=0.72$, $P(A\cap B\mid C)=0.25$, and they are not conditionally independent because $P(A\mid C)P(B\mid C)=25/72\neq0.25$.)
A condition $C$ has probability $0.40$. Given $C$, events $A$ and $B$ are conditionally independent with $P(A\mid C)=0.50$ and $P(B\mid C)=0.30$. Find $P(A\cap B\cap C)$ and $P(A\mid B\cap C)$. (Answer: $0.0600$ and $0.5000$, respectively.)
A two-group model has $P(C)=0.25$. Given $C$, $P(A\mid C)=0.20$, $P(B\mid C)=0.40$, and the events are conditionally independent. Given $C^c$, $P(A\mid C^c)=0.60$, $P(B\mid C^c)=0.50$, and the events are conditionally independent. Are $A$ and $B$ independent overall? (Answer: No; $P(A)=0.50$, $P(B)=0.475$, $P(A\cap B)=0.245$, while $P(A)P(B)=0.2375$.)
Events $A$ and $B$ are independent with $P(A)=0.20$ and $P(B)=0.70$. Let $C=A\cup B$. Compare $P(A\cap B\mid C)$ with $P(A\mid C)P(B\mid C)$. (Answer: $P(A\cap B\mid C)=7/38$, while $P(A\mid C)P(B\mid C)=175/722$; they are not conditionally independent given $C$.)

What You'll Learn