Turning Multi-Stage Narratives Into Events And Random Variables

Start by Translating the Story

A multi-stage probability problem can look like a calculation problem before it is clear what should be calculated. A device may be inspected, tested, and then assigned a status; a claim may be reviewed and then approved or denied. The stages describe how outcomes occur, but the question may ask about a particular sequence, a count of outcomes, a threshold, or an expected amount. The first task is to express that request precisely.

Keep three objects distinct. An event is a set of outcomes, such as “the first inspection passes.” A random variable assigns a number to each outcome, such as the number of stages that pass. A realized value is the number the random variable takes in one observed outcome. Confusing these objects can lead to using a correct formula on the wrong target.

Definition: An event is a collection of outcomes to which a probability can be assigned. A random variable is a numerical function of the outcome of a random experiment. For an event $A$, its indicator random variable is $I_A$, defined by $I_A=1$ when $A$ occurs and $I_A=0$ otherwise.

For sequential events $A$ and $B$, the event “$A$ occurs and then $B$ occurs” is $A\cap B$. If $P(A)>0$, the chain rule gives

P(A\cap B)=P(A)P(B\mid A).

For three stages, if the relevant conditioning events have positive probability, the same rule can be applied repeatedly:

P(A\cap B\cap C)=P(A)P(B\mid A)P(C\mid A\cap B).

The conditional probability at a stage must match the path leading to that stage. In general, $P(C\mid A\cap B)$ need not equal $P(C\mid A)$, and neither is automatically equal to $P(C)$.

A Reliable Translation Process

Before calculating, write down the stages, name the events, and translate the question into an event or random variable. Then use the given conditional probabilities along the paths that make up the target. This separates interpretation from arithmetic and makes it easier to detect missing cases.

List the stages and outcomes.
Record what can happen at each stage, including the order in which stages occur.

Name the events.
Use one event for each outcome of interest at a stage, and state what the event means in context.

Translate the question.
Write the target as an intersection, union, complement, indicator, or function of indicators, as appropriate.

Match probabilities to paths.
Use conditional probabilities for the history specified by each path. Add probabilities only for disjoint paths.

Check the result.
Confirm that every relevant path is included exactly once and that the answer is between 0 and 1 when it is a probability.

Worked Examples

Worked Example: Identifying the Exact Stage Outcomes

A hypothetical parcel facility first checks whether a parcel passes a weight inspection. Let $A$ be the event that it passes, with $P(A)=0.72$. A second sensor then records an alert; let $B$ be the event that an alert occurs. The sensor has $P(B\mid A)=0.85$ and $P(B\mid A^c)=0.25$. Find (a) the probability that the parcel passes inspection and has no alert, (b) the probability of exactly one of $A$ and $B$, and (c) the probability of an alert. The sensor result is specified by its conditional probabilities; do not assume the two events are independent.

Method. Translate each request into event notation. Use the chain rule for each path, and add probabilities only when the paths are disjoint.

(a) “Passes inspection and has no alert” is $A\cap B^c$. Since $P(B^c\mid A)=1-0.85=0.15$,

P(A\cap B^c)=P(A)P(B^c\mid A)=0.72(0.15)=0.1080.

(b) “Exactly one” means either $A\cap B^c$ or $A^c\cap B$. These two events cannot occur together, so their probabilities add. First, $P(A^c)=1-0.72=0.28$, and

P(A^c\cap B)=P(A^c)P(B\mid A^c)=0.28(0.25)=0.0700.

Therefore,

P(\text{exactly one of }A,B)=0.1080+0.0700=0.1780.

(c) An alert can occur along either of the disjoint paths $A\cap B$ and $A^c\cap B$. The first path has probability $0.72(0.85)=0.6120$; the second has probability $0.0700$, calculated above. Thus,

$$ P(B)=0.6120+0.0700=0.6820. $$

Check. The four paths $A\cap B$, $A\cap B^c$, $A^c\cap B$, and $A^c\cap B^c$ have probabilities $0.6120$, $0.1080$, $0.0700$, and $0.2100$. They sum to $1.0000$. A common mistake is to use $0.72(0.85)$ as the probability of any alert; that counts only alerts after a pass and omits the $A^c\cap B$ path.

Worked Example: Defining and Using a Count Random Variable

A hypothetical three-stage quality check records a pass or fail at each stage. Let $A$, $B$, and $C$ denote passing stages 1, 2, and 3, respectively. The model specifies $P(A)=0.60$, $P(B\mid A)=0.70$, and $P(B\mid A^c)=0.20$. The final-stage pass probabilities depend on both earlier results: $P(C\mid A\cap B)=0.90$, $P(C\mid A\cap B^c)=0.40$, $P(C\mid A^c\cap B)=0.50$, and $P(C\mid A^c\cap B^c)=0.10$. Find the probability of exactly two passes and the expected number of passes.

Method. Define one indicator for each pass and add them to form a count. For the probability of exactly two passes, identify the three possible pass/fail paths and calculate each path in stage order.

Let $K=I_A+I_B+I_C$. Each indicator equals 1 for a pass at its stage and 0 otherwise, so $K$ is the number of passes. Exactly two passes occur on $A\cap B\cap C^c$, $A\cap B^c\cap C$, or $A^c\cap B\cap C$. These paths are mutually exclusive.

For the first path, the conditional failure probability at stage 3 is $1-0.90=0.10$:

P(A\cap B\cap C^c)=0.60(0.70)(0.10)=0.0420.

For the second path, $P(B^c\mid A)=1-0.70=0.30$ and $P(C\mid A\cap B^c)=0.40$:

P(A\cap B^c\cap C)=0.60(0.30)(0.40)=0.0720.

For the third path, $P(A^c)=0.40$:

P(A^c\cap B\cap C)=0.40(0.20)(0.50)=0.0400.

Adding the three disjoint path probabilities gives

$$ P(K=2)=0.0420+0.0720+0.0400=0.1540. $$

For the expectation, use $E[I_A]=P(A)$ and linearity of expectation. This rule does not require the indicators to be independent. First calculate the marginal probabilities of the stage passes:

$$ P(B)=0.60(0.70)+0.40(0.20)=0.5000, $$

and, by adding the four disjoint histories through stage 2,

$$ P(C)=0.60(0.70)(0.90)+0.60(0.30)(0.40) +0.40(0.20)(0.50)+0.40(0.80)(0.10)=0.5220. $$

Therefore,

$$ E[K]=E[I_A]+E[I_B]+E[I_C]=0.6000+0.5000+0.5220=1.6220. $$

Check. The count is between 0 and 3, and an expected count of $1.6220$ is within that range. A common mistake is to multiply the stage pass probabilities as though all stages were independent. The model instead gives stage-specific conditional probabilities; the path calculations preserve them.

Worked Example: Turning a Sequential Cost into a Random Variable

A hypothetical service team first checks whether a warning is present. Let $W$ denote a warning, with $P(W)=0.25$. A follow-up visit is made with probability $0.80$ when a warning is present and probability $0.10$ when it is absent. Let $M$ denote that a visit is made. The total cost, in dollars, is $C=12I_W+30I_M$: the warning assessment costs \$12 if there is a warning, and a visit costs \$30 if it occurs. Find $E[C]$ and $P(C>30)$.

Method. Express the cost as a sum of event indicators for its expectation. For the strict threshold probability, list the combinations of $W$ and $M$ and compare each cost with 30.

First, $P(M\mid W)=0.80$ and $P(M\mid W^c)=0.10$. Thus,

P(M)=P(W)P(M\mid W)+P(W^c)P(M\mid W^c) =0.25(0.80)+0.75(0.10)=0.275.

By linearity of expectation,

E[C]=12E[I_W]+30E[I_M] =12P(W)+30P(M)=12(0.25)+30(0.275)=\$11.25.

To find $P(C>30)$, evaluate the four possible combinations. If both $W$ and $M$ occur, $C=12+30=42$, and the path probability is $0.25(0.80)=0.2000$. If $W$ occurs without $M$, $C=12$; if $M$ occurs without $W$, $C=30$; and if neither occurs, $C=0$. Only the first combination has cost strictly greater than 30. Therefore,

P(C>30)=P(W\cap M)=0.25(0.80)=0.2000.

Check. The four combination probabilities are $0.2000$, $0.0500$, $0.0750$, and $0.6750$, which sum to 1. A common mistake is to assume $W$ and $M$ are independent and calculate $P(W\cap M)$ as $P(W)P(M)=0.25(0.275)=0.06875$. That conflicts with the stated conditional model: the correct joint probability is $P(W)P(M\mid W)=0.25(0.80)=0.2000$. Also note that $C>30$ differs from $C\geq30$; the latter would include the $C=30$ case.

Why the Translation Matters

The same narrative can generate several distinct targets. In the quality-check example, $K=2$ asks for the probability of a count equaling two, whereas $E[K]$ asks for the average count over repeated runs. In the service-cost example, $E[C]$ uses the chance of each cost component, while $P(C>30)$ requires identifying which combinations exceed the specified threshold. These questions are not interchangeable, even though they refer to the same stages.

A useful derivation behind these calculations is the chain rule. For events $A$ and $B$ with $P(A)>0$, the definition of conditional probability is

P(B\mid A)=\frac{P(A\cap B)}{P(A)}.

Multiplying both sides by $P(A)$ gives $P(A\cap B)=P(A)P(B\mid A)$. Applying that identity to the event $A\cap B$ and the next-stage event $C$, when $P(A\cap B)>0$, gives

P(A\cap B\cap C)=P(A\cap B)P(C\mid A\cap B) =P(A)P(B\mid A)P(C\mid A\cap B).

This explains why a path probability uses the conditional probability associated with its history. If a conditioning event has probability zero, that conditional probability is not defined by this ratio; do not apply the formula to that path.

For an indicator $I_A$, its expectation is $1\cdot P(A)+0\cdot P(A^c)=P(A)$. Consequently, when a count is defined as a sum of indicators, its expectation is the sum of the corresponding event probabilities. This remains true when the events are dependent. The crucial discipline is to preserve the dependence in joint probabilities, rather than inserting an independence assumption that the model does not provide.

Key takeaway: Before choosing a calculation, write the target in mathematical form. Use intersections for joint occurrence, disjoint unions for alternative paths, indicators for counts or costs, and the conditional probabilities that match each path. This translation-first approach is aligned with the SOA Exam P (Probability) syllabus.

Check Your Understanding

For each problem, translate the requested quantity before calculating. Probabilities are rounded to four decimal places where needed.

A review has a first-stage approval event $A$, with $P(A)=0.40$. A second-stage flag $B$ has $P(B\mid A)=0.75$. What is the probability of approval with no flag? (Answer: $0.1000$)
A two-stage test has $P(A)=0.50$, $P(B\mid A)=0.60$, and $P(B\mid A^c)=0.40$. What is $P(A\cap B)$? (Answer: $0.3000$)
A two-step process has $P(A)=0.30$, $P(B\mid A)=0.50$, and $P(B\mid A^c)=0.20$. What is the probability exactly one of $A$ and $B$ occurs? (Answer: $0.2900$)
A hypothetical inspection costs \$8 when event $A$ occurs and \$20 when event $B$ occurs, so $C=8I_A+20I_B$. Given $P(A)=0.40$, $P(B)=0.30$, and $P(A\cap B)=0.10$, find $E[C]$ and $P(C=28)$. (Answer: $E[C]=\$9.20$, $P(C=28)=0.1000$)
Three stages have pass events $A,B,C$, with $P(A)=0.50$, $P(B\mid A)=0.80$, $P(B\mid A^c)=0.20$, $P(C\mid A\cap B)=0.50$, $P(C\mid A\cap B^c)=0.25$, $P(C\mid A^c\cap B)=0.50$, and $P(C\mid A^c\cap B^c)=0.10$. Define $K=I_A+I_B+I_C$. Find $P(K=3)$. (Answer: $0.2000$)

What You'll Learn