Conditioning on Partitions and Total Expectation

Why Split an Expectation Into Cases?

An overall average can be difficult to calculate directly when a random outcome is generated in stages. A claim count may depend on the type of year; a service cost may depend on which desk handles a case; or a payment may depend on a category selected before the payment is determined. In each situation, the categories provide a way to organize the calculation: find the mean within each category, then weight that mean by the category's probability.

The categories must form a partition: they cannot overlap, and together they must include every possible outcome. The resulting weighted average is useful precisely because it does not require the conditional means to be equal. It also does not require independence between the category and the random variable.

Definition: A finite collection of events $A_1,\ldots,A_k$ is a partition of the sample space if the events are mutually exclusive, their union is the entire sample space, and each has positive probability. If $X$ has a finite expectation, then $$ E[X]=\sum_{i=1}^{k} E[X\mid A_i]P(A_i). $$ Each conditional mean $E[X\mid A_i]$ is multiplied by the probability of the event on which it is conditioned. The probabilities $P(A_i)$ sum to 1.

The identity applies to discrete or continuous $X$, provided $X$ is integrable—that is, $E[|X|]$ is finite. Conditional means can be positive, zero, or negative, so $X$ need not be a nonnegative payment. If an event in a proposed partition has probability zero, its conditional mean is not defined by the usual event-conditioning formula; such events should be omitted from the calculation.

Why the Weighted-Average Formula Holds

The key fact is that exactly one event in a partition occurs. For each outcome, the value of $X$ can therefore be assigned to one and only one partition event. In indicator notation, where $1_{A_i}$ is 1 when $A_i$ occurs and 0 otherwise, this gives

X=\sum_{i=1}^{k}X1_{A_i}.

Take expectations on both sides. Linearity of expectation gives the sum of the expectations of the terms. For an event $A_i$ with positive probability, the definition of conditional expectation on an event gives $E[X1_{A_i}]=P(A_i)E[X\mid A_i]$. Substitution yields the result:

\begin{aligned} E[X] &=E\left[\sum_{i=1}^{k}X1_{A_i}\right]\\ &=\sum_{i=1}^{k}E[X1_{A_i}]\\ &=\sum_{i=1}^{k}P(A_i)E[X\mid A_i]. \end{aligned}

This argument also explains what the formula does not require: the events need not be independent of $X$. In fact, the conditional means are allowed to differ because the distribution of $X$ may change from one partition event to another.

Using the Formula in Mixture Models

A mixture model describes a random variable whose distribution depends on a randomly selected category. For example, a population may contain several types of policyholders, or a claim count may follow one distribution in a low-activity year and another in a high-activity year. The unconditional distribution is formed by combining the conditional distributions using the category probabilities. For expectations, the same principle reduces to combining the conditional means.

Worked Example: Expected Cost Under a Poisson Mixture

Method. Partition by the type of year, calculate the conditional expected claim count in each case, and weight by the probabilities of those cases. A hypothetical insurer models its annual claim count $N$ as Poisson with parameter $0.5$ in a low-activity year and Poisson with parameter $1.5$ in a high-activity year. The probability of a low-activity year is $0.80$, so the probability of a high-activity year is $0.20$. Each claim costs exactly $\$200$, and annual cost is $X=200N$. There is no discounting or mortality assumption in this one-year model.

Let $L$ denote a low-activity year and $H$ a high-activity year. These events form a partition, and the Poisson parameter is the conditional mean of $N$:

E[N\mid L]=0.5,\qquad E[N\mid H]=1.5.

First calculate the unconditional expected number of claims:

\begin{aligned} E[N] &=E[N\mid L]P(L)+E[N\mid H]P(H)\\ &=(0.5)(0.80)+(1.5)(0.20)\\ &=0.40+0.30\\ &=0.70\text{ claims}. \end{aligned}

Because each claim costs $\$200$, the conditional expected costs are $\$200(0.5)=\$100$ in a low-activity year and $\$200(1.5)=\$300$ in a high-activity year. Weighting these conditional costs gives

E[X]=(0.80)(\$100)+(0.20)(\$300)=\$80+\$60=\$140.

Check and mistake avoided. The weights $0.80$ and $0.20$ sum to 1, and the higher conditional mean receives the smaller, but nonzero, weight. Using $0.80$ for both states would ignore the fact that a high-activity year occurs with probability $0.20$. The expected annual cost is $\$140$.

Staged Outcomes: Condition on the First Stage

In a staged experiment, the first stage determines which conditional model applies to the next stage. To find the overall mean, use the possible first-stage outcomes as the partition. Within each case, calculate the conditional mean using the conditional probabilities and values supplied for that case. Then weight those means using the probabilities of the first-stage outcomes—not the probabilities of individual second-stage outcomes as though they were unconditional.

Worked Example: Expected Cost After a Case Is Routed

Method. Condition first on the desk assignment, then calculate the expected cost for each desk. In a hypothetical claims operation, a case is routed to desk A with probability $0.70$ and to desk B with probability $0.30$. At desk A, a case is complex with conditional probability $0.20$; at desk B, it is complex with conditional probability $0.60$. A simple case costs $\$30$, and a complex case costs $\$90$, regardless of desk. Let $X$ be the processing cost. No investment, interest, or mortality assumption is involved.

The routing events A and B are mutually exclusive and exhaustive. At desk A, the conditional probability of a simple case is $1-0.20=0.80$, so

\begin{aligned} E[X\mid A] &=(0.80)(\$30)+(0.20)(\$90)\\ &=\$24+\$18\\ &=\$42. \end{aligned}

At desk B, the conditional probability of a simple case is $1-0.60=0.40$, giving

\begin{aligned} E[X\mid B] &=(0.40)(\$30)+(0.60)(\$90)\\ &=\$12+\$54\\ &=\$66. \end{aligned}

Now use the desk probabilities to combine the conditional means:

\begin{aligned} E[X] &=P(A)E[X\mid A]+P(B)E[X\mid B]\\ &=(0.70)(\$42)+(0.30)(\$66)\\ &=\$29.40+\$19.80\\ &=\$49.20. \end{aligned}

Check and mistake avoided. The probabilities of complex cases, $0.20$ and $0.60$, are conditional on the desk and are used to calculate each desk's mean. They are not the weights for combining desk means. The weights for that final step are $P(A)=0.70$ and $P(B)=0.30$. The result, $\$49.20$, lies between the two desk-specific means, $\$42$ and $\$66$, as it should.

Conditional Payment Distributions

The conditional mean within a partition event may itself require an expectation calculation. A useful procedure is to separate the two levels: first calculate each conditional mean from the possible values and their conditional probabilities; next combine those means using the probabilities of the partition events. This is particularly useful when a payment distribution changes across categories.

Worked Example: Expected Payment Across Severity Categories

Method. Compute the mean payment within each severity category, then take the probability-weighted average across categories. Consider a hypothetical warranty program with one payment $X$ per covered incident. Incidents are classified as minor with probability $0.50$, moderate with probability $0.30$, and severe with probability $0.20$. Conditional on a minor incident, the payment is $\$0$ with probability $0.20$ and $\$100$ with probability $0.80$. Conditional on a moderate incident, it is $\$100$ or $\$300$, each with probability $0.50$. A severe incident always produces a $\$500$ payment. These are hypothetical payment assumptions; no interest or mortality adjustment is made.

For a minor incident, the conditional mean is

E[X\mid \text{minor}]=(0.20)(\$0)+(0.80)(\$100)=\$80.

For a moderate incident, the conditional mean is

E[X\mid \text{moderate}]=(0.50)(\$100)+(0.50)(\$300)=\$50+\$150=\$200.

The severe-incident payment is fixed, so $E[X\mid \text{severe}]=\$500$. The three categories form a partition because they are mutually exclusive and their probabilities sum to $0.50+0.30+0.20=1.00$. Therefore,

\begin{aligned} E[X] &=(0.50)(\$80)+(0.30)(\$200)+(0.20)(\$500)\\ &=\$40+\$60+\$100\\ &=\$200. \end{aligned}

Check and mistake avoided. The $\$80$ and $\$200$ are conditional means, not individual payment outcomes. Weighting the original payment values directly by the category probabilities would mix conditional probabilities with unconditional probabilities. The two-stage calculation gives an expected payment of $\$200$.

Checks That Prevent Weighting Errors

Before using a partition, identify what each probability is conditional on. The outer weights must be the probabilities of the partition events. Probabilities inside a conditional distribution belong in the calculation of its conditional mean; they do not replace the outer weights. A reliable calculation therefore has two visible stages whenever there are two stages in the model.

Check the partition. Confirm that the events are mutually exclusive, cover all possibilities, and have probabilities summing to 1.
Label each mean. Write $E[X\mid A_i]$ next to the event it describes before combining any values.
Use the right weights. Multiply each conditional mean by $P(A_i)$, not by a probability conditional on a different event.
Check the scale. If $X$ is a dollar amount, every conditional mean and the final expectation must also be in dollars.
Use a reasonableness check. For a finite partition, the overall mean is a probability-weighted average of the conditional means. It must lie between the smallest and largest conditional means.

This method is aligned with the SOA VEE Mathematical Statistics syllabus because it is a core tool for expectations under conditional models. It is also useful in actuarial calculations whenever a population, claim process, or payment mechanism is described in distinct categories. The central discipline is to keep the outer partition probabilities separate from the conditional calculations inside each category.

Key takeaway: Partition the possible cases, calculate $E[X\mid A_i]$ within each case, and then weight those means by $P(A_i)$. The partition probabilities—not the conditional probabilities used inside each case—form the final weighted average.

Check Your Understanding

For each problem, identify the partition and apply the weighted-mean calculation. All scenarios are hypothetical.

A hypothetical annual claim count is Poisson with parameter $0.4$ in a quiet year, which occurs with probability $0.75$, and Poisson with parameter $1.2$ otherwise. Each claim costs $\$150$. Find the expected annual claim cost. (Answer: $\$90$)
A case is routed to desk C with probability $0.60$ and desk D with probability $0.40$. Conditional on C, the case is complex with probability $0.25$; conditional on D, it is complex with probability $0.50$. Simple and complex cases cost $\$40$ and $\$100$, respectively. Find the expected cost. (Answer: $\$61$)
A hypothetical portfolio category is type 1 with probability $0.30$ and type 2 with probability $0.70$. Conditional on type 1, an annual signed gain $X$, measured in thousands of dollars, has mean $12$. Conditional on type 2, its mean is $8$. Find $E[X]$. (Answer: $9.2$ thousand dollars)
A warranty payment is $\$0$, $\$40$, or $\$100$ according to a category. The category probabilities are $0.20$, $0.50$, and $0.30$, and the corresponding conditional expected payments are $\$0$, $\$40$, and $\$100$. Find the overall expected payment. (Answer: $\$50$)
A hypothetical repair payment $X$ has $X\mid A$ uniform on $[0,10]$ thousands of dollars and $X\mid B$ uniform on $[4,16]$ thousands of dollars. The category probabilities are $P(A)=0.40$ and $P(B)=0.60$. Use the conditional means of the uniform distributions to find $E[X]$. (Answer: $8$ thousand dollars)

What You'll Learn