Conditioning Strategies for Expected Values

Why the Choice of What to Condition On Matters

An expected value can be hard to calculate directly when the outcome depends on several stages, a hidden category, or a changing set of possibilities. Often the calculation becomes manageable after dividing the situation into cases. The key is to choose cases that simplify what remains: a first trial in a sequence, an unknown type that determines risk, or the first draw from a finite population.

Conditioning is not a new assumption about the process. It is a way to organize the same probability model into manageable pieces. Each piece gives a conditional expected value; the pieces must then be combined using the probabilities of the conditions. This approach is aligned with the SOA Exam P (Probability) syllabus.

Definition: Let $C_1,\ldots,C_k$ be mutually exclusive events whose union is the entire sample space. For any integrable random variable $X$, and with zero-probability events omitted, $$ E[X]=\sum_{j=1}^{k}P(C_j)E[X\mid C_j]. $$ The events form a partition: exactly one of them occurs. The formula says to weight each conditional mean by the probability of its condition.

The same idea applies when conditioning on a discrete random variable $Z$: use its possible values as the partition events. Thus $E[X]=\sum_z P(Z=z)E[X\mid Z=z]$. For a continuous conditioning variable, the corresponding formula is an integral: when $Z$ has density $f_Z$, $E[X]=\int E[X\mid Z=z]f_Z(z)\,dz$, provided the expectation exists and the conditional mean is defined in the model.

A General Justification

The partition formula does not require $X$ itself to be discrete. Let $\mathbf{1}_{C_j}$ equal 1 when $C_j$ occurs and 0 otherwise. Since exactly one partition event occurs, $\sum_j\mathbf{1}_{C_j}=1$. Multiplying by $X$, taking expectations, and using linearity gives

E[X] =E\!\left[X\sum_{j=1}^{k}\mathbf{1}_{C_j}\right] =\sum_{j=1}^{k}E[X\mathbf{1}_{C_j}] =\sum_{j=1}^{k}P(C_j)E[X\mid C_j].

The last equality follows from the definition $E[X\mid C_j]=E[X\mathbf{1}_{C_j}]/P(C_j)$ when $P(C_j)>0$. Integrability ensures the expectations exist. This argument works whether $X$ is discrete, continuous, or has a mixed distribution.

A Practical Strategy for Choosing a Condition

A useful condition is one that makes the conditional model simpler, not merely one that is easy to name. In sequential problems, condition on the first outcome or the state reached after the first stage. In a latent-type model, condition on the type because the type often makes the outcome probabilities known. In sampling without replacement, condition on the first draw because it determines the composition of the remaining population.

Name the target.
Identify the random variable whose expectation is required, including what it counts or measures.

Find a simplifying condition.
Choose a partition or conditioning variable that makes the remaining probabilities or outcomes straightforward.

Calculate each conditional mean.
Use the model after the condition is known; do not silently reuse unconditional probabilities if the condition changes them.

Weight and check.
Multiply each conditional mean by its condition probability, add the contributions, and check that the condition probabilities sum to 1.

Worked Example: Expected Attempts in a Changing Sequence

A hypothetical device is tested until it passes. It passes the first test with probability 0.4. If it fails the first test, it passes the second with conditional probability 0.7. If it fails both the first and second tests, it passes the third test with certainty. Let $N$ be the number of tests performed. Find $E[N]$.

Model and method. The stated probabilities are conditional on reaching the relevant test; no assumption of identical or independent test outcomes is needed. There is no interest or mortality assumption. Condition on the first test. This is useful because a pass ends the process, while a failure leaves a shorter sequential problem.

If the first test passes, $N=1$. If it fails, one test has already been used, and the expected number of additional tests from the second stage is

E[\text{additional tests}\mid\text{first test fails}] =0.7(1)+0.3(2)=0.7+0.6=1.3.

The two outcomes at the second stage are a pass, requiring one additional test, and a failure, requiring two additional tests because the third test is then certain to pass. Therefore, after a first-test failure, the expected total number of tests is $1+1.3=2.3$. Combining the first-test cases gives

$$ E[N]=0.4(1)+0.6(2.3)=0.4+1.38=1.78. $$

Check. The probabilities of stopping after tests 1, 2, and 3 are $0.4$, $0.6(0.7)=0.42$, and $0.6(0.3)=0.18$, respectively. They sum to $1$, and the direct check is $1(0.4)+2(0.42)+3(0.18)=0.4+0.84+0.54=1.78$ tests. The common mistake is to use the second-test pass probability as if it applied unconditionally; it applies only after a first-test failure.

Conditioning on a Latent Type

A latent type is a category that affects the outcome probabilities but is not necessarily observed when an expected value is first requested. If the type is known in the model, calculating a conditional expectation within each type can be much easier than constructing the overall distribution directly. The final answer must still account for how likely each type is.

Worked Example: Expected Payment With an Unknown Risk Type

A hypothetical policy covers a loss $X$. The insured has type A with probability 0.75 and type B with probability 0.25. Conditional on type A, $X$ is 0, 100, 400, or 800 dollars with probabilities 0.4, 0.3, 0.2, and 0.1, respectively. Conditional on type B, those losses have probabilities 0.2, 0.2, 0.3, and 0.3. The payment is $Y^P=\min(\max(X-d,0),u)$, with deductible $d=\$100$ and payment limit $u=\$300$. Find the expected payment. No interest or mortality assumption is involved.

Method. Condition on type, because the loss distribution is specified separately for each type. First apply the deductible and limit to each possible loss. The resulting payment is $\$0$ at losses of $\$0$ and $\$100$, and $\$300$ at losses of $\$400$ and $\$800$. In particular, the limit is a maximum payment after the deductible has been subtracted.

For type A, the probability of a $\$300$ payment is $0.2+0.1=0.3$. Thus

E[Y^P\mid A]=0.3(\$300)+0.7(\$0)=\$90.

For type B, the probability of a $\$300$ payment is $0.3+0.3=0.6$, so

E[Y^P\mid B]=0.6(\$300)+0.4(\$0)=\$180.

Now weight these means by the type probabilities:

E[Y^P]=0.75(\$90)+0.25(\$180)=\$67.50+\$45.00=\$112.50.

Check. The overall probability of a payment is $0.75(0.3)+0.25(0.6)=0.225+0.150=0.375$. Every positive payment is $\$300$, so the expectation is $0.375(\$300)=\$112.50$, as calculated. A common mistake is to average the two type-specific means equally, obtaining $(90+180)/2=\$135$. The types are not equally likely, so their conditional means must be weighted by 0.75 and 0.25.

Conditioning on the First Draw

In sampling without replacement, later draws depend on earlier ones because the population composition changes. Conditioning on the first draw makes the remaining composition explicit. The conditional probabilities for the next draw must use the reduced population, not the original proportions.

Worked Example: A Two-Draw Prize With an Order Bonus

A hypothetical prize is determined by drawing two tokens, without replacement, from a bag containing 4 green and 3 gold tokens. The base prize is $\$14$ if both tokens have the same color and $\$6$ if they have different colors. An additional $\$4$ is awarded if the first token is gold, regardless of the second token. Find the expected prize. No interest or mortality assumption is involved.

Method. Condition on the color of the first token. Given a green first token, 3 green and 3 gold tokens remain. The conditional expected prize is

E[Y\mid\text{first green}] =\frac{3}{6}(\$14)+\frac{3}{6}(\$6) =\$7+\$3=\$10.

Given a gold first token, 2 gold and 4 green tokens remain. A second gold token gives a same-color base prize plus the order bonus, or $\$14+\$4=\$18$. A second green token gives a different-color base prize plus the bonus, or $\$6+\$4=\$10$. Hence

E[Y\mid\text{first gold}] =\frac{2}{6}(\$18)+\frac{4}{6}(\$10) =\$6+\frac{\$20}{3}=\frac{\$38}{3}.

The first token is green with probability $4/7$ and gold with probability $3/7$. Therefore,

E[Y]=\frac{4}{7}(\$10)+\frac{3}{7}\left(\frac{\$38}{3}\right) =\frac{\$40}{7}+\frac{\$38}{7} =\frac{\$78}{7} \approx \$11.14.

The result is rounded to the nearest cent only at the final step. Check. The two first-token cases have probabilities $4/7$ and $3/7$, which sum to 1. The common mistake is to take an unweighted average of the two conditional means. That would treat the first colors as equally likely, although their probabilities are $4/7$ and $3/7$.

Common Traps and Checks

The strongest conditioning choice is the one that produces simple conditional calculations while preserving the full probability model. Before finalizing a result, check both the conditions and the weights.

Using the wrong weights. Conditional means are not automatically equally likely. Multiply each one by the probability of its condition.
Forgetting what the condition changes. After a draw without replacement, update the population. After observing a failure, use probabilities conditional on having reached that stage.
Confusing a conditional probability with an unconditional probability. A probability specified for a later stage may apply only given earlier outcomes.
Applying a payment rule in the wrong order. For a deductible followed by a payment limit, first calculate the amount above the deductible, then cap the payment at the limit.
Skipping a check. Confirm that partition probabilities sum to 1 and, when feasible, compare the conditional calculation with a direct calculation.

Conditioning is particularly useful when the target's direct distribution is cumbersome but the conditional distributions are simple. It does not make every calculation shorter: a poor condition can create more cases without simplifying any of them. Choose a condition that exposes the mechanism generating the outcome.

Check Your Understanding

For each problem, identify a useful condition before calculating the requested expectation.

A hypothetical inspection process uses one inspection if the item passes, and otherwise continues. It passes inspection 1 with probability 0.6. Given failure at inspection 1, it passes inspection 2 with probability 0.5; if it fails inspection 2, inspection 3 passes with certainty. Find the expected number of inspections. (Answer: 1.6000 inspections)
A hypothetical portfolio contains standard risks with probability 0.6 and elevated risks with probability 0.4. A standard risk has claim probability 0.1 and a fixed claim payment of $\$200$. An elevated risk has claim probability 0.4 and a fixed claim payment of $\$500$. Find the expected payment per risk. (Answer: $\$92.00$)
A bag contains 5 amber and 3 teal tokens. Two are drawn without replacement. The prize is $\$9$ for matching colors and $\$3$ for different colors, plus a $\$2$ bonus if the second token is teal. Find the expected prize. (Answer: $\$6.54$)
A machine has a reliable type with probability 0.8 and a faulty type with probability 0.2. Conditional on its type, failures on different days are independent. The daily failure probabilities are 0.05 for a reliable machine and 0.25 for a faulty machine. Given that the machine failed on day 1, find the expected number of failures over the next 3 days. (Answer: 0.4833 failures)

What You'll Learn