Why Conditional Means Matter
An overall expected loss summarizes a portfolio before additional information is observed. Once a risk indicator, exposure feature, or event is known, the relevant summary may instead be the expected loss given that information. Comparing these conditional means shows how the average loss changes across observed circumstances. It does not describe every feature of the conditional loss distribution, but it is a useful and interpretable risk summary.
Let \(X\) and \(Y\) be random variables, where \(Y\) is a loss and \(X\) is another observed risk characteristic or loss. The conditional mean \(E[Y\mid X=x]\) is the mean of \(Y\) under the conditional distribution associated with \(X=x\). For discrete \(X\), this conditions on an outcome with positive probability. For continuous \(X\), it is calculated from a conditional density; although \(P(X=x)=0\) at an exact point, a conditional distribution can be defined from the joint and marginal densities.
In either case, the calculation uses the distribution of \(Y\) after conditioning, not its unconditional distribution. The conditional probabilities or density must first be normalized for the given value of \(X\). The result is an average loss, expressed in the same units as \(Y\); it is not the loss that must occur in that situation.
Discrete Conditional Means
For discrete losses, a joint probability table provides the ingredients. First identify the row or column corresponding to the observed value of \(X\). Divide each joint probability in that row or column by its marginal probability to obtain the conditional probabilities. Then multiply each possible value of \(Y\) by its conditional probability and add.
Worked Example: Expected Loss at Each Risk-Indicator Level
A hypothetical insurer groups annual property losses \(X\) into two indicator levels, 0 and 1, and records a related annual equipment loss \(Y\), in thousands of dollars. The joint probabilities are shown below. Find \(E[Y\mid X=0]\) and \(E[Y\mid X=1]\), and interpret the change.
| \(Y=0\) | \(Y=4\) | \(Y=12\) | Total | |
|---|---|---|---|---|
| \(X=0\) | 0.30 | 0.20 | 0.10 | 0.60 |
| \(X=1\) | 0.05 | 0.15 | 0.20 | 0.40 |
| Total | 0.35 | 0.35 | 0.30 | 1.00 |
Method. Normalize each row to obtain the distribution of \(Y\) conditional on that value of \(X\), then take its weighted mean.
For \(X=0\), the row total is \(P(X=0)=0.60\). Thus the conditional probabilities for \(Y=0,4,12\) are \(0.30/0.60=1/2\), \(0.20/0.60=1/3\), and \(0.10/0.60=1/6\). They sum to 1. Therefore, $$ E[Y\mid X=0]=0\left(\frac12\right)+4\left(\frac13\right)+12\left(\frac16\right) =\frac43+2=\frac{10}{3}=3.3333\text{ thousand dollars}. $$
For \(X=1\), the row total is \(P(X=1)=0.40\). The conditional probabilities are \(0.05/0.40=1/8\), \(0.15/0.40=3/8\), and \(0.20/0.40=1/2\). They also sum to 1. Hence, $$ E[Y\mid X=1]=0\left(\frac18\right)+4\left(\frac38\right)+12\left(\frac12\right) =1.5+6=7.5\text{ thousand dollars}. $$
The conditional mean is higher by \(7.5-3.3333=4.1667\) thousand dollars when the indicator is 1. This describes a difference in average equipment loss between the two groups; it does not say that every policy with \(X=1\) has a larger \(Y\). As a check, the marginal probabilities of \(Y=0,4,12\) are \(0.35,0.35,0.30\), so \(E[Y]=0(0.35)+4(0.35)+12(0.30)=5\) thousand dollars.
Common mistake avoided: Using the joint probabilities directly as though they were conditional probabilities would give \(0(0.30)+4(0.20)+12(0.10)=2.0\) for the first row. Those probabilities sum to \(0.60\), not 1. Dividing by the row total before taking the mean gives \(2.0/0.60=3.3333\), the correct conditional mean.
Conditional Means from a Joint Density
For continuous risks, a conditional mean is obtained by integrating against a conditional density. The support matters: after observing \(X=x\), only the values of \(Y\) allowed by the joint model remain possible. In particular, the conditional density need not have the same range or shape as the marginal density of \(Y\).
Worked Example: A Conditional Loss on a Triangular Support
Suppose the hypothetical normalized loss pair \((X,Y)\) has joint density \(f_{X,Y}(x,y)=2\) on \(0<x<1\) and \(0<y<x\), and zero elsewhere. Find the conditional density of \(Y\) given \(X=x\), then find the conditional mean when \(x=0.8\). Losses are expressed in normalized units.
Method. Integrate the joint density over the allowed \(y\)-values to find the marginal density of \(X\). Divide the joint density by this marginal, then integrate \(y\) against the resulting conditional density.
For a fixed \(x\) between 0 and 1, the support is \(0<y<x\). Thus $$ f_X(x)=\int_0^x 2\,dy=2x. $$ The conditional density on that support is therefore $$ f_{Y\mid X}(y\mid x)=\frac{2}{2x}=\frac{1}{x},\qquad 0<y<x. $$ It integrates to 1 because \(\int_0^x (1/x)\,dy=1\). The conditional mean is $$ E[Y\mid X=x]=\int_0^x y\frac{1}{x}\,dy =\frac{1}{x}\left[\frac{y^2}{2}\right]_0^x =\frac{x}{2}. $$
At \(x=0.8\), this gives \(E[Y\mid X=0.8]=0.8/2=0.4\) normalized units. The conditional distribution is uniform over the interval from 0 to 0.8, so its mean is at the midpoint, 0.4. As a check on the marginal, \(\int_0^1 2x\,dx=1\), as a probability density must.
Common mistake avoided: The unconditional support of \(Y\) is between 0 and 1, but after observing \(X=0.8\), the joint support restricts \(Y\) to values below 0.8. Treating the conditional distribution as uniform from 0 to 1 would ignore the joint support and produce the wrong conditional mean.
Conditioning on an Event and Checking with Total Expectation
Sometimes the information observed is an event involving \(X\), rather than one exact value. For a discrete event \(A\) with \(P(A)>0\), the conditional mean is calculated using the conditional distribution given \(A\). If \(A\) combines several values of \(X\), their relative probabilities within \(A\) determine how the corresponding conditional means are weighted.
The law of total expectation provides a useful check. For discrete \(X\), if the relevant expectations exist, then
To see why this holds, write each joint probability as \(P(X=x,Y=y)=P(Y=y\mid X=x)P(X=x)\). Then
The inner sum is the conditional mean for that value of \(X\); the outer sum averages those conditional means using the probabilities of the conditioning values. A mismatch between this result and the mean calculated directly from the marginal distribution often signals a normalization or arithmetic error.
Worked Example: Combining Conditional Means Over an Event
For a hypothetical portfolio, \(X\) is a count of weather alerts and \(Y\) is the associated annual loss, in thousands of dollars. The joint probabilities are:
| \(Y=0\) | \(Y=2\) | \(Y=8\) | Total | |
|---|---|---|---|---|
| \(X=0\) | 0.30 | 0.10 | 0.10 | 0.50 |
| \(X=1\) | 0.05 | 0.15 | 0.10 | 0.30 |
| \(X=2\) | 0.02 | 0.08 | 0.10 | 0.20 |
| Total | 0.37 | 0.33 | 0.30 | 1.00 |
Find \(E[Y\mid X>0]\), and check it by combining the conditional means for \(X=1\) and \(X=2\).
Method. Add the probabilities for the rows in the event \(X>0\), normalize the resulting distribution of \(Y\), and calculate its mean. Then check by weighting the separate conditional means by their probabilities within the event.
The event probability is \(P(X>0)=0.30+0.20=0.50\). Within that event, the joint probabilities for \(Y=0,2,8\) are \(0.05+0.02=0.07\), \(0.15+0.08=0.23\), and \(0.10+0.10=0.20\). The conditional probabilities are therefore \(0.07/0.50=0.14\), \(0.23/0.50=0.46\), and \(0.20/0.50=0.40\), which sum to 1. Thus, $$ E[Y\mid X>0]=0(0.14)+2(0.46)+8(0.40)=0.92+3.20=4.12\text{ thousand dollars}. $$
For the check, the conditional mean when \(X=1\) is $$ E[Y\mid X=1]=0\left(\frac{0.05}{0.30}\right)+2\left(\frac{0.15}{0.30}\right)+8\left(\frac{0.10}{0.30}\right) =1+\frac{8}{3}=\frac{11}{3}. $$ When \(X=2\), $$ E[Y\mid X=2]=0\left(\frac{0.02}{0.20}\right)+2\left(\frac{0.08}{0.20}\right)+8\left(\frac{0.10}{0.20}\right) =0.8+4=4.8. $$ Within \(X>0\), the probabilities of \(X=1\) and \(X=2\) are \(0.30/0.50=0.60\) and \(0.20/0.50=0.40\). Therefore, $$ E[Y\mid X>0]=0.60\left(\frac{11}{3}\right)+0.40(4.8)=2.20+1.92=4.12\text{ thousand dollars}, $$ agreeing with the direct calculation. As an additional check, the unconditional mean is \(0(0.37)+2(0.33)+8(0.30)=3.06\) thousand dollars. Weighting the conditional means for \(X=0,1,2\) by \(0.50,0.30,0.20\) gives \(0.50(2.0)+0.30(11/3)+0.20(4.8)=3.06\).
Common mistake avoided: The conditional weights for \(X=1\) and \(X=2\) given \(X>0\) are 0.60 and 0.40, not their unconditional probabilities 0.30 and 0.20. Conditional weights must sum to 1 within the event being considered.
Interpreting Conditional Means Carefully
A conditional mean answers a specific question: given the stated information, what is the average value of \(Y\) under the model? A pattern in \(E[Y\mid X=x]\) can summarize how average loss varies with \(x\), but it does not by itself establish a causal effect. Nor does a conditional mean reveal the conditional variance, skewness, or probability of an extreme loss. Two groups can have the same mean and very different loss distributions.
For discrete conditioning values, check that conditional probabilities sum to 1 before taking a weighted mean. For continuous variables, check both the conditional support and that the conditional density integrates to 1. Retain exact fractions or unrounded probabilities during intermediate calculations, and round only the reported result; multiplying rounded conditional probabilities may produce a small discrepancy from the exact mean.
Check Your Understanding
For each problem, identify the conditioning information and calculate the requested conditional mean. Answers are given to two decimal places where needed.
- A hypothetical joint distribution has \(P(X=0,Y=0)=0.20\), \(P(X=0,Y=3)=0.30\), and \(P(X=0,Y=9)=0.10\). The remaining probability has \(X=1\). Find \(E[Y\mid X=0]\). (Answer: 3.00)
- A joint density is \(f_{X,Y}(x,y)=2\) for \(0<x<1\) and \(0<y<1-x\), and zero elsewhere. Find \(E[Y\mid X=0.4]\). (Answer: 0.30)
- A loss \(Y\), in thousands of dollars, has conditional probabilities \(P(Y=0\mid X=1)=0.20\), \(P(Y=5\mid X=1)=0.50\), and \(P(Y=15\mid X=1)=0.30\). Find \(E[Y\mid X=1]\). (Answer: 7.00 thousand dollars)
- Suppose \(P(X=1)=0.25\), \(P(X=2)=0.75\), \(E[Y\mid X=1]=4\), and \(E[Y\mid X=2]=10\). Find \(E[Y]\). (Answer: 8.50)
- A hypothetical joint table has row totals \(P(X=1)=0.40\) and \(P(X=2)=0.60\), with \(E[Y\mid X=1]=3\) and \(E[Y\mid X=2]=7\). Find \(E[Y\mid X>1]\), then explain why it differs from \(E[Y]\). (Answer: \(E[Y\mid X>1]=7\); \(E[Y]=5.40\), since \(X>1\) selects only the \(X=2\) group.)