Nested Conditioning in Multi-Stage Experiments

Introduction

In a sequential experiment, information arrives in stages. An initial observation $Y$ may be followed by a further observation $Z$, after which a quantity $X$ is evaluated. The conditional expectation $E[X\mid Y,Z]$ uses both stages of information. If only the first-stage information is to be retained, the correct operation is to average that conditional expectation over the possible values of $Z$, using the distribution of $Z$ given $Y$.

This is the practical meaning of the tower property for nested conditioning. It lets us simplify a calculation without discarding the dependence between stages. The method is aligned with the SOA VEE Mathematical Statistics syllabus.

Definition: If $X$ is integrable and the information from $Y$ is contained in the information from $(Y,Z)$, then $$ E\big[E[X\mid Y,Z]\mid Y\big]=E[X\mid Y]\quad\text{with probability 1}. $$ The inner conditional expectation uses both $Y$ and $Z$. The outer conditional expectation averages that result over the conditional distribution of $Z$ given $Y$.

Reading the Conditioning Stages

Conditioning on $Y$ means using the information available from $Y$ alone. Conditioning on $(Y,Z)$ means using the information from both variables. The second information set contains the first: knowing $Y$ and $Z$ includes knowing $Y$. Consequently, after calculating $E[X\mid Y,Z]$, conditioning again on $Y$ averages out the additional information from $Z$, while retaining what $Y$ tells us.

For discrete variables, this averaging can be written as a conditional sum. At a value $y$ with $P(Y=y)>0$,

E[X\mid Y=y] =\sum_z E[X\mid Y=y,Z=z]P(Z=z\mid Y=y).

For continuous $Z$, if $Z$ has conditional density $f_{Z\mid Y}(z\mid y)$, the corresponding expression is

E[X\mid Y=y] =\int E[X\mid Y=y,Z=z]f_{Z\mid Y}(z\mid y)\,dz.

The sum or integral is over the conditional support of $Z$ given $Y=y$. These value-by-value expressions presume the relevant conditional distributions are defined; for discrete conditioning, the stated condition $P(Y=y)>0$ ensures the conditional probability is defined by the usual ratio. The general tower property is stated for conditional expectations as random variables and holds with probability 1.

Why the Tower Property Holds

The result follows from the defining property of conditional expectation. Let $\mathcal H$ denote the information generated by $Y$, and let $\mathcal G$ denote the information generated by $(Y,Z)$. Then $\mathcal H$ is contained in $\mathcal G$. Conditional expectation given $\mathcal G$ preserves the expectation of $X$ on every event whose occurrence can be determined from $\mathcal G$. In particular, it preserves that expectation on any event determined by $Y$.

Choose an event described by the first-stage information.
For any event $A$ determined by $Y$, the event is also determined by $(Y,Z)$, because knowing both variables includes knowing $Y$.

Use the defining property at the second stage.
Since $A$ is determined by $(Y,Z)$, integrating $E[X\mid Y,Z]$ over $A$ gives the same result as integrating $X$ over $A$.

Use the defining property at the first stage.
Since $A$ is determined by $Y$, integrating $X$ over $A$ gives the same result as integrating $E[X\mid Y]$ over $A$.

Combining these statements gives

\int_A E[X\mid Y,Z]\,dP =\int_A X\,dP =\int_A E[X\mid Y]\,dP.

The outer conditional expectation $E[E[X\mid Y,Z]\mid Y]$ is itself determined by $Y$. The equation above shows that it has the same integral as $E[X\mid Y]$ over every event determined by $Y$. By the uniqueness property of conditional expectation, the two are equal with probability 1. Integrability of $X$ ensures these expectations exist.

Worked Examples

Worked Example: A Two-Stage Hypothetical Claim Review

A hypothetical review process first assigns a claim to type $Y=A$ or $Y=B$, with $P(Y=A)=0.3$ and $P(Y=B)=0.7$. A second-stage review records $Z$. The table gives the conditional probabilities of the second-stage result and the conditional mean claim amount $E[X\mid Y,Z]$, in dollars. Find $E[X\mid Y=A]$, $E[X\mid Y=B]$, and the overall mean $E[X]$. Assume all amounts are measured in dollars at the same time, with no interest adjustment.

First-stage type	Second-stage result	$P(Z\mid Y)$	$E[X\mid Y,Z]$
A	1	0.5	$20
A	2	0.5	$44
B	1	0.6	$30
B	2	0.4	$40

Method. Apply the conditional-sum form of the tower property, weighting each second-stage conditional mean by its probability given the first-stage type.

For type A,

E[X\mid Y=A] =20(0.5)+44(0.5) =10+22 =32\text{ dollars}.

For type B,

E[X\mid Y=B] =30(0.6)+40(0.4) =18+16 =34\text{ dollars}.

Now average over the first-stage types:

E[X] =32(0.3)+34(0.7) =9.6+23.8 =33.4\text{ dollars}.

Check. The four joint weights are $0.3(0.5)=0.15$, $0.3(0.5)=0.15$, $0.7(0.6)=0.42$, and $0.7(0.4)=0.28$; they sum to 1. Weighting the four means directly gives $20(0.15)+44(0.15)+30(0.42)+40(0.28)=3+6.6+12.6+11.2=33.4$ dollars. The two-stage and direct calculations agree.

Common mistake avoided. The second-stage probabilities are conditional on the first-stage type. Using one unconditional distribution for $Z$ inside both type-specific calculations would use the wrong weights.

Worked Example: A Randomized Hypothetical Payment

A hypothetical payment rule has a first-stage value $Y$ equal to 4 with probability 0.75 and 2 with probability 0.25. Given either value of $Y$, a second-stage indicator $Z$ equals 1 with probability 0.1 and equals 0 otherwise. The payment, in dollars, is $X=5Y+20YZ$. Find the conditional mean payment for each value of $Y$, then find $E[X]$. Assume no interest adjustment.

Method. First condition on both stages; then average over the conditional distribution of $Z$ given $Y$. Since $Z$ is an indicator, its conditional mean given $Y$ is 0.1.

E[X\mid Y,Z]=5Y+20YZ, \qquad E[X\mid Y]=5Y+20Y E[Z\mid Y] =5Y+20Y(0.1)=7Y.

When $Y=4$, the conditional mean is $7(4)=28$ dollars. Directly, the payment is $5(4)+20(4)(0)=20$ dollars when $Z=0$, and $5(4)+20(4)(1)=100$ dollars when $Z=1$. Thus $20(0.9)+100(0.1)=18+10=28$ dollars.

When $Y=2$, the conditional mean is $7(2)=14$ dollars. The corresponding payments are $5(2)=10$ dollars when $Z=0$ and $5(2)+20(2)=50$ dollars when $Z=1$. Hence $10(0.9)+50(0.1)=9+5=14$ dollars.

Finally,

E[X]=28(0.75)+14(0.25) =21+3.5 =24.5\text{ dollars}.

Check. The unconditional probabilities of $(Y,Z)=(4,0),(4,1),(2,0),(2,1)$ are $0.675,0.075,0.225,0.025$. The direct weighted mean is $20(0.675)+100(0.075)+10(0.225)+50(0.025)=13.5+7.5+2.25+1.25=24.5$ dollars.

Common mistake avoided. Do not use the probability of $Z=1$ as though it were the probability of the larger payment itself. The payment also depends on $Y$; the conditional mean must retain that factor.

Worked Example: Averaging Over a Continuous Second Stage

Let $Y$ be uniformly distributed on $[0,2]$. Given $Y=y$, let $Z$ be uniformly distributed on $[0,y+1]$. Suppose $X=2Y+3Z+\varepsilon$, where $X$ and $\varepsilon$ are integrable and $E[\varepsilon\mid Y,Z]=0$. Find $E[X\mid Y=y]$, evaluate it at $y=1.2$, and calculate $E[X]$.

Method. The zero conditional mean of $\varepsilon$ gives $E[X\mid Y,Z]=2Y+3Z$. Then average over the conditional uniform density of $Z$.

For fixed $y$, the conditional density of $Z$ is $1/(y+1)$ on $0\le z\le y+1$. Therefore,

\begin{aligned} E[X\mid Y=y] &=\int_0^{y+1}(2y+3z)\frac{1}{y+1}\,dz\\ &=2y+\frac{3}{y+1}\left[\frac{z^2}{2}\right]_0^{y+1}\\ &=2y+\frac{3(y+1)}{2} =\frac{7}{2}y+\frac{3}{2}. \end{aligned}

At $y=1.2$, the conditional mean is $(7/2)(1.2)+3/2=4.2+1.5=5.7$. For the unconditional mean, $E[Y]=1$ because $Y$ is uniform on $[0,2]$. Also $E[Z\mid Y]=(Y+1)/2$, so $E[Z]=(E[Y]+1)/2=1$. Thus

$$ E[X]=2E[Y]+3E[Z]=2(1)+3(1)=5. $$

As a check, averaging the conditional mean gives $E[(7/2)Y+3/2]=(7/2)(1)+3/2=5$. No distribution for $\varepsilon$ beyond the stated integrability and conditional-mean assumption is needed.

Common mistake avoided. The density used in the integral is conditional on $Y=y$, and its support ends at $y+1$. Replacing it with a fixed interval or an unconditional density would not represent the second-stage experiment.

A Reliable Calculation Procedure

For a multi-stage problem, write down which information is available at each stage before substituting values. This prevents the conditional distribution in the averaging step from being confused with a marginal distribution.

Identify the first-stage variable $Y$ and the additional variable $Z$.
Find or simplify $E[X\mid Y,Z]$, using the model for $X$ given both stages.
Average over $Z\mid Y$: use a conditional sum for discrete $Z$, or a conditional density and integral for continuous $Z$.
If an unconditional expectation is required, average $E[X\mid Y]$ over the distribution of $Y$.
Check the calculation by expanding the weights into joint probabilities or by evaluating the corresponding iterated integral.

The order matters: first average over the later information while holding the earlier information fixed, and only then average over the earlier information if the question asks for an unconditional mean.

Common Traps and Why the Method Matters

Using marginal probabilities instead of conditional probabilities. In the inner averaging step, use $P(Z=z\mid Y=y)$, not $P(Z=z)$, unless independence has been established.
Dropping the first-stage value. If $E[X\mid Y,Z]$ depends on $Y$, keep that dependence while averaging over $Z$.
Confusing the inner and outer expectations. $E[X\mid Y,Z]$ is a quantity that still varies with both stages; $E[X\mid Y]$ varies with the first stage alone.
Reversing the averaging order without justification. A sequential model may specify $Z\mid Y$, so its conditional distribution is the natural one for averaging out $Z$ at fixed $Y$.

Nested conditioning is useful whenever a model describes a quantity through successive information updates: a risk class followed by a review result, a preliminary measurement followed by a confirmatory measurement, or a sampled group followed by an observation within that group. The intermediate conditional expectation may be much simpler than the full distribution of $X$, and the tower property ensures that simplifying it does not change the mean.

Key takeaway: When $Y$ is part of the information in $(Y,Z)$, average $E[X\mid Y,Z]$ over the conditional distribution of $Z$ given $Y$. The result is $E[X\mid Y]$, provided $X$ is integrable.

Check Your Understanding

For each problem, keep the conditional distribution of the later-stage variable attached to the earlier-stage value. Answers are provided for checking.

$Y$ is 0 with probability 0.4 and 1 with probability 0.6. Given $Y=0$, $P(Z=1\mid Y)=0.25$; given $Y=1$, $P(Z=1\mid Y)=0.5$. If $E[X\mid Y,Z]=10+4Y+8Z$, find $E[X\mid Y=0]$, $E[X\mid Y=1]$, and $E[X]$. (Answer: $12,\ 18,\ 15.6$.)
$Y=1$ with probability 0.2 and $Y=3$ with probability 0.8. Given $Y=1$, $P(Z=1\mid Y)=0.5$; given $Y=3$, $P(Z=1\mid Y)=0.25$. If $E[X\mid Y,Z]=Y(2+Z)$, find the two conditional means and the unconditional mean. (Answer: $2.5,\ 6.75,\ 5.9$.)
$Y\sim\operatorname{Uniform}(0,4)$, and given $Y=y$, $Z\sim\operatorname{Uniform}(0,2y)$. If $E[X\mid Y,Z]=Y+Z^2$, find $E[X\mid Y=y]$ and $E[X]$. (Answer: $y+\frac{4}{3}y^2,\ \frac{82}{9}$.)
A first-stage class $Y$ is $A$ with probability 0.4 and $B$ with probability 0.6. For class A, the second-stage outcomes have probabilities 0.2 and 0.8 and conditional means $15$ and $25$; for class B, they have probabilities 0.5 and 0.5 and conditional means $10$ and $30$. Find the conditional mean for each class and the overall mean. (Answer: $23,\ 20,\ 21.2$.)
$Y$ equals 2 or 5, each with probability 0.5. Given $Y=2$, $P(Z=1\mid Y)=0.2$; given $Y=5$, $P(Z=1\mid Y)=0.6$. If $X=Y^2+3YZ$, find $E[X\mid Y=2]$, $E[X\mid Y=5]$, and $E[X]$. (Answer: $5.2,\ 34,\ 19.6$.)

What You'll Learn