Conditional Independence and Exchangeability

Why the Repeated-Observation Assumption Matters

The Bühlmann model describes repeated observations for a risk, such as annual claim counts or losses, while allowing different risks to have different underlying expected outcomes. Its central assumption is that observations for the same risk are conditionally independent and identically distributed given the risk parameter. The conditioning matters: observations may be dependent before the risk parameter is known, even though they are independent once it is fixed.

This is a modeling assumption, not a claim that every observed year must look alike. A risk can have a high-claim year followed by a low-claim year and still be modeled with conditionally identical observations. The relevant question is whether the data-generating process has a stable conditional distribution and no remaining dependence after accounting for the risk parameter.

Definition: Let $\Theta$ denote a risk parameter for one risk, and let $X_1,\ldots,X_n$ be its observations. The observations are conditionally iid given $\Theta$ if, for every possible value $\theta$, their conditional joint distribution factors as $$ P(X_1\in A_1,\ldots,X_n\in A_n\mid \Theta=\theta) =\prod_{j=1}^{n}P(X_j\in A_j\mid \Theta=\theta), $$ and each $X_j$ has the same conditional distribution given $\Theta=\theta$. In a discrete model, the corresponding statement is $$ P(X_1=x_1,\ldots,X_n=x_n\mid\Theta=\theta) =\prod_{j=1}^{n}p(x_j\mid\theta). $$

Conditional independence means that, once $\Theta$ is fixed, learning the observation in one period gives no additional information about another observation. Conditional identical distribution means that the period labels do not change the conditional distribution: the same $p(x\mid\theta)$ applies in each period. These are separate requirements. Observations can be conditionally independent but have different distributions because their exposures differ.

In the actuarial examples below, the models concern property and casualty observations. No interest-rate or mortality assumption is involved.

From Conditional iid to Exchangeability

A sequence is exchangeable when its joint distribution is unchanged by a permutation of the observation labels. For example, for two observations, exchangeability requires $P(X_1=a,X_2=b)=P(X_1=b,X_2=a)$ for all $a,b$. It does not require the observations to be independent.

Conditional iid observations are exchangeable after the risk parameter is averaged over its distribution. The reason is that the conditional product has the same factor for each observation, so reordering its factors does not change it.

P(X_1=x_1,\ldots,X_n=x_n) =\int \prod_{j=1}^{n}p(x_j\mid\theta)\,dP_\Theta(\theta).

For any permutation $\pi$ of $1,\ldots,n$, the product $\prod_{j=1}^{n}p(x_{\pi(j)}\mid\theta)$ contains the same factors as $\prod_{j=1}^{n}p(x_j\mid\theta)$. Integrating either product over the distribution of $\Theta$ therefore gives the same probability. This proves exchangeability.

The converse is not automatic for a finite sequence: a finite exchangeable distribution need not be representable as a mixture of iid distributions. More importantly for practical modeling, exchangeability alone does not establish conditional independence given a particular risk parameter.

Key distinction: Conditional iid observations are exchangeable, but exchangeable observations need not be independent. If the conditional mean $m(\Theta)=E[X_j\mid\Theta]$ varies across risks and conditional variances are finite, then for distinct periods $j$ and $k$, $$ \operatorname{Cov}(X_j,X_k)=\operatorname{Var}(m(\Theta)). $$ The conditional independence removes the within-$\Theta$ covariance; the common random risk parameter creates dependence in the unconditioned observations.

Worked Examples

Worked Example: A Binary Claim Indicator Under Risk Mixing

A hypothetical portfolio has two equally likely risk types. For a policy of type A, the probability of a claim in each year is $0.10$; for type B, it is $0.30$. Conditional on type, the two annual claim indicators $X_1$ and $X_2$ are independent Bernoulli variables with the same annual probability. Find the marginal claim probability, the probability of claims in both years, and the covariance of the two indicators. This is a discrete Bernoulli model with equal type probabilities and parameters $0.10$ and $0.30$; interest and mortality are not modeled.

Method. Calculate conditional probabilities within each type, then average over the two types. This avoids treating the annual indicators as unconditionally independent.

For either year, $$ P(X_1=1)=0.50(0.10)+0.50(0.30)=0.05+0.15=0.20. $$ The same calculation gives $P(X_2=1)=0.20$.

Conditional independence gives the probability of claims in both years for type A as $0.10(0.10)=0.010$, and for type B as $0.30(0.30)=0.090$. Averaging those conditional probabilities, $$ P(X_1=1,X_2=1)=0.50(0.010)+0.50(0.090)=0.005+0.045=0.050. $$

The product of the marginal probabilities is $0.20(0.20)=0.040$. Therefore, $$ \operatorname{Cov}(X_1,X_2)=E[X_1X_2]-E[X_1]E[X_2] =0.050-0.040=0.010. $$ As a check, the covariance also equals the variance of the conditional claim probability: $E[\Theta]=0.20$, $E[\Theta^2]=0.050$, and $\operatorname{Var}(\Theta)=0.050-0.20^2=0.010$. The joint distribution is exchangeable because swapping the two years does not change any conditional product.

Common mistake avoided. Multiplying the marginal probabilities gives $0.040$, not the joint probability under this mixture. The correct joint probability is $0.050$, because both years share the same unknown risk type. In particular, $0.50(0.30^2)=0.045$, not $0.090$; the latter is the type B conditional joint probability before applying its $0.50$ portfolio weight.

Worked Example: Different Exposures Preserve Independence but Break Identical Distribution

For a hypothetical risk with fixed conditional annual frequency parameter $\Theta=0.40$ claims per unit exposure, claim counts in two periods are conditionally independent Poisson variables. The first period has exposure $e_1=1$, and the second has exposure $e_2=2$. Find each period's conditional mean and probability of zero claims, and determine whether the raw counts are conditionally iid. The distribution is Poisson with mean $e_j\Theta$; interest and mortality are not modeled.

Method. Apply the exposure multiplier to the frequency parameter separately in each period. Conditional independence is given, but identical distribution must be checked rather than assumed.

The first count has mean $e_1\Theta=1(0.40)=0.40$, so $N_1\mid\Theta=0.40\sim\operatorname{Poisson}(0.40)$. The second has mean $e_2\Theta=2(0.40)=0.80$, so $N_2\mid\Theta=0.40\sim\operatorname{Poisson}(0.80)$. The counts are independent by the model, but their different means show that they are not identically distributed.

The zero-count probabilities are $$ P(N_1=0\mid\Theta=0.40)=e^{-0.40}=0.6703, \qquad P(N_2=0\mid\Theta=0.40)=e^{-0.80}=0.4493. $$ Independence gives $$ P(N_1=0,N_2=0\mid\Theta=0.40) =e^{-0.40}e^{-0.80}=e^{-1.20}=0.3010. $$ The product check agrees with the joint probability. The answer is therefore conditional independence without conditional identical distribution. Dividing counts by exposure does not make their distributions identical: the scaled counts still have exposure-dependent variability.

Common mistake avoided. Using the same Poisson mean for both periods ignores the doubled exposure. Independence does not imply identical distribution; each period's exposure must be included in its mean.

Worked Example: Exchangeability Despite a Shared Shock

Consider two hypothetical monthly claim indicators for policies in the same class. A common monthly shock $H$ occurs with probability $0.10$; if it occurs, both indicators equal 1. If it does not occur, the indicators are independent Bernoulli variables with claim probability $1/9$. Treat the class parameter as fixed, and do not include $H$ in that parameter. Find the marginal claim probability, the probability of claims in both months, and whether the observations are conditionally independent given the class parameter. This is a two-period Bernoulli mixture model with shock probability $0.10$ and no-shock claim probability $1/9$; interest and mortality are not modeled.

Method. Condition on the common shock and combine the two cases. A shared event can make the observations dependent even when the model is symmetric in the month labels.

For either month, the claim probability is $$ P(X_j=1)=0.10(1)+0.90\left(\frac{1}{9}\right) =0.10+0.10=0.20. $$ For claims in both months, the shock case contributes $0.10$. In the no-shock case, independence gives probability $(1/9)^2=1/81$. Thus $$ P(X_1=1,X_2=1)=0.10+0.90\left(\frac{1}{81}\right) =0.10+\frac{1}{90} =\frac{1}{9}=0.1111. $$ If the two indicators were independent given the class parameter, their joint claim probability would be $0.20^2=0.0400$. Since $0.1111\ne0.0400$, they are not conditionally independent given that parameter. Their covariance is $0.1111-0.0400=0.0711$, to four decimal places.

The construction treats the months symmetrically, so swapping them does not change the joint distribution: the pair is exchangeable. Yet it is not conditionally independent given the class parameter. If the common shock were included in the conditioning information, the conditional model would instead be independent in each shock state (including the degenerate state where both claims are certain).

Common mistake avoided. Symmetry between periods establishes exchangeability here, not conditional independence. Check the joint probability against the product of the relevant conditional marginals.

Diagnosing the Assumptions in Practice

A data pattern can raise a concern about conditional iid assumptions, but a finite sample cannot by itself prove that the assumptions are false. Random variation can produce apparent trends or clusters. Diagnosis requires considering how the observations were generated and whether the model's risk parameter captures persistent differences between risks.

Identify the observation unit.
Specify whether each $X_j$ is an annual count, a loss amount, or another consistently defined quantity. Changes in the definition across periods can mimic changes in the process.

Check conditional distributions.
Ask whether exposure, coverage, limits, inflation, or the operating environment changed. If these affect the distribution and are not handled in the model, periods may not be identically distributed given $\Theta$.

Look for remaining dependence.
After accounting for the risk parameter and relevant conditions, investigate whether a shared event, reporting process, or other common influence connects observations across periods.

Interpret patterns cautiously.
Repeated runs of high outcomes or a time trend may signal a missing factor, but sampling variation is also possible. Assess the mechanism and data, rather than diagnosing a violation from one pattern alone.

Several practical patterns deserve particular attention:

Changing exposure: counts can remain independent while their distributions differ because expected counts scale with exposure. Exposure-adjusted modeling may be appropriate, but it does not automatically make all response distributions identical.
Trend or structural change: changes in rates, policy terms, claims handling, or external conditions can make the conditional distribution period-dependent. A single stable $\Theta$ may not describe the entire observation window.
Shared shocks: weather events, outages, or other common causes can create dependence among observations after conditioning on a persistent risk classification. A relevant shared factor may need to be represented in the conditioning information or modeled explicitly.
Selection and reporting effects: if inclusion or reporting in one period depends on prior outcomes, the observed sequence may not have the same conditional distribution as the underlying sequence.

The purpose of the assumptions is not to require identical observed outcomes. It is to support a coherent repeated-observation model: once the risk parameter and any necessary conditions are fixed, the periods are governed by the same distribution and do not retain unexplained dependence. This is why the assumptions matter for credibility: mixing over persistent risk differences can generate dependence between observations, while unmodeled time variation or common events can create additional patterns that the basic model does not represent.

Key takeaway: Conditional iid is stronger than exchangeability. It requires both a common conditional distribution and conditional independence given the risk parameter. Check exposures and time stability for the first condition, and shared causes or dependence mechanisms for the second.

Check Your Understanding

For each problem, distinguish what is known conditionally from what follows after averaging or comparing periods.

A risk has a Bernoulli claim probability $\Theta$, where $\Theta=0.20$ or $0.60$, each with probability $0.50$. Given $\Theta$, two annual indicators are iid. Calculate $P(X_1=1)$, $P(X_1=1,X_2=1)$, and their covariance. (Answer: $0.4000,\ 0.2000,\ 0.0400$)
Conditional on a fixed frequency parameter $\Theta=0.25$, two independent period counts are Poisson with exposures 1 and 3. Give their means, the probability both are zero, and state whether they are conditionally iid. (Answer: means $0.25$ and $0.75$; joint zero probability $e^{-1}=0.3679$; independent but not identically distributed)
A risk parameter takes values $0.25$ and $0.55$ with equal probabilities. Given it, two indicators are iid Bernoulli with that parameter. Calculate the probability of two claims and the covariance. (Answer: $P(X_1=1,X_2=1)=0.1825$; $\operatorname{Cov}(X_1,X_2)=0.0225$)
Two period outcomes have a joint distribution unchanged when the period labels are swapped. Does this alone establish conditional independence given a risk parameter? State the distinction. (Answer: No. It establishes exchangeability, but conditional independence given a specified risk parameter is an additional assumption.)
A hypothetical insurer observes a sharp increase in claims after a coverage change, and exposure also doubled. Name two separate checks before treating the observations as conditionally iid. (Answer: Check whether the coverage change altered the conditional claim distribution, and account for the doubled exposure when comparing period counts; also assess whether any dependence remains after conditioning.)

What You'll Learn