Why the Bühlmann Model Uses a Risk Parameter
Repeated claims from one insured are not usually treated as if they came from unrelated risks. A policyholder with persistently high underlying claim propensity may tend to have higher losses in several periods. The Bühlmann collective model represents this by assigning each risk a latent parameter and describing the observations conditional on that parameter.
The word collective refers to the population from which the risks are drawn: their latent parameters follow a common structural distribution. The word does not mean that every risk has the same parameter. Risks may differ, while still being modeled as members of one population. This tutorial specifies that model and its assumptions; it does not calculate a Bühlmann credibility estimate.
For example, \(X_j\) might be the claim count or loss amount for one risk in period \(j\). The model distinguishes variation between risks, represented by different values of \(\Theta\), from variation between observations for the same risk, represented by the conditional distribution given \(\Theta\). A risk parameter is a modeling device; it need not be directly observed or have a simple physical interpretation.
Assumptions in Precise Form
For one risk, write the conditional mean as \(m(\Theta)=E[X_j\mid\Theta]\). In the standard collective model, this conditional mean is the same for each observation \(j\). More fully, conditional on \(\Theta\), the observations have one common distribution, not merely the same mean. Thus the standard model assumes conditional identical distribution as well as conditional independence.
- Common structural distribution: each risk parameter is drawn from the same distribution \(F_\Theta\), with no change in that distribution across the risks being modeled.
- Conditional independence: given a risk's \(\Theta\), its observations in distinct periods are independent.
- Conditional identical distribution: given \(\Theta\), those observations share a common conditional distribution, and hence a common conditional mean \(m(\Theta)\).
- Finite moments when needed: means require finite first moments. Covariance and variance calculations require finite second moments, such as \(E[X_1^2]<\infty\) and \(E[X_2^2]<\infty\).
The common structural distribution is an assumption about the risks included in a collective, not a claim that their observed histories are identical. If the book combines materially different populations, a single structural distribution may not describe them appropriately. Likewise, observations from different periods cannot be treated as identically distributed if a trend, coverage change, or other systematic change alters their conditional distributions and that change is not represented in the model.
The model's shared overall mean is a consequence of the common structural distribution and common conditional mean function. When the relevant expectations exist, define \(\mu=E[m(\Theta)]\). Then \(E[X_j]=\mu\) for every modeled period \(j\). This does not imply that the observations for a given risk are independent without conditioning.
Why Conditional Independence Does Not Mean Overall Independence
A latent parameter can make repeated observations associated even when they are independent after that parameter is fixed. A high value of \(\Theta\) can raise the conditional mean in several periods at once. The following identity makes the role of that shared mean precise.
This identity requires finite second moments. In particular, assume \(E[X_1^2]\) and \(E[X_2^2]\) are finite, and that the conditional means are the same function \(m(\Theta)\). These conditions ensure that the covariance and variance in the identity are defined and finite.
Conditional independence gives \(E[X_1X_2\mid\Theta]=E[X_1\mid\Theta]E[X_2\mid\Theta]=m(\Theta)^2\). The second-moment condition makes the required expectations finite.
Taking expectations yields \(E[X_1X_2]=E[m(\Theta)^2]\). Also, \(E[X_1]=E[X_2]=E[m(\Theta)]\) by iterated expectation.
\(\operatorname{Cov}(X_1,X_2)=E[X_1X_2]-E[X_1]E[X_2]=E[m(\Theta)^2]-E[m(\Theta)]^2=\operatorname{Var}(m(\Theta))\).
The result explains why unconditioned observations from the same risk need not be independent: unless \(m(\Theta)\) is constant almost surely, they share a random conditional mean. In the special case where that mean is constant, the covariance is zero under these assumptions. The identity is not justified merely by saying that conditional independence holds; the finite-second-moment condition is essential.
Worked Examples: Applying the Model Assumptions
Worked Example: A Poisson Claim Count With Two Risk Types
A hypothetical portfolio has a risk parameter \(\Theta\) equal to 2 with probability 0.4 and 5 with probability 0.6. Given \(\Theta\), the claim counts \(X_1\) and \(X_2\) in two periods are conditionally independent Poisson random variables, each with parameter \(\Theta\). Find their shared overall mean, verify finite second moments, and calculate their covariance. Name the method and the mistake avoided.
Method. Specify the conditional model first, then average its moments over the structural distribution. For a Poisson random variable with parameter \(\lambda\), \(E[X\mid\lambda]=\lambda\) and \(E[X^2\mid\lambda]=\lambda+\lambda^2\).
Shared overall mean. The conditional mean is \(m(\Theta)=\Theta\), so \(E[X_j]=E[\Theta]=0.4(2)+0.6(5)=0.8+3.0=3.8\), for either period \(j\).
Second moments. Given \(\Theta=2\), \(E[X_j^2\mid\Theta=2]=2+2^2=6\). Given \(\Theta=5\), \(E[X_j^2\mid\Theta=5]=5+5^2=30\). Therefore, \(E[X_j^2]=0.4(6)+0.6(30)=2.4+18=20.4\), which is finite for both periods.
Covariance. The shared conditional mean has \(E[m(\Theta)^2]=0.4(2^2)+0.6(5^2)=1.6+15=16.6\), and \(E[m(\Theta)]^2=3.8^2=14.44\). Thus \(\operatorname{Cov}(X_1,X_2)=16.6-14.44=2.16\).
Check. The covariance is nonnegative, as it must be because it is the variance of \(m(\Theta)\). The common mistake avoided is treating conditional independence as unconditional independence and setting the covariance to zero.
Worked Example: Bernoulli Observations and Shared Risk
For a hypothetical small commercial policy group, let \(\Theta\) equal 0.2 or 0.8, each with probability 0.5. In each of two periods, let \(X_j=1\) indicate a claim and \(X_j=0\) indicate no claim. Conditional on \(\Theta\), the observations are independent Bernoulli random variables with success probability \(\Theta\). Find the overall claim probability, \(E[X_1X_2]\), the covariance, and the marginal variance. Name the method and the mistake avoided.
Method. Use the Bernoulli conditional model and average over the two-point structural distribution. For a Bernoulli variable with success probability \(p\), its conditional mean is \(p\) and its conditional second moment is also \(p\), since \(X^2=X\).
Overall claim probability. The shared conditional mean is \(m(\Theta)=\Theta\), so \(E[X_j]=0.5(0.2)+0.5(0.8)=0.1+0.4=0.5\).
Product expectation. Conditional independence gives \(E[X_1X_2\mid\Theta]=\Theta^2\). Therefore, \(E[X_1X_2]=0.5(0.2^2)+0.5(0.8^2)=0.5(0.04)+0.5(0.64)=0.02+0.32=0.34\).
Covariance. By the definition of covariance, \(\operatorname{Cov}(X_1,X_2)=0.34-(0.5)(0.5)=0.34-0.25=0.09\). Equivalently, \(\operatorname{Var}(\Theta)=0.5(0.04)+0.5(0.64)-0.5^2=0.34-0.25=0.09\).
Marginal variance and check. Since \(X_1\) is Bernoulli with overall success probability 0.5, \(\operatorname{Var}(X_1)=0.5(1-0.5)=0.25\). The covariance, 0.09, is less than this marginal variance and is positive. The common mistake avoided is multiplying the overall probabilities as if the period outcomes were unconditionally independent: that would incorrectly give \(E[X_1X_2]=0.5^2=0.25\).
Worked Example: Checking Whether Two Territories Share One Collective
A hypothetical insurer studies claim indicators from two territories. Within either territory, each risk has a latent type, low or high. The conditional claim probabilities are 0.1 for low type and 0.5 for high type; observations are conditionally independent Bernoulli variables across periods. In Territory A, 75% of risks are low type and 25% high type. In Territory B, the proportions are 25% low and 75% high. Can the two territory-specific populations be treated as having the same structural distribution and shared overall mean? Name the diagnostic method and the mistake avoided.
Method. Compare the structural distributions and average conditional means separately for each territory. The conditional distribution is Bernoulli with the stated type-specific probability, and all probabilities and mixture proportions are held constant across periods.
Territory A. Its mean claim probability is \(0.75(0.1)+0.25(0.5)=0.075+0.125=0.200\). Its structural probabilities for low and high type are 0.75 and 0.25.
Territory B. Its mean claim probability is \(0.25(0.1)+0.75(0.5)=0.025+0.375=0.400\). Its structural probabilities for low and high type are 0.25 and 0.75.
Conclusion. The territory-specific structural distributions differ, and their overall means are 0.200 and 0.400. Therefore these two separate populations do not satisfy the same-structural-distribution and shared-overall-mean assumptions when modeled as one undifferentiated collective. The common mistake avoided is assuming that conditional independence within each risk is enough to justify pooling different populations. A larger portfolio could instead be modeled with territory represented explicitly in its risk structure, but that would require specifying an appropriate common population model.
Model Checks and Exam Traps
Before using the Bühlmann collective model, identify the unit being modeled and the meaning of each repeated observation. For claim counts, for example, the risk might be a policyholder and the observations might be annual counts. Changing the unit to a territory or policy year changes what “within a risk” means and can change whether the assumptions are plausible.
Then check the assumptions in order. Ask whether the risks plausibly come from a common structural distribution; whether the conditional distribution is stable across the periods; and whether the observations are independent after conditioning on the risk parameter. If the data have a shared external shock, such as a catastrophe affecting many policies at once, conditional independence may fail unless the model represents that shared factor. If exposure differs by period, identical conditional distributions may also fail unless the observations have been put on a comparable basis or the model has been extended.
- Wrong: “The observations are conditionally independent, so their covariance is zero.”
Correction: Conditional independence is not unconditional independence. With finite second moments and a common conditional mean, the covariance equals \(\operatorname{Var}(m(\Theta))\). - Wrong: Use the covariance identity without checking whether second moments exist.
Correction: Verify finite second moments, such as \(E[X_1^2]\) and \(E[X_2^2]\), before interpreting covariance or variance formulas. - Wrong: Assume a common structural distribution means all risks have equal expected experience.
Correction: A common distribution allows risk parameters to vary; it requires the risks to be drawn from the same distribution. - Wrong: Pool populations with different mixes of risk types and assume the shared overall mean follows automatically.
Correction: Check that the populations can reasonably be represented by the same structural distribution and conditional model.
Check Your Understanding
For each problem, first identify the structural distribution and the conditional model, then apply only the assumptions stated.
- A hypothetical group has \(\Theta=1\) with probability 0.25 and \(\Theta=3\) with probability 0.75. Given \(\Theta\), two period claim counts are independent Poisson(\(\Theta\)). Find the shared overall mean and covariance. (Answer: mean \(=2.5\); covariance \(=0.75\).)
- For a hypothetical risk, \(\Theta\) equals 0.3 or 0.7 with equal probability. Conditional on \(\Theta\), two claim indicators are independent Bernoulli(\(\Theta\)). Find \(E[X_1X_2]\) and \(\operatorname{Cov}(X_1,X_2)\). (Answer: \(E[X_1X_2]=0.29\); covariance \(=0.04\).)
- A model has conditional mean \(m(\Theta)=\Theta\), but \(E[X_1^2]\) is infinite. Is the finite covariance identity established by conditional independence and the shared conditional mean alone? (Answer: No; those facts do not establish a finite, defined covariance. A suitable finite-second-moment condition is also needed.)
- In a hypothetical portfolio, Territory C has 60% low-risk and 40% high-risk policies, while Territory D has 40% low-risk and 60% high-risk policies. Low and high types have conditional claim probabilities 0.05 and 0.35. Find each territory's mean claim probability and state whether the two separate territory populations have the same structural distribution. (Answer: C mean \(=0.17\); D mean \(=0.23\); no, their structural distributions differ.)
- A risk's annual observations are conditionally independent given \(\Theta\), but the conditional claim probability is 0.10 in the first period and 0.20 in the second for every value of \(\Theta\). Does this satisfy the standard conditionally identical distribution assumption? (Answer: No; the conditional distributions differ between periods.)