The Bühlmann-Straub Model for Unequal Weights

Why Unequal Weights Need a Model

Experience records are not always equally informative. One hypothetical policyholder may contribute a full year of exposure in each observation period, while another contributes only a fraction of a year. Treating those observations as equally precise ignores a material difference: under a standard exposure model, more exposure produces less variability in an observed average.

The Bühlmann-Straub model extends the Bühlmann collective risk model to observations with differing positive weights. Its central assumption is that, conditional on a risk’s underlying parameter, each observation has the same expected value, while its conditional variance is inversely proportional to its assigned weight. The weight describes relative exposure or precision; it is not itself an observation or a credibility factor.

Definition: For risk $i$, let $\Theta_i$ be its unobserved risk parameter. For observation period $j$, let $X_{ij}$ be the observed quantity and $w_{ij}>0$ its known exposure weight. The Bühlmann-Straub model assumes $$ E[X_{ij}\mid\Theta_i]=\mu(\Theta_i), \qquad \operatorname{Var}(X_{ij}\mid\Theta_i)=\frac{\sigma^2(\Theta_i)}{w_{ij}}. $$ Conditional on $\Theta_i$, observations from that risk are assumed independent. The function $\mu(\Theta_i)$ is the risk’s conditional mean, and $\sigma^2(\Theta_i)$ is its conditional variance parameter at unit weight.

What the Assumptions Mean

The conditional mean does not change with the weight: a larger weight does not mean a larger expected loss cost per unit. Rather, it makes the observed quantity less variable around that expected value. If one observation has four times the weight of another, its conditional variance is one quarter as large, provided both observations have the same conditional variance parameter.

Weights are positive and specified on a consistent scale. They may represent exposure units, such as policy-years, or another justified measure of relative precision. The model does not require weights to be integers. The choice must make sense for the quantity being modeled: an exposure weight is not automatically the number of claims, policies, or years unless that is how the data and model define it.

The standard formulation does not require a particular probability distribution. Normal conditional distributions are sometimes convenient for examples, but the defining moment assumptions are the conditional mean and variance relationships above, together with conditional independence. The model also does not say that every risk has the same mean or variance parameter: these may vary with $\Theta_i$.

Definition: The expected process variance per unit weight is $\phi=E[\sigma^2(\Theta)]$. The variance of hypothetical means is $a=\operatorname{Var}(\mu(\Theta))$, where $\Theta$ represents the population distribution of risk parameters. The overall mean is $m=E[\mu(\Theta)]$. These population quantities separate variation within a risk’s experience from variation in the underlying means across risks.

Here, $\phi$ describes the average conditional variance parameter, and $a$ describes how much the conditional means differ across risks. For a particular risk with parameter $\Theta_i$, use $\sigma^2(\Theta_i)$ in its conditional variance. The population quantity $\phi$ is an expectation across the parameter distribution, not necessarily the value for every individual risk.

Worked Example: Conditional and Unconditional Variation

A hypothetical portfolio uses annual claim-cost observations. Let $\Theta$ take the values 900 and 1,500 with probability $1/2$ each. Given $\Theta$, observations are independent and normally distributed with mean $\Theta$ and variance $360{,}000/w$, where $w$ is the exposure weight. A risk contributes observations with weights 1, 2 and 4. Find the population mean and variance of the risk means, then find the conditional and unconditional variance of each observation and the covariance between two different observations for the same risk.

Model and parameters. The conditional distribution is $X_j\mid\Theta\sim N(\Theta,360{,}000/w_j)$. There is no interest or mortality assumption in this example. Thus $\mu(\Theta)=\Theta$, $\sigma^2(\Theta)=360{,}000$, and the conditional independence assumption applies.

Population mean and variance of means. The two equally likely values give $m=(900+1{,}500)/2=1{,}200$. Their deviations from the mean are $-300$ and $300$, so $a=[(-300)^2+(300)^2]/2=90{,}000$. Since $\sigma^2(\Theta)$ is constant, $\phi=360{,}000$.

Conditional variances. For weights 1, 2 and 4, the conditional variances are respectively $360{,}000/1=360{,}000$, $360{,}000/2=180{,}000$, and $360{,}000/4=90{,}000$.

Unconditional variances. The law of total variance gives $\operatorname{Var}(X_j)=E[\operatorname{Var}(X_j\mid\Theta)]+\operatorname{Var}(E[X_j\mid\Theta])$. The conditional variance averages to $360{,}000/w_j$, and the variance of the conditional mean is $a=90{,}000$. Therefore the unconditional variances are $450{,}000$, $270{,}000$, and $180{,}000$, in weight order 1, 2 and 4.

Covariance check. For two distinct observations $j$ and $k$, conditional independence makes their conditional covariance zero. Their unconditional covariance is the variance of their common conditional mean, $90{,}000$. The increasing weights reduce the process-variance component, but do not remove the shared variation in the risk’s underlying mean.

Check. Each unconditional variance equals its conditional variance plus $90{,}000$, and each pair’s covariance equals $a$. The calculation is consistent with both components of the model.

Consequences of the Assumptions

Two useful unconditional moment results follow directly from the model. For an observation of weight $w$, the law of total expectation gives $E[X]=m$. Applying the law of total variance gives

\operatorname{Var}(X)=\operatorname{Var}(E[X\mid\Theta])+E[\operatorname{Var}(X\mid\Theta)] =a+\frac{\phi}{w}.

This separates total variation into between-risk variation, $a$, and average process variation, $\phi/w$. The first component does not depend on the observation’s weight; the second decreases as weight increases.

For two distinct observations $X_j$ and $X_k$ from the same risk, the law of total covariance gives

\operatorname{Cov}(X_j,X_k) =E[\operatorname{Cov}(X_j,X_k\mid\Theta)] +\operatorname{Cov}(E[X_j\mid\Theta],E[X_k\mid\Theta]) =a.

The first term is zero by conditional independence. The second is $\operatorname{Var}(\mu(\Theta))=a$, because both observations share the same conditional mean. Thus conditional independence does not imply marginal independence: observations from the same risk share uncertainty about $\Theta$.

Worked Example: A Weighted Summary Under the Model

For a hypothetical service-risk class, suppose $\Theta$ equals 8 or 12 with equal probability. Conditional on $\Theta$, two observations are independent normal variables with mean $\Theta$ and variance $200/w$. Their weights are 2 and 5, and their observed values are 9 and 11. Calculate the exposure-weighted average, its conditional variance, and its unconditional variance.

Model and parameters. Here $\mu(\Theta)=\Theta$, $\sigma^2(\Theta)=200$, $m=(8+12)/2=10$, and $a=[(8-10)^2+(12-10)^2]/2=4$. There is no interest or mortality assumption; the conditional distribution is normal as stated.

Weighted average. The total weight is $2+5=7$. The observed weighted average is $(2\cdot9+5\cdot11)/7=(18+55)/7=73/7\approx10.4286$.

Conditional variance. Given $\Theta$, the variance of this weighted average is $(2/7)^2(200/2)+(5/7)^2(200/5)$, because the observations are conditionally independent. The terms are $400/49$ and $1{,}000/49$, so the total is $1{,}400/49=200/7\approx28.5714$.

Unconditional variance. The conditional expectation of the weighted average is $\Theta$, so its variance across risks is $a=4$. Its unconditional variance is therefore $4+200/7=228/7\approx32.5714$.

Check. The weights sum to 7, and the conditional variance reduces to $200/7$, matching the inverse-total-weight result for this weighted average. Adding the variance of the conditional mean gives $32.5714$, as required.

Weights Must Use a Consistent Exposure Scale

The numerical value of a weight depends on its unit. If an exposure is recorded in years in one calculation and months in another, both the weight and the variance parameter must be converted consistently. The implied conditional variance must remain unchanged; changing the weight’s unit alone changes the model.

Worked Example: Converting Exposure Units Correctly

A hypothetical insurer models annual loss-cost observations with conditional normal distribution and variance $400/w$, where $w$ is measured in exposure-years. One observation has weight 0.25 years. Re-express the same exposure in months and determine the variance under the equivalent monthly-weight convention. Show the error from retaining the original variance parameter.

Model and assumptions. Conditional on the risk parameter, the mean is constant across the two representations, observations are conditionally independent, and the annual-scale variance parameter is 400. There is no interest or mortality assumption.

Annual convention. With $w=0.25$ years, the conditional variance is $400/0.25=1{,}600$.

Monthly convention. The same exposure is $0.25\cdot12=3$ months. To preserve variance, the monthly-scale parameter must be $400\cdot12=4{,}800$. The variance is then $4{,}800/3=1{,}600$, unchanged.

Common mistake and correction. Keeping the parameter at 400 while changing the weight to 3 gives $400/3\approx133.3333$, which incorrectly changes the implied variance. The correction is to scale the variance parameter by 12 along with the weight. The annual and monthly formulations then agree exactly.

Check. Both consistent formulations give conditional variance 1,600. The error arose from changing the unit of weight without changing the variance parameter.

Common Interpretation Traps

Making variance proportional to weight. The model specifies $\sigma^2(\Theta)/w$, not $w\sigma^2(\Theta)$. Greater exposure reduces conditional variance.
Changing the conditional mean with the weight. Weight affects precision in this model, not the expected value of the observed quantity.
Confusing conditional and unconditional variation. Conditional variance is $\sigma^2(\Theta)/w$. Unconditional variance also includes variation in the risk means.
Assuming marginal independence. Conditional independence does not eliminate the dependence caused by a shared risk parameter.
Mixing exposure units. Weights and their variance parameter must be defined on compatible scales.
Treating any volume measure as a valid weight. A weight must represent exposure or relative precision in a way that supports the inverse-variance assumption.

Key takeaway: Bühlmann-Straub models unequal precision by assigning each observation a positive weight and setting its conditional variance equal to the risk-specific variance parameter divided by that weight. The common conditional mean, inverse-weight variance, and conditional independence assumptions determine how observations relate; the weight alone does not determine the risk’s expected value.

Check Your Understanding

For each problem, use the stated conditional model and keep the exposure units consistent.

A hypothetical risk has conditional variance parameter $\sigma^2(\Theta)=500$. Two observations have weights 2 and 8. Find their conditional variances. (Answer: 250 and 62.5, respectively.)
A risk parameter is equally likely to produce conditional means 50 and 90. The conditional variance parameter is 900, and an observation has weight 3. Find $a$, $\phi$, and the observation’s unconditional variance. (Answer: $a=400$, $\phi=900$, and $\operatorname{Var}(X)=700$.)
Two conditionally independent observations have weights 1 and 3, conditional variance parameter 250, and observed values 4 and 7. Find their exposure-weighted average and its conditional variance. (Answer: weighted average $=6.25$; conditional variance $=62.5$.)
An exposure is 0.25 years, with annual-scale variance parameter 400. Convert to months and find the equivalent monthly-scale variance parameter and conditional variance. (Answer: weight $=3$ months, monthly-scale parameter $=4{,}800$, conditional variance $=1{,}600$.)
A risk’s conditional mean is equally likely to be 2 or 6, and its conditional variance parameter is 16. Two distinct observations have weights 2 and 8. Find their covariance and their respective unconditional variances. (Answer: covariance $=4$; unconditional variances $=12$ and $6$, respectively.)

What You'll Learn