Why Unequal-Exposure Variance Estimation Needs Care
In Bühlmann-Straub credibility, each risk can have multiple observations, and each observation can represent a different amount of exposure. The exposure weights affect both the variability of a risk’s weighted mean and the amount of information it contributes. Method-of-moments estimation therefore has two distinct tasks: estimate process variation from variation within risks, then estimate variation among risk means after accounting for unequal total exposures and process noise.
The calculation is valid only under specific assumptions. In particular, the errors must be uncorrelated across risks as well as across periods within a risk. Without that condition, the usual correction for process noise in the between-risk sum of squares need not hold.
For risk \(i\), let \(n_i\) be its number of observed periods, \(W_i=\sum_j w_{ij}\) its total exposure, and \(\bar X_i=\sum_jw_{ij}X_{ij}/W_i\) its exposure-weighted mean. Let \(r\) denote the number of risks with positive exposure. These are the quantities used to construct the two moment estimates.
The Two Moment Estimators
The process variance estimate uses weighted deviations of observations from their own risk’s weighted mean. The variance of hypothetical means estimate uses deviations of risk means from an exposure-weighted center, with a correction for the process noise contained in those risk means.
To define the second estimate, let \(W=\sum_i W_i\), \(\bar X_W=\sum_iW_i\bar X_i/W\), and
Then the method-of-moments estimate of \(v\) is
When both estimates are positive, the estimated structural constant is \(\widehat K=\widehat a/\widehat v\). It is expressed in the exposure units used for the weights. The denominator \(C\) is positive when at least two risks have positive exposure; if all exposure is concentrated in one risk, there is no between-risk information for estimating \(v\).
Why the Corrections Work
For one risk, the weighted residual sum of squares is \(\sum_jw_{ij}(X_{ij}-\bar X_i)^2\). Since \(\bar X_i\) is a weighted average, its conditional variance is \(a/W_i\). The expected residual sum of squares is therefore
Summing over risks gives expectation \(a\sum_i(n_i-1)\). Dividing the observed within-risk sum of squares by that degrees-of-freedom total gives the stated unbiased moment estimate of \(a\). A risk with only one observation contributes zero within-risk degrees of freedom.
For the between-risk calculation, write \(\bar X_i=\mu_i+\bar\epsilon_i\). The risk-effect part of \(\sum_iW_i(\bar X_i-\bar X_W)^2\) has expectation \(vC\). This follows by writing the weighted centered sum as \(\sum_iW_i\mu_i^2-(\sum_iW_i\mu_i)^2/W\): the first term has expectation \(vW+m^2W\), and the second has expectation \(v\sum_iW_i^2/W+m^2W\). Their difference is \(vC\).
The weighted risk-mean error has variance \(a/W_i\). Its contribution to the centered sum has expectation
The cross term between risk effects and errors has expectation zero under the stated assumptions. Thus the expected between-risk sum is \(vC+(r-1)a\), which explains both the process-noise subtraction and the unequal-exposure denominator. Replacing \(a\) by \(\widehat a\) yields the method-of-moments estimate of \(v\).
Applying the Estimators
Worked Example: Estimating Both Variance Components
A hypothetical portfolio has three risks, each observed in two periods. The exposure weights and observed rates are: Risk A, weights \(1,3\) and rates \(10,14\); Risk B, weights \(2,2\) and rates \(8,12\); Risk C, weights \(1,3\) and rates \(12,16\). Assume the model and independence conditions stated above. Calculate \(\widehat a\), \(\widehat v\), and \(\widehat K\).
Step 1: Calculate the risk means and within-risk sums. Risk A has \(\bar X_A=(1(10)+3(14))/4=13\), and its within-risk sum is \(1(10-13)^2+3(14-13)^2=9+3=12\). Risk B has \(\bar X_B=(2(8)+2(12))/4=10\), and its sum is \(2(8-10)^2+2(12-10)^2=8+8=16\). Risk C has \(\bar X_C=(1(12)+3(16))/4=15\), and its sum is \(1(12-15)^2+3(16-15)^2=9+3=12\).
Step 2: Estimate process variance. The total within-risk sum is \(12+16+12=40\). There are six observations and three risks, so the degrees-of-freedom total is \((2-1)+(2-1)+(2-1)=3\). Therefore \(\widehat a=40/3=13.3333\).
Step 3: Calculate the between-risk quantities. Each \(W_i=4\), so \(W=12\) and \(\sum_iW_i^2=48\). Thus \(C=12-48/12=8\). The exposure-weighted center is \(\bar X_W=(4(13)+4(10)+4(15))/12=38/3\). The between-risk sum is \(4(13-38/3)^2+4(10-38/3)^2+4(15-38/3)^2=456/9=50.6667\).
Step 4: Estimate \(v\) and \(K\). The process-noise correction is \((3-1)(13.3333)=26.6667\). Hence \(\widehat v=(50.6667-26.6667)/8=3.0000\), and \(\widehat K=13.3333/3.0000=4.4444\), rounded to four decimal places.
Check and method. The corrected between-risk sum is \(24\), and \(24/C=24/8=3\), as calculated. This method uses the unequal-exposure moment equations; omitting the process-noise correction would overstate the estimate of \(v\).
Worked Example: Unequal Total Exposures Change the Denominator
For a second hypothetical portfolio, three risks have these observations: Risk A has weights \(1,1\) and rates \(8,12\); Risk B has weights \(1,3\) and rates \(13,17\); Risk C has weights \(2,2\) and rates \(14,10\). Estimate the variance components and structural constant under the stated model.
Step 1: Find the weighted means and residual sums. Risk A has mean \((8+12)/2=10\) and within sum \(1(8-10)^2+1(12-10)^2=8\). Risk B has mean \((13+3(17))/4=16\) and within sum \(1(13-16)^2+3(17-16)^2=12\). Risk C has mean \((2(14)+2(10))/4=12\) and within sum \(2(14-12)^2+2(10-12)^2=16\).
Step 2: Estimate \(a\). The within sum is \(8+12+16=36\), and the degrees-of-freedom total is \(3\). Thus \(\widehat a=36/3=12\).
Step 3: Use the actual total exposures. The risk totals are \(2,4,4\), so \(W=10\), \(\sum_iW_i^2=4+16+16=36\), and \(C=10-36/10=6.4\). The weighted center is \((2(10)+4(16)+4(12))/10=132/10=13.2\). The between-risk sum is \(2(10-13.2)^2+4(16-13.2)^2+4(12-13.2)^2=20.48+31.36+5.76=57.6\).
Step 4: Correct for process noise. The correction is \((3-1)(12)=24\), so \(\widehat v=(57.6-24)/6.4=33.6/6.4=5.25\). Therefore \(\widehat K=12/5.25=2.2857\), rounded to four decimal places.
Check and method. The moment equation predicts a between-risk sum of \(5.25(6.4)+24=57.6\), matching the observed sum. Using \(W-r\) as the denominator would be incorrect: unequal risk totals require \(C=W-\sum_iW_i^2/W\).
Worked Example: Estimation From Risk-Level Summaries
A third hypothetical dataset contains four risks with differing numbers of periods. A summary table gives each risk’s period count, total exposure, weighted mean, and contribution to the within-risk sum. Assume the model conditions hold.
| Risk | Periods \(n_i\) | Total exposure \(W_i\) | Weighted mean \(\bar X_i\) | Within-risk sum |
|---|---|---|---|---|
| A | 2 | 2 | 20 | 6 |
| B | 2 | 3 | 25 | 8 |
| C | 3 | 4 | 18 | 12 |
| D | 2 | 5 | 30 | 24 |
Step 1: Estimate \(a\). The within-risk sum is \(6+8+12+24=50\). The degrees-of-freedom total is \(1+1+2+1=5\), so \(\widehat a=50/5=10\).
Step 2: Calculate \(C\) and the between-risk sum. Here \(W=2+3+4+5=14\), and \(\sum_iW_i^2=4+9+16+25=54\). Thus \(C=14-54/14=71/7\). The weighted center is \((2(20)+3(25)+4(18)+5(30))/14=337/14\). To calculate the centered sum, use the equivalent expression \(\sum_iW_i\bar X_i^2-(\sum_iW_i\bar X_i)^2/W\). The first term is \(2(400)+3(625)+4(324)+5(900)=8471\); the second is \(337^2/14=113569/14\). The between-risk sum is \(8471-113569/14=5025/14=358.9286\).
Step 3: Estimate \(v\) and \(K\). The process correction is \((4-1)(10)=30\). Therefore \(\widehat v=(5025/14-30)/(71/7)=(4605/14)/(142/14)=4605/142=32.4296\). Then \(\widehat K=10/32.4296=1420/4605=0.3084\), rounded to four decimal places.
Check and method. Using the unrounded value, \(\widehat v C+3\widehat a=(4605/142)(142/14)+30=4605/14+420/14=5025/14\), which recovers the observed between-risk sum.
Interpretation, Limits, and Exam Traps
The structural constant compares process variance with variance among hypothetical means. Its interpretation depends on the units of exposure: changing the exposure unit changes the numerical value of \(a\) and \(K\), so the weights and resulting constant must be expressed consistently. The method-of-moments estimates are sample estimates, not guaranteed nonnegative parameters. If the corrected between-risk sum is negative, then \(\widehat v<0\); that is a warning about sampling variation, weak between-risk evidence, or a model/data mismatch. It is not a meaningful negative variance. A modeling or estimation policy must be specified before replacing such an estimate by a boundary value.
The standard equations also do not automatically apply if process variance differs by risk, exposure weights do not scale variance inversely, periods are dependent, or risk effects are not drawn from a common population with a common variance. Those conditions change the expected sums of squares and therefore require a different estimator or model.
A reliable calculation sequence is to compute each weighted risk mean and within-risk sum first, estimate \(a\), then calculate \(W\), \(\sum_iW_i^2\), \(C\), and the centered between-risk sum. Apply the process correction only after those quantities are established. This order makes it easier to detect an exposure-weighting error or an omitted correction.
Confirm fixed positive exposure weights, a common process variance parameter, and uncorrelated process errors across all observations.
Sum exposure-weighted within-risk squared deviations and divide by \(\sum_i(n_i-1)\).
Use \(C=W-\sum_iW_i^2/W\) and subtract \((r-1)\widehat a\) from the between-risk sum.
Divide by \(C\) to obtain \(\widehat v\), then calculate \(\widehat K=\widehat a/\widehat v\) when both estimates are positive. Check whether the estimates are plausible under the assumptions.
These moment estimates translate an unbalanced risk-period dataset into estimates of within-risk noise and between-risk heterogeneity. The central safeguards are to retain the exposure weights throughout and to use the cross-risk independence assumption when applying the process-noise correction. This estimation framework is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
Check Your Understanding
For each problem, assume the stated Bühlmann-Straub moment model applies unless the question asks you to assess an assumption. Report numerical estimates to four decimal places where needed.
- A hypothetical dataset has three risks, each observed in two periods. Its total exposure-weighted within-risk sum of squares is 18. Calculate \(\widehat a\). (Answer: \(\widehat a=6\))
- Three risks have total exposures \(1,2,3\), weighted means \(5,8,9\), and total within-risk degrees of freedom 4. The within-risk sum of squares is 20. Calculate \(\widehat a\), \(C\), the between-risk sum, \(\widehat v\), and \(\widehat K\). (Answer: \(\widehat a=5,\ C=11/3,\ B=12,\ \widehat v=6/11=0.5455,\ \widehat K=55/6=9.1667\))
- Four risks have total exposures \(1,2,3,4\), weighted means \(10,12,15,20\), and total within-risk degrees of freedom 6. The within-risk sum of squares is 24. Calculate \(\widehat a\), \(C\), the between-risk sum, \(\widehat v\), and \(\widehat K\). (Answer: \(\widehat a=4,\ C=7,\ B=134.9,\ \widehat v=17.5571,\ \widehat K=0.2278\))
- Two risks each have one observation of weight 1. Their observed values are \(10+Z\) and \(14+Z\), where \(E[Z]=0\) and \(\operatorname{Var}(Z)=a\). What is the centered between-risk sum contributed by the errors, and can the usual \((r-1)a\) correction be used? (Answer: The error contribution is 0; no, the perfectly shared errors violate the required cross-risk uncorrelatedness.)
- Four risks have total exposures \(2,3,4,5\), weighted means \(20,25,18,30\), and within-risk sums \(6,8,12,24\), respectively. Their period counts are \(2,2,3,2\). Calculate \(\widehat a\), \(C\), \(\widehat v\), and \(\widehat K\). (Answer: \(\widehat a=10,\ C=71/7,\ \widehat v=32.4296,\ \widehat K=0.3084\))