Advanced Credibility Theory for Exam MAS-II

Hierarchical Credibility

Hierarchical credibility (Buhlmann-Straub with multiple levels) handles data organized in nested groups, such as policies within territories within states. At each level, credibility weights blend the group estimate with the next higher level's estimate. For a two-level hierarchy, the territory estimate is Z_territory * (territory mean) + (1-Z_territory) * (state estimate), and the state estimate is Z_state * (state mean) + (1-Z_state) * (countrywide mean). The credibility factors at each level use the within-group and between-group variance components appropriate to that level.

The logic behind hierarchical credibility follows naturally from the single-level Buhlmann-Straub model. At the lowest level, individual risk data is blended with the group (territory) average. At the next level, territory averages are blended with the state average. At the top level, state averages are blended with the countrywide average. Each blending step uses credibility weights determined by the volume of data and the relevant variance components at that level of the hierarchy.

Variance Components in Hierarchical Models

A two-level hierarchical model has three variance components. The process variance v measures randomness within a single risk. The between-risk variance a1 measures heterogeneity among risks within the same group (territory). The between-group variance a2 measures heterogeneity among groups (territories) within the larger pool (state). The credibility factor at the risk level depends on v and a1, while the credibility factor at the group level depends on the effective variance at that level and a2.

Estimating these variance components from data requires careful decomposition of the total sum of squares. The within-risk sum of squares estimates v. The between-risk, within-group sum of squares estimates a combination of v and a1. The between-group sum of squares estimates a combination of all three components. Solving these equations simultaneously yields unbiased estimators. As with the basic Buhlmann-Straub model, negative variance estimates are set to zero by convention.

Multidimensional Credibility

Multidimensional credibility handles multiple correlated observations per risk. For example, an insurer might observe both claim frequency and average severity for each policyholder, or might track losses across multiple coverage types (liability, collision, comprehensive). Rather than applying separate scalar credibility factors to each dimension, multidimensional credibility uses a credibility matrix Z that accounts for the correlations between dimensions.

The credibility premium vector is Z times the vector of individual observations plus (I minus Z) times the vector of collective means, where Z is a matrix rather than a scalar. The credibility matrix is Z = T * (T + S)^{-1}, where T is the covariance matrix of the hypothetical means (between-risk variation) and S is the expected value of the process covariance matrix divided by the number of observations. This formulation allows information from one dimension to improve the estimate in another dimension when the dimensions are correlated.

The Hachemeister Regression Model

Regression credibility (the Hachemeister model) allows the mean to depend on covariates, with credibility applied to the regression coefficients. For each risk i observed over time, the model is X_it = beta_i' * z_t + epsilon_it, where z_t is a vector of covariates (such as time, for modeling trends) and beta_i is a risk-specific coefficient vector. The credibility estimate of beta_i is Z_i * b_i + (I minus Z_i) * b, where b_i is the individual OLS estimate and b is the collective (pooled) estimate.

The Hachemeister model is particularly useful when risks exhibit different trends over time. For example, if territory A has loss costs increasing at 5% per year while territory B has loss costs increasing at 8% per year, the Hachemeister model can produce credibility-weighted trend estimates for each territory that blend its own observed trend with the overall average trend. This is more nuanced than applying a single credibility factor to the mean, which would miss the trend differences.

Connection to Mixed Effects Models

Advanced credibility models have a deep connection to mixed effects models in statistics. The Buhlmann model corresponds to a random intercept model where each risk has a random mean. The Hachemeister model corresponds to a random coefficient model where each risk has random regression coefficients. Hierarchical credibility corresponds to a nested random effects model with multiple grouping levels. Understanding this connection provides theoretical grounding for credibility methods and connects them to a broader statistical framework.

In practice, this connection means that credibility estimates can be obtained using standard mixed effects model software, which can handle more complex data structures (unbalanced designs, missing data, additional covariates) than the classical credibility formulas. However, the classical formulas remain important for Exam MAS-II because they provide closed-form solutions that are testable in a written exam setting.

Common Pitfalls

Students often struggle with the notation and bookkeeping in hierarchical credibility calculations. Keep careful track of which level of the hierarchy each variance component belongs to. Another common error is confusing the credibility matrix in multidimensional credibility with a scalar applied independently to each dimension. The matrix formulation allows cross-dimensional information sharing, which a collection of independent scalars does not.

In the Hachemeister model, a frequent mistake is applying credibility to the predicted values rather than to the regression coefficients. The correct approach credibility-weights the coefficients, then uses the credibility-weighted coefficients to produce predictions. This distinction matters because it ensures that the credibility-weighted model produces sensible predictions at covariate values not observed in the individual risk's data.

Study Tips

Exam MAS-II tests hierarchical models, multidimensional credibility, and the Hachemeister regression approach. For hierarchical credibility, practice computing variance components from data organized in nested groups and producing credibility-weighted estimates at each level. For multidimensional credibility, work through examples with two dimensions (such as frequency and severity) and verify that the matrix formula reduces to the scalar formula when the dimensions are independent. For the Hachemeister model, practice with simple trend models (intercept plus slope) and compute credibility-weighted trends for individual risks.