This guide connects statistical assumptions with calculations and model interpretation for candidates preparing for the Casualty Actuarial Society Modern Actuarial Statistics II Examination (CAS MAS-II). Each concept explains a rule, resolves an original example and identifies a specific error to avoid. The foundations progress from linear models and Bayesian inference to credibility, simulation, mixed models and multivariate analysis.
Linear models and Bayesian foundations
1. Likelihood conditions on the observed data
A sampling distribution describes possible observations for a fixed parameter. A likelihood uses the same expression as a function of the parameter after observations are fixed. Factors constant in the parameter can be omitted when comparing parameter values, but parameter-dependent factors must remain.
Worked example: For three Poisson events over two exposure units, the likelihood for rate λ is proportional to λ³ exp(−2λ). Its maximum occurs at λ = 3/2.
Mistake to avoid: Treating a likelihood as a probability distribution over parameters without specifying a prior.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
2. Least squares is a projection
Ordinary least squares projects the response onto the column space of the design matrix. With full column rank, the coefficient estimate is unique and residuals are orthogonal to every design column. An intercept therefore makes the residuals sum to zero; it does not make every residual small.
Worked example: For x = 0, 1, 2 and y = 1, 3, 2, the fitted line is 1.5 + 0.5x. Residuals are −0.5, 1, −0.5.
Mistake to avoid: Assuming coefficients are uniquely identifiable when one design column duplicates another.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
3. Interactions change conditional effects
A regression coefficient describes a change holding the other modeled predictors fixed. When an interaction is present, a predictor's effect depends on the interacting variable. Interpret the complete expression rather than reading the main-effect coefficient as a universal effect.
Worked example: In E[Y | x,z] = 10 + 2x + 3z + xz, increasing x by one changes the mean by 2 when z = 0 and by 3 when z = 1.
Mistake to avoid: Reporting the coefficient 2 as the effect of x for every value of z.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
4. Mean uncertainty differs from prediction uncertainty
Under a normal linear model, uncertainty about a fitted mean reflects coefficient estimation. Predicting a new independent response also includes its residual variance. Leverage affects mean uncertainty, so predictions far from the observed predictor combinations can be less precise even when the fitted line appears stable.
Worked example: If fitted-mean variance is 0.4 and new-response residual variance is 4, prediction variance is 4.4, compared with 0.4 for the mean.
Mistake to avoid: Using a fitted-mean interval as an interval for an individual future observation.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
5. Bayesian updating multiplies odds by evidence
For two competing hypotheses, posterior odds equal prior odds multiplied by the likelihood ratio. This separates prior plausibility from how strongly the observations favor one hypothesis. The likelihood ratio must compare the probability or density of the same observed evidence under both hypotheses.
Worked example: Prior odds of high risk to low risk are 1:4. Evidence is three times as likely under high risk, giving posterior odds 3:4 and high-risk probability 3/7.
Mistake to avoid: Multiplying prior probabilities directly by a likelihood ratio and failing to renormalize.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
6. Beta-binomial updating adds successes and failures
For independent Bernoulli observations conditional on probability p, a Beta(a,b) prior gives a Beta(a+s,b+f) posterior, where s and f are observed successes and failures. The posterior mean combines both kinds of evidence. The parameterization matters because a and b play different roles.
Worked example: A Beta(2,8) prior and three claims among ten independent policies produce Beta(5,15). The posterior mean claim probability is 5/20 = 0.25.
Mistake to avoid: Adding the total number of policies to b instead of adding only the seven failures.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
7. Gamma-Poisson updating respects exposure
Suppose a count has a Poisson distribution with mean eλ, where e is known exposure. Using a Gamma prior with shape a and rate b gives posterior shape a+y and rate b+e. Counts and exposure enter different parameters, and rate must not be confused with scale.
Worked example: With prior shape 2 and rate 1, seven events over exposure 3 give posterior shape 9 and rate 4. The posterior mean rate is 2.25.
Mistake to avoid: Adding exposure to the shape parameter or interpreting rate 4 as scale 4.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
8. Normal updating weights information by precision
With a normal prior and a normal observation having known sampling variance, posterior precision equals prior precision plus data precision. The posterior mean is their precision-weighted mean. Here precision means reciprocal variance, not reciprocal standard deviation.
Worked example: A prior mean of 100 with variance 25 and an observed estimate of 120 with variance 100 give posterior mean 104 and variance 20.
Mistake to avoid: Giving the noisier observation more weight merely because its numerical value is larger.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
9. Credible intervals depend on the posterior distribution
A Bayesian credible interval contains a specified posterior probability for the parameter under the model and prior. Equal-tailed and highest-density intervals can differ for skewed or multimodal posteriors. A frequentist confidence procedure instead describes repeated-sampling coverage; its interpretation is different.
Worked example: For an exponential posterior with rate 1, an equal-tailed 90% interval is approximately [0.0513, 2.9957], obtained from its 5th and 95th percentiles.
Mistake to avoid: Assuming equal-tailed intervals always contain the values with greatest posterior density.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
10. Posterior prediction integrates parameter uncertainty
A posterior predictive distribution averages the future-response distribution over the posterior parameter distribution. Its variance follows the law of total variance: expected conditional variance plus variance of the conditional mean. Substituting a posterior mean parameter generally omits the second component.
Worked example: For a future Poisson count with posterior rate mean 3 and variance 2, predictive mean is 3 and predictive variance is 3 + 2 = 5.
Mistake to avoid: Calling a Poisson distribution with mean 3 the complete predictive distribution in this example.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
11. The loss function determines the Bayesian estimate
A Bayesian action minimizes posterior expected loss. Squared-error loss selects the posterior mean; absolute-error loss selects a posterior median. Asymmetric absolute loss selects an appropriate posterior quantile. The best action therefore depends on the consequence of error as well as the posterior distribution.
Worked example: A posterior assigns probabilities 0.6 and 0.4 to losses 0 and 10. Squared-error estimation gives 4; absolute-error estimation gives 0.
Mistake to avoid: Assuming the posterior mode, mean and optimal decision must be the same.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
12. Prior predictive checks examine implied observations
A prior predictive distribution combines the prior with the sampling model before using the observed data. It reveals whether a prior implies implausible response magnitudes or variability. Sensitivity analysis then checks how reasonable alternative priors change quantities relevant to the decision.
Worked example: A Poisson rate with Gamma shape 2 and rate 0.1 implies prior predictive mean 20 for one exposure unit. That conflicts with a proposed prior belief of roughly two events.
Mistake to avoid: Judging a prior solely by how its parameter-density plot looks.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
Credibility and heterogeneous insurance experience
13. Process variation differs from risk heterogeneity
Repeated experience can fluctuate within a risk even when its underlying parameter is unchanged. Different risks can also have different conditional means. The law of total variance separates these sources into expected conditional variance and variance of conditional means, providing the foundation for credibility models.
Worked example: If within-risk variance averages 9 and risk means have variance 4, a single observation has unconditional variance 13.
Mistake to avoid: Attributing all observed portfolio variability to differences between underlying risk levels.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
14. EPV and VHM describe different populations
Expected process variance, EPV, averages conditional observation variance across the risk population. Variance of hypothetical means, VHM, measures dispersion of conditional risk means. Credibility depends on their ratio, so the population from which risks are assumed to arise must be defined consistently.
Worked example: Two equally likely risk types have means 2 and 6 and conditional variances 1 and 9. EPV is 5, while VHM is 4.
Mistake to avoid: Calculating VHM from conditional variances instead of from conditional means.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
15. Bühlmann credibility balances individual and collective means
For conditionally independent observations with a common conditional mean and variance within each risk, Bühlmann credibility predicts the risk mean as Z times its sample mean plus (1−Z) times the collective mean. With positive VHM, Z = n/(n+K), where K = EPV/VHM.
Worked example: EPV = 12, VHM = 3 and n = 4 give K = 4 and Z = 0.5. Individual mean 8 and collective mean 5 produce estimate 6.5.
Mistake to avoid: Reversing EPV and VHM when calculating K.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
16. Bühlmann-Straub credibility uses exposure weights
Bühlmann-Straub models allow observations with different known weights. Conditional variance is inversely proportional to weight, and the risk's experience mean is exposure-weighted. Credibility uses total weight rather than the number of periods. Weights must genuinely represent information under the assumed variance structure.
Worked example: Rates 4 and 7 with weights 1 and 2 give weighted mean 6. With total weight 3, K = 3 and collective mean 5, the credibility estimate is 5.5.
Mistake to avoid: Averaging the two rates equally while calculating credibility from their total exposure.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
17. Observed variation between sample means includes noise
For independent risks with equal sample size n, the variance of their sample means equals VHM + EPV/n. Estimating VHM therefore requires removing the within-risk sampling contribution. Sampling error can make a raw variance-component estimate negative; that signals estimation uncertainty rather than a negative population variance.
Worked example: Estimated variance of risk means is 5, EPV is 8 and n = 4. The implied VHM estimate is 5 − 8/4 = 3.
Mistake to avoid: Using variance 5 directly as VHM and overstating between-risk heterogeneity.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
18. Full credibility is tied to a probability criterion
Limited-fluctuation full credibility requires an estimate to fall within a specified relative error with a specified probability. For a homogeneous Poisson count with sufficiently large mean, a normal approximation gives required expected count approximately (z/r)², where r is relative tolerance and z matches the chosen central probability.
Worked example: Using z = 1.96 and r = 0.10 gives a required expected count of approximately 384.16.
Mistake to avoid: Applying this Poisson-count calculation directly to aggregate severity without accounting for severity variability.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
19. Partial credibility is a selected convention
A common limited-fluctuation convention assigns Z = min(1, √(n/n₀)), where n₀ is the full-credibility volume under the stated criterion. This convention differs from Bühlmann's minimum-mean-square-error weight. State the method and volume definition rather than treating every credibility factor as interchangeable.
Worked example: With volume 100 and full-credibility volume 400, square-root partial credibility is 0.5. An experience estimate of 12 blended with reference 8 gives 10.
Mistake to avoid: Replacing Bühlmann's n/(n+K) with the square-root rule without changing the model.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
20. Credibility requires comparable experience
Credibility blends estimates on a common measurement basis. Exposure definitions, coverage, inflation and portfolio mix can make historical and collective experience incomparable. Normalize or model these differences before applying a weight; otherwise additional observations can increase confidence in a systematically biased comparison.
Worked example: A historical cost of 100 requires a 10% trend adjustment to the target period. Its comparable value is 110; equal blending with reference 120 gives 115.
Mistake to avoid: Blending unadjusted historical cost 100 with target-period cost 120.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
MCMC computation and diagnostics
21. MCMC can target an unnormalized posterior
Many MCMC methods need a function proportional to the posterior rather than its normalizing constant. Ratios cancel a constant shared by current and proposed states. The remaining function must still define a proper target distribution, and every parameter-dependent contribution must be retained.
Worked example: For target density proportional to exp(−θ²/2), moving from θ = 1 to θ = 2 gives target ratio exp(−1.5).
Mistake to avoid: Dropping a parameter-dependent determinant because it resembles a normalizing term.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
22. Metropolis-Hastings corrects proposal asymmetry
Metropolis-Hastings accepts a proposal with probability min(1, target ratio × reverse-proposal density / forward-proposal density). The proposal correction is essential when movement probabilities are asymmetric. A rejected proposal leaves the chain at its current state; that repeated state belongs in the retained sequence.
Worked example: Current and proposed target densities are 0.2 and 0.4. Reverse and forward proposal densities are 0.1 and 0.4, giving acceptance probability 0.5.
Mistake to avoid: Omitting the proposal ratio or deleting repeated states after rejection.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
23. Proposal scale controls exploration
For symmetric random-walk proposals, the proposal ratio cancels. Very small steps may be accepted often while exploring slowly; very large steps may be rejected often. Evaluate movement and dependence in the quantities of interest rather than treating a high acceptance rate as proof of efficient sampling.
Worked example: A positive-rate model proposes λ = −0.2 from λ = 0.1. Because the proposal lies outside the target support, it is rejected.
Mistake to avoid: Clipping an invalid proposal to zero without accounting for the changed proposal mechanism.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
24. Gibbs sampling updates full conditional distributions
A Gibbs step samples one parameter or block from its distribution conditional on all other current values and the data. Sequential updates immediately use the newly sampled values. Drawing full conditionals avoids a separate acceptance decision, but strong dependence can still make exploration slow.
Worked example: For parameters a and b, first draw a* conditional on the old b, then draw b* conditional on a*. The second update uses a*, not the old a.
Mistake to avoid: Updating both parameters from old conditioning values and assuming this is ordinary sequential Gibbs sampling.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
25. Blocking can address correlated parameters
A blocked update samples related parameters jointly rather than one at a time. This can help a chain follow a narrow correlated posterior region. The block must use the correct joint conditional distribution or a valid joint Metropolis-Hastings step; independent draws from marginal distributions generally destroy dependence.
Worked example: Two unit-variance normal parameters with correlation 0.95 have conditional variance 1−0.95² = 0.0975. A joint update can move along their shared ridge.
Mistake to avoid: Sampling each marginal independently and calling the result a valid correlated block.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
26. Transformations require a Jacobian
Reparameterizing a constrained quantity can improve sampling, but the transformed density must include the absolute Jacobian determinant. For λ = exp(η), the density for η is the density for λ evaluated at exp(η), multiplied by exp(η). On the log scale, this adds η.
Worked example: If λ has an exponential density with rate 1, the transformed log density is η−exp(η)+constant. At η = log(2), its variable part is log(2)−2.
Mistake to avoid: Using only −exp(η) and thereby targeting the wrong distribution.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
27. Effective sample size measures dependence for a quantity
Effective sample size compares the precision of a correlated-chain average with that of independent draws. It depends on autocorrelations of the particular monitored quantity, so a mean, tail indicator and variance estimate can have different effective sample sizes. The number of stored iterations alone does not establish precision.
Worked example: For an idealized stationary AR(1) sequence with correlation 0.6, integrated autocorrelation time is 4. Thus 2,000 draws provide approximately 500 effective draws.
Mistake to avoid: Assigning one parameter's effective sample size to every posterior summary.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
28. Multiple chains reveal incomplete exploration
Chains started in dispersed plausible regions can expose different modes, slow movement or disagreement in location and scale. Inspect traces and suitable between-chain diagnostics together. Agreement supports computational assessment but cannot prove that every important region was reached, especially when all starts are similar.
Worked example: Two chains remain near −3 while two remain near 3. Their pooled mean is near zero, although no chain explores both regions.
Mistake to avoid: Reporting the pooled mean as stable merely because it changes little when iterations are added.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
29. Monte Carlo error differs from posterior uncertainty
Posterior standard deviation describes uncertainty under the statistical model. Monte Carlo standard error describes numerical uncertainty in an estimated posterior summary. For a mean, an approximate Monte Carlo standard error is posterior standard deviation divided by the square root of its effective sample size.
Worked example: Posterior standard deviation 2 and effective sample size 100 give approximate mean-estimation error 0.2. More simulation reduces 0.2, not the posterior spread 2.
Mistake to avoid: Presenting Monte Carlo error as the uncertainty interval for the underlying parameter.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
30. Conditional averaging removes unnecessary simulation noise
When estimating a predictive event probability, average its conditional probability over posterior draws if that probability is available. This avoids extra randomness from simulating an event for every draw. It does not resolve poor posterior exploration or eliminate the numerical error caused by a finite correlated chain.
Worked example: Three posterior states imply event probabilities 0.1, 0.5 and 0.9. Their conditional average is 0.5 without generating three additional binary outcomes.
Mistake to avoid: Using extra predictive simulations when the desired conditional expectation can be calculated directly.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
Linear mixed models
31. Fixed effects and random effects have different roles
A linear mixed model writes Y = Xβ + Zb + ε. Fixed effects β describe modeled systematic means; random effects b represent structured deviations drawn from an assumed population distribution. Their distinction concerns the model's estimand and pooling assumptions, rather than whether the observed grouping labels change over time.
Worked example: A model uses a fixed fleet-size effect and a random fleet intercept. Fleets share the size relationship while retaining modeled fleet-specific deviations.
Mistake to avoid: Choosing random effects solely because there are many grouping categories.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
32. Random intercepts induce within-group covariance
For Yᵢⱼ = xᵢⱼ'β + bᵢ + εᵢⱼ with independent random intercepts and residuals, two distinct observations in the same group have covariance τ². Each observation has variance τ²+σ². Their intraclass correlation is τ²/(τ²+σ²), assuming the stated common variance structure.
Worked example: Random-intercept variance 2 and residual variance 6 give total variance 8 and within-group correlation 0.25.
Mistake to avoid: Treating observations sharing a random intercept as independent after marginalizing over that intercept.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
33. Random slopes make covariance depend on predictors
With a random intercept and slope, within-group covariance depends on both observations' predictor values. If their covariance matrix has entries τ₀₀, τ₀₁ and τ₁₁, covariance is τ₀₀+(xⱼ+xₖ)τ₀₁+xⱼxₖτ₁₁. An intercept-only correlation pattern cannot generally reproduce this structure.
Worked example: With τ₀₀ = 1, τ₀₁ = 0 and τ₁₁ = 0.25, observations at x = 1 and x = 3 have covariance 1.75.
Mistake to avoid: Assuming a random slope adds the same covariance to every pair of observations.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
34. Nested and crossed groups encode different membership
Grouping factors are nested when each level of one belongs to only one level of the other. They are crossed when membership can overlap across levels. This determines which observations share random effects and therefore which covariance terms are present. Identify membership from the data structure before selecting model notation.
Worked example: Adjusters permanently assigned to one branch are nested within branches. Adjusters handling cases in several branches create crossed adjuster and branch effects.
Mistake to avoid: Declaring nesting simply because one grouping factor has fewer categories.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
35. Generalized least squares uses marginal covariance
Given covariance V, generalized least squares estimates fixed effects using V⁻¹ rather than treating every observation as independent with equal variance. In a random-intercept model, additional observations within one group have diminishing information about the population mean because they share a group deviation.
Worked example: Independent group means 10 and 16 have variances 3 and 1.5. Their inverse-variance weighted estimate is 14, with normalized weights 1/3 and 2/3.
Mistake to avoid: Weighting correlated group data solely by the number of individual observations.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
36. BLUPs can borrow information across random coefficients
With known covariance components, the best linear unbiased predictor is GZ'V⁻¹(Y−Xβ). It predicts random effects from fixed-effect-adjusted residuals. Correlation within G can make an observation informative about a random coefficient that does not directly enter that observation.
Worked example: An intercept-only residual of 3 has residual variance 2. Random intercept and slope variances are 1 with covariance 0.5. Their predicted effects are 1 and 0.5.
Mistake to avoid: Predicting each random coefficient independently when the model specifies correlated coefficients.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
37. REML comparisons require compatible fixed effects
Maximum likelihood fits covariance parameters using the full response likelihood. Restricted maximum likelihood accounts for estimation of fixed effects through error contrasts. REML criteria depend on the fixed-effect design, so their values should not be directly compared across models with different fixed-effect structures.
Worked example: To compare models with and without a fixed age effect using likelihood criteria, fit both by maximum likelihood. REML can then estimate covariance in the selected structure.
Mistake to avoid: Selecting fixed predictors by comparing REML likelihood values from different design matrices.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
38. Contrasts need coefficient covariances
A fixed-effect contrast c'β has estimated variance c'Cov(β̂)c. Covariances matter when combining coefficients, so individual standard errors cannot simply be added or subtracted. First define the scientific comparison, then construct its contrast on the fitted coefficient scale.
Worked example: Estimates 2 and 3 have variances 0.4 and 0.9 and covariance 0.2. Their difference is −1 with variance 0.4+0.9−0.4 = 0.9.
Mistake to avoid: Ignoring the covariance term when calculating uncertainty for a coefficient difference.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
39. Marginal and conditional residuals diagnose different structure
Marginal residuals subtract the fixed-effect prediction. Conditional residuals also subtract predicted random effects. The former retain group deviations; the latter help assess remaining observation-level structure. Residuals may need standardization or whitening according to the fitted covariance before comparing patterns across observations.
Worked example: An observation is 12, its fixed prediction is 10 and its predicted group effect is 1.5. Marginal residual is 2; conditional residual is 0.5.
Mistake to avoid: Reading expected group patterns in marginal residuals as evidence that random effects are absent.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
40. Zero variance estimates lie on a boundary
Variance components cannot be negative, making zero a boundary of the parameter space. Ordinary interior-parameter likelihood-test approximations need not apply when testing a zero random-effect variance. A boundary estimate may reflect weak information or an unnecessary component; inspect identifiability and model structure before interpreting it.
Worked example: An estimated random-intercept variance of zero removes that component's within-group covariance. Residual variance 4 still permits substantial individual variation.
Mistake to avoid: Interpreting a zero group-variance estimate as proof that the response has no variability.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
41. Existing-group and new-group predictions have different information
An existing-group prediction can use that group's observed history to predict its random effect. A new-group mean prediction averages over the random-effect distribution, whose mean is commonly zero. A predictive distribution for the new group's response must still include random-effect variability and residual variability.
Worked example: Fixed prediction 10 and predicted existing-group effect 2 give mean prediction 12. A new group has mean prediction 10, with additional group-level uncertainty.
Mistake to avoid: Setting a new group's random effect to zero and also removing its predictive variance.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
42. Residual variance can differ across observations
A mixed model can specify a residual covariance matrix R that allows unequal variances or additional dependence. This is separate from random-effect covariance G. Incorrect residual assumptions can distort weighting, uncertainty and variance-component estimates even if the fixed-effect mean structure is appropriate.
Worked example: Independent observations 10 and 20 share a mean but have known variances 1 and 9. Their inverse-variance weighted mean is 11.
Mistake to avoid: Using an unweighted mean of 15 while claiming to use the specified unequal-variance model.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
Principal components and latent factors
43. Centering and scaling define the PCA question
PCA typically centers variables so components describe variation around their means. Covariance-based PCA preserves original units and emphasizes high-variance variables. Standardizing variables produces correlation-based PCA, giving each variable unit variance. The choice should reflect the intended analysis rather than a universal preference.
Worked example: Variances 10,000 for cost and 0.04 for a proportion make cost dominate covariance PCA. Standardization removes that unit-driven variance disparity.
Mistake to avoid: Changing measurement units and assuming covariance PCA must produce unchanged components.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
44. Principal component scores are projections
For centered data with covariance matrix S, principal directions are unit eigenvectors of S. A score is the projection of an observation onto a direction. The direction associated with the largest eigenvalue maximizes projected variance among unit-length linear combinations.
Worked example: For S = [[2,1],[1,2]], the first direction is (1,1)/√2 with eigenvalue 3. Centered observation (2,0) has first-component score √2.
Mistake to avoid: Using an unnormalized direction and comparing its variance with unit-length alternatives.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
45. Explained variance measures representation
A component's explained-variance fraction is its eigenvalue divided by the sum of eigenvalues. Retained components preserve that share of total variation under the chosen scaling. This describes data representation, not necessarily predictive usefulness; a low-variance direction can contain information about a response.
Worked example: Eigenvalues 6, 3 and 1 give total variance 10. The first two components retain 90% of the variation.
Mistake to avoid: Concluding that retaining 90% of predictor variance guarantees retaining 90% of predictive information.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
46. Uncorrelated components need not be independent
Orthogonal eigenvector directions produce uncorrelated component scores under the covariance used to construct them. Independence is stronger and follows from zero covariance only under additional assumptions, such as joint normality. Nonlinear dependence can remain even when every pairwise covariance is zero.
Worked example: Let U be standard normal and V = U²−1. Their covariance is zero, yet knowing U determines V exactly.
Mistake to avoid: Treating decorrelated scores as independent inputs without checking distributional assumptions.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
47. Component signs and loading conventions need interpretation
An eigenvector and its negative describe the same principal axis, so component sign is arbitrary. Interpretation comes from relative directions and magnitudes. The word loading may refer to eigenvector coefficients or to scaled quantities such as variable-component correlations; identify the convention before making numerical comparisons.
Worked example: Directions (0.8,0.6) and (−0.8,−0.6) produce opposite scores but the same one-dimensional reconstruction.
Mistake to avoid: Calling a sign reversal across two software outputs a substantive change in the component.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
48. Truncated PCA reconstructs a low-rank approximation
Reconstruction projects centered observations onto retained principal directions and then restores the variable means. Under ordinary Euclidean squared error, retaining the leading components gives the best approximation of the chosen rank. Dropped eigenvalues quantify aggregate reconstruction variation on the analyzed scale.
Worked example: For direction (1,1)/√2, centered observation (2,0) reconstructs as (1,1). Its residual is (1,−1), with squared error 2.
Mistake to avoid: Interpreting reconstructed standardized values as original units without reversing the scaling.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
49. PCA and common-factor analysis target different variation
PCA constructs combinations explaining total observed variance. Common-factor analysis models shared covariance through latent factors while allowing variable-specific variation. A variable with substantial unique noise can strongly influence PCA without providing strong evidence about a common factor. The methods therefore answer different questions.
Worked example: Two measurements share covariance 1, but one has an additional independent noise variance of 20. Covariance PCA favors that noisy variable's total variance.
Mistake to avoid: Treating a leading principal component as proof of an underlying causal latent trait.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
50. Communality and uniqueness partition modeled variance
With orthogonal unit-variance factors, the common-factor model has covariance ΛΛ' + Ψ, where Ψ usually contains variable-specific variances. A variable's communality is the sum of squared loadings in its row. Uniqueness is the remaining modeled variance, subject to the model's scaling assumptions.
Worked example: A standardized variable loading 0.8 on one factor has communality 0.64 and uniqueness 0.36.
Mistake to avoid: Using the loading 0.8 itself as the proportion of variance explained.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
51. Factor rotation preserves fit under compatible transformations
For orthogonal factors, multiplying loadings by an orthogonal rotation preserves ΛΛ' and therefore the fitted covariance. Rotation changes the presentation and interpretation of factors without creating a unique substantive solution. Oblique rotations allow correlated factors, requiring their covariance matrix in calculations of reproduced covariance.
Worked example: Loadings Λ and ΛQ give the same common covariance whenever QQ' = I, because ΛQQ'Λ' = ΛΛ'.
Mistake to avoid: Squaring and summing oblique loadings while ignoring correlations between factors.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
52. Factor scores are estimates of unobserved quantities
Factor analysis does not directly observe each subject's factors. Under a normal factor model, conditional-mean scores use the loadings and uniqueness to estimate latent values, with remaining conditional uncertainty. Different scoring methods can produce different estimates even when the fitted covariance model is unchanged.
Worked example: For x = f+ε, with factor variance 1 and noise variance 3, observing x = 4 gives conditional mean factor score 1 and conditional variance 0.75.
Mistake to avoid: Treating an estimated factor score as an error-free measurement of the latent factor.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
Generalized linear mixed models
53. The link applies to the conditional mean
A GLMM specifies a response distribution conditional on random effects and relates its conditional mean to a linear predictor through a link. The response distribution, link and random-effect distribution perform different roles. A linear predictor can span the real line while its inverse link enforces valid response means.
Worked example: A Bernoulli model with logistic predictor η = log(3) has conditional probability 3/(1+3) = 0.75.
Mistake to avoid: Reading the linear predictor log(3) as the response probability.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
54. Marginal means require integration over random effects
With a nonlinear inverse link, averaging random effects before transforming generally differs from averaging transformed means. For a log-link model with normal random intercept b of variance τ², the marginal mean is exp(x'β+τ²/2). Setting b to its mean zero produces the conditional mean instead.
Worked example: If exp(x'β) = 2 and τ² = 0.4, the marginal mean is 2 exp(0.2), approximately 2.443.
Mistake to avoid: Reporting 2 as the population-average mean after integrating out the random intercept.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
55. Exposure offsets impose proportional expected counts
In a log-link count model, a log-exposure offset fixes exposure's coefficient at one. The remaining predictor models a rate, while multiplying that rate by exposure gives the expected count. This proportionality is an assumption; an offset is not simply another freely estimated predictor.
Worked example: A conditional rate of 0.3 per exposure unit and exposure 4 imply expected count 1.2. Doubling exposure to 8 gives 2.4 under the same rate.
Mistake to avoid: Using raw exposure as a log-link offset instead of its logarithm.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
56. Conditional odds ratios are not marginal odds ratios
In a logistic random-intercept model, exp(β) is an odds ratio conditional on the random effect. Integrating over heterogeneous groups generally changes the population-average odds ratio, even without confounding. Specify the conditioning level when interpreting coefficients or comparing model summaries.
Worked example: Equal groups have baseline probabilities 1/4 and 3/4. Doubling each group's odds gives probabilities 2/5 and 6/7. The marginal odds ratio is 22/13, rather than 2.
Mistake to avoid: Labeling a conditional odds multiplier as the population-average odds multiplier.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
57. Random intensity creates overdispersion and dependence
For a conditionally Poisson count with random intensity Λ, marginal variance is E[Λ]+Var(Λ). If several counts share that intensity and are conditionally independent, their marginal covariance is Var(Λ). Mixing therefore changes both the mean-variance relationship and the dependence structure.
Worked example: An intensity with mean 2 and variance 1 gives marginal count variance 3. Two counts sharing it have covariance 1.
Mistake to avoid: Assuming conditionally independent counts remain independent after integrating out a shared random effect.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
58. GLMM likelihoods often need numerical integration
The marginal likelihood integrates conditional response likelihoods against the random-effect distribution. Quadrature approximates the integral using weighted evaluation points; a Laplace approximation uses local curvature around a mode. Approximation accuracy can depend on group information, distribution shape and random-effect dimension.
Worked example: For one binary observation with a symmetric zero-centered random intercept and zero fixed predictor, the exact marginal success probability is 0.5. This provides a simple integration check.
Mistake to avoid: Treating an approximate likelihood as exact without assessing sensitivity of important estimates.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
59. Validation must match the intended grouping target
Predicting another observation in an existing group differs from predicting observations in entirely new groups. Randomly splitting observations can retain group information on both sides, producing optimistic results for a new-group task. Validation should reproduce the information available for the intended prediction.
Worked example: To predict a new fleet, hold out complete fleets and integrate over their unknown random effects. Holding out one policy from each fitted fleet tests a different task.
Mistake to avoid: Using existing-group validation results to claim accuracy for previously unseen groups.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
60. Excess zeros need a structural explanation
A zero-inflated Poisson model combines a structural-zero probability π with a Poisson distribution for the remaining observations. Extra zeros alone do not establish this mechanism: random-effect heterogeneity and other count models can also increase zero frequency. Compare the proposed mechanism with the data-generating context and predictive behavior.
Worked example: With π = 0.3 and λ = 2, mean count is 1.4 and zero probability is 0.3+0.7 exp(−2), approximately 0.395.
Mistake to avoid: Calling every zero structural or interpreting λ = 2 as the overall mean.
Source: Exam MAS-II — Modern Actuarial Statistics-II | Casualty Actuarial Society
Sources
Exam identity verified:
All study guides