Why Rating Factors Need an Anchor Level

Why an Anchor Is Necessary

A multiplicative rating table describes a rate as a base level multiplied by category factors. In a log-linear representation, the base is the exponentiated intercept and each category factor is the exponentiated category effect. If every category is given a free effect as well as an intercept, those coefficients are not uniquely determined by the fitted rates. The model can shift a constant from the effects to the intercept without changing any prediction.

This is an identifiability issue, not a problem with the predicted rates themselves. An anchor level, usually chosen as the reference category, fixes the coefficient convention: its effect is set to zero on the log scale, or equivalently its relativity is set to one. The intercept then represents the modeled rate for that reference category.

Definition: A parameterization is identifiable when distinct parameter values cannot produce exactly the same model predictions. In a one-factor log-linear rating model with an intercept and a free effect for every category, the intercept and effects are not separately identifiable. Fixing one category effect at zero removes this ambiguity.

The Log-Linear Model and Its Ambiguity

Suppose a rating factor has \(K\) categories. Let \(m_j\) be the modeled positive mean rate for category \(j\), let \(\beta_0\) be an intercept, and let \(\alpha_j\) be the log-scale effect for category \(j\). The multiplicative model is:

\log(m_j)=\beta_0+\alpha_j, \qquad m_j=\exp(\beta_0)\exp(\alpha_j).

The implied base is \(\exp(\beta_0)\), and category \(j\)'s relativity is \(\exp(\alpha_j)\). If all \(K\) effects are free, then for any constant \(c\), define new parameters \(\beta_0^*=\beta_0+c\) and \(\alpha_j^*=\alpha_j-c\). For every category:

\beta_0^*+\alpha_j^* =(\beta_0+c)+(\alpha_j-c) =\beta_0+\alpha_j.

Thus every fitted rate remains exactly the same, even though the intercept and all category effects have changed. The data may determine the category rates, but they cannot decide which of infinitely many equivalent coefficient sets should be reported.

The same issue appears in a design matrix. If the model includes an intercept column and an indicator column for every category, the sum of the category indicators is the intercept column: each observation belongs to exactly one category. Therefore one column is a linear combination of the others, so the design matrix does not have full column rank. Fixing the reference effect at zero is one standard way to remove the redundant parameter.

Anchoring a Reference Category

Choose category 1 as the reference and impose \(\alpha_1=0\). Its relativity is \(\exp(0)=1\), so the model becomes:

m_1=\exp(\beta_0), \qquad m_j=\exp(\beta_0)\exp(\alpha_j) \quad (j=2,\ldots,K).

Now the intercept is the modeled rate for category 1, and each other category effect measures its log rate ratio against category 1. The choice \(\alpha_1=0\) is a coordinate convention; it does not claim that the reference category has no risk, no claims, or a rate of zero. Its rate is \(\exp(\beta_0)\), which is positive.

If known exposure \(E_i\) is included for observation \(i\), a Poisson model may be written as \(\log(\mathbb{E}[N_i])=\log(E_i)+\beta_0+\alpha_j\) for an observation in category \(j\). The known exposure term does not resolve or create the intercept-effect ambiguity: the same shift in \(\beta_0\) and \(\alpha_j\) leaves the expected count unchanged.

Worked Example: Converting Free Effects to a Reference-Level Model

A hypothetical insurer models annual claim counts for three vehicle classes using a Poisson model. The exposure is 100, 50, and 20 policy-years in classes A, B, and C, respectively. A fitted but unanchored representation has \(\beta_0=\ln(0.16)\), \(\alpha_A=\ln(0.5)\), \(\alpha_B=\ln(0.625)\), and \(\alpha_C=\ln(0.9375)\). Find the class rates, the expected counts, and the equivalent model anchored at A. For this example, assume independent Poisson counts with means equal to exposure times rate; interest and mortality assumptions are not applicable.

Method. First combine the intercept and effect for each class. The rates per policy-year are:

m_A=0.16(0.5)=0.08,\qquad m_B=0.16(0.625)=0.10,\qquad m_C=0.16(0.9375)=0.15.

The Poisson means are exposure multiplied by the corresponding rate: class A has \(100(0.08)=8\), class B has \(50(0.10)=5\), and class C has \(20(0.15)=3\). Thus the count distributions are Poisson with parameters 8, 5, and 3, respectively.

To anchor at A, the new intercept must be \(\beta_0^*=\ln(0.08)\), because the reference rate is 0.08. The new A effect is zero. The other effects are log rate ratios to A:

\alpha_A^*=0,\qquad \alpha_B^*=\ln(0.10/0.08)=\ln(1.25),\qquad \alpha_C^*=\ln(0.15/0.08)=\ln(1.875).

Check. The anchored rates are \(0.08(1)=0.08\), \(0.08(1.25)=0.10\), and \(0.08(1.875)=0.15\). They reproduce the original rates and therefore the same Poisson means 8, 5, and 3. The mistake to avoid is changing the anchor effect to zero while leaving the intercept unchanged; doing both would alter the predictions.

Why the Reference Constraint Works

The constraint \(\alpha_1=0\) removes the one free shift that caused the ambiguity. To see this directly, consider any unanchored coefficients and choose category 1 as reference. Set \(\beta_0^*=\beta_0+\alpha_1\) and \(\alpha_j^*=\alpha_j-\alpha_1\). Then \(\alpha_1^*=0\), and for every category:

\beta_0^*+\alpha_j^* =(\beta_0+\alpha_1)+(\alpha_j-\alpha_1) =\beta_0+\alpha_j.

So the conversion preserves every linear predictor and rate. It also produces a unique coefficient set under this chosen reference convention: the reference effect must be zero, the intercept must equal the reference log rate, and each remaining effect must equal the log rate ratio to that reference.

This argument concerns a single categorical factor. In a model with several factors, each factor needs an identifiable coding convention; fixing one reference level for each factor is a common approach. The central point remains the same: do not include a redundant set of free category effects alongside an intercept.

Worked Example: Finding the Correct Shift for a New Anchor

A hypothetical rating model gives class rates through an unanchored log-linear form with \(\beta_0=\ln(0.24)\), \(\alpha_A=\ln(0.25)\), \(\alpha_B=\ln(0.5)\), and \(\alpha_C=\ln(0.75)\). The response is annual claim count, assumed independent Poisson by class. Exposures are 50, 25, and 50 policy-years for A, B, and C. Re-express the model with B as the reference and verify its expected counts. Interest and mortality assumptions are not applicable.

Step 1: Calculate the rates. Multiplying the intercept rate by each category factor gives \(m_A=0.24(0.25)=0.06\), \(m_B=0.24(0.5)=0.12\), and \(m_C=0.24(0.75)=0.18\) claims per policy-year.

Step 2: Shift to the B reference. Add the old B effect to the intercept and subtract it from each effect. The new intercept is \(\ln(0.24)+\ln(0.5)=\ln(0.12)\). The new effects are \(\alpha_A^*=\ln(0.25)-\ln(0.5)=\ln(0.5)\), \(\alpha_B^*=0\), and \(\alpha_C^*=\ln(0.75)-\ln(0.5)=\ln(1.5)\).

Step 3: Verify the rates and means. The anchored rates are \(0.12(0.5)=0.06\), \(0.12(1)=0.12\), and \(0.12(1.5)=0.18\). The Poisson means are \(50(0.06)=3\), \(25(0.12)=3\), and \(50(0.18)=9\). The count distributions therefore have parameters 3, 3, and 9, exactly as under the original coefficients.

Check. The B effect is zero and the intercept corresponds to B's rate, 0.12. Setting the B effect to zero without increasing the intercept from 0.24 to 0.12 would not preserve the fitted rates.

Worked Example: Detecting Redundant Indicator Columns

A hypothetical insurer observes independent Poisson claim counts for three classes: A has 4 claims in 80 policy-years, B has 9 in 60, and C has 6 in 20. It proposes a log-rate model with an intercept and three free class indicators. Find the class-rate estimates, explain the redundancy, and give the equivalent model with A as reference. Assume each count has a Poisson distribution with mean exposure times its class rate. Interest and mortality assumptions are not applicable.

Step 1: Estimate each class rate. For a Poisson count \(N_j\) with mean \(E_jm_j\), the log-likelihood, apart from terms not involving \(m_j\), is \(N_j\ln(E_jm_j)-E_jm_j\). Differentiating with respect to \(m_j\) gives \(N_j/m_j-E_j=0\), so the maximum-likelihood estimate is \(N_j/E_j\). Thus the estimates are \(4/80=0.05\), \(9/60=0.15\), and \(6/20=0.30\).

Step 2: Identify the redundancy. For observations from A, B, and C, the intercept and indicator columns have rows \((1,1,0,0)\), \((1,0,1,0)\), and \((1,0,0,1)\). The sum of the three indicator columns equals the intercept column. Consequently, the four coefficients cannot be uniquely determined from these three independent columns.

Step 3: Anchor at A. Set \(\beta_0^*=\ln(0.05)\) and \(\alpha_A^*=0\). The B and C effects are \(\ln(0.15/0.05)=\ln(3)\) and \(\ln(0.30/0.05)=\ln(6)\). The anchored predictions are \(0.05\), \(0.05(3)=0.15\), and \(0.05(6)=0.30\).

Step 4: Check the count means. Multiplying by exposures gives \(80(0.05)=4\), \(60(0.15)=9\), and \(20(0.30)=6\). These match the observed counts and the fitted Poisson means. The error to avoid is treating four reported coefficients as four independently estimable effects merely because four symbols appear in the formula.

Common Exam Trap: A Zero Effect Is Not a Zero Rate

A common error is to interpret the reference effect \(\alpha_1=0\) as saying the reference class has a rate of zero. The model shows why this is false: its rate is \(\exp(\beta_0)\exp(0)=\exp(\beta_0)\). Another error is to change the reference effect to zero but keep the old intercept. That generally changes every prediction because the intercept must absorb the old reference effect.

For example, in the preceding class-rate model, the unanchored representation used \(\beta_0=\ln(0.30)\) and \(\alpha_A=\ln(1/6)\). The incorrect calculation would set \(\alpha_A=0\) but retain \(\beta_0=\ln(0.30)\), predicting an A rate of 0.30 rather than 0.05. The correction is to shift both parts: \(\beta_0^*=\ln(0.30)+\ln(1/6)=\ln(0.05)\), and \(\alpha_j^*=\alpha_j-\ln(1/6)\). The reference rate is then 0.05 and all class rates are preserved.

Key takeaway: A reference level anchors the coefficient scale, not the risk scale. Its log effect is zero and its relativity is one, while its modeled rate is the positive base rate \(\exp(\beta_0)\).

Check Your Understanding

For each problem, identify the anchor constraint and check that it preserves the fitted rates.

A two-class model has rates 0.04 and 0.10. If the first class is the reference, what is the base rate and what is the second-class relativity? (Answer: Base rate \(0.04\); second-class relativity \(2.5\).)
An unanchored model has \(\beta_0=\ln(0.20)\) and effects \(\ln(0.5)\), \(\ln(1)\), and \(\ln(1.5)\) for classes A, B, and C. Find the rates and the coefficients after anchoring at B. (Answer: Rates \(0.10, 0.20, 0.30\); anchored intercept \(\ln(0.20)\) and effects \(\ln(0.5), 0, \ln(1.5)\).)
A Poisson frequency model has 12 claims in 120 exposure units for A and 18 claims in 60 units for B. Find the rate estimates and the B relativity when A is the reference. (Answer: Rates \(0.10\) and \(0.30\); B relativity \(3\).)
A model has an intercept and indicator columns \(I_A,I_B,I_C,I_D\), where every observation is in exactly one class. State the linear dependence and the number of free class effects to retain with an intercept under reference coding. (Answer: \(I_A+I_B+I_C+I_D=1\), the intercept column; retain three free effects and set the fourth, reference effect to zero.)
An unanchored model has \(\beta_0=\ln(0.40)\) and category effects \(\alpha_A=\ln(0.25)\), \(\alpha_B=\ln(0.50)\), and \(\alpha_C=\ln(0.75)\). Re-anchor at B and verify the rates for A and C. (Answer: New intercept \(\ln(0.20)\); effects \(\ln(0.5),0,\ln(1.5)\); rates A \(0.10\), B \(0.20\), C \(0.30\).)

What You'll Learn