Study guide
4.2 Poisson Regression & Exposure Offsets
CAS Property & Casualty Predictive Analytics PCPA
Track your progress as you read
Finished sections get a check mark.
1 Chapter 1: PCPA Exam & Project Orientation
1.1 PCPA Exam Format, Scoring & Registration 1.2 The PCPA Project: Windows, Deliverables & Grading
2 Chapter 2: Dealing with Data — Sourcing, Importing & Assessing
2.1 Requirements Gathering & Assessing Data Relevance 2.2 Importing, Joining & Reshaping Data Sets 2.3 Exploratory Data Analysis & Evaluating a Data Set 2.4 Data Quality Under ASOP 23 & Modeling Under ASOP 56
3 Chapter 3: Dealing with Data — Conditioning P&C Data for a GLM
3.1 Exposure Bases, Weights & Data Grain 3.2 Missing Data Mechanisms & Imputation Strategies 3.3 Outlier Diagnostics, Trimming & Winsorization 3.4 Variable Transformation & Scaling
4 Chapter 4: Model Diagnostics & Selection — Building a GLM
4.1 Exponential Dispersion Family & Link Functions 4.2 Poisson Regression & Exposure Offsets 4.3 Overdispersion, Quasi-Poisson & Negative Binomial 4.4 Gamma & Inverse Gaussian Models for Claim Severity
5 Chapter 5: Model Diagnostics & Selection — Model Form & Predictors
5.1 Tweedie Compound Poisson-Gamma Modeling 5.2 Binomial & Logistic GLMs for Binary Outcomes 5.3 Categorical Predictors, Base Levels & High-Cardinality Grouping 5.4 Interaction Terms & Non-Linear Effects 5.5 Interpreting Coefficients & Multiplicative Relativities
6 Chapter 6: Model Diagnostics & Selection — Evaluating and Improving the Fit
6.1 Deviance, Likelihood Ratio Tests & Residual Diagnostics 6.2 Stepwise Selection & Information Criteria Penalties 6.3 Multicollinearity, Correlated Predictors & Aliasing 6.4 Under-fitting, Over-fitting & Cross-Validation
7 Chapter 7: Model Diagnostics & Selection — Validation & Model Choice
7.1 Actual vs. Predicted, Calibration & Loss Ratio Charts 7.2 Simple Quantile Plots & Measuring Lift 7.3 Double Lift Charts & Competing Model Evaluation 7.4 Gini Index, Lorenz Curves & ROC Curves
8 Chapter 8: Model Interpretation & Presentation — Tables, Charts & Exhibits
8.1 Designing Tables & Statistical Exhibits 8.2 Chart Selection & Graphical Perception 8.3 How Charts Mislead 8.4 One-Way, Two-Way & Partial Dependence Exhibits
9 Chapter 9: Model Interpretation & Presentation — Communicating the Analysis
9.1 Communicating Methodology & Modeling Decisions 9.2 Explaining Why a Variable Was or Was Not Included 9.3 Communicating to Non-Technical Audiences 9.4 The PCPA Technical Report: Structure & Rubric
4.2 Poisson Regression & Exposure Offsets
Key Takeaways Explain This Topic AI
In casualty frequency modeling, the expected claim count is E[Y_i] = e_i * mu_i, where e_i is earned exposure and mu_i is the annualized claim frequency per unit exposure.
Under a log link, incorporating exposure yields ln(E[Y_i]) = ln(e_i) + x_i^T * beta, establishing ln(e_i) as an offset whose regression coefficient is constrained a priori to exactly 1.0.
Modeling raw claim counts Y_i with an offset of ln(e_i) is mathematically equivalent to modeling empirical claim frequency Y_i / e_i with prior weights equal to e_i.
The Poisson regression model assumes independent claim events, constant occurrence intensity over the exposure period, and strict equidispersion (Var(Y) = E[Y]).
A critical actuarial modeling error is entering raw exposure e_i instead of ln(e_i) as an offset, or including exposure as an unconstrained regression variable, which violates exposure additivity.
Last updated: September 2026
The Poisson Distribution for Insurance Claim Counts
In property and casualty insurance, claim frequency represents the rate at which insured losses occur over a defined period of coverage. The classical starting point for modeling claim counts is the Poisson distribution .
A discrete random variable Y Y Y representing the number of claims reported on a policy follows a Poisson distribution with parameter λ > 0 \lambda > 0 λ > 0 if its probability mass function is given by: P ( Y = y ) = e − λ λ y y ! , y ∈ { 0 , 1 , 2 , … } P(Y = y) = \frac{e^{-\lambda} \lambda^y}{y!}, \quad y \in \{0, 1, 2, \dots\} P ( Y = y ) = y ! e − λ λ y , y ∈ { 0 , 1 , 2 , … }
The defining characteristic of the Poisson distribution is equidispersion , meaning its expected value and variance are mathematically identical: E [ Y ] = λ \mathbb{E}[Y] = \lambda E [ Y ] = λ Var ( Y ) = λ \text{Var}(Y) = \lambda Var ( Y ) = λ
In an insurance portfolio, policyholders do not all share the same exposure to risk. A commercial fleet policy covering 100 power units over an entire policy year has a vastly higher expected claim count than a personal auto policy covering a single vehicle for one month. Therefore, Poisson modeling must explicitly incorporate the volume of earned exposure .
Incorporating Exposure: The Offset Derivation
Let:
Y i Y_i Y i denote the observed claim count for policy i i i ( y i ∈ { 0 , 1 , 2 , … } y_i \in \{0, 1, 2, \dots\} y i ∈ { 0 , 1 , 2 , … } ).
e i e_i e i denote the earned exposure for policy i i i (measured in earned car-years, policyholder years, house-years, or 100 units of payroll).
μ i \mu_i μ i denote the underlying expected claim frequency per unit of exposure (the annualized hazard rate).
The expected number of claims λ i = E [ Y i ] \lambda_i = \mathbb{E}[Y_i] λ i = E [ Y i ] is directly proportional to earned exposure: E [ Y i ] = λ i = e i ⋅ μ i \mathbb{E}[Y_i] = \lambda_i = e_i \cdot \mu_i E [ Y i ] = λ i = e i ⋅ μ i
To relate frequency μ i \mu_i μ i to rating variables x i = [ 1 , x i 1 , … , x i k ] T \mathbf{x}_i = [1, x_{i1}, \dots, x_{ik}]^T x i = [ 1 , x i 1 , … , x ik ] T , we specify a Generalized Linear Model with the standard log link function: ln ( μ i ) = x i T β = β 0 + β 1 x i 1 + ⋯ + β k x i k \ln(\mu_i) = \mathbf{x}_i^T \boldsymbol{\beta} = \beta_0 + \beta_1 x_{i1} + \dots + \beta_k x_{ik} ln ( μ i ) = x i T β = β 0 + β 1 x i 1 + ⋯ + β k x ik
Substituting this into the expected claim count expression and taking the natural logarithm: ln ( E [ Y i ] ) = ln ( e i ⋅ μ i ) = ln ( e i ) + ln ( μ i ) \ln(\mathbb{E}[Y_i]) = \ln(e_i \cdot \mu_i) = \ln(e_i) + \ln(\mu_i) ln ( E [ Y i ]) = ln ( e i ⋅ μ i ) = ln ( e i ) + ln ( μ i ) ln ( E [ Y i ] ) = ln ( e i ) + x i T β \ln(\mathbb{E}[Y_i]) = \ln(e_i) + \mathbf{x}_i^T \boldsymbol{\beta} ln ( E [ Y i ]) = ln ( e i ) + x i T β E [ Y i ] = exp ( ln ( e i ) + x i T β ) = e i ⋅ exp ( x i T β ) \mathbb{E}[Y_i] = \exp\left( \ln(e_i) + \mathbf{x}_i^T \boldsymbol{\beta} \right) = e_i \cdot \exp(\mathbf{x}_i^T \boldsymbol{\beta}) E [ Y i ] = exp ( ln ( e i ) + x i T β ) = e i ⋅ exp ( x i T β )
Definition of an Offset
In statistical regression, an offset is defined as an explanatory variable whose regression coefficient is not estimated from data, but is constrained a priori to equal exactly 1.0 .
Here, ln ( e i ) \ln(e_i) ln ( e i ) acts as an offset. It enters the linear predictor with a fixed parameter of 1.0: η i = 1.0 ⋅ ln ( e i ) + β 0 + β 1 x i 1 + ⋯ + β k x i k \eta_i = 1.0 \cdot \ln(e_i) + \beta_0 + \beta_1 x_{i1} + \dots + \beta_k x_{ik} η i = 1.0 ⋅ ln ( e i ) + β 0 + β 1 x i 1 + ⋯ + β k x ik
Why the Coefficient Must Equal 1.0: Exposure Additivity
Constraining the coefficient of ln ( e i ) \ln(e_i) ln ( e i ) to 1.0 enforces strict exposure proportionality (homogeneity of degree 1) :
If exposure doubles from 1.0 1.0 1.0 car-year to 2.0 2.0 2.0 car-years, the expected claim count must double exactly: E [ Y ] = 2.0 ⋅ μ \mathbb{E}[Y] = 2.0 \cdot \mu E [ Y ] = 2.0 ⋅ μ .
If a policyholder purchases insurance for two consecutive 6-month terms ( e 1 = 0.5 , e 2 = 0.5 e_1 = 0.5, e_2 = 0.5 e 1 = 0.5 , e 2 = 0.5 ), the total expected claim count across both terms equals the expected claim count of a single 12-month policy ( e = 1.0 e = 1.0 e = 1.0 ): E [ Y 1 ] + E [ Y 2 ] = 0.5 μ + 0.5 μ = 1.0 μ = E [ Y full year ] \mathbb{E}[Y_1] + \mathbb{E}[Y_2] = 0.5\mu + 0.5\mu = 1.0\mu = \mathbb{E}[Y_{\text{full year}}] E [ Y 1 ] + E [ Y 2 ] = 0.5 μ + 0.5 μ = 1.0 μ = E [ Y full year ]
If an actuary mistakenly includes ln ( e i ) \ln(e_i) ln ( e i ) as an unconstrained covariate and the regression estimates a coefficient β ^ e = 0.85 \hat{\beta}_e = 0.85 β ^ e = 0.85 , the expected claim count scales as e 0.85 e^{0.85} e 0.85 . A 6-month policy would have an expected count factor of 0.5 0.85 ≈ 0.5548 0.5^{0.85} \approx 0.5548 0. 5 0.85 ≈ 0.5548 . Two consecutive 6-month policies would produce an expected factor of 0.5548 + 0.5548 = 1.1096 0.5548 + 0.5548 = 1.1096 0.5548 + 0.5548 = 1.1096 —charging the policyholder roughly 11 % 11\% 11% more than an identical 12-month policy. This violates fundamental actuarial principles of cost-causation and exposure additivity.
Frequency Modeling Alternatives: Count Offset vs. Weighted Frequency
In actuarial practice, there are two distinct ways to structure a Poisson GLM in statistical software. Both approaches are mathematically equivalent and yield identical parameter estimates, standard errors, and deviance statistics.
Approach 1: Raw Count with Offset
Target variable : Raw claim counts Y i ∈ { 0 , 1 , 2 , … } Y_i \in \{0, 1, 2, \dots\} Y i ∈ { 0 , 1 , 2 , … }
- Distribution family : Poisson with log link
- Exposure specification : offset = ln ( e i ) \text{offset} = \ln(e_i) offset = ln ( e i )
The log-likelihood function is: ℓ ( β ) = ∑ i = 1 n [ y i ln ( e i μ i ) − e i μ i − ln ( y i ! ) ] \ell(\boldsymbol{\beta}) = \sum_{i=1}^n \left[ y_i \ln(e_i \mu_i) - e_i \mu_i - \ln(y_i!) \right] ℓ ( β ) = i = 1 ∑ n [ y i ln ( e i μ i ) − e i μ i − ln ( y i !) ] ℓ ( β ) = ∑ i = 1 n [ y i ln ( e i ) + y i x i T β − e i exp ( x i T β ) − ln ( y i ! ) ] \ell(\boldsymbol{\beta}) = \sum_{i=1}^n \left[ y_i \ln(e_i) + y_i \mathbf{x}_i^T \boldsymbol{\beta} - e_i \exp(\mathbf{x}_i^T \boldsymbol{\beta}) - \ln(y_i!) \right] ℓ ( β ) = i = 1 ∑ n [ y i ln ( e i ) + y i x i T β − e i exp ( x i T β ) − ln ( y i !) ]
Differentiating with respect to the regression parameter β j \beta_j β j gives the score equation: ∂ ℓ ∂ β j = ∑ i = 1 n [ y i x i j − e i x i j exp ( x i T β ) ] = ∑ i = 1 n ( y i − e i μ i ) x i j = 0 \frac{\partial \ell}{\partial \beta_j} = \sum_{i=1}^n \left[ y_i x_{ij} - e_i x_{ij} \exp(\mathbf{x}_i^T \boldsymbol{\beta}) \right] = \sum_{i=1}^n \left( y_i - e_i \mu_i \right) x_{ij} = 0 ∂ β j ∂ ℓ = i = 1 ∑ n [ y i x ij − e i x ij exp ( x i T β ) ] = i = 1 ∑ n ( y i − e i μ i ) x ij = 0
- Approach 2: Empirical Frequency with Prior Weights
- Target variable : Empirical claim frequency F i = Y i e i F_i = \frac{Y_i}{e_i} F i = e i Y i
- Distribution family : Poisson with log link
- Prior weight : ω i = e i \omega_i = e_i ω i = e i (no offset)
In the Exponential Dispersion Family, the dispersion function with prior weights is a ( ϕ ) = ϕ ω i = 1 e i a(\phi) = \frac{\phi}{\omega_i} = \frac{1}{e_i} a ( ϕ ) = ω i ϕ = e i 1 (since ϕ = 1 \phi = 1 ϕ = 1 for Poisson). The weighted quasi-log-likelihood for frequency f i f_i f i is: ℓ i ( β ) = f i θ i − b ( θ i ) a ( ϕ ) = e i [ f i ln ( μ i ) − μ i ] = e i [ ( y i e i ) x i T β − exp ( x i T β ) ] \ell_i(\boldsymbol{\beta}) = \frac{f_i \theta_i - b(\theta_i)}{a(\phi)} = e_i \left[ f_i \ln(\mu_i) - \mu_i \right] = e_i \left[ \left(\frac{y_i}{e_i}\right) \mathbf{x}_i^T \boldsymbol{\beta} - \exp(\mathbf{x}_i^T \boldsymbol{\beta}) \right] ℓ i ( β ) = a ( ϕ ) f i θ i − b ( θ i ) = e i [ f i ln ( μ i ) − μ i ] = e i [ ( e i y i ) x i T β − exp ( x i T β ) ] ℓ ( β ) = ∑ i = 1 n [ y i x i T β − e i exp ( x i T β ) ] \ell(\boldsymbol{\beta}) = \sum_{i=1}^n \left[ y_i \mathbf{x}_i^T \boldsymbol{\beta} - e_i \exp(\mathbf{x}_i^T \boldsymbol{\beta}) \right] ℓ ( β ) = i = 1 ∑ n [ y i x i T β − e i exp ( x i T β ) ]
Differentiating with respect to β j \beta_j β j gives the exact same score equation: ∂ ℓ ∂ β j = ∑ i = 1 n ( y i − e i μ i ) x i j = 0 \frac{\partial \ell}{\partial \beta_j} = \sum_{i=1}^n \left( y_i - e_i \mu_i \right) x_{ij} = 0 ∂ β j ∂ ℓ = i = 1 ∑ n ( y i − e i μ i ) x ij = 0
Because the score equations and the negative Hessian matrices (Fisher Information) are identical up to constants independent of β \boldsymbol{\beta} β , both methods produce identical coefficients β ^ \hat{\boldsymbol{\beta}} β ^ and covariance matrices Var ^ ( β ^ ) \widehat{\text{Var}}(\hat{\boldsymbol{\beta}}) Var ( β ^ ) .
Software Implementation Syntax Comparison
Software Approach 1: Count with Offset Approach 2: Frequency with Weights
R glm(claims ~ x1 + x2, family = poisson(link = "log"), offset = log(exposure), data = df) glm(claims/exposure ~ x1 + x2, family = poisson(link = "log"), weights = exposure, data = df)
Python (statsmodels) sm.GLM(df['claims'], X, family=sm.families.Poisson(), offset=np.log(df['exposure'])).fit() sm.GLM(df['claims']/df['exposure'], X, family=sm.families.Poisson(), freq_weights=df['exposure']).fit()
SAS PROC GENMOD; MODEL claims = x1 x2 / DIST=POISSON LINK=LOG OFFSET=log_exposure; PROC GENMOD; MODEL frequency = x1 x2 / DIST=POISSON LINK=LOG; WEIGHT exposure; POISSON MODELING PATHS
- │
- ┌──────────────────────┴──────────────────────┐
- ▼ ▼
- Approach 1: Count + Offset Approach 2: Weighted Frequency
- Response: Y (Claims) Response: F = Y / Exposure
- Offset: ln(Exposure) Weights: Exposure
- Link: Log Link: Log
- │ │
- └──────────────────────┬──────────────────────┘
- ▼
- Score Equation: Σ (y_i - e_i * mu_i) * x_ij = 0
- Identical Coefficients: beta_hat
- Identical Relativities: exp(beta_hat)
- Underlying Assumptions of the Poisson Regression Model
When applying a Poisson GLM to insurance frequency data, the actuary implicitly relies on four key statistical assumptions:
Independence of Events : The occurrence of a claim on policy i i i is independent of claims on all other policies, and independent of prior claims on the same policy. This assumption is violated during catastrophic weather events (such as hail storms or freezes that damage thousands of homes simultaneously) or multi-claimant vehicle accidents.
Homogeneous Intensity Over Time : Within any policy's exposure period, the instantaneous rate of claim occurrence μ i \mu_i μ i remains constant. In reality, claim intensity fluctuates with seasonal driving patterns, weather changes, and economic cycles.
Equidispersion : The conditional variance of claim counts equals the conditional mean ( Var ( Y i ∣ x i ) = E [ Y i ∣ x i ] \text{Var}(Y_i \mid \mathbf{x}_i) = \mathbb{E}[Y_i \mid \mathbf{x}_i] Var ( Y i ∣ x i ) = E [ Y i ∣ x i ] ). In empirical insurance data, this assumption almost never holds; variance typically exceeds the mean, a condition known as overdispersion .
Infinitesimal Probability : Over an infinitesimally small time interval Δ t \Delta t Δ t , the probability of exactly one claim is approximately μ i Δ t \mu_i \Delta t μ i Δ t , and the probability of two or more claims is negligible ( o ( Δ t ) o(\Delta t) o ( Δ t ) ).
Actuarial Exam Traps & Implementation Pitfalls
Warning
Trap 1: Passing Raw Exposure as the Offset A common programming error is writing offset = exposure instead of offset = log(exposure) . If a policy has an exposure of 0.5 0.5 0.5 car-years, passing 0.5 0.5 0.5 directly into a log-link model multiplies the expected count by exp ( 0.5 ) ≈ 1.6487 \exp(0.5) \approx 1.6487 exp ( 0.5 ) ≈ 1.6487 instead of scaling it by 0.5000 0.5000 0.5000 . The offset must always enter on the scale of the link function: ln ( e i ) \ln(e_i) ln ( e i ) .
Warning
Trap 2: Zero Exposure Policies Causing Numerical Crashes In policy databases, cancelled or null-term policies frequently have earned exposure e i = 0 e_i = 0 e i = 0 . Because ln ( 0 ) = − ∞ \ln(0) = -\infty ln ( 0 ) = − ∞ , fitting a GLM with zero-exposure records causes numerical overflow and crashes the estimation algorithm. All records with e i ≤ 0 e_i \le 0 e i ≤ 0 must be identified and removed during data preparation.
Warning
Trap 3: Modeling Frequency Without Weights Fitting a GLM to empirical frequency Y i / e i Y_i / e_i Y i / e i without specifying weights = exposure treats a 1-day policy ( e = 0.0027 e = 0.0027 e = 0.0027 ) with 1 claim (empirical frequency = 365.0 365.0 365.0 ) as equally credible to a 100-vehicle fleet policy ( e = 100.0 e = 100.0 e = 100.0 ) with 1 claim (empirical frequency = 0.01 0.01 0.01 ). Prior weights equal to exposure are essential to reflect the precision of each observation.