Why Overdispersion Matters to a Frequency Indication
A frequency model supplies estimated claim rates for rating classes. Those rates can be combined with exposures to produce an indicated portfolio frequency. Overdispersion matters because it can make the estimates less precise, but that does not automatically change the estimated rates. Keeping those effects separate is essential when explaining or applying a frequency indication.
For a policy or rating cell \(i\), let \(N_i\) be the claim count, \(e_i\) its exposure, and \(\lambda_i\) its expected claim frequency per exposure unit. Under a log-link frequency model, the conditional mean is \(\mu_i=e_i\lambda_i\), with \(\log(\lambda_i)=\boldsymbol{x}_i^\mathsf{T}\boldsymbol{\beta}\). The exposure term is treated as known. This tutorial focuses on how excess variation around these means affects frequency indications and their uncertainty.
For a Poisson model, \(E[N_i]=\operatorname{Var}(N_i)=\mu_i\). A quasi-Poisson specification instead uses \(E[N_i]=\mu_i\) and \(\operatorname{Var}(N_i)=\phi\mu_i\), where \(\phi\) is a dispersion parameter. Values \(\phi>1\) represent overdispersion relative to Poisson. A negative-binomial model provides a different variance structure; a common form is \(\operatorname{Var}(N_i)=\mu_i+\alpha\mu_i^2\), with \(\alpha>0\). The choice of model matters: a quasi-Poisson scale adjustment and a negative-binomial refit do not have identical effects on estimation.
What Changes and What Does Not
In a log-link quasi-Poisson model with a single constant dispersion parameter, the fitted coefficients and means are the same as those obtained by solving the Poisson score equations for the same data and model structure. The dispersion parameter changes the estimated coefficient covariance matrix: under the usual quasi-likelihood calculation, it is the Poisson covariance matrix multiplied by \(\phi\). Thus standard errors are multiplied by \(\sqrt{\phi}\), not by \(\phi\).
This result assumes the same observations, design matrix, exposure specification, and log-link mean structure, and a common multiplicative dispersion parameter. It does not say that every model which allows overdispersion must give the same fitted means as a Poisson model. A negative-binomial model uses a different likelihood and variance structure; estimating its parameters can change the fitted coefficients and means as well as their uncertainty.
The reason for the quasi-Poisson result can be seen directly from the estimating equation. For a log-link model, \(\partial\mu_i/\partial\boldsymbol{\beta}=\mu_i\boldsymbol{x}_i\). The quasi-score contribution is proportional to \((N_i-\mu_i)\boldsymbol{x}_i/\phi\). If \(\phi\) is common to all observations, the entire score is the Poisson score multiplied by \(1/\phi\). Multiplication by a nonzero constant does not change the coefficient values at which the score is zero. But the variance calculation does depend on \(\phi\), which is why the coefficient estimates can stay fixed while their standard errors increase.
Worked Example: Adjusting a Relativity's Uncertainty
A hypothetical insurer fits a log-link quasi-Poisson frequency model. The comparison-class relativity is 1.40. Its Poisson model-based standard error on the log-relativity scale is 0.12. A common dispersion estimate is \(\phi=2.25\). Find the adjusted standard error and an approximate 95% confidence interval for the relativity. Assume the normal approximation is appropriate.
Method. A common quasi-Poisson dispersion factor leaves the fitted coefficients, and therefore the point estimate 1.40, unchanged. It multiplies the coefficient covariance matrix by \(\phi\), so the standard error is multiplied by \(\sqrt{\phi}\).
First calculate the standard-error multiplier:
The adjusted standard error of the log relativity is therefore \(0.12(1.5)=0.18\). The point estimate on the log scale is \(\log(1.40)=0.33647\). The approximate 95% interval on that scale is:
Exponentiating both endpoints gives the interval for the relativity:
The adjusted point estimate remains 1.40, while the interval is wider than it would be using the unadjusted standard error. This is an approximate interval based on a normal approximation to the log-relativity estimator; it is not a claim that the relativity itself is normally distributed. The common mistake is to multiply the point estimate by 2.25 or 1.5. Dispersion scales uncertainty here, not the fitted relativity.
From Class Frequencies to a Portfolio Indication
For a portfolio with fitted class frequencies \(\hat{\lambda}_j\) and exposures \(e_j\), the total fitted claim count is the sum of exposure times fitted frequency. Dividing that count by total exposure gives the portfolio frequency indication. This is an exposure-weighted average of the class rates, not an unweighted average unless the exposures happen to be equal.
These formulas describe a point indication. If the fitted class rates do not change, a larger dispersion estimate does not change either calculation. It does affect how certain we are about the fitted rates and hence about the indication. If an alternative model refit changes the class-rate estimates, the portfolio point indication can change too.
Worked Example: Aggregating Two Alternative Frequency Fits
Two hypothetical frequency fits have the same baseline frequency, 0.06 claims per exposure unit. Fit A gives the comparison class a relativity of 1.50; Fit B gives it a relativity of 1.25. Each class has 500 exposure units. Calculate each fit's class rates, predicted total count, and portfolio frequency. Treat these as two alternative fitted models, not as a dispersion adjustment to one fixed fit.
Method. Multiply the baseline frequency by the comparison-class relativity, then multiply each class frequency by that class's exposure. Add the predicted counts and divide by total exposure.
For Fit A, the class frequencies are \(0.06\) and \(0.06(1.50)=0.09\). The predicted counts are \(500(0.06)=30\) and \(500(0.09)=45\). Thus the total is \(30+45=75\) claims, and the portfolio frequency is \(75/(500+500)=0.075\).
For Fit B, the class frequencies are \(0.06\) and \(0.06(1.25)=0.075\). The predicted counts are \(500(0.06)=30\) and \(500(0.075)=37.5\). Thus the total is \(30+37.5=67.5\) claims, and the portfolio frequency is \(67.5/1{,}000=0.0675\).
As a check, Fit A's portfolio frequency is the equal-exposure average \((0.06+0.09)/2=0.075\); Fit B's is \((0.06+0.075)/2=0.0675\). Fractional predicted counts are valid as expected values. Fit B's indication is \(0.0675/0.075=0.90\), or 10% below Fit A's. The fitted means differ because the two fits have different relativities; this is not the effect of merely multiplying a fixed quasi-Poisson fit's variance by a dispersion factor.
Common mistake and correction. One might incorrectly report 105 claims for Fit A. But the comparison class has a fitted rate of \(0.06(1.50)=0.09\), so its expected count is \(500(0.09)=45\), not 75. Adding the baseline-class count gives \(30+45=75\), and the portfolio rate is \(75/1{,}000=0.075\).
Uncertainty in a Portfolio Rate Indication
An indicated portfolio frequency is a function of fitted model parameters. Its uncertainty therefore depends on the covariance among those parameters and on how the portfolio indication responds to each one. In general, if an indication is \(g(\hat{\boldsymbol{\beta}})\), the delta method approximates its variance by the gradient of \(g\) multiplied on both sides of the coefficient covariance matrix. Use the covariance matrix adjusted for overdispersion when that adjustment is appropriate.
The derivative terms are important: parameter uncertainty in a rate that applies to most of the portfolio may matter more than equal-sized uncertainty in a rate applying to a small class. Also, coefficient estimates are often correlated. Ignoring covariance is only justified when it is supplied as zero or when a stated approximation deliberately omits it.
Worked Example: Overdispersion's Effect on an Indicated Portfolio Rate
A hypothetical portfolio has 800 exposure units in class A and 200 in class B. Its fitted baseline frequency is \(\hat{\lambda}_0=0.04\), and class B has relativity \(r=1.50\). Under the Poisson covariance calculation, the standard errors of \(\hat{\eta}_0=\log(\hat{\lambda}_0)\) and \(\hat{\eta}_r=\log(r)\) are 0.08 and 0.16, respectively, with covariance zero. A quasi-Poisson dispersion estimate is \(\phi=1.5625\). Calculate the portfolio indication and its approximate standard error before and after dispersion adjustment. Use the delta method, and assume the stated zero covariance.
Method. First calculate the fitted class rates and the exposure-weighted portfolio frequency. Then use the derivatives of that frequency with respect to the log baseline and log relativity. The dispersion adjustment multiplies the coefficient standard errors by \(\sqrt{\phi}\).
The class A rate is \(0.04\); the class B rate is \(0.04(1.50)=0.06\). The predicted counts are \(800(0.04)=32\) and \(200(0.06)=12\). Thus there are \(32+12=44\) expected claims over 1,000 exposure units, giving:
Write the indication as \(g=\lambda_0(0.8+0.2r)\). Its derivatives with respect to the log parameters are \(\partial g/\partial\eta_0=g=0.044\) and \(\partial g/\partial\eta_r=0.2\lambda_0r=0.012\). With zero covariance, the Poisson-based delta-method variance is:
The corresponding standard error is \(\sqrt{0.0000160768}\approx0.004010\). Since \(\sqrt{1.5625}=1.25\), the adjusted standard errors for both log parameters are \(0.08(1.25)=0.10\) and \(0.16(1.25)=0.20\). The adjusted variance is:
The adjusted standard error is \(\sqrt{0.00002512}\approx0.005012\). The indication remains 0.044; its estimated standard error increases from about 0.004010 to 0.005012. This calculation depends on the supplied zero covariance assumption. With nonzero covariance, the delta-method variance also includes twice the product of the two derivatives and their covariance.
Exam Traps and Practical Interpretation
A frequency indication is not a confidence bound, and a larger standard error does not mechanically raise or lower the indicated frequency. It signals less precision. Decisions about how uncertainty should affect a selected rate require additional considerations beyond the fitted frequency itself; do not silently turn an uncertainty adjustment into a change in the point estimate.
- Do not scale the mean by the dispersion factor. In a quasi-Poisson fit, \(\phi\) adjusts the variance calculation, not \(\hat{\lambda}_i\).
- Do not confuse standard errors with variances. If covariance is multiplied by \(\phi\), standard errors are multiplied by \(\sqrt{\phi}\).
- Do not assume an alternative model has identical fitted rates. A negative-binomial refit can change the estimated coefficients and class means, so recompute the portfolio indication from its fitted rates.
- Do not average class rates without regard to exposure. Use \(\sum_j e_j\hat{\lambda}_j/\sum_j e_j\); an unweighted mean is appropriate only for equal exposures.
- Do not omit parameter covariance automatically. When propagating uncertainty to a portfolio indication, use the full covariance matrix unless a zero-covariance assumption is stated.
Check Your Understanding
For each problem, distinguish the fitted point indication from its uncertainty where relevant.
- A quasi-Poisson model has \(\phi=4\), a fitted frequency of 0.035, and a Poisson standard error of 0.07 for a log-relativity. What are the fitted frequency and adjusted standard error? (Answer: Frequency 0.035; adjusted standard error 0.14.)
- A hypothetical portfolio has 300 exposure units at frequency 0.05 and 700 at frequency 0.06. Find the predicted total count and portfolio frequency. (Answer: 57 expected claims; frequency 0.057.)
- A baseline frequency is 0.08 and a comparison relativity is 1.25. Each class has 400 exposure units. Calculate the portfolio frequency. (Answer: \(0.09\).)
- A quasi-Poisson fit has relativity 1.30, dispersion \(\phi=2.25\), and Poisson standard error 0.08 on the log-relativity scale. Give the adjusted point estimate and standard error. (Answer: Relativity 1.30; standard error 0.12.)
- A portfolio has 200 exposure units at baseline frequency 0.03 and 800 at a frequency equal to \(0.03(1.40)\). Find its predicted total count and portfolio frequency. Would multiplying both class rates by \(\phi=1.8\) be the quasi-Poisson adjustment to the point indication? (Answer: 39.6 expected claims; frequency 0.0396; no, quasi-Poisson dispersion adjusts uncertainty, not those fitted rates.)