Multiplicative minimum bias and a Poisson GLM are established routes to the same fitted answer under typical conditions. Give them the same data, exposures, rating variables and unpenalized log-link structure, and their equations balance the same observed and fitted totals.
That result raises a more interesting question than the equivalence itself: does using those equations mean the Poisson—or quasi-Poisson—variance describes the data?
Not necessarily. A model can use the Poisson estimating equations without claim counts literally having a Poisson variance. If the conditional mean is correctly specified, the variance assumption can give us an inefficient route to the right mean without changing where the route ends.
But that answer comes with an uncomfortable qualification. Our mean structure is always a simplification of the real world. Once the mean is imperfect, different variance assumptions can put different weight on its errors and lead to different fitted answers.
So, when the choice is presented as minimum bias versus the wider GLM family, is minimum bias still worthwhile?
The equivalence is the starting point
Take a multiplicative model for claim counts:
Expected claims = exposure × base frequency × classification factors.
The balance version of minimum bias adjusts those factors until observed and fitted claim totals agree within every level of each classification variable. An unpenalized Poisson GLM with a log link and the same mean structure produces the same balance equations.
| Minimum bias language | Poisson GLM language |
|---|---|
| Balance observed and fitted claims within each rating class | Solve the Poisson coefficient equations |
| Same data + same exposure treatment + same multiplicative terms + no credibility or penalty | |
| Same fitted means and equivalent normalized relativities | |
Brown established this connection in Minimum Bias with Generalized Linear Models in 1988. The algorithms may travel differently—minimum bias cycling through sets of factors, a GLM typically updating them together—but, with a finite and identifiable solution, they arrive at the same fit.
This does not cover every procedure called minimum bias. Credibility adjustments, constraints, penalties, different observation weights and other minimum-bias criteria can all change the problem being solved. The equivalence applies when the specifications actually match.
Does balance imply the Poisson variance is true?
No. It tells us which weighting sits behind the balance rule, not whether that weighting accurately describes the data.
In a log-link quasi-likelihood model, the variance function controls how residuals enter the estimating equations. With the Poisson working variance, V(μ) = μ, the weighting simplifies into ordinary marginal balance. Mildenhall develops that correspondence.
These are two different claims:
- A variance structure produces the balance equations.
- That variance structure describes how the data actually vary.
The first does not establish the second. An actuary can choose balance as a fitting objective without believing that variance truly equals the mean.
Quasi-Poisson does not change that conclusion as much as its name might suggest. It assumes variance proportional to the mean, φμ. The common dispersion factor changes the estimated uncertainty but cancels out of the coefficient equations. With the same mean specification, ordinary Poisson and quasi-Poisson therefore produce the same fitted means. Clark gives a useful treatment of this distinction.
An inefficient route to the right mean
If the conditional mean is correctly specified, the Poisson equations can still target the right coefficients when the variance is wrong. The price is generally efficiency: we may take a noisier route to the same destination. Conventional Poisson standard errors can also be wrong, so inference needs a variance estimate suited to the data and their dependence.
This is the idea behind Poisson pseudo-maximum likelihood. The distribution does not need to be literally Poisson for its estimating equations to recover a correctly specified conditional mean, subject to the usual sampling and regularity conditions. Santos Silva and Tenreyro discuss the result directly.
| What is wrong? | What happens? | What matters? |
|---|---|---|
| Variance, while the conditional mean is correctly specified | The fitted mean can still be right, but reached inefficiently | Standard errors and stability |
| The conditional mean itself | Different weighting can select a different imperfect compromise | The fitted factors as well as their uncertainty |
Rejecting a literal Poisson distribution is therefore not enough to reject the fitted relativities. It is enough to question what we claim about their precision.
The mean is the uncomfortable part
The clean theoretical result depends on a correctly specified conditional mean. In practice, I do not think that should be treated as a comfortable resting place.
A rating model is a compressed description of a much more complicated process. We choose a manageable set of variables, group continuous differences into bands, omit unavailable information and usually leave many interactions out. Those simplifications create a real risk that the conditional mean is misspecified, particularly when grouping or omitted interactions leave systematic residual patterns.
Once the mean is imperfect, the variance function is doing more than determining efficiency. It decides how strongly different residuals pull on the fitted model. Different working variances can then lead to different compromises.
Balance has the same limitation. A model can reproduce every territory total and every class total while missing what happens inside their combinations:
| Territory | Class 1 observed / fitted | Class 2 observed / fitted | Total observed / fitted |
|---|---|---|---|
| Territory A | 70 / 60 | 30 / 40 | 100 / 100 |
| Territory B | 20 / 30 | 30 / 20 | 50 / 50 |
| Class totals | 90 / 90 | 60 / 60 | 150 / 150 |
Every margin balances, yet every cell is wrong by 10. That could be noise, or it could be evidence of a missing territory-by-class interaction. Balance cannot tell us which.
This is where the choice of fitting rule can matter. If no simple mean structure can fit all the cells, the rule determines which errors the model is most willing to live with.
That risk has to be tested through residual structure, stability and out-of-sample performance. Balance alone cannot resolve it.
So is minimum bias still worthwhile?
Yes—but for a narrower reason than suggesting it is assumption-free or fundamentally separate from a GLM.
Minimum bias is worthwhile when balance is itself a useful objective. Its updates are intuitive, the resulting factors are easy to reconcile to observed totals, and the method provides a transparent benchmark for a multiplicative rating plan. That can be valuable for implementation, communication and model governance.
Balance is a sensible primary target when exact reconciliation to credible class totals matters, the model is intended to be a transparent benchmark, and holdout results do not suggest that the balanced margins conceal an important pattern. It should yield to—or be modified by—credibility, interactions or different weighting when classes are sparse, factors are unstable, or out-of-sample performance shows that marginal balance is preserving the wrong compromise.
The wider GLM framework is more useful when I want to inspect or change the choices around that objective: alternative working variances, interactions, nonlinear effects, diagnostics, credibility, penalties and explicit measures of uncertainty. Those extensions can also be built around minimum bias—Gross and Evans, for example, combine minimum bias with credibility—but then we are no longer comparing the original equivalent solves.
I would therefore not start by asking which label is better. I would ask whether marginal balance should be the fitting target.
- If yes, minimum bias is a clean and explainable way to express that target. An equivalent Poisson GLM reaches the same fitted means and may offer a more convenient toolbox.
- If no, the important decision is what should replace or modify balance: different weighting, a richer mean structure, credibility, a penalty or some combination of them.
Minimum bias still earns its place because it makes one particular modeling compromise legible. That is useful. It is not magic, and it does not remove the need to question the mean or justify the uncertainty.
The real choice is not minimum bias versus GLM. It is whether balance is the compromise we want.