Choosing Count, Rate, or Binary Likelihoods

Choosing a Likelihood From the Data-Generating Process

A claim dataset may be presented as counts, rates, or indicators such as “at least one claim.” Those summaries are not interchangeable. The correct likelihood depends on what was observed, what the denominator represents, and what outcomes a trial can produce. A count records how many events occurred over exposure; a binary indicator records only whether an event occurred at least once during a defined trial.

The choice matters in credibility calculations because the likelihood determines how exposure and observations update the unknown risk parameter. A rate can be a useful way to report experience, but it does not automatically define a new probability model. The first question is always how the underlying data were generated.

Counts and Exposure-Adjusted Rates

Suppose \(N_j\) is the claim count for observation \(j\), \(e_j\) is its exposure, and \(\lambda\) is the claim rate per unit of exposure. Under a common-rate Poisson model, the conditional distribution is \(N_j\mid\lambda\sim\operatorname{Poisson}(e_j\lambda)\). Its probability mass function is

The exposure unit must be specified: it might be policy-years, vehicle-years, or another appropriate measure. The model assumes the rate is constant over the exposure being combined, the counts are conditionally independent, and the exposure is known. If risk varies materially across observations, a single common \(\lambda\) may not be appropriate without further adjustment or a different model.

For independent observations sharing \(\lambda\), write \(N=\sum_j n_j\) and \(E=\sum_j e_j\). Multiplying their likelihoods gives

Thus the count and total exposure enter through their sums. The maximum-likelihood estimate of the common rate is \(N/E\), the total count divided by total exposure. This is also the exposure-weighted average of the observation-specific rates \(n_j/e_j\).

Why Exposure Weights Rates

An observation based on more exposure represents more opportunities for claims. This is why the combined rate is not generally the simple average of individual rates. The justification follows directly from the likelihood: all count contributions contain \(\lambda^{n_j}\), while all exposure contributions contain \(\exp(-e_j\lambda)\). Multiplying these factors adds counts and exposures separately.

The log-likelihood, ignoring terms that do not depend on \(\lambda\), is \(N\log\lambda-E\lambda\). Differentiating gives \(N/\lambda-E\); setting this equal to zero yields \(\hat{\lambda}=N/E\) when \(N>0\). If \(N=0\), the likelihood is maximized at the boundary \(\hat{\lambda}=0\). This derivation explains why counts and exposures, rather than an unweighted average of rates, determine the estimate.

When the Observation Is Binary

A different data structure arises when each trial has a fixed definition and can be classified as either a success or a failure. For example, a trial might be one policy-year, with success defined as “at least one claim during the year.” If there are \(n\) independent trials, each with the same probability \(q\) of success, then the number \(X\) of successes has a binomial distribution:

This model assumes a fixed number of trials, independent outcomes, and a common success probability. A Beta prior with parameters \(a,b\) has density proportional to \(q^{a-1}(1-q)^{b-1}\). Combining it with the binomial likelihood gives a Beta posterior with parameters \(a+x\) and \(b+n-x\); the posterior mean is \((a+x)/(a+b+n)\). This update is appropriate for a probability of a binary outcome, not automatically for a claim count rate.

A binary “any claim” indicator discards information about the number of claims on a trial that had at least one. If a policy-year can have several claims, the indicator treats one claim and several claims alike. Use the binomial model when the target is the probability of at least one claim under the defined trial, or when the data truly consist of one binary outcome per trial. Use count data when the number of claims is observed and claim frequency is the target.

A Practical Selection Process

Before selecting a likelihood, describe the observation mechanism in words. The same portfolio can support different models depending on how experience was recorded and what quantity is being estimated.

Common Mistakes and Model Limits

The most frequent error is to focus on the column label—“frequency,” “rate,” or “claim incidence”—instead of how the observations were produced. A rate may be a count divided by exposure, while an incidence measure may be a count of trials with any event divided by the number of trials. These have different denominators and different interpretations.

• Averaging rates without exposure weights. If exposures differ, use total claims divided by total exposure. An ordinary average gives equal weight to observations with unequal amounts of experience.
• Using a binomial model for a claim count. Binomial outcomes allow only success or failure per trial. If multiple claims can occur and are observed, a count likelihood preserves that information.
• Using a Poisson model for a binary record without checking the target. A binary record may support an any-claim probability, but it does not reveal the number of claims on positive trials.
• Ignoring units or changing denominators. A rate per policy-year is not numerically interchangeable with a rate per vehicle-year unless the exposure definition is consistent.

The basic models also have limits. A Poisson model with a common rate implies conditional variance equal to the conditional mean for a given exposure. A binomial model with a common probability assumes independent trials with the same success probability. Strong dependence, changing risks, or unexplained variation can undermine these assumptions. Model choice is not settled by arithmetic alone: the likelihood must represent the sampling process sufficiently well for the intended estimate.