Choosing Count and Fractional Response Models

Start With the Outcome and the Question

A model should reflect both the values an outcome can take and the quantity the analysis is intended to describe. A count of clinic visits and a share of a budget spent on one program may both be nonnegative, but they have different supports: the count can grow without a fixed upper limit, while the share must remain between zero and one. Treating these outcomes as interchangeable can produce predictions that do not make sense or answer a different question from the one intended.

The first decision is therefore not which named model is most familiar. It is whether the response is a count or a fraction, how it was observed, and what its conditional mean should represent. Zeros and exposure time can further affect the choice, but neither should be interpreted in isolation from the outcome and the research question.

The distinction depends on what the recorded number means, not just on its appearance. The number of claims filed is a count; the fraction of an account spent on claims is a proportion. A percentage recorded from 0 to 100 is still a fractional response after division by 100. Conversely, an integer such as 0, 1, or 2 can be a category code rather than a count if it labels groups instead of measuring how many events occurred.

Count Outcomes and Exposure

A count model is appropriate when the response is an event total for a defined observational unit and period. A common conditional-mean specification uses a positive mean that depends exponentially on covariates:

This is a candidate mean specification, not a claim that every count is generated by a Poisson distribution. A Poisson regression model makes a stronger distributional assumption: conditional on the covariates, the count has a Poisson distribution with that mean. Whether that assumption is suitable is a separate modeling question. For choosing a model, the key first checks are that the response really is a count and that the conditional mean is the target.

Some units have different opportunities to experience events. A person observed for ten years has more time to accumulate visits than a person observed for one year. If the research question concerns an event rate and the expected count is assumed to be proportional to exposure \(t_i\), the mean can be written as \(t_i\lambda_i\). A log-mean representation then includes \(\log(t_i)\) as a fixed offset:

This specification assumes that doubling exposure, while holding other modeled factors fixed, doubles the expected count. That assumption may be reasonable for a stable event rate, but it is not automatic. If an extra year changes behavior, risk, or the population at risk, exposure may need a different treatment. If exposure is not relevant to the question or does not vary meaningfully, an offset may not be needed.

Fractional Responses: What Was Averaged?

A fractional response is bounded, so its conditional mean must also lie between zero and one. A standard option is a fractional-response mean model using a function such as the logistic function:

Because \(0<G(z)<1\) for finite \(z\), this specification keeps fitted conditional means within the unit interval. It can be used when observed responses include exact zeros and ones; the response does not have to be an individual binary outcome. The model concerns the conditional mean of the fraction. If a covariate pattern implies a mean exactly equal to zero or one, however, this finite-index logistic specification cannot represent it exactly.

A fraction may also be constructed as a number of successes divided by a known number of trials. In that case, a binomial model may be appropriate if the trials and their assumptions match the question—for example, if trials can be treated as conditionally independent with a common success probability. The denominator contains information: observing 0 successes out of 2 trials is not as precise as observing 0 out of 200 under that setup.

But not every proportion is a binomial success rate. A household's share of spending on housing, or a firm's fraction of sales in one region, may be a measured allocation rather than successes out of repeated trials. A fractional-response mean model can target the conditional mean of such outcomes without requiring a binomial trial interpretation. For a fractional response \(y_i\), the familiar Bernoulli-form criterion

can be used as a quasi-likelihood criterion with fractional \(y_i\), where \(\mu_i=G(X_i'\beta)\). With a fractional response it is not a literal Bernoulli log-likelihood. Its use for estimating the mean relies on a correctly specified conditional-mean relationship and suitable exogeneity and sampling conditions; it does not turn a share into a binary observation.

What Zeros Do—and Do Not—Tell You

A zero count is an allowed value of a count response, just as a zero share is an allowed value of a fractional response. A high frequency of zeros does not by itself show that the chosen outcome type is wrong or that a special zero-generating process is required. For counts, many zeros can arise when the conditional mean is low; they may also reflect differences in covariates, exposure, or unobserved heterogeneity. A substantive process that prevents some units from ever experiencing the event is another possibility, but it requires evidence and a clear interpretation.

For fractions, zeros and ones can have different meanings too. A zero could mean that a firm made no sales in a particular category, or that the category was not available to it. These processes imply different questions. A standard fractional-response model focuses on the conditional mean; it does not automatically model a separate probability of being exactly zero or one. Consider a model that treats endpoints separately only when that endpoint behavior is itself central to the question and the data support that distinction.

A Practical Selection Sequence

The following sequence keeps model choice tied to the data structure and the estimand. The model named at each step is a candidate; its assumptions still need to be assessed in the application.

Why the Distinction Matters

A linear conditional mean can predict negative counts or fractions below zero and above one. A support-respecting model avoids those impossible fitted means by construction, but it does not guarantee that the model's assumptions match the data. Count models also need careful attention to exposure and the event-generating context. Fractional models need a clear interpretation of the denominator and of whether the estimand weights observational units equally or weights their underlying trials.

Common pitfall: “There are many zeros, so I should use a fractional model,” or “the response is a percentage, so it must be binomial.” Both conclusions skip the meaning of the outcome. A count with many zeros remains a count; a percentage is binomial only when its numerator and denominator represent trials under a defensible binomial setup. Correct the reasoning by stating the unit, response support, denominator or exposure, and target conditional mean before selecting the model.

Model choice is not a contest to find the most elaborate specification. Start with the outcome's construction and a precise empirical question. Then use zeros, exposure, denominators, and distributional assumptions to decide whether the candidate mean model is credible and whether an additional model component is substantively warranted.