Can fractional logit include observed zeros and ones?

Yes, unlike taking the logit of each observed outcome directly.

Is it the same as ordinary beta regression?

No. They make different modeling assumptions and handle boundaries differently.

Is a coefficient a percentage-point effect?

Usually no. Translate the bounded mean function into a derivative or discrete change for a specified scenario.

AI Glossary / Statistics and probability / Advanced

Fractional model

Definition

A fractional model is a statistical model for an outcome expressed as a fraction or proportion, usually between zero and one. In fractional response regression, the aim is often to model the conditional mean of that outcome as a function of explanatory variables while keeping predicted means within the unit interval. Examples include a firm's participation rate or the share of a resource allocated to an activity. This entry uses the fractional-response meaning; it does not refer to fractional factorial experiments, fractional calculus or fractional-integration time-series models.

What it means

A fractional outcome can take intermediate values such as 0.27, rather than only the binary values zero and one. Modeling it with unrestricted linear predictions can produce values outside its feasible range. A bounded mean function offers a way to respect that range, while still examining associations with predictors. The fraction's origin matters: a proportion formed from a known count of trials may contain information about its denominator, whereas a continuous allocation share may not have that binomial interpretation. Those data-generating differences affect appropriate modeling and uncertainty estimates.

How it works

A common fractional logit specification is E[Y given X] = logistic(X beta), where logistic(z) = 1/(1 + exp(-z)). Estimate the parameters using a Bernoulli quasi-likelihood criterion y log(mu) + (1 - y) log(1 - mu), even when y is fractional rather than Bernoulli. With a correctly specified conditional mean and suitable regularity conditions, quasi-likelihood methods support consistent mean-parameter estimation without requiring a Bernoulli outcome distribution. Use appropriate robust or clustered uncertainty estimates. Fractional probit uses a different bounded link. Observed zeros and ones can be included in fractional logit, while ordinary beta regression typically assumes outcomes strictly inside the interval.

Three practical examples

Participation research: model how a firm's participation share varies with measured characteristics without treating the share as a binary indicator.
Resource analysis: investigate associations between operating conditions and the fraction of capacity utilized.
Service evaluation: model completion proportions across comparable units, while considering varying denominators and dependence within organizations.

Think of it this way

Think of a measuring dial whose needle must stay between empty and full. A bounded mean model keeps its expected reading inside that range. It does not tell you that every dial with the same predictors will show exactly that reading, or explain by itself why the reading changes.

Worked example

Suppose a hypothetical fractional logit model has linear predictor -1 + x. At x = zero, the predicted mean fraction is 1/(1 + exp(1)), approximately 0.26894. At x = one, it is 0.5. The discrete increase across that one-unit change is therefore about 0.23106, or 23.106 percentage points. The coefficient of x is one, but this does not mean that the outcome fraction increases by one or by 100 percentage points. The local derivative at x = zero is mu(1 - mu) times one, approximately 0.19661, which differs from the finite one-unit change because the response is nonlinear. These calculations use stipulated coefficients; they do not estimate a relationship from data or establish a causal effect. They also predict a conditional mean, not the probability that an entire unit has an outcome exactly equal to one.

Illustrative case study

Fictional case: Aspen Learning Analytics models course-completion shares across centers. Analysts preserve the counts underlying each share and investigate whether center size and repeated observations affect estimation. They report predicted mean differences for meaningful scenarios rather than describing logit coefficients as percentage-point changes. When management asks whether changing one predictor would cause improvement, the team distinguishes the fitted association from a causal claim. This invented case illustrates bounded-outcome interpretation without claiming that a particular model solves every participation-rate problem.

Common misconceptions and mistakes

Do not convert a fraction into a binary outcome merely to use a familiar classifier. Do not assume the Bernoulli quasi-likelihood requires the observed fractions to be actual zero-one trials. Conversely, do not ignore available trial denominators when they are relevant to precision and the estimand. A direct logit transformation of the outcome is undefined at zero and one and changes the modeled quantity. Robust standard errors do not repair an incorrect mean specification or endogenous predictors. A bounded prediction does not guarantee good calibration, and a single mean model may be inadequate when boundary masses reflect distinct processes.

Business use

Businesses can use fractional models to analyze participation, utilization and allocation measures while respecting their feasible range. Define what the numerator and denominator represent and whether different units should receive equal or denominator-based weighting for the intended question. Evaluate predictions on appropriate held-out observations and inspect behavior near the boundaries. Report marginal effects or scenario differences with uncertainty and clear units. If several shares must sum to one, separate fractional regressions may not enforce that joint constraint; a compositional or multivariate approach may be more appropriate.

Check your understanding

A model predicts a mean participation share of 0.4 for a center. Does that mean a 40% probability that the center has complete participation?

Answer: No. The prediction concerns the expected fractional outcome; the probability of exactly full participation is a different quantity.

Related terms

Conditional means
Fractional response regression
fractional logit
fractional probit
quasi-maximum likelihood
beta regression
conditional mean
marginal effect
bounded outcome
compositional data

Learn next

Learn how quasi-likelihood differs from specifying a full outcome distribution. Then compare fractional response, binomial proportion and beta models, paying attention to boundary observations, denominators and the intended estimand.

Sources

Papke and Wooldridge, Econometric Methods for Fractional Response Variables, primary working-paper record: https://www.nber.org/papers/t0147 Stata, Fractional response regression manual: https://www.stata.com/manuals/rfracreg.pdf Stata, Fractional outcome regression: https://www.stata.com/features/overview/fractional-outcome-models/ Stata, Logit transformation FAQ: https://www.stata.com/support/faqs/statistics/logit-transformation/

Previous: Fractional factorial design Next: Frame averaging