Zero-Success and All-Success Groups

Why Extreme Counts Need Careful Interpretation

A group of ten trials with no successes may look unusual, but it is a possible outcome of a binomial experiment. The same is true of ten successes in ten trials. Neither count is automatically an error, and neither should be altered simply because it lies at an endpoint. The key questions are whether the recorded counts describe the trials correctly and how the likelihood uses the observation.

In grouped binomial regression, group \(i\) has a trial total \(m_i\), a success count \(y_i\), and a success probability \(p_i\). Conditional on its covariates, the count is modeled as \(Y_i\sim\operatorname{Binomial}(m_i,p_i)\). Under logistic regression, \(p_i=\exp(\eta_i)/(1+\exp(\eta_i))\), where \(\eta_i=x_i^\mathsf{T}\beta\). A zero-success group has \(y_i=0\); an all-success group has \(y_i=m_i\).

For any probability strictly between zero and one, both endpoint counts have positive probability: \(P(Y_i=0)=(1-p_i)^{m_i}\) and \(P(Y_i=m_i)=p_i^{m_i}\). Their probabilities may be small for some \(p_i\) and trial totals, but “unlikely under a proposed model” is not the same as “impossible” or “incorrect.”

Likelihood Contributions at the Endpoints

For a binomial group, the likelihood contribution is the probability of the observed count. Its log-likelihood contribution, including the binomial coefficient, is

When \(y_i=0\), the success term is zero; when \(y_i=m_i\), the failure term is zero. The convention \(0\log(0)=0\) is used when evaluating the likelihood at an endpoint probability. In logistic regression with finite linear predictor, however, \(0<p_i<1\), so the logarithms in the expression are finite.

Score and Information from an Extreme Group

The score describes the local change in log likelihood as the model's linear predictor changes. Write \(\eta_i=\log\{p_i/(1-p_i)\}\), so that \(p_i\) is the logistic function of \(\eta_i\). For fixed \(m_i\) and \(y_i\), the group contribution is

Proof. Since \(p_i=\exp(\eta_i)/(1+\exp(\eta_i))\), its derivative with respect to \(\eta_i\) is \(p_i(1-p_i)\). Differentiating the displayed log likelihood gives

Differentiating again gives

For \(m_i>0\) and finite \(\eta_i\), \(0<p_i<1\), so \(m_ip_i(1-p_i)>0\). Substituting \(y_i=0\) or \(y_i=m_i\) into the score yields the two endpoint expressions. \(\square\)

In a regression model with \(\eta_i=x_i^\mathsf{T}\beta\), the group's contribution to the coefficient score is \(x_i(y_i-m_ip_i)\), and its contribution to the negative Hessian is \(m_ip_i(1-p_i)x_ix_i^\mathsf{T}\). A zero-success group therefore pulls its linear predictor downward locally, while an all-success group pulls it upward. The pull depends on the current fitted probability, the number of trials, and the covariates; it does not mean that the group independently sets its fitted probability to zero or one.

What Shared Fitted Probabilities Do with Extreme Groups

A group’s endpoint count is an observation, not a command to set the model probability to the same endpoint. When several groups share a probability under a common-intercept model, their evidence is combined. The following result makes that point precise, including the cases where the combined count itself is at an endpoint.

Proof. Independence gives the likelihood, up to factors that do not depend on \(p\), as

For \(0<Y<M\) and \(0<p<1\), its log likelihood, omitting constants, is \(Y\log(p)+(M-Y)\log(1-p)\). Its derivative is \(Y/p-(M-Y)/(1-p)\), which is zero exactly when \(Y(1-p)=(M-Y)p\), or \(p=Y/M\). Its second derivative is \(-Y/p^2-(M-Y)/(1-p)^2<0\), so this is the unique maximum. If \(Y=0\), the likelihood is proportional to \((1-p)^M\), which is maximized at \(p=0\). If \(Y=M\), it is proportional to \(p^M\), which is maximized at \(p=1\). \(\square\)

Checking Records Without “Fixing” Valid Extremes

A useful first check is mechanical: for each record, verify that the trial total is a nonnegative integer and that the success count is an integer between zero and that total. For example, \(y=12,m=10\) is outside the binomial support and cannot be a success count for ten trials. It must be investigated before fitting the model. By contrast, \(y=0,m=10\) and \(y=10,m=10\) pass the support check.

Passing the range check does not prove that a record was entered correctly. A valid-looking count may still have the wrong denominator, an incorrectly assigned group, or a reversed success/failure coding. Check the data definition and source records when the context suggests a problem. Do not replace a zero by one, reduce an all-success count, or discard an extreme group merely to avoid an endpoint. Such changes alter the data and the likelihood and require a justified correction, not a preference for less extreme counts.

Extreme observations can also have substantial influence. The score magnitude for a zero-success group is \(m_ip_i\), and for an all-success group it is \(m_i(1-p_i)\). Thus the effect depends on both the group size and the current model probability. Their influence should be interpreted with the covariate pattern and the other observations in mind. A single extreme group is not, by itself, enough to determine its regression probability; the model fits probabilities jointly through its specified predictors.

Common Pitfall

Wrong reasoning: “There were no successes, so the success probability for this group must be zero; I should enter a zero probability or add a small count so the model can be fit.” This confuses the observed count with the probability that generated it. Under a binomial model with \(0<p_i<1\), zero successes in \(m_i\) trials has probability \((1-p_i)^{m_i}>0\). The correction is to retain the valid count, check the record, and let the specified model combine its likelihood contribution with those of the other groups. If an actual count or denominator is inconsistent, correct it from its source rather than applying an arbitrary adjustment.