Information Under Overdispersion

Why Overdispersion Changes Information

A sample-size calculation depends on the variability of the estimator, not just on the number of participants or outcomes recorded. If the outcome varies more than the planning model assumes, the same nominal sample size gives a less precise estimate. In information terms, it contributes less information than the nominal model predicts. This distinction matters in a group sequential trial: interim information targets should reflect the variability actually assumed or observed, rather than treating every outcome as equally informative under a possibly inadequate model.

The simple adjustment developed here applies when overdispersion multiplies the variance for the estimand by a common factor. It is a variance and information adjustment, not a universal correction for every form of dependence or model misspecification. For a particular calculation, first specify the outcome model and the variance being compared.

In the simple setting where nominal variance is proportional to \(1/n\), this definition gives \(n_{\mathrm{eff}}=n/\phi\). Thus, a dispersion factor of 1.5 means that \(n\) observations provide the same variance-based information as \(n/1.5\) nominal observations. This equivalence concerns the specified estimand and variance calculation; it does not mean the data literally contain fewer observations.

Binary Outcomes: A Scalar Variance Adjustment

For a binary outcome with probability \(p\), the nominal variance of a sample proportion from \(n\) independent, identically distributed observations is \(p(1-p)/n\). If a suitable grouped, clustered, or estimating-equation model gives variance \(\phi p(1-p)/n\), then

The overdispersed standard error is \(\sqrt{\phi}\) times the nominal standard error, because standard errors are square roots of variances. This calculation assumes that the same scalar factor applies to the variance of the estimand being studied. For a two-arm risk difference, for example, that condition is met if a common \(\phi\) multiplies each arm's binomial variance; the resulting variance of the difference is also multiplied by \(\phi\).

There is an important modeling qualification. A single Bernoulli observation with a fixed probability \(p\) has variance exactly \(p(1-p)\); it cannot have a larger Bernoulli variance while retaining that same probability. In practice, binary overdispersion can arise in grouped outcomes, through dependence among individuals, or relative to a model that does not capture relevant structure. A scalar adjustment is only appropriate when it adequately describes how that structure changes the variance for the target estimand.

Where a Dispersion Factor Can Come From

One common binary example is a cluster of \(m\) individuals whose outcomes have marginal probability \(p\) and equal pairwise correlation \(\rho\). Let \(S\) be the number of outcomes in one cluster. Since each binary outcome has variance \(p(1-p)\), the variance of their sum is the sum of the individual variances plus the covariances:

Compared with the independent-binomial variance \(mp(1-p)\), the multiplier is \(\phi=1+(m-1)\rho\). With equal-sized clusters and a common correlation, this yields a variance-equivalent effective count of \(n/\phi\). The derivation requires the stated equal-size and equal-correlation assumptions; unequal cluster sizes or more complicated dependence may not be summarized accurately by this expression. It is also the variance for this cluster total under these assumptions, not a claim that all features of clustered data are equivalent to independent observations.

A dispersion factor can also be used in a quasi-binomial variance specification. In either application, the factor must be justified for the analysis being planned. It should not be treated as an automatic property of all binary data.

Count Outcomes: Check the Variance Model

For independent count observations with a common mean \(\mu\), a Poisson model assumes \(\operatorname{Var}(Y)=\mu\). Under a scalar quasi-Poisson model, \(\operatorname{Var}(Y)=\phi\mu\). The sample mean then has variance \(\phi\mu/n\), compared with \(\mu/n\) under the Poisson model. Its variance-based information is therefore reduced by a factor of \(\phi\). For a log-mean or log-rate analysis with equal exposure, the corresponding nominal information is proportional to \(n\mu\), and under this scalar model it is \(n\mu/\phi\).

This calculation presumes independent observations and a common multiplicative dispersion factor for the relevant variance. It is not a general rule that every count model has one constant \(\phi\). For instance, a negative-binomial variance of \(\mu+\alpha\mu^2\), with \(\alpha>0\), has a variance ratio to Poisson of \(1+\alpha\mu\). That ratio changes with the mean, so a single constant effective sample size may not represent information across different mean values.

From Information Targets to Enrollment

When a nominal design specifies a sample size to achieve a target variance or information, a constant dispersion factor provides a direct adjustment. If the nominal sample size is \(n_0\), then the overdispersed variance at sample size \(n\) is \(\phi\) times the nominal variance at that same \(n\). Since nominal variance is proportional to \(1/n\), matching the original target requires increasing the sample size by \(\phi\).

This calculation assumes the factor is known or fixed for planning and remains stable across the design. It adjusts the information available at a given enrollment; it does not by itself specify how a sequential trial should make decisions at interim analyses.

Diagnostics, Alternatives, and Limitations

Before using a scalar adjustment, ask whether the proposed variance model is plausible for the outcome and estimand. For count data, compare observed variability with the Poisson mean-variance assumption using residual or grouped summaries appropriate to the fitted model. For grouped binary data, examine variability across groups relative to the binomial model. In either case, investigate whether apparent extra variability is associated with dependence, different group sizes, unequal exposure, or a mean-dependent variance pattern. A single estimated ratio can conceal those distinctions.

The main alternatives answer different modeling needs. A scalar quasi-likelihood adjustment can be useful when the mean model is retained and variance inflation is approximately proportional. A distribution that models the source and shape of extra variation, such as a beta-binomial model for grouped binary outcomes or a negative-binomial model for counts, can be more suitable when the variance pattern depends on the mean or group structure. A robust sandwich variance estimate can protect inference against some forms of variance misspecification when its assumptions are met, but it does not create additional information or automatically solve a small-number-of-clusters problem. The appropriate choice depends on the data structure and target estimand, not on which adjustment gives the largest standard error.

In sequential planning, a stable scalar factor reduces information at every look by the same proportion. Consequently, ratios of information across looks may remain unchanged if the same factor applies throughout, even though the absolute information is lower. If the factor differs by stage, group composition, or outcome mean, an enrollment fraction need not equal an information fraction. Planning from enrollment alone can then misstate how much information has accrued.

The effective-sample-size calculation is a planning and interpretation device, not a substitute for a suitable analysis model. It summarizes one variance comparison. It does not erase dependence, correct a misspecified mean model, or guarantee that all contrasts in a multi-parameter analysis lose information by the same factor.