NEWS

qountstat 0.2.0 (2026-10-16)

New function Overdispersion.test() for a model-agnostic, simulation-calibrated diagnostic of overdispersion in grouped count data (Daniels et al., 2026). The decision threshold is calibrated by Monte Carlo simulation for the design at hand, which makes the diagnostic applicable to small sample sizes. The test is one-sided, i.e. underdispersion is not evaluated.
Dunnett.GLM gained the argument 'family', which selects a quasi-Poisson (default, as before), a negative binomial or a classical Poisson regression. Dunnett.GLM.power and Dunnett.GLM.bMDD gained the corresponding argument 'Dunnett.GLM.family'. The information object of Dunnett.GLM now reports the distribution, the link function and the dispersion parameter that were used.
CPCAT and Dunnett.GLM gained the argument 'overdispersion.test', which adds the result of the overdispersion diagnostic to the information object. It is switched off by default and does not change the results of the calling function when enabled.

The Hampel identifier in CPCAT is no longer evaluated by default; the default of 'hampel.threshold' changed from 5 to NULL. The identifier is not scale-invariant and flags Poisson-distributed data as overdispersed increasingly often as the counts grow. Use Overdispersion.test for a scale-invariant alternative.
Counts are now rejected if any value is negative. The check used min() instead of any() and therefore only triggered if all values were negative.
family = 'nb' can no longer be combined with zero.treatment.action = 'identity.link' if a treatment consists only of zeros, because the negative binomial dispersion parameter is not identifiable when a fitted mean is zero. Use zero.treatment.action = 'log(x+1)' instead.

NA values in the counts no longer cause an error; rows containing them are excluded as documented. Affects CPCAT, Dunnett.GLM and the power and bMDD functions for count data.
The note on excluded rows reported TRUE instead of the number of rows that were excluded.