AbstractAncillary statistics, proposed by Fisher (1925), can be constructed by forming amixture model(Birnbaum 1962) or can be extracted or derived from atransformation‐parameter model(Peisakoff 1951, Fraser 1961) or from the corresponding error‐basedstructural model(Fraser 1968, 1979); these latter models involve an implicitmixturestructure. Compound models with ancillaries can also be formed by a crossembedding, discussed from a technical viewpoint in this paper. Of the 25 examples in Buehler (1982), 22 are mixtures or implicit mixtures and 3 correspond to cross embedding. The cross embedding examples exemplify the nonuniqueness difficulties with ancillaries. This paper discusses a simple and two generalized versions of cross embedding but makes no general valuations of these for statistical inference; their role within inference is discussed in Evans, Fraser, and Monette (1984, 19