The existence of maximum likelihood estimates for a class of heterosce
dastic regression models is considered. For a given dispersion functio
n we show that, under a weak condition, the likelihood is singular at
points corresponding to nonreplicated observations, causing unrestrict
ed maximum likelihood estimation to break down, whilst for an alternat
ive class of dispersion functions we obtain a much stronger linear ind
ependence condition for the likelihood to be unbounded.