The CHITEST function currently calculates a goodness of fit test if the expected and observed range is an n by 1 or 1 by n range.

If the ranges are m by n with both m and n at least 2, then Excel and OOo appear to use a degree of freedom of (n-1) * (m-1) which is only appropriate if the expected values are calculated from the observed values to perform a test of independence or a test of homogeneity. Of course in this case we should not need to specify the expected values. (Behaviour A)

If the ranges are m by n with both m and n at least 2, one could also make sense out of this function by using degrees of freedom (n*m-1). In that case it would simply be a goodness of fit test (assuming that the expected values were determined apriori (rather than fromthe observed values). (Behaviour B)

Currently in the case of the ranges being m by n with both m and n at least 2, gnumeric uses the degree of freedom (n-1). This does not make any sense at all.

Gnumeric's behaviour should be changed to reflect either Behaviour A or B. 

The strength of Behaviour A is compatibility with Excel (and probably OOo) with the drawback that the results are only meaningful if the expected values are appropriate constructed, 

THe strength of Behaviour B is that it makes the most mathematical sense (unless the expected values are constructed from the observed ones.)