It is known that, in the chi-square test of the goodness of fit, 

the expected frequency of each cells should be greater than $5$ 

for the large sample theory to hold, and thus that, if there 

are cells with small observed frequencies, one should group them 

so that the new grouped cell has the observed frequency greater 

than $5$. In the present paper, we treat the problem of grouping 

of small cells in the test of goodness of fit from the viewpoint 

of a Bayes approach to the decision theoretic framework of model 

fitting proposed by Inagaki(1977b). Then we have two errors, one 

of which is caused by the estimation of probabilities of cells 

and the other by the grouping of small cells, and obtain the 

exact and asymptotic representations of two errors explicitly. 

By using them, we compare a new Bayes grouping rule of this model 

fitting to the usual grouping rule of small cells.