The determination of the cross-over values of the factors was at first hindered because of the poor viability of some of the mutants. If the viability of each mutant type could be determined in relation to the viability of the normal, "coefficients of viability" could serve as corrections in working with the various mutant characters. But it was found (Bridges and Sturtevant, 1914) that viability was so erratic that coefficients might mislead. At the same time it was becoming more apparent that poor viability is no necessary attribute of a character, but depends very largely on the condition of culture. Competition among larvae was found to be the chief factor in viability. Mass cultures almost invariably have extremely poor viability, even though an attempt is made to supply an abundance of food. Special tests (Morgan and Tice, 1914) showed that even those mutants which were considered the very poorest in viability were produced in proportions fairly close to the theoretical when only one female was used for each large culture bottle and the amount and quality of food was carefully adjusted. For the majority of mutants which did well even under heavy competition in mass cultures the pair-breeding method reduced the disturbances due to viability to a point where they were negligible. Later a method was devised (Bridges, 1915) whereby mutations of poor viability could be worked with in linkage experiments fairly accurately and whereby the residual inviability of the ordinary characters could be largely canceled. This method consists in balancing the data of a certain class with poor viability by means of an equivalent amount of data in which the same class occurs as the other member of the ratio. Thus in obtaining data upon any linkage case it is best to have the total number of individuals made up of approximately equal numbers derived from each of the possible ways in which the experiment may be conducted. In the simplest case, in which the results are of the form AB:Ab:aB:ab, let us suppose that the class ab has a disproportionately low viability. If, then, ab occurs in an experiment as a cross-over class, that class will be too small and a false linkage value will be calculated. The remedy is to balance the preceding data by an equal amount of data in which ab occurs as a non-cross-over. In these latter the error will be the opposite of the previous one, and by combining the two experiments the errors should be balanced to give a better approximation to the true value. When equal amounts of data, secured in these two ways, are combined, all four classes will be balanced in the required manner by occurring both as non-cross-overs and as cross-overs. The error, therefore, should be very small. For three pairs of gens there are eight classes, and in order that each of them may appear as a non-cross-over, as each single cross-over, and as the double cross-over, four experiments must be made. {20}