Three bivariate statistical methods to predict the family potential to produce elite progeny were studied to improve the efficiency of a sugarcane (Saccharum spp.) breeding program. Progeny from 15 piparental crosses were evaluated in plant cane and first ratoon seedlings, and in clonal plant cane plots during 1989 and 1990. The bivariate predictions of Brix combined with cane yield components (stalk number, stalk weight, stalk diameter, stalk length, and stool weight) were investigated. The best linear unbiased predictors (BLUPs) and the sum of ranks based on family mean values of two traits (RANK) were repeatable among tests in the estimation of family potential. Bivariate normal probabilities (PROB) estimated with family means, phenotypic standard deviations, and genetic correlations generally demonstrated poor repeatability among tests. The three statistical predictions were compared with the progeny selection rate within the crosses through three selection stages. Predictions were not correlated to the selection rates of eight crosses with smaller initial progeny populations (< 500 progeny). However, when the predictions were compared with the 7 of 15 families over which 1,000 progeny for each cross had been evaluated, the rankings based on BLUP and RANK bivariate predictions of Brix and stool weight identified the better crosses. PROB was inconsistent in this regard. Early selection work is highly subjective. We speculate that near-random selection occurs for stalk number at the initial selection stage and that the high selection rate at this stage (≈5%) generates a first clonal population (10 to 25) that is too small to accurately base selection rates for stalk number. Larger initial progeny populations produce sufficiently large clonal populations (>50) to appraise crosses using selection rates. The study suggested that family evaluations for breeding programs can use bivariate predictions. The comparative ease of calculating the RANK estimate versus the BLUP along with the absence of any apparent loss of predictive value suggests that the RANK method would be the most suitable statistic to use for bivariate predictions.