Anderson (1991) provided a Bayesian analysis of feature predicting; if an object contains feature F and belongs to category k, one can predict a novel feature j by using the following formula: . This is one method for calculating how likely the object is to be in each category k and how likely that category is to contain the property. Thus, one should consider all the categories in order to make the prediction. In short, the analysis suggests that people use multiple categories to make predictions when categorizing is uncertain.

Murphy & Ross (1994) suggested that people make feature predictions on the basis of a single category when categorization is uncertain. They found that even if the participants gave a fairly low rating of confidence in categorizing, they did not use multiple category information to make predictions.

Wang Moyun & Mo Lei (2005) presented another viewpoint—feature predicting is based on overall conditional probability instead of the probability of categorizing.

Molei & Zhao Haiyan (2002) found that the association or separation of the two dimensions would influence feature predicting under uncertain categorizing circumstances. The results suggested that the proportion of the association of the object and the feature (Ak) be incorporated into the Bayesian formula: .

However, we found that most previous experimental data were significantly higher than the value of the two-formula theory. We revised the formula as follows: . In addition, we manipulated one factor to test the new formula.

The results in Experiments 1 and 2 show that when the two feature dimensions are not in conjunction with non-target categories, raising the proportion of the coexistence of the dimensions within the non-target categories will not enhance the feature prediction probability. When two feature dimensions are in conjunction in non-target categories, raising the proportion will enhance the feature prediction probability. The results are not consistent with Murphy & Ross’s single-category viewpoint and Anderson’s Bayesian Rule.

Accordingly, this study introduces the proportion of conjunction of the two feature dimensions as a multiplicative variable into the formula of the Bayesian rule. The result of Experiment 3 is consistent with that of the study and shows that raising the proportion of the coexistence of the two feature dimensions within the target category will not improve the probability of feature prediction.

The experimental outcomes are consistent and a better fit with our new, revised Bayesian rule. The coexistence of the two dimensions within the non-target categories promotes the use of non-target category information; the coexistence of the two dimensions within the target categories promotes the use of non-target category information.