This study attempts to give a more adequate account for Mandarin third tone sandhi (T3S) in the light of the optimality theory. In view of the possible discrepancy between the researcher's intuition and the actual behavior, this study bases its analysis on the real outputs of strings of T3 words produced by fluent speakers of Mandarin. It is found that Mandarin TS domains are the result of the interaction between various constraints, i.e. the Lexicon, NAL, AR/NAR, Maximal Domain & NST3 constraints. The former three are derived from the morpho-syntactic consideration of the utterances concerned and the latter two from the general principle of prosodic structure in Mandarin. As for their rank, the Lexicon is the highest, the NAL the second, and the then NAR/AR, Maximal Domain & NST3 constraints. In general, the subjects in the experiment share strong consensus about the reading patterns of lexemes and frequently cited idioms, but beyond this level, variation arises because of the competition among the proposed constraints. Those tone patterns that violate no constraint or violate less of the lower-ranked constraints are al potential optimal forms. Moreover, it is found that performance and pragmatic constraints may be placed even higher than these formally-oriented constraints. However, to make this finding more convincing, a further study is suggested.