Using data from 2000 and 2004 presidential elections and data from the 2000 population Census, this paper tests the hypotheses regarding the dramatic increase of invalid votes in 2004. First, the less-urbanized counties showed a greater increase in the proportion of invalid votes, and whether the DDP is the county ruling party does not show any significant differences. Second, as the definition of invalid votes became broader in 2004, townships with an above-median proportion of illiterate voters significantly increase dir the proportion of invalid votes by 1.392%. Regression results show that for a 1% increase in the illiteracy rate, the proportion of invalid votes increases by 0.21%. Estimation from King’s ecological inference model also shows that the probability of casting invalid votes for illiterate voters increases significantly in 2004 and is of vote-rigging based on the observation that the share of votes gathered by Chen-Lu and the proportion of invalid votes are positively correlated. It is shown that this correlation is not causal. After controlling for the illiteracy rate, the effect of the share votes gathered by Chen-Lu on the proportion of invalid votes is reversed and becomes negative, although not statistically significant. In sum, the hypothesis that invalid votes are caused by the broader definition of invalid votes is the only hypothesis consistent with the data.