This paper considers interval estimation of the difference of marginal proportions for binary matched-pairs data. We adapt the modified p-value method in Lin and Yang (2006) to obtain improved confidence intervals, which are based on the modified p-value derived by one two-sided score test statistics in Tango (1998) and Tang, Tang, and Chan (2005). The improved confidence interval guarantees its corresponding coverage probability greater than or equal to the nominal confidence level. Compared to the confidence interval derived by the exact unconditional method in Tang et al. (2005), numerical results show that the improved confidence interval has the actual confidence levels closer to the nominal confidence levels and shorter mean confidence widths for small to moderately large sample sizes.