P-wave velocity variations may provide precursory information before large earthquakes. We present results from a simple study on variations of P-wave travel-time residuals before and after the 1999 Chi-Chi, Taiwan, earthquake. The short-period seismic network deployed in Taiwan by the Central Weather Bureau (cwb) is routinely used for earthquake location. The accuracy of earthquake location is often judged by the travel-time residual of seismic waves. Because the velocity model used by the cwb was one-dimensional, the travel-time residuals will show lateral inhomogeneity of the velocity structures and unmodeled vertical structures. Also, the travel-time residuals may change with time if the velocity structure changes temporally. We used the data of the Central Weather Bureau Seismic Network (cwbsn) from 1991 to 2002 to obtain the difference between observed and calculated P-wave travel times, which is defined as the P-wave travel-time residual. The purpose of this study is to find the mean value of the P-wave travel-time residuals at each station and its possible temporal changes at stations near the Chelungpu fault. Our results show that the mean P-wave travel-time residuals at stations 40 km east of the Chelungpu fault changed very little before and after the Chi-Chi earthquake. But the mean P-wave travel-time residuals at the stations immediately west of the Chelungpu fault clearly changed before and after the Chi-Chi earthquake. The anomalous zone is bounded by stations NSY, TWQ1, TCU, WNT, WGK, CHN5, YUS, WDT, and WHF. These temporal changes of P-wave travel-time residuals can be attributed to changes in the velocity structure, which in turn might be caused by crustal deformation east of the Chelungpu fault beginning about 6 years before the Chi-Chi earthquake. In other words, before the Chi-Chi earthquake, there was a long-term 6-year precursor of P-wave velocity decrease possibly due to dilatancy or development of cracks. The variations of P-wave travel-time residuals with time at the anomalous stations are also confirmed by statistical tests that show clear significance. This method can be used to monitor and identify variations of P-wave travel-time residuals which may be precursors to a large earthquake.