Abstract: Objectives： The rapid deployment of LEO mega-constellations, represented by Starlink, has brought unprecedented challenges to space situational awareness and space traffic management. Publicly available orbit data for Starlink primarily consist of Two-Line Elements (TLEs) published by NORAD and Operator Ephemerides (OE) provided by SpaceX, which differ fundamentally in format, update mechanism, and accuracy. A systematic and quantitative evaluation of the error characteristics, covariance realism, update patterns, and prediction accuracy decay of both data sources is conducted across different orbital phases. Methods： Historical orbit data of approximately 6,300 Starlink satellites covering April-May 2024 were collected. A continuous "quasi-truth" reference orbit was constructed by concatenating the most recently published segments from successive 72-hour OE batches, exploiting the approximately 64-hour overlap between consecutive batches. TLE states were transformed from the True Equator, Mean Equinox (TEME) frame to the J2000.0 frame, and a 9th-order Lagrange interpolation was applied to the 60-second sampled OE data to enable state comparison at arbitrary epochs. Satellites were categorized into three orbital phases—orbit-raising, station-keeping, and deorbiting—based on the daily change rate of the semi-major axis, for granular analysis. OE prediction errors were evaluated in the Radial, Transverse, Normal (RTN) frame and compared against the formal covariance. TLE performance was evaluated in terms of update frequency, spatiotemporal distribution of epoch times, initial epoch accuracy, and prediction accuracy decay over 72 hours. A real-world conjunction event in December 2025 was independently reproduced to validate the practical utility of the findings. Results： (1) For station-keeping satellites, the actual OE prediction error is heavily dominated by the along-track component, with Root Mean Square (RMS) values of 0.20 km, 2.45 km, and 6.66 km at prediction horizons of 8, 24, and 48 hours, respectively. A distinct performance discontinuity occurs around 48 hours, which is attributed to a simplification in the propagator model that only considers the J2 perturbation after 48 hours. While the formal position covariance aligns well with actual errors (e.g., actual along-track error of 2.45 km vs. formal uncertainty of 2.32 km at 24 hours), the velocity covariance is significantly underestimated, as exemplified at 72 hours where the actual radial velocity error reaches 9.42 m/s against a formal uncertainty of only 0.41 m/s. (2) For TLEs, the update frequency is highly strategy-driven and strongly coupled with the orbital phase. The median update interval is 8.00 hours for the orbit-raising and deorbiting phases, and 9.57 hours for the station-keeping phase. High-frequency updates with intervals of less than 2 hours account for 8.72% of the total dataset, demonstrating a rapid-response capability. The epoch times of TLEs show distinct spatiotemporal patterns: they are tightly clustered at integer UTC hours during the active orbit-raising and deorbiting phases, but are aligned with the ascending node crossings during the passive station-keeping phase. The median 3D position error at the TLE epoch is 1.01 km across the entire constellation, showing better accuracy during active maneuver phases (0.63 km for deorbiting and 0.87 km for orbit-raising) than the station-keeping phase (1.03 km). However, TLE prediction accuracy degrades rapidly over time; the median along-track errors at 8 and 24 hours are 1.73 km and 16.73 km for orbit-raising, 1.95 km and 6.77 km for station-keeping, and 0.31 km and 5.42 km for deorbiting, respectively. (3) In the independent reproduction of the conjunction event between the Kinetica-1 Y11 (Lijian-1 Y11) rocket payloads (NORAD 66993) and Starlink-6079 (NORAD 56120) on December 2025, the minimum distance was 0.60 km at 03:30:18.93 UTC, consistent with the sub-kilometer scale disclosed by official sources. The maximum collision probability computed using the Foster method was 1.54×10^-5, exceeding the standard safety threshold of 10⁻⁵. Conclusions： The quantitative analysis reveals significant differences in the performance and characteristics of OE and TLE data, which are closely tied to satellite orbital phases. OE provides higher short-term accuracy with generally reliable position uncertainty information, but its velocity uncertainty is underestimated. TLE update frequency and epoch selection are strategy-driven, and its forecast utility diminishes quickly, especially during raising and deorbiting phases. These findings provide quantitative references for orbit data application and collision warning research for current and future LEO mega-constellations.