Abstract: Seismic-induced dynamic responses of tunnel structures are governed by the spatial variability of surrounding soil properties. To quantify the influence of such variability in loess strata, a twin-tube metro tunnel in a loess region was investigated. Cross-covariance matrix decomposition was employed to generate random fields of elastic modulus (E), cohesion (c), density (ρ) and internal friction angle (φ) of the loess. A non-intrusive stochastic finite-element framework combined with Monte-Carlo simulation was then used to examine how the randomisation of these parameters and their coefficients of variation (COVs) affect the seismic response of the tunnel. The results revealed that all parameters produced similar patterns in the internal-force response; stress fluctuations were approximately ±50% relative to the mean-field solution. Elastic modulus and cohesion exerted the dominant influence, followed by density and internal friction angle; therefore, the spatial variability of (E) and c should be prioritised in seismic design. Tunnel response increased monotonically with the COV, and the most critical sections were located at 45°, 135°, 225° and 315° of the lining. A higher COV amplified the stress amplitude by 30 %~40 %, and mean-field analyses were shown to underestimate the seismic hazard at the 95 % confidence level.