Tomlinson-Harashima precoding (THP) is an attractive non-linear interference cancellation method for multipleinput multiple-output (MIMO) system with decentralized receivers. Compared to linear methods, it is able to achieve significant performance improvement while introduce negligible computational complexity increase. However, THP is sensitive to inevitable channel state information (CSI) errors at the transmitter in wireless communications. We statistically model the errors caused by channel estimation and feedback delay. By taking this modeling into account, we propose a flexible lowcomplexity compensation method for THP with imperfect CSI. In contrast to several robust designs optimized for one specific THP implementation, our approach can achieve average robustness for various THP implementations with negligible complexity overhead. Several canonical and emerging preprocessing algorithms, e.g, sorted QR decomposition (SQRD), vertical Bell Laboratories Space Time (V-BLAST) and lattice reduction (LR), are evaluated under a 802.11ac based MIMO OFDM simulation system. The experimental results show that the proposed compensation approach is able to lower the error floors caused by imperfect CSI significantly for various THP implementations.