I will present experimental findings on the decoding failure rate (DFR) of BIKE, a linear code-based fourth-round candidate in the NIST Post-Quantum Cryptography Standardization process, at the 20-bit security level. We directly compute the average DFR for a range of BIKE block sizes and detect the threshold for the "error floor" phenomenon, whereby decoding failure rates sharply decrease as the block size increases, but then abruptly flatten out past a certain point. We then study the influence of various factors on the average DFR, such as the syndrome weight and proximity to three specific classes of near-codeword. I will also discuss work in progress using graph-theoretic techniques to refine our understanding of the error floor, aimed at extrapolation to larger parameter sizes. No background in cryptography will be assumed. This is joint work with Sarah Arpin, Tyler Billingsley, Jun Bo Lau, Ray Perlner, and Angela Robinson.