The effect of level of theory on the imaginary frequency and corresponding tunnelling coefficients has been studied for a test set of hydrogen abstraction reactions: .CH2X + CH3Y --> CH3X + .CH2Y for (X,Y) = (H,H), (H,CN), (H,F), (H,Li) and (F,Li). It is found that the imaginary frequency is very sensitive to the level of theory used, with Hartree-Fock (HF) methods severely overestimating the imaginary frequency compared with high-level CCSD(T)/6-311G(d,p) calculations. The errors for the other methods are smaller but nonetheless significant, with MP2 methods overestimating the imaginary frequency and density functional theory (DFT) methods underestimating it. In the case of the HF methods, this leads to errors in the tunnelling coefficient of several orders of magnitude, while for the better DFT and MP2 methods errors of a factor of 2-3 are observed. To address this problem, an IRCmax procedure for estimating the imaginary frequency has been developed and it is found that IRCmax imaginary frequencies calculated with CCSD(T)/6-311G(d,p) single points along a low-level HF/6-31G(d) minimum energy path provide excellent approximations to the high-level values, at a fraction of the computational cost.