Methods for non-invasive quantification of the cardiorespiratory coupling combine information from the heart rate variability (HRV) and respiratory signals. Building the HRV, however, is problematic in presence of abnormal beats. This hinders the evaluation of the cardiorespiratory coupling, in particular when methods based on autoregressive models are used, due to their sensitivity to outliers. In this work, an approach based on robust regression is proposed. A weighted Least Squares Support Vector Machines regression model is used. The residuals of the prediction are used to derive a parameter that changes according to the strength of the coupling and is robust to outliers. To test the proposed approach, respiratory and ECG signals from 100 sleep apnea patients recorded during full-night polysomnography were used. The HRV was derived interpolating the RR interval time series. Both, HRV and respiration were segmented into 5-minutes epochs. The proportion of power linearly correlated with the respiration in the HRV was derived using subspace projections. An analysis of surrogates confirmed a linear coupling between the signals. Ten levels of coupling were defined to split the segments into 10 groups. Afterwards, the epochs were contaminated with simulated ectopic beats. The coupling was evaluated before and after contamination, as well as with different number of ectopic beats. The results show that despite significant differences (Kruskallwallis, p<0.05) before and after contamination, the visual trends suggest that the robust parameter is more robust compared to subspace projections.