The traditional approach for producing spectral estimates from experimental data is the Fast Fourier Transform. However, in certain circumstances this approach is inappropriate due to factors such as non-stationarity or non-linearity. In this paper, the authors apply the auto–regressive time series modelling approach to produce spectral estimates of two such problems — non-stationary data obtained from the large amplitude response of a cable stayed bridge to wind excitation and non-linear data obtained from modal testing of cracked reinforced concrete beams. Although the auto–regressive approach is very sensitive to the model parameters chosen and so should always be used with caution, the results of these studies have produced time varying spectral data that give insight into the problems addressed.