Identification of continuous-time autoregressive processes from discrete-time data by replacing the differentiation operator by an approximation is considered. A linear regression model can then be formulated. The least-squares method and the instrumental variables method must be used with some care to get parameter estimates of good quality. The bias is studied explicitly in the paper together with the asymptotic distribution, and expressions are presented for the covariance matrix of the estimated parameters. It turns out that there are small differences in the dominating bias term for the different methods, whereas the statistical properties are comparable. Overall, the performance is similar to that of a prediction error method for short sampling intervals.