We address the problem of estimating doubly-selective channels using pilot clusters that are time-division multiplexed with the data. Channel estimation is carried out using different basis expansion models (BEMs), and direct MMSE channel estimation using the channel statistics. For a fixed number of pilot symbols, we attempt to optimize the power and placement of the pilot symbols used in transmission, and the number of BEM coefficients used in channel estimation, in the sense of minimizing the mean-square estimation error (MSE) that includes modelling error, which is normally neglected in existing work. Simulation results confirm that for a wide range of SNR and Doppler spread values, this optimization greatly reduces the MSE and the bit-error rate, and that modelling error should be taken into account when optimizing training. The effects of uncertainly in the channel statistics are also studied.