We study space-time-frequency coded multiple-input-multiple-output (MIMO) orthogonal frequency-division multiplexed (OFDM) systems employing antenna selection at the receiver where optimal training has been utilized for performing linear minimum mean square error (LMMSE) based channel estimation. The maximum signal power antenna selection rule is introduced. By upper bounding the pairwise error probability (PEP) expression, we show that the system with channel estimation error (CEE) still achieves full diversity. The Loss in Coding Gain (LCG) and the Loss in Performance (LP) expressions are derived. It is shown that while equal power training performs about 3 dB worse than the perfect CSI case, the optimal power training performs between the two depending on the number of subcarriers assigned per OFDM word. Simulation results are presented to validate our analysis.