Principal Component Analysis (PCA) has been often used for HRTF compression and individualization. However, there is significant variation in how the input matrix on which PCA is applied is constructed. Here, we study the effect of choices on the selection of independent variables, the domain in which impulse responses are represented, the HRTF database used, and possible smoothing on the compression efficiency and the reconstruction quality of the resulting PCA model. Several findings replicate well across different databases. Results point to a benefit for signal compared to space PCA and for using minimumphase HRIRs or HRTFs. Smoothing HRTFs leads to an increase in compression efficiency and a reduction in spectral distortion and using HRTFs with logarithmic magnitude leads to lower spectral distortion compared to linear.