How optimal subsampling depends on guessed parameter values