When you using 4 lines with 16 cols, last two lines (3 and 4) are moved by 4 chars. It's because there is hard-coded move by 0x14... But that would work for 20x4 displays. For 16x04 memory map is arranged different way.

source: http://web.alfredstate.edu/faculty/weimandn/lcd/lcd_addressing/lcd_addressing_index.html

This fix is only for 16x04. Regarding to source web pages there could be more logic for other memory maps too (like 40x4 etc...)