Because numerical overflow can be useful. You basically calculate the value x mod(High(word)+1).

Code:
var
  x : word;    //16 bit unsigned value
  y : longword;  //16 bit unsigned value
begin
//16 bit
x:=65535;   //upper limit for word
x:=x+10;   //x overflows and is now nine

//32 bit version of above code
y:=65535;
y:=(y+10) mod 65536;  //also nine thanks to mod operation
Hmm... I'm not very good at explaining this. But you see that the 16-bit calculation is simpler and possibly faster.