0.
Some people advise you to round up 0.05 to 0.1. Others say round down. The problem with that approach is as follows:
In the first case
If the number is 0.00 you don't need to round
If the number is 0.01, 0.02, 0.03 or 0.04 you round down
If the number is 0.05, 0.06, 0.07, 0.08 or 0.09 you round up
So 4 downs and 5 ups (and one no change).
Net Result: more upward than downward roundings so you introduce an upward bias.
In the second scenario, you introduce a downward bias through rounding.
Possible solution:
Round up from 0.05 half the time, round down the other half. That satisfies the bias problem but introduces another - that of reproducibility. If someone else were to look at your data would they round up/down the same way you did? Unlikely. You need a system which will round 0.5 up half the time and down half the time but where the decision is made for you.
Hence the best solution is to round up or down so that the new last digit is even.
It is 5.2 when rounded to the nearest tenths.
65.6
It is then 52.8 when rounded to the nearest tenth
It is then 56.5 when rounded to the nearest tenth
It is 232.4 when rounded to the nearest tenth
It is then 20.2 when rounded to the nearest tenth
It is 0.8 when rounded to the nearest tenth
4.4
10775.9
2.7
54.4
28.6