I don't believe it correct to do that, I believe truncation is better.
simple example, 10 measurements of an analogRead()
The integer average with truncating
{ 109 109 109 109 109 109 109 109 109 108 } = 1089/10 = 108
The integer average with rounding
{ 109 109 109 109 109 109 109 109 109 108 } = (1089+5)/10 = 109
The float average
{ 109 109 109 109 109 109 109 109 109 108 } = 1089/10 = 108.9
(for above example)
with rounding the max error = 0.5 so the average error is 0.25
with truncating the max error = 0.9 the average error is 0.45
So there is a factor 1.8 difference in the max & avg error
However I do not know the requirements of your project, so I cannot say if rounding is correct or not.
// the numerical implication of error propagation in a rolling average are a bit more complex
// but it is a faster method than taking dozens of samples (when executed in integer domain)