How reliable is this actually? There are a millions reasons for me to fake which is and web browser in using.

Shouldn’t this effect cause Linux use to be under reported? That is, the real percentage would be higher than 4%?

I would pretty strongly expect significant correlation between people who spoof their user agent and Linux users.