by a_traveler
I believe the subcategory ratings only look at votes by users who have assigned it to one of the subcategories, a significantly-more involved process that is used by a much-smaller portion of the BGG membership. In other words, the pool of votes that is being used to determine a subcategory rating is much smaller than the pool of votes being used to determine the overall rating.
I don't think we know for certain one way or another, but this is almost certainly not correct: the number of people voting for which subcategory a game falls in is several orders of magnitude lower than the number of voters for most games, making the math not work out.
Instead, the leading hypothesis is that the number of "5.5" dummy votes used to determine the geek rating (~750 per game for games in the main ranking) varies per subdomain (from as low as ~100 per game for party games, to ~500 per game for strategic games).
(Numbers derived from looking at a sampling of 'non-controversial' games; for example
Concordia has 5431 votes with an average rating of 7.86. with 750 dummy votes, its geek rating comes to 7.574, extremely close to its real geek rating of 7.575 (5431×7.86+750×5.5)÷(5431+750); in the strategy subdomain its geekrating is 7.649, and the calculated geek rating with 500 votes is 7.661, close enough for demonstration purposes.)
Then on top of that, it's expected that the secret shill busting algorithm (which is run once a month (?), 'discarding' outlier votes (hence the big jump in geek rating for this game on Jan 1st; all the hate ratings from the last month were discarded then)) is extended for the subdomains to also weigh votes less from people who've
rated very few other games in the same subdomain. (So if you've rated only one party game, then your opinion of that game is not a good indicator of how it compares to other party games.)