I think you might be off by a factor of 10 here Joe. We have 39xx games in the database (call it 4000). So #300 is top 7.5%, and 1285 is top 33%. Still not bad, as some people's favorite game is Bingo, which would put them in the top 99%

Different strokes for different folks, just because PR is #1 doesn't mean nobody rates it a 1

People say statistics lie - I don't believe that and I hate when people say that - statistics are just numbers.

...IMO the only thing that the disparity in the quantity of votes for a single title does is improve the Confidence Level for the rating of that game. If you look at the the Top 10, the sample sizes are so small and confidence intervals so large above game rank 5 that the 95% confidence levels overlap and you'd be just as right if you pulled the names from a hat to rank them. In this case, attempting data conversions and ANOVA, the P-Values are too large to infer any significance in the ratings without taking massive amounts of alpha risk. The above analysis is limiting sample size further by trying to equalize them. I for one stand firmly in the camp that says, the larger the sample size behind the rating, the better.

Also, looking at any of the games in question, the histograms show the data to be very non-normal and thus any analysis should treat the data as such. The fact that the game ratings are tied to the mean of a non-normal data set is of itself flawed. Even is that mean is adjusted.

I've run a number of multiple regression analysis on a random 100 game sample data set from the top 1000 games. I've yet to get any useful prediction model from the resulting analysis due to high multicollinearity (VIFs), poor R-square, or non-normal distribution of residuals. (In some instances, residuals that are way out compared to the sample population.) Not that that's stopped me from continuing to look for the factors responsible.