The database (slightly updated) that I used for my statistical study is now online. It is quite large (44M, so a broadband connection is recommended.) This database (normbk03.pgn) has 199,375 games of master level (2300+ elo) players.
Only long time control human v human games. Removed include duplicates, rapid, blitz, blindfold, simultaneous, and Internet games. Also removed are variations, comments, nags and fens. Also games with defective tags were removed. Games of 20 moves (40 plies) or less were removed. All games have results and both elo values. The games are "clean" and "unique". They were played from 1991 to the present.
To download, go to crafty-chess.com (run by Peter Skinner),
downloads, Pollock, then pick normbk03.zip.
Direct link:
http://www.crafty-chess.com/down/Pollock/normbk03.zip