Ah, I see BA, OBP, and SLG in advanced stats for pitchers

Upon closer inspection, Newcombe, Stratton, and Sheib posted great hitting lines those seasons as well...but clearly there's something else going on with the upper-tier pitchers...

Also, unless I missed something, you are not accounting for fielding normalization which would roughly be indicated by year.  Since low ERC# pitchers occur more frequently in the deadball era  and these same pitchers are negatively impacted by fielding normalization which is accounted for in salaty, I think that would explain the positive coefficient.

Why not try including IP/162 as an independent variable and allow salary to be the dependent variable? 

Also, I think there is some double-counting when you include BB/9#, OAV# and WHIP#.  I think you should exclude WHIP.  You should probably exclude ERA and ERC as well, since OAV, HR9, BB9 all inderectly factor into ERA and ERC.

Sorry to keep throwing out suggestions, but why not try running a regression on a similar subset of pitchers, such as pitchers from a single year and league (include partial and full seasons only - not combined).   This will help eliminate the noise.