I wrote back in May about the paradoxical nature of OPTA's possession statistic in MLS---how more possession corresponds to better shot ratios, better shot ratios correspond to better goal differentials, but somehow more possession does not correspond to better goal differentials when we control for certain variables. In fact, I found that once I controlled for the teams playing in a given game, possession had a negative correlation with goal differential and winning. The new data agrees with the old. Correlations suggest that team possession still correlates positively with scoring attempts (p-value = 0.01), scoring attempts still correlate positively to goal differential (p-value = 0.02), and now with more data, possession is also positively correlated to goal differential (p-value = 0.01). That all seems to line up with logic, but the paradox from before still exists.
When I look game-by-game and control for the home and away teams, in-game possession has a positive correlation to shot ratio, but a negative correlation to goal differential. In other words, the team that has more possession in a given game tends to also earn more shot attempts, but still loses more frequently than we would expect. As mentioned in the first article back in May, this seems paradoxical. I had some theories in that article, but reader David Stringer got me to think about another logical explanation.
Teams that develop leads tend to sit back more defensively, and often are satisfied allowing the opponent to possess all it wants in less dangerous parts of the pitch. A team that has a lead in the second half probably got that lead because it was generating more opportunities (read: attempts). It makes sense that the team that eventually went on to win also produced better shot ratios early on before getting the lead. After getting the lead, the team in front was willing to give up extreme possession relative to a more neutral shot rate. Thus it ends the game with poor possession, but a still favorable shot rate.
Just a theory, and I'd love to hear about other ideas! The stats are definitely not lying. These correlations are very real, but the causes for the possession paradox are still elusive.