Last time I took a look at which BGG mechanics were the most published, and how that list has changed over time. This time, I'm interested in which mechanics generate the highest ratings on BGG.

A simple way to do this would be to compute the average rating across all games that contain each mechanic, then rank the mechanics. However, this could cause us to confound the effect of a mechanic with the effect of other, correlated factors. For example, if games with **card drafting** usually also have **hand management**, then the effect of the two mechanics would be confounded. To deal with this, I'll be using **multivariate linear regression** to estimate the independent effect of each mechanic on a game's rating.

Since there are a lot of mechanics to consider (51 in total), I also wanted a method that automatically selects the mechanics that seem to matter the most. To that end, I'll be using the **LASSO** method.

Finally, I'll also be throwing in a few other factors that may influence user ratings: the game's weight, its playing time, and indicators for whether it's a wargame, party game, or abstract strategy. The regression is run on all non-expansion games published between 2004 and 2014, with at least 100 user ratings. The dependent variable is BGG's Bayesian rating, which is what BGG uses to construct board game rankings.*

*The purpose of the Bayesian rating is to push games with fewer user ratings towards the middle. As the number of user ratings becomes large, the effect of this Bayesian averaging becomes minimal. A slightly more detailed explanation can be found here.

Without further ado, here are the mechanics that came out of the LASSO regression and their effects:

**The Effect of Mechanics on BGG Ratings**

Rank | Mechanic | Effect |
---|---|---|

1 | Grid Movement | 0.208 |

2 | Player Elimination | 0.204 |

3 | Worker Placement | 0.154 |

4 | Card Drafting | 0.122 |

5 | Variable Player Powers | 0.111 |

6 | Co-operative Play | 0.106 |

7 | Simultaneous Action Selection | 0.083 |

8 | Set Collection | 0.061 |

9 | Hand Management | 0.051 |

10 | Deck / Pool Building | 0.051 |

11 | Area Control / Area Influence | 0.039 |

12 | Partnerships | 0.027 |

13 | Route/Network Building | 0.023 |

14 | Dice Rolling | 0.011 |

15 | Campaign / Battle Card Driven | 0.009 |

16 | Tile Placement | 0.008 |

17 | Point to Point Movement | 0.004 |

18 | Area Movement | 0.002 |

19 | Simulation | -0.003 |

20 | Trading | -0.005 |

21 | Roll / Spin and Move | -0.040 |

22 | Hex-and-Counter | -0.043 |

*How to read this table: The "effect" can be understood to mean the amount by which a game's BGG rating will increase (out of 10) if it had this mechanic. Of course, this estimate is only "local" to the data in the sense that we can't extrapolate too far from the kinds of games we actually observe. Mechanics not selected by LASSO can be interpreted either to have a zero effect, or to be contained by too few games to have a meaningful impact on the data.

As for the non-mechanic factors, here's what LASSO gave us:

**Other Determinants of BGG Rating**

Rank | Variable | Effect |
---|---|---|

1 | Intercept | 5.368 |

2 | Game Weight | 0.252 |

3 | Party Game | 0.023 |

4 | Wargame | -0.010 |

*How to read this table: The "intercept" can be understood to mean the expected BGG rating of a game that has no listed mechanics, a 0 weight, and is not a party, war, or abstract strategy game. The effect of game weight implies that a 1-point increase in weight (out of 5) predicts a 0.252 increase in BGG rating. The effect of being a party game is a 0.023 increase in rating, while the effect of being a wargame is a 0.01 decrease in rating.

**Discussion**

It's reassuring that the mechanics LASSO picked out conform to what I view as some of the more common and popular mechanics in boardgaming (i.e. it's good that Crayon-Rail System wasn't picked out.) Among the highly rated mechanics, I was not surprised to see **Worker Placement**, **Card Drafting**, **Cooperative Play**, and **Variable Player Powers** all there. I was somewhat surprised to see that **Grid Movement** rates so highly. I wonder why? I was also surprised about **Player Elimination**. I always thought that player elimination was considered undesirable; but this does conform to results from my previous post suggesting that player elimination is one of the fastest growing mechanics in terms of number of games published. Finally, I was surprised to see that **Variable Phase Order** (Puerto Rico; Race for the Galaxy) was not selected by LASSO. I previously showed that it is one of the fastest growing mechanics over the past decade, but I suppose it still doesn't have enough absolute numbers.

Only four mechanics scored negative effects, though it should be mentioned that negative is completely relative here. Out of those four, two stand out: **Roll/Spin-and-Move** and **Hex-and-Counter**. This conforms to results from my previous post showing that these two mechanics declined the most in terms of number of games published over the past decade. I'm not surprised that the roll-and-move mechanic is one of the worst rated mechanics. I'm not sure what to think about hex-and-counter as I've had limited experience with it.

Moving on to the non-mechanic factors, we see that game weight is clearly the most important. A one-point increase in game weight predicts a 0.25 increase in rating. This is understandable, as many users think of game weight as a measure of strategic complexity or decision density. I've always wondered whether the weight-to-playing-time ratio would be a strong predictor of rating. Maybe I'll test that some time. The effect of being a party game or a wargame is small, and abstract strategy dropped out of the LASSO regression altogether.

**Conclusion**

The results show us that some of the fastest growing mechanics over the last decade are also the most highly rated ones, and vice-versa with declining mechanics. A limitation of the current approach is that I've not allowed for potentially important interaction terms. For example, perhaps hex-and-counter would hurt the rating of a non-wargame but would increase the rating of a wargame. If that were true, then the current estimate for hex-and-counter confounds these two effects. A more highly interacted model, however, would introduce many more terms and then variable selection becomes even more of an issue! **Dimension-reduction** (i.e. a simpler categorization scheme) seems like it's going to be important, especially for work on the relatively small dataset that is boardgaming. I pointed to some discussion of these issues in my previous post, and I think I will soon be exploring those topics as well.

What an awesome blog about gaming. I am also trying to explore a little bit the BGG database using some statistics and machine learning tools, but it has being tough to download the data and filter out the bad annotation.

I will take a look at your other posts to see if you have solved some of the problems I have 🙂

Thanks! Check out the Data Mining guild on BGG. There are some old posts floating around about how to get the data.

https://www.boardgamegeek.com/guild/2128

I tried to do a similar analysis, of high and low ranked games, using the top 200 and the 1000-1400 games. Instead of only mechanics I included all features I could use from BGG. I used a decision tree to classify them. I couln't make so cool analysis of my results and show a so nice output as you did here.