Hey everyone,
robin was kind enough to provide me with a lot of game details from the Pinside database. I wanted to review the data and perform a simple regression analysis to see if there are any features of pinball machines that can be directly tied to how the game rates in the Pinside system.
I’ve attached an image of the Excel regression output, but I’ll do my best to explain and keep things simple.
First, I only reviewed an extract of solid state titles, so EM machines did not factor into the analysis.
Second, because I was trying to see if a variety of independent variables impacted Pinside ratings, only those games with a rating in the system could be used. This reduced my count from the total SS listing of 1,039 to 371.
Third, I was just using Excel, not something fancy-pants like SPSS. Excel limits the independent variable count to 16. This means I could only consider 16 categories of information (just as well, since the risk of noise increases with every variable added; those of you more into statistics know about Adjusted R Square and all that). So, I had to pick and choose what to include.
So, what did I consider? The following categories were assessed for impact on Pinside rating:
1: Year (listed year of manufacture)
2: If the game was a Williams
3: If the game had a licensed theme
4: If the game was a designated wide body cabinet (versus say a standard or custom size)
5: If the game has a kickback feature
6: If the game has a multi-level playfield
7: If the game has any spinners
8: If the game has a shaker motor
9: The number of flippers the game has
10: The number of balls in a multiball that the game offers (the largest count possible)
11: The number of drop targets
12: The number of bumpers
13: The number of ramps
14: The number of captive balls
15: The number of cannons
16: The number of magnets
For items that were unknown (for example, having a shaker motor has three options in the data, Yes, No, and Unknown), they were treated as if they did not exist (assumption of “No” for Unknown entries).
Next, I had to set some sort of ground-rule of significance, where we decide whether or not the collection of these variables is predictive, and whether each individual one is predictive. I went with a p-value of 0.01 (so, p-values less than 0.01 were accepted as significant, and anything larger was rejected; this value suggests a 99% probability that any reported relationship is real and not merely due to chance).
The Regression’s Significance F field was well under 0.01, which means we get to say these 16 variables do predict a game’s Pinside ranking in some way. So, we look at the R Square value, which is 32.6%. The regression indicates that of a Pinside rating’s value, 32.6% of its total variance is explained by these 16 variables. Or, said another way, around 1/3 of a game’s rating ties to these objective details, and the rest of the rating is determined some other way.
Okay, since we know the collection is significant, we can look at each of the 16 variables and see which ones are individually significant. Of the 16 variables I examined, 5 came in as significant at our 99% level:
1: The year of manufacture
2: If the game was made by Williams
3: The number of bumpers on the game
4: The number of ramps on the game
5: The number of magnets the game uses
A shaker motor was close, but didn’t quite make the threshold (There is a 95%+ probability that a shaker motor’s presence improves a game’s Pinside rating, but since it couldn’t reach 99% probability I’m rejecting it from consideration).
Now, if you wanted to explore any more, you can use the coefficients to see what the regression is estimating from the image. I started to write some examples on how this works, but it turned into a huge paragraph so I’ll just leave it out there as an option for those ultra-curious.
Broadly speaking, this regression analysis suggests that: the more modern the game, the more Pinsiders like it. That a game made by Williams gets a boost in rating. And, the more bumpers on a playfield, the more ramps on a playfield, and the more magnets a game uses, the higher the game’s rating. But, that all these features only really explain roughly 1/3 of a game’s rating.
Again, my thanks to robin for allowing me access to his stored data sets. It’s been a while since I did an analysis like this, so if anyone catches errors just let me know and I’ll edit the post to be more accurate. I’ve been curious for quite a while if the objective features of pins can actually predict how they are rated on this site, and it appears at least a few of them do tie in.