Who Should I Play On The Snatch Game?
Image by rawpixel.com on Freepik
The "Snatch Game" is a perennial challenge on the popular reality TV show, “RuPaul’s Drag Race”. Contestants impersonate celebrities in drag in a game-show format. The challenge is regarded as a make-or-break one, where quick-witted queens distinguish themselves in order to snatch the crown. In recent years, the cast of characters has exploded to include men, fictional characters, figures from ancient history, and anyone else that queens can play up for the camera.
Contestants can play anyone, as long as they are not under copyright. However, it’s not as easy to see what kind of character can win. For example, fans theorize that characters from the “Real Housewives” or “Golden Girls” are cursed. Similarly, fans think young contestants shouldn't pick viral stars, because RuPaul won't understand their references! To illustrate these theories and to ask other questions, I decided to make an R Shiny App. Take a look at the app here, and view the code here on GitHub.
I researched all of the snatch game episodes in the US series, All-Stars seasons, and English-language international seasons. Then, I created the dataset in Excel in three pages, divided up into episodes, queens, and characters.
Firstly, with one row for each episode, I include basic metadata, the number of queens, and their placements. While every season does placements somewhat differently, and runways factor in as well, I boiled them down to WIN, HIGH, SAFE, BOTTOM, and ELIMINATED, which I also assigned number values (1, .5, 0, -.5, -1).
In similar fashion, the second page features all the queens featured on each episode and their characters. Though most queens play one character, ambitious queens sometimes do two.
Then, the third page classifies each of the characters into a series of genres. At this point, the work was the most subjective, since queens are drawn to iconic people who defy categorization. How do I label someone like Grace Jones, a figure known for modeling, music, and acting, in this situation?
My solution was to do two overarching groupings. One was granular, grouping the characters into 30 categories, and assigning placements to any one that seemed plausible. Then, I made a more-general grouping, featuring only ten groups. In the general grouping, I mostly placed characters into the one that fit them above all. I liked the granular categories for deep dives, but also having general categories to allow for some statistical testing. The approximately normally-distributed results of the placements shown below inspired me to continue asking these sorts of questions.
Bar Chart of All Placements in Dataset: x= placement, y = count
Soon after processing my data, I was able to begin plotting. You can plot two categories simultaneously and compare them, including seeing the number of times they have featured, their average score (between -1 and 1), and breakdown by percentage of placements as well.
These two categories, GG for Golden Girls, and RH for Real Housewives, help confirm fan theories that these characters have never won the Snatch Game, and are, indeed, cursed!
Plotting the specific categories was satisfying, but it was difficult to get any meaningful statistical testing from them. Either the sample sizes were too small, or the larger groups simply mirrored the normal distribution of all the placements. Certain hyphenates like "Singer-Actor" and "Actor-Comedian" tended to go together a lot, so the observations I was trying to test weren't actually independent.
The general categories did work for some testing though. In order to add some heft to my analysis, I included the unadjusted p-values from pairwise t-testing between the categories into the app. For any of these intersections between the general categories that are less than .05, this means that the difference in these groups' performances is statistically significant. So if you're on the fence between two of these groups and the p-value is less than .05, you should really give them a closer look!
Dead or Alive?
The last bit of analysis I included in the app was to ask whether living or dead people performed better. I hypothesized that the dead would do better, due to queens feeling more freedom to be exaggerated. RuPaul would also be more likely to know the character and understand jokes and references in the character's style.
The alive outnumber the dead, but a greater percentage of the dead place high and win. The p-value of Welch Two-Sample T-Test currently sits at .0222, a statistically significant significant. So if you're not sure how to win, play dead!
To conclude, there are several directions I could take this app in the future. I’d like to introduce more filtering options and conditional formatting. Then, I'd like to expand my dataset to include non-English language international franchises.
Thanks for reading. Check out my study list for the NYT Crossword Puzzle!