Yuku free message boards

I think this is where the different backgrounds and types of training will collide haha. I'm in psychology so our application of these things is likely different.

I'm not following any of the "factor analysis" stuff thus far. I only use factor analysis to try to reduce a multi-item scale into a single workable number or set of numbers (like a mean score on a factor or factors). So if you have a 25 item scale of aggressive behavior you can use factor analysis to determine the factor structure of the items and determine which ones tend to hang together (usually with the intent of demonstrating that they all load on a single factor of aggression). Then you can do a confirmatory factor analysis to test that model against other factor structures (2 factors, 3 factors, etc.). I wouldn't use anything random for either of those.

I also don't know what the application for Edgic would be, necessarily, unless we are unsure of the variables to be used for analysis. Then again, I probably just don't use it that way in my field. All of my main analyses are more regression or structural equation model based. Perhaps "winner" is the factor we are looking for all of this to load on somehow? In any case, in my own thinking of it, there are several ways to put this together.

My original conceptualization had every episode's tone, rating, and visibility as 3 variables. 15 or so episodes per season x 3 variables was about 45 meaningful variables for analysis (and the individual scores for each contestant provided the data for each variable). However, I eventually scrapped this because episode 7 in Borneo is the merge and episode 7 in Nicaragua, for example, is not. Furthermore, I don't see how it would be valuable to know that no one that has received a MORM2 in episode 5 has ever won the game. Too specific.

The last time I sat down to do it I had the following idea: 15 or so datasets based on episode number. This mirrors the actual Edgic process. At each episode, you have the current episode's information, as well as all prior episodes. Therefore, the dataset for episode 1 would only contain episode 1 information from all seasons. The dataset for episode 5 would have data from the first 5 episodes of all seasons. The goal is/was to look at predictive models over time and see how they changed and what types of information is most helpful at each time point.

Additionally, I decided to move to a more cumulative way of recording the data. Rather than deal with all of the nominal/ordinal/ratio issues of what these constructs are, I was recording a cumulative count of each. Instead of separate variables for each episode, people got a continuous score for each construct based on their cumulative ratings. This way, I would no longer be gaining information about which specific rating predicted the outcome in a given episode, but more of a sense of how all of the information gained to that point was predictive of a win. Again, this mirrors actual Edgic better, as we are not concerned with the fact that NaOnka got an N in episode 2, but more that she has gotten too much N overall.

In addition to that, I DID include more categorical entries for important points in the season. Each person's specific rating, tone, and visibility was recorded in the premiere, merge, and penultimate episodes (I remember once that someone hypothesized that a bad edit in the second to last episode is a death sentence).

These are the variables I was entering:

C-INV - Cumulative number of INV episodes to this point
C-UTR - See above except for UTR
C-MOR - See above except for MOR
C-OTT - See above except for OTT
C-CP - See above except for CP

C-Pos - Cumulative number of times where the person received a P edit (P = 1, PP = 2)
C-Neg - Cumulative number of times where the person received a N edit (N = 1, NN = 2)
C-Mix - Cumulative number of times where the person received a M edit (M = 1, MM = 2 - if such a thing were to ever exist)
T-Tone - Total number of tone ratings a person received thus far (Add up all tone letters, double letters = 2 points)
C-Tone - Cumulative overall tone of the character (PP = 2, P = 1, Neutral = 0, Mixed = 0, N = -1, NN = -2)

C-Vis - Cumulative visibility score for the season

E1 Rate - Rating received in episode 1 (OTT, CP, UTR, MOR, INV; categorical variable)
E1 Tone - Tone received in episode 1 (categorical variable)
E1 Vis - Visibility received in episode 1 (continuous variable)
M Rate - Rating received in merge episode (categorical variable)
M Tone - Tone received in merge episode (categorical variable)
M Vis - Visibility received in merge episode (continuous variable)
Pen Rate - Rating received in penultimate episode (categorical variable)
Pen Tone - Tone received in penultimate episode (categorical variable)
Pen Vis - Visibility received in penultimate episode (continuous variable)

Possible Dependent Variables:
Win - Did this person win? (0 for no, 1 for yes)
Win type - Based on the observation that there may be different types of winners. Didn't get this far but all of the traditional Edgic winners (Brian, Todd, Parvati, Rich, etc.) would go in one pile, while the WTF winners (Danni, Bob, Natalie, etc.) would go in another. Maybe more piles based on other theories.

Other ideas:
- Separate analyses for males/females
- Separate analyses for male/female winner in season
- Separate analyses for pre-/post-ASS, pre-/post-half ASS
Or to simply add all of the above as variables and use them as predictors.
- Swap ratings?
- Calculating P and N by only counting Ps and Ns or by counting P+M for P and P+M for N (as M suggests presence of both).

This approach still has a lot of problems though. The datasets would not exactly be independent of one another and the errors would not be random over time. Probably would have to do some sort of time-series analysis and be sure that it would not mess with the results. Along with this, the profile of each season changes. The merge is in a different episode in different seasons. Some seasons have tricky or fuzzy merges. Some merge with 12 players. Some with 10. Some with less than that. Some seasons have fewer episodes overall. Some don't have swaps. Some have multiple swaps. Keeping all of this straight without pissing off the software with a ton of "missing data" is going to be difficult.

Another major hurdle is how to transform/use incompatible data. Many early seasons either don't have visibility info, they are based on a 3 point scale, or the visibility information was collected AFTER the winner was known. All of these are a problem. To get consistency we would almost need to take Warrior's formula and just go through and take the strict outcome of the formula with no changes for ALL seasons. Good luck on finding someone to do that! Other issues don't really have much of a fix at all. There is no way to control for changing definitions of ratings, introduction of things like M and INV, possible influences of spoilers, or "they don't win so who cares?" throwaway ratings. In a way, we are also losing data by considering JUST what's on the chart. The charts don't always indicate situations where consensus about a rating was fuzzy or the presence of close calls. Being able to capture all of that would also help make better predictions. In any case, our "Limitations" section would be massive!

Anyway, just ideas. Hope we can put something together, even if it's this summer when people have less to do.

ETA: Guess I should more specifically weigh in on the best way to think about the variables like GiMG was talking about.

As you can probably guess from above, the only variable Edgic works with that I think is a continuous ratio scale is Visibility. I think the other two are too multifaceted.

For tone, I think the current system measures a few things: basic tone direction (Positive vs. Negative), tone complexity (1-dimensional P/N vs. multifaceted M), and tone presence (Neutral vs. all others). Each of these could be extracted from tone ratings. They would be highly interrelated but might provide distinct info. In it's raw form, though, I don't see it as a continuous variable. Just nominal.

Rating is also something I think is purely nominal. There are a few things in there that you could try to make into continuous scales but the way the ratings are done/justified makes them so difficult to pin down.

First, I think you have the INV vs. the rest distinction. With INV you have an absence of whatever construct "rating" is. With the rest, you have something there.
Another is being multifaceted vs. being 1 dimensional. UTR and OTT seem to be more 1-dimensional. MOR and CP seem to be more multifaceted.
Another one that is very common is visibility. Low visibility gets you UTR. Middle gets you MOR. High gets you MOR, OTT, or CP. I don't necessarily like it but that's the way it's always been so that's the data we have. This part of it IS a ratio scale because it's a copy of another ratio scale that's already being used.
Another way you can think of it is the amount of character development. This is the closest to being some sort of unique ordinal scale. INV < UTR < MOR < OTT/CP. Both OTT and CP are a high level of character development, but in different ways. If OTT is less development than CP because it's one dimensional, then why is it above MOR? If CP is more development than OTT because it's complex, why isn't MOR above OTT? Even then, I have been questioning what exactly MOR is in terms of character development. It's often given to semi-complex, mid visibility people AND narrators (0 personal development).
Finally, as alluded to, you can go based purely on complexity. Unfortunately that just gives you the same thing as the idea of some constructs being multifaceted and others not being multifaceted. In terms of complexity you have INV < UTR/OTT < MOR < CP. Still doesn't quite work.

In any case, I feel like that's gotta be a categorical and nominal variable. It's just too complex to unpack into a simple scale, IMO.