Strategic Predictions for ‘Survivor: Vanuatu’

It’s Prof. Sadow here, trying again my exercise in semi-futility to use statistical models to pick placements in Survivor, this occasion of course for Survivor: Vanuatu (nice place, I hear, with an interesting colonial history reflective in its government today, but only have flown over it a couple of times when visiting some of its neighbors). As more and more cases become available, the model should get better and better (not including All-Stars because the dynamics were too dissimilar to the past series’).

Whether improvement has occurred is a matter of some doubt. Using an average absolute deviation for the predicted and actual results of Pearl Islands, the figure was 4.6875 (meaning that’s how many places on average were missed) although some were pretty good picks (four perfect, essentially) but with several bad misses (including predicting winner Sandra to come in 13th). Let’s see how things go this time.

Taking the results from Pearl Islands and reanalyzing (look here and here for the past theory and an explanation of how I build the model) shows little predictive improvement in terms of variance explained (in fact, a slight decrease to 18.3%). More interestingly, the predictive variables changed again for both the discriminant and OLS models. For the former, race seemed to disappear from having predictive ability (probably because the prior model weighed higher results in favor of white players, and Hispanic Sandra messed with that by winning).

To summarize, a stepwise procedure in discriminant analysis yielded the variables of age, openness of dress, and lack of wearing jewelry (the latter two from the biographical pictures posted by CBS) as significant predictors of game position. The underlying theory is that older players play more wisely, more colorful and flowing clothes denotes a flexible personality to better adapt and to adjust to stress, and more jewelry denotes a superficiality or attempt to convey an image which usually does not work to the advantage of that player (note: garish jewelry only counts for women; any kind for men).

For the regression, the two appearance variables plus race were important (the latter still significant despite Sandra’s win, it would seem), as well as a new one: past demonstrated leadership ability, on the basis that past leaders bring this quality that is positive to their success to the game. Not that this model did very well; although significant, it explained only 12 percent of the variance. For the record, it looked like this:

Predicted finish = 16.572 + 1.145*race – 1.667*lead – 1.745*clothes – 2.202*jewelry

As done previously, first, discriminant categories will be assigned, and then within categories the regression model will be used. However, the scenario has changed for this edition, with the players numbering 18 rather than 16, because to discriminate into quartiles no longer makes sense, since 18 is not divisible by 4. Instead, the discriminant analysis will try to place the competitors into three categories, a non-jury (places 10-18), earlier jury (places 5-9) and finals (places 1-4). As indicated above, predicted finish (1-18) was the dependent variable for the regression equation.

Below are the results of the two analyses. Predicted placements below are listed by the category predicted by the discriminant analysis (which overweighed into second), then by placement predicted by regression, then how well the two matched (where a positive score meant discriminant picked a lower group than did regression), a revised categorization, and finally a revised placement that first ordered contestants by groups, then within groups:

Even the casual observer can see despite many variables that the two equations produced some rather conflicting results. In essence, two competitors were two whole categories off, and only 7 of 18 fell into the same category in each equation. Thus, some judgment calls had to be made. By way of example of this, the discriminant analysis put Scout in the area designating finalists, but regression pegged her to be the worst finisher. Without presenting the map, Scout’s position was borderline between the finalist group and the non-jury group, so in terms of a finish I would judge her, given the regression results, to be placed in the non-jury group.